Introduction to

OFDM Receiver Design

and Simulation

6899_Book.indb 1 10/22/19 4:54 PM

 For a listing of recent titles in the
Artech House Mobile Communications Library,
turn to the back of this book.

6899_Book.indb 2 10/22/19 4:54 PM

 Introduction to
OFDM Receiver Design
and Simulation

Y. J. Liu

artechhouse.com

6899_Book.indb 3 10/22/19 4:54 PM

 Library of Congress Cataloging-in-Publication Data
A catalog record for this book is available from the U.S. Library of Congress

British Library Cataloguing in Publication Data

A catalog record for this book is available from the British Library.

ISBN-13: 978-1-63081-738-1

Cover design by John Gomes

© 2020 Artech House

685 Canton St.
Norwood, MA

All rights reserved. Printed and bound in the United States of America. No part of this
book may be reproduced or utilized in any form or by any means, electronic or mechanical,
including photocopying, recording, or by any information storage and retrieval system,
without permission in writing from the publisher.
All terms mentioned in this book that are known to be trademarks or service marks
have been appropriately capitalized. Artech House cannot attest to the accuracy of this
information. Use of a term in this book should not be regarded as affecting the validity
of any trademark or service mark.

10 9 8 7 6 5 4 3 2 1

6899_Book.indb 4 10/22/19 4:54 PM

 To my family and friends for their love and support

6899_Book.indb 5 10/22/19 4:54 PM

6899_Book.indb 6 10/22/19 4:54 PM
 Contents

Preface xv

1 Discrete Time Signals and Discrete Fourier

Transform 1

1.1 Introduction 1
1.2 Discrete Time Signals 2
1.3 Fourier Series Representation 2
1.4 DFT 5
1.5 Sampling Theorem and Signal Interpolation 7
1.6 Properties of the DFT 8
1.7 Time and Frequency Relationship of the DFT 10
1.8 Operations of the DFT 12
1.8.1 Linearity 12
1.8.2 Time Shift 12
1.8.3 Frequency Shift 13
1.8.4 Circular Convolution 13
1.9 Z-Transform 14

vii

6899_Book.indb 7 10/22/19 4:54 PM

 viii Introduction to OFDM Receiver Design and Simulation

1.10 Summary 16
References 17

2 Single-Carrier Modulation 19

2.1 Introduction 19
2.2 Data Transmission Rate 19
2.3 Bandpass Signals 20
2.4 Digitally Modulated Signals 21
2.4.1 PAM 22
2.4.2 PSK 22
2.4.3 QAM 24
2.4.4 FSK 26
2.5 Pulse Shaping 27
2.6 Summary 30
References 31

3 Multicarrier Modulation 33

3.1 Introduction 33
3.2 OFDM Waveform 34
3.3 OFDM Characteristics 35
3.3.1 Orthogonality 35
3.3.2 OFDM Spectrum 36
3.3.3 PAPR 37
3.3.4 Cyclic Prefix 39
3.3.5 FFT 40
3.4 Time and Frequency Parameters 41
3.5 Window Function 42
3.6 Digital Modulation 44
3.7 OFDM Waveform Properties 45
3.8 Summary 46
References 47

6899_Book.indb 8 10/22/19 4:54 PM

 Contentsix

4 OFDM Transmission 49

4.1 Introduction 49
4.2 OFDM Transmitter Architecture 49
4.3 Signal Transmission Format 50
4.4 Preamble Signal for IEEE 802.11a 51
4.4.1 Short Sequence 52
4.4.2 Long Sequence 54
4.5 IEEE 802.11a Header Format 56
4.6 IEEE 802.11a Data Format 58
4.7 OFDM Receiver Architecture 58
4.8 Summary 59
Reference 60

5 Shift Register Sequence and Data Scrambler 61

5.1 Introduction 61
5.2 Binary Field 62
5.3 Galois Field 64
5.4 Sequence Generator 66
5.5 Period of Sequence Generator 70
5.6 Maximum-Length Sequences 72
5.6.1 Properties of the Maximum-Length Sequence 72
5.6.2 Sequence Generator from the IEEE 802.11a 73
5.7 Data Scrambler 74
5.8 Summary 75
References 76

6 Radio-Wave Propagation Model 77

6.1 Introduction 77
6.2 Large-Scale Propagation Model 78
6.2.1 Free-Space Propagation Loss 78

6899_Book.indb 9 10/22/19 4:54 PM

 x Introduction to OFDM Receiver Design and Simulation

6.2.2 Two-Ray Model 80

6.2.3 Empirical Model 81
6.3 Small-Scale Propagation Model 83
6.3.1 Time Dispersion 84
6.3.2 Frequency Dispersion 86
6.3.3 Clark’s Fading Model 89
6.4 Receiver Signal-to-Noise Ratio (SNR) 91
6.4.1 Thermal Noise 91
6.4.2 Noise Factor 92
6.4.3 Amplifier Model 92
6.4.4 Cable Loss Model 93
6.4.5 Equivalent Noise Temperature at Receiver
Front End 94
6.5 Range Determination 96
6.6 SNR 98
6.7 Summary 98
References 100

7 Error-Correcting Codes and Interleaver 101

7.1 Introduction 101

7.2 Linear Block Codes 102
7.2.1 Generator Matrix 102
7.2.2 Parity Check Matrix 104
7.2.3 Syndrome 105
7.2.4 Error Correction 106
7.2.5 Hamming Codes 107
7.3 Cyclic Codes 108
7.3.1 Generator Polynomial 108
7.3.2 Syndrome Polynomial 112
7.4 Convolutional Code 113
7.4.1 Convolutional Encoder 114
7.4.2 Convolutional Decoder and Viterbi Algorithm 117

6899_Book.indb 10 10/22/19 4:54 PM

 Contentsxi

7.4.3 Convolutional Code in the IEEE 802.11a 122

7.4.4 Punctured Convolutional Codes 124
7.5 Interleaver 126
7.5.1 Illustration of an Interleaver 126
7.5.2 Interleaver Used in the IEEE 802.11a 127
7.5.3 Deinterleaver Used in the IEEE 802.11a 129
7.6 Summary 129
References 131

8 Signal Acquisition 133

8.1 Introduction 133

8.2 Direct Conversion to IQ Components 134
8.3 Detection Metric 136
8.3.1 Cross-Correlation 137
8.3.2 MMSE Metric 137
8.3.3 Normalized Cross-Correlation 138
8.4 Maximum Likelihood Detection 139
8.5 Coarse Timing Detection 140
8.5.1 Segment Detection 140
8.5.2 Sample Detection 144
8.6 Fine Timing Detection 147
8.7 Summary 150
References 151

9 Synchronization 153

9.1 Introduction 153

9.2 DC Offset 154
9.2.1 Algorithm Analysis 154
9.2.2 Simulation Examples 157
9.3 CFO 160
9.3.1 Algorithm Analysis 160

6899_Book.indb 11 10/22/19 4:54 PM

 xii Introduction to OFDM Receiver Design and Simulation

9.3.2 Simulation Examples 163

9.4 Frame Timing Offset 168
9.5 SCO 169
9.5.1 Algorithm Development 169
9.5.2 Lease Square Estimation in the Tracking Mode 172
9.5.3 Estimation in the Acquisition Mode 175
9.5.4 Estimation Under Special Conditions 176
9.5.5 Simulation Examples 178
9.6 IQ Imbalance 180
9.6.1 IQ Model 182
9.6.2 Estimation in the Acquisition Mode 183
9.6.3 Compensation of IQ Imbalance 186
9.6.4 Simulation Examples 186
9.7 Summary 191
References 192

10 Channel Estimation and Tracking 195

10.1 Introduction 195

10.2 Pilot Patterns 196
10.2.1 Block-Type Pilot Pattern 196
10.2.2 Comb-Type Pilot Pattern 198
10.3 Channel Estimation 199
10.3.1 Channel Estimation for the Block-Type Pilot
Pattern 199
10.3.2 Channel Estimation for Comb-Type Pilot
Pattern 201
10.4 Channel Tracking 211
10.4.1 The LMS Algorithm 211
10.4.2 The Condition for Convergence 213
10.4.3 Examples 215
10.5 Summary 218
References 218

6899_Book.indb 12 10/22/19 4:54 PM

 Contentsxiii

11 Data Decoding 221

11.1 Introduction 221

11.2 Demodulation 222
11.3 Hard Decision Decoding 222
11.3.1 Conventional Demapper 222
11.3.2 Simplified Demapper 224
11.3.3 Deinterleaver and Viterbi Decoding 233
11.3.4 Descrambler 233
11.4 Soft Decision Decoding 233
11.4.1 Transmitter and Receiver Block Diagram 233
11.4.2 Soft Decoding Using the Euclidean Distance 235
11.4.3 Soft Decoding Using LLR 235
11.4.4 Further Decoding Process 236
11.5 Summary 236
References 237

12 Simulation Study of a Multipath Channel on

OFDM 239

12.1 Introduction 239

12.2 Characterization of Multipath Channel 240
12.3 Computation of SNR 242
12.4 Transmitter Architecture for Simulation 243
12.5 Receiver Architecture for Simulation 244
12.6 Simulation Studies 244
12.6.1 Performance over AWGN 244
12.6.2 Performance over a Multipath Channel 247
12.7 Summary 251
Reference 252

About the Author 253

Index 255

6899_Book.indb 13 10/22/19 4:54 PM

6899_Book.indb 14 10/22/19 4:54 PM
 Preface

Orthogonal frequency-division multiplexing (OFDM) has been widely used

for Wi-Fi communications. It also has application in digital audio broadcast-
ing (DAB), digital video broadcasting (DVB), and the 4G Long Term Evolu-
tion (LTE) cellular system. Its major characteristic is to use many subcarriers
for data modulation. In comparison with a single-carrier scheme, the major
advantage is its ease in handling a dispersive fading channel. Currently, Wi-Fi
is provided in many places including homes, hotels, airports, and shops. A
variety of terminal devices including laptops, personal computers, cameras,
televisions, and media players have a Wi-Fi interface for internet access.
Wi-Fi digital communications are based upon several IEEE 802.11 stan-
dards. The 802.11a specifies a maximum data rate of 54 Mbps. The 802.11g
has a similar OFDM modulation to 802.11a. The 802.11n uses a combina-
tion of OFDM and multiple-input multiple-output (MIMO) technology to
transmit up to four spatial streams to achieve a peak data rate of 600 Mbps.
The 802.11ac is a further upgrade of 802.11n aiming at a data rate in the
gigabit range.
The goal of this book is to provide a fundamental understanding of the
receiver design applying OFDM technology. One distinct feature of this book
is the extensive simulation study of an algorithm. The simulation program
was written in C running on an Ubuntu Linux operating system. Because of
the simulation needs, a transmission waveform must be selected. In spite of
several IEEE standards mentioned above, the OFDM principle remains the

xv

6899_Book.indb 15 10/22/19 4:54 PM

 xvi Introduction to OFDM Receiver Design and Simulation

same. The transmission waveform specified in 802.11a was chosen. Once the
reader has an understanding of the design concepts based upon 802.11a, a
similar principle can be easily extended to other OFDM waveforms.
For example, the next generation cellular wireless digital communica-
tion is 5G (fifth generation), which is emerging. The 5G technology provides
much higher speed and is capable of connecting a massive number of devices.
The Third Generation Partnership Project (3GPP), the standards body behind
cellular 5G, has again adopted OFDM. Therefore, an understanding of basic
OFDM will help with stepping into the 5G era.
From Chapters 8 to 12, many simulation examples are provided. Besides
the simulation study, detailed derivations leading to the final formula for any
algorithm are given. In doing it this way, the reader can clearly understand the
approximations and conditions behind the formulas and apply them in the
correct way. If some important intermediate derivations are omitted, it is very
difficult to understand any algorithm. The emphasis is not on mathematics,
but includes enough to grasp the concepts behind the formula.
For any design topic used in the receiver, there can be many algorithms
in the literature. Only algorithms that were required for illustration and
comparison were provided to not confuse the reader. In most cases, only one
algorithm that is most promising from the simulation study was given. In
addition to the simulations, examples were provided in many cases to help to
understand the design subject. Therefore, this book is ideal for undergraduate
seniors, graduate students, engineers, and even professors who would like to
start understanding the OFDM technology.
This book is organized into 12 chapters. Chapter 1 provides the basics
of digital signal processing. We start with discrete time signals generated by
sampling an analog waveform. Next, Fourier series representation is given.
From its spectrum, the requirement for perfect analog signal reconstruction
is derived and that leads to the sampling theorem. The discrete Fourier trans-
form, inverse discrete Fourier transform, and an interpolation formula that
can reconstruct an analog signal based upon the discrete samples are then
subsequently derived.
The discrete Fourier transform has an important time and frequency
relationship and is presented. In addition, some operations including linear-
ity, time shift, frequency shift, and circular convolution are discussed. The
real-time signal has interesting symmetry properties that are illustrated.
Lastly, the z-transform and its usage to specify a time invariant linear system
are provided.
Chapter 2 reviews the single-carrier modulation techniques. The ana-
log modulated signal is first shown to be a bandpass signal. Its information

6899_Book.indb 16 10/22/19 4:54 PM

 Prefacexvii

carrying a baseband signal is further represented in terms of the in-phase and

quadrature-phase components. Based upon how the phase, frequency, and
amplitude are modulated, different modulation techniques are discussed.
They include phase shift keying (PSK), quadrature phase shift keying (QPSK),
16-QAM (quadrature amplitude modulation), 64-QAM, and frequency shift
keying (FSK). Pulse-shaping waveforms are also discussed. They include sinc
and raised cosine waveforms.
Chapter 3 introduces the multicarrier OFDM modulation. Through
inverse Fourier transform, the serial bit stream is converted into a multicar-
rier modulated waveform. Several OFDM waveform characteristics such
as orthogonality among subcarriers, cyclic prefix, fast Fourier transform
(FFT), classical spectrum, and peak-to-average power ratio (PAPR) are
explained in detail through derived equations. The modified rectangular
window used in IEEE 802.11a is presented. To show its spectral properties,
the discrete Fourier spectrum is derived. Lastly, the single-carrier modula-
tion technique in Chapter 2 is applied to modulate each subcarrier in the
frequency domain.
Chapter 4 presents the OFDM transmitter signal format, which can
be different depending upon the applications and the standard. To give the
reader some basic concepts, the format specified by the IEEE 802.11a is
followed. First, a general transmitter and receiver architecture is discussed.
Next, a preamble signal that aids in signal detection and synchronization is
given. For the IEEE 802.11a, it consists of both a short sequence and a long
sequence that are periodic in nature. After the preamble is a header, which
informs the receiver of the selected modulation method, and finally, is the
source data transmission.
Chapter 5 centers on the shift register sequence and the scrambler. The
source data from the transmitter can be scrambled for security protection.
Since the IEEE 802.11a specifies a pseudo-noise sequence for data scrambling,
its generation technique is given in detail. First, both a binary field having
only two elements and a Galois field having 2m elements are reviewed. The
binary field performs an operation of modulo 2 while a Galois field performs
a modulo operation over a primitive polynomial. Next, a sequence generator
through polynomial division to generate a periodic sequence with a desired
period is presented. The period obtained through an inverse polynomial is
illustrated. The maximum length sequence that generates a maximum period
using a primitive polynomial is also discussed. Having the periodic sequence
generated, the scrambler is just the exclusive-OR operation between the source
data bit sequence and the periodic sequence. The descrambler in the receiver
has just a similar operation.

6899_Book.indb 17 10/22/19 4:54 PM

 xviii Introduction to OFDM Receiver Design and Simulation

Chapter 6 discusses the radio-wave propagation loss through the media

and the receiver front-end noise. Both a large-scale model and a small-scale
model for the propagation loss are discussed. The large-scale model includes a
free-space model and a two-ray mode. An empirical model that can fit many
indoor and outdoor propagation losses is also presented. The small-scale model,
including both time dispersion and frequency dispersion, is given. The time
dispersion further discusses the flat channel and the frequency-selective chan-
nel, while the frequency dispersion discusses both the slow channel and the
fast channel. A Clark fading model with Rayleigh distribution for the received
envelope is also given.
After the propagation loss, the receiver noise is discussed to compute the
signal-to-noise ratio. The amplifier noise model, cable loss model, and thermal
noise from the receiver front end are all presented. The noise factor that can be
used to generate an equivalent temperature for these devices is included. An
equivalent temperature referred to the antenna through a cascade of ampli-
fiers, cables, and receivers is derived. Based upon the transmitter power, the
received power, and the loss prediction, the range estimate can be obtained.
Chapter 7 concentrates on error correction codes that are necessary to
correct the transmission errors from the channel. Two classes of codes are
covered: the block code and the convolutional code. For the block code, the
first is the linear block code, which covers the generator matrix, parity check
matrix, syndrome matrix, and error correction. The second is the cyclic code,
which operates on polynomials. The topics covered are the generator polyno-
mial, syndrome polynomial, and error correction. The Hamming code, which
is a special case of linear block codes, is also discussed.
For the convolutional code, both an encoder and a decoder using a Vit-
erbi algorithm are discussed. The convolutional code used in the IEEE802.11a
is also analyzed. Another special case is the punctured convolutional code,
which is derived from the rate 1/2 convolutional code. Both rate 2/3 and rate
3/4 punctured codes are illustrated in detail.
The last topic covered is the interleaver, which is also widely used to
decrease burst errors. The special interleaver and deinterleaver used in the
IEEE 802.11a are also illustrated.
Chapter 8 covers the OFDM signal detection using the preamble wave-
form. After direct IQ conversion, three detection metrics including cross-corre-
lation, minimum mean square error (MMSE), and normalized cross-correlation
are analyzed. A simplified algorithm from the maximum likelihood detection
is derived and justified through simulation. All three metrics are compared
using this algorithm.

6899_Book.indb 18 10/22/19 4:54 PM

 Prefacexix

The detection includes both coarse detection and fine detection. The
coarse detection is done in two steps. The first step is the segment detection
for fast detecting the periodic segment in the preamble waveform. The second
step is the sample detection to narrow down the detection error in samples,
followed by the fine detection to pinpoint the sample onset.
Chapter 9 deals with synchronization issues in the receiver. The topics
covered include direct current (DC) offset, carrier frequency offset, frame
timing offset, sampling clock offset, and in phase and quadratic phase (IQ)
imbalance. For the sampling clock offset, the formulas in the acquisition mode
using the preamble waveform and the tracking mode from using the pilot
carriers in the data symbols are demonstrated. All the issues were analyzed
and derived in detail without omitting any intermediate steps. The possible
approximations and conditions are clearly illustrated. Simulation examples
are also given to help to understand each topic.
Chapter 10 concentrates on channel estimation. Two types of pilot
patterns are analyzed. They are the block type and the comb type, which are
mainly used for slow- and fast-varying channels, respectively. The channel
estimate for the block type is though discrete Fourier transform (DFT). For
the comb type, it is though interpolation of a few pilot carriers inside each
data symbol. Using the Lagrange interpolation formula, linear interpolation,
parabolic interpolation, and cubic interpolation are discussed. The channel
can also be adaptively tracked through the use of the least mean square (LMS)
algorithm, which is derived in detail using the method of steep descent. The
condition for convergence is also derived.
Chapter 11 talks about the decoding process. After the channel effect
is removed, the data demodulation follows. Both hard decision decoding and
soft decision decoding are discussed. For the hard decision decoding, either
a brute-force approach or a simplified approach can be applied. The brute-
force approach has to compute the Euclidean distances for all the constella-
tion points. However, the simplified approach can determine each in-phase
and quadrature-phase bits individually by using a simple logic that is derived
for QPSK, 16-QAM, and 64-QAM. After the demodulation is complete,
deinterleaving, error correction decoding, and descrambling discussed in the
previous chapters follow. For the soft decision decoding, the received in-phase
and quadrature samples are combined with Viterbi decoding to generate cor-
rected output sequence. Both the Euclidean distance and the log likelihood
ratio (LLR) used for the distance metric are discussed.
Chapter 12 performs the simulation study of the OFDM performance in
a multipath channel. First, the time-varying multipath channel is characterized.

6899_Book.indb 19 10/22/19 4:54 PM

 xx Introduction to OFDM Receiver Design and Simulation

Then the transmitter and receiver architecture used for simulation are reviewed.
The simulation consists of two parts. The first part is the OFDM performance
in an additive white Gaussian noise, (AWGN) channel. The performance of
binary phase shift keying (BPSK), QPSK, 16-QAM, and 64-QAM are com-
pared using a 1/2 convolutional code. The study using punctured convolutional
code is also conducted. The second part is the simulation study of OFDM in a
multipath channel with the use of a 1/2 convolutional code. The channel also
has fading and a simulation model is provided. The block pilot pattern together
with LMS algorithm-channel tracking are of particular interest. Various fad-
ing strength and block pilot periods are used for performance comparisons.

6899_Book.indb 20 10/22/19 4:54 PM

 1
Discrete Time Signals and Discrete
Fourier Transform

1.1 Introduction
In digital communication, the input signal is digitized, processed, and modu-
lated before going through a digital-to-analog converter in order for transmis-
sion. For example, the input speech from a microphone is digitized first before
encoding and modulation. This analog signal can be perfectly reconstructed
if an appropriate sampling rate is specified.
Orthogonal frequency division multiplexing (OFDM) is a frequency-
domain modulation technique and will be discussed in later chapters in detail.
The major operation of OFDM is fast Fourier transform (FFT), which is a
fast computation method of discrete Fourier transform (DFT). In prepara-
tion for understanding OFDM, all the steps from discrete time signals to
DFT and sampling theorem are reviewed. The major operations of DFT are
also discussed.
The z-transform, which is useful in specifying the time-invariant system
response, is also briefly reviewed.

6899_Book.indb 1 10/22/19 4:54 PM

 2 Introduction to OFDM Receiver Design and Simulation

1.2 Discrete Time Signals

An analog signal is a time continuous waveform. If this signal is sampled
with a time period T, then a sequence of samples is obtained. The sampling
frequency fs is just the inverse of the T or 1/T. Figure 1.1 shows the analog
waveform x(t) and the sampled values at nT where n is an integer.
The sampling pulse is assumed to have a negligible width and is repre-
sented by the impulse delta function, p(t). The Dirac delta function δ(t) given
below is 1 at t = 0 and 0 elsewhere:

p(t) = ∑ d(t − nT ) (1.1)

Figure 1.2 shows such an impulse delta function. Note that p(t)is zero
except at the sampling instant nT. The discrete time signal x(t) is then repre-
sented by the following formula [1, 2]:

x(t) = x a (t) p(t) (1.2)

= ∑ x a (nT )d(t − nT ) (1.3)

where xa(t) is a corresponding analog signal. The discrete signal x(t) only has
the sequence of values x(T ), x(2T ), x(3T ), and so forth, assuming that the
sampling time is greater than zero.

1.3 Fourier Series Representation

Since p(t) is a periodic function with period T, it can be represented by a Fou-
rier series given below [2]:

p(t) = ∑ an e jnΩs t (1.4)

Figure 1.1 An analog signal x(t) sampled with period T.

6899_Book.indb 2 10/22/19 4:54 PM

 Discrete Time Signals and Discrete Fourier Transform3

Figure 1.2 Impulse delta function.

where ΩS is the radian sampling frequency defined as

2p
Ω s = 2pf s = (1.5)
T

The coefficient an is in general complex. From (1.4), it is clear that p(t) =

p(t + kT ) where k is an integer. Multiplying both sides of (1.4) by e −jmΩst and
integrating from −T/2 to T/2, we have
T /2 T /2 T /2
⎛ ⎞
∫ p(t)e − jmΩs t dt = ∫ am dt + ∑ ame − j(m−n)Ω t ⎟⎠ dt
∫ ⎜⎝ n≠m s (1.6)
−T /2 −T /2 −T /2

The second term on the right side of (1.6) integrates to 0 for m ≠ n. The
coefficient an is then given by
T /2
1
T −T∫/2
an = p(t)e − jnΩs t dt (1.7)

Substituting (1.1) into (1.7), we have an = 1/T. The impulse function in

(1.4) then becomes

⎛ 1⎞
p(t) = ∑ ⎜ ⎟ e jnΩs t (1.8)
⎝T ⎠

Substituting (1.8) to (1.2), the discrete time function x(t) is then given by

⎛ 1⎞
x(t) = ⎜ ⎟ ∑ x a (t)e jnΩs t (1.9)
⎝T ⎠

6899_Book.indb 3 10/22/19 4:54 PM

 4 Introduction to OFDM Receiver Design and Simulation

Figure 1.3 Band-limited spectrum of the analog signal xa (t).

Equation (1.9) is the Fourier series representation of the discrete time

signal x(t). Let f represent the frequency and then the analog radian frequency
Ω = 2π f. The Fourier transform of y(t)ejat is just Y(Ω − a). Assuming the Fou-
rier transform of xa(t) to be Xa(Ω), then the Fourier transform of xa(t)ejnΩst
becomes Xa(Ω − nΩs). By taking the Fourier transform on both sides of (1.9),
the frequency-domain relationship given by (1.10) is obtained:

⎛ 1⎞
(
X (Ω) = ⎜ ⎟ ∑ X a Ω − nΩs
⎝T ⎠ ) (1.10)

Equation (1.10) shows that the sampled spectrum is a summation of the

shifted spectrum of the analog signal by nΩs. In other words, the sampled spec-
trum is periodic in the frequency domain with period equal to the sampling
frequency Ωs. Figure 1.3 shows an analog spectrum with cutoff frequency Ωc.
Figure 1.4 shows the sampled spectrum under the assumption that
Ωs > Ωc. The spectrum is periodic with a period of Ωs. The analog spectrum
can be completely recovered without aliasing in the baseband. However, there
will be aliasing and spectral distortion if Ωs < Ωc. According to Figure 1.4,
the minimum sampling rate such that no spectral distortion occurs is to meet
the following condition:

Figure 1.4 Sampled spectrum of the analog signal xa (t).

6899_Book.indb 4 10/22/19 4:54 PM

 Discrete Time Signals and Discrete Fourier Transform5

nΩ s − Ωc = (n − 1)Ωs + Ωc (1.11)

After some simple algebra from (1.11), the minimum sampling rate is
given here:
Minimum of Ω s = 2Ωc (1.12)

Equation (1.12) is the well-known result and states that the minimum
sampling rate must be twice the cutoff frequency of the analog signal to com-
pletely reconstruct the original signal without any spectral distortion. The
sampling rate given in (1.12) is also the Nyquist rate.

1.4 DFT
The DFT [1, 2] is derived by first taking the Fourier transform on both sides
of (1.3) to have the following:
∞ ∞ ∞

∫ x(t)e − jΩt dt = ∫∑
− jΩt
x a (nT )d(t − nT )e dt (1.13)
−∞ −∞
−∞

Expressing the left term by X(Ω), we have

X (Ω) = ∑ x a (nT ) ∫ d(t − nT )e − jΩt dt

(1.14)
= ∑ x a (nT )e − jΩnT

Defining a radian frequency ω which is related to the analog frequency

Ω through the following relationship:
w = ΩT (1.15)

Using (1.15), (1.14) then becomes

X (w) = ∑ x a (nT )e − jwn

(1.16)
= ∑ xn e − jwn

where we have represented xa(nT ) by xn which is the nth sample of the analog
signal xa(t). Equation (1.16) is the Fourier transform of the discrete samples xn.

6899_Book.indb 5 10/22/19 4:54 PM

 6 Introduction to OFDM Receiver Design and Simulation

The Fourier transform X(ω ) is periodic with period of 2π . The ω can

be divided into N spectral components. Since ω ranges from 0 to 2π , ω can
be written as
2pk
w= k = 0,1,…, N − 1 (1.17)
N

where k is the kth radian frequency component and ranges from 0 to N − 1.

Substituting (1.17) into (1.16), we have
N −1
Xk = ∑ xne − j2pnk/N k = 0,…, N − 1 (1.18)
n=0

Equation (1.18) defines the DFT of the discrete time sequence xn. The
sequence Xk is periodic with a period of N as evidenced by the fact that
Xk = Xk+N. Therefore, there are only N unique values of Xk.
To obtain the coefficients x(n) from Xk, we multiply both sides of (1.18)
by the complex exponential ej2πmk/N and summing k from 0 to N − 1 and
obtain the following:
N -1 N -1 N -1
 X k e j2pmk/N =   xne - j2pnk/N e j2pmk/N (1.19)
k=0 k=0 n=0

Interchanging the order of summation on the right side of (1.19), we have

N −1 N −1
⎛ N −1 − j2p(n−m)k/N ⎞
∑ X k e j2pmk/N = ∑ n ⎜⎝ ∑ e
x ⎟⎠ (1.20)
k=0 n=0 k=0

Equation (1.20) can be simplified using the following relationship:

{
N −1
N for r = jN , j an integer
∑ e − j2prk/N = 0 otherwise
(1.21)
k=0

Equation (1.20) then becomes

N −1
1
xm =
N ∑ X k e j2pmk/N m = 0,…, N − 1 (1.22)
k=0

Replacing the index m by n, the following equation results:

6899_Book.indb 6 10/22/19 4:54 PM

 Discrete Time Signals and Discrete Fourier Transform7

N −1
1
xn =
N ∑ X k e j2pnk/N n = 0,…, N − 1 (1.23)
k=0

The sequence xn is also periodic with a period N and xn = xn+N. Therefore,

there are only N unique values of xn. Equations (1.18) and (1.23) constitute
the DFT pairs.

1.5 Sampling Theorem and Signal Interpolation

To derive the sampling theorem [1−3], we start with the following continuous
Fourier transform:

⎛ 1 ⎞
x a (t) = ⎜ ⎟ ∫ X a (Ω)e jΩt dΩ (1.24)
⎝ 2p ⎠

From (1.12), the minimum sampling rate is Ωs = 2Ωc. This gives Ωc =

Ωs/2 = π /T. In the baseband, the range of Ω is then −π /T < Ω < π /T from
Figure 1.3. From (1.10), Xa(Ω) = TX(Ω) inside this baseband and the follow-
ing equation results:
p/T
⎛T ⎞
x a (t) = ⎜ ⎟
⎝ 2p ⎠ ∫ X (Ω)e jΩt dΩ (1.25)
−p/T

Substituting (1.14) into (1.25), we have

p/T
⎛T ⎞
∫ ∑ xne e dΩ
− jΩnT jΩt
x a (t) = ⎜ ⎟ (1.26)
⎝ 2p ⎠
−p/T

Interchanging the order of integration and summation in (1.26), the

following results:
p/T
⎛T ⎞
∫ ∑ xne e dΩ
− jΩnT jΩt
x a (t) = ⎜ ⎟
⎝ 2p ⎠
−p/T
(1.27)
⎛T ⎞ ⎛ p/T ⎞
= ⎜ ⎟ ∑ xn ⎜ ∫ e jΩ(t−nT ) dΩ⎟
⎝ 2p ⎠ ⎝ −p/T ⎠

6899_Book.indb 7 10/22/19 4:54 PM

 8 Introduction to OFDM Receiver Design and Simulation

The integral inside the parentheses of (1.27) can be evaluated to give

⎡⎛ p ⎞ ⎤
sin ⎢⎜ ⎟ (t − nT )⎥
⎝ T ⎠
x a (t) = ∑ x(nT ) ⎣ ⎦ (1.28)
⎛p⎞
⎜⎝ T ⎟⎠ (t − nT )

Equation (1.28) is a well-known sampling theorem. It is an interpola-

tion formula that shows how an analog signal can be reconstructed from the
discrete samples x(nT ). The Nyquist rate is used in order to get this interpo-
lation formula. To prevent aliasing, the input signal must be band-limited
between −π /T and π /T.
The interpolation filter in (1.28) is a sinc function that has an infinite
duration. In reality, this function is truncated in order for computation. The
spectrum of the sinc function has a rectangular shape that will be discussed
in detail in Chapter 2.

1.6 Properties of the DFT

The DFT given by (1.18) and (1.23) are given again here:
N −1
Xk = ∑ xne − j2pnk/N k = 0,1,…, N − 1 (1.29)
n=0

N −1
1
xn =
N ∑ X k e j2pnk/N n = 0,1,…, N − 1 (1.30)
k=0

Assume that the input has a real sequence xn. The real and imaginary
parts of Xk are then given by
N −1
( ) ∑ xn cos ⎛⎜⎝ 2pnk
Re X k =
N ⎟⎠
⎞
(1.31a)
n=0

N −1
⎛ 2pnk ⎞
( )
Im X k = − ∑ xn sin ⎜
⎝ N ⎟⎠
(1.31b)
n=0

6899_Book.indb 8 10/22/19 4:54 PM

 Discrete Time Signals and Discrete Fourier Transform9

The Re(Xk) and Im(Xk) satisfy the following relationship:

( )
Re X k = Re X − k ( ) (1.32)

( )
Im X k = −Im X − k ( ) (1.33)

If the input sequence is real, the real part of Xk is then an even func-
tion of the frequency while the imaginary part of Xk is an odd function of
the frequency.
The power spectrum of the kth frequency component is the sum of the
square of the real and imaginary parts of Xk and is given here:

( ) ( )
2 2
Pk = ⎡⎣Re X k ⎤⎦ + ⎡⎣Im X k ⎤⎦ (1.34)

Using (1.32) and (1.33), Pk satisfies the following relationship:

Pk = P− k (1.35)

Therefore, the power spectrum of a real sequence is an even function

of the frequency.
In the frequency domain, the DFT Xk is also periodic with period N.
Using (1.29), the following relationship results:
N −1
X N −k = ∑ xne − j2pn( N −k)/N
n=0
N −1
(1.36)
= ∑ xne j2pnk/N
n=0
= X−k

Changing k by −k in (1.36) we have the familiar relationship of Xk =

X N+k to confirm the periodicity of Xk.
Using (1.32) and (1.33), we can write X–kas follows:

( ) ( )
X − k = Re X − k + jIm X − k
= Re ( X k ) − jIm ( X k ) (1.37)

= ( Xk )
∗

6899_Book.indb 9 10/22/19 4:54 PM

 10 Introduction to OFDM Receiver Design and Simulation

Using both (1.36) and (1.37), we have for a real sequence

( )
∗
X N −k = X k (1.38)

The following relationship is true for both N odd and N even:

( ) (N + 1)
∗
X ( N −1)/2 = X ( N +1)/2 k= and N odd (1.39a)
2

N
( )
∗
X N /2 = X N /2 k= and N even (1.39b)
2

Therefore, for N even, the spectrum at the middle frequency is real.

Using (1.34) and (1.36), it can be seen that PN–k = P–k. Using (1.35) for a real
sequence, then Pk = PN–k. An example is shown in Figure 1.5 for N = 8 and
we have P 7 = P1, P6 = P2, and P5 = P3.
If the input sequence xnis complex, then the DFT Xk is, in general,
complex. However, if the Xk sequence is complex, then the xn sequence is
also complex.

1.7 Time and Frequency Relationship of the DFT

The DFT has an interesting but important time and frequency relationship
[2]. Figure 1.6 is a typical plot of the input sequence in the time domain.
Assuming that the sampling period is T and there are N samples, the time
duration Lt is then given by
Lt = NT (1.40)

dt = T (1.41)

Figure 1.5 Power spectrum P(k) for k from −4 to 8.

6899_Book.indb 10 10/22/19 4:54 PM

 Discrete Time Signals and Discrete Fourier Transform11

where dt is the time duration between two consecutive samples in the time
domain and is also equal to the sampling time T. Similarly, Figure 1.7 is a typi-
cal plot of the DFT in the frequency domain. The bandwidth is then given by
1
Lf = (1.42)
T

1
df = (1.43)
NT

where df is the frequency width between two consecutive samples in the fre-
quency domain. The bandwidth is also equal to the sampling frequency 1/T.
Equations (1.41) to (1.44) also illustrates the following relationship:
1
Lt = (1.44)
df

1
dt = (1.45)
Lf

From (1.44), the time duration Lt is the inverse of the frequency incre-
ment df. From (1.45), the time increment dt is the inverse of the bandwidth Lf.

Figure 1.6 Discrete input sequence in the time domain.

Figure 1.7 DFT in the frequency domain.

6899_Book.indb 11 10/22/19 4:54 PM

 12 Introduction to OFDM Receiver Design and Simulation

Therefore, as the frequency increment decreases, the time duration increases.

However, as the bandwidth decreases, the time increment increases. These
two relationships are useful to define the characteristics of the time sequence
when the bandwidth is known.

1.8 Operations of the DFT

The DFT has a few important operations [1, 2] that are useful. Four opera-
tions, linearity, time shift, frequency shift, and convolution, are considered
here. For ease of presentation, the factor W N is defined here:

WN = e j2p/N (1.46)

1.8.1 Linearity
Consider two periodic sequences, xn and yn. The DFT of xn is Xk and the DFT
of yn is Yk. The two periodic sequences can be linearly combined to generate
a third periodic sequence qn = axn + byn where a and b are constants. Then
clearly, the DFT of axn + byn is aXk + bYk.

1.8.2 Time Shift

Assume that a periodic sequence xn is shifted by m to generate yn = xn–m. Then
its DFT Yk is given here:

Yk = X kWN−mk (1.47)

This can be shown by finding the inverse DFT of Yk:

N −1 N −1
⎛ 1⎞ ⎛ 1⎞
yn = ⎜ ⎟ ∑ YkWNnk = ⎜ ⎟ ∑ X kWNnkWN−mk
⎝ N ⎠ k=0 ⎝ N ⎠ k=0
N −1
⎛ 1⎞
= ⎜ ⎟ ∑ X kWN(n−m)k (1.48)
⎝ N ⎠ k=0
= xn−m

Equation (1.47) implies that a time shift in the time domain is equivalent
to multiplication by a phase factor in the frequency domain.

6899_Book.indb 12 10/22/19 4:54 PM

 Discrete Time Signals and Discrete Fourier Transform13

1.8.3 Frequency Shift

Assume that a periodic sequence xn is multiplied by a phase factor to generate
yn = xnW Nnr. Then its DFT is given by
Yk = X k−r (1.49)

This can be shown by finding the DFT of yn:

N −1
Yk = ∑ ynWN−nk
n=0
N −1
= ∑ xnWN−nkWNnr
n=0 (1.50)
N −1
= ∑ xnWN−n(k−r )
n=0
= X k−r

Equation (1.50) implies that a multiplication in the time domain by a

phase factor is equivalent to a frequency shift in the frequency domain. This
frequency shift is equivalent to a Doppler shift. Numerically, it is equal to r/
(NT ).

1.8.4 Circular Convolution

Consider two periodic sequences xn and yn of period N. Then its circular con-
volution qn = ∑ m=0 xm yn−m has the DFT of XkYk. This can be derived by
N −1

finding the inverse DFT of XkYk:

N −1
⎛ 1⎞
qn = ⎜ ⎟ ∑ X kYkWNkn
⎝ N ⎠ k=0

⎛ 1⎞ ⎛
N −1 N −1
⎞ ⎛ N −1 ⎞
= ⎜ ⎟ ∑ ⎜ ∑ xmWN− km ⎟ ⎜ ∑ yrWN− kr ⎟ WNkn (1.51)
⎝ N ⎠ k=0 ⎝ m=0 ⎠ ⎝ r=0 ⎠
⎛ 1⎞
N −1 N −1
⎛ N −1 ⎞
= ⎜ ⎟ ∑ xm ∑ yr ⎜ ∑ WN(n−m−r )k ⎟
⎝ N ⎠ m=0 r=0 ⎝ k=0 ⎠

Using (1.21), the term in the parentheses is 0 unless n − m − r = jN.

Therefore, r = n − m and (1.51) becomes

6899_Book.indb 13 10/22/19 4:54 PM

 14 Introduction to OFDM Receiver Design and Simulation

N −1
qn = ∑ xm yn−m (1.52)
m=0

The sequence qn is then the circular convolution of the two periodic

sequences xn and yn. Its DFT is then the product of Xk and Yk. Instead of
computing the sum given in (1.52), another way is to compute the product
XkYk first. Its inverse DFT generates the convolution.
It is frequently necessary to find the linear convolution of the sequences
xn and yn. For example, an input signal such as speech has to be lowpass-
filtered to prevent aliasing. In this case, the two sequences of finite duration
N are not periodic. However, its linear convolution can still be implemented
using circular convolution. Since qn in (1.52) has duration 2N − 1, the cir-
cular convolution can be applied by appending N − 1 zeros to the end of
both xn and yn.
More details on both linear and circular convolution can be found in [1].

1.9 Z-Transform
The z-transform [1, 2] is defined by the following equation:
n=∞
X (z) = ∑ xn z −n (1.53)
n=−∞

where z is a complex variable. If z = ejω, then the z-transform is the same as

the DFT given in (1.16):

X (w) = ∑ xn e − jwn (1.54)

The absolute value of z is 1. As ω varies from −π to π , the z traverses

along the unit circle from −1 to 1. The region of convergence is the range of
z such that ∑ n=−∞ xn z −n < ∞ .
n=∞

Example 1.1
Consider the positive sequence given here:

xn = { bn n ≥ 0
0 n<0
(1.55)

6899_Book.indb 14 10/22/19 4:54 PM

 Discrete Time Signals and Discrete Fourier Transform15

The z-transform is then given by

n=∞ n=∞ n
b z
X (z) = ∑ bn z −n = ∑ ⎛⎜⎝ z ⎞⎟⎠ = z − b (1.56)
n=0 n=0

The region of convergence is then (b/z) < 1 or z > b.

The two most important operations of the z-transform are time shift
and convolution. Assume that a time sequence yn = xn–m or yn is a time shift
of xn by m samples. Its z-transform is then given by
n=∞ k=∞
Y (z) = ∑ xn−m z −n = ∑ xk z −(k+m)
n=−∞ k=−∞
(1.57)
k=∞
= z −m ∑ xk z − k = z −m X (z)
k=−∞

where the variable k = n − m is defined. Therefore, a time shift by m in the

time domain is equivalent to a multiplication of z –m in the z domain.
Assume that xn and hn are two sequences. Their convolution is the
sequence yn = ∑ m=−∞ xm hn−m . The z-transform of yn is then given by
m=∞

n=∞ m=∞
Y (z) = ∑ ∑ xm hn−m z −n (1.58)
n=−∞ m=−∞

By defining a new variable k = n − m, (1.58) becomes

k=∞ m=∞
Y (z) = ∑ ∑ xm hk z −m−k
k=−∞ m=−∞

⎛ k=∞ ⎞ ⎛ m=∞ ⎞
= ⎜ ∑ hk z − k ⎟ ⎜ ∑ xm z −m ⎟ (1.59)
⎝ k=−∞ ⎠ ⎝ m=−∞ ⎠
= X (z)H (z)

Assume a linear system H(z) whose time sequence is hn. Passing through
the input sequence xn to H(z) generates an output sequence yn, which is the
convolution of xn and hn. The z-transform of the output is simply the product
of the input z-transform X(z) and the system z-transform H(z). This is also
similar to the case for the DFT.

6899_Book.indb 15 10/22/19 4:54 PM

 16 Introduction to OFDM Receiver Design and Simulation

The system function described in (1.59) has the following general form:
N (z)
H (z) = (1.60)
D(z)

Both N(z) and D(z) are polynomials of z−1 and they may have different
degrees. Assuming that N(z) has the degree u and D(z) has the degree v, they
can be written as

N (z) = b0 + b1z −1 + b2 z −2 + ! + bu z −u (1.61)

D(z) = 1 + a1z −1 + a2 z −2 + ! + av z −v (1.62)

Substituting (1.60), (1.61), and (1.62) into (1.59), we have

Y (z) b0 + b1z −1 + b2 z −2 + ! + bm z −u
= (1.63)
X (z) 1 + a1z −1 + a2 z −2 + ! + an z −v

Equation (1.63) can be rewritten in the following summation form

k=u k=v
Y (z) = ∑ X (z)bk z − k − ∑Y (z)ak z − k (1.64)
k=0 k=1

Using the time shift operations given in (1.57), (1.64) can be converted
into the following difference equations
k=u k=v
yn = ∑ bk xn−k − ∑ ak yn−k (1.65)
k=0 k=1

Equation (1.65) is realizable using tapped delay lines. More details on

the implementation can be found in [2].

1.10 Summary
A discrete time signal is obtained by sampling an analog signal at a periodic
rate. The DFT of these discrete samples is then derived. It was found that the
sampled spectrum is a summation of the shifted spectrum of the analog signal.
To reconstruct the original analog signal, the discrete samples must be taken
at least at the Nyquist rate. For a band-limited analog signal with a cutoff

6899_Book.indb 16 10/22/19 4:54 PM

 Discrete Time Signals and Discrete Fourier Transform17

frequency fc, the Nyquist rate is at least 2fc. Based upon the Nyquist rate, an
interpolation formula was derived to reconstruct the analog signal.
The DFT has a relationship between time and frequency. The band-
width in the frequency domain is the inverse of the sampling period in the
time domain. The signal duration in the time domain is the inverse of the
frequency increment in the frequency domain.
Three important operations of the DFT were illustrated: time shift,
­frequency shift, and convolution. A time shift in the time domain is equivalent
to the multiplication by a phase factor in the frequency domain. A multiplica-
tion in the time domain by a phase factor is equivalent to a frequency shift in
the frequency domain. This frequency shift is equivalent to a Doppler shift.
The circular convolution of two discrete time sequences was defined. Its DFT
is the product of the DFT of the two individual sequences.
The z-transform was also defined. It is useful in specifying the time-
invariant system response. The output generated by passing the input through
such a system can be implemented using a tapped delay line.
In the next chapter, single-carrier modulation is reviewed.

References
[1] Oppenheim, A. V., and R. W. Schafer, Digital Signal Processing, Upper Saddle River,
NJ: Prentice Hall, 1975.
[2] Stanley, W. D., Digital Signal Processing, AU: Provide city of publication: Reston
Publishing Company, 1975.
[3] Proakis, J. G., Digital Communications, New York: McGraw-Hill, 1983.

6899_Book.indb 17 10/22/19 4:54 PM

6899_Book.indb 18 10/22/19 4:54 PM
 2
Single-Carrier Modulation

2.1 Introduction
OFDM is a multicarrier modulation technique. First, let us explain single-
carrier modulation. From the name itself, it is clear there is only one carrier.
Depending upon the type of modulation, the amplitude, phase, or frequency
of this single carrier is modulated.
In digital communication, a sequence of digital bits is first grouped
into bauds. Each baud is then mapped into a unique phase, amplitude, or
frequency to modulate the carrier. The four most commonly applied modula-
tion techniques are discussed. They are pulse amplitude modulation (PAM),
phase shift keying (PSK), quadrature amplitude modulation (QAM), and
frequency shift keying (FSK).
The analog signal can be reconstructed through interpolation and was
discussed in Chapter 1. Two interpolation filters, rectangular and raised cosine,
are given.

2.2 Data Transmission Rate

The transmission rate depends upon the number of bits k sent in a baud interval.
A baud is defined to have k bits. The number k is an integer and is different

19

6899_Book.indb 19 10/22/19 4:54 PM

 20 Introduction to OFDM Receiver Design and Simulation

for different modulation schemes. As will be shown in later sections, k = 1, 2,

4, and 6 for binary phase shift keying (BPSK), quadrature phase shift keying
(QPSK), 16-QAM, and 64-QAM, respectively.
The simplest case is to have k = 1 and there are only two possible levels.
During each baud interval, there are only two possible amplitudes, phases,
or frequencies according to PAM, PSK, or FSK. The resulting modulation is
called binary modulation.
In general, there are M = 2k possible levels where k is an integer. In other
words, k binary digits are mapped to M possible transmitted waveforms. These
M possible levels are converted into complex numbers representing constella-
tion points of the digital modulation. If the input bit rate is R bits/sec, then
the baud rate is R/k. Assuming that the baud interval is T, then the resulting
bit transmission rate is k/T.

2.3 Bandpass Signals

The modulated digital signal can in general be represented in the follow-
ing form:

(
x(t) = a(t)cos 2pf c t + q(t) ) (2.1)

where a(t) is the amplitude, θ (t) is the phase, and fc is the carrier frequency
of the modulated signal. Equation (2.1) can be represented in complex form
as given here:

(
x(t) = Re a(t)e jq(t )e j2pf c t ) (2.2a)

(
= Re s(t)e j2pf c t ) (2.2b)

where s(t) is the lowpass signal bearing the information of the input digital
bit stream

s(t) = a(t)e jq(t ) (2.3)

The spectrum of x(t) can be obtained by taking the Fourier transform

on both sides of (2.2b) to obtain

6899_Book.indb 20 10/22/19 4:54 PM

 Single-Carrier Modulation21

Figure 2.1 Spectrum of the lowpass signal s(t).

X( f ) =
1
2∫
( )
s(t)e j2pf c t + s∗ (t)e − j2pf c t e − j2pft dt
(2.4)
1
(( )
= S f − fc + S∗ f + fc
2
( ))
where S(f ) is the spectrum of s(t). Figure 2.1 is a plot of the lowpass signal s(t)
and Figure 2.2 is a plot of the bandpass signal x(t). Clearly, X(f ) is a frequency
shifted version of the baseband signal x(t).

2.4 Digitally Modulated Signals

In each baud interval, a(t), θ (t), or both can only take a set of discrete levels.
Therefore, s(t) can also be represented in the following form [1]:
∞
s(t) = ∑ sn q(t − nT ) (2.5)
n=−∞

where T is the baud interval, q is the pulse-shaping waveform, and sn is the

nth sample of the information bearing signal s(t). The term sn is, in general,

Figure 2.2 Spectrum of the bandpass signal X(f ).

6899_Book.indb 21 10/22/19 4:54 PM

 22 Introduction to OFDM Receiver Design and Simulation

complex and can be represented as

sn = an e jqn (2.6)

where an is the nth sample of a(t) and θ n is the nth sample of θ (t). The in-phase
component In and the quadrature-phase component Qn can then be written as

( )
I n = an cos qn (2.7a)

( )
Q n = an sin qn (2.7b)

Using (2.7a) and (2.7b), sn defined in (2.6) can be written as

sn = I n + jQ n (2.8)

To make sure the average transmitted power is the same independent of

the types of modulation, the amplitude an can be adjusted properly.
The sncan only take one of possible M levels for each of the k = log2M
input bits. Depending on how an and θ n are changed in each baud interval
T, different types of modulated digital signals are generated.

2.4.1 PAM
For PAM, only the amplitude an is changed while the phase θ n remains con-
stant. The simplest case is the on-off keying where k = 1 and M = 2. If the
input bit is 1, there is a transmission. However, if the input bit is 0, there is no
transmission. In general, there are M different amplitude levels for every k bits.

2.4.2 PSK
For PSK, the amplitude an remains constant while only the phase θ n is changed.
There are two popular PSK modulations. One is BPSK and the other is QPSK.
For BPSK, k = 1 and M = 2. Figure 2.3 shows the signal constellation.
The phase angle is either 180° or 0° and sn is either 1 or −1. Table 2.1 lists the
complex amplitude sn.
For QPSK, k = 2 and M = 4, and there are four possible phases. The
phase angles can be encoded as 45°, 135°, 225°, and 315°. Figure 2.4 shows
the signal constellation and Table 2.2 lists the complex amplitude 2sn [2].
The factor 2 is a constant to normalize the average energy.

6899_Book.indb 22 10/22/19 4:54 PM

 Single-Carrier Modulation23

Figure 2.3 Signal constellation of BPSK.

Table 2.1
BPSK Encoding Table

Input Bits sn

0 −1
1 1

Figure 2.4 Signal constellation of QPSK.

