OFDM AIR-INTERFACE DESIGN FOR

MULTIMEDIA COMMUNICATIONS
OFDM AIR-INTERFACE DESIGN FOR
MULTIMEDIA COMMUNICATIONS

Proefschrift

ter verkrijging van de graad van doctor

aan de Technische Universiteit Delft,
op gezag van de Rector Magnificus Prof. dr. ir. J. T. Fokkema
voorzitter van het College van Promoties,
in het openbaar te verdedigen op dinsdag 2 april 2002 om 10.30 uur

door

Klaus WITRISAL

Diplomingenieur der Elektrotechnik,

Technische Universitt Graz,
geboren te Graz, Oostenrijk.
Dit proefschrift is goedgekeurd door de promotoren:
Prof. dr. ir. L. P. Ligthart
Prof. dr. R. Prasad

Toegevoegd promotor:
Dr. ir. G. J. M. Janssen

Samenstelling promotiecommissie:
Rector Magnificus voorzitter
Prof. dr. ir. L. P. Ligthart Technische Universiteit Delft, promotor
Prof. dr. R. Prasad Aalborg University, Denmark, promotor
Dr. ir. G. J. M. Janssen Technische Universiteit Delft, toegevoegd promotor
Prof. ir. W. Dik Technische Universiteit Delft
Prof. dr. H. Rohling Technische Universitt Hamburg-Harburg, Duitsland
Prof. dr. ir. W. C. van Etten Universiteit Twente
Prof. dr. ir. J. W. M. Bergmans Technische Universiteit Eindhoven

ISBN: 90-76928-03-7

Copyright 2002 Klaus Witrisal

All rights reserved. No part of the material protected by this copyright notice may be
reproduced or utilized in any form or by any means, electronic or mechanical, inclu-
ding photocopying, recording or by any information storage and retrieval system,
without permission from the author Klaus Witrisal.

Cover Design by Arnold Zwanenburg (http://www.polypo.com)

Printed in Austria
 To Silke

and my parents
Summary

The aim of this dissertation is the investigation of the key issues encountered in the
development of wideband radio air-interfaces. Orthogonal frequency-division multi-
plexing (OFDM) is considered as the enabling technology for transmitting data at
extremely high rates over time-dispersive radio channels. OFDM is a transmission
scheme, which splits up the data stream, sending the data symbols simultaneously at a
drastically reduced symbol rate over a set of parallel sub-carriers.
The first part of this thesis deals with the modeling of the time-dispersive and frequen-
cy-selective radio channel, utilizing second order Gaussian stochastic processes. A
novel channel measurement technique is developed, in which the RMS delay spread of
the channel is estimated from the level-crossing rate of the frequency-selective channel
transfer function. This method enables the empirical channel characterization utilizing
simplified non-coherent measurements of the received power versus frequency.
Air-interface and multiple access scheme of an OFDM-based communications system
are proposed and investigated in part two of this work. Cumulative data rates up to
155 Mbit/s are reached under optimum channel conditions, in indoor and short range
outdoor scenarios at low mobility (pedestrian speed). Wireless LANs (local area net-
works) are a typical application for the system.
Synchronization and channel estimation algorithms are developed and evaluated, utili-
zing a known training symbol, which is periodically transmitted in the beginning of the
fixed frame structure. It has been concluded that robust and efficient synchronization
and channel estimation schemes critical tasks for an OFDM receiver are enabled
by this training symbol, at the cost of a very small overhead.
Detailed topics in synchronization include the analysis of a fine timing-offset estima-
tion algorithm over multipath channels, and the analysis of the impact of DC-offsets
and carrier feed-through on a popular frequency-synchronization scheme. A remedy is
found for the latter issue.
For the up-link, pre-equalization is suggested in a time-division duplexing (TDD)

i
ii OFDM Air-Interface Design for Multimedia Communications

scheme to pre-compensate for the frequency-selectivity of the radio channel and there-
by to simplify the data detection at the base station. Synchronization is mostly done by
the mobiles, which implies a distribution of the high complexity involved. Concepts
are presented for keeping the power of the up-link signal constant and for estimating
the remaining synchronization-offsets.
The main signal processing algorithms for the OFDM transceivers have been imple-
mented and validated on a DSP-based experimental platform, which operates in real-
time, however, at drastically downscaled data rate.
Forward error correction coding is an essential part of OFDM schemes, because fre-
quency-diversity is exploited by spreading the coded data symbols over the large
signal bandwidth. The performance of coded OFDM systems is evaluated, indicating
that increased system bandwidth and channel delay spread (the latter under certain
constraints) lead to enhanced performance. A novel antenna diversity technique is pro-
posed, which can improve the performance at low computational complexity, if the
system bandwidth and/or the channels delay spread are small.
Generally, it has been concluded that the OFDM scheme is an efficient and robust
method for transmitting data at very high rates. However, some critical hardware
issues, as for instance the linearity of amplifiers and the phase noise of local
oscillators, have to be solved.
Table of Contents

Summary .................................................................................................................... i

Table of Contents ..................................................................................................... iii

Chapter 1 General Introduction.............................................................................1

1.1 Towards Fourth Generation Mobile Systems .................................................... 1
1.2 Wideband Air-interface Design using OFDM................................................... 3
1.3 Framework and Goal of this Ph.D. Project........................................................ 6
1.4 Organization of this Thesis............................................................................... 6
1.4.1 Part I: Channel Characterization................................................................ 7
1.4.2 Part II: OFDM System Proposal and Evaluation........................................ 7
1.5 Problems Addressed in this Dissertation........................................................... 9
1.6 References...................................................................................................... 11

Part I: Channel Characterization ...........................................................................15

Chapter 2 Modeling of the Frequency-Selective Radio Channel ........................17

2.1 Introduction.................................................................................................... 17
2.2 Characterization of the Mobile Radio Channel ............................................... 18
2.2.1 Components of a Multipath Channel Model ............................................ 19
2.2.2 Definitions .............................................................................................. 20
2.2.3 Variation of Channel Parameters Due to Bandwidth Limitation .............. 24
2.3 Frequency-Domain Channel Modeling ........................................................... 26
2.3.1 The WSSUS Channel Model ................................................................... 26
2.3.2 Channel Description................................................................................ 29

iii
iv OFDM Air-Interface Design for Multimedia Communications

2.3.3 Relation to (Physical) Channel Parameters .............................................. 30

2.4 Frequency-Domain Channel Simulation ......................................................... 33
2.4.1 Model Description................................................................................... 33
2.4.2 Implementation of the Simulation Scheme .............................................. 34
2.4.3 FD-Simulation Results ............................................................................ 35
2.4.4 Differences to Time-Domain Simulation Schemes .................................. 39
2.5 Application to mm-Wave Radio Channels...................................................... 39
2.5.1 Discussion of Measurement Results ........................................................ 40
2.5.2 Discussion of Channel Parameters........................................................... 42
2.5.3 Overview of Channel Models.................................................................. 45
2.5.4 Applicability of the FD-model................................................................. 47
2.6 Conclusions.................................................................................................... 48
2.7 References...................................................................................................... 50

Chapter 3 Channel Measurement Technique based on the FD-Level Crossing

Rate...........................................................................................................................55
3.1 Introduction.................................................................................................... 55
3.2 Frequency-Domain Level Crossing Rate ........................................................ 57
3.2.1 Derivation of the LCRf from the Continuous FD-Channel Model............ 57
3.2.2 LCRf for a Deterministic Two-Ray Channel............................................ 64
3.2.3 Derivation of the LCRf for the Sampled Case.......................................... 65
3.2.4 Discussion and Summary ........................................................................ 72
3.3 Application to Channel Measurements ........................................................... 73
3.3.1 Channel Measurement Procedure ............................................................ 73
3.3.2 Validation of the Method using Measurement Results............................. 74
3.3.3 Validation by Time-domain Channel Simulations ................................... 76
3.3.4 Discussion of the Measurement Method .................................................. 79
3.4 Analysis of the Influence of Noise.................................................................. 79
3.4.1 Mathematical Modeling .......................................................................... 81
3.4.2 Derivation of LCRf from the Continuous FD-Channel Model.................. 82
3.4.3 Discrete-Frequency Analysis for Rayleigh Channels ............................... 86
3.4.4 Evaluation and Application of the Analytical Results.............................. 87
3.4.5 A Robust Measurement Procedure .......................................................... 92
3.4.6 Extended Measurement Procedure .......................................................... 96
3.5 Conclusions and Recommendations ............................................................... 96
3.6 References...................................................................................................... 97
Table of Contents v

Part II: OFDM System Proposal and Evaluation.................................................101

Chapter 4 OFDM Introduction and System Modeling......................................103

4.1 Introduction...................................................................................................103
4.2 OFDM Introduction and System Model.........................................................104
4.2.1 OFDM Introduction and Block Diagram ................................................104
4.2.2 Design of the OFDM Signal ...................................................................106
4.2.3 OFDM System Model ............................................................................109
4.2.4 Synchronization Errors...........................................................................114
4.3 Performance of an Uncoded OFDM System..................................................119
4.3.1 Mathematical Modeling .........................................................................120
4.3.2 Analytical Evaluation of the BER...........................................................121
4.3.3 Performance Results...............................................................................127
4.4 Conclusions and Recommendations ..............................................................131
4.5 References.....................................................................................................133

Chapter 5 OFDM System Proposal and Emulation System..............................135

5.1 Introduction...................................................................................................135
5.2 OFDM Based System Proposal .....................................................................136
5.2.1 Air Interface Physical Layer...................................................................137
5.2.2 Up- and Down-link Multiple Access Scheme .........................................145
5.2.3 Architecture of the Transceivers.............................................................147
5.2.4 Forward Error Correction Coding...........................................................148
5.2.5 Link Budget ...........................................................................................151
5.3 The Emulation System ..................................................................................152
5.3.1 Description of the Emulation System .....................................................154
5.3.2 Hardware Characteristics........................................................................158
5.3.3 Implementation of the Channel Simulator ..............................................160
5.4 Summary and Conclusions ............................................................................163
5.5 References.....................................................................................................165

Chapter 6 DSP Algorithm Development for the Down-Link ............................167

6.1 Introduction...................................................................................................167
6.2 Overview of Synchronization Steps...............................................................169
6.2.1 Classification of Synchronization Techniques ........................................169
6.2.2 Design of the Training Symbol...............................................................170
6.2.3 Frame Timing Synchronization ..............................................................171
6.2.4 Fractional Frequency-Offset Synchronization ........................................175
vi OFDM Air-Interface Design for Multimedia Communications

6.2.5 Application of the FFT ...........................................................................177

6.2.6 Integer Frequency-Synchronization ........................................................178
6.2.7 Remaining Timing-Offset Synchronization ............................................181
6.2.8 Sampling Frequency-Offsets ..................................................................185
6.2.9 Carrier Phase-Offset...............................................................................187
6.2.10 Summary and Conclusions .....................................................................189
6.3 Impact of DC-Offsets and Carrier Feed-Through on Fractional Frequency-
Synchronization.............................................................................................189
6.3.1 Mathematical Modeling and Definitions.................................................191
6.3.2 Analysis of the Synchronization Algorithm ............................................191
6.3.3 Extension of the Algorithm ....................................................................196
6.3.4 Analytical and Simulation Results..........................................................197
6.3.5 Conclusions and Recommendations .......................................................201
6.4 Channel Estimation .......................................................................................202
6.4.1 Wiener Filtering for Noise Reduction.....................................................203
6.4.2 Computational Results ...........................................................................206
6.4.3 Concluding Remarks ..............................................................................212
6.5 Experimental Results.....................................................................................213
6.5.1 Performance Results for the Full-Rate Receiver .....................................214
6.5.2 Performance Results for the Quarter-Rate Receiver................................218
6.5.3 Discussion of the Channels Time-Variability ........................................220
6.5.4 Conclusions and Recommendations .......................................................222
6.6 References.....................................................................................................223

Chapter 7 DSP Algorithm Development for the Up-Link.................................227

7.1 Introduction...................................................................................................227
7.2 Pre-Equalization in OFDM............................................................................228
7.2.1 OFDM System Model ............................................................................228
7.2.2 Impact of Synchronization Errors...........................................................228
7.2.3 Channel Reciprocity and Pre-Equalization .............................................229
7.2.4 Is the Channel Reciprocal? .....................................................................229
7.3 Power Limiting Strategies for Pre-equalization .............................................230
7.3.1 Phase Pre-Equalization for Phase-Modulation Schemes .........................230
7.3.2 Phase and Magnitude Pre-Equalization with Power Limiting .................231
7.4 Performance Results......................................................................................234
7.4.1 Application of Phase Pre-Equalization for QPSK ...................................235
7.4.2 Application of Pre-Equalization with Power Limiting ............................235
7.5 Synchronization Parameter Estimation on the Up-link...................................237
Table of Contents vii

7.5.1 Magnitude of the Constellation Values...................................................238

7.5.2 Estimation of the Up-link Timing-Offset................................................239
7.6 Experimental Results.....................................................................................244
7.6.1 Performance over Different Radio Channels ..........................................244
7.6.2 Impact of the I/Q-Modulator and -Demodulator .....................................245
7.7 Conclusions...................................................................................................246
7.8 References.....................................................................................................248

Chapter 8 Performance Evaluation and Enhancement of COFDM.................249

8.1 Introduction...................................................................................................249
8.2 Performance of a Coded OFDM System........................................................250
8.2.1 Review of the Concept of Effective Eb/N0 ..............................................251
8.2.2 Assessment of the Concept of Effective Eb/N0 ........................................256
8.2.3 PDF of the Effective Eb/N0 .....................................................................258
8.2.4 Performance Results and Discussion ......................................................262
8.2.5 Summary................................................................................................268
8.3 Performance Enhancement using Antenna Diversity .....................................270
8.3.1 Antenna Diversity for OFDM Using Cyclic Delays................................271
8.3.2 Even/odd Sub-carrier Transmitter Diversity ...........................................275
8.3.3 Performance ...........................................................................................277
8.3.4 Discussion of the Diversity Schemes......................................................280
8.4 Conclusions and Recommendations ..............................................................282
8.5 References.....................................................................................................283

Chapter 9 Conclusions and Recommendations .................................................285

9.1 Part I: Channel Characterization ....................................................................285
9.1.1 Modeling of the Frequency-Selective Radio Channel .............................285
9.1.2 Channel Measurement Technique using the FD-Level Crossing Rate.....286
9.2 Part II: OFDM System Proposal and Evaluation............................................287
9.2.1 OFDM System Modeling .......................................................................287
9.2.2 OFDM Air-Interface and Multiple Access Scheme Proposal ..................288
9.2.3 Issues in utilizing the 60 GHz Frequency-Band ......................................289
9.2.4 Synchronization and Channel Estimation on the Down-Link..................289
9.2.5 Pre-Equalization for the Up-Link ...........................................................291
9.2.6 Emulation System and OFDM Implementation ......................................292
9.2.7 Performance Evaluation and Enhancement of COFDM..........................293

Appendix A Correlation Coefficient for the Discrete Impulse Response .........295

viii OFDM Air-Interface Design for Multimedia Communications

A-1 Definitions ....................................................................................................295

A-2 Calculation of the Correlation Coefficient .....................................................296
A-3 References.....................................................................................................297

Appendix B FD-Level Crossing Rate in the Presence of Noise .........................299

B-1 Derivation of the FD-Level Crossing Rate.....................................................299
B-2 Approximation 1 ...........................................................................................300
B-3 Approximation 2 ...........................................................................................301
B-4 References.....................................................................................................302

Appendix C Analysis of Fine Timing-Offset Estimation ...................................303

C-1 Review of Estimation Technique and System Model.....................................303
C-2 Estimation Bias over Dispersive Channels.....................................................305
C-3 Estimation Variance on the AWGN Channel.................................................306
C-4 Estimation Variance on Rayleigh Fading Channels .......................................308
C-5 References.....................................................................................................310

Appendix D PDF of Wide-band Average Received Power................................311

D-1 PDF of Average Received Power for Rayleigh Fading Channels...................311
D-1.1 Approximations......................................................................................313
D-1.2 Analytical and Simulation Results..........................................................315
D-2 PDF of Average Power for Ricean Fading Channels .....................................318
D-2.1 Approximations......................................................................................319
D-2.2 Analytical and simulation results............................................................320
D-3 References.....................................................................................................324

List of Acronyms ....................................................................................................325

Publications by the Author ....................................................................................329

Journal Papers ......................................................................................................329
Conference Papers ................................................................................................329
Reports .................................................................................................................331
Relation to this Thesis ..........................................................................................331

Samenvatting..........................................................................................................333

Acknowledgements.................................................................................................335

Curriculum Vitae ...................................................................................................337

Chapter 1 General Introduction

1.1 Towards Fourth Generation Mobile Systems

The expected convergence of the internet and mobile telephony fuels major research
and development efforts in the telecommunications industries. Third generation (3G)
cellular systems, called IMT-2000 (International Mobile Telecommunications in the
year 2000) or UMTS (Universal Mobile Telecommunications System), are currently
deployed to meet this demand, supporting data rates up to 2 Mbit/s for local coverage
and at least 144 kbit/s for wide-area coverage [1][3].
But will those systems become as successful as their ancestors, particularly the (digi-
tal) second-generation systems like GSM (Global System for Mobile Communica-
tions), which for many years largely exceeded the expected growth rates, reaching
penetration factors of well above 50 % in most western European countries? Are the
new services offered sufficient motivation to buy new, in the beginning certainly ra-
ther expensive, mobile phones? Will the enormous cost of rolling-out such systems
including license fees ever be amortized?
In my opinion, mobile telephones are the prerequisite to achieve large penetration
rates with internet services. Many disadvantages of current internet access methods are
eliminated: there is no need to buy or possess a computer, to cable a modem, to sub-
scribe with an internet service provider, to go through lengthy start-up and log-in pro-
cedures, etc. Moreover, niche-time can be used to browse the Web; for instance the
time spent on public transport, commuting to ones work. Advanced services can make
effective use of the mobility aspect: Location-based services can provide the user with
information related to his whereabouts (restaurant guide, cultural program, etc.), which
is most useful in an unknown city but can be interesting in ones home-town as well.
For the mass-market, main ingredients for a successful deployment of 3G systems are:
cheap, easy-to-use terminals,

1
2 Chapter 1 General Introduction

large color displays, and

attractive billing schemes (flat rates).
The huge success of the Japanese i-mode system, which implements the above-men-
tioned features (however, providing low data rates compared to UMTS), proves the
interest of the consumer in mobile internet services [4]. On the other hand, voice tele-
phony (possibly augmented by still images and video) will remain an important appli-
cation in future systems, where enhanced capacity and spectral efficiency need to be
achieved.

While the roll-out of 3G systems is under progress, research activities on the fourth
generation (4G) have already started [5][9]. At the time being, however, there is no
clear vision which ingredients will define this future system generation. Certainly,
transmission rates will be further increased bit rates in the order of 100 Mbit/s are
considered , but many doubt that it will ever be feasible (affordable) to provide such
data rates with nation-wide coverage. For local coverage, on the other hand, current
wireless local area networks (W-LAN) standards can already provide data rates up to
54 Mbit/s. (Those W-LAN standards are: IEEE 802.11a in the USA, HIPERLAN/2 in
Europe, and MMAC in Japan [10], [11]).
Therefore, a popular vision suggests to combine W-LAN systems for high peak data
rates with cellular systems (GSM, UMTS) for wide area-coverage, and to allow inter-
system handovers [12]. Technical aspects of air-interface standards for multi-standard
terminals supporting W-LAN and cellular technologies are discussed in [13]. How-
ever, the more important factor for the user may be the simple fact that a W-LAN can
be installed and operated by the user, free of cost for subscription and call-charges.
That is, the user may own a part of the system infrastructure, and eventually pro-
vided the required billing mechanisms are in place even charge foreign terminals
for accessing the system at his premises. This factor may become a key-ingredient of
4G systems. Other challenges to be solved in order to realize multi-standard systems
include hardware issues for terminals supporting multiple air-interface standards, the
above mentioned inter-system handovers, billing aspects, and security/privacy.
But is such an integration of systems enough reason to speak of a new system genera-
tion, particularly if only current air-interface standards are considered?
The scenarios weakly supported by current technology point out some limitations. Ad-
hoc networking for instance, where a number of terminals form a small wireless net-
work passing on information from node-to-node without the aid of an access point or a
base station, is a concept that will become increasingly important [14]. Bluetooth,
intended as a cable-replacement [15], supports this idea at somewhat limited bit-rates
up to about 800 kbit/s and at very limited ranges of a few meters. At higher rates, the
IEEE 802.11 W-LAN standard is considered by many as an enabling technology for
ad-hoc networks [16]. Serious problems are encountered, however, when its current
multiple access control (MAC) protocol is applied in such systems [16].
1.2 Wideband Air-interface Design using OFDM 3

A wireless digital recording or television studio requires the support of multiple con-
stant-rate data streams at specified bit-error-rates and low delay a scenario that is
very different to the previous one. A centralized, scheduled MAC could be most effi-
cient for such applications, which may also be supported by 4G systems.
Last but not least, a real 4G air-interface to be developed may support data rates in
the order of 100 Mbit/s at full mobility, i.e., at velocities up to 200300 km/h [5][9],
[12]. The term Mobile Broadband Systems (MBS) refers to this type of technology, in
a number of references [5][7].

In this thesis, radio air-interface technology for future wide-band communications sys-
tems is studied, starting with the mathematical modeling of the fading radio channel.
This work has been motivated on the one hand by the huge potential market of wide-
band communications systems, particularly of wireless LANs, and on the other hand
and more importantly , by the technological challenge of developing air-interfaces for
transmitting such large data rates over the hostile mobile radio channels. The follow-
ing section elaborates on the technology aspect.

1.2 Wideband Air-interface Design using OFDM

Multipath propagation is the primary issue in the air-interface design for wideband
(high data-rate) communications systems: multiple replicas of the transmitted signal
arrive at the receiver with various propagation delays, due to reflections on all kinds of
objects and obstacles in the environment. Therefore, if a high-rate data stream is trans-
mitted on such a channel, multiple data symbols interfere with each other, making the
data recovery difficult. At 155 Mbit/s, for instance, the symbol period is about 13 ns
using QPSK (quadrature phase shift keying) and neglecting error correction coding.

Magnitude Magnitude

Delay time Frequency

(a) (b)
Figure 1-1: (a): Impulse response of a multipath radio channel. (b): Frequency-selective
channel transfer function. The Fourier transform relates the impulse response and
the transfer function.
4 Chapter 1 General Introduction

Channel Signal bandwidth

impulse response
Channel
Symbol period transfer function

frequency

time (a)
Symbol period
Sub-carrier
bandwidth

frequency

time (b)

Figure 1-2: Symbol period and signal bandwidth in comparison with the channel impulse res-
ponse and channel transfer function. (a): Single carrier system (serial transmis-
sion); (b): Multi-carrier system (OFDM) with parallel sub-channels.

This value can be compared with the (relative) delay of a reflected path with, say,
30 m path length difference, being 100 ns, which demonstrates that the combination of
multiple data symbols is observed at the receiver at any given time instant. This
phenomenon is called inter-symbol-interference (ISI). Receivers have to eliminate
the ISI. Mathematically, the influence of the channel can be described as a convolution
of the transmitted signal by the channel impulse response depicted in Figure 1-1a.
Figure 1-2a illustrates the relation of the symbol period and the time-extent of the
channel impulse response for a serial wideband transmission system.
Another characteristic property of a multipath radio channel is the frequency-selecti-
vity of its transfer function (TF), as shown in Figure 1-1a. (The TF is the response of
the channel to a narrow-band signal as a function of the frequency.) It is noted that the
TF is the Fourier transform of the channel impulse response. The comparison of the
signal bandwidth of the serial data stream and the channel transfer function demon-
strates that a wide-band signal gets distorted when it is transmitted over such a channel
(see Figure 1-2a).
1.2 Wideband Air-interface Design using OFDM 5

M agnitude

Frequency

sub-carrier
frequency spacing

Figure 1-3: The overlapping spectra (sinc-functions) of four adjacent OFDM sub-carriers. At
one sub-carriers center frequency, all other spectra are zero, demonstrating the
sub-carrier orthogonality.

Equalization is the standard method to combat inter-symbol-interference in a single-

carrier system [17]. Equalizers consist of linear filters whose purpose is to combine the
signal components arriving at various delay times. The main challenge is to adapt the
filter coefficients to the time-variant channel conditions. The methods for achieving
this adaptation are computationally extremely demanding, particularly if long filters
are required as in our case, where the channel impulse response typically spans many
data symbols.

Orthogonal frequency division multiplexing (OFDM) can drastically simplify the

equalization problem [11], [18][23]. In OFDM, the high-rate serial data stream is
split up into a number (several dozens up to a few thousand) of parallel data streams at
a much lower (common) symbol rate, which are modulated on a set of sub-carriers
(frequency division multiplexing). Figure 1-2b illustrates the parallel transmission
principle. High spectral efficiency is achieved by selecting a specific (orthogonal) set
of sub-carrier frequencies. Inter-carrier-interference is avoided due to the orthogona-
lity, although the spectra of the sub-carriers actually overlap (see Figure 1-3). The idea
is to make the symbol period long with respect to the channel impulse response in or-
der to reduce ISI. This implies that the bandwidth of the sub-carriers gets small (with
respect to the channels coherence bandwidth [17]), thus the impact of the channel is
reduced to an attenuation and phase distortion of the sub-carrier symbols (flat fa-
ding), which can be compensated by efficient one-tap equalization.

Digital signal processing is used to generate a complex-valued baseband signal contai-

6 Chapter 1 General Introduction

ning all OFDM sub-carrier signals. (The block diagram of an OFDM transmission
chain is shown in Chapter 4.) Next to the correct estimation of the channel transfer
function, which is required for equalization, the following problems require particular
attention in the implementation of OFDM modems:
Linear power amplifiers are needed at the transmitter (and pre-amplifiers at the re-
ceiver) to avoid any distortion of the OFDM signal. Distortion would lead to inter-
carrier-interference (ICI), implying performance degradation.
At the receiver, synchronization issues are of prime importance. The start of the
OFDM symbols and the exact location of the sub-carrier frequencies have to be
found to be able to recover the data symbols (time- and frequency-synchroniza-
tion).
Phase noise of any mixer oscillators also yields ICI and thus has to be
avoided/minimized.

1.3 Framework and Goal of this Ph.D. Project

The main topic of this dissertation is the design of the air-interface of an OFDM-
based, wide-band mobile communications system for indoor and low-range, low-mo-
bility outdoor scenarios. Wireless computer networks (W-LAN) are the intended appli-
cation of such systems, enabling wireless multimedia communications. This work was
conducted under the framework of a cooperative research program between Delft Uni-
versity of Technology (Delft, The Netherlands) and Korea Telecom (Seoul, South Ko-
rea). Funding was provided by Korea Telecom.
The emphasis of this study lay on the physical radio link. Wireless transmission of
ATM (asynchronous transfer mode) cells should be supported at peak data rates of 155
Mbit/s, according to the target specifications of the air-interface multiple access
scheme under development. The 60 GHz mm-wave frequency band was considered,
mainly because the required bandwidth in the order of several hundreds of MHz is
only available at these frequencies [24]. The mathematical modeling of the radio chan-
nel was the first goal of this project. Another important aspect was the implementation
and validation of the proposed system on a hardware platform to be developed.

A number of students have contributed to the project and to this Ph.D. thesis, working
on their graduation projects or on internships [25][35].

1.4 Organization of this Thesis

This Ph.D. thesis is divided in two parts. In Part I, the wide-band radio channel is
modeled and investigated, while the OFDM system design is presented in Part II.
1.4 Organization of this Thesis 7

1.4.1 Part I: Channel Characterization

The radio channel model is probably the most important tool for the design of a com-
munications system. It has to appropriately describe the relevant properties of the phy-
sical channel, and it should be suitable for computer simulations and analytical studies
of the system under development. As in OFDM information is transmitted over a set of
parallel sub-carriers, the main aspect of interest is the frequency-selectivity of the
channel transfer function caused by multipath propagation.
The frequency-domain channel model proposed and studied in Chapter 2 of this dis-
sertation directly describes this property by characterizing the variations of the transfer
function via second-order stochastic properties. The models behavior is compared to
physical propagation mechanisms and its parameters are related to parameters of the
radio channel. Throughout this thesis, the model is applied for the design and evalua-
tion of the OFDM system.

In Chapter 3, the frequency-domain level crossing rate (LCRf) of the channel transfer
function is analyzed. The LCRf specifies the number of up-going level crossings over a
given threshold per unit of bandwidth. Originally, the motivation for this study was to
derive parameters equivalent to the (time-domain) level crossing rate and average fade
duration [36], which are considered to be useful in the development and evaluation of
mobile communications systems [37].
The most relevant application, however, was found in a slightly different field. It was
discovered in this Ph.D. research that the LCRf can be used for estimating the RMS
delay spread of the channel, which is the most significant single parameter for charac-
terizing the channels time-dispersion (and frequency-selectivity). This relationship
enables channel measurements using a rather simple measurement setup, because it is
sufficient to scan the power transfer function of the channel versus frequency to
determine the LCRf. The novel measurement technique is extensively studied in
Chapter 3.

1.4.2 Part II: OFDM System Proposal and Evaluation

Chapter 4 gives a brief introduction to the OFDM transmission technique. System
models are derived for the analysis of various aspects of OFDM. In Chapter 4, for in-
stance, they are used to evaluate the bit-error-rate (BER) performance of an uncoded
OFDM system, considering various modulation- and detection schemes, and different
channel conditions. The results will serve as benchmarks for the evaluation of the
OFDM receivers investigated in this thesis.

The air-interface and multiple access scheme of a novel OFDM-based wide-band com-
munications system are described in Chapter 5. The system supports the transmission
of single asynchronous transfer mode (ATM) cells at bit-rates up to 155 Mbit/s. To ef-
ficiently transmit such short data packets (one ATM cells consists of just 53 bytes) at
8 Chapter 1 General Introduction

defined quality of service requirements, a fixed (but potentially asymmetric) frame

structure is employed with time-division duplexing (TDD). Training symbols periodi-
cally transmitted on the down-link are used by the mobile for synchronization and
channel estimation, while pre-equalization is considered for the up-link to enable cohe-
rent detection at the base station without introducing additional training symbols.
Chapter 5 also describes a digital signal processor (DSP) based experimental platform,
the so-called emulation system, which has been developed for demonstrating the
OFDM air-interface. All parameters have been largely downscaled to decrease the re-
quired processing speed (and to thereby simplify the software and hardware develop-
ment), enabling the implementation of the transmitter and the receiver on single (but
separate) DSP chips. A third DSP is used to simulate the multipath radio channel. Ana-
log hardware performs in-phase/quadrature (I/Q) -modulation (to an intermediate fre-
quency) and -demodulation, introducing frequency-offsets. Thus, real-time frequency-
synchronization algorithms can be demonstrated. All system components are intercon-
nected by analog, complex-valued baseband signals (I/Q-signals).

In Chapter 6, the signal processing steps for the down-link are investigated and evalu-
ated. The estimation of synchronization parameters and of the channel transfer func-
tion are thoroughly described, utilizing the periodically transmitted training symbol.
Novel contributions in this chapter concern the analysis of a highly accurate timing-
synchronization scheme, and the investigation of the impact of DC-offsets and carrier
feed-through on a popular class of frequency-synchronization techniques. The latter
study leads to an extension of the technique with improved robustness against these
impairments.

To save transmission power, to enhance the spectral efficiency, and to simplify the
symbol detection, pre-equalization has been proposed for the up-link of the OFDM
system. That is, the up-link symbols are pre-distorted using channel knowledge from
the down-link, in order to compensate for the phase rotations and attenuations of the
data symbols, introduced by the multipath radio channel. Channel reciprocity and slow
time-variability are assumed. In Chapter 7, a number of basic issues of this principle
are investigated, as for instance synchronization steps and techniques for limiting the
transmitted power on the up-link. Moreover, it is discussed whether the channel reci-
procity can be exploited for pre-equalization as proposed.

Forward error correction coding is a crucial component of most OFDM systems. Er-
rors caused by the frequency-selective channel on severely attenuated sub-carriers can
be corrected using the reliable data of strong(er) sub-carriers. That is, the frequency-
diversity of the wide-band radio channel is exploited.
In Chapter 8, the performance of coded OFDM systems is evaluated using the con-
cept of effective Eb/N0 [38]. In this method, the fading pattern of the radio channel is
converted to a scalar value, the effective Eb/N0, which quantifies the signal-to-noise
1.5 Problems Addressed in this Dissertation 9

ratio on an AWGN channel resulting in equivalent error rates. Channel simulations

have to be performed to generate realistic fading patterns that can then be transformed
to error rate results in this way. A novel extension to this concept is presented in this
thesis. The probability density function of the effective Eb/N0 is related to channel pa-
rameters, allowing for the analytical computation of average error rates and outage
probabilities.
Novel antenna diversity schemes are proposed in the second part of Chapter 8 to en-
hance the performance in cases where the frequency-diversity of the channel is small
(flat fading channels).

General conclusions and recommendations are summarized in Chapter 9.

1.5 Problems Addressed in this Dissertation

The following problems are analyzed in this Ph.D. thesis:
Modeling of the frequency-selective radio channel
Characterization of frequency-selective, Ricean fading radio channels
Modeling of the frequency-selective channel using second order statistics of
complex Gaussian random processes
Simulation of the frequency-selective channel transfer function in the frequen-
cy-domain
Application of the channel model to mm-wave radio channels
Selection of typical parameters for 60 GHz indoor and outdoor channels
Measurement of channel parameters based on the frequency-domain level crossing
rate (LCRf) of the channel transfer function
Derivation of the LCRf of a frequency-selective Ricean channel from its second
order stochastic model
Impact of channel parameters on the LCRf
Impact of the frequency-domain sampling interval (for Rayleigh channels)
Estimation of the RMS delay spread (a measure for the length of the channel
impulse response) from the LCRf
Independence of the LCRf from the channel impulse response for Rayleigh
channels with given RMS delay spread
Analysis of the influence of noise on the RMS delay spread estimation tech-
nique
OFDM basics
Consideration of hardware aspects in the design of the OFDM system
10 Chapter 1 General Introduction

Modeling of an idealized OFDM system and the impact of (small) synchroniza-

tion errors
Bit-error-rate performance evaluation of an uncoded OFDM system over
Ricean channels using several modulation schemes and coherent/differential
detection
Proposal of an OFDM-based air-interface and multiple access scheme for the trans-
mission of ATM cells at data rates up to 155 Mbit/s
Discussion of the link budget of the system at 60 GHz
Development of a DSP-based emulation system for the demonstration of the air-
interface
Analysis of important hardware characteristics
Development of a channel simulator
Development and analysis of synchronization algorithms for OFDM receivers
Design of a training symbol
Development and analysis of synchronization steps for time, carrier-frequency,
sampling-frequency, and carrier-phase synchronization
Implementation of the synchronization algorithms on the emulation system
Performance analysis of a timing-offset estimation scheme in Ricean channels
Analysis of the impact of DC-offsets and carrier feed-through on a frequency
synchronization scheme
Proposal of an enhanced frequency-synchronization scheme that is robust
against DC-offsets and carrier feed-through
Development and analysis of a computationally efficient channel estimation
scheme using the training symbol
Reduction of the computational complexity of a Wiener filter used for minimiz-
ing the estimation error
Optimization of a fixed filter for various channel conditions
Performance evaluation in time-variant channels using the emulation system
Simulation of a channel prediction scheme for performance enhancement in
time-variant channels
Proposal and analysis of a pre-equalization technique for the up-link of a time-divi-
sion-duplex OFDM system
Proposal and evaluation of transmit power limitation strategies required at the
mobile
Development and evaluation of synchronization algorithms for the up-link
Implementation of the pre-equalization scheme and the synchronization algo-
1.6 References 11

rithms on the emulation system

Performance evaluation of convolutionally coded OFDM systems (with bit-level
interleaving) using the concept of effective Eb/N0 [38]
Assessment of the concept of effective Eb/N0
Application to OFDM systems
Modeling the probability density function of the effective Eb/N0 for given chan-
nel- and OFDM system parameters
Proposal and evaluation of antenna diversity techniques for the OFDM transmitter
and/or receiver

1.6 References
[1] IEEE Personal Communications, Special Issue on IMT-2000: Standards Efforts
of the ITU, vol. 4, no. 4, Aug. 1997.
[2] T. Ojanper and R. Prasad, An Overview of Air Interface Multiple Access for
IMT-2000/UMTS, IEEE Communications Magazine, vol. 36, no. 9, pp. 8295,
Sept. 1998.
[3] M. W. Oliphant, The Mobile Phone Meets the Internet, IEEE Spectrum, vol.
36, no. 8, pp. 2028, Aug. 1999.
[4] K. Enoki, i-mode: the mobile Internet service of the 21st century, in Proc.
ISSCC 2001 (Solid-State Circuits Conference), 2001, pp. 1215.
[5] L. M. Correia and R. Prasad, An Overview of Wireless Broadband Communi-
cations, IEEE Communications Magazine, vol. 35, no. 1, pp. 2833, Jan. 1997.
[6] M. Dinis and J. Fernandes, Provision of Sufficient Transmission Capacity for
Broadband Mobile Multimedia: A Step Toward 4G, IEEE Communications
Magazine, vol. 39, no. 8, pp. 4654, Aug. 2001.
[7] M. Prgler, C. Evci, and M. Umehira, Air Interface Access Schemes for Broad-
band Mobile Systems, IEEE Communications Magazine, vol. 37, no. 9, pp.
106115, Sept. 1999.
[8] H. Rohling, R. Grnheid, and D. Galda, OFDM Air Interface for the 4th Genera-
th
tion of Mobile Communication Systems, in Proc. 6 international OFDM-
Workshop (InOWo01), Hamburg, Sept. 2001, pp. 0-10-28.
[9] NTT DoCoMo, The Path to 4G Mobile, IEEE Communications Magazine, vol.
39, no. 3, pp. 3841, March 2001 (Advertisement).
[10] R. van Nee, G. Awater, M. Morikura, H. Takeshi, M. Webster, and K. W. Hal-
ford, New High-Rate Wireless LAN Standards, IEEE Communications Maga-
zine, vol. 37, no. 12, pp. 8288, Dec. 1999.
12 Chapter 1 General Introduction

[11] R. van Nee and R. Prasad, OFDM for Wireless Multimedia Communications.
Boston: Artech House, 2000.
[12] L. M. Pereira, Fourth Generation: Now, it is personal!, in Proc. PIMRC 2000
(11th International Symposium on Personal Indoor Mobile Radio Communica-
tions), London, Sept. 2000, pp. 10091016.
[13] J. Kalliokulju, P. Meche, M. J. Rinne, J. Vallstrm, P. Varshney, and S.-G.
Hggman, Radio Access Selection for Multistandard Terminals, IEEE Com-
munications Magazine, vol. 39, no. 10, pp. 116124, Oct. 2001.
[14] S. Giordano and W. W. Lu, Challenges in Mobile Ad Hoc Networking, IEEE
Communications Magazine, vol. 39, no. 6, p. 129, June 2001 (Guest Editorial).
[15] R. Schneiderman, Bluetooths Slow Dawn, IEEE Spectrum, vol. 37, no. 11, pp.
6165, Nov. 2000.
[16] S. Xu and T. Saadawi, Does the IEEE 802.11 MAC Protocol Work Well in
Multihop Wireless Ad Hoc Networks? IEEE Communications Magazine, vol.
39, no. 6, pp. 130137, June 2001.
[17] J. G. Proakis, Digital Communications, 3rd ed. New York: McGraw Hill, 1995.
[18] R. Prasad, Universal Personal Communications. Boston: Artech house, 1998, ch.
10.
[19] O. Edfors, M. Sandell, J. J. van de Beek, D. Landstrm, F. Sjberg, An
Introduction to Orthogonal Frequency-Division Multiplexing, Research Report
TULEA 1996:16, Division of Signal Processing, Lule University of Technology,
http://www.sm.luth.se/ csee/sp/publications.html.
[20] M. Speth, S. A. Fechtel, G. Fock, and H. Meyr, Optimum Receiver Design for
Wireless Broad-Band Systems Using OFDMPart I, IEEE Trans. Commun.,
vol. 47, no. 11, pp. 16681677, Nov. 1999.
[21] S. B. Weinstein and P. M. Ebert, Data Transmission by Frequency-Division
Multiplexing Using the Discrete Fourier Transform, IEEE Trans. Commun.
Techn., vol. COM-19, no. 5, pp. 628634, Oct. 1971.
[22] J. A. C. Bingham, Multicarrier Modulation for Data Transmission: An Idea
Whose Time has Come, IEEE Communications Magazine, pp. 514, May 1990.
[23] L. J. Cimini, Analysis and Simulation of a Digital Mobile Channel using
Orthogonal Frequency Division Multiplexing, IEEE Trans. Commun., vol.
COM-33, no. 7, pp. 665675, July 1985.
[24] P. F. M. Smulders, Broadband Wireless LANs: A Feasibility Study. PhD Thesis,
Eindhoven University of Technology, Eindhoven, The Netherlands, 1995.
1.6 References 13

[25] V. Tutucu, Channel Estimation for OFDM System in Multipath Fading

Environments for Wireless Broadband Communications, M.Sc. Thesis, IRCTR
S-028-98, Delft University of Technology, Sept. 1998.
[26] J. Purwaha, Wide-band Channel Measurements at 60 GHz in Indoor Environ-
ments, M.Sc. Thesis, Delft University of Technology (IRCTR ), Aug. 1998.
[27] R. el Hattachi, Measurements and Modeling of the 18 GHz Radio Channel,
M.Sc. Thesis, IRCTR S-029-98, Delft University of Technology, Aug. 1998.
[28] P. Teneva, Pseudo Real-time Simulation of an OFDM System for Wireless
Broadband Communications: OFDM Transmitter, Research Report, IRCTR,
Delft University of Technology, Jan. 1999.
[29] K. Bke, DSP Manual, Research Report, IRCTR S-025-99, Delft University of
Technology, July, 1999.
[30] K. Bke, Assessment of OFDM Based Air-interface Techniques Using an
Emulation Platform: Investigation and Implementation of OFDM Synchroniza-
tion Algorithms, M.Sc. Thesis, IRCTR S-001-00, Delft University of Technolo-
gy, Jan. 2000.
[31] I. Gultekin, DSP Software Implementation for a Broadband Air-interface
Emulation Platform: Onderzoek naar de serile interace tussen de PC an de DSP
board en het maken van een GUI voor de DSP applicaties, Graduation Thesis,
Haagse Hogeschool, June 2000 (in Dutch).
[32] K. S. Lidsheim, A survey of peak-to-average power reduction methods for the
OFDM transmission schemes, Research Report KWATT, IRCTR S-024-00,
Delft University of Technology, Sept. 2000.
[33] A. Snijders, Emulator: De testopstelling voor OFDM, Graduation Thesis,
IRCTR S-030-00, Technische Hogeschool Rijswijk, Nov. 2000 (in Dutch).
[34] D. Murargi, Channel Estimation Enhancement in OFDM Systems for Wireless
Multimedia Communications, M.Sc. Thesis, IRCTR S-028-98, Delft University
of Technology and Technical University of Lisbon, July 2001.
[35] G. Landman, Frequency Domain Study of the Wide-Band Mobile Propagation
Channel, M.Sc. Thesis, Delft University of Technology (IRCTR), Aug. 2001.
[36] W. C. Jakes Jr., Microwave Mobile Communications. New York: Wiley-Inter-
science, 1974.
[37] J. D. Parsons, The mobile radio propagation channel. New York: Wiley-Inter-
science, 2000.
[38] S. Nanda and K. M. Rege, Frame error rates for convolutional codes on fading
channels and the concept of effective Eb/N0, IEEE Trans. Veh. Technol., vol. 47,
no. 4, pp. 12451250, Nov. 1998.
Part I: Channel Characterization

15
Chapter 2 Modeling of the Frequency-
Selective Radio Channel

2.1 Introduction
The channel model is the bread and butter for the telecommunications engineer
this is how Ramjee Prasad, former Professor for Mobile Communications at Delft
University of Technology, used to emphasize in his lectures the importance of the
channel model for designing radio interfaces for wireless communications systems.
And this was not just a phrase to keep the students attention. The channel models
really are the foundation, mobile communications systems are built on.

As the main topic of this thesis is the design of OFDM air-interfaces, the goal of this
chapter is the description and discussion of an appropriate channel model for such
systems. This model must allow for
analytical treatment of OFDM related problems and for
efficient computer simulation schemes,
to address two general requirements. According to the systems key specifications, it
should fit to physical radio channels in the mm-wave frequency band, for indoor (in-
room) and short-range outdoor environments.

Radio propagation in a mobile radio channel is determined mainly by its multipath

nature. Multiple reflections, and sometimes a line-of-sight (LOS) component of the
transmitted signal arrive at the receiver via different propagation paths and therefore
with different amplitudes and delay times. As an effect of this, the narrowband-
received power fluctuates dramatically when observed as a function of location (or
time) and frequency. In the early days of mobile systems, the communications

17
18 Chapter 2 Modeling of the Frequency-Selective Radio Channel

engineer was mainly interested in the time-variability of narrowband channels, which

were thus studied extensively (see e.g. [1]). By that time, transmission bandwidths
were small, thus flat-fading was a reasonable assumption. As the systems evolved,
demand for higher transmission rates has been increasing, making the channels time
dispersion (which is equivalent to its frequency-selectivity) a major issue.
In OFDM, the channels variability in the frequency-domain (FD) has a similar role as
the time-variance in a (flat-fading) narrowband system. Usually, the channel can be
assumed to be static during the transmission of at least one OFDM symbols. In indoor
wireless local area networks (WLAN), the channel is even considered quasi stationary
during up to a whole data packet or frame period.

The following section reviews the propagation mechanisms that have to be

characterized by the channel model. The main properties of interest for the OFDM
system design are emphasized and important channel parameters are defined.
The so-called frequency-domain channel model (FD-model) is proposed and analyzed
in Section 2.3. The model describes the frequency-selective fading by the delay power
spectrum (DPS) of the channel, the Fourier transform of the spaced-frequency
correlation function [2]. This approach is dual to defining the time-variability by the
Doppler power spectrum [3], which is often referred to as Jakes fading model [1], [4].
Expressions are given, relating the DPS (being specified by just two to four
parameters) to the most important physical channel parameters. Rayleigh and Ricean
fading channels are considered.
A direct implementation of the FD-channel model in a computer simulation scheme is
proposed in Section 2.4. The outputs of this simulator are (complex-valued)
frequency-selective channel transfer functions. The differences are emphasized
between this approach and (conventional) time-domain simulators, which generate
channel impulse responses.
Section 2.5 summarizes the basic results of a number of measurement campaigns.
Some of them were performed at Delft University of Technology, others were found in
the literature. We elaborate on the suitability of the proposed channel model for
describing the radio channels investigated.
Conclusions and recommendations are given in Section 2.6.

2.2 Characterization of the Mobile Radio Channel

This section starts with a qualitative description of the main propagation mechanisms
resulting from multipath wave interference. Section 2.2.2 reviews mathematical
definitions that are useful for the characterization of wide-band, frequency-selective,
mobile radio channels. Important channel parameters are introduced and their physical
interpretation is developed.
2.2 Characterization of the Mobile Radio Channel 19

2.2.1 Components of a Multipath Channel Model

For the mathematical description of a multipath radio channel, it is convenient to
distinguish three mechanisms; namely:
path loss,
shadowing, and
multipath interference.
The former two are described by large-scale channel models, which essentially
provide information about the average received power at a certain location. Path loss
strictly describes the dependency of this average power on the distance between
transmitter and receiver, while shadowing accounts for the fluctuations observed at a
fixed distance, due to geometric features of the propagation environment. These
fluctuations occur for instance because of the blocking of relevant propagation paths,
e.g., the line-of-sight component, as the mobile moves around.
Highly sophisticated large-scale models that typically employ geographical
information system (GIS) databases in order to account for topographical features are
incorporated in the cell-planning tools used by mobile system operators. In indoor
environments, ray-tracing models are often used to predict the received signal strength
at a given location.

Such models are not relevant, however, for the design of new transmission techniques.
For this application, the description of the effects of multipath interference is required,
since the air-interface has to cope with them. These effects are often referred to as
small-scale fading.
Small-scale models are valid within (small) local areas, where the signal fluctuations
due to shadowing and path loss can be neglected. The dimension of such a local area is
therefore limited to approximately 5 40, where is the wavelength of the radio
frequency (RF) carrier. (Due to the small wavelengths below one centimeter, this
range may be even larger in the mm-wave band.)
The channel model investigated in this chapter is limited to the description of small-
scale effects. A set of average parameters specifies the channels behavior within a
local area. These parameters are the normalized received power1, P0 , the Ricean K-
factor, K, and the RMS delay spread (RDS), rms. Note, however, that each realization
obtained from the model has varying instantaneous parameters denoted {P0 , K ,rms } ,
since the model is a stochastic one. (To be specific, it is a Gaussian wide-sense sta-
tionary uncorrelated scattering (WSSUS) model, as shown in Section 2.3, and [5], [6]).
The amount of variation of these parameters from the local-area parameters depends in

1
The (dimensionless) normalized received power is defined as the ratio of the received power Prx
and the transmitted power Ptx. Equivalently, the absolute received power Prx [W] could be used for
channel description.
20 Chapter 2 Modeling of the Frequency-Selective Radio Channel

particular on the observed bandwidth. When the bandwidth is much greater than the
coherence bandwidth, then the multipath is completely resolved and the channel para-
meters vary little, since the individual multipath amplitudes do not change rapidly
within a local area. However, if the system is narrowband, then multipath is not re-
solved, and the path amplitudes at each resolvable (delay) time-bin (being spaced by
the reciprocal of the bandwidth) vary due to multipath interference. This leads to the
fluctuation of the instantaneous channel parameters within the local area (cf. [6], [7]).
The mathematical definitions of the channel parameters are given in the following sec-
tion. Thereby, the behavior indicated above will be revisited based on the equations
presented.
The selection of these parameters is an attempt to specify the main characteristics of
the frequency-selective channel with a minimum number of variables. A very strong
indication for the significance and suitability of the parameters chosen will follow
from the analysis of the FD-level crossing rate in Chapter 3.

2.2.2 Definitions

2.2.2.1 Channel Impulse Response

It is most illustrative to start with the definition of the channel impulse response (IR),
which is the straightforward formulation of the sum of discrete multipath components
impinging at the receiver. In complex lowpass equivalent notation, the IR is written as

h( ) = i e j i ( i ) , (2-1)
i

where {i}, {i}, and {i} are the propagation paths amplitudes, phases, and delays,
respectively, and is the delay-time variable. Normally, the delay of the first (shortest)
ray is defined as 0 = 0, because the absolute delay-times are not important, only the
time-dispersion is. Therefore is called the excess delay-time, and it follows that i > 0
for i > 0, i.e., the channel impulse response is causal.
Note that in a real environment, the parameters {i}, {i}, and {i} are time-variant.
For the sake of simplicity, this time dependency was omitted in (2-1). Within a local
area, i.e., for displacements in the order of a few wavelengths , the ray amplitudes
{i} and the delays {i} can be considered relatively static corresponding to the
assumption of a negligible change of the shadowing. The ray phases {i}, however,
change unpredictably within the interval [0, 2), because they are related to the
absolute path-lengths2.