Table 2.2
QPSK Encoding Table

2sn
Input bits

00 −1 − j
01 −1 + j
11 1+j
10 1−j

6899_Book.indb 23 10/22/19 4:54 PM

 24 Introduction to OFDM Receiver Design and Simulation

2.4.3 QAM
For QAM, both the amplitude an and the phase θ n are changed. In the IEEE
802.11a standard [2 ], there are two types of QAM: 16-QAM and 64-QAM.
Depending on the phase angle, the amplitude can be different to make sure
that the average transmitted power remains the same.
In 16-QAM, k = 4 and M = 16. In other words, four input bits are
encoded into 16 different discrete levels. Figure 2.5 shows the signal constel-
lation while Table 2.3 lists the complex amplitude sn/C1 where C1 is a normal-
ization constant. The encoding of the constellation point is based upon the

Figure 2.5 Signal constellation of 16-QAM.

Table 2.3
16-QAM Encoding Table

Input Bits sn /C1 Input Bits sn /C1

0000 −3 − 3j 1000 3 − 3j
0001 −3 − j 1001 3−j
0010 −3 + 3j 1010 3 + 3j
0011 −3 + j 1011 3+j
0100 −1 − 3j 1100 1 − 3j
0101 −1 − j 1101 1−j
0110 −1 + 3j 1110 1 + 3j
0111 −1 + j 1111 1+j

6899_Book.indb 24 10/22/19 4:54 PM

 Single-Carrier Modulation25

concept of Gray coding. In other words, any two adjacent constellation points
differ by only 1 bit. For example, 3 + j encoded as 1011 and 3 − j encoded as
1001 are adjacent but differ only in the third bit.
In 64-QAM, k = 6 and M = 64. Therefore, six input bits are encoded
into 64 discrete levels. Figure 2.6 shows the signal constellation while Table
2.4 lists the complex amplitude sn/C2 where C2 is a normalization constant.
The same Gray coding is applied.
The constant C1 in Table 2.3 and the constant C2 in Table 2.4 can be
determined by normalizing the average energy to be a constant 1. To deter-
mine C1 in Table 2.3 to normalize the average energy, the following is done:
15
⎛ 1⎞
⎜⎝ 16 ⎟⎠ ∑ si = 1
2
(2.9)
i=0

Table 2.4
64-QAM Encoding Table

Input Input Input Input

Bits sn /C2 Bits sn /C2 Bits sn /C2 Bits sn /C2

000000 −7 − 7j 010000 −1 − 7j 100000 7 − 7j 110000 1 − 7j

000001 −7 − 5j 010001 −1 − 5j 100001 7 − 5j 110001 1 − 5j
000010 −7 − j 010010 −1 − j 100010 7−j 110010 1−j
000011 −7 − 3j 010011 −1 − 3j 100011 7 − 3j 110011 1 − 3j
000100 −7 + 7j 010100 −1 + 7j 100100 7 + 7j 110100 1 + 7j
000101 −7 + 5j 010101 −1 + 5j 100101 7 + 5j 110101 1 + 5j
000110 −7 + j 010110 −1 + j 100110 7+j 110110 1+j
000111 −7 + 3j 010111 −1 + 3j 100111 7 + 3j 110111 1 + 3j
001000 −5 − 7j 011000 −3 − 7j 101000 5 − 7j 111000 3 − 7j
001001 −5 − 5j 011001 −3 − 5j 101001 5 − 5j 111001 3 − 5j
001010 −5 − j 011010 −3 − j 101010 5−j 111010 3−j
001011 −5 − 3j 011011 −3 − 3j 101011 5 − 3j 111011 3 − 3j
001100 −5 + 7j 011100 −3 + 7j 101100 5 + 7j 111100 3 + 7j
001101 −5 + 5j 011101 −3 + 5j 101101 5 + 5j 111101 3 + 5j
001110 −5 + j 011110 −3 + j 101110 5+j 111110 3+j
001111 −5 + 3j 011111 −3 + 3j 101111 5 + 3j 111111 3 + 3j

6899_Book.indb 25 10/22/19 4:54 PM

 26 Introduction to OFDM Receiver Design and Simulation

Figure 2.6 Signal constellation of 64-QAM.

The constant C1 is then found to be 1/ 10 . Similarly, the constant C2

is found to be 1/ 42 .

2.4.4 FSK
For FSK, the information sn is used to change the frequency of the signal
waveform s(t). Assuming the frequency deviation is given by ∆fsn, the s(t) is
then given by [1]
∞
s(t) = ∑ e j2p(Δf )ts q(t − nT )
n
(2.10)
n=−∞

Since sncan take M different levels, there are M possible frequencies from
which to choose. The end result is the transmitted carrier frequency is shifted
based upon the input information sn.

6899_Book.indb 26 10/22/19 4:54 PM

 Single-Carrier Modulation27

2.5 Pulse Shaping

The analog signal xa(t) can be reconstructed from the discrete samples x(nT )
according to the sampling theorem given by (1.28). That equation is rewrit-
ten here:

⎡⎛ p ⎞ ⎤
sin ⎢⎜ ⎟ (t − nT )⎥
⎝ T ⎠
x a (t) = ∑ x(nT ) ⎣ ⎦ (2.11)
⎛p⎞
⎜⎝ T ⎟⎠ (t − nT )

Assume that the signal cutoff frequency is f = fc as shown in Figure 1.3.

The minimum sampling rate is the Nyquist rate or 2fc. By setting the baud
period T = 1/2fc, (2.11) becomes

( )(
⎛ n ⎞ sin ⎡⎣ 2pf c t − n/2 f c ⎤⎦
x a (t) = ∑ x ⎜
)
(2.12)
⎝ 2 f c ⎟⎠ ( )(
2pf c t − n/2 f c )
Equivalently, each discrete sample is multiplied by a pulse-shaping func-
tion given here:

⎛ pt ⎞
sin ⎜ ⎟
⎝T⎠
q(t) = (2.13)
pt
T

The pulse q(t), which is defined in (2.13), is 1 at t = 0 and 0 at t = nT

where n is an integer. As can be seen from (2.11), xa(kT ) = x(kT ). Therefore,
to reconstruct an analog signal xa(t) without intersymbol interference (ISI),
the following condition must be satisfied:

q(kT ) = { 01 k=0
k≠0 (2.14)

Equation (2.11) is just the convolution of the discrete signal x(t) with
the pulse q(t). Therefore, the analog spectrum is the product of the pulse
spectrum and the signal spectrum. In order to have no spectral distortion,
the ideal frequency spectrum of the function q(t) is the rectangular function,
Q(Ω) given here:

6899_Book.indb 27 10/22/19 4:54 PM

 28 Introduction to OFDM Receiver Design and Simulation

Figure 2.7 The frequency spectrum of the sinc function versus Ω = 2 π f.

⎧⎪ p p
Q (Ω) = ⎨ T − T < Ω < T (2.15)
⎩⎪ 0, otherwise

The function Q(Ω) is also plotted in Figure 2.7. As shown next, the
inverse Fourier transform of (2.15) generates q(t):
p/T
⎛T ⎞
q(t) = ⎜ ⎟
⎝ 2p ⎠ ∫ Q (Ω)e jΩt dΩ
−p/T
p/T
⎛T ⎞
=⎜ ⎟
⎝ 2p ⎠ ∫ e jΩt dΩ (2.16)
−p/T

⎛ pt ⎞
sin ⎜ ⎟
⎝T⎠
=
pt
T

Figure 2.8 Pulse having a sinc function.

6899_Book.indb 28 10/22/19 4:54 PM

 Single-Carrier Modulation29

Figure 2.8 is a plot of sinc function given in (2.16). As can be seen, q(t)
is 1 at t = 0 and 0 when t/T has integral values.
Even though the sinc function in the time domain generates no ISI, it
has infinite duration and cannot be realized. In practice, the sinc function is
truncated to approximate the ideal function. If enough side lobes are kept,
negligible ISI can result.
The sinc function drops at the rate of 1/t. One method frequently used
in digital communication is to use the raised cosine function defined next [1]:

⎛ pt ⎞ ⎛ bpt ⎞
sin ⎜ ⎟ cos ⎜
⎝T⎠ ⎝ T ⎟⎠
q(t) = (2.17)
pt 1 − 4 b2 t 2
T T2

Its spectrum Q(f )/T has the following closed form [1]:

⎧ (1 − b)
⎪1 0≤ f ≤
2T
⎪ ⎛ ⎞
Q( f ) ⎪ ⎛ ⎛ 1 ⎞⎞
=⎨ pT ⎜ f −
T ⎜ ⎜ ⎝ 2T ⎟⎠ ⎟ ⎟ (1 − b) (1 + b) (2.18)
⎪ 0.5⎜1 − sin ⎜ ⎟⎟ ≤ f ≤
⎪ ⎜ b 2T 2T
⎜ ⎟⎟
⎪⎩ ⎝ ⎝ ⎠⎠

Figure 2.9 is a plot of the raised function against t/T for β = 0, 0.5, and
1. For β = 0, the raised cosine function is the same as the sinc function. As
β increases, the side lobe also decreases. For β = 1, the side lobe is negligibly

Figure 2.9 Pulse having a raised cosine function.

6899_Book.indb 29 10/22/19 4:54 PM

 30 Introduction to OFDM Receiver Design and Simulation

Figure 2.10 Frequency spectrum of the raised cosine function versus fT.

small after the main lobe. Figure 2.10 is a plot Q(f )/T against fT. When β
= 0, the spectrum converges to the ideal rectangle with bandwidth fT = 0.5
or Ω = 2π f = π /T.
The raised cosine function also satisfies the conditions for reconstruction
without ISI as defined in (2.14).

2.6 Summary
In digital communication, the traditional way is to have only one carrier to
carry the data information. A sequence of k input bits is first grouped to gen-
erate a baud that has a total of M = 2k levels. Each level is mapped onto the
phase, amplitude, or frequency of this single carrier for transmission.
Four modulation techniques, PAM, PSK, QAM, and FSK were pre-
sented. In PAM, only the amplitude is modulated. In PSK, only the phase
is modulated. If M = 2 or k = 1, it is BPSK. If M = 4 or k = 2, it is QPSK.
In QAM, both amplitude and phase are modulated. If M = 16 or k = 4, it
is 16-QAM. If M increases to 64 or k = 6, it is then 64-QAM. In FSK, M
different frequencies corresponding to k bits can be selected to change the
carrier frequency.
From the discrete digital samples, the analog signal can be reconstructed
by applying an interpolation filter. Two pulse-shaping waveforms were given.
One has a rectangular spectrum implying a sinc function in the time domain.
This sinc function has an infinite duration requiring truncation for implemen-
tation. To reduce the side lobes for fast convergence, a raised cosine function
can be applied. A parameter β was defined in this function to control the rate

6899_Book.indb 30 10/22/19 4:54 PM

 Single-Carrier Modulation31

of convergence. For β = 0, it is the same as the sinc function. As β increases,

the side lobe also decreases.
In the next chapter, multicarrier modulation is introduced.

References
[1] Proakis, J. G., Digital Communications, New York: McGraw-Hill, 1983.
[2] Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY)
Specifications, IEEE Standard 802.11a, 2007.

6899_Book.indb 31 10/22/19 4:54 PM

6899_Book.indb 32 10/22/19 4:54 PM
 3
Multicarrier Modulation

3.1 Introduction
OFDM is a modulation technique that applies multiple carriers to be modu-
lated by an input data stream. In a single-carrier modulation, the input data
is carried by only one carrier and the modulation is carried out in the time
domain. In a multicarrier modulation, the input data is split into bauds and
carried by multiple subcarriers and the modulation is carried out in the fre-
quency domain. Any two subcarriers are orthogonal in a symbol interval. This
is also why OFDM is so named.
A transmission through a multipath channel can cause ISI and signal
distortion. One characteristic of OFDM is its insertion of guard interval
between adjacent symbols to eliminate this degradation. This guard interval
is also called cyclic prefix to maintain the orthogonality of the subcarriers.
As will become clear in later chapters, the multipath degradation and fading
can be removed much more easily in the frequency domain using OFDM.
Even though it is easier to remove the multipath interference using
OFDM, there are also some disadvantages. One is the peak-to-average power
ratio (PAPR). This is because the peak power can be much higher than the
average power. When this happens, the OFDM signal is clipped to cause
distortion. In other words, the efficiency of the power amplifier is reduced.

33

6899_Book.indb 33 10/22/19 4:54 PM

 34 Introduction to OFDM Receiver Design and Simulation

Two other major characteristics of the OFDM are the FFT and the
classical OFDM spectrum. A DFT is an essential tool used in this multicar-
rier modulation. However, the FFT is a fast computation technique of DFT.
Except for the computation speed, there is no difference between the DFT
and the FFT. The classical OFDM spectrum is a summation of the sinc func-
tion from each subcarrier. It is a sinc function because of the finite window
length in the time domain.
To understand the OFDM waveform, we therefore start with the DFT.
Subsequently, cyclic prefix, FFT, classical OFDM spectrum, PAPR, and
orthogonality are all analyzed. After discussing the OFDM characteristics,
time and frequency relationship, window function, and digital modulation
are given in sequence in the subsequent sections.

3.2 OFDM Waveform

In a multicarrier waveform such as OFDM, each carrier is modulated by k bits
having one of M = 2k possible levels. This waveform is defined by the inverse
discrete Fourier transform given in (1.30) and is rewritten here:
N −1
1
xn =
N ∑ X k e j2pnk/N n = 0,1,…, N − 1 (3.1)
k=0

In (3.1), k is the frequency index, n is the time index, and Xk is the kth
frequency component. During the modulation process, the amplitude, phase,
or frequency of Xk is modified. Therefore, the whole process is performed in
the frequency domain in contrast to the time-domain modulation of the
single-carrier waveform.
To get more insight into the multicarrier waveform, define the kth radian
frequency component given here:

2pk
wk = (3.2)
N

Substituting (3.2) into (3.1), we have

N −1
1
xn =
N ∑ X k e jnwk
n = 0,1,…, N − 1 (3.3)
k=0

6899_Book.indb 34 10/22/19 4:54 PM

 Multicarrier Modulation35

In general, Xk is complex and can be written as

X k = Ak e jΨ k (3.4)

Substituting (3.4) into (3.3), we have

N −1
⎛ 1⎞
xn = ⎜ ⎟ ∑ Ak e ( k k )
j nw +Ψ
(3.5)
⎝ N ⎠ k=0

It is clear from (3.5) that there are N carriers. These carriers are also
called subcarriers. Each subcarrier has an amplitude Ak, a phase Ψk, and a
radian frequency wk.
The discrete samples xn can be reconstructed to generate x(t) as long as
the sampling rate is at least twice the baseband cutoff frequency. Assuming
that the radio frequency (RF) carrier frequency is fc, then the transmitted
analog waveform for the jth OFDM symbol is given by

(
y(t) = Re x(t)e j2pf c t ) jNT < t < ( j + 1)NT (3.6)

In the same way as the single-carrier case, the input bit stream is grouped
into bauds. Each baud has k bits where k = log2M and M is the total number
of levels. Each baud of k bits is then mapped onto one of M possible phases,
frequencies, or amplitudes.

3.3 OFDM Characteristics

As mentioned in the introduction, the most important characteristics of
OFDM are orthogonality, sinc spectrum, PAPR, and FFT. Each of these
subjects is discussed next.

3.3.1 Orthogonality
In OFDM, any two subcarriers are orthogonal. To show that, assume any two
subcarriers, yi and yk, given below:

yk = X k e j2pkn/N (3.7)

yi = X i e j2pin/N (3.8)

6899_Book.indb 35 10/22/19 4:54 PM

 36 Introduction to OFDM Receiver Design and Simulation

Then the following relationship must be true:

⎛ 1⎞
N −1
⎪⎧ 0 if k ≠ i
⎜⎝ N ⎟⎠ ∑ yk yi = ⎨ X 2 if k = i
∗
(3.9)
n=0 ⎪⎩ k

Substituting (3.7) and (3.8) into (3.9), we have

N −1 N −1
⎛ 1⎞ ⎛ 1⎞
⎜⎝ N ⎟⎠ ∑ yk yi = ⎜⎝ N ⎟⎠ ∑ X k X i e
∗ ∗ j2pn(k−i)/N

n=0 n=0
(3.10)
⎛ 1⎞ ⎛ N −1 ⎞
= ⎜ ⎟ X k X i∗ ⎜ ∑ e j2pn(k−i)/N ⎟
⎝N⎠ ⎝ n=0 ⎠

However, the term inside the brackets is 0 unless k = i as given in (1.21);

we then have

⎛ 1⎞ ⎪⎧ 0 if k ≠ i
N −1 (3.11)
⎜⎝ N ⎟⎠ ∑ yk yi = ⎨ X 2 if k = i
∗
(3.12)
n=0 ⎩⎪ k

Therefore, any two subcarriers are orthogonal in OFDM.

3.3.2 OFDM Spectrum

OFDM transmits a large number of subcarriers and each subcarrier has the
same spacing of 2π /N in radians or fc/N in frequency. Assume each subcarrier
is not modulated, then Xk is 1. In the digital domain, each subcarrier then has
the form (ej2πnk/N)/N. Converting it to the analog domain, it has the equivalent
form (ej2πkfst/N)/N where fs is the sampling frequency. Since the DFT interval
is NT, the classical spectrum X(f ) is then given here:
N −1 NT /2
⎛ 1⎞
X( f ) = ⎜ ⎟ ∑ ∫ e j2pkf s t/N e − j2pft dt (3.13)
⎝ N ⎠ k=0
− NT /2

The integral inside the summation of (3.13) can be easily evaluated to

give the following:

X( f ) N −1
sin(pfNT − pk)
T
= ∑ pfNT − pk
(3.14)
k=0

6899_Book.indb 36 10/22/19 4:54 PM

 Multicarrier Modulation37

Figure 3.1 OFDFM spectrum for N = 5 and fs = 20 MHz.

Equation (3.14) shows that the OFDM spectrum is a summation of sinc

function from different subcarriers.

Example 3.1
Assuming that fs = 20 MHz, then T = 0.05 µ s. Assume that also N = 5 and
k ranges from 1 to 4. The subcarrier at k = 0 is assumed to be unused. The
spectrum is plotted in Figure 3.1. The spectrum is maximum at frequency 4
MHz, 8 MHz, 12 MHz, and 16 MHz corresponding to k = 1, 2, 3, and 4. It
can be seen also that the maximum at a given subcarrier is the zero-crossing
for all the rest of subcarriers. This also means there is no intercarrier interfer-
ence at the peak frequencies.

3.3.3 PAPR
The PAPR is defined to be the ratio of maximum power to the average power
in an OFDM symbol interval:

Pmax
PAPR = (3.15)
Pavg

⎛ N −1 ∗ ⎞
Pmax = max ⎜ ∑ xn xn ⎟ (3.16)
⎝ n=0 ⎠

6899_Book.indb 37 10/22/19 4:54 PM

 38 Introduction to OFDM Receiver Design and Simulation

⎛ N −1 ⎞
Pavg = E ⎜ ∑ xn xn∗ ⎟ (3.17)
⎝ n=0 ⎠

where Pmax represents the maximum power, Pavg represents the average power,
and E is the expected value. As was mentioned in the introduction, PAPR can
be large to cause detrimental effects to the OFDM signal.
To get some idea about the magnitude of PAPR, both Pmax and Pavg
can be evaluated by substituting (3.1) into (3.16) and (3.17). To simplify the
analysis, each subcarrier is assumed to be unmodulated and Xk = 1. Under
this assumption, (3.16) becomes

1 N −1⎛ N −1 ⎞ ⎛ N −1 ⎞
Pmax = 2 ∑ ⎜ ∑
N n=0 ⎝ k=0
max e (
j2pnk/N
⎟ ⎜ ∑)
⎠ ⎝ k=0
( )
max e − j2pnk/N ⎟
⎠
(3.18)
1 N −1
= 2 ∑(N2) = N
N n=0

Similarly, (3.17) becomes

1 N −1 N −1⎛ N −1 j2pn( k− j )/N ⎞

N2 ∑ ∑⎜ ∑e
Pavg = ⎟⎠ (3.19)
k=0 j=0 ⎝ n=0

Using the orthogonality relationship given in (1.21), the integral inside

the parentheses is 0 unless k = j; (3.19) can then be further simplified to give

N2
Pavg = =1 (3.20)
N2

Substituting (3.18) and (3.20) into (3.15), we then have the following:
PAPR = N (3.21)

In (3.21), N is defined to be the number of modulated subcarriers and

can be quite large. This proves PAPR in OFDM can be quite significant.
To minimize the impact of PAPR, many approaches are available. They
can be classified either as signal scrambling techniques or as signal distortion
techniques [1]. On signal distortion techniques, one is to perform clipping,
filtering, and peak windowing [2]. On signal scrambling techniques, one is
to perform block coding [3]. There are many others [1].

6899_Book.indb 38 10/22/19 4:54 PM

 Multicarrier Modulation39

Example 3.2
For IEEE 802.11a, there are 52 modulated subcarriers as will be given later,
and then N = 52. The PAPR in decibels is then 10log52 = 10∗1.71 = 17.1 dB.
This means that the power amplifier must have an output back-off of 17 dB
due to the power surge.

3.3.4 Cyclic Prefix

When an OFDM signal travels through the communication media, the
received signal can be degraded due to the multipath interference. For exam-
ple, the receiver may receive multiple delayed copies of the same transmitted
waveform. If the channel impulse response has a length of Nc samples, then
the received signal can also be prolonged by Ncsamples.
An OFDM symbol is defined to have N samples. If two adjacent OFDM
symbols are transmitted without any gap in between, then the tail end of one
symbol may interfere with the leading edge of the subsequent symbol. To
remove this type of ISI, the delay spread of the channel must first be deter-
mined. A gap of Ng samples must be inserted between two adjacent OFDM
symbols such that Ng ≥ Nc. This time gap is also called the guard interval.
Figure 3.2 is a plot of the OFDM signal without any guard interval
between three adjacent symbols j − 1, j, and j + 1. For simplicity of notation,
each symbol here refers to one FFT interval without including the guard
interval. Each OFDM symbol has N samples. Figure 3.3 is a plot of the
OFDM signal with a guard interval of Ng samples inserted between any two
adjacent symbols.
The guard interval has a size of Ng samples. Note that an insertion of
guard interval is a waste of transmission resources. However, the price paid
is a reward of reduced ISI.

Figure 3.2 OFDM signal without time gap between adjacent symbols.

Figure 3.3 OFDM with time gap between adjacent symbols.

6899_Book.indb 39 10/22/19 4:54 PM

 40 Introduction to OFDM Receiver Design and Simulation

Transmissions don’t have to happen inside the guard interval. If there is

no transmission, the advantage is to save transmission power. However, with
channel delay spread, different delayed subcarriers are added together and the
orthogonality is lost. To solve this problem, a cyclic prefix is transmitted during
the guard interval. From (3.1), it is seen that each OFDM symbol transmits
N samples and N is the period. To make sure OFDM signal is periodic to
maintain the orthogonality, the last Ng samples from the current symbol are
inserted into the gap preceding the beginning of the current symbol.
Figure 3.4 is a plot of how the cyclic prefix is generated. In the graph,
there are two adjacent symbols j − 1 and j with a guard interval of Ng samples
in between. Assume that the symbol j has samples from n = 0 to n = N − 1.
The dark region in Figure 3.4 has the last Ng samples from n = N − Ng to n
= N − 1 of symbol j. These Ng samples are copied to the dark region from n
= −Ng to n = −1 between symbols j − 1 and j. With this insertion of cyclic
prefix, there is no ISI if the channel delay spread is shorter than Ng samples.

3.3.5 FFT
So far, the DFT has been used to describe OFDM characteristics. However,
it is time-consuming in using DFT for direct computation. Equation (1.18)
defined the DFT and is rewritten here for easy reference:
N −1
Xk = ∑ xne − j2pnk/N k = 0,…, N − 1 (3.22)
n=0

From (3.22), there are N complex multiplications and additions for each
point of Xk. Since there are N points, the total number of complex multiplica-
tions and additions becomes N2.
To reduce the number of computations, the FFT is designed. The prin-
ciple is easily described when N is even using the decimation-in-time technique.

Figure 3.4 OFDM transmission of cyclic prefix.

6899_Book.indb 40 10/22/19 4:54 PM

 Multicarrier Modulation41

Initially, the N points are split into two groups. One group has N/2 even points
and the other group has N/2 odd points. In other words, a N-point DFT is
split into two N/2-point DFTs. These N/2 even points and odd points can be
further split into N/4 even points and odd points. This process continues until
a total of log2N stages are reached. For each stage, the number of complex
multiplications and additions is N [4]. Therefore, the total number of com-
plex multiplications and additions becomes Nlog2N. A detailed quantitative
analysis can be found in [4].
The above qualitative description shows there are significant computation
savings. The computation reduction ratio from the DFT to the FFT becomes
N/log2N. This ratio is 4 for N = 16 and 32 for N = 256. As the size of FFT
increases, the computation time savings are even more obvious.

3.4 Time and Frequency Parameters

There are some important time and frequency parameters defined in OFDM.
They are T, Tfft, Tgi, ∆f, and Tsym specifying the sampling interval, FFT interval,
guard interval, subcarrier spacing, and OFDM symbol interval, respectively.
Equivalently, T, Tfft, and ∆f correspond to dt, Lt, and df defined in (1.41),
(1.40), and (1.43).
Table 3.1 lists these parameter values for channel spacing of 20 MHz,
10 MHz, and 5 MHz. The parameter N is the number of samples for per-
forming FFT. The sampling interval is just the inverse of channel spacing.
The Fourier transform interval is then NT. The subcarrier spacing ∆f is chan-
nel spacing divided by N. The OFDM symbol interval Tsym is the sum of the
FFT interval and the guard interval or Tsym = Tfft + Tgi. For IEEE 802.11a,
the guard interval is specified as 1/4 of the FFT interval. It can be seen from

Table 3.1
OFDM Time and Frequency Parameters from IEEE 802.11a

20-MHz Channel 10-MHz Channel 5-MHz Channel

Parameter Spacing Spacing Spacing

N 64 64 64
T 0.05 μs 0.1 μs 0.2 μs
Tfft 3.2 μs 6.4 μs 12.8 μs
∆f 0.3125 MHz 0.15625 MHz 0.078125 MHz
Tgi 0.8 μs 1.6 μs 3.2 μs
Tsym 4 µs 8 μs 16 μs

6899_Book.indb 41 10/22/19 4:54 PM

 42 Introduction to OFDM Receiver Design and Simulation

this table the symbol interval increases and the subcarrier spacing decreases
as the channel spacing decreases.
Unless specified, a symbol interval refers to the data portion only without
including the guard interval in the future chapters.

3.5 Window Function

The OFDM symbol is normally multiplied by a window function in the time
domain. The purpose of the window is to smooth the transition at the sym-
bol boundary so as not to create higher spectral side lobes of the transmitted
waveform. Another is to meet the spectral mask requirements. For operations
using a 20-MHz bandwidth, the IEEE 802.11a specifies a spectrum mask to
be 28 dB down at a 20-MHz frequency offset from the carrier.
The IEEE 802.11a specifies a rectangular window. However, it has sharp
transitions at the symbol boundary. To smooth the transition, the following
modified rectangular window is defined [5]:

2⎛p b b
⎧ tp ⎞
⎪ sin ⎜⎝ 4 + 2b⎟⎠ −
2
<t<
2
⎪
⎪ b b
h(t) = ⎨ 1 ≤t <P− (3.23)
⎪ 2 2
⎪ sin2 ⎛⎜ p − p(t − p)⎞⎟ P−
b
≤t <P+
b
⎪⎩ ⎝4 2b ⎠ 2 2

where β is defined to extend the original rectangular window of time duration

P. The sine function is defined to smooth the transition during the transition
period of length β on both edges of the rectangular window. Note that as β
becomes 0, the window function h(t) converges to the rectangular window
of length P given here:
h(t) = 1 0≤t<P (3.24)

The spectrum can be found by taking the Fourier transform. For the
rectangular window given in (3.24), the spectrum is given here:
P
H ( f ) = ∫ e − jΩt dt
0
(3.25)
sin(pfP)
(
= P e − jpfP ) pfP

6899_Book.indb 42 10/22/19 4:54 PM

 Multicarrier Modulation43

where Ω = 2π f and f is the frequency. For the modified window given from
(3.23), the spectrum is derived here:
b/2 P− b/2
p tp ⎞ − jΩt
2⎛
H ( f ) = ∫ sin ⎜ + ⎟⎠ e dt ∫ e − jΩt dt
− b/2
⎝ 4 2b b/2
(3.26)
P+ b/2
⎛ p p ( t − P ) ⎞ − jΩt
+ ∫ sin2 ⎜ −
⎝4 2b ⎟⎠
e dt
P− b/2

After some lengthy algebra, (3.26) becomes

⎡ ⎤
⎢ sin ( pf (P − β )) sin(pf b)cos(pfP) 4pf cos(pf b)sin(pfP) ⎥
H ( f ) = e − jpf P ⎢ + − ⎥
⎢ pf pf ⎛ 2 2 p2 ⎞ ⎥
⎢ ⎜⎝ 4p f − b2 ⎟⎠ ⎥
⎣ ⎦
(3.27)

If β = 0, (3.27) converges to (3.25). However, if f = 0, both spectra have

a magnitude of P.
The original rectangular window is from 0 to P. Figure 3.5 is a plot of
the modified window function defined in (3.23). This window overlaps with
the adjacent window by β . The effective window length is shortened due to
the β transition period on both ends. From (3.27), the normalized power
spectrum is computed as 10log (H(f )/P)(H(f )/P)∗. Figure 3.6 is a plot of this
power spectrum against frequency in megahertz for β = 0, 0.6 μs, and 0.9
μs. The channel spacing is 20 MHz and the extended symbol width P is 4
μs. As the figure shows, the side lobe decreases as β increases, even though
the main lobe does not change much.

Figure 3.5 Modified rectangular window function.

6899_Book.indb 43 10/22/19 4:54 PM

 44 Introduction to OFDM Receiver Design and Simulation

μs
μs
μs

Figure 3.6 Power spectrum of the modified rectangular window at a 20-MHz channel
spacing.

3.6 Digital Modulation

The complex amplitude of the kth carrier given in (3.4) is rewritten here:

X k = Ak e jΨ k
( ) ( )
= Ak cos Ψ k + jAk sin Ψ k (3.28)
= I k + jQ k

where Ak is the spectral amplitude, Ψk is the spectral phase, Ik is the in-phase

spectral component, and Qk is the quadrature-phase spectral component.
Depending on how Ak and Ψk are utilized, different modulated waveforms
are generated.
The IEEE 802.11a utilizes three different modulation techniques: PAM,
PSK, and QAM. For PAM, only the amplitude Ak is changed. For PSK, only
the phase Ψk is changed. The two most commonly used are BPSK and QPSK.
For BPSK, there are only two-phase levels and, for QPSK, there are four phase
levels. For QAM, both amplitude Akand phase Ψk are changed. For 16-QAM,
there are 16 different Xk levels. As Table 2.3 and Figure 2.5 show, there are
four different Ik levels and four different Qk levels. Therefore, the total num-
ber of Xk levels is 4∗4 = 16. For 64-QAM, there are 64 different Xk levels. As

6899_Book.indb 44 10/22/19 4:54 PM

 Multicarrier Modulation45

Table 2.4 and Figure 2.6 show, there are 8 different Ik levels and 8 different
Qk levels. Therefore, the total number of Xk levels is 8∗8 = 64.
Not all of the N subcarriers in an OFDM symbol are used for modula-
tion. The number of subcarriers actually used is M and M < N. The subcar-
rier at DC is normally set null to minimize the impact of a DC component
generated from the receiver front end. Other subcarriers not used are normally
around the band edges. This can reduce the interference from side lobes of the
OFDM transmission. For IEEE 802.11a, N = 64 and M = 52, there are 12
unused subcarriers including the DC. Figure 3.7 shows the region of unused
subcarriers in a dark color. The subcarriers numbered from −26 to −1 and
from 1 to 26 are used.
To modulate the subcarriers, the input bit stream is first grouped into
bauds. Each baud has k bits. The total number of levels is M = 2k or k = log2M.
Based upon the baud, each available subcarrier is modulated using one of the
M levels according to the desirable modulation technique.

Example 3.3
Assume that the input bit stream is 001011011000 using QPSK modulation.
For QPSK, k = 2 and M = 4. Grouping every two bits together, the baud
stream becomes 023120. Using Table 2.2 of the QPSK encoding table, the
Xk sequence becomes 1 2 (−1 − j), 1 2 (1 − j), 1 2 (1 + j), 1 2 (−1 + j),
1 2 (1 − j), and 1 2 (−1 − j).

3.7 OFDM Waveform Properties

All the important operations in OFDM such as linearity, time shift, frequency
shift, and convolution are the same as described in Section 1.8 and are not
repeated here.
The sequence Xnis normally complex. As given in (1.32) and (1.33), Xn
can be real if its spectrum has some symmetry properties. In other words,

Figure 3.7 Unused subcarriers indices in IEEE 802.11a.

6899_Book.indb 45 10/22/19 4:54 PM

 46 Introduction to OFDM Receiver Design and Simulation

Figure 3.8 Unused subcarriers indices in IEEE 802.11a.

the real part of the spectrum has to be even and the imaginary part of the
spectrum has to be odd.
The transmission of the OFDM signal through the media normally suf-
fers from multipath delay and the Doppler shift. The multipath delay is the
operation of a time shift. The impact is the multiplication by a phase factor
in the frequency domain as given in (1.47). The Doppler shift is equivalent
to the multiplication by a phase factor in the time domain. The impact is the
frequency shift in the frequency domain as given in (1.49).

Example 3.4
The null subcarrier region can be plotted in the positive frequency domain
using the periodic property X–k = X N–k. Since N = 64, null subcarriers from a
frequency index −31 to −27 are the same as those from a frequency index 33
to 37. Figure 3.7 is plotted again in Figure 3.8.

3.8 Summary
OFDM utilizes the multicarrier modulation technique through an inverse
DFT, which has a summation of N terms. Each term represents one subcarrier
and there are a total of N subcarriers. Several important characteristics were
discussed: orthogonality, OFDM spectrum, PAPR, cyclic prefix, and FFT.
The orthogonality characteristics showed that any two subcarriers are
orthogonal in an OFDM symbol duration. The OFDM spectrum is a sum-
mation of sinc function from each subcarrier. At the peak for any subcarrier,
the rest of subcarriers exhibit zero crossing. One disadvantage of OFDM is
PAPR because the peak power in a single OFDM symbol can be N times
larger than the average power. This effect can saturate the power amplifier and
cause signal distortion. To reduce the number of computations, the DFT is
performed using the FFT. The number of complex multiplications and addi-
tions for DFT against FFT is N/log2N.

6899_Book.indb 46 10/22/19 4:54 PM

 Multicarrier Modulation47

A cyclic prefix was inserted between two adjacent symbols to remove

the ISI. Assume that this guard interval has a duration of Ng samples and the
channel has a multipath spread of Nc samples. To remove this ISI, a necessary
requirement is for Ng > Nc.
Each OFDM symbol was multiplied by a window function in the time
domain for the purpose of reducing the higher spectral side lobes of the trans-
mitted waveform. This window function used in IEEE 802.11a is a modified
rectangular function. Its frequency spectrum was derived. A parameter was
defined in the window function to control the drop rate of these side lobes.
On digital modulation, the input bit stream was applied to change the
amplitude, phase, or frequency of the complex spectral amplitude. Following
IEEE 802.11a, PSK, QPSK, 16-QAM, and 64-QAM were described. The
signal constellation is very similar to that of the single-carrier modulation. If
the DFT has a dimension N, then there is a maximum of N subcarriers. In
reality, the number of subcarriers used is normally smaller than N to remove
interference of side lobes from the band edges.
In the next chapter, the transmission waveform specified by the IEEE
802.11a is presented.

References
[1] Grangwar, A., and M. Bhardwaj, “An Overview: Peak to Average Power Ratio in
OFDM System & Its Effect,” International Journal of Communication and Computer
Technologies, Vol. 1, No. 2, 2012, pp. 22−25.
[2] O’Neill, R., and L. B. Lopes, “Envelope Variations and Spectral Splatter in Clipped
Multicarrier Signals,” Proceedings of the IEEE International Symposium on Personal,
Indoor, and Mobile Radio Communications, Vol. 1, 2005, pp. 71−75.
[3] Jones, A., and T. A. Wilkinson, “Block Coding Scheme for Reduction of Peak to Mean
Envelope Power Ratio on Multicarrier Transmission Schemes,” Electronic Letters, Vol.
30, No. 25, 1994, pp. 2098−2099.
[4] Oppenheim, A. V., and R. W. Schafer, Digital Signal Processing, Upper Saddle River,
NJ: Prentice Hall, 1975.
[5] Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY)
Specifications, IEEE Standard 802.11a, 2007.

6899_Book.indb 47 10/22/19 4:54 PM

6899_Book.indb 48 10/22/19 4:54 PM
 4
OFDM Transmission

4.1 Introduction
Before data is transmitted through OFDM, the preamble and header fields
are normally sent first. A preamble is usually in a periodic sequence with a
known data structure. Its purpose is to help the receiver in signal detection,
channel estimation, timing synchronization, and frequency offset estima-
tion. A header field normally contains a rate field to inform the receiver of
the transmission data rate.
To help understand the preamble, header, and data structure, the frame
format specified in IEEE 802.11a is used for illustration. However, any other
useful format can be used to meet the specific requirements.
Before getting into the signal transmission format, a general transmit-
ter block diagram is presented first to have an idea of the overall transmitter
architecture. Subsequently, the frame format for preamble, header, and data
is discussed in detail for IEEE 802.11a. Lastly, a typical receiver architecture
is also given in preparation for future discussion of all the major components.

4.2 OFDM Transmitter Architecture

Figure 4.1 is the block diagram of a typical OFDM transmitter. The input
bit sequence can be scrambled first for transmission security. The military

49

6899_Book.indb 49 10/22/19 4:54 PM

 50 Introduction to OFDM Receiver Design and Simulation

Figure 4.1 Typical OFDM transmitter block diagram.

communication normally has a complicated encryption algorithm in place to

scramble the input data. The propagated OFDM signal through the media can
suffer channel degradation such as multipath and fading that cause bit errors
after reception. The scrambled bit sequence is therefore encoded with error
correcting codes in order for the receiver to correct the bit errors. However,
the bit errors can appear in bursts and it is difficult for the receiver to correct.
The encoded bits then go through an interleaving procedure to spread the bits
so as to make the error correction much easier.
The next four blocks, modulation, inverse FFT, cyclic prefix, and win-
dowing, were discussed in Chapter 3. These are key blocks to generate OFDM
discrete samples. These discrete samples are then converted to an analog signal
by passing through a digital-to-analog converter (DAC). At last, this baseband
signal modulates an RF carrier to be transmitted.
Before the data transmission, a preamble signal is normally transmitted
first to help the receiver for signal detection, channel estimation, and tim-
ing synchronization.

4.3 Signal Transmission Format

OFDM signal transmission consists of three parts: preamble, signal header,
and data. They are shown in Figure 4.2. The preamble is a periodic signal. For
IEEE 802.11a, it consists of a short sequence followed by a long sequence. Its
purpose is to help the receiver to perform signal detection, frequency offset
estimation, time synchronization, and channel estimation.
The header field usually takes one OFDM symbol length. The processes
of error correction encoding, interleaving, modulation, inverse FFT, and cyclic
prefix insertion as shown in Figure 4.1 are applied. Their purpose is to send
some key transmission parameters to the receiver. The data modulation format

6899_Book.indb 50 10/22/19 4:54 PM

 OFDM Transmission51

Figure 4.2 OFDM signal transmission format.

and the total length are required for the receiver to demodulate. The tail bits
can be inserted to make a full OFDM symbol. The header field is normally
modulated with BPSK. However, the header bits are not necessarily scrambled.
After both the preamble and header transmission, the user data starts.
The detailed frame format of preamble, header, and data is discussed in the
following sections.

4.4 Preamble Signal for IEEE 802.11a

The preamble is a transmission of fixed signal patterns and is also known to
the receiver. For OFDM, it is normally a periodic signal with a fixed signal
pattern. For each subcarrier k, the complex amplitude Xk = AkejΨk given in (3.4)
is completely specified. The periodic signal plus the unique OFDM character-
istics make the just mentioned receiver tasks easier to perform.
For IEEE 802.11a, the preamble consists of a short sequence and a long
sequence [1]. The short sequence is designed for coarse timing estimation
while the long sequence is used to make further refinements. These two-step
processes can extract the necessary OFDM parameters for the demodulation
of future data signals. Figure 4.3 is the preamble design of IEEE 802.11a.
The short sequence consists of 10 periodic segments and each segment has
the same number of 16 samples [1]. The long sequence consists of two OFDM
symbols. Two guard intervals are followed by two consecutive FFT intervals.

Figure 4.3 Preamble signal format of IEEE 802.11a.

6899_Book.indb 51 10/22/19 4:54 PM

 52 Introduction to OFDM Receiver Design and Simulation

Figure 4.4 The phase of one short sequence segment in the frequency domain.

4.4.1 Short Sequence

Figure 4.4 is a plot of the phase of one fundamental short sequence segment
from frequency index equal to −26 to 26. There are only 12 nonzero subcar-
riers. The complex amplitude Xk is either A(1 + j) or A(−1 − j). The constant
A is selected so that the average power per subcarrier is 1. The total power of
the 12 carriers is 24A2. The average power per subcarrier is 24A2/52 since the
total number of subcarriers from frequency index −26 to 26 is 52. To normal-
ize this average power, A is solved to be 13/6 . The phase angle Ψk is either
π /4 or −3π /4 and the amplitude Ak is a constant equal to 13/6 . Note the
phase is nonzero for every fourth frequency index. Therefore, the samples in
the frequency domain can be downsampled by a factor of 4.
The short sequence specifies 52 subcarriers with the index ranging from
−26 to −1 and from 1 to 26. The data from −26 to −1 can be copied to an
index from 38 to 63. The data from an index from 27 to 37 including index
0 has zero value. After the downsampling by a factor of 4 starting from index
0, only 12 samples are left. Figure 4.5 is a plot of 16 frequency samples against
the frequency index. From this figure, it can be seen that there are only 12
nonzero samples.

Figure 4.5 The phase of downsampled short sequence segment in the frequency
domain.

6899_Book.indb 52 10/22/19 4:54 PM

 OFDM Transmission53

Taking the inverse FFT of the 16 frequency-domain samples, the time-

domain samples for each fundamental segment are obtained. The short
sequence is a repetition of 10 such fundamental segments. The time duration
tshort is given here:

⎛ T fft ⎞ 10NT
tshort = 10 ⎜ = (4.1)
⎝ 4 ⎟⎠ 4

Equation (4.1) shows that the duration increases with increasing sam-
pling period or decreasing channel spacing. The number of samples, Nshort,
is then given here:
10N
N short = (4.2)
4

Therefore, the number of samples is independent of the channel spac-

ing. Assuming a 20-MHz channel spacing, we have T = 0.05 μs and N = 64.
Equation (4.1) gives tshort = 8 μs and Nshort = 160. Each fundamental segment
has 16 samples of duration 0.8 μs.
Before transmission, the 160 time samples must be multiplied by a time
window shown in Figure 3.5. Repeating each fundamental time segment 10
times and multiplying by such a time window, the short sequence is obtained.
Figure 4.6 is an amplitude plot, and Figure 4.7 is a phase angle plot of four

Figure 4.6 The amplitude of the first four segments of a short sequence in the time
domain.

6899_Book.indb 53 10/22/19 4:54 PM

 54 Introduction to OFDM Receiver Design and Simulation

Figure 4.7 Phase angle of the first four segments of a short sequence in the time
domain.

such fundamental segments in a short sequence. Note the amplitude of the

first sample of the first segment is down by a half due to the multiplication
by the window function. Both figures clearly show the pattern repeats every
16 samples.

4.4.2 Long Sequence

The long sequence has two OFDM symbols [1]. Each OFDM symbol consists
of one inverse FFT interval and one guard interval. The duration of the long
sequence is then the following:

( ) ⎛
t long = 2 T fft + Tgi = 2 ⎜ NT +
⎝
NT ⎞
4 ⎟⎠
= (2.5)NT (4.3)

The number of samples Nlong is again independent of the bandwidth

and is given here:
t long
N long = = (2.5)N (4.4)
T

Assume again 20-MHz channel spacing, Equation (4.3) gives tlong = 8 μs

and Nlong = 160. Therefore, the short sequence and the long sequence have the
same number of samples. Also, for a long sequence, the guard interval has a
total of 2 ∗ N/4 = 32 samples while the two FFT intervals have a total of 2
∗ N = 128 samples.
The complex amplitude Xk is either 1 or −1. Therefore, the amplitude
Ak is 1 and the phase angle Ψk is 0° or 180°. Figure 4.8 is a plot of the long

6899_Book.indb 54 10/22/19 4:54 PM

 OFDM Transmission55

Figure 4.8 The complex amplitude of the long sequence in the frequency domain.

sequence amplitude against a frequency index from −26 to 26. Except at DC

or frequency index zero, the complex amplitude is nonzero at all other fre-
quency indexes.
Taking an inverse FFT, 64 time-domain samples can be obtained. Fol-
lowing the design of Figure 4.3, the long sequence of 160 samples is obtained.
After multiplying by a time window, Figure 4.9 is a plot of the long sequence
amplitude and Figure 4.10 is a plot of the long sequence phase in the time
domain. Note that the amplitude and phase are the same between frequency
index from 32 to 95 and a frequency index from 96 to 159. Note also that
the amplitude and phase of the guard interval from a frequency index 0 to
31 are copied from those of a frequency index from 128 to 159. The ampli-
tude at the time index zero is multiplied by a factor of 0.5 because of the
window function.

Figure 4.9 The amplitude of the long sequence in the time domain.

6899_Book.indb 55 10/22/19 4:54 PM

 56 Introduction to OFDM Receiver Design and Simulation

Figure 4.10 The phase angle of the long sequence in the time domain.

4.5 IEEE 802.11a Header Format

For IEEE 802.11a, the header format is shown in Figure 4.11 [1]. There are a
total of 5 fields. The RATE field has 4 bits and can specify a maximum of 16
different combinations of modulation and coding rate. The LENGTH field
has 12 bits and specifies the total data length in bytes. There are six TAIL bits
attached at the end to make for a total of 24 bits. The modulation is BPSK
using convolutional code at a rate of 1/2. Therefore, the total coded bits per
symbol are 48.
Table 4.1 lists eight different data rates specified by IEEE 802.11a. The
convolutional code is selected for the error correction. Based upon a convolu-
tional coding rate R, the other parameters can be either specified or derived.
There are only four different modulation techniques and they are BPSK,
QPSK, 16-QAM, and 64-QAM. The coded bits per subcarrier Nbpsc are 1, 2,
4, and 6, respectively. The coded bits per OFDM symbol Ncbps and data bits
per OFDM symbol Ndbps are then computed as here:
N cbps = N sd N bpsc (4.5)

(
N dbps = Rcoding N cbps ) (4.6)

Figure 4.11 Header format of IEEE 802.11a.

6899_Book.indb 56 10/22/19 4:54 PM

 OFDM Transmission57

where Nsd is the number of data subcarriers and Rcoding is the coding rate.
For a convolutional encoder, the coding rate Rcoding specifies the ratio of
data bits against the coded bits. For IEEE 802.11a. there are 52 subcarriers
with a frequency index ranging from −26 to −1 and from 1 to 26. However,
subcarriers −21, −7, 7, and 21 are reserved for pilots. That means the total
number of data subcarriers is Nsd = 48. The data rate, Rdata is then computed
as follows:
N dbps
Rdata = (4.7)
T fft + Tgi

The parameters Tfft and Tgi are given in Table 3.1.

Example 4.1
Assuming that QPSK and coding rate R = 3/4, then Nbpsc = 2 from Table 4.1
and there are Nsd = 48 data subcarriers using IEEE 802.11a. From (4.5), Ncbps
= 48 ∗ 2 = 96. From (4.6), Ndbps = 96 ∗ 3/4 = 72. At 20 MHz, the symbol
duration is Tfft + Tgti = 4 μs from Table 3.1. From (4.7), the data rate is then
Rdata = 72/4 = 18 Mbps.

Example 4.2
For the header field, the modulation is BPSK and the coding rate is 1/2. From
Table 4.1, we have Ncbps = 48 and Ndbps = 48 ∗ 1/2 = 24. The data rate at 20
MHz is then 24/4 = 6 Mbps.

Table 4.1
IEEE 802.11a Modulation Parameters at 20-MHz Channel Spacing

Modulation R Nbpsc Ncbps Ndbps Data Rate (Mbps)

BPSK 1/2 1 48 24 6
BPSK 3/4 1 48 36 9
QPSK 1/2 2 96 48 12
QPSK 3/4 2 96 72 18
16-QAM 1/2 4 192 96 24
16-QAM 3/4 4 192 144 36
64-QAM 2/3 6 288 192 48
64-QAM 3/4 6 288 216 54

6899_Book.indb 57 10/22/19 4:54 PM

 58 Introduction to OFDM Receiver Design and Simulation

Figure 4.12 Data field of IEEE 802.11a.

4.6 IEEE 802.11a Data Format

Figure 4.12 shows the data format of IEEE 802.11a [1]. There are four fields,
SERVICE, MESSAGE, TAIL, and PAD. The PAD field is used so the total
data length is an integral multiple of Ndbps.
The total number of data bits is given here:
⎛
N data = N dbps ⎜ Ceiling ⎜
(
⎛ 16 + 8N
message
+6 ) ⎞⎟ ⎞⎟ (4.8)
⎜ ⎜ N dbps ⎟⎟
⎝ ⎝ ⎠⎠

where Nmessage is the number of message bytes specified in the LENGTH

field of the header. In (4.8), 16 is the number of SERVICE bits while 6 is the
number of TAIL bits. The number of PAD bits Npad is then computed here:

(
N pad = N data − 16 + 8N message + 6 ) (4.9)

The number of PAD bits is a variable and depends upon the data length
in bytes specified in the header field.

Example 4.3
Assume that the incoming message has 125 bytes using the QPSK and 1/2
coding rate. The data field has 16 + 6 + 8 ∗ 125 = 1,022 bits. From Table
4.1, Ndbps = 48. From (4.8), Ndata = 48 Ceiling (1,022/48) = 48 ∗ 22 = 1,056.
From (4.9), Npad = 1,056 − 1,022 = 34. Therefore, 34 bits must be padded.
After data demodulation, a total of Ncbps bits are collected for each
OFDM symbol. When the processes of deinterleaver, error correction decod-
ing and descrambling are completed, the desired number of Ndbps data bits
are then recovered.

4.7 OFDM Receiver Architecture

Before getting into the various topics related to the receiver design, a typical
receiver architecture is presented first in Figure 4.13. In this diagram, the
major components inside a receiver are given.

6899_Book.indb 58 10/22/19 4:54 PM

 OFDM Transmission59

Figure 4.13 A typical OFDM receiver architecture.

After propagating through the media, the received analog signal goes
through an RF front end, which performs filtering, automatic gain control
(AGC), and the recovery of in-phase and quadrature-phase (IQ) signals. The
IQ signal is subsequently digitized for further processing.
There are two different types of signal to be received: preamble and data.
If the received signal is a preamble, the channel estimate and synchronization
are performed. However, if the input is data, the cyclic prefix is first removed
before performing the FFT. During the data transmission, not all the subcarri-
ers are dedicated to data. Some are pilot subcarriers to help channel estimation
and synchronization. This is true not just for 802.11a but also for 4G-LTE.
Subsequently, the channel degradation is removed based upon an earlier
channel estimate. The channel estimation can also be updated concurrently
using the LMS algorithm. After that, the received IQ signal is demodulated
to get the bit stream. This bit stream subsequently goes through the deinter-
leaver, error correction decoder, and, finally, the descrambler to get the original
transmitted data.
The detailed processing of every box will be discussed in subse-
quent chapters.

4.8 Summary
A general transmitter block diagram was first given to have a picture of the over-
all OFDM architecture. The transmission frame format including preamble,

6899_Book.indb 59 10/22/19 4:54 PM

 60 Introduction to OFDM Receiver Design and Simulation

header, and data was subsequently discussed. The frame format specified in
IEEE 802.11a was used for illustration.
A preamble consists of a short sequence and a long sequence in IEEE
802.11a. The short sequence has 10 periodic segments and a total of 160 sam-
ples. The long sequence also has 160 samples including two guard intervals of
32 samples and two FFT intervals of 128 samples.
The header field specifies the data rate to inform the receiver for demodu-
lation; it also has a LENGTH field specifying the total data length in bytes
for transmission. The data field has tail bits to allow the convolutional encoder
to return to the zero state. It also inserts some pad bits so the total duration
has an integral number of data bits per OFDM symbol.
Finally, a typical receiver architecture was presented to show the major
components inside a receiver. Each component in both the transmitter and
receiver architecture will be discussed in subsequent chapters.
In the next chapter, the shift-register sequence, data scrambler, and
descrambler are discussed.

Reference
[1] Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY)
Specifications, IEEE Standard 802.11a, 2007.

6899_Book.indb 60 10/22/19 4:54 PM

 5
Shift Register Sequence and
Data Scrambler

5.1 Introduction
In digital communication, the input data is often scrambled before the modula-
tion and transmission. The receiver knows exactly how to generate the scram-
bler sequence used in the transmitter to recover the original data. For military
communication, the sequence period is extremely long and is difficult for the
enemy to descramble. For commercial communication, it is long enough to
protect the security of the user data.
IEEE 802.11a also defines a scrambler generated from a shift register
sequence. Depending upon the number of shift registers, this sequence has
a unique period of maximum length. To understand the process of sequence
generation from the shift registers, we start with a binary field, and a Galois
field is also discussed. The details of sequence generation from the shift reg-
isters and its period determination are subsequently presented.
Finally, a data scrambler is discussed. The process of data scrambling in
the transmitter and descrambling in the receiver is also discussed.

61

6899_Book.indb 61 10/22/19 4:54 PM

 62 Introduction to OFDM Receiver Design and Simulation

5.2 Binary Field

A Galois field [1] with p elements is normally represented as GF(p) where p is
a prime. If p = 2, it is a binary field GF(2) having only two elements. These
two elements are 0 and 1. The operation of multiplication and addition are
defined in Tables 5.1 and 5.2.
The multiplication has the symbol “.” while the addition has the symbol
“+.” Both addition and multiplication are performed under a modulo-2 opera-
tion. For example, 1 + 1 (modulo-2) = 0. The field is closed since both addi-
tion and multiplication generate an element of either 0 or 1 in the same field.
The operation of subtraction “−” and division “/” can also be defined.
Assuming that a and b are two elements in a field, then a − b = a + (−b). The
−b is defined such that b + (−b) = 0. The element −b is also called the additive
inverse element. Since the operation is modulo-2, −b is then equal to 2 − b.
For b = 1, −b = 1, and for b = 0, −b = 0.
The division of a by b is defined as a/b = a.b−1. The element b−1 is also
called a multiplicative inverse element. It is defined such that b.b−1 = 1. For b
= 1, b−1 = 1. The element 0 is excluded from having a multiplicative inverse.
Table 5.1 also shows that the addition is commutative and associative.
Assuming that a, b, and c are three numbers from GF(2), then clearly a + b
= b + a and a + (b + c) = (a + b) + c. Tables 5.1 and 5.2 also show that the
multiplication is distributive since a(b + c) = (a.b) + (a.c).

Table 5.1
Binary Addition

+ 0 1

0 0 1

1 1 0

Table 5.2
Binary Multiplication

. 0 1

0 0 0

1 0 1

6899_Book.indb 62 10/22/19 4:54 PM

 Shift Register Sequence and Data Scrambler63

A polynomial whose coefficients are from GF(2) can also be defined.

Assuming that f(x) of degree n is such a polynomial, it is given by

f (x) = a0 + a1 x + a2 x 2 + a3 x 3 + … + an x n (5.1a)

where an = 0 or 1. Similarly, a polynomial g(x) of degree m is defined here:

g(x) = b0 + b1 x + bx 2 + bx 3 + … + bm x m (5.1b)

where bm = 0 or 1. The addition of f(x) and g(x) is then given here:

( )
f (x) + g(x) = a0 + b0 + ( a1 + b1 ) x + ( a2 + b2 ) x 2 + … (5.2)

The subtraction of g(x) from f(x) can be similarly defined:

( )
f (x) − g(x) = a0 − b0 + ( a1 − b1 ) x + ( a2 − b2 ) x 2 + …
(5.3)
( ( )) ( ) (
= a0 + −b0 + a1 + ( −b1 ) x + a2 + ( −b2 ) x + … ) 2

Example 5.1
Assuming that f(x) = 1 + x 2 + x3 and g(x) = x + x 2 + x3 + x4, then

The multiplication of f(x) by g(x) is given here:

(
f (x)g(x) = a0 + a1 x + a2 x 2 + a3 x 3 + … + an x n )( b
0
+ b1 x + b2 x 2 + b3 x 3 + … + bm x m )
( ) ( )
= a0b0 + a0b1 + a1b0 x + a0b2 + a1b1 + a2b0 x + …anbm x2 n+m

= c0 + c1 x + c2 x + c3 x + … + cn+m x
2 3 n+m

(5.4a)
where cn is in general given by
cn = a0bn + a1bn−1 + a2bn−2 + … + anb0 (5.4b)

Example 5.2
Assume f(x) = 1 + x + x 2 and g(x) = 1 + x 2, then the product of f(x) and g(x)
is given by

6899_Book.indb 63 10/22/19 4:54 PM

 64 Introduction to OFDM Receiver Design and Simulation

f (x)g(x) = 1 + x + x 2 + x 2 + x 3 + x 4 = 1 + x + (1 + 1)x 2 + x 3 + x 4
= 1 + x + x3 + x 4

Equations (5.4a) and (5.4b) can also be used to compute the product. In
this case, a 0 = 1, a1 = 1, a2 = 1, a3 = 0, a4 = 0 and b0 = 1, b1 = 0, b2 = 1, b3 = 0,
b4 = 0. From (5.4b), we have c 0 = a 0b0 = 1, c1 = a 0b1 + a1b0 = 1.0 + 1.1 = 1, c 2
= a 0b2 + a1b1 + a2b0=1.1 + 1.0 + 1.1 = 0, c3 = a 0b3 + a1b2 + a2b1 + a3b0 = 1.0
+ 1.1 + 1.0 + 0.1 = 1 and c4 = a2b2 = 1. Therefore, f(x)g(x) = 1 + x + x3 + x4
and is the same as given earlier.
The division in GF(2) can be performed through long division. If the
quotient is q(x) and the remainder is r(x), then f(x) = q(x)g(x) + r(x). The
degree of r(x) is less than q(x). The process of long division is shown through
the following example.

Example 5.3
Assuming that f(x) = x4, g(x) = 1 + x + x 2, then f(x)/g(x) is given here:
x2 + x
x2 + x + 1 x 4
x 4 + x3 + x2
x3 + x2
x3 + x2 + x
x

Therefore, the quotient q(x) is x 2 + x and the remainder r(x) is x. To prove

its correctness, we have f(x) = q(x)g(x) + r(x) = (x 2 + x) (x 2 + x + 1) + x = x4 +
x 3 + x 2 + x 3 + x 2 + x + x = x 4.
If r(x) = 0, then f(x) = q(x)g(x) and f(x) is divisible by g(x) If x is a root
of f(x), then f(x) = 0. As an example, x = 1 is a root of f(x) = x4 + x3 + x 2 + x
since f(1) = 0. Through long division, it can be shown that f(x) = x4 + x3 + x 2
+ x = (x + 1)(x3 + x). This shows that f(x)may have roots not in GF(2). It is
similar to the case that a real polynomial may have complex roots.

5.3 Galois Field

In general, a Galois field has more than two elements. Of particular inter-
est is the field GF(2m) and m > 1 [1]. To construct such a field, a primitive

6899_Book.indb 64 10/22/19 4:54 PM

 Shift Register Sequence and Data Scrambler65

polynomial needs to be defined. Assume that p(x) is a polynomial with degree

m over GF(2). It is primitive if the smallest positive n such that p(x) divides
xn + 1 is n = 2m − 1and p(x) is not divisible by any polynomial of degree less
than m. For example, p(x) = 1 + x + x3 is primitive since m = 3 is the smallest
degree such that p(x) divides x 7 + 1. Through long division, it can be shown
that x 7 + 1 = (x4 + x 2 + x + 1)(x 3 + x + 1). The polynomial p(x) is also not
divisible by any polynomial of degree less than 3. Therefore, p(x) is primitive.
In a binary field of GF(2), it is a modulo-2 operation. In the extension
field of GF(2m), it is an operation by performing modulo p(x) and p(x) is a
primitive polynomial. Equivalently, p(x) = 0 and is similar to the binary field
by setting 2 to zero. Therefore, no polynomial in GF(2m) can have a degree
m m m
greater than m. Because x 2 −1 +1 is divisible by p(x), x 2 −1 + 1 = 0 or x 2 −1 =
1. Assume that y is an element of GF(2m). To construct this field, we start
with the number 1 and continue multiplying by y. A new element is generated
m
with every multiplication by y. This process continues until y2 −2 is reached.
Including 0, a total of 2m elements is then generated. Using the property p(y)
m
= 0 and y2 −1 = 1, each field element can be reduced to a polynomial with a
degree less than m.

Example 5.4
Assume that the primitive polynomial is p(x) = 1 + x + x3. Then we have x 3 =
x + 1 and x 7 = 1. The 8 field elements are then generated here:

0
1
y
y2
y3 = y + 1
y 4 = y( y + 1) = y2 + y
y5 = y ( y 2 + y ) = y 3 + y 2 = y 2 + y + 1
y6 = y ( y2 + y + 1) = y3 + y2 + y = y + 1 + y2 + y = y2 + 1

Since y7 = 1, the process repeats and there are at most 8 elements in GF(8).
To prove y 7 = 1, we have y 7 = y(y2 + 1) = y3 + y = y + 1 + y = 1. Also, from the
listing above, every element polynomial has degree of 2 or less.
If the primitive polynomial has degree of m, then all the field element
polynomials can be represented as a0 + a1y + a2y2 + ⋯ + am–1ym−1. The coefficient

6899_Book.indb 65 10/22/19 4:54 PM

 66 Introduction to OFDM Receiver Design and Simulation

an is over GF(2) and is either 0 or 1. The total number of terms is m. Therefore,

the total number of field elements is 2m.
The multiplicative inverse element and additive inverse element also exist
in GF(2m). The additive inverse element is the element itself. Therefore, the
additive inverse element of yi is yi. The multiplicative inverse element of yi is y−i
m m
= y2 −1y−i = y2 −i−1. From Example 5.4, the multiplicative inverse element of yi
is y 7−i. For example, the multiplicative inverse element of y3 is y4 and so forth.
The Galois field GF(2m) is also closed under the addition and multipli-
cation [1]. This means that the addition or multiplication of any two field ele-
ments is also an element in GF(2m). For example, consider two field elements
y and y5. From Example 5.4, the addition of y and y5 is y2 + 1 = y6, which is
also a field element. From Example 5.4 again, the multiplication of y and y5
is y 6, which is clearly a field element. However, this is true for any two field
elements. This illustrates that the Galois field GF(2m) is closed under the addi-
tion and multiplication.
The Galois field GF(2m) also satisfies the commutative, associative, and
distributive laws. Every polynomial in GF(2m) has coefficients over GF(2). As
has been shown before, these polynomials satisfy these laws. Therefore, these
laws are also satisfied over GF(2m).
Another interesting property of GF(2m) is that every field element of
m
the form yi where i < 2m − 1 is a root of f(x) = x 2 −1 + 1 [1]. This can be easily
proven as follows:

f (x) = f ( y i ) = y i(2 ( ) +1 = 1 +1 = 1+1 = 0

i
m
−1) m
−1)i m
−1
+ 1 = y(2 + 1 = y2 i

m
Therefore, the polynomial sequence 1, y, y2 + ⋯ + y2 −2
are all the roots
m
of x 2 −1 + 1 = 0.

5.4 Sequence Generator

The shift registers can be used to generate a pseudonoise (PN) sequence with
the desired period. Figure 5.1 is a logical diagram to generate such a sequence
through polynomial division [1, 2]. The circle with a + sign inside represents
summation and the circle with z−1 inside means a delay by one sample. Finally,
the circle with a coefficient gn inside represents a multiplication by gn. On the
rightmost shift register, yn is an output and input to the shift register. On the
leftmost register, there is a feedback path back to the input. All the polynomial
summations and multiplications are performed under modulo-2 arithmetic.

6899_Book.indb 66 10/22/19 4:54 PM

 Shift Register Sequence and Data Scrambler67

Figure 5.1 Linear feedback shift register performing polynomial division.

Based upon Figure 5.1, we have the following difference equation:

yn = g1 yn−1 + g 2 yn−2 + … + g r−1 yn−r+1 + g r yn−r + A0 (5.5)

where A0 is the initial condition on the shift register. Performing a z-transform

using the time shift operation given by (1.57) on both sides of (5.5), we have

(
Y (z) = Y (z) g1z −1 + g 2 z −2 + … + g r−1z −r+1 + g r z −r ) (5.6)
+ A(z)

where A(z) is given by the following equation:

A(z) = a0 + a1z −1 + a2 z −2 + … + ar−1z −r+1 (5.7)

The coefficients from a0 to ar–1 are initial loads to the r shift registers.
The coefficient a 0 is the initial load to the leftmost shift register while ar–1 is
the initial load to the rightmost shift register. The coefficient an of the nth
shift register is either 0 or 1. Combining terms with Y(z) together, we have

( )
Y (z) 1 − g1z −1 − g 2 z −2 + … + g r−1z −r+1 − g r z −r = A(z) (5.8)

Since gn is binary, (5.8) after a modulo-2 operation becomes

A(z)
Y (z) = −1 −2
(5.9)
1 + g 1z + g 2 z + …+ g r−1z −r+1 + g r z −r

6899_Book.indb 67 10/22/19 4:54 PM

 68 Introduction to OFDM Receiver Design and Simulation

Substituting x = z−1 into (5.9), we have

a0 + a1 x + a2 x 2 + … + ar−1 x r−1
y(x) = (5.10)
1 + g1 x + g 2 x 2 + … + g r−1 x r−1 + g r x r

Assuming that the nominator polynomial is A(x) and the denominator

polynomial is G(x), we have

A(x) = a0 + a1 x + a2 x 2 + … + ar−1 x r−1

G(x) = 1 + g1 x + g 2 x 2 + … + g r−1 x r−1 + g r x r (5.11)
A(x)
y(x) =
G(x)

Note that a 0 is the content of the leftmost shift register while ar–1 is the
content of the rightmost shift register [2]. Before performing the division, A(x)
is modified to B(x) such that B(x) is the polynomial of A(x)G(x) after remov-
ing all terms with a degree greater than r − 1 [2].

Example 5.5
Assume that G(x) = 1 + x + x3 and the initial load to the shift registers is 001
as illustrated in Figure 5.2. There are two methods to get the output. The first
is to step through the shift register operations cycle by cycle and the second
is through the polynomial divisions. Using the first method, the shift register
content and output are given in Table 5.3.
During each cycle, the content of the shift register shifts one place to
the left. The initial condition is that the shift register content is 001 at the first
cycle. In the second cycle, the content of the third register becomes 0 while
the content of the second register becomes 1. The summation result is shifted
to the first register and becomes 0 + 1 = 1 and the output is 1.

Figure 5.2 Sequence generator for Example 5.5.

6899_Book.indb 68 10/22/19 4:54 PM

 Shift Register Sequence and Data Scrambler69

Table 5.3
Register Content and Output for Example 5.5

Cycle Register 3 Register 2 Register 1 Output

1 0 0 1 1
2 0 1 1 1
3 1 1 1 1
4 1 1 0 0
5 1 0 1 1
6 0 1 0 0
7 1 0 0 0

Therefore, the content of the shift registers becomes 011 after the first
cycle. For the other cycles, the register content can be obtained in the same
way. After the seventh cycle, the content of the shift register repeats so the
period is 7.
Using the second method, the polynomial B(x) needs to be determined
first. From the initial load, A(x) = x 2 and A(x)G(x) = x 2(1 + x + x3) = x 2 + x3
+ x5. After removing all the terms with a degree greater than 2, we have B(x)
= x 2. The long division is then performed here:

x2 + x3 + x 4 + x6 + x9
1 + x + x3 x2
x2 + x3 + x5
x +
3
+ x5
x +x
3 4
+ x6
x 4 + x5 + x 6
x 4 + x5 + x7
x6 + x7
x6 + x7 + x9
x9

Observe that the coefficients of the quotient polynomial starting at x 9

repeats. Therefore, the long division generates a periodic output of 1110100,
which is the same as that from method 1. Since there are only three registers,
the maximum number of unique register contents is 23 − 1 = 7. Therefore, 7
is also the maximum period.

6899_Book.indb 69 10/22/19 4:54 PM

 70 Introduction to OFDM Receiver Design and Simulation

Figure 5.3 Sequence generator for Example 5.6.

Example 5.6
Consider another example with G(x) = 1 + x3 illustrated in Figure 5.3. The
first method is used to find the output with different initial register contents.
The results are given in Table 5.4.
As can be seen from Table 5.4, different initial register content gener-
ates different output and period. For the initial register content 001, 111, 000,
110, the output sequence is 100, 111, 000, 011, respectively, and the period
is either 1 or 3.

5.5 Period of Sequence Generator

Assume a sequence generator has any polynomial G(x) of degree r defined by
(5.11). An inverse polynomial G −1(x) can be defined as follows:

⎛ 1⎞
G −1(x) = x rG ⎜ ⎟ (5.12)
⎝ x⎠

where r is the number of shift registers. It can be shown [2] the maximum
period N of a shift register sequence with polynomial G(x) is the smallest N
such that xN + 1 is divisible by G −1(x). This is demonstrated by the follow-
ing example.

Table 5.4
Register Content for Example 5.6

Initial Register
Contents/Cycles 001 111 000 110

1 010 111 000 101

2 100 111 000 011
3 001 111 000 110

6899_Book.indb 70 10/22/19 4:54 PM

 Shift Register Sequence and Data Scrambler71

Figure 5.4 Sequence generator for Example 5.7.

Example 5.7
Let G(x) = 1 + x + x3 + x4 for the sequence generator illustrated in Figure 5.4.
The inverse polynomial is then given by

⎛ 1⎞
G −1(x) = x 4G ⎜ ⎟ = x 4 ( x −4 + x −3 + x −1 + 1) = 1 + x + x 3 + x 4
⎝ x⎠

Through long division, it can be shown that

x 6 + 1 = ( x 4 + x 3 + x + 1)( x 2 + x + 1)
x12 + 1 = ( x 4 + x 3 + x + 1)( x 8 + x 7 + x 6 + x 2 + x + 1)

Both x6 + 1and x12 + 1 are divisible by (x4 + x3 + x + 1), but the smallest
degree is 6 and the period is then 6. This can be verified by going through the
state sequence as given by Table 5.5.
The maximum number of states for 4 registers is 15 and all the 15 states
appear in Table 5.3 and there are no other periods. This proves that the maxi-
mum period is 6.

Table 5.5
Register Content of Example 5.7

Initial Register
Contents/Cycles 0000 0101 1111 1011 0011 0010

1 0000 1010 1111 0110 0111 0100

2 0101 1101 1110 1001
3 1011 1100 0010
4 1000
5 0001
6 0011

6899_Book.indb 71 10/22/19 4:54 PM

 72 Introduction to OFDM Receiver Design and Simulation

5.6 Maximum-Length Sequences

The period of a shift register sequence for any polynomial G(x) has been defined
in the previous section. It is the smallest N such that xN + 1 is divisible by the
inverse polynomial G −1(x). If G(x) is a primitive polynomial as discussed in
Section 5.3, it has been shown [2] that G −1(x) is also primitive. If the primi-
tive polynomial has a degree r, the smallest degree N such that G(x) divides
xN + 1 is N = 2r − 1. Since the number of shift registers is also r, the maxi-
mum number of possible state sequences is N = 2r − 1. Under this condition,
the maximum period is 2r − 1. The shift register sequence that generates the
maximum period is then called the maximum-length sequence.
Table 5.6 lists some representative primitive polynomials of degrees less
than 10 [1]. A more comprehensive listing of primitive polynomials can be
found in [2].

5.6.1 Properties of the Maximum-Length Sequence

In Example 5.5, G(x) = 1 + x + x 3 is a primitive polynomial. The output
sequence is 1110100 if the initial state is 001. As indicated, the period is 7.
An examination of the output sequence shows that there are 4 ones and 3
zeros. The number of ones is one more than the number of zeros. This is, in
general, true for any maximum length sequence. Since the state sequence goes
through all the possible states, the number of odd numbers is one more than

Table 5.6
Listing of Some Primitive Polynomials

Degree Primitive Polynomial Maximum Period

2 1 + x + x2 3
3 1+x+x 3 7
4 1+x+x 4 15
5 1 + x2 + x 5 31
6 1+x+x 6 63
7 1 + x3 + x 7 127
8 1+x +x +x +x
2 3 4 8 255
9 1+x +x
4 9 511
10 1 + x3 + x10 1023

6899_Book.indb 72 10/22/19 4:54 PM

 Shift Register Sequence and Data Scrambler73

the number of even numbers. Examining the register content of Figure 5.2,
it can be seen that there are 4 odd numbers and 3 even numbers. Since the
output is shifted to the rightmost register, the number of ones is one more
than the number of zeros.
Including two periods, the output sequence of Example 5.5 is
11101001110100. If the initial content of the register is changed to 110, the
output sequence becomes 01001110100111. Comparing the two sequences,
the second sequence starts from the fourth digit of the first sequence. There-
fore, the same sequence is generated except for the phase difference. That
this is true is because the maximum-length sequence goes through all the
possible states. Therefore, for a fixed primitive polynomial, the maximum
length sequence is independent of the register content except for the phase.
The period is also always the same independent of the content of the initial
register state.

5.6.2 Sequence Generator from the IEEE 802.11a

In the IEEE 802.11a, a sequence generator is specified to scramble the input
data. The primitive polynomial used in this sequence generator has the seventh
degree. From Table 5.6, it is given here:

G(x) = 1 + x 3 + x 7 (5.13)

The inverse polynomial is also primitive and is from (5.12) with r = 7:

G −1(x) = x 7 (1 + x −3 + x −7 ) = 1 + x 4 + x 7 (5.14)

The IEEE 802.11a uses the inverse polynomial defined in (5.14) as the
generator and is shown in Figure 5.5. The period is then 27 − 1 = 127. The
initial state is all ones and the output repeats every 127 digits.

Figure 5.5 Sequence generator for the IEEE 802.11a.

6899_Book.indb 73 10/22/19 4:54 PM

 74 Introduction to OFDM Receiver Design and Simulation

Figure 5.6 Data scrambler.

5.7 Data Scrambler

For military communications, it is required that the incoming data be
encrypted before being transmitted. The encryption algorithm is sophisticated
so the enemy receiver cannot decrypt it. For commercial communications,
it is also necessary to protect sensitive data, which is often scrambled before
transmission. Independent of the complexity of the sequence generator, the
data scrambler is shown in Figure 5.6.
The data scrambler consists of two parts. The first part is to generate the
random bit stream. The second part is the modulo-2 sum of the input data
and the random bit stream. For military communications, the random bit
generator has an extremely long period. For commercial communications, a
maximum-length sequence generator is frequently used to generate the pseudo-
random sequence. The IEEE 802.11a uses the polynomial defined in (5.14) to
generate a maximum length sequence with a period of 127 bits.
To recover the original data, the receiver has to know the exact random
bit generator used in the transmitter. The operation is the reverse process used
in the transmitter. The modulo-2 sum of the received data and the output
stream from the random-bit generator generates the original data. Figure 5.7
shows the process of data descrambling.

Figure 5.7 Data descrambler.

6899_Book.indb 74 10/22/19 4:54 PM

 Shift Register Sequence and Data Scrambler75

Example 5.8
Assume that the sequence generator uses G(x) = 1 + x + x3 defined in Example
5.5. The pseudo-random sequence with period 7 is 1110100. Assuming that the
input data is 01101100101000, then the scrambler output is 10000101011100
by performing the modulo-2 sum of 11101001110100 and 01101100101000.
In the receiver, the modulo-2 sum of the received data 10000101011100 and
11101001110100 generates the original data of 01101100101000.

5.8 Summary
A binary field of GF(2) has only two elements, 0 and 1. The operation of
addition, subtraction, multiplication, and division can be defined. The addi-
tion is commutative and associative while the multiplication is distributive. A
polynomial whose coefficients over GF(2) was then defined. The polynomial
addition, multiplication, and division were illustrated.
A Galois field of GF(2m) has a total of 2m elements. The operation in this
field is over a primitive polynomial whose coefficients are over GF(2). Based
upon this primitive polynomial, every element can be generated. This primi-
tive polynomial has a degree m and no element polynomial can have a degree
greater than m. In this field, the commutative, associative, and distributive
laws are all satisfied and the operation of addition and multiplication is closed.
Another interesting property is that every element in this field is also a root of
m
the polynomial f(x) = x2 −1 + 1 where m is the order of the primitive polynomial.
A sequence generator through shift registers was then discussed. There
are two methods to get the output. The first is to step through the shift reg-
ister operations cycle by cycle. The second is through polynomial divisions
over a generator polynomial. The output through polynomial division was
derived in detail.
The generator polynomial has an inverse polynomial. The maximum
period of a shift register sequence with a given generator polynomial depends
upon this inverse polynomial. Based upon this inverse polynomial, the period
depends upon the initial condition of the shift register.
A maximum-length sequence was then defined. To generate this sequence,
the generator polynomial is primitive. If the primitive polynomial degree is
r, the maximum period is 2r − 1. There are two interesting properties. The
first is that the number of ones is one more than the number of zeros in the
output. The second is that the output is independent of the initial condition
of the shift register except for the phase.

6899_Book.indb 75 10/22/19 4:54 PM

 76 Introduction to OFDM Receiver Design and Simulation

Using the maximum-length sequence, a scrambler and descrambler were

then discussed. Both operations are through a simple modulo-2 operation.
The next chapter centers on wave propagation loss. Both a large-scale
model and a small-scale model are discussed, and the noise model for the
amplifier, cable, and receiver is presented.

References
[1] Lin, S., and D. Costello, Error Control Coding, Upper Saddle River, NJ: Prentice Hall,
1983.
[2] Peterson, R. L., R. Ziemer, and D. Borth, Introduction to Spread Spectrum, Upper
Saddle River, NJ: Prentice Hall, 1995.

6899_Book.indb 76 10/22/19 4:54 PM

 6
Radio-Wave Propagation Model

6.1 Introduction
In a wireless communication system, an RF carrier is modulated by the base-
band signal and propagates through the air. Along the propagation path, it
may suffer three major types of obstacles: reflection, diffraction, and scattering.
The reflection and scattering are caused by objects with a dimension greater
than and smaller than the wavelength, respectively. The reflection can occur
from the ground, buildings, or mountains. If scattering occurs, the output
wave may travel in different directions in a random manner. The diffraction
is caused by an object with sharp edge and the output wave may even travel
behind the object.
As a result of these three mechanisms, the receiver may receive signals
traveling from different directions to cause multipath distortion. This phe-
nomenon can be characterized from the point of view of large-scale path loss
or small-scale fading.
The large-scale propagation loss centers on how the received power varies
as a function of the distance between the transmitter and the receiver. As the
propagation distance increases, the received power decreases. However, the
signal may be received from different paths with varying phase and amplitude
fluctuations. The small-scale fading model concentrates on how the received
power fluctuates with distance around a few wavelengths.

77

6899_Book.indb 77 10/22/19 4:54 PM

 78 Introduction to OFDM Receiver Design and Simulation

6.2 Large-Scale Propagation Model

Two well-known models, free space and two-ray, are presented next. The
empirical model can also be used through curve fitting of the experimental
data. A frequently used empirical formula is also discussed.

6.2.1 Free-Space Propagation Loss

If the electromagnetic wave propagates through a free space, then the power
received by a receiver at a distance d from the transmitter is related by the
following formula [1],
PtGtGr c 2
Pr = (6.1)
Ω2 d 2

where Pr is the received power, Pt is the transmitted power, Gt is the transmitter

antenna gain, Gr is the receiver antenna gain, c is the velocity of light, Ω is the
radian frequency, and d is the propagation distance. If the carrier frequency is
f, then Ω = 2π f. Therefore, the received power is inversely proportional to the
square of both distance and frequency. Lower frequency and shorter distance
can increase the power received.
Equation (6.1) can also be written using the effective isotropic radiated
power (EIRP). The EIRP is the power radiated from an isotropic antenna with
unit gain uniformly in all directions and is defined here:
EIRP = PtGt (6.2)

Using (6.2), the received power then becomes

(EIRP)Gr c 2
Pr = (6.3)
Ω2 d 2

The free-space propagation loss in decibels is defined here:

⎛P⎞
PL(dB) = 10log ⎜ t ⎟
⎝ Pr ⎠ (6.4)
(
= 20log(d) + 20log(Ω) − 10log GtGr c 2 )
Example 6.1
Assuming Gt = Gr = 1, Figure 6.1 is a plot of propagation loss with respect
to the distance from 1 km to 4 km as the frequency increases from 400

6899_Book.indb 78 10/22/19 4:54 PM

 Radio-Wave Propagation Model79

Figure 6.1 Free-space propagation loss in decibels.

MHz to 1.2 GHz. Clearly, the loss increases with the increasing distance
and frequency.

Example 6.2
The received power in decibels from (6.4) is 10log(Pr) = 10log(Pt) − PL (dB). If
the transmitted power is 1W or 30 dBm, the received power is then 30 dBm
− PL (dB). Using the same parameters as in Figure 6.1, Figure 6.2 shows the
received power in dBm.

Figure 6.2 Power received using 1W of transmission power.

6899_Book.indb 79 10/22/19 4:54 PM

 80 Introduction to OFDM Receiver Design and Simulation

6.2.2 Two-Ray Model

The free space propagation model does not take into account any obstacles
along the propagation path. This model is good if the line-of-sight path is
high above the ground. The well-known two-ray model considers two propa-
gation paths. The first is the direct path from the transmitter to the receiver
and the second is a path reflected from the ground. The situation is illustrated
in Figure 6.3.
At the receiver, the total electric field is the sum received from both
paths. The power received at receiver R is given by the following formula [1]

PtGtGr ht2 hr2

Pr = (6.5)
d4

where ht is the transmitter antenna height, hr is the receiver antenna height,

and d is the distance. Equation (6.5) shows that the received power decreases
against the distance much faster than the free-space loss. In the free space,
the path loss is proportional to d2. However, in the two-ray model, it is pro-
portional to d 4.
The path loss in decibels is 10log(Pt/Pr) and can be written as

( )
PL (dB) = 40log(d) − 10log GtGr − 20log ht hr ( ) (6.6)

As is evident from (6.6), the path loss increases with increasing distance d
but decreases with increasing antenna gain and antenna height. One drawback
of (6.6) is that there is no frequency dependence. However, the two-ray model
is found to predict the path loss reasonably well at large distances.

Figure 6.3 Two-ray model.

6899_Book.indb 80 10/22/19 4:54 PM

 Radio-Wave Propagation Model81

Example 6.3
Assume that both the transmitter and receiver have unit gain and the antenna
height for both transmitter and receiver is 1m. The path loss is plotted against
d in Figure 6.4 by setting Gt = Gr = 1 and ht = hr = 1m. Comparing Figures
6.1 and 6.4 at the same distance, it is evident that the two-ray model has a
much larger path loss than that predicted by free space.

6.2.3 Empirical Model

The free-space and two-ray models predict the propagation loss either in free
space or close to the ground. The actual propagation loss is quite complicated
and depends upon many factors. Even for propagation close to the ground,
the loss may depend upon ground conductivity, farmland, and desert. It may
also depend upon the environment such as forests, open space, and weather.
The indoor propagation or obstacles such as buildings or mountains will also
impact the propagation loss. In reality, the propagation loss is measured in
the field and predicted through curve fitting.
The long-distance propagation loss that fits many theoretical models can
generally be given by the following formula:

⎛d⎞ ⎛ f ⎞
( )
PL(d, f ) = PL dr , f r + 10nlog ⎜ ⎟ + 10mlog ⎜ ⎟ + O (dB)
⎝ dr ⎠ ⎝ fr ⎠
(6.7)

Figure 6.4 Path loss of two-ray model against the distance.

6899_Book.indb 81 10/22/19 4:55 PM

 82 Introduction to OFDM Receiver Design and Simulation

where n and m are variables. The parameter n depends upon the environment
and is normally greater than 2 for urban areas and with building obstacles.
Since the propagation loss increases with increasing frequency, the third term
is added to emphasize frequency dependence. The last term, O, is added so as
to include any additional variations against frequency, distance, and others.
The parameter dr is the reference distance and the parameter f r is the reference
frequency. If d = dr and f = f r, then PL(d, f ) reduces to PL(dr, f r). The refer-
ence distance dr is normally set at 1 km and PL(dr, f r) is measured in the field
using the desired reference frequency f r. If only one frequency is of interest,
then f = fr and the third term is 0. However, if the frequency is changed to a
different frequency, then (6.7) can be used to predict the path loss using the
new frequency f. If there is no other complicated dependence, then the last
term is also 0.
To see how (6.7) can be applied to the theoretical model, consider first
the free-space formula given in (6.4). By performing PL(d) − PL(dr), we have
the following

⎛d⎞ ⎛ f ⎞
( )
PL(d) = PL dr + 20log ⎜ ⎟ + 20log ⎜ ⎟
d
⎝ r⎠ ⎝ fr ⎠
(6.8)

Comparing (6.7) against (6.8), we have n = 2, m = 2. Note that the O

term is zero since there is no other dependence on frequency and distance.
Consider next the two-ray formula given in (6.6). By again performing
PL(d) − PL(dr), the following is obtained

⎛d⎞
( )
PL(d) = PL dr + 40log ⎜ ⎟
⎝ dr ⎠
(6.9)

Comparing (6.7) against (6.9), we have n = 4 and m = 0, and the O

term is 0.
A more complicated path loss in the urban area using frequency from
150 MHz to 1,500 MHz is given below by Hata [1]

(
PL(d, f ) = 69.55 + 26.16 log f − 13.82log hte − a hre , f ) (6.10)
( )
+ 44.9 − 6.55log hte log d

where f is the frequency in megahertz, hte is the effective transmitter antenna

height, and hre is the effective receiver antenna height. Both hte and hre are in

6899_Book.indb 82 10/22/19 4:55 PM

 Radio-Wave Propagation Model83

units of meters. The variable a(hre, f ) is a correction factor depending upon

the city size and frequency. By setting f = fr and d = dr, (6.10) then becomes

(
PL(dr , f r ) = 69.55 + 26.16 log f r − 13.82log hte − a hre , f r )
(6.11)
( )
+ 44.9 − 6.55log hte log dr

Subtracting (6.11) from (6.10), we have the following:

⎛ f ⎞
(( ) (
PL(d, f ) = PL(dr , f r ) + 26.16 log ⎜ ⎟ − a hre , f − a hre , f r
⎝ fr ⎠
))
(6.12)
⎛d⎞
( )
+ 44.9 − 6.55log hte log ⎜ ⎟
⎝ dr ⎠

Comparing (6.12) against (6.7), we have n = 4.49 − 0.655 loghte, m = 2.616

and, O = a(hre, f r ) − a(hre, f ).
In reality, the measured path loss could be different at different locations
even if the distance and frequency are the same. To represent this phenom-
enon, the path loss is rewritten in the following form:

⎛d⎞ ⎛ f ⎞
( )
PL(d, f ) = PL dr , f r + 10nlog ⎜ ⎟ + 10mlog ⎜ ⎟ + O (dB) + X s (6.13)
⎝ dr ⎠ ⎝ fr ⎠

where Xσ is a random variable having a normal distribution with zero mean and
standard deviation σ . Since Xσ has a normal distribution in the log domain,
this phenomenon is also called log-normal shadowing.
Through field measurement, the parameters n and m can be determined
through least square curve fitting to minimize the difference between the
theoretical prediction and the measured data. The parameters n and m are not
necessarily 2 as shown in (6.12). The standard deviation σ is also determined
through curve fitting. Equation (6.13) emphasizes the fact that the measured
path loss at a specified distance and frequency f is random following a log-
normal distribution.

6.3 Small-Scale Propagation Model

The small-scale propagation model deals with the amplitude, phase, and
frequency variations in a short period of time. Because of these variations,

6899_Book.indb 83 10/22/19 4:55 PM

 84 Introduction to OFDM Receiver Design and Simulation

the received power fluctuates with respect to time. When a signal propagates
through the media, it may be reflected from the surrounding structures such as
the ground, building, and mountain. These reflected waves may suffer different
propagation delays and arrive at the receiver at different times. The resulting
signal is the vector sum of all these component signals. Due to the amplitude
and phase difference of each arriving signal, signal distortions can occur.
However, the relative motion between the transmitter and the receiver
and the time variation of the channel can cause a Doppler shift in the received
carrier frequency. Therefore, this Doppler shift can vary with time. With
respect to the motion, the Doppler shift is at its highest when the receiver
moves away or toward the transmitter.
These two phenomena can be grouped into time dispersion and frequency
dispersion. The time dispersion accounts for a multipath channel, while the
frequency dispersion accounts for the Doppler shift from the relative motion
between the transmitter and the receiver. Both will be discussed in detail in
the following sections. The Clark model, which has Rayleigh distribution for
the received signal envelope, is also presented.

6.3.1 Time Dispersion

The time dispersion is due to the reception of the same signal from several
paths having different propagation delays. Assume that there are N paths and
the delay of the ith path is τ i and let ri(t, τ i) represent the signal received from
the ith path, the overall received signal r(t) is then given by
i= N −1
r(t) = ∑ ri ( t,ti ) (6.14)
i=0

Assume that the transmitted baseband signal is s(t). Since the propaga-
tion delay causes a time shift of s(t), (6.14) can also be written as [1, 2]
i= N −1
r(T ) = ∑ ai (t)e jq (t ) s ( t − ti )
i
(6.15)
i=0

where α i(t) and θ i(t) are amplitude and phase perturbations associated with
the ith path. More details will be given in Chapter 12. Because the phase
θ i(t) is random, the multipath signal may sometimes add constructively and
sometimes add destructively. This amplitude variation is called signal fading.
The instantaneous power of the received multipath signal is given by

6899_Book.indb 84 10/22/19 4:55 PM

 Radio-Wave Propagation Model85

P(T ) = r(t)r ∗ (t)

N −1
∑ ai (t) s ( t − ti )
2 2
= (6.16)
i=0

{ ( )s(t − t ) s
(t − t ) }
N −1 N −1 ∗
j qi (t )−q j (t )
+ 2∑ ∑ Re ai (t)a ∗j (t)e i j
i=0 j>i

Both the channel induced amplitude perturbation α i(t) and phase per-
turbation θ i(t) are random variables. If each path is independent, then the
amplitude and phase in one path are uncorrelated with those of the other
path. Performing the time average, the second term of (6.16) then becomes
0. Under this assumption, (6.16) becomes

⎛ N −1 2⎞
Pavg = E ⎜ ∑ ai (t) si t − ti
⎝ i=0
2
( ) ⎟⎠ (6.17)

where E represents the ensemble time average. If the transmitted signal is

unmodulated, there is no loss of generality to assume s(t) to be 1. Equation
(6.17) then simplifies to

⎛ N −1 2⎞
Pavg = E ⎜ ∑ ai (t) ⎟ (6.18)
⎝ i=0 ⎠

Equation (6.18) shows the time average power is just the summation of
the square of average amplitude from each path.
The received power P(t) is also called the intensity profile of the chan-
nel. The width of this intensity profile with an essentially nonzero value is the
delay spread, Δτ . The Fourier transform of P(t) generates the spaced frequency
correlation function ϕ (Δf ) defined below [2]:

f( Δf ) = ∫ P(t)e − j2pΔft dt (6.19)

The width of ϕ (Δf ) is called the coherence bandwidth, Bc. The delay
spread is proportional to the standard deviation of the multipath delay, σ τ. If
the delay is widely distributed to generate a large standard deviation, then the
delay spread Δτ is also large. However, the coherence bandwidth Bc is inversely
proportional to the delay spread Δτ or the standard deviation of delay σ τ. If the
delay spread increases, then the coherence bandwidth decreases and vice versa.

6899_Book.indb 85 10/22/19 4:55 PM

 86 Introduction to OFDM Receiver Design and Simulation

The coherence bandwidth can be thought of as the frequency separation

such that the channel can pass the signal with little distortion. If the coher-
ence bandwidth is wide, it is a good channel and the delay spread is small.
However, it is a bad channel for a narrow coherence bandwidth and the delay
spread is large.
Associated with time dispersion, there are flat fading and frequency
selective fading. This is related to the signal bandwidth and coherence band-
width. It is called flat fading if the signal bandwidth is much smaller than
the coherence bandwidth. Under this condition, the signal can pass through
the channel with little amplitude or phase distortion. However, the signal
bandwidth can be wider than the coherence bandwidth. When this happens,
different frequency components will have different channel degradations. If
the signal passes through this channel, it can suffer severe distortion. Figure
6.5 illustrates the situation. In this figure, Bs1 is the signal 1 bandwidth, which
is larger than the coherence bandwidth and the signal will pass through the
channel with distortion. In the second case, Bs2 is the signal 2 bandwidth,
which is smaller than the coherence bandwidth and the signal will pass through
the channel with little distortion.

6.3.2 Frequency Dispersion

The source and receiver may not be stationary when the signal is transmit-
ted. Therefore, the relative motion between the transmitter and the receiver

Figure 6.5 Coherence bandwidth in relation to signal bandwidth.

6899_Book.indb 86 10/22/19 4:55 PM

 Radio-Wave Propagation Model87

Figure 6.6 Illustration of Doppler shift.

can generate Doppler shift given as fd . Because of this Doppler frequency,

the received carrier frequency moves from fc to fd ± fd . The magnitude of fd
depends upon the motion velocity v and the angle between the direction of
motion and the signal propagation direction θ .
Figure 6.6 illustrates the effect of Doppler shift. In this figure, the mobile
moves from point X to point Y within a time period of Δt. The source T trans-
mits from T to the mobile. The angle between direction of motion from X to
Y and the wave propagation direction from T to Y or from T to X is θ . This
is because source is assumed to be far away and the lines from T to X and
from T to Y are, in essence, parallel. The distance from X to Y is R. The path
difference is then Rcosθ which generates a phase shift φ given by
2pRcosq
j= (6.20)
l

The phase shift φ corresponds to a frequency shift fd given by

j Rcosq vcosq
fd = = =
(2pΔt) (lΔt) l
(6.21)
vf cosq
= c
c

where v is the velocity, c is the speed of light, and λ is the wavelength.

Equation (6.21) indicates that the Doppler shift depends both upon the
velocity v and the angle θ . The Doppler shift fd is maximum when the angle θ
is either 0° or 180°. If the angle is 0°, the mobile moves toward the source and
the received frequency increases by a maximum amount. If the angle is 180°,
the mobile moves away from the source and the received frequency decreases
by a maximum amount. If the angle is 90°, there is no Doppler shift even if
motion exists.

6899_Book.indb 87 10/22/19 4:55 PM

 88 Introduction to OFDM Receiver Design and Simulation

Example 6.4
For a walkie-talkie handheld radio, the carrier frequency is around 462 MHz.
Assuming that angle θ = 0, Figure 6.7 is a plot of Doppler shift in hertz against
the car speed from 2 miles/hour to 65 miles/hour. Also shown in the figure are
two additional curves for carrier frequency 800 MHz and 1.2 GHz. Clearly,
the Doppler shift increases linearly with motion speed. It also increases with
carrier frequency. At 2 miles/hour or human walking speed, the Doppler shift
is around 1.5 Hz at 462 MHz and is negligibly small. For a car moving at 65
miles/hour and the same carrier frequency of 462 MHz, the Doppler shift
is around 49 Hz, which may need to be taken into account depending upon
the signal bandwidth.
The Doppler spread, f m, is the maximum frequency shift of the chan-
nel. In Example 6.4, the Doppler spread is 49 Hz at 462 MHz, 85 Hz at 800
MHz and 127 Hz at 1.2 GHz assuming the maximum car speed is 65 miles/
hour. If the Doppler spread is much smaller than the signal bandwidth, it can
then be neglected. The receiver should take Doppler spread into account if it
is not small in comparison to the signal bandwidth. The coherence time, Tc,
is a measure of the time variations of the channel and is proportional to the
inverse of the Doppler spread. An empirical formula exists which relates the
coherence time to the Doppler spread as given below [1]:
0.423
Tc = (6.22)
fm

Figure 6.7 Doppler shift against car speed.

6899_Book.indb 88 10/22/19 4:55 PM

 Radio-Wave Propagation Model89

The Doppler spread changes little for a large coherence time and changes
a great deal for a small coherence time. If the Doppler spread is small, the
channel can be considered as time invariant and the signal can pass through
with little distortion. However, the coherence time is small if the Doppler
spread is large. If the coherence time is smaller than the signal duration, the
signal will suffer significant distortion after passing through this channel.
For a time-varying channel, it can be classified as a slow channel or a
fast channel. For a fast channel, the coherence time is smaller than the signal
duration. Because the channel changes faster with time, it causes Doppler
spread that is significant in comparison to the signal bandwidth. This is also
called time-selective fading or frequency dispersion.
For a slow channel, the coherence time is longer than the signal dura-
tion. Because this channel is essentially flat, the Doppler spread approaches a
delta function in the frequency domain. In other words, the Doppler spread
is much smaller than the signal bandwidth. This time-invariant nature of the
channel is considered to be a good channel. The signal can then pass through
it with negligible distortion.

6.3.3 Clark’s Fading Model

When a signal passes through a fading channel, it may be scattered in many
directions. Clark developed a model based upon flat fading [3]. It is assumed
N scattered signals are received by a mobile. Each signal has random amplitude
An, phase ϕ n and Doppler shift fn. From (6.21), the Doppler shift is given here:

vf c cosan
fn = (6.23)
c

where α n is the angle between the nth scattered wave and the direction of the
moving mobile. Because of flat fading, all the scattered signals are assumed
to arrive at the same time with the same delay. The received signal strength is
then the sum of these N scattered signals as given here [1]:
N −1
r(t) = R0 ∑ An Re ⎡ e ( )
j2p f c + f n t+φn (t ) ⎤
(6.24)
n=0
⎣ ⎦

where R 0 is the average signal amplitude and is assumed to be a constant for

every scattered signal. Taking the real part of (6.24), the following is obtained

6899_Book.indb 89 10/22/19 4:55 PM

 90 Introduction to OFDM Receiver Design and Simulation

N −1
( (
r(t) = R0 ∑ An cos j2p f c + f n t + φn (t)
n=0
) )
(6.25)
( ) (
= U c (t)cos 2pf c t − U s (t)sin 2ppf c t )
where the Uc(t) and Us(t) are defined here:
N −1
(
U c (t) = R0 ∑ An cos 2pf n t + fn(t) ) (6.26)
n=0

N −1
(
U s (t) = R0 ∑ An sin 2pf n t + fn(t) ) (6.27)
n=0

From (6.26) and (6.27), the envelope of the received signal is then
given by

( )
0.5
E(t) = U c (t)2 + U s (t)2 (6.28)

The Doppler angle fn is assumed to be small in comparison with fc. There-

fore, Uc(t) and Us(t) are assumed to be narrowband random process having
Gaussian distribution with zero mean and equal variance of σ 2 = R 20/2 [1].
The envelope r = E(t) then has the following Rayleigh distribution [1]
r −r 2 /s 2
p(r) = e (6.29)
s2

The power spectral density S(f ) of the received signal was developed by
Gans and is given by the following formula [1, 4]:
1.5
S( f ) = (6.30)
pf m 1 − b2

where β = (f − fc)/f m and f m is the maximum Doppler shift. The maximum

frequency is f = fc + f m corresponding to β = 1 while the minimum frequency is
f = fc − f m corresponding to β = −1. Figure 6.8 is a plot of π f mS(f )/1.5 against
the normalized frequency β . The spectrum becomes infinite corresponding to
f = fc ± f m. This happens when the angle of arrival is either 180° or 0°. Between
these two extremes, the spectrum is almost flat.
In Clark’s model, the received signal envelope has a Rayleigh distribution.
This means the spectrum given by (6.30) dictates the time-domain channel

6899_Book.indb 90 10/22/19 4:55 PM

 Radio-Wave Propagation Model91

Figure 6.8 Spectrum of Clark’s model.

waveform due to Rayleigh fading. In other words, the channel with Rayleigh
fading must have a spectrum specified by (6.30).

6.4 Receiver Signal-to-Noise Ratio (SNR)

In the receiver, there are two types of major losses. The first is the propagation
loss from the source through the media. This propagation loss can be predicted
using either the theoretical or empirical model discussed earlier. The second is
the thermal noise at the antenna input. However, there are also amplifier noise
and cable loss, which can be referred to at the input. Based upon the received
power and the accumulated receiver noise, the SNR can then be predicted.

6.4.1 Thermal Noise

The thermal noise is generated due to random electron motion inside a resistor
or active devices. The thermal noise energy is given by [5]
N 0 = kT (6.31)

where T is the temperature in Kelvin and k is the Boltzmann’s constant equal

to 1.38 × 10 −23 joule/K. Therefore, the thermal noise is proportional to the
temperature, and the higher the temperature, the larger the thermal noise.

6899_Book.indb 91 10/22/19 4:55 PM

 92 Introduction to OFDM Receiver Design and Simulation

Equation (6.31) defines the noise energy N0 or the power density in

watts/hertz. If the receiver has a bandwidth of B, then the noise power in
watts is given by
N = kTB (6.32)

6.4.2 Noise Factor

The receiver noise due to amplifier or cable can be referred to its input to
generate an equivalent thermal temperature Te. The noise power density is
then given by [5]
N 0 = kTe (6.33)

An alternative way to express amplifier noise is to define a noise factor

F, which is related to N0 by the following relationship:
N 0 = k(F − 1)T0 (6.34)

where T0 is the room temperature. Equating (6.33) and (6.34) together, we have

Te = (F − 1)T0 (6.35)

where F is dimensionless. If the amplifier noise factor F is known, then its

equivalent noise temperature Te can be used to compute the noise power
density given by (6.33). From (6.35), the noise factor F can also be given as
Te
F = 1+ (6.36)
T0

The noise factor is normally defined in decibels and is given by [5]

FdB = 10Log F (6.37)

6.4.3 Amplifier Model

Figure 6.9 shows an amplifier noise model in the receiver front end. Assume
this amplifier has a power gain G and equivalent temperature Te. The power
gain G is defined to be the ratio of output power to the input power. Assum-
ing also that the noise power density referred to its input is N1 and the output
noise power density is N2, then we have

6899_Book.indb 92 10/22/19 4:55 PM

 Radio-Wave Propagation Model93

Figure 6.9 Amplifier noise model.