2
It is an open issue whether the assumption of discrete paths is viable. Generally, each reflection
will show some time-dispersion, and therefore a frequency-dependent magnitude. However, for a gi-
ven observation bandwidth, such physical paths can normally be approximated by (a number of)
discrete Dirac-impulses.
2.2 Characterization of the Mobile Radio Channel 21

2.2.2.2 Channel Parameters

All channel parameters introduced here are defined from the (static) power delay
profile (PDP), which is a function derived from the channel IR (2-1). The PDP
specifies the ray-power versus delay-time structure of the IR, being

p( ) = i2 ( i ) . (2-2)
i

As the ray phases are dropped in this equation, the channel parameters must be
(largely) constant within the local area, provided that the propagation paths are fully
resolvable.
The first parameter is the (normalized) received power, being the sum of the ray
powers

P0 = i2 . (2-3)
i

The Ricean K-factor is the ratio of the dominant paths power to the power in the
scattered paths, defined as

i2,max
K= , where i ,max = max{ i } . (2-4)
P0 i2,max i

It will be seen that the K-factor specifies the depth of the fades within a local area, as
the Ricean probability density function (PDF) will be used to characterize the
amplitude distribution of the channel response. Larger K-factors relate to shallower
fades.
Note that in the presence of a line-of-sight, the first ray is the dominant one, implying
that i,max = 0 at 0 = 0.

Finally, the RMS delay spread is introduced, which is the second central moment of
the (power normalized) PDP, written as

rms = 2 2 , where m = im i2 P0 , m = {1, 2}. (2-5)

i

rms is considered to be the most important single parameter for specifying the time-
extent of the dispersive channel. It also characterizes the frequency-selectivity, since
rms is related to the average number of fades per bandwidth, and to the average
bandwidth of the fades (see Section 3.2).

Smulders states, based on channel measurements over bandwidths of 2 GHz in the 60 GHz band,
that mm-waves have sufficiently small wavelengths to be modeled as rays following discrete paths
(see [8], p. 432 f.).
22 Chapter 2 Modeling of the Frequency-Selective Radio Channel

2.2.2.3 Channel Transfer Function

An equivalent description of the time-dispersive channel is obtained by applying the
Fourier transform to the IR, yielding the channel transfer function (TF). This step will
demonstrate that a time-dispersive channel is also frequency-selective.
Firstly, the time variability is re-introduced to the IR (2-1),

h( , t ) = i (t )e
j i ( t )
( i (t )) , (2-6)
i

leading to the time-variant TF,

d = i (t )e
j [2f i ( t )+ i ( t ) ]
h( , t )e
j 2f
H ( f , t) = . (2-7)
i

The magnitude of this function shows rapid variations with respect to both, the time-
and frequency variables. H(f,t) can be seen as the vector sum of the ray amplitudes
{i(t)}, with vector-angles [2f i (t ) + i (t )] . As the ray phases {i(t)} change rapidly
for small displacements, the vector sum changes, causing the location and time-
variability. The frequency-dependency is due to the different delay times {i(t)},
which, at different frequencies, also lead to drastic changes in the vector sum. The
phases at two specific frequencies differ more with larger excess delay times {i(t)}.
This suggests a dependency of the time-extent of the impulse response (which is
characterized by the RMS delay spread), and the number of fades per unit of band-
width. In Chapter 3, Section 3.2, this relation is extensively studied.

2.2.2.4 Magnitude Distribution

Due to the quasi-random phases of the terms of (2-7), H(f,t) can be seen as the sum of
a (large) number of random variables (RV) with amplitudes {i(t)}, and uniformly
distributed phases over [0, 2). Assuming that a considerable number of rays has
similar magnitudes (except for possibly one dominant ray), the central limit theorem
leads to the conclusion that H(f,t) has a complex Gaussian distribution. Without the
dominant ray it is zero-mean, otherwise it is non-zero mean. The magnitude R = |H(f,t)|
of the complex Gaussian process is described by the Ricean PDF
(r2 + 2 )
r r
p R (r ) = e 2 0
I 0 , (2-8)
0 0
where 0 is the common variance of the real and imaginary components of the
complex Gaussian process, is the amplitude of the mean of H(f,t), = |E{H(f,t)}|,
and I0() is the zero-th order modified Bessel function of the first kind. For the zero-
mean case ( = 0), the Ricean PDF reduces to the Rayleigh PDF 3.

3
If the dominant component i,max occurs at a delay time different to 0 = 0 (or at a non-zero
2.2 Characterization of the Mobile Radio Channel 23

The parameters of (2-8) are related to the channel parameters P0 and K as

K 1
2 = i2, max = P0 and 2 0 = P0 i2,max = P0 . (2-9)
K +1 K +1
Note that 2 is the power of the dominant component, while 20 is the power of the
scattered components. If the central limit theorem (plus dominant path) is not perfectly
valid, then the parameters given in eq. (2-9) may still express a best fit of the Ricean
distribution to the given channel. However, in this case, the magnitude of the dominant
path i,max may be rather seen as an equivalent dominant path gain, which does not
strictly relate to one physical propagation path.

2.2.2.5 Band-Limiting the Transfer Function and Sampling the Impulse Response
For computer simulation schemes, a sampled version of the channel IR is required,
which implies the band-limitation of the respective TF. Let us first introduce the latter.
A sampling interval Ts in the time-domain limits the bandwidth to BW2 = 21T . s

Multiplication of the TF (2-7) by a rectangular window WBW(f) applies such band-

limitation,

1 if f BW
H BW ( f , t ) = H ( f , t ) WBW ( f ) , where WBW ( f ) = 2
. (2-10)
0 if f >
BW
2

This step is equivalent to a convolution of the IR by a sinc-function,

i (t )
hBW ( , t ) = h( , t ) sinc( / Ts ) = i (t )e j i (t ) sinc , (2-11)
i Ts

sin x
if x 0
where sinc x = x . Clearly, rays stop being resolvable, if the delay-time
1 if x = 0
separation between adjacent rays is in the range of Ts or below.
Sampling in the time-domain can be seen as a multiplication by a train of Dirac-
impulses with period Ts. It therefore has the effect of convoluting the frequency-
domain representation by a pulse train with period 1/Ts = BW [11]. The prior band-
limitation keeps the thereby duplicated spectra from overlapping (i.e., aliasing is
avoided), which essentially means that no information is lost through the sampling.
The sampled IR becomes
n Ts i (t )
hBW ,n (t ) = i (t )e j i (t ) sinc , (2-12)
i Ts

Doppler frequency), then the mean will become zero as well, as a (deterministic) complex harmonic
component results. However, the amplitude distribution is still appropriately described by the Ricean
distribution. (This case is described by Rice as the Distribution of Noise plus Sine Wave [9], [10]).
24 Chapter 2 Modeling of the Frequency-Selective Radio Channel

with n = {, 1, 0, 1, 2, } being the discrete delay time index. From this equation,
one can observe that the IR has contributions of all propagation paths at any time-bin
n. (Except if a ray has an excess delay of i = kTs, where k is an integer to be exact).
Even at negative delay times, some leakage of the (causal) IR is evident. From
(2-12) it also becomes clear that, for limited time-resolution or bandwidth, the sampled
IR (at any time-bin n) is rapidly time-variant, due to the time-dependency of the
superimposed rays phases {i(t)}. Calculating channel parameters from this sampled
IR results in instantaneous parameters {P0 , K ,rms } , which are time-variant, even within
a local area, as discussed in Section 2.2.1. The variability of these parameters is shown
in Section 2.2.3, based on simulation results.
The application of the central limit theorem again leads to the conclusion that complex
Gaussian processes appropriately model the coefficients {hBW ,n (t )} (cf. [6], [12]).

Their variances follow the so-called average power delay profile, which usually decays
with increased delay-time. In various channel models, the IR is described in this way
(see e.g., [24][16]). The complex Gaussian distribution also applies to the ray gains
of IRs derived from the FD-channel model, which is proposed in Section 2.3.

The above analysis is an attempt to describe theoretically the behavior of the time-
variant, frequency-selective radio channel. It focuses on the aspects that are important
for a deeper understanding of the FD-channel model. Therefore, particularly the
frequency-selectivity of a band-limited, quasi-static channel has been discussed.

2.2.3 Variation of Channel Parameters Due to Bandwidth Limitation

Simulation results are presented in this section, of the variability of instantaneous
channel parameters within a local area.
Channel realizations were generated with a time-domain (TD) simulation scheme,
which produces channel impulse responses. The simulation model assumes a line-of-
sight ray at = 0, a Poisson process of ray-arrivals (of approx. 60 rays), an
exponentially decaying average power delay profile, and complex Gaussian ray
amplitudes (compare [16] for one cluster; see Section 2.5.3). In a second step, the
impulse responses were normalized to get the required K-factor K, rms = 1 and P0 = 1
[15], allowing for simple evaluation of the estimation error. Applying the Fourier
transform to the generated IRs, (complex-valued, discrete-frequency) TFs were
obtained, with arbitrary bandwidth.
The variability of rms and P0 within a local area is depicted in Figure 2-1a and Figure
2-1b, respectively. For analyzing rms , the simulated TFs were transformed back to the
delay time-domain using the inverse discrete Fourier transform (DFT) without win-
dowing. Consecutively, rms was determined from the positive- part of the obtained,
sampled channel IR, using eq. (2-5). P0 is simply the average power of the band-
2.2 Characterization of the Mobile Radio Channel 25

RMS delay spread estimated from the impulse response; OS = 1

rms
25
standard deviation
mean error
20 K=0
K=1
relative estimation error of rms [%]

K=4
K = 10
15

10

5
0 1 2
10 10 10
observation bandwidth in 1/rms

(a)
normalized received power P0; OS = 1
45
K=0
40 K=1
K=4
K = 10
35
relative estimation error of P [%]
0

30

25

20

15

10

0
0 1 2
10 10 10
observation bandwidth in 1/rms

(b)
Figure 2-1: (a): Bias and standard deviation of the instantaneous RMS delay spread rms with-
in a local area due to band-limitation. (The bias is caused by leakage effects). (b):
Standard deviation of the instantaneous normalized received power P0 within a
local area.

limited TFs.
26 Chapter 2 Modeling of the Frequency-Selective Radio Channel

The standard deviations of these parameters decrease with increasing bandwidth, be-
cause individual propagation paths become gradually more resolvable. The estima-
tion bias in rms (see Figure 2-1a) is due to leakage effects.
Reduced variance and bias for higher K-factor are intuitively explained by the fact that
the (deterministic) dominant path largely determines Ricean channels. Note that K has
most influence on the instantaneous values of the average power P0. This behavior can
be anticipated, since K directly relates to the depth of the fades. I.e., a channel with a
high K-factor (which has shallow fades) shows less variation in this parameter, than
for instance a Rayleigh fading channel (which has quite deep fades).

2.3 Frequency-Domain Channel Modeling

The channel model proposed in this section describes the correlation properties of the
channel transfer function in the frequency-domain. Starting from the definition of the
channel correlation functions (and power spectra), this so-called frequency-domain
channel model is derived. Mathematical expressions are given, relating the models
parameters to (physical) channel parameters.

2.3.1 The WSSUS Channel Model

The channel correlation functions and power spectra are a set of functions defining the
small-scale characteristics of multipath fading channels in more detail than the channel
parameters given above. Introducing some assumptions will lead to the channel model
used throughout this work. In particular, we concentrate in this work on the correlation
properties of the time-variant transfer function H(f,t) (see (2-7)), because this function
determines the channels impact on an OFDM system modeled as a set of parallel
Gaussian channels (see Section 4.2, [17]). Considering the mobile radio channel as a
linear time-variant system, it is seen that the TF H(f,t) is only one possible channel
representation (from the family of Bellos system functions [5], [6]). Another one is,
for instance, the time-variant impulse response given by (2-6).

Let us first define the channel correlation functions assuming that those functions are
wide-sense stationary. This means that the autocorrelation function

H ( f1 , f 2 , t1 , t 2 ) = E{H * ( f1 , t1 ) H ( f 2 , t 2 )} (2-13)
depends only on the frequency-separation f = f1 f2 and on the time-separation
t = t1 t2, but not on the absolute observation frequencies {f1, f2} and times {t1, t2}. In
other words, the time-variant transfer function H(f,t) is wide-sense stationary (WSS)
with respect to both variables f and t. The channel is thus characterized for all times
and all frequencies by the so-called spaced-frequency, spaced-time correlation func-
tion
2.3 Frequency-Domain Channel Modeling 27

H (f , t ) = E{H * ( f , t ) H ( f + f , t + t )}. (2-14)

It can be shown that this assumption is equivalent to the introduction of the wide-sense
stationary and uncorrelated scattering (WSSUS) channel (see e.g., [5], [6], [2]). In the
WSSUS channel, the WSS-property applies to the time variability of the IR h(,t). The
uncorrelated scattering (US) property is based on the assumption that the attenuation
and phase of a propagation path at delay time i is uncorrelated to the attenuation and
phase at delay time k, for i k.
In order to apply the concept of the WSSUS channel to real radio channels, the quasi-
WSSUS channel (QWSSUS) was introduced by Bello [5]. A QWSSUS channel has
the properties of a WSSUS channel within a local area, and for a limited bandwidth
and time.
Furthermore, it should be noted that for Gaussian processes, the WSS property implies
stationarity in the strict sense. If the distribution of the TF H(f,t) is complex Gaussian,
with zero- or non-zero mean, then the amplitude distribution is Rayleigh or Ricean,
respectively. As this agrees to the channel properties derived in Section 2.2.2, and as
Gaussian processes generally simplify any stochastic mathematical analysis, the com-
plex Gaussian case will be assumed.

In Figure 2-2, an overview is given of the most commonly used correlation functions
and power spectra defining the stochastic properties of the time-variant channel IR,
and TF. These system functions are found in the center of the figure, surrounded by
their second order moments, which are interrelated by Fourier transforms. As men-
tioned above, our focus lies on the spaced-frequency, spaced-time correlation function
depicted just above the center of this figure.

2.3.1.1 Special Cases

Most of the analysis presented in this thesis concentrates on the case of the time-
invariant frequency-selective channel. The channel is then described by the TF H(f),
which is a WSS complex Gaussian stochastic process in f, according to the above
assumptions. The second order statistical functions characterizing H(f) are the spaced-
frequency correlation function H (f ) = H (f ,0) and its Fourier transform (FT), the
delay power spectrum (DPS) h() (see Figure 2-2). A mathematical description of the
DPS will be the basis of the so-called frequency-domain (FD) channel model.
More familiar is the dual approach of modeling the time variability of a narrowband
channel as a WSS complex Gaussian stochastic process H(t). An example for this
method is widely known in the literature as Jakes fading model [1]. Compared to the
FD model, the frequency variable is exchanged with the time variable, and the second
order statistics are the spaced-time correlation function H (t ) = H (0, t ) and the
Doppler power spectrum SH() for Doppler frequency , which are a Fourier pair as
well (see Figure 2-2).
28 Chapter 2 Modeling of the Frequency-Selective Radio Channel

h() H(f)
multipath intensity profile FT spaced-frequency wide-band characterization
delay power spectrum ( f) correlation function (time-invariant channel)
max max. delay spread (f)c coherence bandwidth
characterization of time
t = 0 t = 0 variations (narrow-band)

h(;t) H(f;t) H(t)

delay cross-power spectral spaced-frequency, spaced- f = 0 spaced-time correlation
FT
density ( f) time correlation function function
(t)c coherence time

ACF ACF
(WSSUS) (WSSWSS)

h(;t) H(f;t)
equivalent lowpass FT time-variant
time-variant ( f) transfer function
impulse response
FT FT FT
(t ) (t ) (t )

S(;) SH(f;) SH()

Scattering function Doppler cross-power f = 0 Doppler power spectrum
FT
( f) spectral density fm max. Doppler freq.;
Doppler spread

Figure 2-2: Overview of the two time-variant system functions described the channel impulse
response and the channel transfer function and a set of correlation functions
(second order moments) describing their stochastic properties.

2.3.1.2 Additional Channel Parameters

Figure 2-2 also introduces some additional channel parameters, which are derived
from the correlation functions and power spectra.
Coherence-time and -bandwidth indicate the ranges (in time and frequency) over
which the TF H(f,t) shows significant correlation. They are defined as the time- or
frequency-separations t and f, where the spaced-time or -frequency correlation
functions, respectively, drop below 0.9. (Sometimes 0.5 is used for this threshold.)

Related to the power spectra, the maximum delay spread and the Doppler spread are
defined, corresponding to the maximum delay-time and frequency-components in
these spectra.
Often, mathematical relations are given in-between these parameters, i.e., between the
coherence-bandwidth and the (reciprocal of the) maximum delay spread or the RMS
delay spread, and between the coherence-time and the (reciprocal of the) Doppler
spread. However, these relations loose significance in the Ricean case, since the
dominant component (leading to the non-zero mean of the Gaussian distribution)
2.3 Frequency-Domain Channel Modeling 29

causes a constant additive term in the channel correlation functions [15]. Therefore,
these relationships should be used with care.

2.3.2 Channel Description

The delay power spectrum (DPS) characterizes the frequency-selectivity in the FD-
channel model. In agreement with measurements reported in [24], the shape of the
DPS is defined as shown in Figure 2-3. It is specified by four parameters: 2 the
normalized power of the direct ray; [1/s] the normalized power density of the
constant-level part; 1 [s] the duration of the constant level part; and [1/s] the
decay exponent of the exponentially decaying part. Mathematically, the DPS can be
written as

0 <0
2 ( ) =0

h ( ) = . (2-15)
0 < 1
e ( 1 ) > 1

In many cases, the number of (free) parameters can be further decreased. The
exponentially decaying DPS is a good approximation for most practical channels,
which is implemented by letting 1 = 0. The existence of a line-of-sight (LOS) ray at
= 0 implies that the channel TF has non-zero mean, thus the fading envelope
distribution is Ricean. Rayleigh fading channels have = 0.
For the analysis it is appropriate to define u = 1, being a single parameter to account
for the shape of the DPS. u can take values u [0, ], where the two extreme cases
u = 0 and u = describe an exponentially decaying and a rectangular DPS, respective-
ly. Note that in the latter case (rectangular DPS), the maximum excess delay will be
much smaller than for u = 0, thus u can be used to adjust this parameter (see below).
Relations between the model parameters defined above and the channel parameters are
presented in Section 2.3.3.

h() [dB]
2

1 Excess delay [s]

Figure 2-3: Model of the delay power spectrum (DPS).

30 Chapter 2 Modeling of the Frequency-Selective Radio Channel

2.3.3 Relation to (Physical) Channel Parameters

For the application of the FD-channel model, it is most important to relate its
parameters {2, ,,1} to the channel parameters defined in Section 2.2.2: the
normalized received power P0, the Ricean K-factor K, and the RMS delay spread rms.

The channel parameters derived from the channel model are the local-area means, as
discussed in Section 2.2. Finite bandwidth realizations or measurements within a local
area have instantaneous channel parameters {P0 , K ,rms } spread around those means.
Table 2-1 gives an overview of expressions relating the model parameters {2, ,,1}
to the channel parameters {P0,K,rms} and vice versa. The derivation of these equations
is outlined below. For notational convenience we introduce u1 = u + 1 , u 2 = u 2 2 + u + 1 ,
and u 3 = u 3 3 + u 2 + 2u + 2 , with u = 1. An important special case is given by u = 0,
the exponentially decaying DPS, which is an appropriate description for many
practical channels. Table 2-1 also lists the simplified expressions for this case.

2.3.3.1 Derivation of Channel Parameters

From the continuous DPS h() defined by (2-15), the analytical expressions given in

Table 2-1: Relation between model and channel parameters. (The symbols are defined in the
text).

model channel
u = 1 [0,] u=0

P0 = 2 + u1 P0 = 2 +

2 2
K= K=
u1

1 1 u3 1 u2 2 1 2K + 1
rms = rms =
K + 1 u1 ( K + 1) 2 u1 2 K +1

channel model
u = 1 (must be known) u=0
K K
2 = P0 2 = P0
K +1 K +1
1 1 u3 1 u2 2 1 2K + 1
= =
rms K + 1 u1 ( K + 1) 2 u12 rms K +1

P0 P0
= =
K + 1 u1 K +1
2.3 Frequency-Domain Channel Modeling 31

Table 2-1 can be derived for the expected values of normalized received power P0,
Ricean K-factor K, and RMS delay spread rms. P0 relates to the DPS as

1
P0 = h ( )d = 2 + 1 + . (2-16)
0

The K-factor is used to characterize the amplitude distribution of Ricean channels,

relating the power of the direct path to the power of the scattered paths.

2 2
K= = (2-17)
P0 2 ( 1 + 1 )

The RMS delay spread rms is the single most important parameter characterizing the
frequency-selectivity. It can be interpreted as the centralized second moment of the
normalized DPS

rms = 2 ( )2 , (2-18)
where

h ( ) 2 1
= d = 1 + 1 + 2 , and (2-19)
0
P0 2

h ( ) 3 2 2 2
= 2
2
d = 1 + 1 + 21 + 3 . (2-20)
0
P0 3

2.3.3.2 Spaced-Frequency Correlation Function

The spaced-frequency correlation function is used repeatedly throughout this thesis to
implement the channel-behavior in the mathematical analysis of the radio channel and
in the analysis of OFDM system aspects. It is derived from the DPS (2-15) via the
Fourier transform:

H (f ) = E{H * ( f ) H ( f + f )} = F { h ( )} =
1 (2-21)
2 + 1 sinc( 1f )e j1f + e j 21f
+ j 2f

For 1 = 0, i.e., for the special case of an exponentially decaying DPS, the spaced-
frequency correlation function can be written as

P0 1
H (f ) = K + , (2-22)
K +1 1 + j 2f rms K1

where, K1 = ( K + 1) 2K + 1 .
32 Chapter 2 Modeling of the Frequency-Selective Radio Channel

2.3.3.3 Maximum Excess Delay

The shape factor u introduces another degree of freedom in the channel model, which
allows the variation of the maximum excess delay max by a certain factor, for a given
RMS delay spread rms. Strictly speaking, the maximum delay spread is infinite due to
the exponentially decaying part of the DPS, which never becomes zero. In practice,
however, multipath components can be neglected that are attenuated very significantly.
We therefore define the maximum excess delay as the delay time, where the exponen-
tially decaying part has decreased by about 43 dB. Such attenuation is reached, if the
duration of the exponentially decaying part is exactly exp = 10/, leading to the maxi-
mum delay spread max = 1 + exp = 1 + 10 / . Expressed in terms of channel parame-
ters this is
u1 ( K + 1)
max = rms (u + 10) , (2-23)
u1u3 ( K + 1) u 22

which simplifies for u = 0 (i.e., 1 = 0) to

K +1
max = rms 10 = 10 rms K1 . (2-24)
2K + 1

It is seen that max and rms are related by a factor, which is a function of K and u.
Figure 2-4 illustrates this factor. According to this definition, max is exactly ten times
larger than rms at K = 0, and u = 0. Larger K-factors generally increase this factor,

relation between the maximum delay spread and the RMS delay spread
25
u = 0 (exponential DPS)
u=2
normalized maximum delay spread max/rms

u = 10
20 u = (rectangular DPS)

15

10

0
0 2 4 6 8 10
Ricean Kfactor

Figure 2-4: Factor between the maximum delay spread max and the RMS delay spread rms, as
a function of the Ricean factor K and with the shape factor u as a parameter.
2.4 Frequency-Domain Channel Simulation 33

larger parameters u decrease it. For instance, max is only about three times rms for the
rectangular DPS, at low K-factors.
As the maximum delay spread max defines the maximum frequency component of
the DPS, it is this parameter which defines the Nyquist frequency when a sampled ver-
sion of the channel transfer function is needed in measurements or computer simula-
tions. That is, the sampling interval in the frequency-domain must be smaller than
1/(2max).

2.4 Frequency-Domain Channel Simulation

The discussion of simulation schemes in this section is restricted to the case of static
(time-invariant) frequency-selective channels. Such simulations are for instance
appropriate for the study of OFDM systems, with a system model that reduces the
channel including the IFFT at the OFDM transmitter and the FFT at the receiver to a
set of parallel Gaussian (sub-) channels (see Chapter 4, [17]). These sub-channels have
complex attenuation factors given by the channels TF H(f) at the frequency instants of
the OFDM sub-carriers f = nF, where F [Hz] is the sampling interval in the frequency-
domain and n = {0, 1, 2, N 1}. The simulation scheme presented in this section
directly generates H(f) for well defined channel parameters.
In some cases, for instance for evaluating channel estimation schemes, the time-
variability of the TF is also of great importance. The extension of the static simulation
scheme to a time-variant one is discussed. A discussion of the channel-variability in an
OFDM-based wireless LAN system is given in Section 5.2.1.2. The induced
performance degradation is studied in Section 6.5.3.

2.4.1 Model Description

The simulation system for time-invariant channels is shown in Figure 2-5. Real-valued
white (or wide-band) Gaussian random processes W(f) in the frequency-domain are
generated by a noise source. The appropriate spaced-frequency correlation is obtained
by FD-filtering of W(f) with a (low-pass) filter g(f). The output of this filter is the real-
valued, colored noise process rH'(f) = W(f) g(f), where denotes convolution. The
inverse FT of rH'(f) (in delay-time-domain representation) is complex-valued and
hermitian, i.e., symmetric with respect to the = 0 axis. It is not causal, in contrast to
the impulse response (IR) of a real channel. The required causality in the time-domain
is obtained by applying the Hilbert transform to rH'(f) and adding the result iH'(f) as
H'(f) = rH'(f) jiH'(f). Doing this, the negative part of the IR is canceled.
The amplitude of the TF |H'(f)| is Rayleigh distributed since H'(f) is a complex
Gaussian noise process. A Ricean fading channel may be simulated by adding a
j
complex constant e to H'(f), representing the LOS path at = 0.