N1 = kTe (6.38)

N 2 = GN1 (6.39)

If we know the amplifier output noise power density N2, the input noise
power density N1 is then given by N1 = N2/G.
If the amplifier noise factor is F, then (6.38) becomes after using (6.35)
N1 = k(F − 1)T0 (6.40)

In other words, the amplifier noise power density referred to its input is
known as long as its noise factor is available.

6.4.4 Cable Loss Model

When a signal passes through the cable, it suffers some power loss. The loss
factor L is defined to be the ratio of input power to output power. Therefore,
N2 is less than N1. Figure 6.10 shows the cable loss model and Te is the equiva-
lent noise temperature. It can be shown [5] that the loss factor L is equal to
the noise factor F at room temperature as given here:
L= F (6.41)

Based upon (6.41), the following are true for the cable loss model
N1 = kTe = k(L − 1)T0
(6.42)
N1 = LN 2 (6.43)

Figure 6.10 Cable loss model.

6899_Book.indb 93 10/22/19 4:55 PM

 94 Introduction to OFDM Receiver Design and Simulation

6.4.5 Equivalent Noise Temperature at Receiver Front End

Figure 6.11 is a general block diagram of the receiver front end. It consists
of antenna, cable, amplifier, and receiver in cascade. Between the cable and
the receiver, there are n amplifiers. For amplifier k, the gain is Gak, the noise
equivalent temperature is Tak, and the noise factor is Fak. For the cable, the loss
factor is L and the noise equivalent temperature is Tc. Finally, for the receiver,
the gain is Gr, the noise equivalent temperature is Tr, and the noise factor is
Fr. Even though each device has its own noise equivalent temperature, what is
of interest is the overall noise equivalent temperature Te at the antenna input.
For the cable, the noise power density referred to its input is simply

Pc = kTc = k(L − 1)T0 (6.44)

For the first amplifier, the noise power density referred to its input is
given by

(
Pa1 = kTa1 = k Fa1 − 1 T0 ) (6.45)

At the antenna input, the noise power density referred from the first
amplifier is then

( )
Pa1,ant = Lk Fa1 − 1 T0 (6.46)

where we have used the relationship (6.43). For the second amplifier, the noise
power density referred to its input is

Figure 6.11 Block diagram of the receiver front end.

6899_Book.indb 94 10/22/19 4:55 PM

 Radio-Wave Propagation Model95

Pa2 = kTa2 = k Fa2 − 1 T0 ( ) (6.47)

From the second amplifier to the antenna, there is an amplifier with a

gain of Ga1 and a cable with a loss factor of L, and the noise power density
referred to the antenna input is then

Pa2,ant =
(
kL Fa2 − 1 T0 ) (6.48)
G1

For any amplifier i, the noise power density referred to the antenna input
can be derived in the same way and is given here:

Pai ,ant =
(
kL Fai − 1 T0 ) i≥2 (6.49)
∏ j=1G j
i−1

For the receiver, the noise power density referred to the antenna input
is derived similarly and is given by

Pr ,ant =
(
kL Fr − 1 T0 ) (6.50)
∏ j=1G j
n

Using all the equivalent noise temperatures referred to the antenna input
by combining (6.44) to (6.50), we have

kTe = kTant + k(L − 1)T0 + Lk Fa1 − 1 T0 ( )

(6.51)
n
Fai − 1 Lk ( Fr − 1)T0
+ LkT0 ∑ +
∏ j=1G j ∏ j=1G j
i−1 n
i=2

Removing the common term k from (6.51), we have

( )
Te = Tant + (L − 1)T0 + L Fa1 − 1 T0 + LT0 ∑
n
Fai − 1
+
( )
L Fr − 1 T0
(6.52)
∏ j=1G j ∏ j=1G j
i−1 n
i=2

Equation (6.52) is the equivalent noise temperature referred to the

antenna input.

6899_Book.indb 95 10/22/19 4:55 PM

 96 Introduction to OFDM Receiver Design and Simulation

Example 6.5
Assume that the receiver has one amplifier with a gain of 30 dB and an
equivalent noise temperature of 125K. After the amplifier is a receiver with a
noise figure of 15 dB. The room temperature is 293K and the antenna noise
temperature is 40K. There is no cable loss. Based upon these assumptions, n
= 1 and L = 1 and (6.52) becomes

) ( rG ) 0
F −1 T
(
Te = Tant + T0 Fa1 − 1 +
1

= Tant + Ta1 +
( Fr − 1)T0
G1

The amplifier gain G1 = 1030/10 = 1,000. The receiver noise figure Fr =

10 15/10
= 31.62. Substituting these numbers into the above equation, we have

(31.62 − 1)293
Te = 40 + 125 + = 40 + 125 + 8.97 = 173.97
1000

Example 6.6
Assume the same configuration as in Example 6.5 except there is a cable loss
of 4 dB. Since L is not 0 anymore, (6.52) becomes

Te = Tant + (L − 1)T0 + LTa1 +

( )
L Fr − 1 T0
G1

The cable loss factor L = 100.4 = 2.51. Substituting all the numbers into
the above equation, we have
2.51(31.62 − 1)293
Te = 40 + (2.51 − 1)293 + 2.51 × 125 +
1000
= 40 + 442.43 + 313.75 + 22.51 = 818.69

6.5 Range Determination

For a given transmitter power, the radio range can be determined once
the received power is measured. The empirical path loss formula in (6.7) is
given here:

6899_Book.indb 96 10/22/19 4:55 PM

 Radio-Wave Propagation Model97

⎛d⎞ ⎛ f ⎞
( )
PL(d, f ) = PL dr , f r + 10nlog ⎜ ⎟ + 10mlog ⎜ ⎟ + O
d
⎝ r⎠ ⎝ fr ⎠
(6.53)

Expressing PL in terms of Pt and Pr, we have

⎛d⎞ ⎛ f ⎞
( )
Pt (dB) − Pr (dB) = PL dr , f r + 10nlog ⎜ ⎟ + 10mlog ⎜ ⎟ + O
⎝ dr ⎠ ⎝ fr ⎠
(6.54)

Based upon (6.54), the range d can be determined once Pr is measured.

Another receiver parameter frequently used is the receiver signal strength
indicator (RSSI). It is the measurement of the minimum received power in
order for signal detection. If the received power is less than RSSI, then there is
no signal detection. It is normally expressed in dBm or power in milliwatt (mw).
If the receiver sensitivity is high, then RSSI is low. In other words, a
good receiver normally has a low RSSI so even a weak signal can be detected.

Example 6.7
Using the two-ray model and the parameter values given in Example 6.3, we
have n = 4, m = 0, PL(dr, f r) = 40log(dr), and O(f, d) = 0. Equation (6.54)
then becomes

40log(d) = Pt (dB) − Pr (dB)

Assuming that the transmitted power is 1W = 30 dBm, the range is

given in Table 6.1 for various received power. It is clear from Table 6.1 that
the range increases as the received power drops.

Table 6.1
The Range at Various Received Power

PT (dBm) Pr (dBm) Range (m)

30 −80 562.3
30 −85 749.9
30 −90 1,000
30 −95 1,333.5

6899_Book.indb 97 10/22/19 4:55 PM

 98 Introduction to OFDM Receiver Design and Simulation

6.6 SNR
Based upon (6.54), the received power is given by

⎛d⎞ ⎛ f ⎞
( )
Pr (dB) = Pt (dB) − PL dr , f r − 10nlog ⎜ ⎟ − 10mlog ⎜ ⎟ − O (dB)
⎝ dr ⎠ ⎝ fr ⎠
(6.55)

Assume that the noise equivalent temperature referred to the receiver

antenna is Te. Then the noise power is given by

N (dB) = 10log kTe B ( ) (6.56)

The SNR is obtained by subtracting (6.56) from (6.55) to have

⎛d⎞ ⎛ f ⎞
( )
SNR = Pt (dB) − PL dr , f r − 10nlog ⎜ ⎟ − 10mlog ⎜ ⎟
d
⎝ r⎠ ⎝ fr ⎠
(6.57)
(
− O (dB) − 10log kTe B )

Example 6.8
Using the two-ray model as in Example 6.7, we have the SNR given by

SNR = Pt (dB) − 40log(d) − 10log kTe B ( )

Using the same data from Example 6.7 and Example 6.6, we have Te =
818.69K and Pt = 30 dBm. Figure 6.12 is a plot of SNR against the bandwidth
in decibels for a propagation distance of 1,000m and 2,000m. The bandwidth
is from 30 dB to 60 dB or from 1,000 Hz to 1 MHz. Clearly, as the bandwidth
increases, there is more receiver noise in the passband and the SNR decreases.
However, as the distance increases, there is greater propagation loss and the
SNR decreases as well.

6.7 Summary
For wave propagation through a communication channel, the received sig-
nal may be distorted. There are two models to account for this effect. One is
the large-scale propagation model and the other is the small-scale propaga-
tion model.

6899_Book.indb 98 10/22/19 4:55 PM

 Radio-Wave Propagation Model99

Figure 6.12 The SNR against bandwidth in decibels and propagation distance.

The large-scale model accounts for the power loss after the radio wave
travels a distance through the media. Three different theoretical models were
presented. The first is the free-space propagation model. The second is the two-
ray model, which takes into account the radio wave reflected from the ground.
The third is an empirical model developed from field measurements. A general
formula that fits many theoretical propagation models was given. The frequency
and distance dependence can be determined from field measurements.
The small-scale model accounts for the amplitude, phase, and frequency
fluctuations as the radio wave propagates the media. There are two major
phenomena. One is the time dispersion and the other is the frequency disper-
sion. The time dispersion is due to the receiver receiving several signals with
different delays. The delay spread is related to the coherence bandwidth and
there are flat fading and frequency selective fading. For the flat fading, the
signal bandwidth is much smaller than the coherence bandwidth and the sig-
nal can pass through with little distortion. For the frequency-selective fading,
the signal bandwidth is much wider than the coherence bandwidth and the
received signal can suffer severe degradations.
The frequency dispersion is due to the time variation of the channel and
the relative motion between the transmitter and the receiver. This Doppler
effect is related to the coherence time and there are both fast fading and slow
fading. For slow fading, the coherence time is larger compared to the signal

6899_Book.indb 99 10/22/19 4:55 PM

 100 Introduction to OFDM Receiver Design and Simulation

duration and the signal can pass through with little distortion. It this case,
the channel can be considered as time invariant. For fast fading, the reverse
is true and the signal will suffer significant distortion after passing through
this channel.
Next, a noise model referred to the antenna input was discussed. A
general receiver front end has an antenna, cable, amplifiers, and receiver. A
general noise energy formula was derived based upon this front-end system.
An equivalent thermal temperature referred to at the antenna input was sub-
sequently derived. Based upon this equivalent temperature, the SNR was then
derived. From the received power measurement, the propagation range can
also be predicted.
In the next chapter, two classes of error correction codes are covered.
They are block code and convolutional code. The Viterbi algorithm used for
convolutional decoding is analyzed in detail. The punctured convolutional
code, which is derived from the 1/2 convolutional code, is also discussed. At
last, an interleaver used to reduce the block errors is also presented.

References
[1] Rappaport, T. S., Wireless Communications, New York: IEEE Press, 1996.
[2] Proakis, J. G., Digital Communications, New York: McGraw-Hill, 1983.
[3] Clarke, R. H., “A Statistical Theory of Mobile-Radio Reception,” Bell Systems Technical
Journal, Vol. 47, No. 6, 1968, pp. 957−1000.
[4] Gans, M. J., “A Power Spectral Theory of Propagation in the Mobile Radio Environment,”
IEEE Transactions on Vehicular Technology, Vol. 21, No. 1, 1972, pp. 27−38.
[5] Roddy, D., Satellite Communications, New York: McGraw-Hill, 1995.

6899_Book.indb 100 10/22/19 4:55 PM

 7
Error-Correcting Codes and Interleaver

7.1 Introduction
In digital communications, a sequence of input bits can be scrambled and
encoded before going through the modulation process. The scrambling is to
protect the data security and was discussed in Chapter 5. The encoding is to
allow the receiver to detect or correct the bit errors due to the transmission
through a noisy channel.
There are many different types of error correction schemes. Two types
are discussed in this chapter. The first type is the block codes. A block of k
information bits is transformed to n bits and n > k. The coding rate is k/n. The
additional n − k parity bits increase the transmission overheads. However, the
added advantage is to enhance the reception accuracies. Both linear block codes
and cyclic codes are given. The linear block code is operated through matrix
manipulations while the cyclic code is operated using algebraic polynomials.
From the generator polynomial of the cyclic codes, the generator matrix of
the linear block codes can be derived.
The second type is the convolutional codes. Again, k input bits are con-
verted into n coded bits. However, the input bit sequence is infinite in nature
and the output bits are obtained through shift register operations. The coding
rate is still defined to be k/n. The Viterbi algorithm is used for decoding. For the
convolutional code with a coding rate R = 1/2, the transmission redundancy

101

6899_Book.indb 101 10/22/19 4:55 PM

 102 Introduction to OFDM Receiver Design and Simulation

is 50%. To reduce the overhead bits, the punctured codes are discussed to
reduce the overhead transmission. Based upon the 1/2 code, both R = 2/3
and R = 3/4 punctured codes can be derived. Even though the transmission
data rate is increased, the price paid is a slight drop of reception accuracy. If
the drop is not significant, it is still a good approach.

7.2 Linear Block Codes

There are only two elements, 0 and 1, in a binary field. In the vector space of
n binary bits, there are a total of 2n such vectors. In its subspace, there exist
k independent vectors, g i , i = 0, k − 1. Each g i has n bits. Any vector in this
k-dimensional subspace can be created by linearly combining these k inde-
pendent vectors. Supposing that c is any vector in this subspace, then we have
k−1
c = ∑ ai g i (7.1)
i=0

where ai = 0 or 1. Both c and g i are vectors of n bits. Since the number of

binary coefficients ai is k, there are 2k such vectors. Equation (7.1) can be
thought of as transforming k information bits ai into a code word of n bits.
These 2k code words are referred to as the (n, k) linear block code. The extra
n − k bits are called the parity check bits.

7.2.1 Generator Matrix

In the matrix form, (7.1) can be rewritten as

⎡ g0 ⎤
c = a G = ⎡⎣a0 ,a1 ,…,ak−1⎤⎦ ⎢ ! ⎥ (7.2)
⎢ g ⎥
⎣ k-1 ⎦

where G is called the generator matrix. Using (7.2), the k information bits are
multiplied by the generator matrix G to create the code word c. Any k linear
independent row vectors g i , i = 0, k − 1 can be used to create the generator
matrix G. However, it is more convenient to put the matrix G in the follow-
ing standard form [1]:

G = ⎡⎣ P I ⎤⎦ (7.3)

6899_Book.indb 102 10/22/19 4:55 PM

 Error-Correcting Codes and Interleaver103

where P is a k × n − k matrix while I is a k × k identity matrix. Specifically,

P and I are in the following form:

⎡ p0,0 ! p0,n−k−1 ⎤
P=⎢ " # " ⎥ (7.4)
⎢ p ! pk−1,n−k−1 ⎥
⎣ k−1,0
⎦

⎡1 ! 0⎤
I=⎢ " # " ⎥ (7.5)
⎢⎣ 0 ! 1 ⎦⎥

The generator matrix G has a dimension of k × n and the code word

c has a dimension of 1 × n. The advantage of the form (7.3) is that the code
word c has the following form:

c = [n − k parity check bits, k information bits] (7.6)

The above equations can be understood more easily by considering the

following example.

Example 7.1
A (15,11) linear code has the following generator matrix

⎡ 1 1 0 0 1 0 0 0 0 0 0 0 0 0 0 ⎤
⎢ 0 1 1 0 0 1 0 0 0 0 0 0 0 0 0 ⎥
⎢ 0 0 1 1 0 0 1 0 0 0 0 0 0 0 0 ⎥
⎢ 1 1 0 1 0 0 0 1 0 0 0 0 0 0 0 ⎥
⎢ 1 0 1 0 0 0 0 0 1 0 0 0 0 0 0 ⎥
G=⎢ 0 1 0 1 0 0 0 0 0 1 0 0 0 0 0 ⎥ (7.7)
⎢ 1 1 1 0 0 0 0 0 0 0 1 0 0 0 0 ⎥
⎢ 0 1 1 1 0 0 0 0 0 0 0 1 0 0 0 ⎥
⎢ 1 1 1 1 0 0 0 0 0 0 0 0 1 0 0 ⎥
1 0 1 1 0 0 0 0 0 0 0 0 0 1 0
⎢ 1 0 0 1 0 0 0 0 0 0 0 0 0 0 1 ⎥
⎣ ⎦

In the matrix G, the first n − k = 4 columns are the 11 × 4 matrix P

while the next k = 11 columns are the 11 × 11 identity matrix I. The ith row
is the vector g i . For example, row 3 and row 10 are g 3 = [1 1 0 1 0 0 0 1 0 0
0 0 0 0 0] and g 10 = [1 0 0 1 0 0 0 0 0 0 0 0 0 0 1]. Let a = [0 1 0 1 0 1 0 1
0 0 1] be the message to be encoded. Using (7.2) to perform the matrix mul-
tiplication, the code word c becomes

6899_Book.indb 103 10/22/19 4:55 PM

 104 Introduction to OFDM Receiver Design and Simulation

c = ⎡⎣ 0 1 0 1 0 1 0 1 0 0 1 ⎤⎦G
= ⎣⎡ 0 0 0 0 0 1 0 1 0 1 0 1 0 0 1 ⎤⎦

In the code word c, the last 11 bits are exactly the message [0 1 0 1 0
1 01 0 01].

7.2.2 Parity Check Matrix

Corresponding to the k × n generator matrix G, there exists another n − k × n
matrix H such that G and H are orthogonal. This matrix H is called the par-
ity check matrix. The matrix H can be created from the G matrix and put in
the following standard form [1]:

H = ⎡⎣ I n−k P t ⎤⎦ (7.8)

where I is the n − k × n − k identity matrix while P T is a n − k × k matrix

created by the transpose of the matrix P. To show G and H are orthogonal
to each other, we have

GH t = ⎡⎣ P I k ⎤⎦ ⎡⎢ n−k ⎤⎥
I
⎣ P ⎦ (7.9)
= P+P=0

Combining (7.2) and (7.9), we have

c H t = aG H t = a 0 = 0 (7.10)

Therefore, any code word c is orthogonal to the parity check matrix H.

Since H is an n − k × n matrix, it is the generator matrix for the (n, n − k)
linear block code. Any code word d generated by H can be written as
d =bH (7.11)

where d is a 1 × n matrix, b is a 1 × n − k matrix and H is an n − k × n matrix.

The row matrix b has n − k message bits. Any code word c generated by G
and code word d generated by H are orthogonal to each other. This is easily
shown by combining (7.2), (7.9), and (7.11) as follows:

cd t = a G H t bt = 0 (7.12)

6899_Book.indb 104 10/22/19 4:55 PM

 Error-Correcting Codes and Interleaver105

Example 7.2
The parity check matrix H corresponding to (7.7) is the following:

⎡ 1 0 0 0 1 0 0 1 1 0 1 0 1 1 1 ⎤
H =⎢ 0 1 0 0 1 1 0 1 0 1 1 1 1 0 0 ⎥ (7.13)
⎢ 0 0 1 0 0 1 1 0 1 0 1 1 1 1 0 ⎥
⎣ 0 0 0 1 0 0 1 1 0 1 0 1 1 1 1 ⎦

The first n − k = 4 columns are the identity matrix. The next 11 columns
are the transpose of matrix P in G. By performing the matrix multiplication,
it can be shown that GH t = 0. Let b = [1 0 0 1] be the message to be coded
by H, and then d = bH = [1 0 0 1 1 0 1 0 1 1 1 1 0 0 0]. From Example
7.1, c = [0 0 0 0 0 1 0 1 0 1 0 1 0 0 1]. Performing the matrix multiplication
again, we have c d t = 0.

7.2.3 Syndrome
Suppose c is the transmitted code word and r is the received code word. If
there are no transmission errors, r = c. Since c is a code word, it must satisfy
c H t = 0. If there are no transmission errors, then s = r H t = c H t = 0. The
product s = r H t is called the syndrome of r and is a 1 × n − k row matrix.
If the code word c is transmitted through a noisy channel, there can
be transmission errors and the received code word r is different from the
transmitted code word c. The difference between r and c is the error vector
e and we can write
r =c+e (7.14)

Multiplying both sides of (7.14) by H t , the syndrome s then becomes [1]

s = rHt = cHt + eHt = eHt (7.15)

Equation (7.15) shows that the syndrome s of the received vector r is

equal to the syndrome of the error vector e.
If the syndrome s ≠ 0, then it is clear there are transmission errors.
However, if s = 0, there is no guarantee that there are no transmission errors.
This is because e could be equal to another code word and the linear combi-
nation of two code words is also a code word. When this happens, the error
is undetectable.

6899_Book.indb 105 10/22/19 4:55 PM

 106 Introduction to OFDM Receiver Design and Simulation

7.2.4 Error Correction

Equation (7.15) shows there is a one-to-one correspondence between s and
e. Once the received syndrome s of r is computed, the error vector e can be
identified. This is because s is equal to the syndrome of the error vector e. A
table can be created by associating each correctable error vector with its syn-
drome. Through table look-up, an error vector e corresponding to the received
syndrome is located. Once the error vector e is known, the transmitted code
word c is then given by
c =r+e (7.16)

Let C represent a linear block code. The weight of any code word c in
C is the number of nonzero bits in c. For example, the weight of code word
c = [0 0 0 0 0 1 0 1 0 1 0 1 0 0 1] is 5. The distance d(c, v) between any two
code words c and v of C is the number of places that they differ or the weight
of c + v. Since c + v is also a code word, the minimum weight of a block code
C is then given here:
dmin = min(weight of c , where c is any codeword of C ) (7.17)

Assume that t errors can be corrected in a linear block code C . Then it

can be shown that it is given by [1]

t=
( dmin − 1) (7.18)
2

where t is the largest integer no greater than (dmin − 1)/2.

Example 7.3
The generator matric G in Example 7.1 has k independent rows and each row
of n elements is a possible code word of C . From these k code words, it is seen
that the minimum weight is 3. Therefore, the (15,11) code can correct 1 error.
Example 7.4
From Example 7.3, the (15,11) code can correct 1 error. Since each code word
has 15 bits and each bit location is a possible error, there are 15 correctable
error patterns. These 15 correctable error patterns and their corresponding
syndromes are given in Table 7.1. From (7.15), the syndrome s of each error
pattern e is computed from s = e H t . Since there is only one nonzero element
in each error pattern e, the syndrome s is actually equal to the column vectors
of the parity check matrix H given in (7.13).

6899_Book.indb 106 10/22/19 4:55 PM

 Error-Correcting Codes and Interleaver107

Table 7.1
Syndrome and Correctable Error Patterns of a (15,11) Code

Syndrome Correctable Error Patterns

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

Example 7.5
From Example 7.1, assume that c = [0 0 0 0 0 1 0 1 0 1 0 1 0 0 1] is the trans-
mitted code word. Assume that there is a single bit error in the third bit of c
and the received code word r = [0 0 1 0 0 1 0 1 0 1 0 1 0 0 1]. The syndrome
s of r is computed here:

s = r H t = ⎣⎡ 0 0 1 0 ⎤⎦

From Table 7.1, the error pattern is found to be e 2 = [0 0 1 0 0 0 0 0 0 0

0 0 0 0 0]. Adding e 2 to r, the transmitted code word becomes [0 0 0 0 0 1 0
1 0 1 0 1 0 0 1], which is exactly equal to c and the single error is corrected.

7.2.5 Hamming Codes

Hamming codes [1] are special classes of an (n, k) linear block code. For any
integer m ≥ 3, n = 2m − 1 and k = 2m − 1 − m. The number of parity check

6899_Book.indb 107 10/22/19 4:55 PM

 108 Introduction to OFDM Receiver Design and Simulation

bits is n − k = m. The linear (15,11) code given in Example 7.1 is a Hamming

code with m = 4, n = 24 − 1 = 15 and k = 15 − 4 = 11.
The parity check matrix H of the (15,11) Hamming code is shown in
(7.13). The columns of H have all the 4-bit patterns. The sum of any two
columns does not add to zero and the error correcting capability is one [1].
Actually, all the Hamming codes can correct 1 error.

7.3 Cyclic Codes

A cyclic code is also a special class of an (n, k) linear block code. If c = (c0c1 …
cn–1) is a code word, then a cyclic shift to the right generates another code
word c 1 = (cn–1c0 … cn–2).
A cyclic code word is represented as a code polynomial c(x) given here:

c(x) = c0 + c1 x + c2 x 2 + … + cn−2 x n−2 + cn−1 x n−1 (7.19)

Every element of the code word c is a coefficient of a polynomial c(x).

The highest degree of a code polynomial is n − 1. Depending upon the poly-
nomial coefficients, the degree could be less than n − 1. A cyclic shift to the
right generates the following code polynomial:

c (1) (x) = cn−1 + c0 x + c1 x 2 + … + cn−3 x n−2 + cn−2 x n−1 (7. 20)

7.3.1 Generator Polynomial

Assume g(x) is a code polynomial of minimum degree in a cyclic code. Then
g(x) is a generator polynomial and is unique [1]. An ith shift to the right is
equivalent to multiplying g(x) by xi and xig(x) is also a code polynomial.

Example 7.6
Let c = (1 1 0 1 0 0 0) be a code word of a (7, 4) linear code. It is also of
minimum degree and g(x) = 1 + x + x3. A cyclic shift two places to the right
of c generates c = (0 0 1 1 0 1 0) and g(2)(x) = x 2 + x3 + x5. Multiplying g(x)
by x 2 gives g(2)(x) = x 2g(x) and g(2)(x) is also a code polynomial.
The message code vector a = (a 0a1 −− ak–2ak–1) of an (n, k) linear code
has k elements. The highest degree of a message polynomial is then k − 1. The
message polynomial a(x) is then given here:

a(x) = a0 + a1 x + … + ak−2 x k−2 + ak−1 x k−1 (7.21)

6899_Book.indb 108 10/22/19 4:55 PM

 Error-Correcting Codes and Interleaver109

Assume that g(x), xg(x), …., and xk−1g(x) are all code polynomials such
that the degree of xk−1g(x) is less than or equal to n − 1. Since g(x) is a code
polynomial of a minimum degree [1], all the code polynomials xig(x), 0 < i ≤
k − 1 have a degree larger than that of g(x). A linear combination of all these
code polynomials must also be a code polynomial c(x), we then have

( )
c(x) = a0 + a1 x + … + ak−2 x k−2 + ak−1 x k−1 g(x)
(7.22)
= a(x)g(x)

Assuming that the highest degree of g(x) is r, then r + k − 1 = n − 1 and

we obtain r = n − k. The g(x) is called the generator polynomial and is given by

g(x) = 1 + g1 x + g 2 x 2 + … + g n−k−1 x n−k−1 + x n−k (7.23)

The coefficient of xn−k must be 1; otherwise, another code polynomial of

minimum degree exists. From (7.22), every code polynomial is the product of
message polynomial a(x) by the generator polynomial g(x).
Multiplying g(x) by x, x 2 − xk−1 generates code polynomials of maximum
degree n − 1. They are given here:

xg(x) = x + g1 x 2 + g 2 x 3 + … + g n−k−1 x n−k + x n−k+1 = g (1) (x)

x 2 g(x) = x 2 + g1 x 3 + g 2 x 4 + … + g n−k−1 x n−k+1 + x n−k+2 = g (2) (x)
.........................................................
x k−1 g(x) = x k−1 + g1 x k + g 2 x k+1 + … + g n−k−1 x n−2 + x n−1 = g (k−1) (x)

However, multiplying g(x) by xk gives the following

x k g(x) = x k + g1 x k+1 + g 2 x k+2 + … + g n−k−1 x n−1 + x n

= 1 + x k + g1 x k+1 + g 2 x k+2 + … + g n−k−1 x n−1 + ( x n + 1) (7.24)
= ( x n + 1) + g (k) (x)

Since g(k)(x) is a code word, g(k)(x) can be written as g(k)(x) = b(x)g(x) and
(7.24) becomes [1]

x n + 1 = ( x k + b(x)) g(x) (7.25)

Equation (7.25) shows that the generator polynomial g(x) must be a fac-
tor of xn + 1 and the degree is n − k.

6899_Book.indb 109 10/22/19 4:55 PM

 110 Introduction to OFDM Receiver Design and Simulation

Example 7.7
The generator polynomial of a (15,11) cyclic code can be found from the fac-
tors of x15 + 1.

x15 + 1 = ( x 4 + x + 1)( x11 + x 8 + x 7 + x 5 + x 3 + x 2 + x + 1)

The first term has a degree n − k = 4 and must be a generator polynomial

of the (15,11) cyclic code. Therefore, g(x) = x4 + x + 1.
Let a(x) be the message polynomial. Dividing xn−ka(x) by g(x) gives
the following:

x n−k a(x) = d(x)g(x) + b(x) (7.26)

The remainder b(x) has a degree less than n − k. The term d(x)g(x) is a
code polynomial and we have [1]

c(x) = b(x) + x n−k a(x) (7.27)

However, the term xn−ka(x) has a degree greater than or equal to n − k.

Therefore, the information bits are shifted to the highest k location after encod-
ing a(x). Equation (7.27) provides a way to encode the message polynomial
such that the information bits come after the parity bits.

Example 7.8
In Example 7.1, the message vector is a = [0 1 0 1 0 1 0 1 0 0 1]. The
equivalent message polynomial is a(x) = x + x3 + x5 + x 7 + x10. The generator
polynomial from Example 7.7 is g(x) = 1 + x + x4. Dividing x4a(x) by g(x),
we have

x 4 a(x) = x 5 + x 7 + x 9 + x11 + x14

= ( x 5 + x 6 + x10 ) g(x) + 0

The remainder happens to be zero. Therefore, the code polynomial is

c(x) = x5 + x 7 + x 9 + x11 + x14. The equivalent code word is c = [0 0 0 0
0 1 0 1 0 1 0 1 0 0 1], which is the same as that given in Example 7.1.

Example 7.9
From Example 7.7, the generator polynomial is g(x) = x4 + x + 1. From (7.26),
the division of x4, x5, x 6, x 7, x 8, x 9, x10, x11, x12, x13, and x14 by g(x) gives
the following:

6899_Book.indb 110 10/22/19 4:55 PM

 Error-Correcting Codes and Interleaver111

x 4 = g(x) + 1 + x
x 5 = xg(x) + x + x 2
x 6 = x 2 g(x) + x 2 + x 3
x 7 = ( x 3 + 1) g(x) + 1 + x + x 3
x 8 = ( x 4 + x + 1) g(x) + 1 + x 2
x 9 = ( x 5 + x 2 + x ) g(x) + x + x 3
x10 = ( x 6 + x 3 + x 2 + 1) g(x) + 1 + x + x 2
x11 = ( x 7 + x 4 + x 3 + x ) g(x) + x + x 2 + x 3
x12 = ( x 8 + x 5 + x 4 + x 2 + 1) g(x) + 1 + x + x 2 + x 3
x13 = ( x 9 + x 6 + x 5 + x 3 + x + 1) g(x) + 1 + x 2 + x 3
x14 = ( x10 + x 7 + x 6 + x 4 + x 2 + x + 1) g(x) + 1 + x 3

Moving d(x)g(x) to the left, we have

g(x) = 1 + x + x 4
xg(x) = x + x 2 + x 5
x 2 g(x) = x 2 + x 3 + x 6
( x 3 + 1) g(x) = 1 + x + x 3 + x 7
( x 4 + x + 1) g(x) = 1 + x 2 + x 8
( x5 + x 2 + x ) g(x) = x + x 3 + x 9
( x 6 + x 3 + x 2 + 1) g(x) = 1 + x + x 2 + x10
( x 7 + x 4 + x 3 + x ) g(x) = x + x 2 + x 3 + x11
( x 8 + x5 + x 4 + x 2 + 1) g(x) = 1 + x + x 2 + x 3 + x12
( x 9 + x 6 + x5 + x 3 + x + 1) g(x) = 1 + x 2 + x 3 + x13
( x10 + x 7 + x 6 + x 4 + x 2 + x + 1) g(x) = 1 + x 3 + x14
The left sides are all code polynomials. The right sides can be written in
matrix forms as follows:

6899_Book.indb 111 10/22/19 4:55 PM

 112 Introduction to OFDM Receiver Design and Simulation

⎡ 1 1 0 0 1 0 0 0 0 0 0 0 0 0 0 ⎤
⎢ 0 1 1 0 0 1 0 0 0 0 0 0 0 0 0 ⎥
⎢ 0 0 1 1 0 0 1 0 0 0 0 0 0 0 0 ⎥
⎢ 1 1 0 1 0 0 0 1 0 0 0 0 0 0 0 ⎥
⎢ 1 0 1 0 0 0 0 0 1 0 0 0 0 0 0 ⎥
G=⎢ 0 1 0 1 0 0 0 0 0 1 0 0 0 0 0 ⎥
⎢ 1 1 1 0 0 0 0 0 0 0 1 0 0 0 0 ⎥
⎢ 0 1 1 1 0 0 0 0 0 0 0 1 0 0 0 ⎥
1 1 1 1 0 0 0 0 0 0 0 0 1 0 0
⎢ 1 0 1 1 0 0 0 0 0 0 0 0 0 1 0 ⎥
⎢ 1 0 0 1 0 0 0 0 0 0 0 0 0 0 1 ⎥
⎣ ⎦

The matrix G is the same as that given in (7.7). This also points out a
way to generate the generator matrix.

7.3.2 Syndrome Polynomial

If the received polynomial r(x) is the same as the transmitted polynomial, then

r(x) = c(x) = a(x)g(x) (7.28)

In other words, r(x) must be divisible by g(x) and the remainder is zero.
In a noisy channel, the received polynomial r(x) may not be the same as the
transmitted polynomial and the remainder s(x) may not be zero. Under this
condition, r(x) can be written as

r(x) = b(x)g(x) + s(x) (7.29)

The polynomial s(x) is called the syndrome of r(x). The syndrome s(x)
is the remainder after dividing r(x) by g(x). If the syndrome s(x) is not zero,
r(x) is not a code polynomial and there are transmission errors. Therefore, the
syndrome s(x) can be used to detect whether there are any received bit errors.
Assuming that the difference between the received polynomial r(x) and
the transmitted polynomial c(x) is e(x), we can write

r(x) = c(x) + e(x) (7.30)

Substituting (7.22) and (7.29) into (7.30), we have

e(x) = ( a(x) + b(x)) g(x) + s(x) (7.31)

Therefore, the syndrome s(x) is also the remainder of dividing e(x) by g(x).

6899_Book.indb 112 10/22/19 4:55 PM

 Error-Correcting Codes and Interleaver113

Based upon (7.31), a table can be created to correlate the error patterns e(x)
with the corresponding syndromes s(x). After receiving r(x), it is first divided
by g(x) to get the syndrome s(x). From the syndrome s(x), the corresponding
error pattern e(x) is found from the table. The error pattern e(x) is then added
to the received r(x) to recover the transmitted code polynomial c(x) or
c(x) = r(x) + e(x) (7.32)

The e(x) could be equal to the code polynomial and r(x) is also divis-
ible by g(x). When this happens, the syndrome s(x) is 0 and the error bits are
not correctable. The error correction capability is the same as the linear code.
Assuming that t errors can be corrected, then t is given by

t=
( dmin − 1) (7.33)
2

where dmin is the weight of the cyclic code.

Example 7.10
A (15,11) cyclic code with a generator polynomial of g(x) = 1 + x + x4 has a
minimum weight of 3. Therefore, it can correct one error. There are a total
of 15 error patterns with one bit error. These error patterns together with the
corresponding syndromes are listed in Table 7.2.
The syndrome on the second column is the remainder of dividing e(x)
in the first column by g(x). Comparing Tables 7.1 and 7.2, it can be seen that
they are really identical.
Example 7.11
From Example 7.8, the transmitted polynomial is c(x) = x5 + x 7 + x 9 + x11 +
x14. Assume that the received polynomial is r(x) = x 2 + x5 + x 7 + x 9 + x11 + x14
and an error has occurred in the second bit. The syndrome s(x) of r(x) is the
remainder of dividing r(x) by g(x) = 1 + x + x4. After the division, we have
s(x) = x 2. From Table 7.2, the error pattern is e(x) = x 2. Adding e(x) to r(x), we
have r(x) + e(x) = x 2 + x5 + x 7 + x 9 + x11 + x14 + x 2 = x5 + x 7 + x 9 + x11 + x14.
Therefore, the original c(x) is recovered.

7.4 Convolutional Code

Similar to the (n, k) linear block code, the convolutional code is defined to
be a rate R = k/n. For every k-bit input message, there is an n-bit output.

6899_Book.indb 113 10/22/19 4:55 PM

 114 Introduction to OFDM Receiver Design and Simulation

Table 7.2
Error Pattern and Syndrome of a (15,11) Cyclic Code

Error Pattern Syndrome

1 1
x x
x 2
x2
x3 x3
x 4
1+x
x 5
x + x2
x 6
x2 + x3
x 7
1 + x + x3
x8 1 + x2
x 9
x + x3
x10 1 + x + x2
x 11
x + x2 + x3
x12 1 + x + x2 + x3
x13 1 + x2 + x3
x 14
1 + x3

The information bits pass through a shift register to generate the output. In
general, the shift register has a total of Lk bits. The number L is called the
constraint length [2].
Each output bit is generated from a binary adder with input connections
to the shift register. If there are n output bits, there are n such binary adders.
Similar to the sequence generator described in Chapter 5, the connections to
generate each output bit can be specified by a generator polynomial.

7.4.1 Convolutional Encoder

Consider the convolutional encoder given in Figure 7.1. Each input bit (k
= 1) generates 2 (n = 2) output bits. This is then called a 1/2 convolutional
encoder. The shift register has a total of 2 bits. Since k = 1, the constraint
length is equal to 2. As can be seen in Figure 7.1, each output bit is associated
with a binary adder. Each adder can be described by a generator polynomial.
Let g 1(x) and g 2(x) represent the two polynomials associated with output bits
1 and 2; we then have

6899_Book.indb 114 10/22/19 4:55 PM

 Error-Correcting Codes and Interleaver115

Figure 7.1 A 1/2 convolutional encoder.

g1(x) = 1 + x 2 (7.34)

g 2 (x) = 1 + x + x 2 (7.35)

The convolutional encoder can be described by a state-transition dia-

gram. Figure 7.2 shows the state-transition diagram of a 1/2 convolutional
encoder. The content of the shift register represents the state. Since the shift
register has 2 bits, there are four possible states ranging from 0 to 3. All the
numbers in Figure 7.2 are in decimal format. Therefore, the equivalent binary
representations of 0, 1, 2, and 3 are 00, 01, 10, and 11. The number inside each
circle represents the state. The number beside each branch is the output. The

Figure 7.2 State transition diagram of a 1/2 convolutional encoder.

6899_Book.indb 115 10/22/19 4:55 PM

 116 Introduction to OFDM Receiver Design and Simulation

four possible 2-bit outputs are 00, 01, 11, and 10, which can be represented
as bauds of 0, 1, 3, and 2. Assume the initial state is 0. If the input bit is 0,
the next state remains at zero and the output baud is zero. If the input bit is
1, the state transitions from 0 to 2 and the output baud is 3. The dashed line
represents the transition due to a bit 0 input while the solid line represents the
transition due to a bit 1 input. Figure 7.2 also shows that there are two pos-
sible ways to reach any state. For example, state 1 can be reached from state
2 due to a bit 0 input and from state 3 due to a bit 0 input.
The convolutional encoder can also be represented by a trellis diagram.
Figure 7.3 shows the trellis diagram of a 1/2 convolutional encoder. The left-
most column shows the four possible states of 0, 1, 2, and 3. The top two rows
show the input bit sequence of either bit 0 or 1 and the corresponding output
sequence of 0, 1, 2, or 3. On each branch, there is a symbol of the form a/b.
The letter “a” represents the output while the letter “b” represents the single
bit input. For example, the symbol 3/1 means a bit 1 input generating an
output baud 3. The trellis diagram starts at state 0 with a depth of d = 0 and
proceeds to the right. The depth d increments by 1 for each new bit input.
At d = 0, there are two possible transitions. A bit 0 input transits from state
0 to state 0 and outputs 0. A bit one input transits from state 0 to state 2 and
outputs 3. At depth 1, there are then two possible states. Since each state has
two possible transitions, there are four possible states at depth d = 2. After

Figure 7.3 Trellis diagram of a 1/2 convolutional encoder.

6899_Book.indb 116 10/22/19 4:55 PM

 Error-Correcting Codes and Interleaver117

depth 2, two major events are observed. The first is that the number of states
remains at 4. The second is that there are two possible ways to reach any state.
For example, at d = 3, state 1 can be reached from state 2 due to a bit 0 input
and from state 3 also due to a bit 0 input. Following the transition path, an
input sequence of 01100 generates an output baud sequence of 03223. The
path traced through due to the input sequence is marked by dark lines. After
d = 3, there are only two paths reaching each state, as will be discussed next.

7.4.2 Convolutional Decoder and Viterbi Algorithm

Figure 7.3 shows the possible number of paths through the trellis increases
with the depth. There are two paths at d = 1, four paths at d = 2, and eight
paths at d = 3, and, in general, there are 2k paths at d = k. Assume the received
semi-infinite symbol sequence is Y = y 0y1 … yj …. Every possible path through
the trellis diagram generates a possible output code sequence. Assume that
the output code sequence of the nth path is Xn = xn0xn1 … xnj …. Based upon
the received symbol sequence, the convolutional decoder finds a path with
the highest probability of being followed through the trellis. The metric used
is then to maximize ln(p(Y/Xk)). The probability p(Y/Xk) can be written as

⎛ Y ⎞ j=∞ ⎛ y j ⎞
p⎜ = ∏ p⎜ ⎟ (7.36)
⎝ X k ⎟⎠ j=0 ⎝ x kj ⎠

The metric then becomes

⎛ ⎛ Y ⎞⎞ j=∞ ⎛ ⎛ yj ⎞⎞
ln ⎜ p ⎜ ⎟=
⎝ ⎝ X k ⎟⎠ ⎠
∑ ⎜ p⎜ x ⎟ ⎟
ln (7.37)
j=0 ⎝ ⎝ kj ⎠ ⎠

Out of all the possible paths through the trellis, the code sequence Xk
such that ln (p(Y/Xk)) is the maximum is selected to be the code sequence
input to the channel [3].

Example 7.12
From Figure 7.3, assume there are no channel errors and the received baud
sequence is Y = 03223. The code sequence through the trellis that most likely
matches with Y is the path k such that Xk = 03223. This path is marked with
the dark line in Figure 7.3. Based upon Xk, the input message sequence is then
found to be 01100. No other path can have a code sequence that matches so
perfectly. For example, the code sequence corresponding to the binary input

6899_Book.indb 117 10/22/19 4:55 PM

 118 Introduction to OFDM Receiver Design and Simulation

message sequence 01110 is 03212, which does not match with Y as closely as
Xk. In other words, ln (p(Y/Xk)) is the maximum.
In practice, it is not so easy to accurately determine p(Y/Xk). For hard
decision decoding, the metric used is to minimize the Hamming distance
between Y and code sequence Xk. Assuming that X and Z are two binary
sequences, the Hamming distance is then obtained by performing the exclu-
sive operation of finding the number of places that they differ. For example,
the Hamming distance between X = 11011 and Z = 01111 is 2. It can also be
computed as the weight of X + Z. For example, the weight of X + Z = 10100
is also 2. Table 7.3 lists the Hamming distance between 2-bit bauds and the
number inside the parentheses is the baud representation.
The received convolutional codes can be decoded using the Viterbi algo-
rithm [3], which is illustrated in Figure 7.4 using the encoder given in Figure
7.1. Associated with each branch, there is a symbol in the form a/b/c. The
additional letter c gives the accumulated Hamming distance at the ending
state. The receiver has a replica of the encoder and knows all the possible
transmitted codes. Before all four states are generated, all the possible paths
are kept. Therefore, four paths are kept before d = 2. On each branch, the
Hamming distance between branch symbol a and the received symbol is com-
puted first. For example, the output state 1 at d = 3 is reached from the input
state 2 at d = 2. The branch symbol is the encoder output baud a = 1 after a
bit 0 input and the accumulated Hamming distance c needs an update. The
Hamming distance between a = 1 and the received baud 2 is 2. The input
state 2 is reached from state 0 with an accumulated Hamming distance 0.
The cost at d = 2 and state 2 is then zero. The new accumulated Hamming
distance from state 2 to state 1 is then 0 + 2 = 2. The letter c is then set to
2. The branch symbol becomes 1/0/2. All the other branch symbols can be
computed in a similar manner.
From depth d = 2 to d = 3, a total of eight possible paths exist. The
path is expressed in terms of state sequence. To reach state 0, for example,

Table 7.3
Distance Between 2-Bit Bauds

00(0) 01(1) 10(2) 11(3)

00(0) 0 1 1 2
01(1) 1 0 2 1
10(2) 1 2 0 1
11(3) 2 1 1 0

6899_Book.indb 118 10/22/19 4:55 PM

 Error-Correcting Codes and Interleaver119

there are two paths 0000 and 0210. The other 6 paths are 0021, 0231, 0002,
0212, 0233, and 0023. If all the paths are kept, the number of paths increases
exponentially. There are two paths to reach each state at d = 3 and only one
path with a smaller Hamming distance is kept. All the kept paths are in dark
colors. For example, to reach state 0, the path 0000 has accumulated Ham-
ming distance of 3 while the path 0210 has an accumulated distance of 4. The
path 0000 is then kept and the path 0210 is discarded. The discarded path
will never be used again. After d = 3, the same algorithm is used to keep or
discard paths. At d = 3 and state 1 for example, the path 0021 is kept and the
branch symbol 1/0/2 is used to update the accumulated distance to transition
from state 1 to state 0 and state 2. Doing it this way, a maximum of 4 paths is
maintained at each depth. What has been discussed is one way of discarding
paths although other possible ways exist.
At each depth and each state, there are only two paths leading to it. At
the end of d = 6 and state 0, for example, the two possible paths are from state
0 and state 1 at d = 5. When all the symbols are received, the decoder finds the
minimum accumulated distance. The minimum distance is zero and the path
0023100 is the final selected path. The encoder output baud sequence and the
input bit sequence associated with this path can be found from the symbol
on each branch. For example, the path 0023100 has an encoder output baud

Figure 7.4 Illustration of the Viterbi decoding algorithm.

6899_Book.indb 119 10/22/19 4:55 PM

 120 Introduction to OFDM Receiver Design and Simulation

sequence 032230 and an input bit sequence of 011000. From Figure 7.4, the
bit sequence of 011000 is indeed the transmitted message. Figure 7.5 shows
the trellis diagram after removing the discarded paths.
Out of the many possible paths through the trellis diagram, the Viterbi
algorithm selects the final path with minimum Hamming distance. In this
sense, it performs the maximum likelihood decoding. To show how the Vit-
erbi algorithm corrects errors, Figure 7.6 shows the trellis diagram with one
transmission error. The input bit sequence is still 011000 and the transmitted
encoder baud sequence is 032230. However, the received baud sequence is
033230 and there is a 1-bit error in the third baud. At each depth and state,
only one path is chosen. If two paths have the same distance, any path can be
selected. One such case is state 1 at depth 5. The branch from state 2 at depth
4 to state 1 has a distance of 3. The same is true for the branch from state 3 at
depth 4 to state 1. In Figure 7.6, the branch from state 3 is arbitrarily selected.
At d = 6, the selected path at states 0, 1, 2, and 3 has an accumulated distance
of 1, 3, 3, and 3, respectively.
The minimum accumulated distance is 1 and the path 0023100 is
selected. This is indeed the correct path and the input bit sequence is cor-
rectly decoded to be 011000. Figure 7.7 shows the trellis diagram after all the
discarded paths are removed.

Figure 7.5 Illustration of Viterbi algorithm after removing the discarded paths.

6899_Book.indb 120 10/22/19 4:55 PM

 Error-Correcting Codes and Interleaver121

Figure 7.6 Illustration of Viterbi decoding with a single bit error.

Figure 7.7 Illustration of Viterbi decoding with a single bit error after removing the
discarded path.

6899_Book.indb 121 10/22/19 4:55 PM

 122 Introduction to OFDM Receiver Design and Simulation

In comparing Figure 7.5 with Figure 7.7, it is seen that not all the dark
lines or paths are the same. This does not matter as long as the correct path
has the minimum distance. If there is no transmission error, the correct path
has a distance of 0. If there are indeed bit errors, the minimum distance is not
0. However, the minimum distance still selects the optimum path, which can
decode the received code word to get the input message stream.

7.4.3 Convolutional Code in the IEEE 802.11a

In the IEEE 802.11a, the convolutional encoder is shown in Figure 7.8. For
every single bit (k = 1) input, two output bits (n = 2) are generated. There-
fore, this is still a 1/2 (k/n = 1/2) convolutional encoder. The total number of
bits in the shift register is 6. Since k = 1, the constraint length is 6. The total
number of states is then 64. The two generator polynomials are defined by
the following equations:

g1(x) = 1 + x + x 3 + x 4 + x 6 (7.38)

g 2 (x) = 1 + x 3 + x 5 + x 6 (7.39)

One property of this encoder is that every output state can only be gen-
erated from two unique input states. Each input state generates two unique
output states. One comes from a bit 0 input and the other comes from a bit

Figure 7.8 Convolutional encoder in the IEEE 802.11a.

6899_Book.indb 122 10/22/19 4:55 PM

 Error-Correcting Codes and Interleaver123

Figure 7.9 Relationship of input and output states.

1 input. Therefore, there exist two input states and two output states that
depend upon each other. Figure 7.9 shows the relationship. In this figure,
both state j and state j + 1 transition to state k after a bit 0 input. However,
if the input bit is 1, both input state j and j + 1 transition to the same output
state i. In other words, there exist two unique input states j and j + 1 that can
only transition to state k and state i. For example, state 8 (j = 8) and state 9
(j + 1 = 9) are two input states that can only transition to state 4 (k = 4) and
state 36 (i = 36). For this generator, the two input states differ by 1 while the
two output states are generally not consecutive. Table 7.4 lists all the input
and output state pairs. For example, input state pair 18 and 19 transitions to
output state pair 9 and 41.