34 Chapter 2 Modeling of the Frequency-Selective Radio Channel

Noise
(real-valued) W(f) shaping filter rH'(f) H'(f)
Noise source (real) (cmplx.)
G(), g(f) (real) H(f)

Hilbert iH'(f)
transform (real)
j
e
j
(cmplx.)
PSDs:
SH'() (causal)
SW() (flat) SrH'() (colored) h'() h() (+ LOS)

Figure 2-5: Frequency-domain simulation of the frequency-selective radio channel.

2.4.2 Implementation of the Simulation Scheme

To obtain a computer simulation program producing TFs with the desired DPS, two
elements of the above simulation scheme must be appropriately designed; the noise-
shaping filter g(f) and the variance of the noise source W2 . The simulator produces a
sampled version of the TF, H(nF), where F [Hz] is the sampling interval in the fre-
quency-domain and n = {0, 1, 2, , N 1}. F must be selected according to the sam-
pling theorem, i.e., F < 1/(2max).
The power spectral density (PSD) of the output of the simulation scheme (which is in
-domain) has to match the continuous DPS defined by (2-15). This is achieved by
designing the filter g(nF) to have a TF G() proportional to the DPS (for > 0; i.e.
skipping the LOS component). Any classic filter design method can be used in this
process [11]. By definition we let |G()| = 1 during the constant-level part (or at = 0+
if there is no constant-level part), which leads to the variance W2 = (4 F ) , as de-
rived below.

2.4.2.1 Derivation of the Variance of the Noise Source

The noise source produces independent, real-valued noise samples with variance W2 .
The sequence W(nF) thus has a (periodic) spectrum with constant PSD

SW ( ) = W2 F . (2-25)
Applying these samples to the noise shaping filter with amplitude TF

1 1
G ( ) = ( 1 ) , (2-26)
e > 1

leads to the PSD of Re{H'(nF)} written as

2.4 Frequency-Domain Channel Simulation 35

S r H ' ( ) = W2 F G ( ) . (2-27)

The next step in the simulation scheme is the addition of the Hilbert transformed (HT)
sequence, which increases the PSD for > 0 by a factor of four. (The HT cancels the
negative- part of the Fourier spectrum, while doubling the positive- part, resulting in
four-fold power for > 0). This yields the PSD to be compared with the model (the
DPS) as

S H ' ( ) = 4 W2 F G ( ) h ( ), for > 0 , (2-28)

which yields W2 = /(4 F ) .

2.4.2.2 Extension to a Time-Variant Channel Simulator

In order to extend this static simulation scheme to a time-variant one, the TF H(f1,t)
must have the required Doppler spectrum when the time-variations are investigated at
any given frequency f = f1. This may be achieved by generating a number of
independent TFs H(f,t = kTs), k = {1, 2, 3, } and filtering them in time-direction at
each frequency sample, according to a specific Doppler spectrum. (Separability of the
joint time-frequency correlation function H(f,t) is thereby assumed4). A set of N
filters is required for applying time-variability to the transfer functions in this way.
It should be noted that this simulation scheme gets rather complex. It might thus be
preferable to use a conventional fading simulator one that generates a (time-variant)
IR and transform the IR to the frequency-domain, if required. Usually, the IR is
defined by much less than N coefficients, therefore the complexity is reduced.

2.4.3 FD-Simulation Results

In Figure 2-6, a simulated TF is shown (Figure 2-6a) and compared to a measured one
(Figure 2-6b). The two channels IRs are given in Figure 2-7a and b, both derived
from the respective TFs using the inverse discrete FT (IDFT) without windowing. The
measurement was performed with a network analyzer, observing a bandwidth of 1
GHz around a center frequency of 11.5 GHz 5. The channel parameters P0 = 62.1 dB,
K = 1.9 dB, and rms = 9.0 ns were extracted from the measured TF (Figure 2-6b), and
(with 1 = 0) used to generate the simulated TF (see Figure 2-6a). (The method
proposed in Section 3.3 of this thesis was employed to estimate the channel parame-

4
Separability of the two-dimensional spaced-frequency, spaced-time correlation function H(f,t)
means that it can be written as a product H(f,t) = H(f)H(t). This assumption is valid if
maxfm << 1 [18], which is given for practical propagation channels (max denotes the maximum excess
delay; fm is the Doppler spread).
5
The author would like to thank Dr. G. J. M. Janssen for providing measurement results for the
validation of the proposed methods [14]. The measurements were conducted at the TNO Physics and
Electronics Laboratory in The Hague, The Netherlands, in the period of August December 1991.
36 Chapter 2 Modeling of the Frequency-Selective Radio Channel

50

amplitude |H(f)| [dB]

60

70

80

90
0 200 400 600 800 1000
frequency [MHz]

4
phase arg(H(f)) [rad]

4
0 200 400 600 800 1000
frequency [MHz]

(a)

50
amplitude |H(f)| [dB]

60

70

80

90
0 200 400 600 800 1000
frequency [MHz]

4
phase arg(H(f)) [rad]

4
0 200 400 600 800 1000
frequency [MHz]

(b)
Figure 2-6: (a): Simulated transfer function (TF); (b): TF measured with a network analyzer
(corrected for linear phase shift);

ters.) Both TFs have a length of 801 samples. A 15-tap FIR filter was used for the
noise shaping filter g(nF) in the simulation scheme.
The TF is obtained from a stochastic simulation model. Therefore we do not expect it
to be identical to the measured TF. However, it is clearly seen that the characteristic of
the fading is well reproduced. Originally, a linearly increasing phase shift was evident
in the measured TF corresponding to the propagation delay of the shortest path. In the
2.4 Frequency-Domain Channel Simulation 37

65

70

75

80
amplitude [dB]

85

90

95

100

105

110

115
20 0 20 40 60 80 100 120 140 160
excess delay time [ns]

(a)

65

70

75

80
amplitude [dB]

85

90

95

100

105

110

115

20 0 20 40 60 80 100 120 140 160

excess delay time [ns]

(b)
Figure 2-7: (a) Impulse response (IR) derived by IDFT from the simulated TF shown in Figure
2-6a. (b): Impulse responses derived from the measured TF (see Figure 2-6b).

illustration this was compensated, to have the first component arrive at (excess) delay
= 0, in agreement with the simulation model.

The probability density function (PDF) and the cumulative distribution function (CDF)
of the simulated amplitude TF are shown in Figure 2-8 and compared to the Rayleigh
distribution. Because of the low K-factor (K = 1.9 dB), good agreement is evident.
38 Chapter 2 Modeling of the Frequency-Selective Radio Channel

0.8
pdf of data
Rayleigh

PDF p(|H(f)|)
0.6

0.4

0.2

0
0 1 2 3 4 5
amplitude |H(f)|
0
10
Pr(|H(f)| < abscissa)

1
10

2
10
cdf of data
Rayleigh
3
10
25 20 15 10 5 0 5 10
amplitude [dB]

Figure 2-8: PDF and CDF of the amplitude of the simulated transfer function.

Second order statistical properties estimated from simulated TFs are shown in Figure
2-9. The power spectrum obtained by averaging over periodograms of 100 simulated
TFs agrees well with the used DPS model (upper plot). The lower plot shows (spaced-
frequency) correlation properties and compares them to the theoretical function given
by (2-21). Coherence bandwidths are determined by solving numerically for the fre-

10
0 simulation
amplitude [dB]

model
10
20
30
40
50
20 0 20 40 60 80 100 120
excess delay time [ns]

1
correlogram
correlation coeff.

0.8 correl. fct.

analytical
0.6
0.4
0.2
0
100 50 0 50 100
frequency seperation [MHz]

Figure 2-9: Delay power spectrum (DPS) and spaced-frequency correlation function for the
FD-simulation model. Upper plot: DPS and estimated power spectrum. Lower
plot: Correlogram; estimated and analytical correlation functions; markers {:
Coherence bandwidths 0.5 and 0.9.
2.5 Application to mm-Wave Radio Channels 39

quency-separations, where the correlation functions magnitude drops to 0.9 (or 0.5
according to the definition).

2.4.4 Differences to Time-Domain Simulation Schemes

The equally spaced tap-gains of the sampled (and band-limited) impulse response are
not uncorrelated, according to the analysis shown in Section 2.2.2 (see eq. (2-12)).
That is, there exists a certain autocorrelation between different delay-bins of the IR.
This correlation results from the band-limitation needed for the time-quantization,
which implies a convolution of the discrete, non-sampled impulse response with a
sinc-function. Another effect of this convolution is visible in spectral components at
negative delay-values, due to leakage effects. The FD-simulation scheme shows these
properties (see Figure 2-7a and Figure 2-9).
Many time-domain simulators, however, implement the channel impulse response by
simply generating independent, complex-valued path gains at the (sample-spaced)
delay bins. Leakage effects, i.e., components at negative delays, are not considered
either (see e.g. [6], Figure 1.12; [19]). (One sampled simulation model, which does
consider those effects is described in [18].)
In particular, these simplifications are used, when the channel models are applied for
the design of digital radio interfaces. Normally, the resulting differences are negligible,
but there are cases, where the impact gets important. An example is discrete Fourier
transform (DFT) based channel estimation schemes for OFDM. Such channel
estimators determine firstly a coarse estimate of the channel TF, for instance from a
training sequence. In order to reduce the mean-square-error (MSE) of the estimate, the
next step is a transform of the transfer function to the delay time domain, yielding a
noisy channel impulse response. In this form, likely noise components at negative or
very large delay values can be identified and set to zero, followed by a back
transformation to the frequency-domain. The result is an estimate of the channel TF
with enhanced signal-to-noise ratio (SNR) and hopefully reduced MSE. The
simulation of such a scheme suggests excellent performance, if a so-called sample-
spaced channel simulator is used, because then the channel IR is indeed zero at the
sample-bins set to zero. On a real channel, however, important information is lost, as
channel taps are set to zero which correspond to leakage-components. This leads to
irreducible error floors in terms of bit-error-rate and MSE [20][22]. In this respect,
the proposed FD-simulation model has an inherent advantage over conventional,
sampled TD-models, because the correlation among channel taps and leakage effects
are considered, in closer agreement to reality.

2.5 Application to mm-Wave Radio Channels

This section has two main purposes. Firstly, the suitability of the proposed FD-channel
model and simulation scheme is verified, and secondly, parameters are found for the
40 Chapter 2 Modeling of the Frequency-Selective Radio Channel

model. These goals are approached through a discussion of measurement campaigns

reported in the literature.
In particular, our focus lies on mm-wave radio channels, which is the considered
frequency-band for the communications system described in Part II of this thesis.
Within the mm-wave frequencies, the 60 GHz band has received most attention in the
literature, for the following main reasons: Large amounts of bandwidth are unallocated
in this band; bandwidths that are required for communications systems at the intended
data rates of 100 Mbit/s and above. Another advantage of 60 GHz is due to a physical
property of the propagation channel at this specific frequency. Oxygen absorption
leads to attenuation above 11 dB/km, between 57 and 63 GHz. This attenuation
(additional to the path loss) is believed to enable shorter reuse distances in cellular
systems because it counteracts co-channel interference. Over short distances, the
additional attenuation can be neglected.
It is a general property of mm-wave propagation that the behavior of propagation rays
is well characterized by geometric optics. That is, waves do not penetrate through
walls or other obstacles, and wave reflection is the main mechanism leading to
multipath. Scattering, diffraction, and wave guiding are considered far less important
[23].
This section starts with a discussion of measurement results (Section 2.5.1). In Section
2.5.2, typical channel parameter values are given. The influence of features of the
propagation environment on those parameters is discussed. Channel models suggested
in the literature are treated separately, in Section 2.5.3. Section 2.5.4 deals with the
applicability of the newly proposed FD-channel model to (mm-wave) radio channels
the validation of the FD-model.

2.5.1 Discussion of Measurement Results

A major activity in the field of mm-wave propagation has been conducted in the
framework of the European RACE (Research into Advanced Communications systems
in Europe) project 2067, Mobile Broadband Systems (MBS) [23], [24]. The
measurement campaigns described include material characterization, and indoor and
outdoor propagation studies. Ray-tracing models have been developed for predicting
propagation parameters and for investigating the impact of environment features,
antenna characteristics, etc.. An extensive list of literature on mm-wave propagation is
found in the Final report on propagation aspects of the RACE-MBS project [23].
Partly related is the activity carried out within the COST 231 (European Co-Operation
in the field of Scientific and Technical research) program [8]. This study also covers
indoor and outdoor channels. A major contribution on indoor propagation originates
from the research of Smulders, conducted at Eindhoven University of Technology
[24].
Other work on indoor channel is found in [12], [14], [15], and [25][35], outdoor
studies are presented in [36][38]. Note that most of the work has been done on indoor
2.5 Application to mm-Wave Radio Channels 41

channels and their modeling, probably also because of the range limitation of mm-
wave propagation.

The main parameters of interest for applying the FD-model to mm-wave channels are
the normalized received power (NRP), P0, the Ricean K-factor, K, and the RMS delay
spread, rms. For the air interface design, the latter two parameters, {K, rms} are gene-
rally sufficient. The NRP is required for link budget considerations (see Chapter 5).
While most studies present results of rms and the NRP, the K-factor is unfortunately
commonly not investigated.

Generally, it is difficult to compare measurements conducted by various research

groups, because of
differences in the measurement equipment and method used,
different antenna characteristics and configurations,
different parameters measured and presented, and
different environments investigated.
We try to organize this comparison and overview by discussing the parameters of
interest and elaborating on the impact of some of the above listed factors. Only wide-
band measurements are considered, because of the importance of characterizing the
time-dispersive and frequency-selective nature of the channel. The modeling of these
channel properties is essential for the air-interface design, the intended application of
the channel model under development.

2.5.1.1 Measurement Set-Ups and Techniques

Most indoor measurement campaigns use vector network analyzers to scan the channel
transfer function (phase and magnitude) versus frequency (see e.g., [8], [12][15],
[31], and [33][35]). The conditions to use such equipment are short distances, be-
cause a phase reference must be provided between the transmitting and the receiving
sides, and a (quasi) static channel, because of the time it takes to acquire the frequency
transfer function. These conditions are feasible in indoor scenarios. The main advan-
tages of this approach are the high time-resolution achieved by scanning over a large
bandwidth, and the good SNR, because a narrowband (continuous wave) signal is
transmitted in which the whole transmit power is concentrated. The delay-time resolu-
tion investigated is normally around 1 ns, corresponding to a scanning bandwidth of 1
2 GHz. In [38], a network analyzer was used for outdoor measurements.

Correlation type channel sounders were developed for the extensive measurement
campaigns performed in the RACE-MBS project [23], [24]. For outdoor channels, a
wide-band test signal (chirp) was generated by rapidly sweeping a carrier over a
bandwidth of up to 200 MHz [36], [39]. A separate indoor channel sounder is based on
42 Chapter 2 Modeling of the Frequency-Selective Radio Channel

the transmission of a pseudo random binary sequence and a sliding-correlator on the

receivers side (see [40]). Similar equipment was employed in [32] for indoor mea-
surements, and in [37] for outdoor measurements.

Within this Ph.D. research, a novel, non-coherent channel measurement technique was
developed that can estimate the NRP, K-factor, and of rms from swept-frequency
power measurements. No phase measurement is required, which simplifies the equip-
ment needed. The measurement technique is described in detail in Chapter 3 of this
thesis. Measurement campaigns conducted with this method at TU-Delft are described
in [25][28]. Indoor and outdoor channels were studied, at 17 and 60 GHz.

2.5.2 Discussion of Channel Parameters

The RMS delay spread (RDS) rms and the Ricean K-factor are the two most important
parameters for specifying the channels frequency-selective nature, in the context of
air-interface design. It will be seen from the study of the frequency-domain level
crossing rate that the RDS determines the number of fades per bandwidth, while the K-
factor specifies the depth of the fades (see Chapter 3). The normalized received power
(NRP) just determines the average signal-to-noise ratio (SNR). Since the SNR is
usually considered as a variable in any kind of system studies, absolute values of NRP
are not of major importance for the air-interface design.

2.5.2.1 RMS Delay Spread

The following main features of the propagation environment influence the RDS. Note
that the mentioned properties are applicable for indoor channels only. Similar features,
however, will also have an impact on the RDS in outdoor scenarios.
Room size
Generally, the RDS increases with the room size. Such behavior was for instance
reported in the work by Smulders [24], who measured typical values of RDS between
15 and 45 ns in small rooms with dimensions 24114.5 m3, and values between 30
and 70 ns in larger rooms.
These values are rather large, compared to the results from many other indoor
measurement campaigns found in the literature. Main reason for the large values is
next to the large rooms investigated, the antenna design implemented. The biconical
horn antennas, having an omnidirectional radiation pattern in the azimuth plane, a 3-
dB beam-width of 9 in the elevation plane, and a directivity of 9 dBi, were designed
such that the NRP hardly depends on the position within a room. Therefore, they
radiate quite a large fraction of the transmitted power towards the walls, leading to
strong first reflections and long delay spreads. Although the delay spreads are quite
large, this design might be of advantage, because self-shadowing effects become less
harmful. That is, signal loss due to the obstruction of the LOS-path by the user (see
2.5 Application to mm-Wave Radio Channels 43

[41]) is assumed to be less significant for such antenna set-ups.

Antenna directivity
Directive antennas attenuate parts of the impinging reflected waves. Therefore, the
RDS usually decreases, as more directive antennas (in the azimuth plane) are
employed.
Such behavior is clearly seen from measurements and ray-tracing simulations
performed by T. Manabe et al. [31]. In a room with dimensions of 13.582.6 m, they
measured typical RDS-values of 18, 14, 5, and 1 ns, for, respectively, an omni-
directional antenna (/2-dipol), and antennas with 3 dB beam-widths of ~60, ~10,
and ~5.
A similar study based on a ray-tracing tool is presented in the final report of the
RACE-MBS project [23]. In the investigated room of approximate dimensions 1173
m3, different antenna configurations were evaluated, leading to RDS-values of 20 25
ns for the less directional antennas, and values (significantly) lower than 5 ns for the
most directional ones.
In order to investigate this anticipated dependency between the RDS and the antenna
characteristics, Smulders has conducted some additional measurements using a 15 dBi
circular-horn antenna (in stead of the 9 dBi omnidirectional biconical horn antenna) on
one side of the measured link [24]. His results confirm the expected behavior if
median values of RDS are considered (RDS decreases from ~40 ns to ~25 ns).
However, the max. RDS values observed were even larger than for the standard
antenna configuration (increase from ~48 ns to ~60ns). A similar behavior was
reported by Bultitude et al. [33], who performed a measurement campaign at 40 GHz,
in a large open office environment. It is a possible explanation that the more directive
antenna, which also has the higher gain, may emphasize some reflected paths with a
rather large delay time. Such paths contribute strongly to the RDS.
Building material
The reflectivity of building material is expected to be another important factor
influencing the RDS. This behavior was e.g. reported by Smulders [24], who measured
higher RDS values in a small room with metal walls (room dimensions ~1093 m3;
rms 45 ns), than in a much larger auditorium room with walls covered by wood and
acoustically soft material (room dimensions ~30216 m3; rms 35 ns). In a small
room (~1394 m3) with wood-covered walls, RDS-values of ~20 ns were measured.
Outdoor measurements
Measurements in seven different streets downtown Oslo were reported in [23], [24],
[36] (MBS-RACE project). The RDS is typically lower than 20 ns, except for one
measurement where a major reflection source (tourist bus) was located on the street. In
the latter case the RDS was < 50 ns. The maximum delay spread (the sliding delay
window, i.e., the shortest period of the IR containing 90 % of the received energy) is
44 Chapter 2 Modeling of the Frequency-Selective Radio Channel

less than 45 ns for 90 % of the measurement points. However, maximum values up to

270 ns were observed. Results from measurements at city squares show higher values
of RDS and maximum delay spread.
Outdoor measurement results reported in [25], [26] (for 17 GHz), [38] (for 60 GHz),
[37] (for 40, 60 GHz) also show RMS delay spread values starting below 20 ns and
occasionally going up to about 100 ns and above [38]. Little work has been done on
outdoor propagation in mm-wave bands.

2.5.2.2 Ricean K-factor

Parameter pairs of RMS delay spread, rms and the Ricean K-factor are required for
modeling multipath radio channels using the FD-channel model introduced in Section
2.3. While statistics of the RDS are found in most propagation studies, the Ricean K-
factor is often not (explicitly) investigated. Many studies assume Rayleigh fading
amplitude distributions, i.e., K-factors of zero. In situations where the line-of-sight
(LOS) between transmitter and receiver is obstructed, this assumption may be
reasonable. However, as a LOS path is often required for reliable transmission at mm-
wave frequencies [41], the K-factor becomes an important channel parameter. Typical
values of K are given below. The influence of a LOS path and the influence of the
antenna characteristics are investigated.
Influence of a line-of-sight path
Two measurement campaigns that consider the K-factor were conducted by Janssen
[14], [15] and by Bohdanowicz [25], [26]. Although those measurements were
performed at lower frequency-bands (at 2.4, 4.75, and 11 GHz, and at 17 GHz,
respectively), the results are interesting for modeling the 60 GHz channel. One
important reason is that most of the measurement situations were similar to the
expected scenarios for 60 GHz systems, where both, the transmitter and the receiver
are typically located within the same room. Moreover, results for all these frequency-
bands are quite similar, suggesting that a shift to the 60 GHz band would not have a
large impact, either. The comparative study of a 1.7 GHz and a 60 GHz channel
presented in [32] confirms the latter.
Characteristic channel parameters reported by Janssen ([14], [15]) are rms 10 ns, K
2.5 dB in LOS situations, and rms 15 ns, K 3 dB without LOS. All measurements
were performed in relatively small rooms.
The 17 GHz channel study by Bohdanowicz ([25], [26]) gives typical K-factors
between 0.3 and 2.5 dB for LOS indoor scenarios (rms 5 17 ns), and values
around 1 dB for indoor non-LOS situations (rms 9 ns). Larger K-factors of 3 5
dB were determined from outdoor LOS measurements, where rms 20 30 ns.
Clearly, increased K-factors are observed in the presence of a (dominant) LOS path,
corresponding to amplitude distributions with shallower fades. While K-factors below
3 dB can be well represented by the Rayleigh distribution, higher values should be
2.5 Application to mm-Wave Radio Channels 45

modeled by the Ricean distribution. In particular if K 0 dB, i.e., the dominant path
carries greater or equal power than all the reflected paths, the Ricean model must be
used.
Antenna directivity
It is expected that more directive antennas yield higher K-factors, because if the
antennas are pointed towards one another, the dominant path is amplified while the
reflected ones are attenuated. Inspection of the impulse responses shown by Manabe et
al. in [31] confirm such behavior. Unfortunately, no values of K-factors are given
there.
The channel model parameters given by Smulders [24] and by Kunisch et al. [12] can
be used to estimate the Ricean K-factor and investigate the impact of the antennas
directivity. Those model parameters were obtained from 60 GHz channel
measurements.
Smulders model parameters [24] imply that even for the 15 dBi directive antenna, and
in the presence of a LOS path, the K-factor would be less or equal to 6 dB, and
therefore well described by the Rayleigh model. Note that in this study the directive
(receive) antenna was not pointed towards the transmitter. This may be a partial
explanation for this unexpected result.
Kunischs model parameters [12] correspond to Ricean K-factors between 7.3 dB and
25 dB (and rms between 5.7 and 1 ns, respectively). Kunischs measurement set-up
used an 8 dBi antenna at the transmitter and two receiver antennas pointed towards the
transmitter, with respective gains of 20 and 22 dBi. It appears that such an antenna
configuration can effectively reduce the multipath fading. Adaptive antennas (beam-
forming) can avoid the need of pointing the antenna manually.
Larger K-factors reduce rms, when the decay exponents of the average power delay
profile remain constant. This is also seen from the equations given in Table 2-1, where
should be considered constant. The model parameters given by Kunisch [12] confirm
that such a dependency may exist, at least within one room.

2.5.3 Overview of Channel Models

Most of the (stochastic) channel models proposed in the literature for mm-wave
channels are based on the indoor propagation model presented by Saleh and
Valenzuela [16]. This section first reviews their model. Secondly, a number of
modifications are discussed for its application to mm-wave channels. The suitability of
the FD-channel model for this frequency band is studied in Section 2.5.4.

2.5.3.1 Review of the Saleh and Valenzuela Model

The Saleh and Valenzuela model is a method to generate time-discrete channel im-
pulse responses as defined by eq. (2-1). Stochastic processes are specified to model the
46 Chapter 2 Modeling of the Frequency-Selective Radio Channel

ray arrival times {i}, the ray amplitudes {i}, and the ray phases {i}.
The ray phases are considered to be independent random variables that are uniformly
distributed over [0,2), because the phases vary over that range when the path-lengths
change by just one wavelength.
Two Poisson processes implement the ray-arrival process. Reflections are assumed to
arrive in clusters, where the first Poisson process models the arrival times of the
clusters with some fixed rate [1/s] 6. Subsequent ray-arrivals within the clusters are
realized by the second Poisson process with rate >> . Per definition, the first ray
and the first cluster arrive at = 0. A Poisson process of (ray) arrivals implies
exponentially distributed inter-arrival times, written as
p( ) = exp[ ( )] , (2-29)
where is the delay time difference between consecutive paths of the same cluster.
The probability distribution of the path gains {i} is a Rayleigh distribution.
(Therefore, the path gains including the uniformly distributed path phases { i e j } i

follow a complex Gaussian distribution.) Introducing the variables l and k for indexing
the cluster and ray-within-cluster, respectively, the mean square values of the
magnitudes { kl2 } are written

kl2 = 002 e Tl / e kl / SV , (2-30)

where {Tl} and {kl} are the cluster and ray-within-cluster arrival times, respectively,
and and SV are the corresponding power decay time-constants. This function is
called the average power delay profile (PDP), because it characterizes the average ray
power of the impulse response as a function of the excess delay-time. It is composed
of a set of exponentially decaying parts, one for each cluster of rays.
For more details on the Saleh and Valenzuela model, the reader is referred to [16].

2.5.3.2 Modifications to the Model

Several authors have applied a number of modifications to the above-described model
in order to match it to mm-wave channels.
Most of the implementations found in the literature reduce the number of clusters to
one (see e.g. [12][15]). This simplification is made, since in a typical indoor mm-
wave channel, the reflections originate all from within one room, leading to a single,

6
According to [16], the formation of clusters is related to the building superstructure, i.e., clusters
of rays typically originate from (steel-reinforced) exterior or interior walls or large metal doors or
objects. The rays within the clusters are due to reflections in the vicinity of the transmitter or receiver.
Clustering of rays is therefore a property of indoor channels at longer ranges and at lower carrier
frequencies, where propagation through walls is possible. Clustering of rays also occurs in outdoor
channels [42][44].
2.5 Application to mm-Wave Radio Channels 47

dense cluster of ray arrivals. Remember that mm-wave frequencies hardly penetrate
through building material. An exception is the work of Park [35], who gives a set of
parameters for the original, multi-cluster version of Saleh and Valenzuelas model.
Park investigated indoor channels at 60 GHz.
In several cases, the model has been augmented by a separately specified path at 0 = 0
([12][15]) in order to extend the model to Ricean channels by introducing a
(dominant) LOS path.
Smulders [24] proposes a composite average PDP, where the exponentially decaying
part of the single cluster is preceded by a constant-level part. The reason for
introducing this part is to better describe first-order reflections arriving at similar
strength due to the antenna design chosen. Such a constant-level part is also
implemented in the frequency-domain channel model proposed in this Thesis (see
Section 2.3.2).
Janssen shows in [15] how to adjust the generated discrete-time impulse responses in
order to exactly realize a given set of channel parameters {P0,K,rms}. (Note that those
are local-area mean parameters; see Section 2.2.2). He also suggests a method to
incorporate small-scale fading effects resulting from movements within a local area.
That mechanism is based upon ray-arrival directions relative to the assumed
transceiver movement.

2.5.4 Applicability of the FD-model

The proposed frequency-domain (FD) channel model characterizes the mobile radio
channel by its delay power spectrum (DPS), the Fourier transform of the spaced-
frequency correlation function. This model agrees well with the modified (single-
cluster) versions of the Saleh and Valenzuela model introduced above, because the
DPS of the FD-model is described in an almost equivalent way to the average power
delay profile of the (single-cluster) Saleh and Valenzuela model. In the FD-model,
Ricean channels may be implemented using the discrete, direct path at 0 = 0.
Moreover, a constant-level part is incorporated as in [13], which allows for a better
match to certain channel impulse responses, and which also enables varying the
maximum excess delay in some range, as investigated in Section 2.3.3.
According to the overview of channel models presented above, a single cluster of rays
is an appropriate description of mm-wave indoor channels, where the transmitter and
receiver are typically located within the same room. The suitability of the proposed
channel model for such scenarios was also confirmed by the comparison of
measurement results to computer simulations (see Section 2.4.3).
Clustering of rays can be implemented in the proposed model by modifying the DPS
accordingly, i.e., by defining a DPS consisting of multiple exponentially decaying
parts. Similarly, arbitrary outdoor channels could be realized. However, using the
model for the design of OFDM systems, I am confident that the simple model
48 Chapter 2 Modeling of the Frequency-Selective Radio Channel

introduced in Section 2.3.2 is applicable to a much wider range of actual environ-

ments. For such systems, the most important channel properties are the correlation
among (adjacent) sub-carriers and the distribution of their amplitudes (or powers).
These properties are well preserved by the model as long as the channel parameters are
matched to the environment under investigation. And these parameters can be freely
(and easily) chosen in the proposed model. It can be even proven that, for Rayleigh
fading channels and for small frequency-separations, the correlation coefficient in
frequency-domain is independent of the channel model (see Appendix A). (The corre-
lation coefficient is the normalized auto-correlation function of the squared magni-
tudes of the TF.) This finding strongly supports the claim that such a simple stochastic
model is sufficient for many air-interface design problems.

2.6 Conclusions
The main novelties discussed in this Chapter concern the so-called frequency-domain
(FD) channel model and its implementation on a computer simulation scheme. The
FD-model is the frequency-domain dual of Jakes Doppler-spectrum model [1], [3],
[4]. Just as in Jakes model the (narrowband) channels time-variability is described by
the spaced-time correlation function and by the Doppler spectrum, in the FD-model
the (time-invariant) channels frequency-selectivity is described by the spaced-fre-
quency correlation function and by the delay power spectrum. (The power spectra and
correlation functions are inter-related by Fourier transforms.) The simulation scheme
introduced directly generates realizations of channel transfer functions with well-de-
fined channel parameters. Note that a frequency selective channel is equivalent to a
time-dispersive (multipath) channel. The major advantages of the proposed models
are:
Good agreement with physical propagation channels, in particular in mm-wave fre-
quency bands and in indoor environments (see Section 2.5).
Availability of analytical expressions relating model parameters to physical chan-
nel parameters and vice versa, allowing to straightforwardly match the model to
any given environment (see Section 2.3.3).
Suitability for OFDM system design, the goal of this research (see Part II).
The simplicity of the model allows for the mathematical analysis of many aspects
of transmission schemes, like the performance evaluation and optimization of bit-
error-rates, synchronization, and channel estimation schemes (see Part II).
Availability of an efficient simulation model (see Section 2.4).
However, the extension of the simulation model from the static version presented, to a
time-variant version is rather complex. This may be a disadvantage of the FD-model.

The (physical) channel parameters specifying the FD-model are elaborately discussed.
2.6 Conclusions 49

The channel at a local area of dimensions of a few wavelengths (approx. 5 40 ) is

defined by a set of fixed parameters: the normalized (or average) received power P0,
the Ricean K-factor K, and the RMS delay spread rms. At a limited observation band-
width, however, these parameters appear to be time- (or location) variant within a local
area, because individual propagation paths are not resolvable and multipath interfer-
ence between them leads to rapid (small-scale) variations of the resulting channel im-
pulse response. Reduced-bandwidth simulations performed with the FD-model also
show a variability of these instantaneous parameters among realizations.

It was suggested that the RMS delay spread rms and the Ricean K-factor are equally
important for the characterization of frequency selective multipath radio channels. In
the next chapter, it will be shown that rms effectively specifies the number of fades per
bandwidth and their average bandwidth, while the K-factor describes the depth of
fades. Most experimental studies, however, investigate rms only. The K-factor is ana-
lyzed in rather few cases, although line-of-sight conditions and directional antennas
are commonly considered , two factors that are anticipated to increase the K-factor.
Rayleigh fading channels have a K-factor of zero.

Channel parameters depend on a number of features of the propagation environment

and of the antenna set-up. Larger rooms and more reflective building materials gener-

Table 2-2: Typical channel parameters of frequency-selective, mm-wave radio channels.

Title Comments and Antenna configu- RMS delay spread Ricean K-factor
Reference ration
Small / medium 10 ns 2.5 dB
room, LOS measurem. at 2.4, ~2.5 dBi bi-
4.75, and 11 GHz conical antennas
Small / medium [14], [15] 15 ns 3 dB
(~100 beam-
rooms, non-LOS (Rayleigh)
width)
Outdoor 17 GHz [25], [26] 30 ns 3 dB
Medium room, High-gain antenna BS: 8 dBi 5 ns 10 dB
directional ant. pointed at BS; 60 PS: ~20 dBi
GHz [12]
Computer room 45 ns 0 (Rayleigh)
(~1093 m3)
Large Hall 60 ns 0 (Rayleigh)
3 9 dBi bi-conical
(~43417 m )
60 GHz [13] antennas (~9
Corridor beam-width) 75 ns 0 (Rayleigh)
(~452.53 m3)
Lecture room 20 ns 0 (Rayleigh)
(~1394 m3)
BS: Base Station; PS: Portable station
50 Chapter 2 Modeling of the Frequency-Selective Radio Channel

ally increase the RMS delay spread rms. Higher antenna directivity decreases rms and
increases the K-factor. The presence of a line-of-sight between the transmit- and re-
ceive-antennas leads to larger K and sometimes to lower rms. Within the same envi-
ronments and with similar antenna set-ups, the frequency band has surprisingly little
influence on those parameters. A list of typical parameter-values with short descrip-
tions of the main features of the corresponding environments is given in Table 2-2.
Mm-wave channels (e.g. 60 GHz) are considered for the multimedia communications
system studied in Part II of this thesis. Most investigations of these channels conclude
that a line-of-sight between the transmitter and the receiver is required for reliable
communications. However, the results from [13], where a special antenna design was
used, suggest that the reflections can be sufficient as well.

Since the channel parameters are influenced by many factors and in ways that are hard
to predict, a method is desirable to measure them in a cheap and simple way. The next
chapter presents a method that can be used to accurately estimate these parameters
{P0, K, rms} from scans of the channels power response versus frequency. Standard
laboratory equipment can be used to apply that scheme.

2.7 References
[1] W. C. Jakes Jr., Microwave Mobile Communications. New York: Wiley-Inter-
science, 1974.
[2] J. G. Proakis, Digital Communications, 3rd edition. New York: McGraw Hill,
1995.
[3] M. J. Gans, A Power-Spectral Theory of Propagation in the Mobile-Radio
Environment, IEEE Trans. Veh. Technol., vol. VT-21, no. 1, pp. 2738, Feb.
1972
[4] R. H. Clarke, A Statistical Theory of Mobile-Radio Reception, Bell Syst. Tech.
J., vol. 47, pp. 9571000, JulyAug. 1968.
[5] P. A. Bello, Characterization of randomly time-variant linear channels, IEEE
Trans. on Commun. Systems, vol. CS-11, pp. 360-393, Dec. 1963.
[6] R. Steele, Mobile Radio Communications. New York: John Wiley and Sons,
1992.
[7] T. S. Rappaport, Wireless Communications: Principles and Practice, Upper Sad-
dle River: Prentice-Hall, 1996.
[8] European Commission, Cost Action 231, Digital mobile radio towards future
generation systems, Final Report, EUR 18957, Luxembourg, ISBN 92-828-5416-
7, 1999.
2.7 References 51

[9] S. O. Rice, Mathematical Analysis of Random Noise, Bell Syst. Tech. J., vol.
23, pp. 282332, July 1944; vol. 24, pp. 46156, Jan. 1945.
[10] S. O. Rice, Statistical Properties of a Sine Wave Plus Random Noise, Bell Syst.
Tech. J., vol. 27, pp. 109157, 1948.
[11] A. V. Oppenheim and R. W. Schafer, Discrete-time Signal Processing, 2nd edi-
tion. New Jersey: Prentice-Hall, 1999.
[12] J. Kunisch, E. Zollinger, J. Pamp, and A. Winkelmann, MEDIAN 60 GHz
Wideband Indoor Radio Channel Measurements and Model, in Proc. IEEE
Vehic. Techn. Conf. (VTC99-fall), Amsterdam, The Netherlands, Sept. 1999, pp.
23932397.
[13] P. F. M. Smulders, Broadband Wireless LANs: A Feasibility Study. PhD Thesis,
Eindhoven University of Technology, Eindhoven, The Netherlands, 1995.
[14] G. J. M. Janssen, P. A. Stigter, and R. Prasad, Wideband indoor channel
measurements and BER analysis of frequency selective multipath channels at 2.4,
4.75 and 11.5 GHz, IEEE Trans. on Commun., vol. 44, no. 10, pp. 1272-1288,
Oct. 1996.
[15] G. J. M. Janssen, Robust receiver techniques for interference-limited radio chan-
nels, Ph.D. Thesis, Delft Univ. of Techn., Delft, The Netherlands, June 1998.
[16] A. A. M. Saleh and R. A. Valenzuela, A statistical model for indoor multipath
propagation, IEEE J. Select. Areas Commun., vol. 5, no. 2, pp. 128137, Feb.
1987.
[17] O. Edfors, M. Sandell, J. J. van de Beek, D. Landstrm, and F. Sjberg, An
introduction to orthogonal frequency-division multiplexing, Division of Signal
Processing, Lule University of Technology, Research Report TULEA 1996:16
(http://www.sm.luth.se/csee/sp/publications.html).
[18] Weimin Zhang, Simulation and modelling of multipath mobile channels, in
Proc. VTC94 (IEEE Vehicular Technology Conference), Stockholm, Sweden,
1994, pp. 160-164.
[19] K. Pahlavan and A. H. Levesque, Wireless Information Networks. New York:
John Wiley and Sons, 1995.
[20] A. Chini, M. S. Tanany, and S. A. Mahmoud, Transmission of high rate ATM
packets over indoor radio channels, IEEE J. Select. Areas Commun., vol. 14, no.
3, pp. 469476, Apr. 1996.
[21] O. Edfors, Low-complexity algorithms in digital receivers. PhD Thesis, Lule
University of Technology, Lule, Sweden, Sept. 1996.
52 Chapter 2 Modeling of the Frequency-Selective Radio Channel

[22] J.-J. van de Beek, O. Edfors, M. Sandell, S. K. Wilson, and P. O. Brjesson, On

channel estimation in OFDM systems, in Proc. IEEE Vehic. Technol. Conf.,
Chicago, IL, July 1995, pp. 815819.
[23] L. M. Correia, et al., Final Report on Propagation Aspects, RACE 2067,
Deliverable R2067/IST/2.2.5/DS/P/070.b1, RACE Central Office, European
Commission, Brussels, Dec. 1995.
[24] S. A. Mohamed, G. Lvnes, E. Antonsen, R. Rkken, B. Nigeon and J. J. Reis:
Report on Propagation Measurements, RACE 2067, Deliverable
R2067/BTL/2.2.2/DS/P/ 035.b1, RACE Central Office, European Comission,
Brussels, Dec. 1994.
[25] A. Bohdanowicz, G. J. M. Janssen, S. Pietrzyk, Wideband indoor and outdoor
multipath channel measurements at 17 GHz, in Proc. VTC99-fall (IEEE
Vehicular Technology Conference), Amsterdam, The Netherlands, Sept. 1999,
pp. 19982003.
[26] A. Bohdanowicz, Wideband indoor and outdoor radio channel measurements at
17 GHz, Delft Univ. of Technol., UbiCom-Technical Report/2000/2, Jan. 2000
(http://www.ubicom.tudelft.nl/docs/).
[27] R. El Hattachi, J. M. M. de Nijs, K. Witrisal, and R. Prasad, Characterization
th
and simulation of the 18 GHz radio channel, in Proc. IEEE Benelux 6 Sympo-
sium on Vehicular Technology and Communications, Brussels, Belgium, Oct.
1998.
[28] J. Purwaha, A. Mank, D. Matic, K. Witrisal, and R. Prasad, Wide-band channel
th
measurements at 60 GHz in indoor environments, in Proc. IEEE Benelux 6
Symposium on Vehicular Technology and Communications, Brussels, Belgium,
Oct. 1998.
[29] J. J. G. Fernandes, J. C. Neves, and P. F. M. Smulders, MM-Wave Indoor Radio
Channel Modelling vs. Measurements, Wireless Personal Comm., vol. 1, no. 3,
pp. 211219, Kluwer, 1995.
[30] A. Kato, T. Manabe, et al., Measurements of Millimeter Wave Indoor Propaga-
tion and High-Speed Digital Transmission Characteristics at 60 GHz, in Proc.
PIMRC97 (IEEE 8th Intern. Symp. on Personal Indoor Mobile Radio Commun.),
pp. 149 154, Helsinki, Sept. 1997.
[31] T. Manabe, Y. Miura, and T. Ihara, Effects of Antenna Directivity on Indoor
th
Multipath Propagation Characteristics at 60 GHz, in Proc. PIMRC95 (IEEE 6
Intern. Symp. on Personal Indoor Mobile Radio Commun.), pp. 10351039, Sept.
1995.
2.7 References 53

[32] R. Davies, M. Bensebti, M. A. Beach, and J. P. McGeehan, Wireless Propaga-

tion Measurements in Indoor Multipath Environments at 1.7Ghz and 60Ghz for
st
Small Cell Systems, in Proc. 41 IEEE Veh. Techn. Conf., pp. 589-593, St.
Louis, USA, May 1991.
[33] R. J. C. Bultitude, R. F. Hahn, and R. J. Davies, Propagation Considerations for
the Design of an Indoor Broad-Band Communications System at EHF, IEEE
Trans. Veh. Technol., vol. 47, no. 1, pp. 235-245, Feb. 1998.
[34] J. Hbner, S. Zeisberg, K. Koora, and A. Finger, Simple channel model for 60
GHz indoor wireless LAN design based on complex wideband measurements,
in Proc. IEEE Vehic. Techn. Conf. (VTC97), 1997, pp. 10041008.
[35] J.-H. Park, Y. Kim, Y.-S. Hur, K. Lim, and K.-H. Kim, Analysis of 60 GHz
Band Indoor Wireless Channels with Channel Configurations, in Proc.
PIMRC98 (IEEE 9th Intern. Symp. on Personal Indoor Mobile Radio Commun.),
Taiwan, Sept. 1998, pp. 617620.
[36] G. Lvnes, J. J. Reis, and R. H. Rkken, Channel Sounding Measurements at
th
59 GHz in City Streets, in Proc. PIMRC94 (IEEE 5 International Symposium
on Personal Indoor Mobile Radio Communications), The Hague, The Nether-
lands, Sept. 1994, pp. 496500.
[37] S. W. Wales and D. C. Rickard, Wideband propagation measurements of short
range millimetric radio channels, Electronics and Commun. Eng. Journal, pp.
249254, Aug. 1993.
[38] N. Daniele, D. Chagnot, and C. Fort, Outdoor millimetre-wave propagation
measurements with line of sight obstructed by natural elements, IEE Electronics
Letters, vol. 30, no. 18, pp. 15331534, Sept. 1994.
[39] G. Lvnes, S. E. Paulsen, and R.H. Rkken, A versatile channel sounder for
th
millimetre wave measurements, in Proc. PIMRC93 (IEEE 4 International
Symposium on Personal Indoor Mobile Radio Communications), Yokohama, Ja-
pan, Sept. 1993.
[40] D. C. Cox, Delay Doppler Characteristics of Multipath Propagation at 910 MHz
in a Suburban Mobile Radio Environment, IEEE Trans. Ant. and Prop., vol. 20,
no. 5, pp. 625635, Sept. 1972.
[41] M. Flament, On 60 GHz Wireless Communication Systems. PhD Thesis,
Chalmers Univ. of Techn., Gteborg, Sweden, 2000.
[42] G. L. Turin, F. D. Clapp, T. L. Johnston, S. B. Fine, and D. Lavry, A statistical
model of urban multipath propagation, IEEE Trans. Veh. Technol., vol. VT-21,
pp. 19, Feb. 1972.
[43] H. Suzuki, A statistical model for urban radio propagation, IEEE Trans. Com-
mun., vol. COM-25, pp. 673680, July 1977.
54 Chapter 2 Modeling of the Frequency-Selective Radio Channel

[44] H. Hashemi, Simulation of the urban radio propagation channel, IEEE Trans.
Veh. Technol., vol. VT-28, pp. 213224, Aug. 1979.
Chapter 3 Channel Measurement Tech-
nique based on the FD-Level Crossing Rate

3.1 Introduction
Channel measurements are required to obtain parameters for the channel model pro-
posed in the previous chapter. A novel technique for conducting such measurements is
introduced here, which is based on the frequency-domain level crossing rate (LCRf) of
the fading radio channel.

Usually, the level crossing rate (LCR) is defined and investigated for time-dependent
stochastic processes, where it specifies the number of up-going level crossings through
a given threshold. In this chapter, the LCRf of the transfer function (TF) of a fre-
quency-selective channel is studied, specifying the average number of fades per band-
width.
From the analysis of the LCRf, which is one of the main topics of this chapter, it was
recognized that the LCRf is proportional to the RMS delay spread (RDS) rms of the
multipath-fading channel. That is, the average number of fades per bandwidth (and
also the average bandwidth of the fades) is related to rms by a given factor. This factor
was studied on the basis of the frequency-domain (FD) channel model introduced in
the previous chapter. It has been noticed that the factor does depend on the K-factor of
the Ricean fading channel, but the actual form of the channel model has little or no
impact.
This property can be used for estimating the RDS from the LCRf, which can be deter-
mined from non-coherent channel measurements (power-vs.-frequency sweeps of the
channel TF). Since the normalized received power P0 and the Ricean K-factor can be
derived from such data as well [1], full sets of channel parameters {P0, K, rms} can be

55
56 Chapter 3 Channel Measurement Technique based on the FD-Level Crossing Rate

obtained from pure power measurements. Hence, standard laboratory equipment may
be used to conduct channel measurements, as e.g. a swept-frequency continuous-wave
(CW) signal generator and a power meter or spectrum analyzer. Its simplicity makes
the method particularly useful at extremely high frequencies (> 30 GHz; millimeter
wave band), where for instance network analyzers become very cumbersome and
expensive. Moreover, large distances can be allowed between the transmitter and the
receiver, because no reference connection is required.

Since the proposed measurement method is based on a statistical model, a sufficiently

large bandwidth must be observed to obtain high accuracy, just like a large frequency-
band must be scanned with a network-analyzer in order to obtain a certain time-reso-
lution. An advantage of our method is that because of ergodicity the observation
bandwidth can be increased by analyzing the combined data of a number of narrow-
band measurements performed in a local area. The size of this local area must be
selected sufficiently small for the channel parameters not to vary due to shadowing.
Well de-correlated spectra can be obtained when the receivers location is changed in
the order of one or a few wavelengths . Especially at the millimeter wave band with
wavelengths below 1 cm, many spectrum samples can thus be taken within small
areas.

Note that due to the lack of phase information, the Fourier transform cannot be used
for transforming a magnitude TF to the delay time-domain, which would allow deter-
mining delay spread parameters as the RDS directly. However, the causality of the im-
pulse response implies that the Hilbert transform describes the relationship between
the real and imaginary components of the complex valued TF. Donaldson et al. have
applied this property for analyzing magnitude TFs [2], yielding estimates of the chan-
nels impulse responses. Their method can be an alternative way of determining delay
spread parameters using this type of measurements.

This chapter begins with the analysis of the LCRf based on the FD-channel model in-
troduced in the previous chapter (see Section 3.2). For the Rayleigh fading case it is
proven that the channel impulse response has no influence on the factor relating rms
and the LCRf. Some impact of the FD sampling interval will be seen, because level
crossings may be overlooked if it is selected too large.
A measurement procedure derived from the relation of rms and the LCRf is described
in Section 3.3, and its performance is investigated. We find that an observation band-
width of 10/rms leads to estimation errors with standard deviations in the order of 10
15 %.
The methods sensitivity to additive noise is of major concern for its practical applica-
tion, since additional level crossings caused by the noise lead to a systematic overesti-
mation of the RDS. Such noise may be due to measurement inaccuracies, when scan-
3.2 Frequency-Domain Level Crossing Rate 57

ning the channel TF. This problem is extensively studied in Section 3.4.
The Chapter is concluded in Section 3.5, where also recommendations for further work
are given.

3.2 Frequency-Domain Level Crossing Rate

The level crossing rate (LCR) is usually defined for time-domain fading (or other
time-dependent stochastic processes) as the average number of crossings per second at
which the envelope of a signal (t) crosses a specified level r in an up-going direction.
Its dimension is [s1].
Considering the TF in the frequency-domain, the LCRf gives the average number of
crossings per Hertz bandwidth at which the amplitude R(f) = |H(f)| of the TF crosses a
level r in an up-going direction. This LCRf will be denoted by NR(r), its dimension
being [s].

The derivation of the LCRf is firstly conducted on the basis of the FD-channel model,
which describes the frequency-selectivity as a continuous stochastic process (see
Chapter 2, Section 2.3). A solution is found for Ricean and Rayleigh fading channels.
We apply this result for analyzing the impact of channel parameters (rms and Ricean
K-factor) and model parameters (the shape of the delay power spectrum (DPS)) on the
LCRf. A proportional relationship between the LCRf and rms is found. In Section
3.2.2, the significance of this relation is assessed by analyzing a deterministic two-ray
channel.
As mentioned above, the LCRf can be used to estimate the RDS of time-dispersive ra-
dio channels. In order to employ this relationship for channel investigations, the power
response of the channel has to be scanned versus frequency, which is usually done at
discrete frequency points. Selecting thereby the sampling interval in the frequency-
domain too large, some level crossings may be overlooked, leading to a bias in the es-
timated rms. In order to analyze the impact of sampling, the LCRf is also derived for
the sampled case (see Section 3.2.3). This study is limited to Rayleigh channels, how-
ever, because for the Ricean case, the mathematical expressions involved require nu-
merical solutions.
Moreover, it will be shown that for Rayleigh channels the proportionality relationship
between rms and LCRf is independent of the channel impulse response.

3.2.1 Derivation of the LCRf from the Continuous FD-Channel Model

The following derivation is based on the FD-channel model introduced in Chapter 2,
Section 2.3, which characterizes the frequency-selectivity (time-dispersive nature) of
the multipath radio channel based on its delay power spectrum (DPS):
58 Chapter 3 Channel Measurement Technique based on the FD-Level Crossing Rate

0 <0
2 ( ) =0

h ( ) = . (3-1)
0 < 1
e ( 1 ) > 1

For notational convenience we introduce u = 1, expressing the shape of the DPS,

and u1 = u + 1 , u 2 = u 2 2 + u + 1 , and u 3 = u 3 3 + u 2 + 2u + 2 .