Table 7.4
Generation of Input and Output State Pairs

Input/Output Input/Output Input/Output Input/Output

0,1 /0, 32 16,17/8,40 32,33/16,48 48,49/24,56

2,3/1,33 18,19/9,41 34,35/17,49 50,51/25,57
4,5/2,34 20,21/10,42 36,37/18,50 52,53/26,58
6,7/3,35 22,23/11,43 38,39/19,51 54,55/27,59
8,9/4,36 24,25/12,44 40,41/20,52 56,57/28,60
10,11/5,37 26,27/13,45 42,43/21,53 58,59/29,61
12,13/6,38 28,29/14,46 44,45/22,54 60,61/30,62
14,15/7,39 30,31/15,47 46,47 /23,55 62,63/31,63

6899_Book.indb 123 10/22/19 4:55 PM

 124 Introduction to OFDM Receiver Design and Simulation

In the Viterbi decoding algorithm, several different ways of discarding

paths are possible. One way discussed before and also justified through simu-
lation in Chapter 12 is summarized here:

• Generate the initial 64 states and keep all the paths through the trel-
lis. During the generation process, compute the Hamming distance
between the received baud and the encoder output baud for each
branch, and update the accumulated Hamming distance according
to the following formula

k+1 k
(
C s = C s + d sk+1 ,sk ) (7.40)

where sk represents the state at depth k, d is the Hamming distance

during a state transition, and C is the accumulated Hamming distance.
• After all the initial paths are generated, the total number of states
remains at 64. There are two possible paths to reach a new state.
Only one path with the minimum accumulated Hamming distance
is selected. Mathematically, it is given as

( ) ( ( )
C sk+1 = min C sk+1 , the two paths from sk to sk+1 ) (7.41)

If the two paths have the same Hamming distance, a path can be
selected randomly.
• The process continues until the input message is exhausted. At each
new depth, the number of states is 64 and only 64 paths are kept.

7.4.4 Punctured Convolutional Codes

Higher data rates can be generated by employing the puncturing technique.
This means certain output bits are stolen without being transmitted. In the
receiver, a dummy bit is inserted into the stolen bit location and its distance
metric is ignored. The advantage is that the data rate increases by reducing the
transmission overhead. The disadvantage is a slight increase in the bit error
rate. If the channel is not too bad, the degradation is minimal.
There are two punctured encoders specified in the IEEE 802.11a. The
first has a rate of R = 2/3 and the second has a rate of R = 3/4. Both punctured
encoders use 1/2 convolutional coding as a base to generate the desired rate.
Table 7.5 shows the punctured coding at a rate of R = 2/3. Each column
represents the two output bits for every single bit input. The symbol bij indicates

6899_Book.indb 124 10/22/19 4:55 PM

 Error-Correcting Codes and Interleaver125

Table 7.5
Rate 2/3 Punctured Coding

b 00 b10 b20 b 30

b 01 b11 b21 b31

the first (j = 0) and second (j = 1) output bit corresponding to the ith input
bit. In other words, the first index is the input bit index while the second
index is either the first output bit or the second output bit. For example, b30
is the first output bit due to the third input bit while b31 is the second output
bit due to the same third bit input. The bold text is the output bit that will
not be transmitted. Therefore, every fourth output bit is omitted. Table 7.5
shows a 4-bit input stream [4]. After the first 2-bit input, the output bit b11
is omitted and after the next 2-bit input, the output bit b31 is omitted. The
transmitted bit stream is then b00b01b10b20b21b30. Instead of generating 2 bits
for every single bit input for a rate R = 1/2, 3 bits are generated for every 2-bit
input to generate a rate of R = 2/3.
Table 7.6 [4] shows the punctured coding at rate = 3/4. After every
3-bit input, the fourth and fifth output bits are omitted. Table 7.6 shows
that bits b11 and b20 are omitted after the first 3-bit input and bits b41and
b50 are omitted after the next 3-bit input. The transmitted bit stream is then
b00b01b10b21b30b31b40b51Therefore, for every 3-bit input, only 4 bits are trans-
mitted and the rate becomes 3/4.
In the receiver, the convolutional decoding follows exactly the same
procedure. However, the computation of the Hamming distance is slightly
different. Assume after the dummy bits are inserted, bk0 and bk1 are the received
bits and ak0 and ak1 are the encoder generated output bits before any bits are
stolen. If bk1 is the inserted dummy bit, then both ak1 and bk1 are ignored
and the Hamming distance is the exclusive or operation between bk0 and ak0.
However, if ak0 is omitted and bk0 is the inserted dummy bit, the Hamming
distance is the exclusive or operation between ak1 and bk1. For example, assume

Table 7.6
Rate 3/4 Punctured Coding

b 00 b10 b20 b 30 b40 b50

b 01 b11 b21 b 31 b41 b 51

6899_Book.indb 125 10/22/19 4:55 PM

 126 Introduction to OFDM Receiver Design and Simulation

that 11 is the encoder generated output baud. Assume also that the received
baud is 10 where 0 is the inserted dummy bit. The distance is then the exclu-
sive operation between the first bit 1 of the encoder output baud (11) and the
first bit 1 of the received baud (10) and the result is 0.
Except for the slight modification in computing the cost, the Viterbi
algorithm can be followed in exactly the same way. The way the cost com-
putation as modified will be justified in the simulation given in Chapter 12.

7.5 Interleaver
Even though the error correction discussed in the previous section can correct
bit errors in digital communications, the error performance can be further
enhanced through an interleaver. This is because the errors often occur in
bursts and a group of continuous bits are in error. When this happens, it is
rather difficult for the error correction codes to handle. An interleaver is used
to rearrange the bits following certain rules to prevent a block of bit errors.

7.5.1 Illustration of an Interleaver

A block interleaver is commonly used in digital communications. Assume
that bk,k = 0, N − 1 are the N input bits to the interleaver after error correc-
tion coding. Without an interleaver, these N bits are sent in sequence for data
modulation. With an interleaver, these N bits are rearranged and sent in a
sequence following a given rule.
For illustration purpose, consider a block of N = 14 bits. The bit index of
the interleaver output is bit reversed from that of the input. Table 7.7 shows the
sequence of the input and output bit index. It can be seen that the output bit
index of the interleaver is no longer continuous. Following the third column,
the bit transmission sequence is then b0, b8, b4, b12, and so forth.
Assume that the beginning two bits b0 and b8 are in error after reception.
Before error correction, the bit stream is deinterleaved. The deinterleaving is
just another bit reversal to restore the original bit stream index as given in the
first column of Table 7.7. Note that b 0 and b8 are no longer adjacent and are
now separated by 8-bit positions. Assume that a Hamming (7,4) block code
is used for error correction. In the first 7 bits, only one bit is in error and this
error can be corrected. Without the interleaver, there are two bits in error
and it is beyond the error correction capability of the Hamming code. The
advantage of an interleaver can now be clearly seen.

6899_Book.indb 126 10/22/19 4:55 PM

 Error-Correcting Codes and Interleaver127

Table 7.7
Bit Index of the Interleaver Input and Output

Decimal Input Binary Input Decimal Output Binary Output

Bit Index Bit Index Bit Index Bit Index

0 0000 0 0000
1 0001 8 1000
2 0010 4 0100
3 0011 12 1100
4 0100 2 0010
5 0101 10 1010
6 0110 6 0110
7 0111 14 1110
8 1000 1 0001
9 1001 9 1001
10 1010 5 0101
11 1011 13 1101
12 1100 3 0011
13 1101 11 1011

7.5.2 Interleaver Used in the IEEE 802.11a

In the IEEE 802.11a, the interleaver is defined by a two-step permutation.
The first permutation converts the input bit index k to index i according to
the following formula:
N cbps ⎛ k⎞
i= (kmod16) + int ⎜ ⎟ k = 0,1,…, N cbps …1 (7.42)
16 ⎝ 16 ⎠

where Ncbps is the number of coded bits per symbol, int denotes the integral
part, and mod is the remainder after the division.
The second permutation converts the index i to index j according to the
following formula:

⎛ ⎛ i ⎞⎞ ⎛ ⎛ 16i ⎞ ⎞
j = n ⎜ int ⎜ ⎟ ⎟ + ⎜ i + N cbps − int ⎜ ⎟ ⎟ mod n (7.43)
⎝ ⎝ n⎠⎠ ⎝ ⎝ N cbps ⎠ ⎠

6899_Book.indb 127 10/22/19 4:55 PM

 128 Introduction to OFDM Receiver Design and Simulation

Table 7.8
Input Bit Indexes of an Interleaver in the IEEE 802.11a

0 6 12 18 24 30 36 42

1 7 13 19 25 31 37 43

2 8 14 20 26 32 38 44

3 9 15 21 27 33 39 45

4 10 16 22 28 34 40 46

5 11 17 23 29 35 41 47

where n depends upon the number of bits per subcarrier, Nbpsc, and is given
here:

⎛ N bpsc ⎞
n = max ⎜ ,1⎟ (7.44)
⎝ 2 ⎠

After the two-step permutation, the input bit index k is converted to

bit index j.
For BPSK used in the IEEE 802.11a, each OFDM symbol has 24 data
bits and 48 coded bits after 1/2 convolutional coding. Table 7.8 shows these

Table 7.9
Output Bit Indexes of an Interleaver in the IEEE 802.11a

0 18 36 7 25 43 14 32

3 21 39 10 28 46 17 35

6 24 42 13 31 2 20 38

9 27 45 16 34 5 23 41

12 30 1 19 37 8 26 44

15 33 4 22 40 11 29 47

6899_Book.indb 128 10/22/19 4:55 PM

 Error-Correcting Codes and Interleaver129

48 input bit indexes before an interleaver. Inside each square is the bit index
k. The bit stream is sent column-wise from the first column to the eighth
column. Without an interleaver, the bit indexes are adjacent and continuous.
With an interleaver, the bit indexes are no longer adjacent. Table 7.9 shows the
output bit index j and each bit is reassigned. Bit bk may be assigned to square
m and m ≠ k. For example, bit 1 is assigned to square 16. No two bits can be
assigned to the same square. If bk is assigned to square m and bj is assigned to
square i, then m ≠ i. Table 7.9 clearly shows the bits are reassigned such that
a block of adjacent bit errors can be prevented.

7.5.3 Deinterleaver Used in the IEEE 802.11a

For the IEEE 802.11a, the deinterleaver is the reverse process of the inter-
leaver given in Section 7.5.2. It is completed after two permutations. The
first permutation converts the index j to index i and is given by the follow-
ing equation:

⎛ ⎛ j⎞⎞ ⎛ ⎛ 16 j ⎞ ⎞
i = n ⎜ int ⎜ ⎟ ⎟ + ⎜ j + int ⎜ ⎟ ⎟ mod n j = 0,1,…, N cbps − 1 (7.45)
⎝ ⎝ n⎠⎠ ⎝ ⎝ N cbps ⎠ ⎠

The second permutation converts the intermediate index i back to the

original index k and is given by the following equation:

⎛ 16i ⎞
( )
k = 16i − N cbps − 1 int ⎜ ⎟
⎝ N cbps ⎠
i = 0,1,…, N cbps − 1 (7.46)

Example 7.13
For BPSK defined in the IEEE 802.11a, the interleaver shown in Section 7.5.2
converts index k = 1 to index i = 16. In the deinterleaver, the reverse process
converts the index i = 16 back to index k = 1. In other words, Table 7.9 is
converted back to Table 7.8.

7.6 Summary
For the block codes, both linear block codes and cyclic codes were discussed.
Associated with the linear block codes are a generator matrix and a parity

6899_Book.indb 129 10/22/19 4:55 PM

 130 Introduction to OFDM Receiver Design and Simulation

matrix. The generator matrix is orthogonal to the parity matrix. Every code
word can be generated from the generator matrix. If there are transmission
errors, a nonzero syndrome is created. Associated with every syndrome is a
corresponding error pattern. Based upon the syndrome, the error can be cor-
rected. A special (n, k) Hamming code was also defined such that n = 2m − 1
and k = n − m.
For the cyclic code, a cyclic shift of every code word generates another
code word. Associated with the cyclic code is a generator polynomial and
every code word can be generated through this generator polynomial. For
an (n, k) cyclic code, the generator polynomial is a factor of the polynomial
xn + 1. As with the linear block code, a syndrome polynomial was defined.
If there are no transmission errors, the syndrome polynomial is zero. If
there are transmission errors, every syndrome polynomial has a correspond-
ing error pattern. Based upon the syndrome polynomial, the bit errors can
be corrected.
For the convolutional code of rate k/n, every k information bits generate
n coded bits. The input bit sequence can be infinite in length and the output
bit sequence is generated through the contents of a shift register and input
bits. The input and output bits can be traced through a trellis diagram. Each
possible path is associated with a cost defined to be the Hamming distance
between the received baud and the output coded baud. The Viterbi algorithm
used for decoding is to find the best path through the trellis such that the
cost between the received bits and the encoded output bits is at a minimum.
A special convolutional code defined in the IEEE 802.11a was also discussed
in more detail.
To reduce the transmission overhead, a punctured convolutional code
was defined. It can be generated by taking away a certain number of output
bits. For the 1/2 convolutional code, a punctured code of R = 2/3 and R =
3/4 can be generated. For R = 2/3, one output bit is dropped for every two
input bits. For R = 3/4, two output bits are dropped for every 3 input bits. By
ignoring the cost associated with the dropped bits, the Viterbi algorithm can
be followed in exactly the same way.
An interleaver is also frequently applied to reduce burst errors. The idea
is to distribute the input bits following a certain rule. In such a case, a block
of consecutive bits is transmitted far apart.
Upon reception, these errors can then be easily corrected. A special
interleaver used in the IEEE 802.11a was discussed.
In the next chapter, the OFDM signal detection using the preamble
waveform is discussed. Both coarse detection and fine detection are covered.

6899_Book.indb 130 10/22/19 4:55 PM

 Error-Correcting Codes and Interleaver131

References
[1] Lin, S., and D. Costello, Error Control Coding, Upper Saddle River, NJ: Prentice Hall,
1983.
[2] Proakis, J. G., Digital Communications, New York: McGraw-Hill, 1983.
[3] Viterbi, A. J. “Error Bounds for Convolutional Codes and an Asymptotically Optimum
Decoding Algorithm,” IEEE Transactions on Information Theory, Vol. 13, No. 2, pp.
260−269.
[4] Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY)
Specifications, IEEE Standard 802.11a, 2007.

6899_Book.indb 131 10/22/19 4:55 PM

6899_Book.indb 132 10/22/19 4:55 PM
 8
Signal Acquisition

8.1 Introduction
For OFDM, a preamble is generally sent before the data transmission. The
preamble normally consists of a repetition of signals. As shown in Chapter
4, the preamble in the IEEE 802.11a has a short sequence followed by a long
sequence. The short sequence is a repetition of 16 segments of 16 samples
each. The receiver must be able to first detect the presence of such a repeti-
tion of signals.
Since the preamble signal repeats, adjacent segments must have strong
correlations. This property can be used to define several metrics for signal
detection. Three metrics are given in this chapter. They are cross-correlation,
normalized cross-correlation, and minimum mean square error (MMSE). Their
performances are compared through examples from the simulation study. The
simulation is based upon transmitted signals specified in the IEEE 802.11a.
The results are compared running on a standard compliant C-program.
From the central limit theorem, each of the three metrics can be con-
sidered to have a Gaussian distribution. Using this assumption, a maximum
likelihood detection approach is used for signal detection. This assumption is
justified from the simulation data.
The front-end electronics must first convert the received signal into
the in-phase I component and the quadrature-phase Q component. These

133

6899_Book.indb 133 10/22/19 4:55 PM

 134 Introduction to OFDM Receiver Design and Simulation

IQ signals are then digitized for further processing for signal detection and
others. The traditional heterodyne receiver needs an intermediate frequency
(IF) stage to recover the baseband signal. In this chapter, the received signal
is directly converted without going through an IF stage.
The detection can be done in two steps. The first step is a coarse tim-
ing detection and the second step is a fine timing detection. In order for fast
detection, the coarse detection is also further divided into two steps. The first
is the segment detection and the second is the sample detection.
Each repetition of the preamble signal is called a segment. Initially, there
is no way of knowing in which segment the signal starts. Instead of computing
the correlation for each sample, a fast detection strategy is to first compute the
correlation for each segment. After detecting the signal segment, the sample
detection can be further performed using samples within the detected segment.
Each segment in the IEEE 802.11a has 16 samples. The maximum number
of correlations for the detected segment is then 15.
After the coarse detection, the sample at which the preamble starts is
roughly known. To fine-tune the sample location, a fine detection algorithm is
further applied. In the IEEE 802.11a, a long sequence follows immediately after
the short sequence. The fine timing detection can be applied using either the
long sequence or short sequence. The long sequence is applied for illustration.
The received signals may also suffer impairments from the front-end
receiver. Only random noise is discussed in this chapter through simulation
study. Others are deferred to Chapter 12.
The normalized cross-correlation is found to have the best performance
and is selected for further studies across various SNR levels.

8.2 Direct Conversion to IQ Components

From (2.2), the transmitted signal can, in general, be written as

(
x(t) = Re s(t)e j2pf c t ) (8.1)

where fc is the carrier frequency. Assuming that the baseband signal s(t) has
amplitude a(t) and phase θ (t), we have

s(t) = a(t)e jq(t ) = x I (t) + jx Q (t) (8.2)

The in-phase component xI (t) and quadrature-phase component xQ (t)

are then given by

6899_Book.indb 134 10/22/19 4:55 PM

 Signal Acquisition135

x I (t) = a(t)cos( q(t)) (8.3a)

x Q (t) = a(t)sin ( q(t)) (8.3b)

Substituting (8.2) into (8.1), x(t) can be rewritten as:

(
x(t) = a(t)cos 2pf c t + q(t) ) (8.4)

Using (8.3a) and (8.3b), x(t) can be written in terms of xI (t) and xQ (t)

( ) (
x(t) = a(t)cos( q(t)) cos 2pf c t − a(t)sin ( q(t)) sin 2pf c t ) (8.5)
( ) (
= x I (t)cos 2pf c t − x Q (t)sin 2pf c t )
Assume that the impulse response of the channel is given by h(t). The
received baseband signal y(t) is then just the convolution between the trans-
mitted signal, s(t), and the channel impulse response, h(t).

y(t) = yI (t) + jyQ (t) = ∫ ( xI (τ ) + jxQ (τ ))h(t − t) dt (8.6)

where yI (t) and yQ (t) are just the convolution of h(t) and s(t)

yI (t) = ∫ x I (τ )h(t − t) dt (8.7)

yQ (t) = ∫ x Q (τ )h(t − t) dt (8.8)

The received signal r(t) is then given by

(
r(t) = Re y(t)e j2pf c t ) (8.9)
( )
= yI (t)cos 2pf c t − yQ (t)sin 2pf c t ( )
In (8.9), we have ignored the additive random noise for analysis simpli-
fication, However, the noise impact will be simulated in later sections. In the
RF front end of the receiver, yI (t) and yQ (t) can be recovered by passing the
received signal through a mixing stage. Without any intermediate frequency,
the bandpass signal can be recovered as shown in Figure 8.1. In this mixer,
there are two locally generated signals. Assuming, for now, that there is no IQ

6899_Book.indb 135 10/22/19 4:55 PM

 136 Introduction to OFDM Receiver Design and Simulation

imbalance, the first is l1(t) = 2cos(2π fct) and the other is l2(t) = −2sin(2π fct).
The signal l2(t) has exactly a 90° phase shift from l1(t). Multiplying (8.9) by
l1(t), we have

( ) ( ) ( ) (
r(t)l1(t) = 2 yI (t)cos 2pf c t cos 2pf c t − 2 yQ (t)sin 2pf c t cos 2pf c t )
(8.10)
= yI (t)(1 + cos( 4pf c t )) − yQ (t)sin ( 4pf c t )

After passing through the lowpass filter (LPF), the double frequency
terms are filtered. The output becomes yI (t). Similarly, multiplying (8.9) by
l2(t), we have

( ) ( ) ( ) (
r(t)l2 (t) = −2 yI (t)cos 2pf c t sin 2pf c t + 2 yQ (t)sin 2pf c t sin 2pf c t )
(8.11)
( ) ( (
= − yI (t)sin 4pf c t + yQ (t) 1 − cos 4pf c t ))
Similarly, the double frequency term is filtered and the output becomes
yQ (t). Therefore, the upper branch recovers the in-phase component yI (t) while
the bottom branch recovers the quadrature-phase component yQ (t).

8.3 Detection Metric

The received samples after going through direct conversion becomes
y(t) = yI (t) + jyQ (t) (8.12)

This analog signal then goes through an analog-to-digital converter to

generate digitized samples for further processing. The kth such sample is writ-
ten as yk = yIk + jyQk and yk is, in general, complex.

Figure 8.1 Direct conversion receiver without IQ imbalance.

6899_Book.indb 136 10/22/19 4:55 PM

 Signal Acquisition137

Assume that each repetition segment of the preamble has M samples.

In the IEEE 802.11a, the short sequence has M = 16 and there are 10 such
segments. Since the signal repeats, the following must be true assuming no
channel degradation
yi+k = yi+k+ M (8.13)

where i is the start sample index, k is the sample index, and M is the period
in sample count. Based upon (8.13), several detection metrics can be utilized.

8.3.1 Cross-Correlation
In this scheme, two signals separated by M samples apart are correlated. The
metric m is given here [1]:
k= M −1
mi = ∑ ∗
yi+k yi+k+ M
(8.14)
k=0

In (8.14), the first signal is R1 = yi+k while the second signal is R 2 = yi+k+M.
The separation between these two signals is M samples.
The correlation must be low if both signals are noise. The correlation
increases if the signal appears in either R1 or R 2. The correlation reaches a
maximum when both R1 and R 2 are signals. Assuming that there is no noise,
then (8.13) applies and (8.14) becomes
k= M −1
∑
2
mi = yi+k
k=0

In this case, mi is just the energy in one segment.

8.3.2 MMSE Metric

An MMSE metric utilizes the characteristics of (8.13). In other words, the dif-
ference between yi+k and yi+k+M must be a minimum when the signal appears.
The metric can then be given as [2]

k= M −1
∑
2
mi = yi+k − yi+k+ M (8.15)
k=0

6899_Book.indb 137 10/22/19 4:55 PM

 138 Introduction to OFDM Receiver Design and Simulation

Expanding (8.15), the following is obtained:

k= M −1
∑ ( yi+k − yi+k+ M )( yi+k − yi+k+ M )
∗
mi =
k=0
k= M −1 k= M −1 k= M −1 k= M −1
∑ ∑ ∑ ∑
2 2 ∗ ∗
= yi+k + yi+k+ M − yi+k yi+k+ M
− yi+k yi+k+ M
k=0 k=0 k=0 k=0
k= M −1 k= M −1 k= M −1
= ∑
2
yi+k + ∑ yi+k+ M − 2
2
∑ ( ∗
Re yi+k yi+k+ M )
k=0 k=0 k=0

(8.16)

From (8.15), it can be seen that the metric is always positive. When the
signal is present and without noise, (8.13) applies. Then from (8.15), mi = 0
and this is also the minimum value of mi.

8.3.3 Normalized Cross-Correlation

The metric in (8.16) can be normalized with respect to the signal energy
given here:
k= M −1 k= M −1
∑ ∑
2 2
Ei = yi+k + yi+k+ M (8.17)
k=0 k=0

Dividing mi in (8.16) by (8.17), we have [2]

k= M −1

mi
2 ∑ ( ∗
Re yi+k yi+k+ M )
= 1− k=0 (8.18)
Ei Ei

Minimizing mi/Ei is equivalent to maximizing the second term of (8.18).

The normalized metric is then defined below:
k= M −1
2 ∑ ( ∗
Re yi+k yi+k+ M )
mi (normalized) = k=0
(8.19)
Ei

Using (8.13) under the condition without noise, mi in (8.19) is then

equal to one. With a signal present, the metric is close to 1. With only noise

6899_Book.indb 138 10/22/19 4:55 PM

 Signal Acquisition139

present, the correlation is even smaller due to the further division by the energy.
Therefore, (8.19) is a good metric to detect the signal in noise.
For all three metrics given in (8.14), (8.16), and (8.19), the search is to
find either the maximum or minimum when the signal appears. Since there
is no way of knowing when the signal is transmitted, the metric is initially
computed each time by advancing the index i by the period M. In other
words, we compute mi, mi+m, mi+2M until mi+lM such that the signal segment
is detected. In doing so, more precise signal timing is not known in exchange
for fast detection.

8.4 Maximum Likelihood Detection

Even though the transmitter does not send mi directly, it can be computed
from the received samples. To detect the signals in noise, it becomes equiva-
lent to detect the metric in noise. There are now two hypotheses. In the first
hypothesis, H1, the signal is present while in the second hypothesis, H0, only
noise is present.
Let y represent the received metric sample. Based on y, either H1 is true
or H0 is true. If H1 is true, the probability density function of y is p(y/H1).
However, the probability density function of y is p(y/H0) if H0 is true. The
decision rule is then to choose H1 if

p( y/H1 )
>l (8.20)
(
p y/H 0 )
where λ is a detection threshold [3]. The ratio on the left side of (8.20) is called
the likelihood ratio. A test based upon this ratio is called a likelihood ratio test.
Under hypothesis H1 and H0, the variable y is given below:
H1 : y = u + n (8.21)

H0 : y = n (8.22)

where n is the random noise and u is the average value of the metric. Assum-
ing that the random noise n to be Gaussian distributed with mean zero and
standard deviation σ , the probabilities p(y/H1) and p(y/H0) are then given here:

⎛ y ⎞ 1 ⎡ −( y−u )2 /2s 2 ⎤
p⎜ = e⎣ ⎦ (8.23)
⎝ H1 ⎟⎠ 2ps

6899_Book.indb 139 10/22/19 4:55 PM

 140 Introduction to OFDM Receiver Design and Simulation

⎛ y ⎞ 1
e ⎡⎣− y /2s ⎤⎦
2 2
p⎜ = (8.24)
⎝ H 0 ⎟⎠ 2ps

Substituting (8.23) and (8.24) into (8.20), we have

e(2 yu−u )/2s > l

2 2
(8.25)

Taking logarithm on both sides of (8.25), we have

u2 + 2s 2 ln l
y> (8.26)
2u

The decision rule can be to choose H1 if p(y/H1) > p(y/H0). In this case,
λ = 1 and (8.26) becomes
u
y> (8.27)
2

Equation (8.27) is a simple decision rule. If the computed metric is greater

than one-half of its average signal value, the signal is detected.

8.5 Coarse Timing Detection

The short sequence defined in Section 4.4.1 is used for the simulation of the
coarse detection. The transmitted signal is assumed to go through no channel
degradations and synchronization errors. That also means the channel impulse
response h(t) = 1 in (8.7) and (8.8). However, the random noise impact on the
accuracy of signal detection is provided through simulation study.
The coarse detection is done in two steps. The first is the segment detec-
tion to quickly locate the segment the signal appears. The second is the sample
detection to find the onset of signal samples.

8.5.1 Segment Detection

In the segment detection, the goal is to find the segment where the signal is
located. The segment size is M and the metric is computed by advancing the
samples by M each time. In other words, only metrics m 0, m16, m32, m48, and
so forth are computed. In the following examples, two different noise levels
are considered. One has SNR = 37.8 dB and the second has SNR = 2.8 dB.

6899_Book.indb 140 10/22/19 4:55 PM

 Signal Acquisition141

Example 8.1
In this example, the SNR is set to be 37.8 dB. Figures 8.2 to 8.4 are plots of
cross-correlation, normalized cross-correlation, and MMSE against the seg-
ment number. Each segment of the short sequence has 16 samples. The index
i of the metric mi is advanced by 16 each time and has numbers 16, 32, 48,
and so forth. The segment number is just i/16 and has numbers 1, 2, 3, and
so forth. The short sequence starts at segment 26.
Figure 8.2 shows that the noise correlation is negligibly small. The cross-
correlation is 0.1 at segment number 26 and reaches a maximum of 0.2 for
segment number ≥ 27. Since the short sequence samples start within segment
26, the cross-correlation is not at maximum. The average signal level u in
(8.27) is then 0.2 and the threshold is u/2 = 0.1. Clearly, the signal is correctly
detected at segment 26. Figure 8.3 shows that the normalized cross-correlations

Figure 8.2 Cross-correlation against segment number at SNR = 37 dB.

Figure 8.3 Normalized cross-correlation against segment number at SNR = 37.8 dB.

6899_Book.indb 141 10/22/19 4:55 PM

 142 Introduction to OFDM Receiver Design and Simulation

Figure 8.4 MMSE against the segment number at SNR = 37.8 dB.

maximum is 1. The threshold is then 0.5. The signal is also correctly detected
at segment 26 with a normalized cross-correlation of 0.66. Since the signal is
normalized, the noise correlation is not small anymore and the value can be
either positive or negative as expected.
Figure 8.4 shows the MMSE and the detection is to find the minimum
and (8.27) does not apply. Because of signal and noise mismatch at segments
25 and 26, the MMSE has a larger value. When the signal matches again
starting in segment 27, the MMSE does reach a minimum. Since there is
a peak at segment 26, it cannot be detected as the starting signal segment.
Another issue is that the minimum is almost indistinguishable for noise, and
thererfore, confusion can arise.

Example 8.2
In this example, the SNR is set to be 2.8 dB. Figures 8.5 to 8.7 are plots of
the three metrics against the segment number. For cross-correlation, Figure
8.5 shows at segment 26, the metric is 0.16 and is greater than the threshold
0.1. Even with significant noise, the signal is still correctly detected. At seg-
ment 25, the metric is 0.069 and there is no false alarm. Figure 8.6 shows the
normalized cross-correlation. At segment number 25, the metric is −0.029,
and at segment number 26, the metric is 0.586. Therefore, the signal is still
correctly detected at segment 26. Figure 8.7 shows the MMSE. At segment
number 25, the signal has a peak and drops immediately starting at segment
26. Again, the miniumum is almost indistinguishable from the noise.
The simulation justifies the maximum likelihood detection strategy.
Both cross-correlation and normalized cross-correlation work well. However,

6899_Book.indb 142 10/22/19 4:55 PM

 Signal Acquisition143

Figure 8.5 Cross-correlation against the segment number at SNR = 2.8 dB.

Figure 8.6 Normalized cross-correlation against segment number at SNR = 2.8 dB.

Figure 8.7 MMSE against the segment number at SNR = 2.8 dB.

6899_Book.indb 143 10/22/19 4:55 PM

 144 Introduction to OFDM Receiver Design and Simulation

these are by no means the only methods to detect the signal in noise. Another
standard method is to estimate the noise level and its standard deviation. For
signal detection, the signal level must be greater than the noise level by at least
the noise standard deviation.
For cross-correlation, the threshold parameter u in (8.27) can be set at
its maximum value. As will become clear later, this value is difficult to set in
advance. For normalized cross-correlation, it is theoretically set at 1 and is a
constant. In practice, it can be set at slightly less than 1 in order not to miss
the sample onset due to interference and others.

8.5.2 Sample Detection

After locating the signal segment, the next step is to locate the starting sample.
In the segment detection, the index i of the metric mi increases by 16 each
time. In the sample detection, i has to advance by 1. In other words, only m1,
m2, m3, and so forth are computed. The question is what should be the value
of M or the number of samples used for computing the correlation. If M is set
to the maximum of 16, the correlation can be too large to distinguish between
noise and signal. This is because too many signal samples are included in the
computation. This means that M should be set small.
The impact of SNR and M are investigated through the following two
examples. Again, the SNR is either 37.8 dB or 2.8 dB. Two different M val-
ues are used. One is 16 and the other is 2. Since the MMSE has the worst
performance, only cross-correlation and normalized cross-correlation are used
for the simulation study.

Example 8.3
From previous examples, the signal segment is 26. The starting sample is 16 ∗
25 = 400 since the first segment is 1. Figure 8.8 is a plot of cross-correlation
against sample number within segment 26 at SNR = 37.8 dB. The top figure
uses M = 16 for correlation while the bottom figure uses M = 2 for correlation.
Since the signal appears somewhere after a certain sample number, the signal
u can be set at a sample number of 15. The threshold of the top figure is then
0.5m15 = 0.1. However, the correlation is greater than 0.1 for all the sample
numbers and it is not clear where the signal starts. In the bottom figure, the
correlation is much smaller since only 2 samples are used in the correlation.
The threshold is then 0.5m15 = 0.5 ∗ 0.022 = 0.011. The curve shows the cor-
relation is much less than 0.011 until sample number 8. Since sample number
8 has a correlation greater than the threshold 0.011, it is correctly identified as
the signal starting sample. Figure 8.9 is a plot of normalized cross-correlation

6899_Book.indb 144 10/22/19 4:55 PM

 Signal Acquisition145

Figure 8.8 Plot of cross-correlation against the sample number at SNR = 37.8 dB.

against a sample number. Since the maximum correlation is 1 and the thresh-
old is 0.5. Again, for M = 16, it is not clear where the signal starts. For M = 2,
the correlation is greater than 0.5 for sample numbers greater than 8. Again,
sample number 8 is correctly identified as the signal starting point. It is also
obvious that the percentage jump at signal onset for the normalized cross-
correlation is much higher than that for the cross-correlation.

Example 8.4
To see the impact of noise, the SNR is set to be 2.8 dB. Figure 8.10 is a plot
of cross-correlation. The top figure uses M = 16. To estimate the detection

Figure 8.9 Normalized cross-correlation against the sample number at SNR = 37.8 dB.

6899_Book.indb 145 10/22/19 4:55 PM

 146 Introduction to OFDM Receiver Design and Simulation

Figure 8.10 Cross-correlation against the sample number at SNR = 2.8 dB.

threshold, the sample value at m15 is used. The threshold is then approximately
given by m15/2 = 0.24/2 = 0.12. Figure 8.10 shows that the cross-correlation is
always greater than 0.12 and it is not clear again where the signal starts. The
bottom figure uses M = 2 and the threshold is approximately set at 0.024. The
correlation is greater than the threshold at sample 0 and at sample number
greater than 10. Sample 0 cannot be the signal starting point since future cor-
relations can be smaller than the threshold. Therefore, sample 10 is selected as
the signal onset and there is an error of 2 samples. Since the threshold has to
be set before the detection can start, this shows the disadvantage of the cross
correlation using the maximum likelihood detection strategy.
Figure 8.11 is a plot of normalized cross-correlation. The threshold is
always 0.5. The top figure is for M = 16 and it is not clear where the signal

Figure 8.11 Normalized cross-correlation against the sample number at SNR = 2.8 dB.

6899_Book.indb 146 10/22/19 4:55 PM

 Signal Acquisition147

starts. The bottom figure is for M = 2. The correlation is greater than 0.5 at
sample 1, 5 and all samples greater than 9. Sample 1 and sample 5 cannot be
the choice since future correlations can drop below 0.5. Therefore, sample 9
is the choice and there is an error of 1 sample.

8.6 Fine Timing Detection

After the starting signal sample location is roughly detected, the next step is
to make some potential refinements. Since the preamble is a known signal
sent from the transmitter, the receiver has a duplicate copy of this signal. For
the IEEE 802.11a, the preamble consists of two parts. The first part is the
short sequence, while the second part is the long sequence. Since the short
sequence is used for coarse signal detection, the long sequence can be used
for fine signal detection.
Assume that the clean preamble waveform is given by ck where k = 0,
…, M − 1 and M is the sample size. This clean waveform then starts at k = 0.
After a coarse timing estimate, assume that yL is aligned with c0. Using the
normalized cross-correlation as the metric, the following are defined:
k= M −1 k= M −1
∑ ∑
2 2
Ed = ck + yk+d+ L (8.28a)
k=0 k=0

k= M −1
2 ∑ ( ∗
Re c k yk+d+ L )
md (normalized) = k=0
(8.28b)
Ed

where the parameter δ is the sample deviation index from k = 0 of the clean
preamble waveform. If there is a perfect match, δ = 0 and the following is true:
c k = yk+ L k = 0,…, M − 1 (8.29)

Applying (8.29) to (8.28b), we have

k= M −1
∑
2
E0 = 2 ck (8.30)
k=0

k= M −1
2 ∑ (
∗
Re c k yk+ L )
m0 (normalized) = k=0
=1 (8.31)
E0

6899_Book.indb 147 10/22/19 4:55 PM

 148 Introduction to OFDM Receiver Design and Simulation

This means the metric has the maximum value of 1 with a perfect match.
After the coarse timing estimate is performed, the value δ should not be
too far away from zero. The value δ can be assumed to have the range from
−α to α . In other words, the metric given in (8.26) is computed a total of 2α
+ 1 times to find the best match. The index δ that generates the maximum
metric matches with clean preamble waveform.

Example 8.5
Using the IEEE 802.11a waveform, the long sequence has a total of 160 samples
at a 20-MHz channel spacing. It starts with 32 guard samples followed by
the first symbol of 64 samples and the second symbol of 64 samples. At SNR
= 37.8 dB, the detection of short sequence finds that it starts at sample 408.
The second long symbol then starts at 408 + 32 + 64 = 504. The parameter L
is then set to 504. In other words, the received sample index 504 is assumed
to match with sample index 0 of the clean waveform. At SNR = 2.8 dB, the
short sequence is detected to start at sample index 409. In this case, sample
index 505 is assumed to match with sample 0 of the clean waveform.
Figure 8.12 is a plot of the normalized cross-correlation against the sam-
ple deviation index δ . The dark curve is for SNR = 37.8 dB and the gray curve

Figure 8.12 Normalized cross-correlation with clean waveform against sample

deviation.

6899_Book.indb 148 10/22/19 4:55 PM

 Signal Acquisition149

is for SNR = 2.8 dB. The parameter α is set to 5 and a total of 11 searches are
performed. As expected, the maximum for SNR = 37.8 dB is 1 appearing at δ
= 0. The first OFDM symbol of the long sequence then starts at sample index
504. At SNR = 2.8 dB, the maximum appears at sample deviation index −1.
The maximum is not 1 due to the background random noise. Since the value
δ = 0 corresponds to the received sample index 505, the first OFDM symbol
of the long sequence then starts at received sample index 504.
Even though the coarse timing estimate may misalign the starting sample
by 1 at SNR = 2.8 dB, the fine timing estimate can indeed correct the error
and again make the correct adjustment.

Example 8.6
Based upon the simulation studies, the normalized cross-correlation has the
best performance. This metric is then selected for further studies in various
SNR levels. Figure 8.13 is a plot of the performance of coarse sample detec-
tion in random noise. The vertical axis shows the number of samples deviated
from the correct signal sample onset. The horizontal axis shows the SNR in
decibels. For SNR above 3.1 dB, there is no sample detection error and the
sample error count is 0. For SNR between −2.08 dB and 3.01 dB, there is
one sample detection error. For SNR between −4.18 dB and −2.18 dB, there

Figure 8.13 The noise performance of coarse sample detection.

6899_Book.indb 149 10/22/19 4:55 PM

 150 Introduction to OFDM Receiver Design and Simulation

are two sample detection errors. For SNR below −4.18 dB, the signal segment
cannot be detected and the signal acquisition fails.
For segment detection, the segment detection error is always zero, as long
as the SNR is above −4.18 dB. As is mentioned before, no signal segment can
be detected for SNR below −4.18 dB. Even though there is either a one-sam-
ple or two-sample detection error from the coarse sample detection, the fine
sample detection can always correct the error. That means there is no sample
detection error for the fine detection as long as the SNR is above −4.18 dB.

8.7 Summary
The received radio waveform is first processed to remove the carrier. After
going through a direct conversion, in-phase and quadrature-phase signals
are obtained. These two signals can then be digitized for further processing.
For OFDM waveform, a periodic preamble signal is normally sent to help
the receiver for signal detection. This preamble waveform used in the IEEE
802.11a is a short sequence followed by a long sequence. The short sequence
has 16 periodic segments and is used for coarse timing detection. The long
sequence has two periodic segments and can be used for fine timing detection.
For any detection algorithm, the detection metric must be defined.
Three such metrics were discussed. They are cross-correlation, normalized
cross-correlation, and MMSE. The cross-correlation and normalized cross-
correlation find the correlation maximum while MMSE finds the correlation
minimum. A detection algorithm based upon maximum likelihood criterion
was derived. The simple algorithm shows a signal detected if the metric is
greater than a threshold. For normalized cross-correlation, this threshold is
simply set at 0.5 theoretically.
The detection algorithm follows a two-step procedure. The first step
is coarse timing detection followed by fine timing detection. Since the pre-
amble is periodic, we develop further a two-step approach for fast detection
of a coarse timing. The first step is the segment detection in which the cor-
relation is advanced by a segment. For the IEEE 802.11a, the short sequence
has 16 samples in each segment. In other words, the correlation between two
consecutive segments is 16 samples apart. The idea is to find which segment
the signal starts. After segment detection, the sample detection in which the
correlation is performed on a sample basis follows immediately. The idea is to
locate which sample the signal starts. From the simulation study, it was found
that the signal sample is difficult to locate if the correlation size is large. It was
found that the best correlation size is M = 2.

6899_Book.indb 150 10/22/19 4:55 PM

 Signal Acquisition151

The three metrics were compared from the simulation study. The MMSE
metric was found to have the worst performance since the minimum can be
indistinguishable from noise. The cross-correlation somehow works; however,
the detection threshold is difficult to set and depends upon the correlation size
and SNR. The normalized cross-correlation not only has the best performance,
but is also the easiest metric from the computational point of view. The deci-
sion threshold from the theoretical analysis is simply 0.5 and is independent
of SNR and correlation size.
After the coarse detection is completed, the fine detection follows. The
fine detection computes the correlation between a clean preamble and that
from the received waveform. Since the transmitted preamble is known to
the receiver, this known waveform can be stored in the receiver for the fine
detection. After the fine detection, the signal sample onset can be accurately
determined. Since normalized correlation has the best performance, it is only
used for the simulation study. Based upon the simulation in the examples, the
coarse detection could accurately locate the signal sample at SNR = 37.8 dB,
but with one sample error at SNR = 2.8 dB. However, the fine detection can
accurately correct the error.
Lastly, the signal detection performance in random noise was investi-
gated. As long as the SNR is above −4.18 dB, there is no sample detection error
after the fine sample detection. For SNR below −4.18 dB, no signal segment
can be located and the detection fails.
The next chapter centers on synchronization issues in the receiver. The
topics covered include DC offset, carrier frequency offset, frame timing offset,
sampling clock offset, and IQ imbalance.

References
[1] Schmidt, T., and D. Cox, “Robust Frequency and Timing Synchronization for OFDM,”
IEEE Transactions on Communications, Vol. 45, No. 12, December 1997, pp. 1631−1621.
[2] Minn, H., V. Bhargava, and K. Letaief, “A Robust Timing and Frequency Synchronization
for OFDM Systems,” IEEE Transactions on Wireless Communications, Vol. 2, No. 4,
2003, pp. 822−839.
[3] Whalen, A. D., Detection of Signals in Noise, San Diego: Academic Press, 1971.

6899_Book.indb 151 10/22/19 4:55 PM

6899_Book.indb 152 10/22/19 4:55 PM
 9
Synchronization

9.1 Introduction
The received OFDM signal normally suffers two types of degradations. The
first is due to the propagation effects from the channel and is unavoidable. The
second is due to the imperfect front-end electronics that can cause frequency
and phase mismatches between the transmitter and the receiver. Since the
circuit parts can never completely match, some compensation schemes must
be taken to minimize its impact. The synchronization deals with both issues.
Regarding the channel effect, the received OFDM signal may suffer
Doppler spread. It can be caused by relative motion between the transmit-
ter and the receiver. The received carrier frequency is then different from the
transmitted frequency by the Doppler shift. The second is due to the channel
variations that cause frequency dispersion or time-selective fading. The fad-
ing can be fast if the coherence time is less than the signal duration. Another
distortion that the OFDM transmission suffers is frequency-selective fading.
This is because signals with different delays are often received due to reflec-
tions from the ground or obstacles. This delay spread can be longer than the
OFDM guard interval or the channel coherence bandwidth can be shorter
than the signal bandwidth to cause serious signal distortions.
Regarding the imperfect circuit parts, the local receiver clock can run
slightly differently in comparison with the transmitter clock. This can cause

153

6899_Book.indb 153 10/22/19 4:55 PM

 154 Introduction to OFDM Receiver Design and Simulation

some problems. Because of the different clocks, the local oscillator (LO) can-
not generate the same carrier frequency and phase as the received signal. First,
it causes the carrier frequency offset (CFO). Second, it causes IQ imbalance
because the two LOs used to recover the in-phase and quadrature-phase
components of the received signal run at a different frequency and phase. The
received signals can also be sampled at different time instants in comparison
with the transmitter to cause misalignments of the frame boundary. Another
issue is the DC bias caused by the leakage current from the LO to the front-
end linear amplifier.
All these synchronization issues including DC offset, CFO, sampling
clock offset (SCO), sampling time offset, and IQ imbalance are analyzed in
detail in this chapter. The parameters associated with each synchronization
issue are estimated. A number of examples from computer simulations are
also given to show the accuracy of the estimate.

9.2 DC Offset
After passing through the RF front end, the digitized baseband signals fre-
quently have a DC component. The DC signal is not random and can be
wrongly detected as the transmitted signal, thereby causing a false alarm.
Therefore, it is important to remove the DC bias before further processing.
In an OFDM receiver, direct conversion of the RF band signals to base-
band is frequently used. However, one of the drawbacks is DC bias. In this
section, the DC bias is analyzed and the compensation scheme is also presented.

9.2.1 Algorithm Analysis

In a direct-conversion receiver, the converted IQ signal may have a DC offset.
This phenomenon is depicted in Figure 9.1 [1, 2]. The received signal passes
through a linear amplifier (LNA) and mixes with an LO. The output from

Figure 9.1 LO leakage.

6899_Book.indb 154 10/22/19 4:55 PM

 Synchronization155

the mixer is lowpass filtered and subsequently digitized using an analog-to-

digital converter.
If the isolation between the LO and the mixer input or LNA input is not
perfect, a leakage path may exist from LO to A (LNA input) or from LO to
B (mixer input). The leakage then mixes with LO again to generate the DC
offset. Assume that the LO signal is cos(2π fct) and the leakage is α cos(2π fct).
The mixing then produces the following:

( )( (
cos 2pf c t a cos 2pf c t )) = α cos2 (2pf c t )
(9.1)
a a
= + cos( 4pf c t )
2 2

The first term of (9.1) is the DC offset.

Besides the leakage to the LNA or mixer input, the leakage can also go
to the antenna. The leakage radiated from the antenna can be reflected from
the surrounding obstacles and received again by the antenna. The received
reflection then mixes with the LO to generate the DC offset.
For OFDM, the data demodulation is done in the frequency domain. The
DC offset only generates an extra zero frequency term in the frequency domain.
Therefore, as long as the signal detection can still find the correct signal onset
location, the DC offset cannot affect the accuracy of data demodulation.
On signal detection, the DC offset should have minimal impact on the
MMSE metric. Assume that the DC offset is the constant c. Then the MMSE
metric becomes
k= M −1
∑ ( zi+k + c ) − ( zi+k+ M )
2
mi = +c
k=0
k= M −1
(9.2)
∑
2
= zi+k − zi+k+ M
k=0

The DC offset term is cancelled out and the metric is unchanged.

Using the metric of normalized cross correlation for signal detection, the
impact can be significant. With the addition of a DC offset, the noise correla-
tion between adjacent segments becomes highly correlated. The background
noise can be mistakenly detected to be the signal to cause false alarm. From
(8.17) to (8.19), we have the following by considering the case without any
input signal
yj = nj + d (9.3)

6899_Book.indb 155 10/22/19 4:55 PM

 156 Introduction to OFDM Receiver Design and Simulation

where d is the DC offset. The correlation between two segments of M samples

apart becomes
k= M −1
Ri = 2 ∑ ( ∗
Re yi+k yi+k+ M )
k=0
k= M −1
=2 ∑ ( )(
∗
Re ni+k + d ni+k+ M
+d ) (9.4)
k=0
k= M −1 k= M −1 k= M −1
=2 ∑ ( *
Re ni+k ni+k+ M )
+ 2d ∑ (
Re ni+k + ni+k+
*
M
+2 ) ∑ d2
k=0 k=0 k=0

Since the noise average is zero and there is no correlation between two
different noise segments, the first and second terms of (9.4) are both zero. We
then have the following:

Ri = 2Md 2 (9.5)

Similarly, the energy from (8.17) can be shown as given here:

k= M −1 k= M −1
∑ ∑
2 2
Ei = yi+k + yi+k+ M
k=0 k=0 (9.6)
= 2Md + 2 EN 2

where EN is the noise energy and is defined here:

k= M −1
EN = ∑ nk2 (9.7)
k=0

The normalized correlation metric then becomes

Ri Md 2
mi = = (9.8)
Ei Md 2 + EN

If the background noise energy is very small, then mi is close to 1. Under

this condition, the noise will be wrongly detected to be the received signal.
To avoid this problem, the DC bias needs to be removed. In the analog
domain, the AC coupling can be used [2]. In the digital domain, one way is
to estimate the DC bias and to remove it. However, the estimation can be

6899_Book.indb 156 10/22/19 4:55 PM

 Synchronization157

difficult if there are other channel related degradations such as multipath

and fading. The easiest way is to pass the received signal through a highpass
filter to remove the DC bias. The FIR filter has the advantage of linear phase.
Assume that this filter has length N and coefficients, hk, k = 0, …, N − 1. The
filter output zk due to an input yk is then given here:
N −1
zn = ∑ hk yn−k (9.9)
k=0

9.2.2 Simulation Examples

The preamble waveform from the IEEE 802.11a was used for the study. At 20
MHz, the short sequence has 10 segments of 160 samples and each segment
has 16 samples. The long sequence has also 160 samples, which start with 2
guard intervals of 32 samples and are followed by two OFDM symbols of
128 samples. There are no channel degradations and other synchronization
issues. However, background noise is mixed with the preamble waveform to
the generated SNR at 37.8 dB and 2.8 dB. The coarse detection is done using
the short sequence while the fine detection is done using the long sequence.
As expected, the noise was detected as the received signal without using a
highpass filter. In Example 9.1, the advantage of using a highpass filter to filter
out the DC bias is illustrated.

Example 9.1
To design a highpass filter, the frequency sampling design was used [3]. This
filter has 65 coefficients with h0 = 0.984615 while h1 to h64 all have the same
value of −0.015385. The frequency response is shown in Figure 9.2 against
the frequency index. A frequency index of 65 corresponds to 2π . Clearly, the
frequency response is 0 at DC and 1 at all the other frequency indexes.
The computation of the output response can be simplified by utilizing
the fact that all filter coefficients are the same except h0. Assuming that hk =
hc, k = 1, …, N − 1, yk is the input, and zk is the output, we have
k= N −1
zn = ∑ hk yn−k
k=0
(9.10)
( )
= h0 yn + yn−1 + ! + yn− N +2 + yn− N +1 hc

Similarly, we have the following for zn+1:

6899_Book.indb 157 10/22/19 4:55 PM

 158 Introduction to OFDM Receiver Design and Simulation

Figure 9.2 Highpass filter frequency response.