3.2.1.1 Proof of the Proportionality of the LCRf and the RMS Delay Spread
An analytical expression for the LCR of Ricean processes with cross-correlated real-
and imaginary parts of the underlying complex Gaussian process H ' ( f ) = r H ' ( f ) +
j i H ' ( f ) is given as (cf. [3], and [4], [5])

{ }
r2+2 / 2
r 2 r
cosh cos e ( sin ) + sin( ) erf( sin ) d ,
2 0
N R (r ) = 3 / 2 e
2
(3-2)
0 0 0
where 0 = 12 H ' (0) = 12 u1 / is the variance of the real or imaginary component of
0 02
H'(f) (i.e. half of the power of the scattered rays), and = and = 0
0 2 0
account for the second order statistics of H'(f). ( 0 and 0 are given in (3-4).) H'(f) is
defined by the DPS (3-1) after subtraction of the LOS-component 2(), or equiva-
lently by its auto-correlation function (ACF):

H ' (f ) = F { h ' ( )} = 1 sinc( 1f )e j1f +
1
e j 21f (3-3)
+ j 2f
Note that the real and imaginary parts of this ACF denote respectively the ACF of the
real or imaginary component of H'(f), and the cross-correlation function (CCF) be-
tween its real- and imaginary components, written as H ' (f ) = 2[ H ' (f ) + r

j r H 'i H ' (f )] . To calculate and , the curvature of the ACF r H ' (f ) and the gradient
of the CCF r H 'i H ' (f ) have to be evaluated at f = 0, yielding

d2
0 = r H ' (f ) = 2 2 3 u3 and (3-4a)
d f 2
f =0

d
0 = r H 'i H ' (f ) = 2 u 2 . (3-4b)
d f f = 0

Next, it will be shown that the LCRf for the FD-model can be expressed in the form
3.2 Frequency-Domain Level Crossing Rate 59

N R (r ' ) = rms f ( K , u, r ' ) , (3-5)

where f(K,u,r') is the proportionality factor between rms and the LCRf. Note that this
factor is constant at a given K-factor, normalized threshold level r', and DPS shape de-
fined by u. It will be seen below that the impact of the channel model (written by u) is
usually negligible.
With (3-4) we get

u2
= and
u 3 u1 u 2 u1
2 2

(3-6)
2
u2
2

= 2 3 u 3 .
u1

The threshold level r is related to the square root of the normalized received power P0
(the RMS amplitude of the TF), to eliminate the influence of P0 from the LCRf:

r' = r P0 = r 2 + u1 . (3-7)

Using (3-6), (3-7), 0 = 12 u1 , and the expression for rms(K,,u) from Table 2-1
(Section 2.3.3, page 30) yields with (3-2)

{ }
/2
N R (r ' ) = rms a e b cosh(c cos ) e ( d sin ) + d sin erf( d sin ) d ,
2
(3-8)
0

where

u3u1 u 2
2
4
a= r ' ( K + 1) 3 / 2
u3u1 ( K + 1) u 2
2

b = r '2 ( K + 1) + K
. (3-9)
c = 2r ' K ( K + 1)
u2
d= K
u1u 3 u 2
2

It is observed from (3-9) that {a, b, c, d} are expressed as functions of {K, u, r'}, i.e.,
they are independent of rms, which proves (3-5).

This result can be used for estimating rms from the LCRf, which can be obtained from
non-coherent wide-band measurements (i.e. from wide-band power measurements)
[6]. It enables the wide-band characterization of the radio channel using a very simple
measurement principle, as explained in the introduction. Section 3.3 discusses this
method in detail.
60 Chapter 3 Channel Measurement Technique based on the FD-Level Crossing Rate

3.2.1.2 Average Bandwidth of Fades

The average bandwidth of fades (ABF), BR(r), is the mean value for the bandwidth
over which the amplitude R(f) of the TF is below a specified level r. Equivalently to
the respective time-domain parameter (the average duration of fades [7]), the ABF is
written as
FR (r )
BR ( r ) = , (3-10)
N R (r )

where FR(r) denotes the (Ricean) CDF of the signal envelope, i.e., the probability that
R(f) is below the level r.
2 x2
1 2 0
r x
FR (r ) = Pr{R( f ) r } = e 0 xe 2 0
I dx (3-11)
0 0
0
In this equation, I0() designates the zeroth-order modified Bessel function of the first
kind.

Computational results of the normalized level crossing rate and the average bandwidth
of fades are shown in Figure 3-1a and b, respectively. They are compared to computer
simulations generated with the FD-simulation scheme introduced in Chapter 2, Section
2.4. The comparison clearly demonstrates that the analytical expressions describe the
statistical properties of the simulated channel appropriately.

3.2.1.3 Influence of Channel and Model Parameters

Using the analytical expressions, the influence of the channel parameters {P0, K, rms}
and the influence of the shape of the DPS (expressed by u = 1) on the LCRf and on
the ABF is studied.
One of the main results is that the LCRf is proportional to rms, as seen from the
mathematical analysis above. For this reason it is appropriate to show LCRf and ABF
normalized to rms. Furthermore, the result is independent of the NRP P0, if the thresh-
old variable r is normalized to P0 . Therefore, the shape of the LCRf is characteristic
for particular K-factors (and parameters u) as seen from the factor f(K,u,r') in (3-5).
Normalized LCRf and ABF are depicted in Figure 3-2a and b, respectively, as a func-
tion of r' and for various K-factors and parameters u.
It is observed from these figures that even in the extreme cases u = 0 (exponentially
decaying DPS) and u = (rectangular DPS), the LCRf and ABF remain similar, pro-
vided rms and K are kept constant. The dependency on u disappears completely for
K = 0, i.e., for Rayleigh fading channels. For any K, the LCRf at r' = 1 ( r = P0 )
shows little variation, which is also evident from Figure 3-3a, illustrating the factor
f(K,u,r' = 1) as a function of K with parameter u. Figure 3-3b depicts the systematic
3.2 Frequency-Domain Level Crossing Rate 61

1
10
theory
simulation
100 sims.

rms
normalized level crossing rate N (r)/ 0
R 10

1
10

2
10

3
10
40 30 20 10 0 10
threshold level r [dB], normalized to RMS amplitude

(a)
3
10
theory
simulation
100 sims.
average bandwidth of fades B (r) [MHz]

2
10
R

1
10

0
10

1
10
40 30 20 10 0 10
threshold level r [dB], normalized to RMS amplitude

(b)
Figure 3-1: (a): Normalized level crossing rate for Ricean K-factor K = 7.5 dB. Analytical re-
sults compared with results from one single simulation and averaged results from
100 simulations (simulated bandwidth 1.28 GHz; rms = 25.3 ns); (b): Average
bandwidth of fades for the same simulations.

estimation error that would yield from using (3-5) with u = 0 for estimating rms of
channels with u = {2, }. It is seen that the difference of f(K,u,r' = 1) for a rectangular
DPS and an exponentially decaying one is less than 4% at any given K-factor. The
other curves in Figure 3-3 are described in Section 3.2.2.
62 Chapter 3 Channel Measurement Technique based on the FD-Level Crossing Rate

1
10
u=0
u=2
u=

rms
normalized level crossing rate N (r)/ K = 6 dB
0
R 10 K = 6 dB
K = 9 dB
K = 15 dB

1
10

2
10

3
10
40 30 20 10 0 10
threshold level r [dB], normalized to RMS amplitude

(a)
1
10
u=0
rms

u=2
normalized average bandwidth of fades B (r)

u=
K = 6 dB
R

0
10 K = 6 dB
K = 9 dB
K = 15 dB

1
10

2
10

3
10
40 30 20 10 0 10
threshold level r [dB], normalized to RMS amplitude

(b)
Figure 3-2: (a): LCRf for various K-factors and parameters u; (b): ABF for the same parame-
ters. Both figures are normalized to rms and P0 . The small influence of the shape
of the delay power spectrum (expressed by the parameter u) is observed.

From the behavior of the LCRf-curves, conclusions can be drawn on the significance
of the channel parameters {P0, K, rms} used in this study. Each of them has a very dis-
tinct impact on the LCRf, thus one might expect them to have different impact on per-
formance results as well.
3.2 Frequency-Domain Level Crossing Rate 63

11

at the RMS amplitude

u = 0 (exponential DPS)
10
u=2
u = (rectangular DPS)
9 tworay model
2ray; K statistically eval.
8

7
rms
proportionality factor N (r = 1)/

6
close up
2.5
5
R

4 2

3
1.5
2 1.3041
1
1 0 1 2 3 4 5

0
0 10 20 30 40 50 60 70 80 90 100
Ricean Kfactor

(a)
0
10

max (1.2 dB,36 %)

max (0 dB,30 %)
1
10
systematic estimation error

max (1.4 dB,3.2 %)

max (0.96 dB,1.7 %)
2
10

3
10
DPS has u = 2
DPS has u =
tworay channel
tworay ch.; K estimated
4
10
10 5 0 5 10 15 20
Ricean Kfactor [dB]

(b)
Figure 3-3: (a): The factor f(K,u,r' = 1) = NR(r' = 1)/rms as a function of the Ricean K-factor for
various channel models. (b): Estimation error of rms, when f(K,u = 0,r' = 1) is used,
but the channels are characterized by other models.

The Ricean K-factor characterizes the depth of the fades about the mean power given
by P0. Therefore, the distribution of the signal-to-noise ratio is related to P0 and K,
which generally determines the bit error rate (BER) achievable, at a given noise and
interference power level.
64 Chapter 3 Channel Measurement Technique based on the FD-Level Crossing Rate

The RMS delay spread specifies the number of fades per bandwidth and the average
bandwidth of the fades. Considering multi-carrier transmission systems (e.g. coded
OFDM), one would expect a dependency of the BER on the number of fades within
the transmission bandwidth. From the LCRf, this parameter is seen to be strictly re-
lated to rms. Performance evaluations of OFDM systems have confirmed these obser-
vations (see Sections 4.3 and 8.2).
For (non-equalized) single-carrier modulation schemes, the relationship between rms
and the BER is even more obvious, since the delay spread determines the amount of
inter-symbol-interference, which itself impacts on irreducible error floors. Numerous
studies are available for various modulation and detection schemes, reporting on
qualitative and quantitative relations between rms and the BER. Although no general
result is known, rms is probably the most important single parameter for characterizing
the time-dispersion or frequency-selectivity of the wide-band radio channel.

3.2.2 LCRf for a Deterministic Two-Ray Channel

In this section, a brief analysis is presented in order to assess the validity of the pro-
portionality factor f(K,u,r') (which relates the LCRf to rms) for different channel mod-
els. A deterministic two-ray model is investigated for this purpose. The IR of such a
channel is defined as

h( ) = 0 e j 0 ( ) + 1e j 1 ( ) , (3-12)

where 0 1 are the ray amplitudes, {0,1} are the ray phases, and > 0 is the rela-
tive delay among the two paths. Applying the FT leads to the amplitude TF

R( f ) = H ( f ) = 02 + 12 + 2 0 1 cos(2 f + 1 0 ) . (3-13)

From (3-13), the LCRf is seen to be constant:

0 1 r 0 + 1
N R (r ) = (3-14)
0 otherwise

rms must be calculated for this model to obtain the normalized LCRf, which is the pro-
portionality factor required. Analyzing the IR yields

0 1
rms = = , (3-15)
0 + 1
2 2
+1

where = 02 / 12 is the power ratio of the two rays. Taking as K-parameter, (3-14)
and (3-15) can be used to derive the proportionality factor as a function of K (see
Figure 3-3a, {{).
In fact, the Ricean distribution is not describing the amplitude distribution of (3-13),
thus comparing to the Ricean K-factor in (3-5) might be inappropriate. One method
3.2 Frequency-Domain Level Crossing Rate 65

of deriving K from a set of amplitude values R is to calculate E{R} and E{R2}. The
ratio E{R}2 E{R 2 } can then be related to K as elaborated in [1]. Using this definition
of K, the proportionality factor is found as indicated by ++ in Figure 3-3a.
The similarity of all the results shown in this graph confirms the significance of the
relationship found between rms and the LCRf. It suggests that the proposed measure-
ment method can be applied quite generally, i.e., even if the investigated propagation
channel does not match to the model defined by eq. (3-1). This statement is further
evaluated below.
In Figure 3-3b, the error of rms is depicted, resulting from strictly using
f(K,u = 0,r' = 1) when estimating rms for channels described by the two-ray model
(and by the FD-model for u = {2,}). If a two-ray channel is evaluated with this
method, the maximum error is 36 % when both rays have equal powers, and it drops
below 10 % when is above 6 dB.

3.2.3 Derivation of the LCRf for the Sampled Case

The derivation of the LCRf for the sampled channel TF is the goal of this section. Pur-
pose of this analysis is to evaluate the impact of the sampling interval, which if se-
lected too large may lead to systematic errors in the rms estimation, because level
crossings in-between sampling instants may be overlooked. Note that the following
derivations are limited to Rayleigh channels, because for the Ricean case the mathe-
matical expressions do not yield analytical solutions.

The probability of a level crossing between adjacent samples is the probability that the
current samples magnitude Rn is larger than a specified threshold value, Rn r, while
the preceding sample Rn1 was smaller, Rn1 < r. The LCRf is thus written as
N R (r ) = Pr (Rn r , Rn1 < r ) F , (3-16)
where F [Hz] is the sampling interval in the frequency-domain, and Rn and Rn1 denote
correlated random variables. Knowledge of the bivariate cumulative distribution func-
tion (CDF) of {Rn,Rn1}, FR , R (r1 , r2 ) , is required to obtain the LCRf from
n n 1

Pr(Rn r , Rn 1 < r ) = Pr (Rn 1 < r ) Pr(Rn < r , Rn 1 < r ) = FRn 1 (r ) FRn , Rn 1 (r , r ) , (3-17)

where FR (r ) = FR , R (, r ) is the CDF of any one sample, e.g. Rn1. Using an expres-
n 1 n n 1

sion of the bivariate Rayleigh CDF given in [8] (eq. (10-10-3)), the probability (3-17)
becomes

2 2 c 2 c
Pr(Rn r , Rn 1 < r ) = e r ' Q1
2 2
r', r ' Q1 r', r ' , (3-18)
1 c 1 c 1 c
1 c

where Q1(a,b) is the Marcums Q-function (see [9], (2-1-123)), r' is the normalized
66 Chapter 3 Channel Measurement Technique based on the FD-Level Crossing Rate

threshold level r ' = r 2 0 = r P0 , and c is the correlation coefficient of the squared

magnitudes defined as c = cov(Rn2 , Rn21 ) var( Rn2 ) var( Rn21 ) , 0 c < 1. c is related to
the auto-correlation of the underlying complex Gaussian process Zn (|Zn| = Rn) by
c = 1 02 , where m = 12 E{Z n Z n*+ m }. An alternative expression for (3-18) is given as
2

(cf. [10])
1+ c sin
2 r '2
1 c
Pr(Rn r , Rn 1 < r ) = e
1
e
r '2 1 c
d . (3-19)
2 1 + c + 2 c sin

It is seen that the crossing probability (3-18), (3-19) is solely determined by the corre-
lation coefficient c and by r'. Calculating c based on the stochastic or deterministic
model of a Rayleigh distributed process (e.g. for the FD-channel model defined by
(3-1)) thus leads to the level crossing rate.

This section continues with the derivation of an approximation for (3-18) (and (3-19))
for the case that c 1, which is for instance given when the sampling interval ap-
proaches zero, F 0. Note that in this limit, the sampled case approaches the con-
tinuous case analyzed above (Section 3.2.1).
Secondly, the correlation coefficient is derived from the FD-channel model and in
Appendix A from the discrete impulse response defined by eq. (2-1), Section 2.2.2.
It will become evident that a common expression relates c to rms in the limit F 0.
Therefore, there is no influence of the channel impulse response on the proportionality
factor between rms and the LCRf for the continuous case and for Rayleigh fading
channels.
Based on the analytical results, the impact of sampling on the LCRf is evaluated.

3.2.3.1 Approximation of the Crossing Probability for the Sampled Case

Both expressions for the sampled version of the LCRf, eqs. (3-18) and (3-19), are diffi-
cult to evaluate if c is close to one. The goal of this derivation is to find an approxi-
mation for this case.
In order to find an asymptotic expression for (3-18) in the limit c 1 , we use the
relation of the Marcums Q-function to the CDF of a Ricean random variable (see [9],
(2-1-142))
b
1 Q1 (a, b) = ve ( v +a2 ) / 2
2
I 0 (av)dv . (3-20)
0

When av becomes large, I0(av) may be replaced by its asymptotic expression, as sug-
gested in [4], eq. (3.10-19). This yields the following approximation for the Ricean
CDF, being valid for ab >> 1 and a >> |b a| (see [4]), which is fulfilled for c 1 .
3.2 Frequency-Domain Level Crossing Rate 67

1 1 ba 1 (b a )2 2 b a 1 + (b a ) 2
1 Q1 (a, b) + erf e 1 a + (3-21)
2 2 2 8 a 4 8a 2
Replacing the error function by the first terms of its power series expansion, the most
important terms of (3-21) can be identified,

erf
ba
=
2

b a (b a) 3
+  . (3-22)
2 2 6 2
For the two Q-functions in (3-18), b a and a are

#(
(b a) I , II = 1 c ) 12 c
r' # 1 c
2
r ' << 1
c
, (3-23)
1 2 2
a I , II = r' r ' >> 1
c 1 c 1 c

respectively, where the approximations and the inequities hold for c 1 . Keeping
the most significant terms yields

1 1 1 1 1 1 c 1 c 1
Q1 (a, b) I , II (b a ) I , II + r '+ . (3-24)
2 2 8 a I , II 2 2 4r '

Substituting these expressions for the Q-functions in (3-18), the approximation

r 'er'
2

Pr( Rn r , Rn 1 < r ) 1 c , (3-25)

is obtained, which becomes exact in the limit c 1 . This condition is fulfilled
strictly for F 0 , i.e., for an infinitely small sampling interval, and approximately, if
the sampling theorem holds. The systematic error is less than ~1 % and ~5 %, for
c 0.9 and c 0.65, respectively, and for thresholds r' between 6 dB and 6 dB (see
Figure 3-4). Larger negative errors are evident for smaller r', since the sampling inter-
val gets more impact as the fades get deeper and narrower.

3.2.3.2 Calculation of the Correlation Coefficient c

The correlation coefficient is obtained from c = 1 02 , where m = 12 E{Z n Z n*+ m } is

2

the auto-correlation function of the discrete complex Gaussian process, which is un-
derlying the sampled Rayleigh process. Therefore, m has to be determined from the
channel model in order to calculate the LCRf for the sampled case. This calculation is
given here for the FD-channel model, and in Appendix A, for a generic discrete chan-
nel impulse response.

The spaced-frequency correlation function for the Rayleigh case (see eq. (3-3)), leads
68 Chapter 3 Channel Measurement Technique based on the FD-Level Crossing Rate

error of approximated level crossing prob. [%] 6

r = 6 dB
6 r = 3 dB
r = 0 dB
8 r = 3 dB
r = 6 dB
10
0.5 0.6 0.7 0.8 0.9 1
correlation coefficient c

Figure 3-4: Error of the approximated level crossing probability for the sampled case. Relative
error [%] as a function of the correlation coefficient c, with r' as a parameter.

to c for the FD-channel model, with m = 12 H ' (mF ) . For the two important special
cases of an exponentially decaying DPS (u = 0) and a rectangular DPS (u = ), c be-
comes

1
for u = 0
c = 1 + (2 rms F ) 2 . (3-26)
(
sinc 2 2 3 F
rms ) for u =

Introducing the series expansions of the functions involved in the above expressions,
the common approximation

c 1 (2 rms F ) 2 (3-27)
is obtained, in the limit F 0 .
It is shown in Appendix A that the same approximation (3-27) holds for any arbitrary
channel IR. Therefore, there is no dependency of the level crossing probability on any
of the channel model parameters, provided the channel is a Rayleigh channel. The va-
lidity of the approximations introduced is discussed below.

3.2.3.3 Approximated LCRf

Inspection of eqs. (3-26) and (3-27) suggests that the LCRf is proportional to rms, for
the following reasons. Provided that the sampling theorem is not violated (i.e., strictly
speaking, for F 0 ), the LCRf must be independent of the sampling interval F.
Therefore, the probability (3-18) must be proportional to F, to yield a constant LCRf
3.2 Frequency-Domain Level Crossing Rate 69

with (3-16). This implies that (3-18) is also proportional to rms, because it is seen from
(3-27) that F and rms have the same influence on c. Thus the LCRf is proportional to
rms.
Based on the approximation (3-25), this observation can be confirmed mathematically.
With (3-27) and (3-16), the LCRf for Rayleigh fading channels becomes

N R (r ' ) 2 r ' e r ' rms ,
2
(3-28)
which clearly shows the proportionality.
Note, moreover, that (3-28) is identical to the result of the continuous-frequency analy-
sis presented in Section 3.2.1 [cf. eq. (3-2), for K = 0 (i.e., = 0)]. Therefore, the
difference of the approximation (3-28) to the exact LCRf for the sampled case (which
can be calculated from (3-16) and (3-18) or (3-19), with (3-26)) quantifies the impact
of a finite sampling-interval. This impact is analyzed below.

3.2.3.4 Discussion of the Impact of Sampling

Results of LCRf vs. rms are depicted in Figure 3-5a. Due to sampling effects, there is a
deviation between the exact LCRf for the sampled case (+ + for u = 0 and
for u = ) and the linear relation obtained from the continuous model and the
approximations ({ {). An increasing number of level crossings is missed, when
the sampling interval is selected too large with respect to the channels rms.
Figure 3-5b illustrates the systematic estimation error resulting from the application of
the linear relation (3-5) for estimating rms from the power-frequency-scan of a channel
with a certain sampling interval F. To analyze this error, the exact LCRf is calculated
from (3-16) and (3-18), with (3-26) for a specific rmsF. (As seen from (3-26), the
product rmsF determines c, therefore, rms can be normalized in this way). The pro-
portional relation (3-5) (and (3-28)) is applied to calculate the erroneous estimate rms
for this observed LCRf, leading to the relative estimation error

rms LCR f , sampled case

= 1 = 1. (3-29)
rms LCR f , continuous case

Considering the Nyquist theorem, the sampling interval should be F 1/(2max), which
means at K = 0 (and u = 0) that rmsF 1/20, following eq. (2-24), Section 2.3.3. In
this range of rmsF, the maximum bias is below 4 %, at r' = 1. The errors increase for
smaller thresholds r', where the fades get deeper and narrower, and also for larger ones
(see Figure 3-5b; curves and ). Investigating the channel model with the
rectangular DPS, i.e., u = , the errors decrease, since channels having shorter impulse
responses imply higher oversampling (see Figure 3-5b; curve , for r' = 1).
70 Chapter 3 Channel Measurement Technique based on the FD-Level Crossing Rate

0.14
proportional relation (continuous case)
sampling interval F; u = 0 (exponential DPS)

normalized FDlevel crossing rate N (r = 1)F

0.12 sampling interval F; u = (rectangular DPS)

R
0.1

0.08

0.06

0.04

0.02

0
0 0.02 0.04 0.06 0.08 0.1
normalized RMS delay spread and sampling interval rmsF

(a)

10
relative error [%]

15

20
u = 0; r = 0 dB
u = 0; r = 3 dB
25
u = 0; r = 6 dB
u = 0; r = 10 dB
u = 0; r = 3 dB
30
u = 0; r = 6 dB
u = ; r = 0 dB
35
0 0.02 0.04 0.06 0.08 0.1
normalized sampling interval rmsF

(b)
Figure 3-5: Influence of sampling on the proportionality of rms and the LCRf. Parameters:
Rayleigh fading; F = 1 Hz. (a): LCRf vs. rms at r' = 1, u = {0, }. Continuous and
sampled cases. (b): Relative error of the proportional relationship (continuous
case). All results for u = 0 except for , where u = .
3.2 Frequency-Domain Level Crossing Rate 71

3.2.3.5 Independence from the Channel Impulse Response

A number of simulation results are depicted in Figure 3-6 in order to support the
claimed independence of the relation between the LCRf and rms from the channel IR.
Sets of 500 impulse responses were generated for this purpose, where each IR con-
sisted of L = 15 rays with unit variance, Rayleigh distributed magnitudes, and arrival
times being uniformly distributed within a unit time interval. The IRs were then nor-
malized with respect to power and rms. Next, the exact LCRf was calculated for each
IR (for r' = 1), using equations (3-16) and (3-18) (or (3-19)), with the correlation coef-
ficient obtained from (A-6) (see Appendix A).
Figure 3-6 illustrates the error compared with the proportionality relationship as a
function of the (normalized) sampling interval rmsF. The errors mean, minimum and
maximum values are indicated as well as their standard deviation.
A systematic error is evident from this figure, which is zero at F = 0 and which in-
creases with F. Remarkable are also the errors small standard deviations, meaning
that c is largely independent of the structure of the IR. Minimum and maximum
values are evidence for an asymmetric distribution of errors about the mean. Thereby,
above average values of c lead to larger negative errors of the LCRf. A simulation for
a smaller number of rays would show similar mean errors, but increased standard
deviations.
To obey to the sampling theorem, F < 1 (2 max ) 1 (20 rms ) should be given. In this
range of F, the mean error is less than ~2 %. The mean error as a function of F is com-

2
relative error of LCRf [%]

10 minimum error to maximum error range

mean of errors
standard deviation of errors
12
0 0.02 0.04 0.06 0.08 0.1
normalized sampling interval F
rms

Figure 3-6: Relative error of the LCRf compared with the proportional relation (at r' = 1).
These results are for simulated random impulse responses with L = 15 rays.
72 Chapter 3 Channel Measurement Technique based on the FD-Level Crossing Rate

parable to the curve shown in Figure 3-5 for the case of the rectangular DPS. Indeed,
the ensemble of simulated IRs agrees with this model of a rectangular average power
delay profile.

3.2.4 Discussion and Summary

In Section 3.2.1, it was shown that a strict proportional relationship exists between the
level crossing rate of a frequency-selective radio channel in the frequency domain
(LCRf, NR(r')), and the channels RMS delay spread rms. This relation is written as
N R (r ' ) = rms f (r ' , K , u ) (3-5), where the proportionality factor f (r ' , K , u ) is a function
of the threshold level at which the LCRf is observed (r'), the Ricean K-factor of the
channel (K), and channel parameters (expressed by u). It was suggested to use this re-
lation for estimating the channels rms using simple swept-frequency power measure-
ments, from which the LCRf can be determined. The measurement principle is further
discussed in Section 3.3.
The Ricean K-factor can be determined prior to applying (3-5). However, the impact
of the current channel on f (r ' , K , u ) (expressed by u) remains an uncertainty, and
might thus be a source of systematic estimation errors.
In Appendix A it is shown for Rayleigh fading channels that the channel model does
not influence the proportionality factor between the LCRf and rms. This finding
strongly supports the claim that the LCRf is a valuable means for estimating rms in a
simple way.
Analytical results from eq. (3-5) have shown that the proportionality factor does de-
pend on the channel model in the general Ricean case. The factor has been compared
for widely varying channel models, suggesting that the impact is very small and can
thus be neglected in many cases. So is the difference of f (r ' , K , u ) for a rectangular
delay power profile and an exponentially decaying one less that 4 % at any K-factor,
and at r' = 1.
A deterministic two-path channel model was analyzed in Section 3.2.2 to assess the
relationship between the LCRf and rms. Although this model is very different to the
Rayleigh and Ricean stochastic models on which eq. (3-5) was based, similar propor-
tionality factors were found.

The proportionality relationship (3-5) was derived from the FD-channel model (see
Section 2.3), which describes the frequency-selective transfer function of the multipath
channel as a continuous, WSS stochastic process. In Section 3.2.3, the LCRf was de-
rived for the sampled case, i.e., for transfer functions given at discrete frequency-in-
stants. When the sampling interval is selected too large, then some level-crossings in-
between sampling instants may be overlooked, and the LCRf deviates from the value
suggested by (3-5). The systematic errors introduced have been analyzed.
3.3 Application to Channel Measurements 73

3.3 Application to Channel Measurements

In this Section, we discuss the application of the relationship between the LCRf and
the RMS delay spread rms for estimating rms. A measurement procedure is introduced
and validated using experimental data. Finally, the estimation accuracy is evaluated
based on computer simulations.

3.3.1 Channel Measurement Procedure

A practical channel measurement procedure based on the power response in the fre-
quency domain is outlined in this section. Although the procedure operates only on
power measurements, where no information about the phase of the received signal is
available, all the important channel parameters (i.e. the average received power P0,
Ricean K-factor, RMS delay spread rms) and their statistics can be obtain.
The measurement setup comprises of a transmitter and a receiver. The transmitter is
made up of a sine wave generator, an up-converter, an amplifier, and a transmit anten-
na. The receiver consist of an antenna, a preamplifier, a down-converter and a power
meter (e.g. a spectrum analyzer) [11][14].
Calibration measurements are necessary in order to evaluate the amount of noise and
inaccuracy caused by the system itself. This stage gives calibration data, which can
then compensate for the frequency response of the measurement system. It also quanti-
fies the noise caused by the system, which is necessary to evaluate the influence of
noise on the rms estimation (see Section 3.4).
Once the calibration data has been gathered, the channel measurements can be per-
formed. For each point within the frequency range to be measured (e.g. 10/rms band-
width) the transmitter sends a sine signal, the power of which is measured by the
power meter tuned to the frequency of the transmitted wave. At the end of the proce-
dure a sampled magnitude transfer function in the measured frequency range is com-
posed Rn = |H(nF)|, n = {1,2,,N}. From the data, the channel parameters are obtained
in the following steps:
1. Compensation for the measurement systems characteristics is performed by sub-
tracting the calibration data [dB] from the measurement data.

N
2. The average received power P0 = 1
N n =1
Rn2 is estimated.

3. The Ricean K-factor is determined, e.g. by the method given in [1], based on the
average power P0 and the average amplitude R =
N
1
N n =1
Rn .

4. The RMS delay spread rms is estimated from the LCRf at threshold level r = P0
(i.e., at r' = 1) 7, using (3-5) and (3-30).

7
Other threshold levels could be selected as well, but r' = 1 is most simple to determine, it is close
74 Chapter 3 Channel Measurement Technique based on the FD-Level Crossing Rate

The above procedure describes the steps to be followed to obtain the channel parame-
ters from one measured TF. Several measurements performed at one particular loca-
tion can be combined, leading to improved estimation results, as elaborated below.

3.3.1.1 Approximation of the Proportionality Factor

For measurements, the following approximation of f(K,u,r') at u = 0 and r' = 1 can be
used, which results in errors below 1 % for all values of K, compared to the exact
f(K,u = 0,r' = 1):

14 K 3 / 2 + 1.3041 K 1
~
f ( K , u = 0, r ' = 1) = K +1 (3-30)
K K + 0.31 K >1

3.3.1.2 Increasing the Effective Measurement Bandwidth

The accuracy obtained depends heavily on the observation bandwidth, since the meas-
urement method is based on a statistical model. By increasing the bandwidth, the
number of detected level crossings is increased, and thus the accuracy of the estimated
RDS is enhanced. A valuable advantage of the proposed method is found in the fact
that because of ergodicity the observation bandwidth can be extended not only by
increasing the bandwidth of the measurement, but also by combining data from several
measurements that are performed within a local area. This local area must be suffi-
ciently small (maximum size 5 40) so that we can assume the channel parame-
ters (and thus the statistical properties of the channel) to be constant. In other words,
the shadowing must be constant (see Section 2.2). Data collected for such a cluster of
measurements is analyzed as follows. The NRP and the K-factor are determined by
simply combining all measured amplitudes and calculating R and P0 for the resulting
data set. Level crossing rates must be calculated for each measurement separately, but
at a common threshold r'. Consecutively, they are averaged to obtain the LCRf for the
combined data set.

The proposed method is evaluated below, using channel measurements performed with
a network analyzer, and using time-domain channel simulations.

3.3.2 Validation of the Method using Measurement Results

Coherent measurements conducted with a network analyzer allow the calculation of
the channels IR, from which a reference-value of rms can be derived. Clusters of six

to the maximum of the NR(r') curves thus yielding nearly optimum accuracy, and the dependency of
f(K,r',u) on u is very small at this r' (cf. Figure 3-2a). One could also estimate rms at a set of threshold
levels and consecutively combine the estimates to enhance the accuracy. In [15], performance results
are depicted for such a scheme, indicating some improvement.
3.3 Application to Channel Measurements 75

1
10
measurement
theory

rms
normalized level crossing rate N (r)/ 0
R 10

1
10

2
10

3
10
40 30 20 10 0 10
threshold level r [dB], normalized to RMS amplitude

(a)

3
at the RMS amplitude

single spectra, 11 GHz

clusters combined, 11 GHz
theoretical factor
2.5

2
rms
proportionality factor N (r = 1)/

1.5
R

0.5

0
0 1 2 3 4 5
Ricean Kfactor

(b)
Figure 3-7: (a): Comparison of measured and theoretical level crossing rates for matched
channel parameters. (b): Indicated points: Empirical proportionality factor be-
tween the estimated LCRf obtained from the amplitude TF and the reference rms
obtained from the IR. Curve (): Theoretical factor f(K,u = 0,r' = 1). The mis-
match indicates the estimation error.

measurements were investigated within the local areas (of diameter = 2). Each TF
scanned had 1 GHz bandwidth around a center frequency of 11.5 GHz [16] (see foot-
note 5 on page 35).
76 Chapter 3 Channel Measurement Technique based on the FD-Level Crossing Rate

Figure 3-7a shows the empirical LCRf of one measured TF as a function of r' and the
analytical curve for the estimated parameters K and rms (assuming u = 0). The good
agreement demonstrates the suitability of the FD-model for characterizing the fre-
quency-selective channel.
To assess the accuracy of the estimated RDS, the proportionality factor N R (r ' = 1) / rms
is depicted in Figure 3-7b as a function of the Ricean K-factor. The estimated LCRf
N R (r ' = 1) and K-factor K were derived from the amplitude TF Rn = |H(nF)| using the
proposed method to be evaluated and (the reference) rms was calculated from the
channel IR using a conventional method [17], [18].
The RDS was estimated for every single measured TF and also for the combined data
sets of each cluster of six measurements. The theoretical factor NR(r' = 1)/rms =
f(K,0,1) is shown in the same figure for comparison. The distance between this curve
and the data points indicates the estimation error. It is observed from this figure that
the estimation error is decreased significantly by investigating the combined data of
the measurement clusters.

3.3.3 Validation by Time-domain Channel Simulations

Channels obtained from a time-domain (TD) channel simulator were analyzed for
evaluating estimation errors. This approach was taken because rms is defined from the
channels IR the output of the TD simulator.

3.3.3.1 Computer Simulation Scheme

The TD simulation model assumes a line-of-sight ray at = 0, a Poisson process of
ray-arrivals (of approx. 60 rays), an exponentially decaying average power delay pro-
file, Rayleigh distributed ray amplitudes, and uniformly distributed ray phases (com-
pare [17] for one cluster; and Section 2.5.3). In a second step, the generated impulse
responses were normalized to get the required K-factor K, rms = 1 and P0 = 1 [19], al-
lowing for simple evaluation of the estimation error. Applying the Fourier transform to
the generated IRs, (complex-valued, discrete-frequency) TFs were obtained.
The FD-sampling interval of these TFs, F, is related to the maximum delay spread max
of the channel impulse responses, which can be written as a function of the Ricean K-
factor and rms (see eq. (2-24) on page 32). Using this expression of max, the oversam-
pling factor OS is defined as

1 1 2K + 1
OS = = . (3-31)
2 max F 20 rms F K +1

In the simulations of this section, OS = 1 was used.

3.3 Application to Channel Measurements 77

3.3.3.2 Accuracy of the Estimation Technique

After generating TFs for well-defined channel parameters, the RMS delay spreads
were estimated using the proposed procedure, yielding rms . The estimation error is de-
fined as
rms rms
= . (3-32)
rms

Performing simulations for different values of K and different observation bandwidths,

the mean and standard deviation of the error are investigated. Results are depicted in
Figure 3-8a, as a function of the normalized observation bandwidth. (The PDF of the
estimation error was observed to be appropriately described by a Gaussian distribu-
tion.)
For the estimation method described above, a small systematic estimation error (~
5 %) and decreasing standard deviation with increased bandwidth can be seen from
the results. The systematic error agrees fairly with the bias introduced by the sampling
of the TF, which was analyzed in Section 3.2.3 for K = 0. (Note that in the simulation
here rmsF = 0.05 at K = 0.) The results given in Figure 3-5b can thus be used for can-
celing this systematic error.
However, the bias is also partly caused by a bias in the estimated K-factors, because
the K-factor estimation is performed prior to the rms estimation. An evaluation of the
accuracy of this estimation step is a recommended topic for further work. It can result
in a compensation table that indicates the required corrections as a function of the
normalized sampling interval and K-factor.
In Figure 3-8b, the standard deviation of the error is shown as a function of the number
of level crossings. This diagram clearly reveals that the number of observed level
crossings determines the estimation accuracy. Note that the variance appears to be
proportional to the reciprocal of the level crossing rate, i.e., the standard deviation
1 N R (r ' ) .

The results given in Figure 3-8 specify the estimation error of rms compared to the lo-
cal area mean parameters, i.e., compared to the constant parameters of the WSSUS
channel model. Considering a limited bandwidth, however, there is a certain variation
of the instantaneous channel parameters within a local area anyway, because the
structure of the impulse response is not resolved completely (see Section 2.2). This
variation of the instantaneous channel parameters was investigated in Section 2.2.3
(see Figure 2-1). The results given there should be compared to the accuracy of the
novel estimation technique for rms (Figure 3-8), because the estimation error of the
novel method includes the variation of the channel parameters due to the limited ob-
servation bandwidth.
78 Chapter 3 Channel Measurement Technique based on the FD-Level Crossing Rate

RMS delay spread estimated from power transfer function; OS = 1

rms
40
standard deviation
35 mean error
K=0
30 K=1
relative estimation error of rms [%]

K=4
25 K = 10

20

15

10

10
0 1 2
10 10 10
observation bandwidth in 1/rms

(a)

RMS delay spread estimated from power transfer function; OS = 1

rms
2
10
K=0
K=1
standard deviation of the rms estimation [%]

K=4
K = 10

1
10

0
10
0 1 2
10 10 10
number of level crossings observed (at r = 1)

(b)
Figure 3-8: Relative estimation error of rms, derived from sets of hundred TD simulations. (a):
Mean and standard deviation of the estimation error as a function of the nor-
malized bandwidth. (b): Standard deviation of the estimation error as a function
of the number of level crossings.

It is seen that the standard deviations of these parameters are in the same order of
magnitude as the estimation errors of the proposed rms estimation technique. The stan-
3.4 Analysis of the Influence of Noise 79

dard deviation of rms due to bandwidth limitation is by a factor of about two below the
standard deviation of the estimation error, at a given bandwidth. The error between the
estimated rms and the instantaneous rms was also evaluated in order to analyze the cor-
relation between the two values. A small decrease of standard deviation compared
with the result shown in Figure 3-8 (about 10 %) indicates that the deviations from
the local-area-mean parameters are partly correlated. A correlation coefficient of about
0.45 was obtained from the computer simulations.

3.3.4 Discussion of the Measurement Method

It is seen that the RDS can be estimated with reasonable accuracy when the observa-
tion bandwidth is larger than 10/rms, or when more than ten level crossings are pre-
sent. While the required bandwidth might be higher than the bandwidth needed for
other measurement techniques, it can be increased easily by combining multiple meas-
urements performed within a small local area. Another advantage lies in the simplicity
of the hardware that can be used. It makes the method particularly interesting at ex-
tremely high frequencies (millimeter wave band), where e.g. network analyzers be-
come very expensive and cumbersome due to the phase reference required, which is
very difficult to provide over large distances. A clear advantage compared to other
channel sounding techniques is that no specific equipment must be designed. A con-
tinuous wave frequency generator and a power meter or spectrum analyzer may be
used to collect measurement data.
Preliminary measurement campaigns performed at Delft University of Technology
([11][14]) have shown the practical suitability of the described methods. But they
have also led to the discovery of an interesting problem. Even a small measurement
noise level may increase the number of level crossings detected. Improved measure-
ment methods have been proposed in these references to reduce this effect. The fol-
lowing section gives the theoretical analysis of the impact of measurement noise.
These investigations show that the frequency-domain sampling interval of the meas-
urements should be selected as large as possible to minimize the influence of noise.

3.4 Analysis of the Influence of Noise

A major practical problem of the novel technique for estimating rms is its sensitivity to
additive noise in the measured power TF. This noise may introduce additional level
crossings and thus lead to systematic overestimation of rms, as illustrated in Figure
3-9. In this section, the influence of noise on the LCRf is analyzed mathematically. In-
troducing the noise to the frequency-domain (FD) channel model, which is modeled as
a continuous stochastic process, quantifies the impact of noise for the general case of
Ricean fading channels. However, this analysis is not sufficient, because a measure-
ment is typically performed on discrete-frequency instants, spaced by the sampling
80 Chapter 3 Channel Measurement Technique based on the FD-Level Crossing Rate

55

56

transfer function magnitude [dB] 57

58

59

60

61

62

63
transfer function without noise
64 threshold
upcrossings
65
450 500 550 600 650 700
frequency [MHz]

(a)

55

56

57
transfer function magnitude [dB]

58

59

60

61

62

63
transfer function with noise
64 threshold
upcrossings
65
450 500 550 600 650 700
frequency [MHz]

(b)
Figure 3-9: Influence of noise on the level crossing rate. (a): Channel transfer function without
additive noise. (b): Channel transfer function with additive noise.

interval F. The result from the analysis of the continuous-time model predicts a large
impact of the sampling interval on the noises influence, but it fails to describe that
impact accurately, due to the inappropriate mathematical model. To elaborate on this
effect, we also investigate the LCRf for the sampled case. This analysis is again limited
to Rayleigh channels, however.
The analytical results presented can be used for evaluating the systematic estimation
3.4 Analysis of the Influence of Noise 81

error due to a given noise level. Unfortunately, the result is less suitable for correcting
this systematic error. Computer simulations have indicated that applying the analytical
result can indeed reduce the systematic estimation error; the errors variance, however,
is boosted at the same time.
More successfully, a method was applied that reduces the influence of noise on the
counted LCRf by introducing a second threshold [12]. Thereby, level crossings are ig-
nored, which are likely due to noise and not due to fades. For a good performance, the
threshold separation must be adapted to the noise and channel parameters. This
method is presented in Section 3.4.5 and basic performance results are given.

3.4.1 Mathematical Modeling

The measurement noise is introduced to the FD-channel model (see eq. (3-1), Section
3.2.1 and Section 2.3) as an additive band-limited Gaussian noise component. To
model the independence of the additive noise samples, band-limitation to 1/(2F) is
assumed, where F [Hz] is the sampling interval in the frequency-domain. Note that the
sampling interval has been introduced although the mathematical model is a continu-
ous one. The magnitude of the noise process is defined by the power density N [1/s],
yielding the noise power PN = N/F = E{|nn|2}. The {nn} denote independent complex
Gaussian noise samples that are added to the complex-valued TF H(nF). The FD-
channel model and the model for the noise process are depicted in Figure 3-10.

h() [dB]
direct path
(line-of-sight) 2
constant-level part

exponentially

decaying part
measurement
noise N Excess
delay [s]
1/(2F) 1 1/(2F)

Figure 3-10: Model of the delay power spectrum (DPS) with additive measurement noise.

3.4.1.1 Definitions
The noise power is related to the variance of the (zero-mean) complex Gaussian noise
process underlying the Ricean fading process. This is written N' = PN/(20), where
0 = 12 H ' (0) = 12 u1 . H'(f) is the auto-covariance of the channel transfer function
(see eq. (3-3)). (Note that the line-of-sight component is excluded from the normaliza-
tion term.)
The additive noise component increases the average power of the observed TF and it
decreases the K-factor as
82 Chapter 3 Channel Measurement Technique based on the FD-Level Crossing Rate

~ N'
P0 = P0 + PN = 2 + 2 0 (1 + N ' ) = P0 1 +
K +1 . (3-33)
~
K = K (1 + N ' )

Independence of the noise processes is assumed. The tilde indicates parameters influ-
enced by the additive noise. Since usually N' << 1, this influence is small and can be
neglected in the mathematical derivations and in channel parameter estimation. That
~ ~
is, K K , P0 P0 , and furthermore ~
r ' r' .
The above definitions specify the first order statistics, i.e., (the change of) the ampli-
tude distribution. Second order statistics, as the LCRf can be derived from the spaced-
frequency correlation function, which for our model is given by
~
H (f ) = H (f ) + PN sinc( f / F );
e j 21f . (3-34)
H (f ) = 2 + 1 sinc( 1f )e j1f +
+ j 2f

It will be seen below that the influence of noise depends strongly on the sampling in-
terval F. Therefore, this parameter must be appropriately defined. According to Ny-
quists sampling theorem, F 1/(2max) must be given, where max is the maximum ex-
cess delay of the channel. The oversampling factor OS, as defined by eq. (3-31) will be
used to specify F in relation to rms and K.

Calculating the level crossing rate for the continuous model described above (using the
equations given in Section 3.2.1) yields a compact expression quantifying the impact
of noise for Ricean channels. These derivations are outlined in Section 3.4.2 and in
Appendix B. In practice, however, the LCRf is determined from measurements taken
at discrete frequency instants. The LCRf is quite sensitive to this sampling; therefore
we also analyze the LCRf for the sampled case, called the discrete-frequency LCRf
(see Section 3.4.3). This analysis is limited to the Rayleigh fading case, however. For
the Ricean case, a constant correction factor is introduced to the result from the con-
tinuous-frequency analysis, which can partly correct for sampling effects (Section
3.4.2.1).

3.4.2 Derivation of LCRf from the Continuous FD-Channel Model

It is shown in Section 3.2.1 that without additive noise, the LCRf, NR(r'), and the RMS
delay spread rms are proportional as N R (r ' ) = rms f (r ' , K , u ) (3-5). The proportionality
factor f(r',K,u) the normalized LCRf (of a channel having rms = 1) is a function of
the Ricean K-factor K, the normalized threshold level r ' = r P0 (normalized to the
average power), and the channel model parameter u. The latter has little impact on
f(r',K,u) and can thus be let u = 0 in practice (see Figure 3-2 and Figure 3-3).
The straight, dashed lines in Figure 3-11 show this proportionality relationship. In or-
3.4 Analysis of the Influence of Noise 83

0.16
K=0

frequencydomain level crossing rate NR(r=1) [s]

0.14 K = 6 dB
K = 10 dB
prop. w/o noise
0.12

0.1

0.08

0.06

0.04

0.02

0
0 0.01 0.02 0.03 0.04 0.05
RMS delay spread rms [s]

Figure 3-11: Relation between rms and the LCRf with and without noise. The asymptotic behav-
ior is seen. Relative noise power N' = 0.01; sampling interval F = 1 Hz. LCRf for r'
= 1 and cs = 1.

der to apply it for estimating rms, the factor f(r',K,u) must be known. It can be calcu-
lated from eq. (3-8) with (3-9), or from the approximation (3-30).

Measurement noise may raise the level crossing rate for a given channel. The idea of
this analysis is to quantify the impact of noise on the LCRf, and to use the resulting
equation for correcting for it.
With the mathematical model and the definitions introduced in Section 3.4.1, and with
two approximations (see Appendix B), a rather simple relation (3-35) between the
~
LCRf influenced by noise, N R (r ' ) , and rms is found.

(N~ R ) 2
(r ' )
N' 2 2
F2
c s h (r ' , K ) = rms f (r ' , K , u ) (3-35)

Figure 3-11 illustrates its behavior for various K-factors. It is seen that noise deter-
mines the level crossing rate at low rms, where the number of level crossings due to
the multipath channel is low.
Just as the factor f(r',K,u) is the normalized LCRf for the multipath radio channel, so is
h(r',K) the LCRf for the additive, uncorrelated measurement noise, scaled by N ' F
(8). Note that this result was obtained from the continuous FD-channel model, consid-
ering noise with a flat, band-limited spectrum. The constant cs is used to correct for

8
This statement becomes clear when the special case of a flat fading channel is considered (i.e.,
rms = 0), where level crossings are due to noise only.
84 Chapter 3 Channel Measurement Technique based on the FD-Level Crossing Rate

0.38

0.36

factor h(r = 1,K)

0.34
h(r=1,K=0) = 0.37646
0.32

0.3

0.28
20 15 10 5 0 5 10 15 20
Ricean Kfactor [dB]

Figure 3-12: Evaluation of h(r',K) at r' = 1.

sampling effects that are overlooked by this approach. A thorough explanation of this
issue and the value of cs are given in Section 3.4.2.1.
Note that the right-hand-side of (3-35) is the LCRf for the noiseless case (compare eq.
(3-5)), therefore, the left-hand-side quantifies the influence of noise on the LCRf. To
evaluate the expression, the factor h(r',K) is needed, which is shown in Figure 3-12 as
a function of K, at r' = 1. It is obtained from (see Appendix B)

h(r ' , K ) =

3
2
( )
r ' ( K + 1) e r ' ( K +1) K I 0 2r ' K ( K + 1) , (3-36)

where I0 is the zero-th order modified Bessel function of the first kind.
The components of the LCRf due to the fading channel and due to measurement noise,
and the total LCRf can be seen as the sides of a right-angled triangle. Its hypotenuse
stands for the LCRf of the noisy measurement, while the adjacent sides are the compo-
nent LCRs, as illustrated in Figure 3-13. It appears that the two uncorrelated noise
processes correspond to LCR-components in orthogonal directions of a plane, while
their vector sums length corresponds to the total LCR. Note that this observation was
made from (3-35). It was not tried to prove mathematically if this is a general property
of the LCR of sums of (independent) stochastic processes.
The nature of the square root in (3-35) indicates that the estimation of rms becomes
~
more difficult when the measured LCRf, N R (r ' ) , gets in the range of the subtracted

~
N R (r ' ) noise
influenced LCRf
N'
c s h( r ' , K )
F
LCRf due to
measurement
noise
rms f (r ' , K , u ) LCRf due to
frequency-selectivity

Figure 3-13: Interrelation of level crossing rates in the noise-influenced case.

3.4 Analysis of the Influence of Noise 85

term ( N ' F )c s h(r ' , K ) . No meaningful result can be obtained when it is smaller than
this value. This difficulty is also seen from Figure 3-11 and Figure 3-13. It corresponds
to the case where the observed LCRf is less than the (expected) LCRf-component due
to the (specified) measurement noise. In Figure 3-11, this value is seen as the LCRf at
rms = 0. In Figure 3-13 it would mean that the hypotenuse becomes shorter than the
side representing the noise component, which is impossible.
A main conclusion drawn from (3-35) is that the sampling interval F has a major im-
pact on the influence of noise. Doubling the sampling interval has the same effect as
reducing the noise power by a factor of four. While the sampling interval can be easily
increased as long as the sampling theorem is not violated , it is usually very diffi-
cult to reduce the noise. Thus the sampling interval should always be selected as large
as possible. An over-sampled measurement should be down-sampled appropriately
prior to the estimation of rms.