(
zn+1 = h0 yn+1 + yn + yn−1 + … + yn− N +2 hc )
(
= h0 yn+1 + yn hc − yn− N +1hc + yn−1 + … + yn− N +2 + yn− N +1 hc ) (9.11)
= zn − h0 yn + h0 yn+1 + yn hc − yn− N +1hc
( ) (
= zn + h0 yn+1 − yn + hc yn − yn− N +1 )
Instead of having N additions and multiplications, (9.11) only has 4
additions and 2 multiplications.
Assuming that the ith preamble signal is yi, the ith random noise sample
is ni, and the DC offset is di, then the output signal zi becomes
zi = yi + ni + di (9.12)

where the random noise has zero mean. The DC offset is assumed to be a
constant and has the following form:
di = c + jc (9.13)

where c is a constant. At SNR = 37.8 dB and 2.8 dB, the DC offset d ranging
from 0 to 0.059 is used. The short sequence has an average amplitude of 0.112,

6899_Book.indb 158 10/22/19 4:55 PM

 Synchronization159

Table 9.1
Deviation of Coarse and Fine Detection at SNR = 37.8 dB

Average Coarse
Clean Short Detection Fine Detection
Sequence DC Offset DC Offset Sample Sample
Amplitude Amplitude Percentage Deviation Deviation

0.112 0.000 0% 0 0
0.112 0.020 17.9% 0 0
0.112 0.04 35.7% 0 0
0.112 0.059 52.7% 0 0

which is the square root of the average short sequence energy. It is assumed
that the addition of a DC offset does not saturate the power amplifier. The
average DC offset amplitude is then 2c. Tables 9.1 and 9.2 list the results at
SNR = 37.8 dB and 2.8 dB, respectively. The SNR is computed without add-
ing DC bias. In both tables, the DC offset percentage is the ratio of DC offset
amplitude against the average short sequence amplitude. The coarse detection
deviation is the number of the sample difference between the coarse detected
sample and the actual sample onset. A plus sign means the detected position
is larger than the true position, while a minus sign means the contrary. The
fine detection deviation is the deviation due to the fine detection.
Both Tables 9.1 and 9.2 show that the DC bias has minimum impact
on the signal detection after highpass filtering.

Table 9.2
Deviation of Coarse and Fine Detection at SNR = 2.8 dB

Average Coarse Fine

Clean Short Detection Detection
Sequence DC Offset DC Offset Sample Sample
Amplitude Amplitude Percentage Deviation Deviation

0.112 0.000 0% 0 0
0.112 0.020 17.9% 0 0
0.112 0.040 35.7% 0 0
0.112 0.059 52.7% 0 0

6899_Book.indb 159 10/22/19 4:55 PM

 160 Introduction to OFDM Receiver Design and Simulation

9.3 CFO
In the front end of the OFDM receiver, the received carrier frequency may
be different from the transmitted carrier frequency. This happens because of
the relative motion between the transmitter and the receiver. The net result is
a Doppler shift and was discussed in Chapter 6.
Besides the Doppler shift, there are also interferences from the neighbor
subcarriers. Equation (3.14) and Figure 3.1 show such an effect. It is clear
from Figure 3.1 that interferences exist in the region between peak subcar-
riers. However, at the peak subcarrier frequency, there are no interferences
from the rest of subcarriers.
This frequency offset due to Doppler shift can degrade system perfor-
mance and must be removed. The acquisition signal in OFDM normally has
a repetitive structure that can be used to find the frequency offset. In this
section, the estimate of frequency offset due to Doppler shift is discussed.

9.3.1 Algorithm Analysis

There are two possibilities to generate the CFO. One is the Doppler shift
caused by the relative motion between the transmitter and the receiver. This
happens even if there is a perfect match between the transmitter and receiver
frequency. The other is that the LO frequency of the receiver may deviate from
the transmitted carrier frequency.
Assuming that the frequency deviation is given by f δ, the received carrier
frequency is then fc + f δ. The received signal r(t) given in (8.9) then becomes

(
r(t) = Re y(t)e j2p( f c + f d )t ) (9.14)
= Re ( z(t)e j2pf c t
)
where z(t) is given by

z(t) = y(t)e j2pf d t (9.15)

The baseband signal that is actually received is then z(t).

Assuming that the sampling frequency is fs and the FFT size is N, the
subcarrier frequency spacing is then

fs
Δf = (9.16)
N

6899_Book.indb 160 10/22/19 4:55 PM

 Synchronization161

Assuming also that the ratio of f δ against Δf is ò, the following then results:
Nf d
ε= (9.17)
fs

After analog-to-digital conversion, the nth sample of z(t) then becomes

zn = yn e j2pεn/N (9.18)

As shown in Section 1.8.3 of Chapter 1, a multiplication by a phase

term in the time domain is equivalent to a frequency shift in the frequency
domain. We have the following:
Zk = Yk−ε (9.19)

where Zk is the discrete Fourier transform of zn and Yk is the discrete Fourier

transform of yn.
To estimate ò, the repetitive structure of the preamble can be utilized.
The frame size M of the preamble can be set to the FFT size N. Since the
period is then N, the following is true:
yn+ N = yn n = 0,…, N − 1 (9.20)

Setting n to n + N in (9.18), we have

zn+ N = yn+ N e j2pε(n+ N )/N

= yn e j2pεn/N e j2pε (9.21)
j2pε
= zn e

where we have used (9.18) and (9.20). Both zn+N and zn are the known received
samples. Therefore, the CFO ò can be estimated based upon (9.21).
The estimate can be inaccurate in noise if just one sample is used. A
better way is to form a cross-correlation between two consecutive segments.
We then have the following:
n= N −1 n= N −1
∑ zn∗ zn+ N = ∑ zn∗ zn e j2pε
n=0 n=0
(9.22)
⎛ n= N −1 2 ⎞
= ⎜ ∑ zn ⎟ e j2pε
⎝ n=0 ⎠

6899_Book.indb 161 10/22/19 4:55 PM

 162 Introduction to OFDM Receiver Design and Simulation

From (9.22), the CFO ò is estimated as follows [1, 4]:

ε=
1
2p
! (∑ n= N −1 ∗
n=0
zn zn+ N ) where !(z) = tan −1 ⎛ Im(z) ⎞
⎜⎝ Re(z) ⎟⎠ (9.23)

In (9.23), we have assumed that the beginning sample index is 0. Also,

(9.23) can be used only if the frequency offset is less than the subcarrier spacing.
In (9.23), ò is estimated using just one segment of the preamble. Using a
concept similar to BLUE [5], a better estimate can be obtained using several
such segments. For the mth and lth segment, we have from (9.18) the following:

zn+mN = yn+mN e j2pε(n+mN )/N n = 0,…, N − 1 (9.24)

zn+lN = yn+lN e j2pε(n+lN )/N n = 0,…, N − 1 (9.25)

Because of the repetitive structure, the following is true:

yn+lN = yn+mN n = 0,…, N − 1 (9.26)

Using (9.26), (9.25) can then be further simplified as follows:

zn+lN = ⎡⎣ yn+mN e j2pε(n+mN )/N ⎤⎦ e j2pε(l −m)N /N

(9.27)
= zn+mN e j2pε(l −m) n = 0,…, N − 1

Performing the cross-correlation between the lth and mth segment,

we have

n= N −1
⎡ n= N −1 2⎤
∑ *
zn+mN zn+lN = ⎢ ∑ zn+mN ⎥ e j2pε(l −m) (9.28)
n=0 ⎣ n=0 ⎦

where we have used (9.27). From (9.28), the frequency offset φ is estimated
as follows:

j(m,l) =
1
2p(l − m)
! (∑ n= N −1 ∗
n=0
zn+mN zn+lN ) (9.29)

6899_Book.indb 162 10/22/19 4:55 PM

 Synchronization163

To reduce the number of computations, any two consecutive segments

are used for averaging. In this casem l = m + 1 and we have

j(m,m + 1) =
1
2p
! (∑ n= N −1 ∗
n=0
zn+mN zn+(m+1)N ) (9.30)

Assuming that there are L segments, the estimate of ò then becomes

the following:

1 L−2
ε= ∑ j(m,m + 1)
L − 1 m=0
(9.31)

9.3.2 Simulation Examples

The simulation has been performed to understand the effect of SNR on the
accuracy of the frequency offset estimate. Except for random noise, CFO,
and DC offset, it is assumed that there are no other channel degradations and
synchronization issues in the following examples. Both the short sequence and
the long sequence from the IEEE 802.11a are used for simulation at 20-MHz
channel spacing.

Example 9.2
In this example, the effect of SNR on the accuracy of the estimate of frequency
offset is investigated. No DC offset is added; however, three different SNR
levels at 37.8 dB, 7.8 dB, and 2.8 dB are used for simulation. From (9.17),
the frequency offset is normalized to the subcarrier spacing and is set to 0.1.
The short sequence is used for simulation and each segment has 16 samples.
Equation (9.31) is used for averaging and L = 9. Only coarse detection is
applied and there is no fine detection. Figure 9.3 is a plot of ò against m in
(9.30). At SNR = 37.8 dB, the estimate is almost flat at ò = 0.1. As the noise
level increases, the estimate also fluctuates. The fluctuation is much higher as
SNR decreases from 7.8 dB to 2.8 dB. At SNR = 37.8 dB and 7.8 dB, there
is no sample misalignment. At SNR = 2.8 dB, there is one sample misalign-
ment. Therefore, at SNR = 2.8 dB, there are two factors contributing to the
estimate inaccuracy. One is the noise and the other is the sample misalign-
ment. Even though the individual estimate for a segment may be inaccurate

6899_Book.indb 163 10/22/19 4:55 PM

 164 Introduction to OFDM Receiver Design and Simulation

Figure 9.3 Effect of SNR on the coarse frequency offset estimate.

Table 9.3
Effect of Averaging on the Frequency Offset Estimate

SNR = 37.8 dB SNR = 7.8 dB SNR = 2.8 dB

Average ò 0.10005 0.10089 0.1022

as SNR drops, the average can have significant improvement. After averaging
by using (9.31), Table 9.3 shows the results.
It is clear from Table 9.3 that averaging does improve the estimate. Even
at SNR = 2.8 dB, the estimate is close.

Example 9.3
This example is a continuation of the previous example using the short
sequence. In this example, the impact of detection accuracy on the estimate
of ò is investigated. Figure 9.4 shows the plot of ò against m at SNR = 2.8
dB. The gray curve uses coarse detection only, while the dark curve follows
the coarse detection with another fine detection. After fine detection, it is
found that there is no sample misalignment. In other words, the only factor
contributing to the estimate inaccuracy is random noise after fine detection.
Even though the individual estimate may still show a large swing, the aver-
age estimate improves from ò = 0.1022 using coarse detection to ò = 0.1008

6899_Book.indb 164 10/22/19 4:55 PM

 Synchronization165

Figure 9.4 Effect of detection accuracy on the frequency offset estimate.

after fine detection. The estimate after fine detection is also the fine estimate
of frequency offset.

Example 9.4
In the previous examples, the estimate is done using the short sequence. How-
ever, the long sequence can also be used for the fine estimate of frequency
offset. Fine detection using the long sequence can still be performed after the
coarse detection using the short sequence. Immediately after fine detection,
the fine estimate of ò is continued. Table 9.4 shows the results.
In Table 9.4, the ò deviation percentage is the computation of the estimate
deviation from the true value and is computed as (ò − 0.1)/0.1. Quite clearly,
the short sequence is a better choice. This is because the short sequence has
10 segments, while the long sequence has only two segments, even though

Table 9.4
Comparison of ò Estimate Using Short and Long Sequence

ò Deviation ò Deviation
ò (SNR = ò (SNR = Percentage Percentage
37.8 dB) 7.8 dB) (SNR = 37.8 dB) (SNR = 7.8 dB)

Long 0.0998 0.0941 −0.20% −5.90%

Short 0.100052 0.10089 0.05% 0.89%

6899_Book.indb 165 10/22/19 4:55 PM

 166 Introduction to OFDM Receiver Design and Simulation

the segment size of the long sequence (64 samples) is larger than that of the
short sequence (16 samples).

Example 9.5
In this example, a range of the SNR values are scanned to find the effect of
frequency offset on the detection accuracy. The SNR ranges from −3.1 dB
to 37.8 dB and the frequency offset ò is set at 0.1. Only the short sequence is
used for both the coarse detection and fine detection. Figure 9.5 is a plot of ò
against the SNR under these two conditions. For the coarse detection, there is
no sample error for the SNR above 5.8 dB, 1 sample error for the SNR from
1.81 dB to 4.8 dB, 5 sample errors for the SNR from −1.18 dB to 0.8 dB, and
6 sample errors for the SNR at −2.18 dB. In other words, for the SNR above
5.8 dB, only noise affects the estimation accuracy of ò. However, for the SNR
below 4.8 dB, both the SNR and detection have the impacts. The sudden
transition on the curve is due to the change on the number of sample errors.
The fine detection follows immediately after the coarse detection. After
the fine detection, there is no sample detection error for the SNR above −1.18
dB. However, there is only one sample error for the SNR at −2.18 dB and that
is why the curve has a sudden transition. For the SNR above −1.18 dB, only
noise affects the estimation accuracy of ò.

Figure 9.5 Effect of both SNR and detection accuracy on the frequency offset
estimate.

6899_Book.indb 166 10/22/19 4:55 PM

 Synchronization167

Figure 9.6 Effect of DC bias on the frequency offset estimate.

Example 9.6
In the previous examples, no DC bias is applied in the simulation. Since DC
bias can also be considered as a periodic function, its presence can definitely
affect the accuracy in estimating the frequency offset. To numerically show
the impact of the DC bias, Figure 9.6 is the plot of two curves using the short
sequence. The top curve is the true frequency offset and is the same for all of
the DC bias. The bottom curve shows the estimate of the frequency offset.
The SNR and frequency offset ò are set at 37.8 dB and 0.1, respectively. As the

Table 9.5
The Frequency Offset Estimate After Highpass Filtering

DC Offset Percentage SNR = 37.8 dB SNR = 2.8 dB

0 0.1001 0.1007
17.7% 0.1001 0.1007
35.3% 0.1001 0.1007
53% 0.1001 0.1007

6899_Book.indb 167 10/22/19 4:55 PM

 168 Introduction to OFDM Receiver Design and Simulation

DC offset percentage increases, the estimate accuracy also drops. This example
indicates DC bias has to be removed in order to estimate the frequency offset.

Example 9.7
This example is a continuation of Example 9.6. The goal is to find the impact
of applying highpass filtering. The result is shown in Table 9.5.
Quite clearly, the frequency offset can be correctly determined after
highpass filtering to remove the DC offset. Since the DC offset is completely
removed, it has no impact on the estimate accuracy.

9.4 Frame Timing Offset

The frame boundary of the received OFDM symbol can be different from its
true location. Figure 9.7 is a timing diagram showing this situation. Between
two consecutive OFDM symbols, a cyclic prefix is inserted. The OFDM sym-
bol has N samples while the cyclic prefix has Ng samples. Assuming that the
duration of the channel response lasts L samples, then D = Ng − L.
Figure 9.7 also shows two windows of an OFDM symbol. One window
is correctly aligned, while the other window is misaligned by δ samples. As
long as δ is less than D, there is no ISI [6]. The effect is just a time delay of
the received waveform. Since the delay is given by δ samples, the received
preamble signal, zn, can be written as
zn = yn−d (9.32)

Taking the discrete Fourier transform on both sides of (9.32) or from

(1.47), we have

Figure 9.7 Timing diagram of frame synchronization.

6899_Book.indb 168 10/22/19 4:55 PM

 Synchronization169

Zk = Yk e − j2pkd/N (9.33)

where Zk and Yk are discrete Fourier transforms of zn and yn, respectively.

Therefore, the impact of symbol timing error is equivalent to a phase shift in
the frequency domain. However, if δ is within L, the ISI occurs. In this case,
the received preamble suffers amplitude attenuation in addition to a phase
shift [6].
Looking at this from a different angle, the timing delay can also be con-
sidered as the channel response from the multipath effect. It can be removed
by first estimating the channel response. This channel equalization technique
will be discussed in the next chapter.

9.5 SCO
Besides CFO and frame timing offset, the receiver may also suffer from SCO.
This is because the receiver clock and the transmitter clock may not be aligned.
This translates to a slightly different sampling period between the transmitter
and the receiver. Assume that the transmitter sampling period is T, the receiver
sampling period is T′, and the deviation ratio is δ . Then T and T′ are related
by the following equation

T ′ = T + dT = T (1 + d) (9.34)

Suppose δ is 100 PPM or 10 −4; then there is one sample misalignment

after 10,000 sampling periods. If the sampling clock is 20 MHz, one sample
is misaligned after 0.5 ms. This shows that the SCO can be significant in the
receiver performance.

9.5.1 Algorithm Development

Let yn and Yk represent the time domain and the corresponding frequency
domain samples, respectively. Assuming no CFO, they are related by the fol-
lowing FFT for the lth symbol [7, 8]:

1 k= N −1 j2pk[tn −( N g +lN s )T ]/NT

N ∑
yn = Yk e (9.35)
k=0

6899_Book.indb 169 10/22/19 4:55 PM

 170 Introduction to OFDM Receiver Design and Simulation

Figure 9.8 Frame and sample relationship.

where Ng is the number of guard samples and Ns is given here:

Ns = N g + N (9.36)

Figure 9.8 shows the relationship between the frames and samples. The
time tn is the time of the nth sample in the lth frame. If there is no SCO, tn
can be written as

(
tn = N g + lN s + n T ) (9.37)

Substituting (9.37) into (9.35), the following familiar inverse FFT equa-
tion results:
k= N −1
⎛ 1⎞
yn = ⎜ ⎟ ∑ Yk e j2pkn/N (9.38)
⎝ N ⎠ k=0

If there is a CFO of ò, it results in a shift of the frequency index k and

(9.35) becomes
j2pεtn k= N −1
⎛ 1⎞
⎜⎝ N ⎟⎠ ∑ Yk,l e
j2pk[tn −( N g +lN s )T ]/NT
zn,l = e NT (9.39)
k=0

In (9.39), we replaced yn with zn,l and Yk with Yk,l to show the spectral
response due to the lth symbol. In the receiver, the sampling clock has a period
of T′ and (9.37) becomes

( )
tn = N g + lN s + n (1 + d)T (9.40)

6899_Book.indb 170 10/22/19 4:55 PM

 Synchronization171

Substituting (9.40) into (9.39), we have

(
j2pε N g +lN s +n (1+d) ) ⎛ 1⎞
k= N −1
⎜⎝ N ⎟⎠ ∑ Yk,l e
j2pk[d( N g +lN s )+n(1+d)]/N
zn,l = e N (9.41)
k=0

Taking the FFT of zn,l generates the frequency response of Zk,l

n= N −1
Zk,l = ∑ zn,l e − j2pkn/N (9.42)
n=0

Substituting (9.41) into (9.42),

q= N −1
1
∑
[ j2pε( N g +lN s )(1+d)]/N j2pq( N g +lN s )d/N
Zk,l = Yq,l e e
N q=0
n= N −1
(9.43)
[ j2pn( q−k+qd+(1+d)ε )]/N
∑ e
n=0

Equation (9.43) can be rewritten in the following form:

n= N −1
1
∑ e[ j2pn( kd+(1+d)ε)]/N + ICI
[ j2pε( N g +lN s )(1+d)]/N [ j2pk( N g +lN s )d]/N
Zk,l = Yk,l e e
N n=0

(9.44)

where intercarrier interference (ICI) is given by

q= N −1
1
∑
[ j2pε( N g +lN s )(1+d)]/N [ j2pq( N g +lN s )d]/N
ICI = Yq,l e e
N q=0,q≠k
n= N −1
(9.45)
[ j2pn( q−k+qd+(1+d)ε )]/N
∑ e
n=0

The phase term inside the summation of (9.44) is very small and the
following approximation is made:

e j2pn( kd+(1+d)ε)/N ≈ 1 (9.46)

Substituting (9.46) into (9.44) and neglecting ICI, we have

6899_Book.indb 171 10/22/19 4:55 PM

 172 Introduction to OFDM Receiver Design and Simulation

[ j2pε( N g +lN s )(1+d)]/N [ j2pk( N g +lN s )d]/N

Zk,l = Yk,l e e (9.47)

Assume the phase of Zk,l is θ and the phase of Yk,l is φ , the following
is then obtained:

⎛ 1⎞
⎝N⎠ ( )
⎛ 1⎞
qk,l = jk,l + ⎜ ⎟ 2pε N g + lN s (1 + d) + ⎜ ⎟ 2pk N g + lN s d
⎝N⎠ ( ) (9.48)

For the symbol l + 1, (9.48) then becomes

⎛ 1⎞
(
qk,l +1 = jk,l +1 + ⎜ ⎟ 2pε N g + lN s + N s (1 + d)
⎝N⎠ )
(9.49)
⎛ 1⎞
(
+ ⎜ ⎟ 2pk N g + lN s + N s d
⎝N⎠ )
Subtracting (9.48) from (9.49), we have

2pεN s (1 + d) 2pkdN s
qk,l +1 − qk,l = jk,l +1 − jk,l + + (9.50)
N N

Since the product of ò and δ is small, (9.50) can be further simplified to be

2pεN s 2pkdN s
qk,l +1 − qk,l = jk,l +1 − jk,l + + (9.51)
N N

The phase difference between two consecutive symbols can be given

as below:
Δqk = qk,l +1 − qk,l (9.52)

Δjk = jk,l +1 − jk,l (9.53)

Using (9.52) and (9.53), (9.51) then becomes

2pεN s 2pkdN s
Δqk = Δjk + + (9.54)
N N

6899_Book.indb 172 10/22/19 4:55 PM

 Synchronization173

Table 9.6
Pilot Phases of the First 5 OFDM Symbols from IEEE 802.11a

Symbol Number k = −21 k = −7 k=7 k = 21

1 1 1 1 −1
2 1 1 1 −1
3 1 1 1 −1
4 −1 −1 −1 1
5 −1 −1 −1 1

For the CFO, the phase shift is the same for all of the subcarriers.
However, for SCO, the phase increases linearly with respect to the subcarrier
frequency index.

9.5.2 Lease Square Estimation in the Tracking Mode

The OFDM symbols are transmitted during the tracking mode. Usually in
the frequency domain, there are a set of pilot subcarriers. The phases of these
pilot subcarriers are known. At a given frequency index, two consecutive
symbols may not have the same phase. For the IEEE 802.11a, each OFDM
symbol has four pilot subcarriers at frequency indexes of −21, −7, 7, and 21.
Table 9.6 shows the phases of these four pilot subcarriers in the first five data
symbols. From symbol 3 to symbol 4, there is a sign change. This means ∆φ k
is not necessarily 0, but it is a known quantity. Moving ∆φ k to the left in
(9.54), we have
2pεN s 2pkdN s
Δ∅k = + (9.55)
N N

where ∆∅k is defined here:

Δ∅k = Δqk − Δjk (9.56)

Before trying to solve (9.55), it is rewritten as here:

N
vk= Δ∅k = dk + ε (9.57)
2pN s

Since pilot locations are not continuous, let kj (j = 1, M) represent the

pilot locations. For the IEEE 802.11a, kj takes the four possible values −21,

6899_Book.indb 173 10/22/19 4:55 PM

 174 Introduction to OFDM Receiver Design and Simulation

−7, 7, and 21. Equation (9.57) has only two unknowns, but M data points
and it can be solved using the method of least squares.
For each data point, the error is given here:
e j = vk − dk j − ε j = 1, M (9.58)
j

Summing up the square of all errors, we have

( )
M M 2
E = ∑ e 2j = ∑ vk − dk j − ε (9.59)
j
j=1 j=1

To find the minimum E, the two derivatives of E against ò and δ are

set to 0. We then have

⎛M ⎞ M
Mε + ⎜ ∑ k j ⎟ d = ∑ vk (9.60)
⎝ j=1 ⎠ j=1
j

⎛M ⎞ ⎛M ⎞ M
⎜ ∑ k j ⎟ ε + ⎜ ∑ k j ⎟ d = ∑ vk j k j (9.61)
2

⎝ j=1 ⎠ ⎝ j=1 ⎠ j=1

Equations (9.60) and (9.61) have two unknowns and can be solved to give

ε=
(∑ v )(∑ k ) − (∑ k )(∑
M
j=1 k j
M 2
j=1 j
M
j=1 j
M
v kj
j=1 k j ) (9.62)
J

d=
M (∑ M
v kj −
j=1 k j ) (∑ v )(∑ k ) M
j=1 k j
M
j=1 j
(9.63)
J

where J is given by
2
⎛M ⎞ ⎛M ⎞
J = M ⎜ ∑ k2j ⎟ − ⎜ ∑ k j ⎟ (9.64)
⎝ j=1 ⎠ ⎝ j=1 ⎠

Substituting vkj = (N/2π Ns)Δ∅kj into (9.62) and (9.63), we have

6899_Book.indb 174 10/22/19 4:55 PM

 Synchronization175

⎛ N ⎞
ε=⎜
(∑ M
j=1
Δ∅k
j
)(∑ k ) − (∑ k )(∑
M 2
j=1 j
M
j=1 j
M
j=1
Δ∅k k j
j
) (9.65)
⎝ 2pN s ⎟⎠ J

⎛ N ⎞M
d=⎜
(∑ M
j=1 j
) (∑
Δ∅k k j −
M
j=1
Δ∅k
j
)(∑ k ) M
j=1 j
(9.66)
⎝ 2pN s ⎟⎠ J

The definition of Ns and Δ∅kj are rewritten here for quick reference
Δ∅k = Δqk − Δjk j = 1, M (9.67)
j j j

Ns = N g + N (9.68)

Equations from (9.65) to (9.66) are then joint estimations of both CFO
and SCO in the tracking mode using the received data symbols.

9.5.3 Estimation in the Acquisition Mode

Previously, pilot subcarriers in the received data symbols were used to derive
a joint estimate of the CFO and SCO. Similar formulas can also be used for
their estimates in the acquisition mode. In the acquisition mode, the repeti-
tive acquisition sequence can be used.
Assume that there are L such repetitive segments. For the short sequence
in the IEEE 802.11a, there are L = 10 segments and each segment has 16
samples. Each segment repeats itself and can be considered similar to one
OFDM symbol. There are two major differences. The first is that there is no
guard interval and the second is a different signal transmission format.
Equations (9.65) and (9.66) can still be used with two modifications.
The first is Ng = 0 and Ns = N. For the IEEE 802.11a, each segment has M
= 12 subcarriers and each subcarrier can be considered as pilot subcarriers.
Since the segment repeats, two consecutive segments have the same phases at
each subcarrier. Therefore, the second modification is ∆φ k = 0. Incorporating
these two modifications, the following equations result:

⎛ 1 ⎞
ε=⎜ ⎟
(∑ M
j=1
Δ∅k
j
)(∑ k ) − (∑ k )(∑
M 2
j=1 j
M
j=1 j
M
j=1
Δ∅k k j
j
) (9.69)
   ⎝ 2p ⎠ J

6899_Book.indb 175 10/22/19 4:55 PM

 176 Introduction to OFDM Receiver Design and Simulation

⎛ 1 ⎞
d=⎜ ⎟
M (∑ M
j=1
∆∅k k j −
j
) (∑ M
j=1
Δ∅k
j
)(∑ k )
M
j=1 j
(9.70)
⎝ 2p ⎠ J

Δ∅k = Δqk j = 1, M (9.71)

j j

Ns = N (9.72)

2
⎛M ⎞ ⎛M ⎞
J = M ⎜ ∑ k2j ⎟ − ⎜ ∑ k j ⎟ (9.73)
⎝ j=1 ⎠ ⎝ j=1 ⎠

As we have shown previously, the CFO can be estimated independently.

Assuming that the CFO is known already, then we have from (9.57):

1
Δq − ε = dk j (9.74)
2p k j

Summing across all subcarriers, (9.74) becomes

⎛M ⎞ 1 ⎛M ⎞
d⎜ ∑kj ⎟ = ⎜ ∑ Δqk ⎟ − Mε (9.75)
⎝ j=1 ⎠ 2p ⎝ j=1 j
⎠

Solving for δ in (9.75), we obtain

d=
1
2p (∑ Δq ) − Mε
M
j=1 kj
(9.76)
(∑ k ) M
j=1 j

For the long sequence used in the IEEE 802.11a, there are two consecu-
tive symbols with no guard intervals in between. These two symbols are identi-
cal to each other. Therefore, the same equations applied to the short sequence
are also valid. The only differences are the parameter values, which are L and
M. The parameter L is 2 because there are only two symbols. The number of
subcarriers in each symbol including DC is 53. Excluding DC, then M = 52.

6899_Book.indb 176 10/22/19 4:55 PM

 Synchronization177

9.5.4 Estimation Under Special Conditions

More insight to the estimation can be obtained under two special conditions.
The first is when there is no SCO and the second is when there is no CFO.
Under both cases, the estimation is significantly simplified.
Assume that first there is no SCO and δ = 0. Setting δ = 0 in (9.66),
we have

⎛M ⎞
⎜ ∑ Δ∅k j k j ⎟ =
(∑ M
j=1
Δ∅k
j
)(∑ k ) M
j=1 j
(9.77)
⎝ j=1 ⎠ M

Substituting (9.77) into (9.65), the CFO becomes

ε=
⎛ N ⎞
⎜⎝ 2pN ⎟⎠
s
(∑ M
j=1
Δ∅k
j
) (9.78)
M

Δ∅k = Δqk − Δjk j = 1,…, M (9.79)

j j j

Ns = N g + N (9.80)

Equation (9.78) is the frequency-domain estimate of CFO in the track-

ing mode. At any given subcarrier pilot index kj, it depends upon the phase
difference between two consecutive OFDM symbols. It does not depend upon
a particular subcarrier pilot index. The average across all possible pilot subcar-
rier indexes minimizes the error.
In the acquisition mode, Ns = N and Δφ j = 0 and the following equa-
tion results:

ε=
(∑ M
j=1
Δθ k
j
) (9.81)
(2pM )

In this case, Δθ kj is just the phase difference between two consecutive

segments in the repetitive acquisition sequence for a given subcarrier index kj.
It is very similar to the frequency-domain estimate given in [4]. The average
across all possible subcarrier indexes improves the estimate.

6899_Book.indb 177 10/22/19 4:55 PM

 178 Introduction to OFDM Receiver Design and Simulation

Assume next that there is no CFO. By setting ò = 0 in (9.65), we then have

M
∑ Δ∅ k =
(∑ M
j=1
Δ∅k)(∑ k )
j
M 2
j=1 j
(9.82)
j=1
kj j
(∑ k ) M
j=1 j

Substituting (9.82) into (9.66), the following equation is obtained:

⎛ N ⎞
d=⎜
(∑ Δ∅ )
M
j=1 kj

(∑ k )
(9.83)
⎝ 2pN s ⎟⎠ M
j=1 j

Equation (9.82) can also be written in the following form:

M
∑ Δ∅k =
(∑ M
j=1 )(∑ k )
Δ∅k k j
j
M
j=1 j
(9.84)
j=1
j

(∑ k ) M 2
j=1 j

Substituting (9.84) into (9.66), δ then becomes

⎛ N ⎞
d=⎜
(∑ Δ∅ k )
M
j=1 kj j
(9.85)
⎝ 2pN s ⎟⎠
(∑ k ) M 2
j=1 j

Both (9.78) and (9.85) are similar to those given in [1].

Equation (9.83) is the frequency-domain estimate of δ in the tracking
mode. Another direct way of getting this formula is by setting ò = 0 in (9.57).
It is seen from there that the phase difference between two consecutive OFDM
symbols is linearly proportional to the pilot subcarrier index. Equation (9.83)
just finds the average slope to improve the estimate.
Again, in the acquisition mode, Ns = N and Δφ kj = 0. Equations (9.83)
and (9.85) are further simplified to give the following

⎛ 1 ⎞
d=⎜ ⎟
(∑ Δq )M
j=1 kj
(9.86)
⎝ 2p ⎠
(∑ k ) M
j=1 j

6899_Book.indb 178 10/22/19 4:55 PM

 Synchronization179

or

⎛ 1 ⎞
d=⎜ ⎟
(∑ Δq k )
M
j=1 kj j

(∑ k )
(9.87)
⎝ 2p ⎠ M 2
j=1 j

If it is known in advance that either CFO = 0 or SCO = 0, the estimate

can be greatly simplified. If there is no such knowledge, then both assumptions
are not true and the original equations given earlier should be used.

9.5.5 Simulation Examples

The short sequence defined in the IEEE 802.11a is used for the simulation. It
is assumed that there are no channel degradations and IQ imbalances. How-
ever, there is possible CFO, SCO, and background noise.

Example 9.8
Assume that the CFO and SCO are ò = 0.05 and δ = 0.001, respectively. The
joint estimate is applied at SNR = 37.8 dB. Assume also that there is no DC
bias. Because of the SCO, the received samples are actually sampled at nT +
nδ T where n is an integer. These time-shifted samples are obtained through
interpolation. The DFT is then applied to find the phase at each subcarrier

Figure 9.9 A plot of phase difference against subcarrier index.

6899_Book.indb 179 10/22/19 4:55 PM

 180 Introduction to OFDM Receiver Design and Simulation

index kj. The phase difference given by (9.52) is subsequently evaluated between
two consecutive segments l and l + 1 where l ranges from 2 to 8. For a short
sequence, there are 12 subcarriers from 1 to 6 and from 10 to 15. Figure 9.9 is
a scatter plot of phase difference against subcarrier index. Also superimposed
on the figure is an ideal curve. From (9.74), it is given as
y = 0.001x + 0.05

where y = (1/2π )Δθ kj and x = kj. It can be seen that the data points are scat-
tered along the straight line of this ideal curve. This also means an average
from all two consecutive segments shall improve the estimate.

Example 9.9
Under the same conditions given in Example 9.8, (9.69) and (9.70) are then
applied to compute the CFO and SCO for the SNR from 7.8 dB to 37.8 dB.
First, they are computed for each two consecutive segments l and l + 1 where
l ranges from 2 to 8. A total of seven such values are generated. The average
CFO and SCO are subsequently computed and plotted in Figures 9.10 and
9.11. Both Figures 9.10 and 9.11 show that the estimate converges to the desired
value as the SNR increases. In Example 9.2, we showed the estimate of the
CFO without the SCO through time-domain analysis. It is clear that the

Figure 9.10 A plot of the CFO against the SNR.

6899_Book.indb 180 10/22/19 4:55 PM

 Synchronization181

Figure 9.11 A plot of the SCO against the SNR.

estimate accuracy of the CFO goes down if there is an SCO. This is because
there are approximations involved during the development of the joint estimate.

Example 9.10
Assume that there is no CFO. Then both (9.86) and (9.87) can be used to
compute the SCO. Figure 9.12 is a plot of both against the SNR. The dark

Figure 9.12 Comparisons of the SCO against the SNR.

6899_Book.indb 181 10/22/19 4:55 PM

 182 Introduction to OFDM Receiver Design and Simulation

curve is a plot using (9.86) while the gray curve is a plot using (9.87). All the
other conditions are the same as in previous examples. As can be seen from
the plot, the difference from both computations is small. However, the com-
putation from (9.87) is slightly better.

9.6 IQ Imbalance
As was shown in Section 8.2, the in-phase and quadrature-phase components
need to be extracted by mixing the received signal with two locally generated
sinusoidal signals. Under ideal conditions, these two signals differ in phase
by 90°. In reality, the circuit parts may not be perfect and there are phase and
amplitude disturbances. The IQ imbalance [9−11] refers to the mismatch of
these two LOs.
Because of the mismatch, the channel estimate is no longer accurate
and the system performances can be affected significantly. Using the train-
ing samples in the acquisition mode, the compensation scheme is discussed
in the following sections.

9.6.1 IQ Model
Assume the received complex amplitude to be y(t), the transmitted complex
amplitude to be x(t) and the channel response to be h(t), then the following
equations result from (8.7) to (8.9):

y(t) = yI (t) + jyQ (t) (9.88)

x(t) = x I (t) + jx Q (t) (9.89)

(
r(t) = Re y(t)e j2pf c t ) (9.90)
( ) (
= yI (t)cos 2pf c t − yQ (t)sin 2pf c t )
yI (t) = ∫ x I (t)h(t − t) dt (9.91)

yQ (t) = ∫ x Q (t)h(t − t) dt (9.92)

6899_Book.indb 182 10/22/19 4:55 PM

 Synchronization183

Figure 9.13 Direct conversion receiver with IQ imbalance.

where r(t) is the received signal and fc is the carrier frequency

Figure 9.13 shows a direct conversion receiver in which the two LOs
are not a perfect match in both amplitude and phase. Assume that the phase
deviation is given by δ and the amplitude deviation is given by ε . The two
LOs can be written as

(
l1(t) = 2cos 2pf c t ) (9.93)

(
l2 (t) = −2(1 − e)sin 2pf c t + d ) (9.94)

(
= −2asin 2pf c t + d ) (9.95)

where we have defined α = 1 − ε . If δ = 0 and ε = 0, then the two LOs are

perfectly matched and Figure 9.13 reduces to Figure 8.1.
Multiplying (9.90) by (9.93) and following the same analysis as given in
Chapter 8, we recover yI (t) and

rI (t) = yI (t) (9.96)

Similarly, multiplying (9.90) by (9.95), we have rQ (t) in the lower branch

as given here:

( ( ) ( ) ( ) (
rQ (t) = −2a yI (t)sin 2pf c t + d cos 2pf c t − yQ (t)sin 2pf c t sin 2pf c (t) + d ))
( ( ) ( )
= a − yI (t)sin 4pf c t + d − yI (t)sin(d) − yQ (t)cos 4pf c (t) + d + yQ (t)cos(d) )
(9.97)

Again, rQ (t) in (9.97) after passing through an LPF becomes

6899_Book.indb 183 10/22/19 4:55 PM

 184 Introduction to OFDM Receiver Design and Simulation

rQ (t) = ayQ (t)cos(δ ) − ayI (t)sin(d) (9.98)

If both ε and δ are 0 and α = 1, then rQ (t) becomes yQ (t).

9.6.2 Estimation in the Acquisition Mode

Equation (9.96) and (9.98) can also be written in the following matrix form [11]:

⎡ rI (t) ⎤ ⎡ 1 0 ⎤ ⎡ yI (t) ⎤
⎢ r (t) ⎥ = ⎢ −a sin(d) a cos(d) ⎥ ⎢ yQ (t) ⎥ (9.99)
⎣ Q ⎦ ⎣ ⎦⎣ ⎦

After the analog-to-digital conversion, the ith digitized samples are

written as
rI ,i = yI ,i (9.100)

rQ ,i = −a sin(d) yI ,i + a cos(d) yQ ,i (9.101)

The amplitude imbalance α is estimated by first taking the square on

both sides of (9.101) and summing across all the acquisition samples. Assum-
ing that there are L such samples, we have
L−1 i= L−1 i= L−1
∑ rQ2 ,i = a 2 sin2 (d) ∑ yI2,i − 2a 2 sin(d)cos(d) ∑ yI ,i yQ ,i
i=0 i=0 i=0
L−1
(9.102)
+ a cos (d)∑
2 2
yQ2 ,i
i=0

During the acquisition mode, the training samples can be designed in

such a way that the following relationship is satisfied:
i= L−1 L−1
∑ yI2,i = ∑ yQ2 ,i (9.103)
i=0 i=0

Similarly, yI,i can be uncorrelated with yQ,i and the following relation-
ship can be made:

6899_Book.indb 184 10/22/19 4:55 PM

 Synchronization185

i= L−1
∑ yI ,i yQ ,i = 0 (9.104)
i=0

Both (9.103) and (9.104) will be justified in Example 9.11 using the short
sequence of the IEEE 802.11a.
Substituting (9.103) and (9.104) into (9.102), we have
L−1 i= L−1
∑ rQ2 ,i = a 2 ( sin2 (d) + cos2 (d)) ∑ yI2,i
i=0 i=0
(9.105)
i= L−1
=a 2
∑ yI2,i
i=0

Summing across all the L samples in (9.100), we have

L−1 i= L−1
∑ rI2,i = ∑ yI2,i (9.106)
i=0 i=0

Dividing (9.105) by (9.106), we have an estimate for the amplitude

imbalance α :

∑ i=0 rQ2 ,i
L−1

a= (9.107)
∑ i=0 rI2,i
L−1

From the definition of α , the amplitude deviation ε is then given by

e = 1− a (9.108)

∑ i=0 rQ2 ,i
L−1

= 1− (9.109)
∑ i=0 rI2,i
L−1

To estimate the phase imbalance δ , the cross-correlation of rI,i and rQ,i

are formed. From (9.100) and (9.101), the following is obtained:

6899_Book.indb 185 10/22/19 4:55 PM

 186 Introduction to OFDM Receiver Design and Simulation

L−1 i= L−1 i= L−1

∑ rI ,i rQ ,i = −a sin(d) ∑ yI2,i + a cos(d) ∑ yI ,i yQ ,i
i=0 i=0 i=0
(9.110)
i= L−1
= −a sin(d) ∑ yI2,i
i=0

where we have used the same approximation given by (9.104). From (9.110)
and (9.100), δ is solved to give

∑ rI ,i rQ ,i P
L−1

sin(d) = − i=0i= L−1 = − est (9.111)

a ∑ i=0 rI2,i a

where we have defined the factor Pest as follows:

∑ rI ,i rQ ,i
L−1

Pest = i=0L−1 (9.112)

∑ i=0 rI2,i
From (9.111), the solution of δ depends upon α . However, α is inde-
pendently determined from (9.107). Once α is computed, it is used in (9.111)
to solve for δ .

9.6.3 Compensation of IQ Imbalance

Once both the amplitude imbalance α and the phase imbalance δ are esti-
mated, they can be used for restoring the received IQ signals. Multiplying
(9.100) by α sin(δ ) and adding to (9.101), we have the following:

rQ ,i + a sin(d)rI ,i = a cos(d) yQi (9.113)

From (9.113), yQ,i is solved to give

rQ ,i + a sin(d)rI ,i
yQ ,i = (9.114)
a cos(d)

Substituting (9.111) into (9.114), the following equation results:

6899_Book.indb 186 10/22/19 4:55 PM

 Synchronization187

Table 9.7
The 12 Subcarriers of the Short Sequence in the IEEE 802.11a

Subcarrier Index k Subcarrier Index k Subcarrier

1 −1.472(1 + j) 5 1.472(1 + j) 12 1.472(1 + j)

2 −1.472(1 + j) 6 1.472(1 + j) 13 −1.472(1 + j)
3 1.472(1 + j) 10 1.472(1 + j) 14 −1.472(1 + j)
4 1.472(1 + j) 11 −1.472(1 + j) 15 1.472(1 + j)

rQ ,i − Pest rIi
yQ ,i = (9.115)
a 2 − Pest2

From (9.100), yI,i is obtained as

yI ,i = rI ,i (9.116)

Both (9.115) and (9.116) can be used to recover the original in-phase
and quadrature-phase samples.

9.6.4 Simulation Examples

The short sequence from acquisition defined in the IEEE 802.11a will be
used for simulation study. The background noise will be mixed to find the
performance of the imbalance estimate in noise. It is assumed that there are
no other degradations.

Example 9.11
There are two major assumptions given in (9.103) and (9.104). Both assump-
tions will be justified using the short sequence of the IEEE 802.11a. There are
12 subcarriers, which are given in Table 9.7.
After performing 16-point IFFT, the time-domain samples are further
divided by 4 for further normalization. The powers of the real part and the
imaginary part are computed to obtain

Real part power = Imaginary part power = 0.101568

Correlation = 0

6899_Book.indb 187 10/22/19 4:55 PM

 188 Introduction to OFDM Receiver Design and Simulation

Table 9.8
The Original and Impaired Short Sequence Subcarriers

YI YQ ZI ZQ

−1.00 −1.00 −1.04 −1.05

−1.00 −1.00 −1.06 −0.87
1.00 1.00 0.91 0.95
1.00 1.00 0.94 0.77
1.00 1.00 0.91 0.95
1.00 1.00 0.94 0.77
1.00 1.00 0.94 0.77
−1.00 −1.00 −1.04 −1.05
1.00 1.00 0.94 0.77
−1.00 −1.00 −1.04 −1.05
−1.00 −1.00 −1.06 −0.87
1.00 1.00 0.91 0.95

Figure 9.14 The constellation of the original and impaired short sequence.

6899_Book.indb 188 10/22/19 4:55 PM

 Synchronization189

Clearly, both (9.103) and (9.104) are satisfied

Example 9.12
Assume that the amplitude imbalance α = 0.9 (ε = 0.1) and the phase imbal-
ance δ = 5° = 0.08727 radian. First, perform the FFT of the short sequence
generated in Example 9.11 and then divide by the normalization factor 1.472
to obtain the original spectrum Yi + jYQ. Next, (9.100) and (9.101) are applied
to generate the time-domain samples with IQ imbalance. The FFT is again
applied to obtain the impaired spectrum Zi + jZQ. Table 9.8 lists all the sub-
carriers after dividing the spectrum by the normalization factor 1.472. As
Table 9.8 shows, the original spectrum has only two different subcarriers. It
is either 1 + j or −1 − j. After the application of IQ imbalance, the impaired
spectrum has now four different subcarriers. They are 0.91 + 0.95j, 0.94 +
0.77j, −1.04 − 1.05j, and −1.06 − 0.87j.
Figure 9.14 is a plot of the constellation of the short sequence with IQ
imbalance. Without IQ imbalance, there are only two spectral points. With
IQ imbalance, there are four modified spectral points. The total number of
unique spectral components then increases from 2 to 4. From Figure 9.14, all
these 6 spectral components are clearly seen. Therefore, the occurrence of IQ
imbalance can cause spectral distortion.

Figure 9.15 Estimate of phase angle imbalance in random noise.

6899_Book.indb 189 10/22/19 4:55 PM

 190 Introduction to OFDM Receiver Design and Simulation

Figure 9.16 Estimate of amplitude imbalance in random noise.

Example 9.13
Example 9.11 shows that both (9.103) and (9.104) are satisfied. This means
that the estimate accuracy does not depend upon the magnitude of the imbal-
ance. Assume first that the amplitude imbalance is fixed at α = 0.9. The phase
angle imbalance δ is allowed to change from 5° to 45° in 5° increments.

Figure 9.17 Estimate of amplitude imbalance against its true input.

6899_Book.indb 190 10/22/19 4:55 PM

 Synchronization191

Figure 9.18 Estimate of phase angle imbalance against amplitude imbalance.

The imbalanced samples are generated based upon (9.100) and (9.101). The
amplitude imbalance α is then estimated from (9.107) while the phase angle
imbalance is estimated from (9.111). Short sequence samples from the second
segment to the ninth segment are used for estimation. The total number of
samples is L = 128. The SNR of 37.8 dB, 27.8 dB, and 17.8 dB is used to
obtain the performance in random noise.
Figure 9.15 is a plot of estimated phase angle imbalance against its true
input. It can be seen that the estimate of phase angle imbalance is even better
at larger angles than with smaller angles. The straight-line curve demonstrates
the superiority of the estimate. As the SNR drops, the estimate accuracy also
drops because (9.103) and (9.104) are no longer valid.
Figure 9.16 is a plot of estimated amplitude imbalance in random noise.
The horizontal line shows that the estimated amplitude imbalance is inde-
pendent of the input phase angle imbalance. However, as the SNR drops,
the estimation accuracy also drops but is still not too far away from the true
input of 0.9.

Example 9.14
This example is similar to Example 9.13. However, the input phase angle
imbalance δ is fixed at 5° while the amplitude imbalance α is allowed to

6899_Book.indb 191 10/22/19 4:55 PM

 192 Introduction to OFDM Receiver Design and Simulation

vary. Just as before, the SNRs of 37.8 dB, 27.8 dB, and 17.8 dB are used
to get the performance of the estimation in noise. Figure 9.17 is a plot of
the estimated amplitude imbalance against the input amplitude imbalance.
Again, the straight-line curve shows that the estimate is excellent in noise. As
the SNR drops, the accuracy also drops slightly when the input amplitude
imbalance is large.
Figure 9.18 is a plot of phase angle imbalance against the input ampli-
tude imbalance. The horizontal line shows the estimate is independent of the
magnitude of the input amplitude imbalance. However, as the SNR drops,
the estimate accuracy also drops.

9.7 Summary
Five major synchronization issues in the receiver front end have been analyzed
in detail in this chapter. They are DC offset, CFO, SCO, sampling time offset,
and IQ imbalance.
The DC offset is caused by leakage from the receiver LO to the input of
either the amplifier or mixer. Its appearance can cause a false alarm through
the use of the normalized cross-correlation metric in the signal detection.
However, it can also be easily removed though highpass filtering after the
analog-to-digital conversion.
The CFO can be caused by the relative motion between the transmitter
and the receiver. Using the periodic nature of the preamble signal, this offset
can be estimated through correlation between periodic segments. The estimate
improves through the averaging of estimates from adjacent segments.
The sampling time offset can cause the frame boundary difference
between the transmitted and the received signal. This offset is simply a phase
shift in the frequency domain. The effect is similar to the signal delay in a
multipath channel and can be removed through channel estimate and adap-
tation discussed in the later chapters.
The SCO is caused by the difference in sampling time period between
the transmitter and receiver clocks. The effect can be analyzed through a
joint estimate of both SCO and CFO in the frequency domain from adja-
cent symbols. In the tracking mode or during data transmission, the known
phases of pilot subcarriers are used. In the acquisition modes or during the
preamble transmission, the periodic nature of adjacent segments and the zero
guard period can be used for simplification. The least square optimization is
used to estimate both SCO and CFO since there are more than two pilots in
the tracking mode and in the acquisition mode. Under the special condition
where either SCO or CFO is 0, a much simpler formula was derived.

6899_Book.indb 192 10/22/19 4:55 PM

 Synchronization193

The IQ imbalance is caused by defects of circuit parts during the direct

conversion process. This results in amplitude and phase disturbances in the
LOs of the in-phase and quadrature-phase components. Both can be estimated
in the acquisition mode much easier if the preamble waveform satisfies two
special conditions. The first is that the in-phase and quadrature-phase signals
have the same power. The second is that the in-phase and quadrature-phase
signals are not correlated. The short sequence specified in the IEEE 802.11a
meets these two conditions. After both phase and amplitude disturbances are
estimated, they can be compensated and the original in-phase and quadrature-
phase signals are restored.
In the next chapter, both block-type and comb-type pilot patterns for
the channel estimation are presented, and channel tracking using the LMS
algorithm is also covered.

References
[1] Chiueh, T. -D., P. Y. Tsai, and I. W. Lai, Baseband Receiver Design for Wireless MIMO-
OFDM Communications, New York: Wiley, 2012.
[2] Razavi, B., “Design Considerations for Direct-Conversion Receivers,” IEEE Transactions
on Circuits and Systems, Vol. 44, No. 6, June 1997, pp. 428−435.
[3] Oppenheim, A. V., and R. W. Schafer, Digital Signal Processing, Upper Saddle River,
NJ: Prentice Hall, 1975.
[4] Moose, P. H., “A Technique for Orthogonal Frequency Division Multiplexing Frequency
Offset Correction,” IEEE Transactions on Communications, Vol. 42, No. 10, 1994, pp.
2908−2914.
[5] Morelli, M., and U. Mangali, “An Improved Frequency Offset Estimator for OFDM
Applications,” IEEE Communication Letters, 1999.
[6] Speth, M., F. Classen, and H. Meyer, “Frame Synchronization of OFDM Systems in
Frequency Selective Fading Channels,” 1997 IEEE 47th Vehicular Technology Conference.
Technology in Motion, Vol. 3, pp. 1807−1811.
[7] Speth, M., S. Stefan, and A. Fechtel, “Optimum Receiver Design for Wireless Broad-
Band Systems Using OFDM-Part I,” IEEE Transactions on Communications, Vol. 47,
No. 11, November 1999, pp. 1668−1677.
[8] Liu, S. Y., and J. W. Chong, “A Study of Joint Tracking Algorithms of Carrier Frequency
Offset and Sampling Clock Offset for OFDM-Based WLANs,” IEEE International
Conference on Communications, Circuits and Systems, Vol. 1, June 2002.
[9] Tubbax, J., et al., “Compensation of Transmitter Imbalance for OFDM Systems,” 2004
IEEE International Conference on Acoustics, Speech, and Signal Processing (2004 ICASSP),
Vol. 2, 2004.

6899_Book.indb 193 10/22/19 4:55 PM

 194 Introduction to OFDM Receiver Design and Simulation

[10] Tseng, H. -Y., et al., “Compensation of Imbalance and DC Offset for OFDM
Transmission over Frequency Selective Channels,” 2008 IEEE International Conference
on Communications, 2008.
[11] Held, I., et al., “Low Complexity Digital IQ Imbalance Correction in OFDM WLAN
Receivers,” 2004 IEEE 59th Vehicular Technology Conference, VTC 2004-Spring, Vol.
2, pp. 1172−1176.

6899_Book.indb 194 10/22/19 4:55 PM

 10
Channel Estimation and Tracking

10.1 Introduction
In a radio channel, there are normally reflecting objects and scatters in the sig-
nal propagation path. These obstacles include mountains, buildings, and even
the ground. The directly transmitted signal is then reflected. The receiver may
then receive multiple versions of the original signal with a different amplitude,
phase, and arrival time. These multipath signals add together and may cause
significant distortion of the received signal.
For OFDM, the signal bandwidth is normally larger than the coherence
bandwidth of the channel. When this happens, the received signal suffers fre-
quency selective fading. In the time domain, this also means the delay spread
is longer than the reciprocal of the signal bandwidth. This is also why a guard
interval is added after each OFDM symbol to minimize the ISI.
The channel equalization technique for OFDM is quite simplified in the
frequency-selective channel. The channel estimation is normally performed
with the transmission of training symbols before the actual signal transmis-
sion. These training symbols are designed using subcarriers with a known
amplitude and phase to the receiver. Based upon the received amplitude and
phase of each subcarrier, the channel is easily estimated.
After the channel estimation, the original signal can be restored. If
the channel is steady and does not change fast over time, the use of training

195

6899_Book.indb 195 10/22/19 4:55 PM

 196 Introduction to OFDM Receiver Design and Simulation

symbols to aid the channel estimation in the data acquisition mode is normally
enough. However, the channel may not be steady and it is necessary to esti-
mate the channel frequently. The latter is done by inserting pilot subcarriers
besides the data subcarriers to aid in the estimation.
If the channel varies fast, it can also be tracked for each received data
symbol. Based upon the estimation error and the received signal, the chan-
nel at each data subcarrier is tracked using the method of steep descent. For
OFDM, only one tap is necessary and the process is also greatly simplified.
In the following sections, the possible pilot structures are introduced
first. For each pilot structure, the channel estimation is presented. After that,
the tracking algorithm is analyzed. The simulation examples are also given to
understand the effectiveness of each algorithm.

10.2 Pilot Patterns

For OFDM, the pilot patterns are usually given in a time-frequency diagram.
Both time and frequency are represented in a group of small circles. Each
circle represents an OFDM symbol in the time axis. In other words, a circle
along the time axis has the time duration of an equivalent OFDM symbol.
The same circle along the frequency axis has the equivalence of a subcarrier.
If a circle is filled, it carries pilot subcarriers.
Many different pilot patterns can be designated. However, the most
commonly used patterns are the block type and comb type. Both types are
given in the following sections.

10.2.1 Block-Type Pilot Pattern

Figure 10.1 shows a typical block type pilot pattern [1]. Along the time axis,
there is a filled node for every four empty nodes. Along the frequency axis, all
the nodes are filled at a given time period. In terms of OFDM terminology,
it means after for every four OFDM symbols, a symbol is transmitted with
all its subcarriers designated as pilot subcarriers.
Using the block type pattern, pilot subcarriers are periodically trans-
mitted for the new channel estimation. Since pilot subcarriers do not carry
data, the data transmission rate is reduced in exchange for better bit error
rate performance. The period of pilot transmission depends upon the system
requirement and the channel bandwidth.
Assuming that Tg is the guard interval, Ts is the symbol interval, Tbpsc is
the number of data bits per subcarrier, Nsd is the number of data subcarriers

6899_Book.indb 196 10/22/19 4:55 PM

 Channel Estimation and Tracking197

Figure 10.1 Block-type pilot pattern.

per symbol, and M is the data symbol count between pilot symbols, then the
data transmission rate R is given here:
MN bpsc N sdC
R= (10.1)
(
(M + 1) Tg + Ts )
where C is the coding rate. If there is no pilot transmission, then the follow-
ing equation results in:
N bpsc N sdC
R= (10.2)
(Tg + Ts )
Example 10.1
Assume that the modulation is QPSK and the channel spacing is 20 MHz.
From Table 3.1, Tg = 0.8 μs, Ts = 3.2 μs, Nbpsc = 2, Nsd = 48, and C = 1/2 for
the convolutional code. If there is no pilot transmission, the data rate is then
2 ∗ 48 ∗ 0.5
R= = 12 Mbps
3.2 + 0.8

Assuming that now that one pilot symbol is sent every 4 data symbols,
then the new data rate becomes

6899_Book.indb 197 10/22/19 4:55 PM

 198 Introduction to OFDM Receiver Design and Simulation

4 ∗ 2 ∗ 48 ∗ 0.5
R= = 9.6 Mbps
5∗ 4

With the transmission of a pilot symbol after 4 data symbols, the data
rate then reduces from 12 Mbps to 9.6 Mbps.
During the time between pilot transmissions, the existing channel esti-
mation is used. Therefore, the block-type pilot pattern is normally used when
the channel is not fast varying. However, it can still be slightly updated through
tracking as will be discussed later.

10.2.2 Comb-Type Pilot Pattern

Figure 10.2 shows a comb-type pilot pattern [2, 3]. In this case, pilot subcarri-
ers are sent during every data symbol transmission. However, inside each data
symbol, only certain subcarriers are designated as pilot subcarriers. In Figure
10.2, subcarriers 0, 4, and 8 are pilot subcarriers and the others are data sub-
carriers. These three pilot subcarriers exist during every symbol transmission.
Strictly speaking, the data rate is still reduced because the pilot subcarriers
for each data symbol do not carry data. The data rate is computed according
to (10.2). Since the pilot subcarriers have the known amplitude and phase
to the receiver, the channel can still be estimated at these pilot subcarriers.
The channel at the data subcarriers then needs to be estimated through other
means such as interpolation.

Figure 10.2 Comb-type pilot pattern.

6899_Book.indb 198 10/22/19 4:55 PM

 Channel Estimation and Tracking199

Since the channel is estimated at each data symbol, this type of pilot
pattern is suitable for a fast varying channel. If the number of pilot carriers is
not large enough, the accuracy of channel estimation at data subcarriers will
suffer. The IEEE 802.11a specifies four subcarriers as pilot signals. They can
be used for frequency offset and sampling clock offset estimation as discussed
in Chapter 9. If necessary, they can also be used for channel estimation.

10.3 Channel Estimation

There are two major types of pilot patterns. One is the block type and the
other is the comb type. The channel estimation for both types is given here.

10.3.1 Channel Estimation for the Block-Type Pilot Pattern

To send the OFDM signal, the frequency-domain subcarriers Xk is first trans-
formed to the time-domain samples xn following the inverse discrete Fou-
rier transform:
N −1
1
xn =
N ∑ X k e j2pkn/N (10.3)
k=0

After propagating through the multipath channel, the signal xn is cor-

rupted. Assume that the impulse response is hn. Then the received sample yn
is given by the following convolution of xn and xn:
N −1
yn = ∑ xi hn−i + wn (10.4)
i=0

where wn is an additive Gaussian noise. The noise type could be thermal

noise, interference noise, or others. To recover the transmitted subcarriers, the
time-domain samples yn are again transformed back to the frequency domain
following the discrete Fourier transform
N −1
Yk = ∑ yne − j2pnk/N (10.5)
n=0

Substituting (10.4) into (10.5), we have

N −1 N −1 N −1
Yk = ∑ ∑ xi hn−i e − j2pnk/N + ∑ wne − j2pnk/N (10.6)
n=0 i=0 n=0

6899_Book.indb 199 10/22/19 4:55 PM

 200 Introduction to OFDM Receiver Design and Simulation

The second term of (10.6) is simply Wk representing the kth noise sample
in the frequency domain. Assuming that the frequency response of the channel
is given by Hk, then it is related to hn by the following inverse Fourier transform
N −1
1
hi =
N ∑ H l e j2pli/N (10.7)
l =0

Replacing i by n − i in (10.7), we have

N −1
1
hn−i =
N ∑ H l e j2pl (n−i)/N (10.8)
l =0

Substituting (10.3) and (10.8) into (10.6), the following is obtained:

N −1 N −1
1 N −1 1 N −1
Yk = ∑ ∑ N ∑ X me j2pmi/N N ∑ H l e j2pl (n−i)/N e − j2pnk/N + Wk
n=0 i=0 m=0 l =0
(10.9)
N −1 N −1 N −1 N −1
⎡1 ⎤⎡ 1 ⎤
= ∑ ∑ XmHl ⎢ ∑ e j2pn(l −k)/N ⎥ ⎢ N ∑ e j2pi(m−l )/N ⎥ + Wk
m=0 l =0 ⎣N n=0 ⎦⎣ i=0 ⎦

However, the summation inside the bracket of (10.9) has the follow-
ing relationship:

{
N −1
N for r = jN , j an integer
∑ e − j2prk/N = 0 otherwise
(10.10)
k=0

Using (10.10), (10.9) then becomes

N −1 N −1
⎡1 ⎤
Yk = ∑ X l H l ⎢ N ∑ e j2pn(l −k)/N ⎥ + Wk
l =0 ⎣ n=0 ⎦ (10.11)
= X k H k + Wk

Equation (10.11) says that the discrete Fourier transform of the received
samples is the addition of noise spectrum and the product of the transmitted
data spectrum and channel spectrum. This relationship is true for every sub-
carrier index k. From (10.11), the channel response is given here:

Y!k
Hk = where Y!k = Yk − Wk (10.12a)
Xk

6899_Book.indb 200 10/22/19 4:55 PM

 Channel Estimation and Tracking201

If the random noise is negligible small or 0, then Hk is exactly recovered

as given here:
Yk
Hk = if Wk = 0 (10.12b)
Xk

Using the pilot subcarriers, Xk is known in the receiver. Therefore, from

the received frequency domain samples, Yk, the channel frequency response
Hk is then obtained.
For block-type pilot patterns, (10.12) gives the newest channel estimation
for every M data symbols. Based upon this estimation, the data subcarriers
can be recovered before any new channel estimation as given here:

Yk
Xk = (10.13)
Hk

10.3.2 Channel Estimation for Comb-Type Pilot Pattern

For the comb-type pilot pattern, there are only a finite number of pilot subcar-
riers. Even though the channel response can be estimated at the pilot subcar-
riers, there is no information available to estimate that at the data subcarriers.
One popular method is to interpolate at the data subcarriers based upon the
known channel response at the pilot subcarriers.
Depending upon the number of points used for interpolation, the inter-
polation can be linear, parabolic, and cubic. The number of points used is 2, 3,
and 4 for linear, parabolic, and cubic interpolators, respectively. The Lagrange
interpolation formula is used for each of these interpolators

10.3.2.1 Linear Interpolation

The easiest way is to use only two points for straight-line interpolation. The
Lagrange interpolation formula is given by the following equation:

H (k) =
( k − k1 ) H ( k ) + k − k0 H ( k ) (10.14)
( k0 − k1 ) 0 ( k1 − k0 ) 1
where k 0 and k1 are the two pilot subcarriers indexes and k is the data subcar-
rier index to be interpolated. The function values H(k 0) and H(k1) are channel
responses at the pilot subcarrier index k 0 and k1. From (10.14), it is clear H(k)
= H(k 0) when k = k 0 and H(k) = H(k1) when k = k1.

6899_Book.indb 201 10/22/19 4:55 PM

 202 Introduction to OFDM Receiver Design and Simulation

Assume that the distance between k 0 and k1 is D and k1 > k 0. Assuming

that k is larger than k 0, then k can be written as
k = k0 + mD (10.15)

k1 = k0 + D (10.16)

Substituting (10.15) and (10.16) into (10.14), we have [1]

( )
H (k) = (1 − m)H k0 + mH ( k1 ) (10.17a)

k = k0 + mD (10.17b)

Equation (10.17) is the linear interpolation at point k using the two

known data points at k 0 and k1. For k < k 0, μ < 0 and (10.17) still applies.

10.3.2.2 Parabolic Interpolation

If there are three data points available, then it is parabolic interpolation. Assum-
ing that these three points are k 0, k1, and k2, then the Lagrange interpolation
formula is given here:

H (k) =
( k − k1 )( k − k2 ) H ( k )
( k0 − k1 )( k0 − k2 ) 0
(10.18)
+
( k − k0 )( k − k2 ) H ( k ) + ( k − k0 )( k − k1 ) H ( k )
( k1 − k0 )( k1 − k2 ) 1 ( k2 − k0 )( k2 − k1 ) 2
Let the distance between consecutive data points be a constant such that
km+1 − km = D for m = 0, 1. Then we have the following:
k = k0 + mD (10.19)

k1 = k0 + D (10.20)

k2 = k0 + 2D (10.21)

Substituting (10.19), (10.20), and (10.21) into (10.18), we have [3]

( m − 1)( m − 2 ) H
H (k) =
2
( k0 ) + m(2 − m)H ( k1 ) + m(m2− 1) H ( k2 ) (10.22a)

6899_Book.indb 202 10/22/19 4:55 PM

 Channel Estimation and Tracking203

k = k0 + mD (10.22b)

Equation (10.22) is the second-order Lagrange interpolation using three

data points k 0, k1, and k2.

10.3.2.3 Cubic Interpolation

The interpolation order can go higher to use four data points, k 0, k1, k2, and
k3. The Lagrange interpolation formula is then given here:

( k − k1 )( k − k2 )( k − k3 ) H k + ( k − k0 )( k − k2 )( k − k3 ) H k
H (k) = ( ) ( )
( k0 − k1 )( k0 − k2 )( k0 − k3 ) 0 ( k1 − k0 )( k1 − k2 )( k1 − k3 ) 1
( k − k0 )( k − k1 )( k − k3 ) H k + ( k − k0 )( k − k1 )( k − k2 ) H k
( k2 − k0 )( k2 − k1 )( k2 − k3 ) 2 ( k3 − k0 )( k3 − k1 )( k3 − k2 ) ( 3 )
+ ( )
(10.23)

Again, the distance between consecutive data points is a constant D.

That means km+1 − km = D for m = 0, 1, and 2. However, we let the interpola-
tion data point k be computed starting from k1. We then have the following:
k = k1 + mD (10.24)

k0 = k1 − D (10.25)

k2 = k1 + D (10.26)

k3 = k1 + 2D (10.27)

Substituting from (10.24) to (10.27) into (10.23), the following is obtained:

( )
H (k) = C0 H k0 + C1H ( k1 ) + C2 H ( k2 ) + C3 H k3( ) (10.28a)

k = k1 + mD (10.28b)

where C 0, C1, C2, and C3 are given here [4]:

m(m − 1)(m − 2)
C0 = − (10.29a)
6

6899_Book.indb 203 10/22/19 4:55 PM

 204 Introduction to OFDM Receiver Design and Simulation

C1 =
( m2 − 1)(m − 2) (10.29b)
2
m(m + 1)(m − 2)
C2 = − (10.29c)
2

m ( m2 − 1)
C3 = (10.29d)
6
The choice of linear, parabolic, or cubic interpolation will certainly
affect interpolation accuracy. There are also three other factors that will have
significant impact on the accuracy. The first is the number of pilot subcarriers
available for interpolation. The second is the degree of channel spectral varia-
tions. If the channel frequency response has too many spectral cycles, then
more pilot subcarriers are needed for interpolation. The last is the effects due
to random noise. The dependence of interpolation accuracy on all of these
factors is illustrated in the following examples.

Example 10.2
To illustrate the dependence of interpolation accuracy on the number of pilot
subcarriers, channel spectral variations, and the interpolation method, a simple
multipath model is used. Assume that the receiver receives signals from two
major paths. One is the main signal xn and the other is the reflected signal
bxn-m where b is the reflected amplitude and m is the sample delay. The received
signal yn is then the sum of these two signals:
yn = xn + bxn−m (10.30)

Equation (10.30) is a special case of the more general formula given in

(12.11).
Taking DFT on both sides of (10.30), we have
Yk = X k + bX k e − j2pmk/N
(10.31)
(
= X k 1 + be − j2pmk/N )
From (10.31), the frequency channel response Hk is then given below:

H k = 1 + be − j2pmk/N
(10.32)
⎛ ⎛ 2pmk ⎞ ⎞ ⎛ 2pmk ⎞
= ⎜ 1 + bcos ⎜ ⎟ − jbsin ⎜
⎝ ⎝ N ⎠⎠ ⎟ ⎝ N ⎟⎠

6899_Book.indb 204 10/22/19 4:55 PM

 Channel Estimation and Tracking205

The real part Hkr and the imaginary part Hki of Hk are then written below:

⎛ 2pmk ⎞
H kr = 1 + bcos ⎜ (10.33)
⎝ N ⎟⎠

⎛ 2pmk ⎞
−H ki = bsin ⎜ (10.34)
⎝ N ⎟⎠

In this example, the channel frequency response is a sine wave. The period
is obtained by setting the cos argument to 2π . We then have
N
Period = (10.35)
m

Clearly, as the delay m increases, the period decreases and it becomes

more difficult to interpolate. This is illustrated in the next few examples.

Example 10.3
In this example, we illustrate how the number of pilot subcarriers may affect
the interpolation accuracy. Parabolic interpolation is used and the number of
pilot subcarriers is 8, 16, and 32. However, the total number of subcarriers is
set to 64. Assuming that m = 3, N = 64, and b = 0.9, Figures 10.3, 10.4, and
10.5 are the plots of both the original Hkr and −Hki/b and their interpolation
against the subcarrier index k.

Figure 10.3 Parabolic interpolation using eight pilot subcarriers.

6899_Book.indb 205 10/22/19 4:55 PM

 206 Introduction to OFDM Receiver Design and Simulation

Figure 10.4 Parabolic interpolation using 16 pilot subcarriers.

Figure 10.5 Parabolic interpolation using 32 pilot subcarriers.

6899_Book.indb 206 10/22/19 4:55 PM

 Channel Estimation and Tracking207

It is clear from the plots that the interpolation accuracy improves as

the number of pilot subcarriers M increases. When M = 32, the interpolated
curve is almost indistinguishable from the original. However, when M = 8,
the interpolation is not considered as being good enough. The period is 64/3
≅ 21 and is confirmed from the figure. Roughly, each cycle requires that at
least 4 pilot subcarriers and 12 pilot subcarriers are needed in this example.
This also shows when M = 16, the interpolation significantly improves.

Example 10.4
In this example, we show how the interpolation method affects the inter-
polation accuracy. When the number of pilot subcarriers is insufficient, the
interpolation cannot have a good match to the original. Under this condition,
the interpolation method really does not matter. It is not surprising that the
linear interpolation may even be better than the higher-order interpolation.
If there are a sufficient number of pilot subcarriers that exist, then the higher-
order interpolation shows some advantage in exchange for more computation
time. However, if there are more than enough pilot carriers, then even linear
interpolation can work well.
Figure 10.6 shows the linear interpolation using 16 pilot subcarriers
and m = 3. The interpolation is not that bad. However, it is certainly not as
good as the parabolic interpolation shown in Figure 10.4. Figure 10.7 shows
linear interpolation using 32 pilot subcarriers and m = 3. The interpolation
also has a fairly close match with the original. This also suggests that a lower-
order interpolation is acceptable when more than enough numbers of pilot
subcarriers exist.

Figure 10.6 Linear interpolation using 16 pilot subcarriers.

6899_Book.indb 207 10/22/19 4:55 PM

 208 Introduction to OFDM Receiver Design and Simulation

Figure 10.7 Linear interpolation using 32 pilot subcarriers.

Example 10.5
In this example we show how the delay may affect the interpolation. If the
delay decreases, the spectral period increases. Then the number of pilot sub-
carriers required also decreases. Figure 10.8 shows the parabolic interpolation
with the delay reduced to m = 1 and the number of pilot subcarriers used is
8. In this case, there is only one spectral cycle and the match is much better
than that given in Figure 10.3.
Figure 10.9 is another plot with the number of subcarriers reduced to
four. The pilot subcarriers selected are at k = 7, 23, 39, and 55. Because of
the lack of pilot subcarriers near the beginning and end of the symbol, the
match is not good there. However, the overall match is not bad. This is also
the early statement that a minimum of 4 pilot subcarriers are required for
each spectral cycle.

Example 10.6
In this example, the effects of random noise on the accuracy of parabolic inter-
polation are investigated. Four different numbers of subcarriers are considered

6899_Book.indb 208 10/22/19 4:55 PM

 Channel Estimation and Tracking209

Figure 10.8 Parabolic interpolation with delay = 1 sample and 8 pilot subcarriers.

Figure 10.9 Parabolic interpolation with delay = 1 sample and 4 pilot subcarriers.

and they are M = 4, 8, 16, and 32. The channel parameters used are the same
as previous examples. They are m = 3, N = 64, and b = 0.9. For simplicity,
assuming that Xk = 1, then Yk = Hk. To compute SNR, the signal amplitude
is then the same as the channel amplitude. Both the channel amplitude and

6899_Book.indb 209 10/22/19 4:55 PM

 210 Introduction to OFDM Receiver Design and Simulation

noise amplitude at these pilot subcarriers can be averaged. The SNR in decibels
is then computed based upon these average values.
Assume that Hkr and Hki represent the true channel response at subcarrier
k. Assume that also H! and H ! represent the estimated channel response at
kr kı
subcarrier k. Then the average deviation δ per subcarrier is computed as follows:
63
⎛ 1⎞
∑ ⎡⎣⎢( H kr − H!kr ) + ( H ki − H!kı ) ⎤⎦⎥
2 2
d=⎜ ⎟ (10.36)
⎝ 8⎠ k=0

Assume also that the peak channel amplitude at a subcarrier is ⎪Hpeak⎪.

Then the deviation percentage ε is computed as

d
e= (10.37)
H peak

Figure 10.10 is a plot of ε in percentage against the SNR and four

different numbers of pilot subcarriers. One common trend is the deviation

Figure 10.10 The deviation percentage of channel estimation against the SNR and the
number of pilot subcarriers.

6899_Book.indb 210 10/22/19 4:55 PM

 Channel Estimation and Tracking211

percentage increases as either the SNR decreases or M decreases. At SNR =

42 dB, the ε is 28.7% for M = 8 and 70.7% for M = 4. The high percentage
deviation at such a high SNR shows that both M = 4 and M = 8 are unaccept-
able. For both M = 32 and M = 16, the deviation percentage is acceptable at
a high SNR but not at a low SNR. If we would like the deviation percentage
to be less than 10%, then the SNR has to be higher than 17 dB.
Also, in this example, it is assumed that m = 3 or 3 sample delay. This
is why M = 16 works reasonably well. As m increases, the number of pilot
subcarriers has to increase further. However, a higher number of pilot subcar-
riers takes away the data bandwidth and reduces the date rate. This could be
a design problem for using the comb-type pattern.
In summary, the number of pilot subcarriers required for interpolation
increases with increasing delay. If the number of subcarriers is insufficient,
then the accuracy of interpolation also suffers.

10.4 Channel Tracking

As discussed previously, the channel frequency response can be exactly solved
using the block-type pilot pattern. The simplicity of (10.12) shows the advan-
tage of OFDM against the single-carrier system over the frequency-selective
channel. However, this pattern requires the use of all subcarriers to be pilot
subcarriers and cannot be used frequently. The comb-type pilot pattern is used
within each OFDM data symbol and is suitable for a fast-varying channel.
Since only a finite number of pilot subcarriers are selected, the channel at other
data subcarriers is estimated through interpolation. The accuracy may not be
great depending upon the multipath delay and the number of pilot subcarriers.

10.4.1 The LMS Algorithm

Another method is to adaptively adjust the channel response based upon
the LMS algorithm using a gradient search [5, 6]. The technique intends to
minimize the mean-square error between the information sequence and its
estimation. At the kth subcarrier, assume that the information data is Xk and
the received data is Yk. The information data is estimated to be GkYk and Gk
is the weighting coefficient. From (10.13), the weight Gk is simply the inverse
of the channel frequency response. The difference between the desired and
estimated response is the error signal ε k:
ek = X k − GkYk (10.38)

6899_Book.indb 211 10/22/19 4:55 PM

 212 Introduction to OFDM Receiver Design and Simulation

In (10.38), all quantities are complex in general. The Gk is determined

by minimizing ⎪ε k⎪2 with respect to Gk. This is done by setting (∂⎪ε k⎪2)/∂Gk
to zero. The partial derivative is given here:
2
∂ εk
= −2ekYk∗ (10.39)
∂Gk

Since Yk is not 0, the only way for the left side of (10.39) to be 0 is to set
ε k to 0. After doing that in (10.38), the Gk is solved to give

Xk
Gk = (10.40)
Yk

Clearly, Gk is the inverse of channel frequency response Hk given in

(10.12). This is just another way of interpreting how Hk is obtained from the
point of view of minimizing the MSE.
Since Xk is not always available, a weight-adjusting procedure is avail-
able to solve the problem. This procedure is the least mean square (LMS)
algorithm using the method of steepest descent. The weight can be adjusted
adaptively without any knowledge of the information sequence. Equation
(10.40) applies to every subcarrier index k. The one-tap equalizer is then given
by the following equation:

G j+1,k = G j ,k + m∇ ⎛ e j ,k ⎞
2
(10.41)
⎝ ⎠

where
Gj,k: weight vector at the current or the jth iteration;
Gj+1,k: the weight vector at the next or the j + 1th iteration;
∇(⎪ε j,k⎪2): the gradient or partial derivative of the error ⎪ε j,k⎪2 against
Gj,k;
μ : a scalar constant that controls the rate of convergence;
ε j,k: error between the estimation Gj,kYk and the reference Xk.
Substituting (10.39) into (10.41), we have

G j+1,k = G j ,k − 2me j ,kYk∗ (10.42)

6899_Book.indb 212 10/22/19 4:55 PM

 Channel Estimation and Tracking213

The adaptation index j can also be considered as the jth received OFDM
symbol. Equation (10.42) shows the adaptation depends upon the estimation
error and the received subcarrier Yk∗.

10.4.2 The Condition for Convergence

The adaptation given by (10.42) needs to be convergent or else it will fail. To
derive the condition for convergence, the error ε j,k can be written in terms of
Gj,k given by (10.38) [6]:

(
G j+1,k = G j ,k − 2m X k − G j ,kYk Yk∗ ) (10.43)

(
= G j ,k 1 + 2m Yk
2
) − 2mX Y k k
∗ (10.44)

Taking the expected value on both sides of (10.44), we have

( ) ( )(
E G j+1,k = E G j ,k 1 + 2m Yk
2
) − 2mX Y k k
∗ (10.45)

Setting j = 0, 1, and 2 and expressing the Gj+1,k in terms of G 0,k, we have

( ) ( )(
E G1,k = E G0,k 1 + 2m Yk
2
) − 2mX Y k k
∗ (10.46)

( ) ( )(
E G2,k = E G0,k 1 + 2m Yk
2 2
) − 2mX Y k k
∗
− 2mX kYk∗ 1 + 2m Yk ( 2
) (10.47)
( ) ( )(
E G3,k = E G0,k 1 + 2m Yk ) − 2mX Y
2 3
k k
∗
− 2mX kYk∗ 1 + 2m Yk ( 2
)
(10.48)
− 2mX kYk∗ (1 + 2m Y )
k
2 2

Repeating the process given in (10.46), (10.47), and (10.48), we have for
the n + 1th iteration,

( )( ) ( )
n
( )
E Gn+1,k = E G0,k 1 + 2m Yk
2 n+1
− ∑ 2mX kYk∗ 1 + 2m Yk
2 i
(10.49)
i=0

6899_Book.indb 213 10/22/19 4:55 PM

 214 Introduction to OFDM Receiver Design and Simulation

If n goes to infinity, the second term converges to the following:

( ) ( )
∞ ∞
2 i 2 i
∑2mX kYk∗ 1 + 2m Yk = 2mX kYk∗ ∑ 1 + 2m Yk
i=0 i=0
(10.50)
2mX kYk∗ X
= 2 =− k
−2m Yk Yk

Equation (10.50) shows that the second term of (10.49) converges to a

finite number. In order for E(Gm+1,k)to converge, the first term of (10.49) must
be 0 and the following condition must be satisfied:
2
−1 < 1 + 2m Yk < 1 (10.51)

The condition for convergence is then the following:

−1
2 <m<0 (10.52)
Yk

Equation (10.52) shows that for convergence, the scalar constant μ is

related to the total power of the received kth subcarrier and it must be less
than 0.
Substituting (10.50) and (10.52) to (10.49) and setting n to ∞, we
finally have
Xk
(
E G∞,k = ) Yk
(10.53)

Equation (10.53) is the same as (10.40). This shows that the gain adapta-
tion given by (10.42) converges to the exact or Wiener-Hopf equation given
by (10.40).
Equation (10.53) is important in the sense that the ideal solution can be
reached by simply performing the adaptation given by (10.42) as long as μ is
chosen to satisfy (10.52). The error ε j,k given by (10.38) requires the knowledge
of the reference Xk which is unknown. However, it can be estimated from
the previous (j − 1)th iteration. Based upon this estimation, the gain can be
estimated in the next iteration.
The adaptation speed depends upon the coefficient μ . If it is too small,
it may take a long time to converge. If it is too large, it may not converge at
all. One way to choose the adaptation coefficient μ is in the following:

6899_Book.indb 214 10/22/19 4:55 PM

 Channel Estimation and Tracking215

−a
m= 2
(10.54)
Yk,max

where Yk,max is the maximum of the kth subcarrier. Substituting (10.54) into
(10.52), we have
2
Yk,max
0<a< 2
(10.55)
Yk

The parameter α given in (10.55) controls the convergence behavior and

a proper selection is necessary.

10.4.3 Examples
In the following examples, we concentrate on a single subcarrier k using
QPSK. Assume the multipath channel is given by (10.33) and (10.34). From
the training sequence, the initial gain G is then given here:
Xk 1 H − jH ki
Gk = = = kr2 (10.56)
Yk H kr + jH ki H kr + H ki2

Later, assume the channel suffers gradual fading which is determined

by the factor b in (10.33) and (10.34). Since b is changed, the initial value of
Gk given by (10.56) is no longer valid. If it is not updated, the receiver per-
formance will degrade.
Using (10.42) for adaptation, the error given by (10.38) can be used to
determine how good the convergence is.

Example 10.7
Let k = 24, b = 0.9, N = 64, and m = 3, and then from (10.33) and (10.34) we
can set some initial values given in (10.57a) to (10.57d):

X 24 = 0.707(1 + j) (10.57a)

H 24 = 1.636 + 0.707 j (10.57b)

Y24 = 0.657 + 1.657 j (10.57c)

G24 = 0.515 − 0.223 j (10.57d)

6899_Book.indb 215 10/22/19 4:55 PM

 216 Introduction to OFDM Receiver Design and Simulation

If the channel does not change, the G24 given by (10.57d) can be used
to achieve a perfect demodulation. Assume now that b drops to 0.8 and is
maintained for at least 20 symbol periods. Since H24 is changed, the gain
G24 is also changed. If the gain is not updated, an error always occurs and
cannot be removed.
Assume now that the gain is updated according to (10.42). Depending
upon the selection of α , there is a different convergence behavior. Figure 10.11
is a plot of the error magnitude ⎪ε k⎪ given in (10.38) against the iteration
number. The plot starts at iteration 1 when b drops to 0.8.
Depending upon the adaptation and α , there are four different curves. If
there is no adaptation, the error remains the same and can never be reduced.
For both α = 0.1 and 0.5, the adaptation converges. However, the conver-
gence speed at α = 0.5 is much faster than at α = 0.1. For α = 0.1, the error is
close to 0 at the end of iteration. For α = 0.5, the error is exactly 0 at iteration
number 5. It has to be careful to continue increasing α . When α is increased
to 1.1, the error continues increasing because the convergence condition given
by (10.55) is not satisfied.

Figure 10.11 The demodulation error against the iteration number without fading.

6899_Book.indb 216 10/22/19 4:55 PM

 Channel Estimation and Tracking217

Example 10.8
In this example, the fading factor b is assumed to drop from 0.9 to 0.8 in 20
steps. In other words, b drops according to the following formula:
b = 0.9 − 0.005n (10.58)

where n is the iteration number. Besides the change of b values, all the other
conditions follow the same as given in Example 10.6. At b = 0.9, all the initial
parameter values are given in (10.57).
Figure 10.12 is a plot of the error magnitude ⎪ε k⎪ against the iteration
number. There are four curves. One is without adaptation. The other three
curves have α = 0.1, 0.5, and 1.12. The error magnitude is the same at iteration
number 1 for all. This is because at iteration number 0, the error magnitude is
zero. Without adaptation, the error continues increasing with increasing itera-
tion number. With adaptation, the error depends upon the choice of α . The
error does not go to 0 because the channel frequency response changes due
to the b variations at every iteration number. For α = 0.5, the error is better
than that for α = 0.2. At α = 1.12, the error is small initially and diverges
eventually, because the convergence condition given by (10.55) is not satisfied.

Figure 10.12 The error magnitude against iteration number with fading.

6899_Book.indb 217 10/22/19 4:55 PM

 218 Introduction to OFDM Receiver Design and Simulation

10.5 Summary
Two pilot patterns were used to estimate the channel. One is the block type
and the other is the comb type. The pilot patterns are usually given in a time-
frequency diagram.
For the block-type pattern, the pilot subcarriers are periodically inserted
at all subcarriers. This type is used when the channel variation is slow. At the
insertion symbol time, there are no data subcarriers. Therefore, the channel
is accurately estimated at all subcarriers. Since the transmission of the pilot
symbol is an overhead, the price paid is a reduction in the data rate.
For the comb-type pattern, the pilots are sent all the time but only a
finite number of subcarriers are dedicated to be pilots. This type is used when
the channel is fast varying. Even though the channel at the pilot subcarri-
ers can be exactly determined, the channel at the data subcarriers has to be
estimated through interpolation. Lagrange interpolation using polynomials
is used. Depending upon the interpolation order, linear, parabolic, and cubic
interpolation formulas were derived. The interpolation accuracy depends upon
the sample delay, interpolation order, the number of pilot subcarriers, and
random noise. Clearly, the interpolation accuracy increases with an increasing
number of available pilot subcarriers. Since the number of pilot subcarriers is
limited and there is uncertainty on the sample delay, the interpolation may
not generate a good channel estimate.
The channel tracking provides an alternative to comb pilot pattern. The
LMS algorithm uses a gradient search to adaptively adjust the channel response.
The channel at every data subcarrier can be independently tracked and only
one tap is necessary. The tracking depends upon the received subcarrier samples
and the mean-square error between the information sequence and its estimate.
As long as the adaptation coefficient is selected to meet the convergence con-
dition, the channel estimation error is significantly reduced in comparison to
without any tracking. Another advantage is adaptive channel tracking does
not need pilot subcarriers, and the data rate will not additionally suffer.
The next chapter concentrates on the decoding process. Both hard deci-
sion decoding and soft decision decoding are discussed.

References
[1] Chiueh, T.-D., P. Y. Tsai, and I. W. Lai, Baseband Receiver Design for Wireless MIMO-
OFDM Communications, New York: Wiley, 2012.

6899_Book.indb 218 10/22/19 4:55 PM

 Channel Estimation and Tracking219

[2] Rinne, J., and M. Renfors, “Pilot Spacing in Orthogonal Frequency Division
Multiplexing Systems on Practical Channels,” IEEE Transactions on Consumer
Electronics, Vol. 42, No. 4, 1996, pp. 959−962.
[3] Hsieh, M. H., and C. H. Wei, “Channel Estimation for OFDM Systems Based upon
Comb-Type Pilot Arrangement in Frequency Selective Fading Channels,” IEEE
Transactions on Consumer Electronics, Vol. 44, No. 1, 1998, pp. 217−225.
[4] Eurp, L., F. M. Gardner, and R. A. Harris, “Interpolation in Digital Modems, Part II,
Implementation and Performance,” IEEE Transactions on Communications, Vol. 41,
No. 6, 1993, pp. 998−1008.
[5] Proakis, J. G., Digital Communications, New York: McGraw-Hill, 1983.
[6] Widrow, B., et al., “Adaptive Antenna Systems,” IEEE Proceedings, Vol. 55, No. 12,
1967, pp. 2143−2159.

6899_Book.indb 219 10/22/19 4:55 PM

6899_Book.indb 220 10/22/19 4:55 PM
 11
Data Decoding

11.1 Introduction
In previous chapters, we have discussed signal detection, synchronization,
channel estimation, and tracking. After the completion of those steps, each
OFDM symbol should be correctly identified. The next step is the demodu-
lation after channel removal to recover the in-phase and quadrature-phase
samples for each subcarrier. These IQ samples are then demapped. Both hard
decision decoding and soft decision decoding are covered.
For the hard decision decoding, both a brute-force approach and a sim-
plified approach are discussed. The brute-force approach has to compute all
the Euclidean distances against all the constellation points. For the simplified
approach, each bit is determined individually based upon a simple metric for
the I bits and Q bits. After demapping and subsequent deinterleaving, Viterbi
decoding follows and the distance metric is based upon computing the Ham-
ming distance. The Viterbi decoder output is then descrambled to recover the
data bit sequence.
For the soft decision decoding, the IQ samples are used together with
Viterbi decoding for demapping and error correction. The difference from the
hard decision decoding is that the distance metric is based upon computing
the Euclidean distance in each branch of the Viterbi decoder. A simplified log
likelihood ratio (LLR) for soft decoding is also discussed. The transmitter and

221

6899_Book.indb 221 10/22/19 4:55 PM

 222 Introduction to OFDM Receiver Design and Simulation

receiver block diagrams are also slightly modified for the purpose of discuss-
ing soft decision decoding.
Both hard decision decoding and soft decision decoding are discussed
in the following sections.

11.2 Demodulation
Figure 4.13 shows a typical receiver block diagram. After ADC and cyclic
prefix removal, the FFT is performed to transform the samples from the time
domain to the frequency domain. The channel effects can then be removed
using the techniques discussed in Chapter 10. After the in-phase I samples
and quadrature-phase Q samples are recovered, the next step is demodulation.
On demodulation, there are two different ways data can be decoded.
One is hard decision decoding and the other is soft decision decoding. In the
hard decision decoding, the I and Q samples are first demapped before pass-
ing to the convolutional decoder. The Hamming distance is used to find the
right path through the trellis diagram to correct errors. In the soft decision
decoding, the I and Q samples are not demapped initially. They are used to
compute the Euclidean distance in the Viterbi decoder to find the right path
for error correction. Since the I and Q samples have a range of possible val-
ues, more information is carried. This additional information provides the
improved error correction capability of the Viterbi decoder.
Both hard decision and soft decision decoding are covered in the fol-
lowing sections.

11.3 Hard Decision Decoding

In the hard decision decoding, the first step is to transform the I and Q samples
into data bits. There are two ways to do it. One way is to use the conven-
tional Euclidean distance, which is time-consuming. The other is a simplified
demapper without loss of accuracy [1]. Both approaches are discussed in the
following sections.

11.3.1 Conventional Demapper

Let Yk represent the kth received subcarrier of an OFDM symbol as given here:
Yk = YI ,k + jYQ ,k (11.1)

6899_Book.indb 222 10/22/19 4:55 PM

 Data Decoding223

where Y I,k and YQ,k represent the in-phase and quadrature-phase sample of the
kth subcarrier. Let Hk represent the channel estimation of the kth subcarrier.
After removing the channel degradation, the restored subcarrier Zk is then
given according to (10.13):
Yk
Zk = (11.2)
Hk

Assume also that there are M constellation points for a given modula-
tion. Let the ith constellation point be given as
AI ,i + jAQ ,i
Ai =
N

where N is the normalization factor. The Euclidean distance between Zk and

Ai is then given here:
2
Di = Zk − Ai
(11.3)
(
⎡ NZ − A
) + ( NZ − AQ ,i ⎤⎥)
2 2

⎢ I ,k I ,i Q ,k
=⎣ 2
⎦
N

If there are M constellation points, there are M such computations

according to (11.3). Since N is a constant, the decision rule can be based upon
the following modified Di:

( ) + ( NZ )
2 2
Di = NZ I ,k − AI ,i Q ,k
− AQ ,i
(11.4)

The lth constellation point Al is selected if Dl is the minimum. Once Al

is determined, the corresponding bit sequence is obtained.

Example 11.1
Figure 11.1 shows the constellation of a QPSK modulation. The normalization
factor is N = 2 . The four constellation points are then A1 = (1 + j)/ 2 , A2
= (1 − j)/ 2 , A3 = (−1 + j)/ 2 , and A4 = (−1 − j)/ 2 . The coordinates of the
constellation points are plotted as NAi. Assume that the received kth subcar-
rier after removing the channel frequency response is Zk = (0.8/ 2 )(1 + j).
From (11.4), the decision metric is then the distance from 0.8 + 0.8j to 1 + j,
1 − j, −1 + j, and −1 − j. They are given by D1 = 0.08, D2 = 3.28, D3 = 3.28,

6899_Book.indb 223 10/22/19 4:55 PM

 224 Introduction to OFDM Receiver Design and Simulation

Figure 11.1 QPSK constellation.

and D4 = 6.48. Since D1 is the minimum, the constellation point selected is

A1. Based upon A1, the data bit sequence is mapped to be 11.
For QPSK, only four Euclidean distances need to be computed. For
64-QAM, the number of distance computations jumps to 64 for each subcar-
rier. Therefore, it is necessary to simplify the computations.

11.3.2 Simplified Demapper

The approach given in the previous section is to first find the best constel-
lation point. Based upon this constellation point, all the sequence bits are
obtained at the same time. In the simplified approach [1], each bit is decoded
separately. This is possible due to the way the bits sequence is assigned to the
constellation point.
To simplify the presentation, let UI,k and UQ,k be defined here:
U I ,k = NZ I ,k (11.5)

U Q ,k = NZ Q ,k (11.6)

Substituting (11.5) and (11.6) into (11.4), we have

( ) + (U )
2 2
Di = U I ,k − AI ,i Q ,k
− AQ ,i (11.7)

6899_Book.indb 224 10/22/19 4:55 PM

 Data Decoding225

To further simplify the presentation, the index i is dropped and we have

( ) + (U )
2 2
D = U I ,k − AI Q ,k
− AQ
(11.8)
= DI + DQ

where DI and DQ are given here:

( )
2
DI = U I ,k − AI (11.9)

( )
2
DQ = U Q ,k − AQ (11.10)

Assume that each constellation point is mapped to 2m bits. As will

become clear later, the first m bits are in-phase bits, while the second m bits
are quadrature-phase bits. Each bit is either bI,i or bQ,i where i ranges from 1
to m. The in-phase bits are determined by DI while the quadrature-phase bits
are determined by DQ.
Each bit is either 0 or 1 and has its own decision region. For the in-phase
bits, they are either R 0I,i or R1I,i where i is the bit number. For the quadrature-
phase bits, they are either R 0Q,i or R1Q,i.

11.3.2.1 Simplified QPSK Demodulation

For the QPSK, each constellation point has 2 bits and m = 1. The 2-bit sequence
is represented by bI,1bQ,1. As can be seen from Figure 11.1, the decision region
is to the left of the y-axis if the first bit is 0 and to the right of the y-axis if
the first bit is 1. In other words, R 0I,1 is for UI,k < 0 while R1I,1 is for UI,k > 0.
Since the decision boundary is the y-axis, the first bit is the in-phase bit and
is determined by DI.
For the second bit, the decision region is the x-axis. The decision region
is above the x-axis if the second bit is 1 and below the x-axis if the second bit
is 0. In other words, R 0Q,1 is for UQ,k < 0 while R1Q,1 is for UI,k > 0. Since the
decision region is the x-axis, the second bit is the quadrature-phase bit and is
completely determined by DQ.
Using DI, the first bit is 1 if the following is true:

(U ) ( )
2 2
I ,k
− 1 < U I ,k + 1 (11.11)

From (11.11), we have UI,k > 0 if the first bit is 1 and UI,k < 0 if the first
bit is 0. This quantitatively justifies the previous argument.

6899_Book.indb 225 10/22/19 4:55 PM

 226 Introduction to OFDM Receiver Design and Simulation

Using DQ , the second bit is one if the following is true:

(U ) ( )
2 2
Q ,k
− 1 < U Q ,k + 1 (11.12)

From (11.12), we similarly have UQ,k > 0 if the first bit is 1 and UQ,k < 0
if the first bit is 0.
In summary, the decision rule is given here:

⎧ 1 if U I ,k > 0
bI ,1 = ⎨ (11.13a)
0 if U I ,k < 0
⎩

⎧⎪ 1 if U Q ,k > 0
bQ ,1 = ⎨ (11.13b)
⎪⎩ 0 if U Q ,k < 0

Example 11.2
In Example 11.1, the received kth subcarrier is Zk = (0.8/ 2 )(1 + j). Multiplying
by the normalization factor 2 , we have Uk = (0.8)(1 + j). The in-phase and
quadrature-phase components are then UI,k = 0.8 and UQ,k = 0.8. From (11.13),
the bits are decoded as bI,1 = 1 and bQ,1 = 1. The sequence of 2 bits is then 11.
Equation (11.13) shows that the computation of Euclidean distance is
completely removed. A simple comparison of the in-phase and quadrature-
phase component against 0 determines the baud bits.

11.3.2.2 Simplified 16-QAM Demodulation

For 16-QAM, each constellation point has 4 bits. The first two bits are in-
phase bits and the last two bits are quadrature-phase bits. The bit sequence is
given by bI,1bI,2bQ,1bQ,2. Figure 11.2 shows the partitions of the first in-phase
bit bI,1. As can be seen, the partition boundary is the y-axis. The partition R 0I,1
is to the left of the y axis while R1I,1 is to the right of the y-axis. Therefore, the
shortest distance is determined by DI.
In each partition, there are two constellation points on the x-axis. The
first step is to choose the constellation point in each partition. In the parti-
tion R1I,1, the two constellation points are 3 and 1. The constellation point 3
is chosen if the following is true:

(U ) ( )
2 2
I ,k
− 3 < U I ,k − 1 (11.14)

6899_Book.indb 226 10/22/19 4:55 PM

 Data Decoding227

Figure 11.2 The partitions of the first in phase bit of 16-QAM.

After some simple algebra, (11.14) leads to the condition UI,k > 2. Simi-
larly, constellation point 1 is chosen if UI,k < 2.
In the partition R 0I,1, there are also two constellation points −1 and −3.
The constellation point −3 is chosen if the following is true:

(U ) ( )
2 2
I ,k
+ 3 < U I ,k + 1 (11.15)

Equation (11.15) leads to the condition UI,k < −2. Similarly, the condi-
tion becomes UI,k > −2 if the constellation point −1 is chosen.
For UI,k > 2, the constellation point in partition R1I,1 is 3 and the closest
constellation point in partition R 0I,1 is −1. Therefore, bI,1 = 1 if the following
is true:

(U ) ( )
2 2
I ,k
− 3 < U I ,k + 1 or U I ,k > 1 (11.16)

Following the similar logic, the following are obtained:

( ) ( )
2 2
U I ,k < −2 and bI ,1 = 0 if U I ,k + 3 < U I ,k − 1 or U I ,k < −1 (11.17)

( ) ( )
2 2
U I ,k < 2 and bI ,1 = 1 if U I ,k − 1 < U I ,k + 1 or U I ,k > 0 (11.18)

6899_Book.indb 227 10/22/19 4:55 PM

 228 Introduction to OFDM Receiver Design and Simulation

(11.19)

From (11.16) to (11.19), bI,1 is decoded based upon the following criteria:
U I ,k − 1 U I ,k > 2 (11.20a)

U I ,k U I ,k < 2 (11.20b)

U I ,k + 1 U I ,k < −2 (11.20c)

For each criterion, the first term determines whether the bit is 1 or 0.
For example, the first criterion is to set bI,1 = 1 for UI,k − 1 > 0 and bI,1 = 0 for
UI,k − 1 < 0 under the condition of UI,k > 2.
Figure 11.3 shows the partitions of the second in-phase bit bI,2. The two
vertical lines y1 and y2 generate three partitions. The partition R 0I,2 is to the
left of y1 and to the right of y2 while the partition R1I,2 is between y1 and y2.
The constellation point is 3 if the partition R 0I,2 is to the right of y2 and −3 if
the partition R 0I,2 is to the left of y1.
For constellation R1I,2, there are two possible constellation points −1 and
1 on the x-axis. To choose 1 as the constellation point, the following inequal-
ity must be satisfied:

Figure 11.3 The partitions of the second in-phase bit of 16-QAM.

6899_Book.indb 228 10/22/19 4:55 PM

 Data Decoding229

(U ) ( )
2 2
I ,k
− 1 < U I ,k + 1 or U I ,k > 0 (11.21)

Similarly, to choose −1 as the constellation point, the condition becomes

UI,k < 0.
Based upon the same logic as before, bI,2 = 0 if the following inequali-
ties are true:

(U ) ( )
2 2
I ,k
− 3 < U I ,k − 1 or − U I ,k + 2 < 0 for U I ,k > 0 (11.22a)

(U ) ( )
2 2
I ,k
+ 3 < U I ,k + 1 or U I ,k + 2 < 0 for U I ,k < 0 (11.22b)

Similarly, bI,2 = 1 if the following inequalities are true:

(U ) ( )
2 2
I ,k
− 1 < U I ,k − 3 or − U I ,k + 2 > 0 for U I ,k > 0 (11.23a)

(U ) ( )
2 2
I ,k
+ 1 < U I ,k + 3 or U I ,k + 2 > 0 for U I ,k < 0 (11.23b)

Equations (11.22) and (11.23) can be combined to have the simple cri-
terion −⎪UI,k⎪ + 2. The interpretation is then bI,2 = 1 if −⎪UI,k⎪ + 2 > 0 and
bI,2 = 0 if −⎪UI,k⎪ + 2 < 0.
Equation (11.20) can be further simplified. For criterion (1), UI,k − 1 is
always greater than 0 if UI,k > 2. Therefore, the criterion is similar to bI,1 = 1
for UI,k > 0. Similarly, for criterion (3), UI,k + 1 is always less than zero if UI,k
< −2. The criterion is then the same as bI,1 = 0 for UI,k < 0. Based upon these
arguments, (11.20) and (11.23) can be combined to give the following criteria:
bI ,1: U I ,k (11.24a)

bI ,2 : − U I ,k + 2 (11.24b)

Similarly, the quadrature-phase bits have the following criteria:

bQ ,1: U Q ,k (11.25a)

bQ ,2 : − U Q ,k + 2 (11.25b)

6899_Book.indb 229 10/22/19 4:55 PM

 230 Introduction to OFDM Receiver Design and Simulation

Equations (11.24) and (11.25) are significantly simplified in comparison

with the computations of 16 Euclidean distances to decode one subcarrier.

Example 11.3
Assume 16-QAM and the received kth subcarrier is Zk = (−0.5 + 3j)/ 10
. The normalization factor is 10 . After multiplying by 10 , the received
Zk = (−0.5 + 3j). Using the brute-force approach, it is necessary to compute
16 Euclidean distances. Figure 11.4 is a plot of these 16 Euclidean distances
against the decimal equivalent of the bit sequence. The minimum is 0.25 cor-
responding to the decimal number 6. The bit sequence is then 0110. The two
in-phase bits are 01 and the two quadrature-phase bits are 10.
Equation (11.24) and (11.25) can be used to quickly get the same answer.
The in-phase and quadrature-phase components are UI,k = −0.5 and UQ,k = 3.
To decode the in-phase bits, we have from (11.24), UI,k < 0 while −⎪UI,k⎪ + 2
= 1.5 > 0. The two in-phase bits are then 01. To decode the two quadrature-
phase bits, we have from (11.25), UQ,k > 0 while −⎪UQ,k⎪ + 2 = −1 < 0. The
two quadrature-phase bits are then 10. The whole bit sequence is then 0110,
which is the same as that from the brute-force computation.

11.3.2.3 Simplified 64-QAM Demodulation

Figure 2.6 shows the constellation for 64-QAM. The constellation points are
1, 3, 5, and 7 on the positive x-axis and −1, −3, −5, and −7 on the negative
x-axis. The same is true on the y-axis. Each constellation point has 6 bits.
There are 3 in-phase bits and 3 quadrature-phase bits. The bit sequence can

Figure 11.4 Plot of Euclidean distances for QAM-16 and Uk = −0.5 + 3j.

6899_Book.indb 230 10/22/19 4:55 PM

 Data Decoding231

be written as bI,1bI,2bI,3bQ,1bQ,2bQ,3. Following the same procedure as given in

the previous two sections for QPSK and 16-QAM, the results are given here:

bI ,1: (1) U I ,k U I ,k ≤ 2
(2) U I ,k − 1 2 < U I ,k ≤ 4
(3) U I ,k − 2 4 < U I ,k ≤ 6
(4) U I ,k − 3 U I ,k > 6 (11.26)
(5) U I ,k + 1 −4 ≤ U I ,k < −2
(6) U I ,k + 2 −6 ≤ U I ,k < −4
(7) U I ,k + 3 U I ,k < −6

bI ,2 : (1) − U I ,k + 3 U I ,k ≤ 2

(2) 4 − U I ,k 2 < U I ,k ≤ 6 (11.27)

(3) − U I ,k + 5 U I ,k > 6

bI ,3 : (1) U I ,k − 2 U I ,k ≤ 4
(11.28)
(2) − U I ,k + 6 U I ,k > 4

Equation (11.26) can be further simplified. Criteria (2), (3), and (4) are
always positive under the specified positive values of UI,k. This is the same
as bI,1 = 1 for UI,k > 0. Similarly, criteria (5), (6), and (7) are always negative
under the specified negative values of UI,k. Again, this is the same as bI,1 = 0
for UI,k < 0. Criterion (1) also suggests bI,1 = 1 for UI,k > 0 and bI,1 = 0 for UI,k
< 0. In other words, bI,1 is completely determined by UI,k.
Equation (11.27) can also be further simplified. Criterion (1) is always
positive for ⎪UI,k⎪ ≤ 2. Criterion (3) always generates negative values for ⎪UI,k⎪
> 6. Both are either always positive or always negative under the second cri-
terion. In other words, bI,2 is completely determined by 4 − ⎪UI,k⎪ without
any conditions.
Equation (11.28) shows that there are two criteria to specify bI,3 under
two separate conditions. These two criteria can be replaced by a single criterion
−⎥⎪UI,k⎪ − 4⎪ + 2 without any attached conditions [1]. By tracing through
all possible values of UI,k, it can be seen that this single criterion is equivalent
to (11.28).

6899_Book.indb 231 10/22/19 4:55 PM

 232 Introduction to OFDM Receiver Design and Simulation

From the above arguments, the further simplified criteria to determine

the in-phase bits are summarized here [1]:
bI ,1: U I ,k
bI ,2 : 4 − U I ,k (11.29)

bI ,3 : − U I ,k − 4 + 2

Similarly, the quadrature-phase bits are determined by the follow-

ing criteria:
bQ ,1: U Q ,k

bQ ,2 : 4 − U Q ,k (11.30)

bQ ,3 : − U Q ,k − 4 + 2

Example 11.4
Assume the modulation is 64-QAM and the received Zk = (1.5 + j)/ 42 . The
normalization factor is 42 . Multiplying by 42 , we have Uk = 1.5 + j. Figure
11.5 is a plot of Euclidean distance against all 64 possible bit sequences. The
minimum is 54 and the corresponding bit sequence is 110110.

Figure 11.5 Plot of Euclidean distance for 64-QAM and Uk = 1.5 + j.

6899_Book.indb 232 10/22/19 4:55 PM

 Data Decoding233

Using the simplified approach, the in-phase and quadrature-phase com-

ponents are UI,k = 1.5 and UQ,k = 1, respectively. The three criteria for the
in-phase bits from (11.29) are 1.5, 2.5, and −0.5, respectively. The in-phase
bit sequence is then 110. For the quadrature-phase bits, the three criteria gen-
erate 1, 3, and −1. The quadrature phase bit sequence is then also 110. The
complete bit sequence is then 110110. The simplified approach generates the
same answer but with much less computation.

11.3.3 Deinterleaver and Viterbi Decoding

After the I and Q samples are demapped, the bit sequence must first be dein-
terleaved. The deinterleaver process is discussed in Chapter 7. The output bit
sequence from the deinterleaver is then fed into the Viterbi decoder. The details
of the Viterbi decoder are also discussed in Chapter 7.
The Viterbi decoder in Chapter 7 is illustrated using Hamming distance
as the distance metric. The received baud is compared with the baud from each
trellis branch to get the distance. For a 1/2 convolutional decoder, each baud
has 2 bits. Table 7.3 lists the Hamming distance for all the possible bauds. In
essence, this is hard decision decoding.

11.3.4 Descrambler
The last step in the decoding process is to pass the corrected bit sequence from
the Viterbi decoder to a descrambler to recover the data bits. The scrambler
in the transmitter is necessary for data security. The scrambler using the shift
register sequence is discussed in Chapter 5.
Since the receiver knows exactly how the data is scrambled in the trans-
mitter, the descrambling process is pretty easy and is illustrated in Figure 5.7.

11.4 Soft Decision Decoding

For soft decision decoding, two slightly modified transmitter and receiver
block diagrams are given first. After understanding the overall architectures,
the soft decision processes then follows.

11.4.1 Transmitter and Receiver Block Diagram

For the soft decision decoding, a typical transmitter block diagram is shown
in Figure 11.6. In comparing with Figure 4.1, the only change is that the error

6899_Book.indb 233 10/22/19 4:55 PM

 234 Introduction to OFDM Receiver Design and Simulation

Figure 11.6 A typical soft decision encoder block diagram.

correction encoder and interleaver are replaced by the outer encoder, inter-
leaver, and convolutional encoder. The outer encoder could be a Reed-Solomon
encoder, which can be applied to reduce block errors [2, 3]. However, it can
be another error correction encoder depending on the specific application. For
example, the outer encoder could be a quaternary convolutional encoder [4].
The output from the outer encoder then passes through an interleaver and then
a convolutional encoder before getting into the modulation process. After the
modulator, it follows what is given in Figure 4.1.
A typical receiver block diagram is shown in Figure 11.7. In compar-
ing with Figure 4.13, the demodulation, deinterleaver, and error correction
decoder are replaced by Viterbi decoder, deinterleaver, and outer decoder. After
the channel effects are removed, the I and Q samples have a range of values
and include valuable information. They will not be demapped immediately.
Instead, they will be used in combination with a Viterbi decoder to generate
the bit sequence.
The output from the Viterbi decoder then goes through deinterleaver
and outer decoder before getting into the descrambler to recover the data
bit sequence.

Figure 11.7 A typical soft decision decoder block diagram.

6899_Book.indb 234 10/22/19 4:55 PM

 Data Decoding235

11.4.2 Soft Decoding Using the Euclidean Distance

Using the soft decision in the Viterbi decoding, the only difference from the
hard decision is the distance metric. As was mentioned before, the metric used
is the Euclidean distance. For simplicity in the illustration, assume that QPSK
is used for the modulation. For each branch in the Viterbi trellis diagram, let
the received sample be given as I + jQ. Each branch output baud is 00, 01, 11,
or 10. Their equivalent values in the constellation are given in Table 2.2. Let
each branch baud be represented as BI + jBQ. Then the Euclidean distance is
simply given here:

( I − BI )2 + ( Q − BQ )
2
D= (11.31a)

To save computation time, D can also be computed as follows:

D = I − BI + Q − BQ (11.31b)

In each branch, the Hamming distance for hard decision decoding is

then replaced by D given in (11.31).
Following the same process as in hard decision decoding, the path score is
accumulated for each additional branch. At the last branch, the path with the
minimum accumulated distance is the optimum path. The input bit sequence
is then determined from this optimum path.

11.4.3 Soft Decoding Using LLR

The simplified demodulation discussed in Section 11.3.2 simplifies the hard
decision demodulation. However, the soft bit values defined are actually derived
from LLR, which is very similar to what has been presented in Section 8.4.
Assume that the constellation point is a in the decision region R1I,j and b in
the decision region R 0I,j for the jth I bit. Then point a is selected if the follow-
ing LLR is true

⎛ y ⎞ ⎛ y ⎞
p⎜ 1 ⎟ > p⎜ 0 ⎟ (11.32)
⎝ RI , j ⎠ ⎝ RI , j ⎠

where y is the received in-phase sample. Assuming that y is Gaussian distrib-

uted with standard deviation σ , (11.32) then becomes

6899_Book.indb 235 10/22/19 4:55 PM

 236 Introduction to OFDM Receiver Design and Simulation

⎡ ( y−a)2 ⎤ ⎡ ( y−b)2 ⎤
1 ⎢− 2s 2 ⎥⎦ 1 ⎢− 2s 2 ⎥⎦
e⎣ > e⎣ (11.33)
2ps 2ps

Removing the common term and taking logarithm on both sides of

(10.33), we have

( y − a)2 < ( y − b)2 if a is selected (11.34)

Comparing the Euclidean distance given in (11.34) is the basis for the
discussion of the hard decision decoding presented in Section 11.3. However,
it is actually derived from the LLR.
Based upon the LLR, the soft bit values are defined for QPSK in (11.13),
for 16-QAM in (11.24) and (11.25) and for 64-QAM in (11.29) and (11.30).
It has been shown these soft bit values can be applied for soft Viterbi decod-
ing successfully [1].

11.4.4 Further Decoding Process

In reference to Figure 11.7, further processes after soft Viterbi decoder are
shown. The corrected data bit sequence is then deinterleaved and further
decoded through the outer decoder. The output bit sequence from the output
decoder is then descrambled to recover the original data bits.

11.5 Summary
The data demodulation starts after the receiver has completed the process of
IQ recovery, synchronization and frequency offset estimation. There are two
different ways for data demodulation. One is hard decision decoding and the
other is soft decision decoding.
On the hard decision decoding, the received I and Q samples have to
be demapped first. There are two different approaches. One is the brute-force
approach by finding the minimum Euclidean distance against all the constel-
lation points. The other is a simplified approach by decoding a single bit each
time based upon it is I bits or Q bits. The decision rules for QPSK, 16-QAM,
and 64-QAM are derived. This greatly simplifies the required computations
and speeds up the demodulation process. After the demapping is complete,
the transmitted bit stream is recovered after additionally going through the
process of deinterleaving, FEC decoding, and descrambling.

6899_Book.indb 236 10/22/19 4:55 PM

 Data Decoding237

On the soft decision decoding, the transmitter and receiver block dia-
grams are slightly modified and discussed. The soft Viterbi decoding is then
described. The I and Q samples are demapped in combination with the Viterbi
decoding. The difference from the hard Viterbi decoding is the computation
of distance metric. The soft decision uses Euclidean distance while the hard
decision uses the Hamming distance.
The soft bit values given in the hard decision for the simplified demodu-
lation are derived from the LLR by comparing two different decision regions
for bit 1 and bit 0. These LLR values can be applied successfully to the soft
Viterbi decoding. Following Figure 11.7, the input data bit sequence can
eventually be recovered.
The next chapter presents the simulation results of the OFDM perfor-
mance in a multipath channel.

References
[1] Tosato, F., and P. Bisaglia, “Simplified Soft-Output Demapper for Binary Interleaved
COFDM with Application to HIPPERLAN/2,” Proceedings of IEEE International
Conference on Communications, 2002, pp. 664−668.
[2] Lin, S., and D. Costello, Error Control Coding, Upper Saddle River, NJ: Prentice Hall,
1983.
[3] Peterson, W. W., and E. J. Weldon, Error Correcting Codes, Cambridge, MA: MIT
Press, 1972.
[4] Singh, M., and I. Wassell, “Comparison Between Soft and Hard Decision Decoding
Using Quaternary Convolutional Encoders and the Decomposed CPM Model,” IEEE
VTS 53rd Vehicular Technology Conference, Vol. 2, 2001, pp. 1347−1351.

6899_Book.indb 237 10/22/19 4:55 PM

6899_Book.indb 238 10/22/19 4:55 PM
 12
Simulation Study of a Multipath
Channel on OFDM

12.1 Introduction
Up to this point, all the major components in the transmitter and the receiver
have been presented. They were provided in Figures 4.1 and 4.13 for the trans-
mitter and receiver block diagrams. The complete system is then simulated to
find out the effectiveness of the OFDM to combat the impact of a multipath
channel and random Gaussian noise.
Both the transmitter and receiver software were written in C language
running on a laptop. The Ubuntu Linux operating system was used to gener-
ate the results. However, the commands and operations are not different from
the Linux operating system.
The transmitter waveform follows exactly what was specified in the IEEE
802.11a. The receiver algorithms have been discussed in the previous chapters.
Another goal for the simulation is to justify these algorithms. On the Vit-
erbi decoding, only hard decision decoding is used. The unique algorithm is
the maximum likelihood detection based upon normalized cross-correlation
for signal acquisition. Other algorithms are the block pilot pattern inserted
at a given period for channel estimation, the convolutional codes including

239

6899_Book.indb 239 10/22/19 4:55 PM

 240 Introduction to OFDM Receiver Design and Simulation

punctured convolutional codes for error corrections, and the LMS algorithm
using the method of steep descent for channel tracking.
In the following, a multipath channel model is presented first. Simula-
tion examples for OFDM performances over both an AWGN and a multipath
channel are then given.

12.2 Characterization of Multipath Channel

From (8.1), the transmitted signal x(t) with a carrier frequency of fc is given here:

(
x(t) = Re s(t)e j2pf c t ) (12.1)

where s(t) is the source signal. To characterize the multipath channel, assume
that there are L(t) paths. The number of paths at any given time is time-
dependent. This means that, at any time, the number of signals the receiver
receives is not a constant. Corresponding to each path, assume the delay is
τ n(t), the Doppler shift is ∈n(t), and the attenuator factor is α n(t). The received
signal r(t) is then given by
L(t )
r(t) = ∑an (t)x ( t − tn (t))e j2pt∈ (t ) n (12.2)
n=1

The Doppler shift, ∈n(t), from each path is also time-dependent. Assum-
ing that the angle between the direction of motion and the signal propagation
path is θ n(t), then the Doppler shift can be written as here:

(
∈n (t) = vf c cos θ n (t) /c) (12.3)

where c is the light velocity and v is the velocity of motion. Substituting (12.1)
into (12.2), the following equation results:

⎧⎪ ⎡ L(t ) ⎤ ⎫⎪
( )
r(t) = Re ⎨ ⎢ ∑ an (t)s t − tn (t) e − j2pf c tn (t )e + j2p∈n (t )t ⎥ e j2pf c t ⎬ (12.4)
⎪⎩ ⎣ n=1 ⎦ ⎪⎭

The quantity inside the bracket is the equivalent lowpass signal z(t) and
is given by
L(t )
z(t) = ∑ an (t)s ( t − tn (t))e − j2p( f t (t )−∈ (t )t )
c n n
(12.5)
n=1

6899_Book.indb 240 10/22/19 4:55 PM

 Simulation Study of a Multipath Channel on OFDM241

Then the phase term in (12.5) can be combined to give

(
jn (t) = −2p f c tn (t) − εn (t)t ) (12.6)

Substituting (12.6) into (12.5), the following results [1]:

L(t )
z(t) = ∑ an (t)e jj (t ) s ( t − tn (t))
n
(12.7)
n=1

Equation (12.7) shows that there is an amplitude perturbation α n(t) and

a phase perturbation φ n(t) corresponding to each path. The phase term φ n(t)
can cause the signal fading. Depending upon the magnitude of phase term
φ n(t), the signals from different paths may add constructively or destructively.
If it is the constructive addition, the combined signal strength increases. How-
ever, if it is the destructive addition, the combined signal strength decreases.
These signal variations are called signal fading.
Equation (12.7) can be further simplified by defining the weighting
factor here:

wn (t) = an (t)e jjn (t ) (12.8)

Substituting (12.8) into (12.7), we have

L(t )
z(t) = ∑ wn (t)s ( t − tn (t)) (12.9)
n=1

From (12.9), the received lowpass signal z(t) is the summation of the
delayed version of the original signal s(t) multiplied by a time-dependent
weighting factor wn(t). This weighting factor is, in general, complex and has
both amplitude and phase variations.
From (12.9), the equivalent channel impulse response h(τ , t) at time t
is given here:
L(t )
h(t,t) = ∑ wn (t)d ( t − tn (t)) (12.10)
n=1

The received signal z(t) is the output of passing the source signal s(t)
through the channel characterized by h(τ , t). Equation (12.10) is justified by
the following convolutional integral:

6899_Book.indb 241 10/22/19 4:55 PM

 242 Introduction to OFDM Receiver Design and Simulation

z(t) = ∫ h(t,t)s(t − t) dt
L(t )
= ∫∑ wn (t)d ( t − tn (t)) s(t − t)dt (12.11)
n=1
L(t )
= ∑ wn (t)s ( t − tn (t))
n=1

From the derivation of (12.11), the channel impulse response given by

(12.10) is justified. Equation (12.9) is used as the basis for the simulation of
a multipath channel.

12.3 Computation of SNR

In a multipath channel, only one is the direct path and the others are due to
reflections. For the direct path, the delay is τ 1(t). Equation (12.11) can then
be rewritten as
L(t )
z(t) = w1(t)s ( t − t1(t)) + ∑ wn (t)s ( t − tn (t)) (12.12)
n=2

In the receiver, the received signal after adding the noise n(t) becomes
y(t) = z(t) + n(t) (12.13a)

Taking the discrete Fourier transform of (12.13a), we have the following:

Yk = Zk + N k (12.13b)

The kjth subcarrier of the mth symbol in the frequency domain is denoted
as Zmkj. Assume that the total number of data subcarriers is NSD. The average
signal energy of these data subcarriers for the mth symbol is then
N
⎛ 1 ⎞ SD 2
Em = ⎜ ∑Z
⎝ N SD ⎟⎠ j=1 mk j
(12.14)

The noise energy in the frequency domain can be computed in a similar

fashion and is given here:

6899_Book.indb 242 10/22/19 4:55 PM

 Simulation Study of a Multipath Channel on OFDM243

⎛ 1 ⎞ NSD 2
Nm = ⎜ ∑N
⎝ N SD ⎟⎠ j=1 mk j
(12.15)

where Nmkj is the noise energy at the subcarrier index kj of the mth symbol
and Nm represents the average noise energy of the mth symbol in the fre-
quency domain.
Assuming that there are a total of N symbols, the average bit SNR Eb/
N0 is then given by
Eb ⎛ 1 ⎞ N Em
= ∑N
N 0 ⎜⎝ N ⎟⎠ m=1 (12.16)
m

12.4 Transmitter Architecture for Simulation

For simulation, the transmitter architecture is the same as given in Figure
4.1. The transmitter waveform follows exactly the specification of the IEEE
802.11a. The channel spacing is 20 MHz. Four different modulation techniques
were studied and they are BPSK, QPSK, 16-QAM, and 64-QAM. The error
control coding selected is the convolutional coder. The three possible coding
rates are 1/2, 3/4, and 2/3. The data scrambling and interleaving follow those
specified in Chapters 5 and 7.
The preamble includes both short sequence and long sequence. The short
sequence has 10 segments and each segment lasts 0.8 μs. The total number of
samples is then 160. The long sequence has two data symbols of 3.2 μs each
and two guard intervals of 0.8 μs each. The short sequence then lasts 8 µs and
the same is true for the long sequence. The total preamble period is then 16 μs.
Each data symbol has a duration of 3.2 μs and a cyclic prefix of 0.8
μs. The total signal duration for each symbol is then 4 μs. The number of
subcarriers is 48 for data and 4 for pilots and the resulting total becomes 52.
The size of IFFT is 64. This means that there are 12 unused subcarriers and
their complex amplitudes are set to zero. Before the data transmission, a signal
symbol of 4 μs is also inserted.
The channel is tracked in two ways. The first is that block type pilot pat-
terns are inserted at a given period. The second is to the use the LMS algorithm
to track the channel at every data subcarrier. For the block-type pilot pattern,
the actual data rate is reduced according to (10.1) due to the overhead trans-
mission. However, there is the advantage of reducing the bit error rate (BER).

6899_Book.indb 243 10/22/19 4:55 PM

 244 Introduction to OFDM Receiver Design and Simulation

12.5 Receiver Architecture for Simulation

The receiver architecture follows that were given in Figure 4.13. Before the
data demodulation, there will be synchronization and signal detection. The
detection algorithm used is the maximum likelihood detection based upon
the simple rule according to (8.27). The decision metric is the normalized
cross-correlation.
The long sequence has two data symbols. The second data symbol is
used for the initial estimation of the channel. The equalization starts at the
signal symbol and continues with the following data symbols. The multipath
channel can have fading and the wn(t) given in (12.10) varies with time. This
means that the channel must also be tracked and the LMS algorithm based
upon (10.42) is used.
The effectiveness of transmitting the block pilot patterns is also simulated.
In other words, the channel is estimated every M data symbols. At the end
of the new channel estimation, it is then tracked until the next pilot period.
After the channel equalization, the sequence of demodulation, dein-
terleaving, descrambling, and convolutional decoding follow to restore the
originally transmitted data.
The performance metric is the BER, defined here:
Total number of bits in error
BER =
Total number of transmitted bits

The unit is normally in percentage.

12.6 Simulation Studies

The simulation of OFDM is grouped into two categories. The first is the addi-
tion of noise, but there are no multipath degradations. The second is with both
noise and multipath degradations.
The details are given in the following sections.

12.6.1 Performance over AWGN

In the following, the transmitter and receiver architecture described in Sec-
tions 12.4 and 12.5 are simulated. The signal detection accuracy can impact
the BER performance. Therefore, the beginning data sample number detected
is checked for every run. As described in Chapter 8, the detection is per-
formed in two steps. The first step is coarse detection and the second step is

6899_Book.indb 244 10/22/19 4:55 PM

 Simulation Study of a Multipath Channel on OFDM245

fine detection. It was found that the beginning sample location is accurately
determined without error for Eb/N0 > 5.2 dB. For Eb/N0 between 2.38 dB and
5.2 dB, there is one sample error for the coarse detection. However, the fine
detection can accurately correct the error. In other words, the signal detection
plays no role in the BER performance.
Figure 12.1 shows the BER performance in decibels over AWGN. For
comparison purposes, four modulation schemes, BPSK, QPSK, 16-QAM, and
64-QAM, are included. The error correction coding applied is the convolu-
tional coder of coding rate = 1/2. The detailed structure is given in Section
7.4.3. As expected, BPSK has the best performance while 64-QAM has the
worst performance.
For BPSK, the BER is 0 at Eb/N0 = 8.71 dB, 0.000385 at Eb/N0 = 7.71
dB, 0.000866 at Eb/N0 = 6.71 dB, and 0.00182 at Eb/N0 = 5.71 dB. The
inclusion of error correction coding and interleaver significantly improve the
error rate performance.
The simulation results also show that the BER is 0 at Eb/N0 = 8.71 dB
for BPSK, 12.71 dB for QPSK, 18.81 dB for 16-QAM, and 26.72 dB for
64-QAM. As the modulation becomes more complicated to increase the data
rate, there is also a requirement for a higher SNR.
Table 12.1 lists the Eb/N0 for BPSK, QPSK, 16-QAM and 64-QAM
at around the same BER. There is about 5.9-dB improvement in Eb/N0 using
BPSK instead of QPSK. If QPSK is used instead of 16-QAM, the improve-
ment of Eb/N0 is about 7.1 dB.

Figure 12.1 BER performance over AWGN using a convolutional of coding rate = 1/2.

6899_Book.indb 245 10/22/19 4:55 PM

 246 Introduction to OFDM Receiver Design and Simulation

Table 12.1
Eb /N 0 Comparisons of Different Modulations at Around the Same BER

Modulation Coding Rate BER Eb /N0

BPSK ½ 0.71% 4.82

QPSK ½ 0.65% 10.71
16-QAM ½ 0.68% 17.81
64-QAM ½ 0.50% 24.81

The punctured convolutional coder has been discussed in Chapter 7. It

is generated based upon the coder rate = 1/2 by dropping bits at a specified
pattern. The net result is an increase in date rate at a cost of a higher BER.
Figure 12.2 a plot of BER performance over AWGN by comparing a rate =
1/2 and a rate = 3/4 convolutional coder. Both QPSK and BPSK modulations
are used in the comparisons.
As expected, the BER of the rate = 1/2 coder is always lower than the
punctured coder of rate = 3/4. At the same BER, that means a higher Eb/N0
is required for rate = 3/4 coder. An example is shown in Table 12.2 by listing
the Eb/N0 at around the BER of 0.012.

Figure 12.2 BER comparisons between QPSK and BPSK.

6899_Book.indb 246 10/22/19 4:55 PM

 Simulation Study of a Multipath Channel on OFDM247

Table 12.2
Eb /N 0 Comparisons Between Rate = 1/2 and 3/4 at Around the Same BER

Modulation Coding Rate BER Eb /N0

QPSK 1/2 1.30% 10.01

QPSK 3/4 1.28% 11.51
BPSK 1/2 1.3% 4.22
BPSK 3/4 1.24% 6.71

For QPSK, the Eb/N0 increase in using the coding rate of 3/4 instead
of 1/2 is 1.5 dB to achieve about the same BER. However, from Table 4.1,
the data rate increases from 48 Mbps to 72 Mbps using a 20-MHz channel
spacing. For BPSK, there is an increase of Eb/N0 by 2.5 dB in using coding
rate of 3/4 instead of 1/2. Again, from Table 4.1, the data rate increases from
24 Mbps to 36 Mbps using a 20-MHz channel spacing. If a higher data rate
is required, then a slight loss in Eb/N0 is not a bad option. The BER becomes
zero at Eb/N0 = 12.71 dB for QPSK and at Eb/N0 = 8.72 dB for BPSK using
a bit coding rate of 3/4.

12.6.2 Performance over a Multipath Channel

In the previous section, the BER performance over background noise was
given. In a real environment, there can be additional multipath delay and signal
fading. To minimize the impact of these degradations, two techniques were
discussed in Chapter 10. The first is the LMS algorithm to track the channel
response. The second is to add a block-type pilot pattern. The efficiency of
both techniques is simulated in this section.
Before the simulation data is given, the simulated multipath channel
and simulation parameters are first presented.

12.6.2.1 Multipath Channel Model for Simulation

A multipath channel characterized by (12.11) is written here
L(t )
z(t) = ∑ wn (t)s ( t − tn (t)) (12.17)
n=1

In (12.17), wn(t) is the signal fading and can have both amplitude and
phase perturbations and τ n(t) is the path delay. For the simulation study, we
consider two multipath channels or L(t) = 2. One path is the main path and

6899_Book.indb 247 10/22/19 4:55 PM

 248 Introduction to OFDM Receiver Design and Simulation

the other path is the reflected path due to obstacles. For the main path or
n = 1, the propagation delay is assumed negligible and is set to zero. For the
reflected path or n = 2, τ 2(t) is assumed to be a constant, τ . Based upon these
two simplifications, (12.17) becomes
z(t) = w1(t)s(t) + w2 (t)s(t − t) (12.18)

In (12.18), the weighting coefficients ω 1(t) and ω 2(t) are complex and
time variant. In the simulation, they are further given by the following
two equations:
w1(t) = w1 p + d1(t) (12.19)

w2 (t) = w2 p + d2 (t) (12.20)

where δ 1(t) and δ 2(t) are small perturbations around their principal values. For
the main path, the principal value is w1p and for the reflected path, the principal
value is w2p. Both δ 1(t) and δ 2(t) are complex and can be written as follows:
dn (t) = dnr (t) + jdni (t) n = 1,2 (12.21)

Both δ nr(t) and δ ni(t) are assumed to be Gaussian random variables with
a zero mean and standard deviation of σ n. This standard deviation can be
considered as the volatility or strength of the fading. The larger the standard
deviation, the stronger is the signal fading.
For each path, a set of two random variables are generated every M data
symbols. One is for the real part, δ nr(t), and the other is for the imaginary
part, δ ni(t). We assume that the channel cannot change that fast. Between
the beginning and ending symbols, the channel is estimated through linear
interpolation at these M symbols. Let t = 0 and t = MT represent the begin-
ning and ending time of any M symbol period where T is the symbol duration.
The symbol duration here includes the guard interval. Assume further that
δ n(0) represents the channel perturbation at t = 0 and δ n(MT ) represents the
channel perturbations at t = MT. Then at t = kT, the channel perturbation,
δ n(kT ), is given by
k
dn (kT ) = dn (0) + (
d (MT ) − dn (0)
M n
) (12.22)

For each symbol, the channel remains the same without any change
throughout the entire duration. After computing δ n(kT ), (12.19) and (12.20)
can be used to compute wn(t).

6899_Book.indb 248 10/22/19 4:55 PM

 Simulation Study of a Multipath Channel on OFDM249

12.6.2.2 Key Parameter Values for Simulation

For the simulation study, the transmitted waveform given by the IEEE 802.11a
is used. The channel spacing is 20 MHz. The modulation is QPSK using 1/2
convolutional error correction coding. The data scrambling and interleaving
are all included in the simulation. Each symbol lasts 4 μs, which includes 3.2
μs for data and 0.8 μs for the guard interval.
Since the LMS algorithm is used to track the channel, the performance
depends strongly on how the tracking coefficient μ defined in (10.54) is
used. If the μ is too large, the channel estimation error diverges and system
performance degrades. However, if it is too small, the convergence speed is
too slow. To obtain an appropriate value, some test runs were applied and the
parameter α given in (10.54) is set to 0.01.
On the block pilot pattern, the known data is inserted with three dif-
ferent periods. They are 5, 10, and 20. If the period is N, that means a known
pilot symbol is inserted after every N symbols. Table 4.1 shows that the data
rate for QPSK using the 1/2 coding rate and 20-MHz channel spacing is 12
Mbps. For periods of 5, 10, and 20, the new data rate is reduced to 10 Mbps,
10.9 Mbps, and 11.42 Mbps, respectively.
On the channel model, the parameter M defined in (12.22) is set to 20
and the delay for the reflected path is set to 0.2 μs. The principal values for
w1p and w2p are set as 1 and 0.5, respectively. The fading strength is given by
its standard deviation, σ n. This means that w1(t) in (12.19) may have values
ranging from 1 − σ n to 1 + σ n and w2(t) in (12.20) may have values ranging
from 0.5 − σ n to 0.5 + σ n. Three different values of σ n, 0.1, 0.25, and 0.3,
are used in the simulation.
12.6.2.3 Simulation Results
As was mentioned before, two approaches are used to minimize the impact
of a multipath channel with signal fading. One is the LMS algorithm and
the other is the block pilot pattern. On the block pilot pattern, the channel
at the data symbols between the inserted pilot patterns is still tracked using
the LMS algorithm.
Figure 12.3 is a plot of BER in percentage against the Eb/N0 (dB) when
the fading standard deviation is 0.1. Three different noise seeds of 93, 40, and
60 are used and the average BER from these three runs is performed. There
are four curves. One curve is for the LMS algorithm without any block pilot
pattern. The other three curves have a block pilot pattern with pilot periods
of 5, 10, and 20 symbols. The LMS algorithm is still used to track the chan-
nel at the data symbols. The LMS algorithm tracks the channel very well as
is evidenced by the negligible error rate percentage when the Eb/N0 is above

6899_Book.indb 249 10/22/19 4:55 PM

 250 Introduction to OFDM Receiver Design and Simulation

Figure 12.3 BER comparisons with fading standard deviation = 0.1.

20 dB. When the Eb/N0 is below 20 dB, the performance advantage of add-
ing a block pilot pattern is obvious. As the period of pilot pattern increases,
the error percentage also increases. When the pilot period is 20 symbols, the
data rate only decreases to 11.42 Mbps from the original 12 Mbps without a

Figure 12.4 BER comparisons with fading standard deviation = 0.25.

6899_Book.indb 250 10/22/19 4:56 PM

 Simulation Study of a Multipath Channel on OFDM251

Figure 12.5 BER comparisons with fading standard deviation = 0.3.

block pilot pattern. However, the BER is significantly improved and especially
at a low Eb/N0.
Figures 12.4 and 12.5 are similar plots except the fading standard devia-
tion is increased to 0.25 and 0.3. The advantage of adding a block pilot pattern
can clearly be seen. There is a small price paid due to data rate decrease, but
the big gain in BER performance is still a worthy approach.

12.7 Summary
A multipath channel was characterized and simulated. First, the BER in
AWGN was compared for four different modulations, BPSK, QPSK, 16-QAM,
and 64-QAM. As the constellation bit encoding becomes more complex, the
performance also degrades. The BPSK has the best performance and there is
a gain of around 5.9 dB in Eb/N0against QPSK while 64-QAM has the worst
performance. Under the same modulation, punctured convolutional coding
can gain in data rate increase with a slight price paid on BER performance.
For QPSK and 1/2 convolutional encoding, the BER performance in a
multipath channel was further performed. Two approaches were studied. One
is the LMS algorithm only. The other is the same LMS algorithm with an
added block pilot pattern. The result shows that the addition of block pilots

6899_Book.indb 251 10/22/19 4:56 PM

 252 Introduction to OFDM Receiver Design and Simulation

can significantly improves the BER, especially at a low Eb/N0. The price paid
is only a slight decrease in date rate.

Reference
[1] Proakis, J. G., Digital Communications, New York: McGraw-Hill, 1983.

6899_Book.indb 252 10/22/19 4:56 PM

 About the Author

Y. J. Liu received his electrical engineering B.S. from the National Taiwan
University, his M.S. from the University of Rochester, and his Ph.D. from
the Ohio State University. He has been working in the industry since 1978.
Initially, he worked for the Lear Siegler Instrument division in LORAN and
OMEGA navigational signal processing and isolated word recognition under
a noisy environment. He then joined Nartron Corporation in developing an
isolated word recognizer for a passenger car. Next, he worked at ITT Exelis for
almost 28 years. He was involved in numerous projects related to the develop-
ment of handheld, wireless, and digital combat tactical radios. Specific working
areas include ad hoc networking and modem and low-rate speech coding. He
has also had many papers published and patents awarded.

253

6899_Book.indb 253 10/22/19 4:56 PM

6899_Book.indb 254 10/22/19 4:56 PM
 Index

Acquisition mode, 175, 184 Code polynomial, 109

Additive inverse, 62 Coding rate, 57
AGC, 59 Coherence time, 88
Amplifier model, 92 Coherence bandwidth, 85
Aliasing, 4, 14 Comb type pilot pattern, 198, 201
Associative, 62, 66 Commutative, 62, 66
AWGN, 244 Compensation of IQ imbalance, 186
Constraint length, 114
Band pass signal, 20 Constellation, 23–26
BER, 245, 251 Constellation point, 226–229
Binary field, 62, 75 Conventional demaper, 222
Bits per subcarrier, 56 Convergence, 213
Block code, 102 Convolution, 15, 135
Block type pilot pattern, 196, 199, 244 Convolutional code, 113, 122
BLUE, 162 Convolutional encoder, 114
Boltzmann’s constant, 91 Convolutional decoder, 117
BPSK, 22, 23, 245, 246, 247 Cross correlation, 131
Cubic interpolation, 203
Cable loss model, 93 Cyclic code, 108
CFO, 160 Cyclic prefix, 39
Channel tracking, 211
Channel estimation, 59, 195, 199, 201 Data bits per symbol, 56
Channel impulse response, 241 Data decoding, 221
Circular convolution, 13 Data transmission rate, 19
Clark’s fading model, 89 Data format, 58
Coarse timing detection, 140 Data scrambler, 74
Coded bits per symbol, 56 Data de-scrambler, 74
Code word, 102 DC offset, 154

255

6899_Book.indb 255 10/22/19 4:56 PM

 256 Introduction to OFDM Receiver Design and Simulation

Decision region, 225 Hard decision decoding, 222

Delay spread, 85 Hamming code, 107
Deinterleaver, 129, 233 Hamming distance, 118, 233
Demodulation, 222 Hata, 82
Descrambler, 233 Header, 56
Detection metric, 136
Diffraction, 77 ICI, 171
Digital modulation, 44 Interleaver, 126–127
Digital modulated signals, 21 Impulse delta function, 2
Digital-to-analog conversion, 1, 50 Interpolation, 8, 201
Direct conversion, 134, 154 Inter-symbol interference, 27, 33
Discrete Fourier transform, 5 Inverse polynomial, 70
Discrete time signals, 2 IQ imbalance, 182
Distributive, 66 Isotropic, 78
Doppler shift, 46, 87, 240
Doppler spread, 88
Lagrange Interpolation, 201–203
EIRP, 78 Large-scale propagation model, 78
Empirical model, 81 Leakage, 155
equivalent noise temperature, 94 Least square estimation, 173–177
Error correction, 95 LENGTH field, 56
Euclidean distance, 223, 235 Likelihood ratio test, 139
Even function, 9 Linearity, 12
Linear amplifier, 154
Fast channel, 89 Linear block code, 102
Fading, 86, 241 Linear interpolation, 201
FFT, 40 Linear system, 15
Fine timing detection, 147 LINUX, 211
Flat fading, 86 LLR, 235
Fourier series, 2 Long division, 64
Frame timing offset, 168 LMS algorithm, 211
Free space propagation loss, 78 Local oscillator, 154, 183
Frequency dispersion, 86 Log-normal shadowing, 83
Frequency shift, 13 Long sequence, 54
Frequency selective fading, 86
FSK, 26 Maximum length sequence, 72
Maximum likelihood detection, 139
Galois field, 64 MESSAGE field, 58
Generator matrix, 102 MMSE metric, 137
Generator polynomial, 108 Multipath channel, 240, 247
Gradient search, 211 Multipath delay, 46, 84, 240
Guard interval, 41 Multiplicative inverse, 62

6899_Book.indb 256 10/22/19 4:56 PM

 Index257

Noise factor, 92 RATE field, 56

Noise energy, 243 Receiver architecture, 58, 244
Noise power density, 92 Rectangular window, 42
Normalization factor, 223 Reed-Solomon encoder, 234
Normalized cross correlation, 138 Reflected path, 248
Nyquist rate, 27
Sample detection, 144
Odd function, 9 SCO, 169
OFDM characteristics, 35 Sampling theorem, 7, 27
OFDM receiver architecture, 58 Scattering, 77
OFDM spectrum, 36 Segment detection, 140
OFDM transmitter architecture, 49 Sequence generator, 66, 73
OFDM waveform properties, 45 SERVICE field, 58
Operation of the DFT, 12 Shift register, 66, 114
Orthogonality, 35 Short sequence, 52
Sidelobe, 43
PAD field, 58 Signal acquisition, 133
PAM, 22 Signal constellation, 23–24, 26
PAPR, 37 Signal fading, 241
PARITY field, 56 Signal transmission format, 50
Parity check, 104 Simplified demapper, 224
Parabolic interpolation, 202 Simplified QPSK demodulation, 225
Parity check matrix, 104 Simplified 16-QAM demodulation, 226
Partition, 226–228 Simplified 64-QAM demodulation,
Period of sequence generator, 70 230
Pilot patterns, 196 Sinc function, 29
Preamble, 51 SNR, 98, 242
Prime, 62 Slow fading, 89
Primitive polynomial, 65 Small-scale propagation model, 83
Propagation loss, 78 Soft decision decoding, 233
Properties of DFT, 8 Spectral distortion, 5
PSK, 22 Steepest descent, 212
Pulse shaping, 27 Synchronization, 59, 153
Punctured convolutional coding, 124, Syndrome, 105
246 Syndrome polynomial, 112

QPSK, 23, 225, 245–247 Tail, 58

QAM, 24, 26, 226, 230 Thermal noise, 91
Quaternary convolutional encoder, 234 Time dispersion, 84
Time selective fading, 89
Raised cosine function, 29 Time and frequency relationship, 10
Range, 96 Time and frequency parameters, 41

6899_Book.indb 257 10/22/19 4:56 PM

 258 Introduction to OFDM Receiver Design and Simulation

Tracking mode, 173 Viterbi algorithm, 117, 233

Time shift, 12
Two-ray model, 80 Weighting factor, 241
Transmission format, 50 Wiener-Hopf equation, 193
Transmitter architecture, 50, 243 Window function, 42

UBUNTA, 239 Z transform, 14

6899_Book.indb 258 10/22/19 4:56 PM

 Recent Titles in the Artech House
Mobile Communications Series
William Webb, Series Editor

3G CDMA2000 Wireless System Engineering, Samuel C. Yang

3G Multimedia Network Services, Accounting, and User Profiles,
Freddy Ghys, Marcel Mampaey, Michel Smouts, and
Arto Vaaraniemi
5G New Radio: Beyond Mobile Broadband, Amitav Mukherjee
5G Spectrum and Standards, Geoff Varrall
802.11 WLANs and IP Networking: Security, QoS, and Mobility,
Anand R. Prasad and Neeli R. Prasad
Achieving Interoperability in Critical IT and Communications Systems,
Robert I. Desourdis, Peter J. Rosamilia,
Christopher P. Jacobson, James E. Sinclair, and
James R. McClure
Advances in 3G Enhanced Technologies for Wireless
Communications, Jiangzhou Wang and Tung-Sang Ng, editors
Advances in Mobile Information Systems, John Walker, editor
Advances in Mobile Radio Access Networks, Y. Jay Guo
Applied Satellite Navigation Using GPS, GALILEO, and Augmentation
Systems, Ramjee Prasad and Marina Ruggieri
Artificial Intelligence in Wireless Communications,
Thomas W. Rondeau and Charles W. Bostian
Broadband Wireless Access and Local Network: Mobile WiMax and
WiFi, Byeong Gi Lee and Sunghyun Choi
CDMA for Wireless Personal Communications, Ramjee Prasad
CDMA Mobile Radio Design, John B. Groe and Lawrence E. Larson
CDMA RF System Engineering, Samuel C. Yang
CDMA Systems Capacity Engineering, Kiseon Kim and Insoo Koo
CDMA Systems Engineering Handbook, Jhong S. Lee and
Leonard E. Miller
Cell Planning for Wireless Communications, Manuel F. Cátedra and
Jesús Pérez-Arriaga
Cellular Communications: Worldwide Market Development, Garry A.
Garrard
Cellular Mobile Systems Engineering, Saleh Faruque
Cognitive Radio Interoperability through Waveform Reconfiguration,
Leszek Lechowicz and Mieczyslaw M. Kokar
Cognitive Radio Techniques: Spectrum Sensing, Interference
Mitigation, and Localization, Kandeepan Sithamparanathan and
Andrea Giorgetti
The Complete Wireless Communications Professional: A Guide for
Engineers and Managers, William Webb
Digital Communication Systems Engineering with Software-Defined
Radio, Di Pu and Alexander M. Wyglinski
EDGE for Mobile Internet, Emmanuel Seurre, Patrick Savelli, and
Pierre-Jean Pietri
Emerging Public Safety Wireless Communication Systems,
Robert I. Desourdis, Jr., et al.
From LTE to LTE-Advanced Pro and 5G, Moe Rahnema and
Marcin Dryjanski
The Future of Wireless Communications, William Webb
Geographic Information Systems Demystified, Stephen R. Galati
Geospatial Computing in Mobile Devices, Ruizhi Chen and
Robert Guinness
GPRS for Mobile Internet, Emmanuel Seurre, Patrick Savelli, and
Pierre-Jean Pietri
GPRS: Gateway to Third Generation Mobile Networks, Gunnar Heine
and Holger Sagkob
GSM and Personal Communications Handbook,
Siegmund M. Redl, Matthias K. Weber, and
Malcolm W. Oliphant
GSM Networks: Protocols, Terminology, and Implementation, Gunnar
Heine
GSM System Engineering, Asha Mehrotra
Handbook of Land-Mobile Radio System Coverage, Garry C. Hess
Handbook of Mobile Radio Networks, Sami Tabbane
High-Speed Wireless ATM and LANs, Benny Bing
Inside Bluetooth Low Energy, Second Edition, Naresh Gupta
Interference Analysis and Reduction for Wireless Systems,
Peter Stavroulakis
Interference and Resource Management in Heterogeneous Wireless
Networks, Jiandong Li, Min Sheng, Xijun Wang, and Hongguang
Sun
Internet Technologies for Fixed and Mobile Networks,
Toni Janevski
Introduction to 3G Mobile Communications, Second Edition,
Juha Korhonen
Introduction to 4G Mobile Communications, Juha Korhonen
Introduction to Communication Systems Simulation,
Maurice Schiff
Introduction to Digital Professional Mobile Radio,
Hans-Peter A. Ketterling
Introduction to GPS: The Global Positioning System,
Ahmed El-Rabbany
An Introduction to GSM, Siegmund M. Redl, Matthias K. Weber, and
Malcolm W. Oliphant
Introduction to Mobile Communications Engineering,
José M. Hernando and F. Pérez-Fontán
Introduction to OFDM Receiver Design and Simulation, Y. J. Liu
Introduction to Radio Propagation for Fixed and Mobile
Communications, John Doble
Introduction to Wireless Local Loop, Broadband and Narrowband,
Systems, Second Edition, William Webb
IS-136 TDMA Technology, Economics, and Services,
Lawrence Harte, Adrian Smith, and Charles A. Jacobs
Location Management and Routing in Mobile Wireless Networks,
Amitava Mukherjee, Somprakash Bandyopadhyay, and Debashis
Saha
LTE Air Interface Protocols, Mohammad T. Kawser
Metro Ethernet Services for LTE Backhaul, Roman Krzanowski
Mobile Data Communications Systems, Peter Wong and
David Britland
Mobile IP Technology for M-Business, Mark Norris
Mobile Satellite Communications, Shingo Ohmori, Hiromitsu
Wakana, and Seiichiro Kawase
Mobile Telecommunications Standards: GSM, UMTS, TETRA, and
ERMES, Rudi Bekkers
Mobile-to-Mobile Wireless Channels, Alenka Zajic′
Mobile Telecommunications: Standards, Regulation, and
Applications, Rudi Bekkers and Jan Smits
Multiantenna Digital Radio Transmission, Massimiliano Martone
Multiantenna Wireless Communications Systems,
Sergio Barbarossa
Multi-Gigabit Microwave and Millimeter-Wave Wireless
Communications, Jonathan Wells
Multipath Phenomena in Cellular Networks, Nathan Blaunstein and
Jørgen Bach Andersen
Multiuser Detection in CDMA Mobile Terminals, Piero Castoldi
OFDMA for Broadband Wireless Access, Slawomir Pietrzyk
Personal Wireless Communication with DECT and PWT, John Phillips
and Gerard MacNamee
Practical Wireless Data Modem Design, Jonathon Y. C. Cheah
Prime Codes with Applications to CDMA Optical and Wireless
Networks, Guu-Chang Yang and Wing C. Kwong
Quantitative Analysis of Cognitive Radio and Network Performance,
Preston Marshall
QoS in Integrated 3G Networks, Robert Lloyd-Evans
Radio Engineering for Wireless Communication and Sensor
Applications, Antti V. Räisänen and Arto Lehto
Radio Propagation in Cellular Networks, Nathan Blaunstein
Radio Resource Management for Wireless Networks, Jens Zander and
Seong-Lyun Kim
Radiowave Propagation and Antennas for Personal Communications,
Third Edition, Kazimierz Siwiak and
Yasaman Bahreini
RDS: The Radio Data System, Dietmar Kopitz and Bev Marks
Resource Allocation in Hierarchical Cellular Systems, Lauro
Ortigoza-Guerrero and A. Hamid Aghvami
RF and Baseband Techniques for Software-Defined Radio,
Peter B. Kenington
RF and Microwave Circuit Design for Wireless Communications,
Lawrence E. Larson, editor
RF Positioning: Fundamentals, Applications, and Tools,
Rafael Saraiva Campos and Lisandro Lovisolo
Sample Rate Conversion in Software Configurable Radios,
Tim Hentschel
Signal Processing Applications in CDMA Communications, Hui Liu
Signal Processing for RF Circuit Impairment Mitigation,
Xinping Huang, Zhiwen Zhu, and Henry Leung
Smart Antenna Engineering, Ahmed El Zooghby
Software-Defined Radio for Engineers, Travis F. Collins,
Robin Getz, Di Pu, and Alexander M. Wyglinski
Software Defined Radio for 3G, Paul Burns
Spread Spectrum CDMA Systems for Wireless Communications, Savo
G. Glisic and Branka Vucetic
Technical Foundations of the IoT, Boris Adryan,
Dominik Obermaier, and Paul Fremantle
Technologies and Systems for Access and Transport Networks,
Jan A. Audestad
Third-Generation and Wideband HF Radio Communications,
Eric E. Johnson, Eric Koski, William N. Furman, Mark Jorgenson,
and John Nieto
Third Generation Wireless Systems, Volume 1: Post-Shannon Signal
Architectures, George M. Calhoun
Traffic Analysis and Design of Wireless IP Networks, Toni Janevski
Transmission Systems Design Handbook for Wireless Networks,
Harvey Lehpamer
UMTS and Mobile Computing, Alexander Joseph Huber and
Josef Franz Huber
Understanding Cellular Radio, William Webb
Understanding Digital PCS: The TDMA Standard,
Cameron Kelly Coursey
Understanding GPS: Principles and Applications, Second Edition,
Elliott D. Kaplan and Christopher J. Hegarty, editors
Understanding WAP: Wireless Applications, Devices, and Services,
Marcel van der Heijden and Marcus Taylor, editors
Universal Wireless Personal Communications, Ramjee Prasad
WCDMA: Towards IP Mobility and Mobile Internet, Tero Ojanperä
and Ramjee Prasad, editors
Wireless Communications in Developing Countries: Cellular and
Satellite Systems, Rachael E. Schwartz
Wireless Communications Evolution to 3G and Beyond,
Saad Z. Asif
Wireless Intelligent Networking, Gerry Christensen,
Paul G. Florack and Robert Duncan
Wireless LAN Standards and Applications, Asunción Santamaría and
Francisco J. López-Hernández, editors
Wireless Sensor and Ad Hoc Networks Under Diversified Network
Scenarios, Subir Kumar Sarkar
Wireless Technician’s Handbook, Second Edition, Andrew Miceli

For further information on these and other Artech House titles,including

previously considered out-of-print books now available through our
In-Print-Forever® (IPF®) program, contact:

Artech House Artech House

685 Canton Street 16 Sussex Street
Norwood, MA 02062 London SW1V 4RW UK
Phone: 781-769-9750 Phone: +44 (0)20 7596-8750
Fax: 781-769-6334 Fax: +44 (0)20 7630-0166
e-mail: [email protected] e-mail: [email protected]

Find us on the World Wide Web at: www.artechhouse.com