3.4.2.1 Correcting for Sampling Effects

The solution given by eq. (3-35) was obtained through the introduction of the measure-
ment noise to the continuous time (and -frequency) model shown in Section 3.4.1. The
measurement is done on discrete frequency instants however, which leads to a system-
atic underestimation of the LCRf compared to the theory, because the continuous
model also considers level crossings in-between sampling points. Such crossings occur
mostly due to the interpolation implied by the flat, band-limited model for the meas-
urement noise (see also [20]).
Computer simulations have been performed in order to quantify this error and to in-

0.38

0.36

0.34
systematic error in h(r = 1, K)
due to sampling effects
0.32
factor h(r = 1,K)

0.3

0.28

0.26

theoretical
0.24 simulation; N = 10 dB
simulation; N = 20 dB
0.22 simulation; N = 30 dB
simulation; N = 40 dB
0.2
0 0.01 0.1 1 10
Ricean Kfactor

Figure 3-14: Simulated factor h(r',K) compared to the theoretical one. A constant offset of
approx. 22% is observed.
86 Chapter 3 Channel Measurement Technique based on the FD-Level Crossing Rate

vestigate the possibility of incorporating it in the constant h(r',K). Results are depicted
in Figure 3-14, where the simulated h(r',K) as a function of K is compared to the theo-
retical one, at r' = 1. The simulation was performed for different values of N'. En-
sembles of flat-fading channels (rms = 0) with Ricean amplitude distribution according
to K were generated. It is observed that the simulated h(r',K) is below the theoretical
one, by approximately 22 %.
Multiplication of h(r',K) in eq. (3-35) by cs = 0.78 can account for this modeling error.
The suitability of this correction factor will be seen from numerical evaluations and
computer simulations presented below.

3.4.3 Discrete-Frequency Analysis for Rayleigh Channels

The correction factor cs can account for the modeling deficiencies of the additive noise
process. However, as the sampling interval is increased, also level crossings may be
overlooked, which are due to fades. For Rayleigh channels, all effects related to the
sampling of the noisy magnitude TF can be extracted from the discrete-frequency
analysis presented in Section 3.2.3.
The LCRf for a discrete Rayleigh process has been shown to be a function of the
threshold level r', and the correlation coefficient between the squared magnitudes of
adjacent samples, c. In order to evaluate the impact of noise, this correlation coeffi-
cient has to be determined for the frequency-selective channel plus additive noise.
c is related to the auto-correlation coefficients of the underlying complex Gaussian
process Zn as c = 1 02 , where i = 12 E{Z n Z n*+i }. Eq. (3-26) gives the correlation
2

coefficient for the noiseless case, for an exponentially decaying and a rectangular de-
lay spectrum. In the limit F 0 , c can be approximated as given in eq. (3-27), re-
gardless of the channel model or channel impulse response.
For the LCRf with noise, we find
c
~c = . (3-37)
(1 + N ' ) 2

The exact expressions for ~c (i.e., (3-37) with (3-26)), together with (3-18) (or (3-19)),
is used in Section 3.4.4 to evaluate the influence of the sampling interval on the LCRf.
Approximation (3-27) is used below to verify the equation for the noise influenced
LCRf obtained from the continuous model, (3-35), and to evaluate cs analytically.

3.4.3.1 Validation of the Result Obtained from the Continuous Model

In the limit c 1 , the level crossing probability for the sampled case can be approxi-
mated by eq. (3-25) (see Section 3.2.3.1). This condition is fulfilled for F 0 (mean-
ing that the continuous case is approached) and for N' << 1. Under these assumptions,
and using (3-37) and approximation (3-27), the LCRf for noise influenced Rayleigh
3.4 Analysis of the Influence of Noise 87

fading channels becomes

~ 2 N' 1
N R (r ' ) 2 r ' e r ' + rms
2
. (3-38)
F 2 2 2
After some manipulations it is seen that this equation is equivalent to (3-35), for K = 0.
Based on this result we can also verify the value of cs for Rayleigh channels. Letting
rms = 0 leads to the special case of the flat channel, where level crossings are caused
by the additive noise only. Comparing (3-38) and (3-35) then gives

c s h(r ' , K = 0) = 2 / r ' e r ' .
2
(3-39)

From this expression and (3-36) (for K = 0) follows c s = 6 0.78 . The simulation
results shown in Figure 3-14 suggest that this correction factor is also appropriate for
any other K-factor.

3.4.4 Evaluation and Application of the Analytical Results

3.4.4.1 Discussion of the Analytical Results for Rayleigh Channels

In this sub-section, we study the applicability and the limitations of the analytical ex-
pressions as derived above, for Rayleigh channels. The relation between rms and the
LCRf is depicted in Figure 3-15a for the noiseless and the noise-corrupted cases. The
results from the continuous analysis (eqs. (3-5) and (3-35)) for the noise-corrupted
case with and without the correction factor cs, i.e., for cs = {1, 0.78} and the exact
results from the discrete analysis ((3-16) and (3-18), with (3-37) and (3-26), for u = 0)
are compared. Figure 3-15b shows the relative systematic error of the simpler and
more general results from the continuous model compared with the exact results from
the discrete analysis.
For all curves, the RMS amplitude was taken as the threshold level for calculating
LCRf, i.e., r' = 1. A fixed sampling interval F = 1 Hz and a fixed, relative noise
strength N' = 0.01 were selected.
A thorough analysis of the noiseless case was given in Section 3.2.3.4. The relative
estimation error is also defined there, in eq. (3-29).

The presence of measurement noise causes an increase of the LCRf. The relative in-
crease is largest in the region of low rms, where the level crossings due to noise get
dominant. The error of (3-35) (continuous case) without correction (i.e., for cs = 1) is
seen to raise towards 22 % as rms 0 (curves in Figure 3-15a and b). This
illustrates the sampling effects for the additive noise that were discussed in Section
3.4.2.1. When the correction factor cs = 0.78 is introduced, the curves for the discrete
and the continuous analysis agree well, if rms 1/(20F) (Figure 3-15a, ++ and
{{). For larger rms, the influence of the sampling is leading to deviations, as in
88 Chapter 3 Channel Measurement Technique based on the FD-Level Crossing Rate

0.14
continuous model; no noise

frequencydomain level crossing rate N (r = 1) [s]

discrete model; no noise
0.12 continuous model; with noise
discrete model; with noise
R cont. mod.; corrected; with noise
0.1

0.08

0.06
with noise

0.04

0.02
noiseless

0
0 0.02 0.04 0.06 0.08 0.1
RMS delay spread [s]
rms

(a)

5
systematic relative error [%] at r = 1

10

15

20
no noise
with noise; original relation (cs = 1)
with noise; corrected relation (cs = 0.78)
25
0 0.02 0.04 0.06 0.08 0.1
RMS delay spread rms [s]

(b)
Figure 3-15: Influence of noise and sampling on the relation of rms and the LCRf. Rayleigh fad-
ing; noise power N' = 0.01; sampling interval fs = 1 Hz; u = 0, r' = 1. (a): LCRf vs.
rms. (b): Relative error of the relation obtained from the continuous model vs. rms.

the noiseless case. Thereby, the relative error is very similar to the one for the noise-
less case (compare {{ and + + in Figure 3-15b). Close inspection reveals a
small residual error < 1 % at rms = 0, which is due to the approximation 1 + N' 1,
used in (3-35). This error reduces for smaller N'.
3.4 Analysis of the Influence of Noise 89

3.4.4.2 Influence of Noise on Estimating rms

In this sub-section we investigate the systematic error resulting from the application of
the conventional, linear relationship for estimating rms from noise-afflicted measure-
ments of the LCRf. That is, (3-5) is used to derive rms from a noise-influenced LCRf,
which is either obtained from the equations presented above (see Figure 3-16a), or
from simulated channel transfer functions (see Figure 3-16b).
The relative errors shown are derived equivalently to (3-29), but this time the errors
due to noise were to be evaluated, i.e.

rms LCR f , noise influenced

= 1 = 1 . (3-40)
rms LCR f , noiseless

All results are shown as a function of the noise power N', at r' = 1. Parameters are the
Ricean K-factor, and the sampling interval expressed by the oversampling factor OS
(3-31).
We first discuss the analytical results depicted in Figure 3-16a. As expected, the addi-
tive measurement noise increases the LCRf if it exceeds a certain value. The important
role of the sampling interval F is noticed. Doubling F has the same effect as reducing
the noise power by a factor of four (6 dB), according to the above equations. At
OS = 1, the systematic error stays below ~10 %, if the noise power expressed by N' is
below ~ 18 dB. Due to the definition of OS as a function of F and K (3-31), the
Ricean K-factor seems to have little influence on the results.
For the Rayleigh fading case, the noise-influenced LCRf was calculated from the
equations for the discrete analysis and from the equations for the continuous model.
The difference among them corresponds to the impact of the sampling interval F if it is
selected too large. (That is, when level crossings due to fades are missed, because the
sampling theorem is violated). As seen from Figure 3-16a, and from Figure 3-15, the
continuous analysis fails to describe these effects. It only describes the impact of addi-
tive noise.
An oversampling factor of 0.5 leads to systematic underestimation of rms by ~12 % (in
the absence of noise). Since noise tends to increase the LCRf, those adverse effects
partly cancel, therefore, about 15 dB more noise can be tolerated for OS = 0.5, com-
pared to OS = 2.
The analytical results have been validated by computer simulations, using the simula-
tion scheme introduced in Section 3.3.3.1 9. Complex noise samples were added to the
generated transfer functions (TF) to introduce the measurement noise. Consecutively,
the measurement procedures were applied to the TFs amplitudes. The mean values
and standard deviations of relative estimation errors were derived from ensembles of

9
The computer simulations in this section have been performed by Giovanni Landman. More simu-
lation results can be found in [21].
90 Chapter 3 Channel Measurement Technique based on the FD-Level Crossing Rate

50
K = 0; cont. model corrected
K = 1; cont. model corrected
40 K = 4; cont. model corrected
K = 10; cont. model corrected
K = 0; discrete model
30
estimation error [%]

OS = 2 OS = 1
20

10

0
OS = 0.5

10

20
40 35 30 25 20 15 10
ratio of noise power to scattered power (N) [dB]

(a)

50
K=0
K=1
% error of simulation compared to theory ( )

K=4

40
K = 10

30

20

10

10

20
40 35 30 25 20 15 10
Ratio of noise power to scattered power (N) [dB]

(b)
Figure 3-16: Bias of the standard estimation method as a function of measurement noise power.
Parameters are the Ricean K-factor, and the sampling interval expressed by the
oversampling factor OS. The LCRf is evaluated at r' = 1. (a): Analytical results;
(b): Simulation results; OS = 2

simulated channels.
In Figure 3-16b, simulation results (mean-errors) are depicted for OS = 2. A good
match with the theoretical results is evident. The errors standard deviations are dis-
cussed below.
3.4 Analysis of the Influence of Noise 91

3.4.4.3 Standard Deviations of Estimation Errors

It was seen from the comparison of the analytical results and computer simulation re-
sults that the analytical expressions provide a reliable prediction of the noise-induced
bias of the rms-estimates. In this section, simulation results are used to evaluate the
standard deviations of the estimates. The standard estimation method employing eq.
(3-5) is analyzed here.
The errors standard deviations depend on the observation bandwidth. Larger band-
width implies that more level crossings can be observed, therefore, the performance of
the statistical method for estimating rms improves. Simulation results as a function of
the bandwidth were shown in Section 3.3.3.2. It is seen that the standard deviation is
approx. 10 15 % for an observed bandwidth of 10/rms. This bandwidth was used in
all further simulations. rms was estimated from the LCRf at r' = 1.
In the presence of noise, a deterioration is seen as N' is increased, because the additive
noise reduces the accuracy of the LCRf determined from the TF (see Figure 3-17; lines
marked by +). The simulation shown is for K = 1 and for two-fold oversampling.
The other lines in this figure are discussed in Sections 3.4.4.4 and 3.4.5.2.

50
Standard method
45 Analytic method with correction
Noise reducing method (with 2nd threshold)

40
Standard deviation [%]

35

30

25

20

15

10
40 35 30 25 20 15 10
Ratio of noise power to scattered power (N) [dB]

Figure 3-17: Standard deviations of the rms estimation errors as a function of the power of the
additive noise. rms is estimated from the LCRf at r' = 1, using different estimation
methods. K = 1, OS = 2.

3.4.4.4 Noise Cancellation by Applying the Analytical Results

In Figure 3-18, the performance of the RMS delay spread estimation method is depic-
ted, this time when noise is accounted for by using (3-35) the analytical result allow-
ing to separate the noises component from the LCRf with cs = 0.78. All results are
for an oversampling factor of two (OS = 2) and rms estimated from the LCRf at r' = 1.
92 Chapter 3 Channel Measurement Technique based on the FD-Level Crossing Rate

10

% error of simulation compared to theory ( )

20

30

40

50

60

70

80
K=0
K=1
90 K=4
K = 10
100
40 35 30 25 20 15 10
Ratio of noise power to scattered power (N) [dB]

Figure 3-18: Systematic estimation error of rms using the analytical relationship for correcting
for noise. OS = 2; cs = 0.78; LCRf determined at r' = 1.

Some improvement can be seen for this extended method (compare Figure 3-18 and
Figure 3-16), however, rms is underestimated at high N'. The underestimation is due to
the application of the non-linear (hyperbolic) relation (3-35) for estimating rms from
the noise-corrupted LCRf (see Figure 3-11). Considering that the measured LCRf-val-
ues have a certain standard deviation (depending on the observation bandwidth), the
non-linear translation curve introduces some bias to the estimates. In particular, when
the measured LCRf is lower than the subtracted noise term, then the square root in
(3-35) gets a negative argument. In these cases, the rms value was taken as zero, which
adds to this bias.
Also the estimates variance is increased when using eq. (3-35), because the transfor-
mation functions (see Figure 3-11) get flatter as the influence of noise increases. This
is confirmed by the standard deviation results shown in Figure 3-17b (curve marked by
). While the systematic errors suggest that approx. 5 dB more noise can be tolerated
when noise is corrected for by using (3-35) with cs = 0.78, the errors standard devia-
tions show that the estimates are thereby getting far less accurate.
We conclude that the original measurement procedure utilizing eq. (3-5) can be used in
the area where the systematic error due to noise is sufficiently small. Eq. (3-35) (with
cs = 0.78) can be used to identify this area. Robust measurement methods as the one
described below should be applied if noise cannot be neglected.

3.4.5 A Robust Measurement Procedure

A robust measurement procedure is briefly discussed in this section, which was sug-
gested by Chris van den Bos and studied by the author in cooperation with Adrian
3.4 Analysis of the Influence of Noise 93

55

56

57
transfer function magnitude [dB]
58

59

60

61

62
transfer function with noise
63 main threshold
validation threshold
64 upcrossings
accepted upcrossings
65
450 500 550 600 650 700
frequency [MHz]

Figure 3-19: Illustration of the robust measurement method using a validation threshold.

Bohdanowicz10. The idea of this method is to neglect during the counting process level
crossings that are likely due to noise and not due to the fading.

3.4.5.1 Introduction of an Additional Threshold

The proposed algorithm augments the standard procedure for determining the FD-level
crossing rate by introducing an additional threshold, which is then used to validate the
level crossings through the standard threshold (see Figure 3-19). The modified algo-
rithm works as follows.
Firstly, the crossings through the standard threshold are identified in the measured
data. Then, for each crossing, the data between two crossings (the one, which is inves-
tigated and the next one) is analyzed. If there is a crossing through the additional
threshold (placed above the standard one) within the analyzed interval, the crossing is
accepted. If, however, the data does not exceed the additional threshold within the
interval, the crossing is rejected. As a result, only the crossings likely caused by the
channel variability are counted.
The method was used by A. Bohdanowicz in his measurement campaign as presented
in [12]. Although the promising potential of reducing the influence of noise from the
calculation of the LCR was presented there, no detailed investigation of the methods
performance was given. In the following section, we investigate this method with a
relation to the parameters of the underlying channel model, based on computer simu-
lations. No attempt was made to describe the performance or the required parameters

10
C. v. d. Bos and A. Bohdanowicz are with the Ubiquitous Communications Program, Delft Uni-
versity of Technology, Department of Electrical Engineering.
94 Chapter 3 Channel Measurement Technique based on the FD-Level Crossing Rate

50
K=0
K=1

% error of simulation compared to theory ()

K=4
40 K=10

30

20

10

10

20
40 35 30 25 20 15 10
Ratio of noise power to scattered power (N) [dB]

(a)

2.5
K=0
K=1
K=4
K=10
2
threshold separation [dB]

1.5

0.5

0
40 35 30 25 20 15 10
Ratio of noise power to scattered power (N) [dB]

(b)
Figure 3-20. (a): Systematic estimation error of rms for the two-threshold method; OS = 2;
LCRf determined at r' = 1. (b): Optimal threshold separations for OS = 2.

analytically.

3.4.5.2 Performance of the Method

A large set of channel impulse responses is generated as described in Section 3.3.3.1.
For each noise level N' and Ricean K-factor, the optimal threshold separation is then
calculated by means of minimizing the error of the estimated rms. The simulation re-
3.4 Analysis of the Influence of Noise 95

50
K=0
40 K=1

% error of simulation compared to theory ( )

K=4

K=10
30

20

10

10

20

30

40

50
40 35 30 25 20 15 10
Ratio of noise power to scattered power (N) [dB]

Figure 3-21. Systematic estimation error of rms with a global threshold separation of 1 dB.
OS = 2; LCRf determined at r' = 1.

sults of the proposed method are depicted in Figure 3-20a, whereas the optimal set-
tings of the proposed filtering method (the optimal threshold separations) are shown in
Figure 3-20b (for OS = 2).
By comparing the results from Figure 3-20a with those presented in Figure 3-16b, one
can conclude that, indeed, the method can be used to reduce the influence of noise on
the RDS estimation. The comparison shows that for higher values of N' the estimation
error is drastically reduced, while for low values, effectively no filtering is required
and the method performs as well as the one based on the standard LCR calculation.
The method can be used to increase the accuracy of the rms estimation at the presence
of a significant noise level, but the threshold separation must be set appropriately (see
Figure 3-20b). It is seen from Figure 3-17 that the impact of noise on the standard de-
viation of the estimates has also improved significantly (curve marked by ), com-
pared to the other methods.
The results presented in Figure 3-20b show that the optimal threshold separation is a
function of two parameters, N' and K. Although we do not provide the explicit formula
for this relation, the curves from Figure 3-20b can be followed in practical applications
to set the correct threshold separation for each measurement. A different oversampling
factor (approximately) shifts the curves for OS = 2 to the left or right by 20log(OS/2)
dB.
The importance of the threshold separation is presented in Figure 3-21, where a com-
mon value was used for all the measurements. Better performance at higher values of
N' (when compared to Figure 3-16b) is paid by an increased underestimation of RDS
in the range of low N' values.
96 Chapter 3 Channel Measurement Technique based on the FD-Level Crossing Rate

3.4.6 Extended Measurement Procedure

The measurement procedure introduced in Section 3.3.1 may be extended if measure-
ment noise is an issue that cannot be neglected. Again the procedure starts with the ac-
quisition of calibration data, from which also the noise power is derived. Consecu-
tively, sampled magnitude transfer functions of the channel Rn = |H(nF)|,
n = {1,2,,N} are measured. From the data, the channel parameters are obtained in
the following steps, which include testing for the necessity of noise-suppression:
1. Compensation for the measurement systems characteristics is performed by sub-
tracting the calibration data [dB] from the measured transfer function.

N
2. The average received power P0 = 1
N n =1
Rn2 is estimated.

3. The Ricean K-factor is determined, e.g. by the method given in [1], based on the
average power P0 and the average amplitude R =
N
1
N n =1
Rn . Note that noise has
negligible influence on this and on the previous step, if N' ~0.01 (compare
(3-33)).
4. Based on the noise power, sampling interval, and Ricean K-factor, the increase of
the LCRf due to noise is evaluated, employing eq. (3-35).
5. The RMS delay spread rms is estimated from the LCRf at threshold level r = P0
(i.e., at r' = 1), using (3-5) and (3-30). If according to the previous step, noise can-
not be neglected, then the LCRf is determined by the robust method described in
Section 3.4.5.

3.5 Conclusions and Recommendations

In this chapter, an elaborate study of the frequency-domain level crossing rate (LCRf)
of a frequency-selective radio channel is presented. The LCRf is the average number of
up-going level crossings of the channel transfer function (with respect to a specified
threshold) per unit of bandwidth. It has been derived analytically from second order
statistical properties of the FD-channel model, which was proposed in Chapter 2, Sec-
tion 2.3.
Using these analytical results, the influence of channel parameters has been investi-
gated. It was observed that the RMS delay spread rms is proportional to the absolute
value of the LCRf, while the Ricean K-factor determines the shape of the LCRf-func-
tion vs. the threshold level. That is, rms specifies the number of fades per bandwidth
and K indicates the depth of the fades. By defining the threshold level relative to the
normalized received power P0 of the channel, any dependency of the LCRf on P0 is
eliminated. Since these three parameters {rms, K, P0} have such distinct and different
effects on the LCRf, it was concluded that they comprise a most significant set of pa-
rameters to specify frequency-selective radio channels.
3.6 References 97

It was observed that other parameters describing the channel model have little impact
on the LCRf, as for instance the shape of the delay power spectrum or the actual
structure of the channel impulse response. For Rayleigh fading channels, it has been
shown that any dependency on the channel impulse response disappears, provided that
the sampling interval is sufficiently small.

Because of the proportionality between the LCRf and rms, rather simple swept-fre-
quency power measurements can be used to estimate the three most relevant channel
parameters {rms, K, P0}. Standard procedures allow determining P0 and K, while the
newly discovered relation leads to estimates of rms. The estimation accuracy of this
technique depends on the observation bandwidth. It can be enhanced by combining
multiple measurements performed within a small local area.

Unfortunately, the method to determine rms is quite sensitive to measurement noise.

The influence of noise on the proportionality relation has been analyzed, yielding a
compact expression that allows the separation of the channels and the noises contri-
butions to the level-crossing rate.
The analytical result shows that increasing the frequency-domain sampling-interval by
a particular factor has the same effect as reducing the noise power by the square of this
factor, which is thus a valuable means of noise reduction. Thereby it is important not
to violate against the sampling theorem. Moreover, the analytical results can be used
for partial compensation of the influence of noise. Unfortunately, this technique in-
creases the standard deviation of the estimated rms.
To enhance the robustness against noise, another method is presented, which uses two
thresholds when counting the level-crossing rate. This allows for identification and
removal of level crossings that are likely caused by noise and not by fades. The draw-
back of this technique is that the threshold separation must be adapted according to
noise and channel parameters to get the optimum result. Selecting the threshold appro-
priately, both the mean and standard deviation of the estimation errors remain close to
the noiseless case. A fixed threshold separation leads to sub-optimum results. Finding
analytic expressions for the optimum threshold separation is subject for further work,
as well as the appropriate estimation of the noise power.

3.6 References
[1] F. van der Wijk, A. Kegel, and R. Prasad, Assessment of a pico-cellular system
using propagation measurements at 1.9 GHz for indoor wireless communica-
tions, IEEE Trans. Veh. Technol., vol. 44, no. 1, pp. 155162, Feb. 1995.
98 Chapter 3 Channel Measurement Technique based on the FD-Level Crossing Rate

[2] B. P. Donaldson, M. Fattouche, and R. W. Donaldson, Characterization of in-

building UHF wireless radio communication channels using spectral energy
measurements, IEEE Trans. on Antennas and Prop., vol. 44, no. 1, pp. 8086,
Jan. 1996.
[3] M. Ptzold, U. Killat, F. Laue and Y. Li, On the Statistical Properties of Deter-
ministic Simulation Models for Mobile Fading Channels, IEEE Trans. Veh.
Technol., Vol. 47, No. 1, pp. 254269, Feb. 1998.
[4] S. O. Rice, Mathematical Analysis of Random Noise, Bell Syst. Tech. J., vol.
23, pp. 282332, July 1944; vol. 24, pp. 46156, Jan. 1945.
[5] S. O. Rice, Statistical Properties of a Sine Wave Plus Random Noise, Bell Syst.
Tech. J., vol. 27, pp. 109157, 1948.
[6] K. Witrisal, Y.-H. Kim, and R. Prasad, RMS delay spread estimation technique
using non-coherent channel measurements, IEE Electronics Letters, vol. 34, no.
20, pp. 19181919, Oct. 1998.
[7] W. C. Jakes Jr., Microwave Mobile Communications. New York: Wiley-Inter-
science, 1974.
[8] M. Schwartz, W. R. Bennett, and S. Stein, Communication Systems and Tech-
niques. New York: McGraw-Hill, 1966.
[9] J. G. Proakis, Digital Communications, 3rd Edition. New York: McGraw Hill,
1995.
[10] M. K. Simon and M.-S. Alouini, A Simple Single Integral Representation of the
Bivariate Rayleigh Distribution, IEEE Commun. Letters, vol. 2, no. 5, pp. 128
130, May 1998.
[11] A. Bohdanowicz, G. J. M. Janssen, S. Pietrzyk, Wideband indoor and outdoor
multipath channel measurements at 17 GHz, in Proc. VTC99-fall (IEEE Ve-
hicular Technology Conference), Amsterdam, The Netherlands, Sept. 1999, pp.
19982003.
[12] A. Bohdanowicz, Wideband indoor and outdoor radio channel measurements at
17 GHz, Delft Univ. of Technol., UbiCom-Technical Report/2000/2, Jan. 2000
(http://www.ubicom.tudelft.nl/docs/).
[13] R. El Hattachi, J. M. M. de Nijs, K. Witrisal, and R. Prasad, Characterization
th
and simulation of the 18 GHz radio channel, in Proc. IEEE Benelux 6 Sympo-
sium on Vehicular Technology and Communications, Brussels, Belgium, Oct.
1998.
3.6 References 99

[14] J. Purwaha, A. Mank, D. Matic, K. Witrisal, and R. Prasad, Wide-band channel

th
measurements at 60 GHz in indoor environments, in Proc. IEEE Benelux 6
Symposium on Vehicular Technology and Communications, Brussels, Belgium,
Oct. 1998.
[15] K. Witrisal, Y.-H. Kim, and R. Prasad, A New Method to Measure Parameters
of Frequency-Selective Radio Channels using Power Measurements, IEEE
Trans. on Commun., vol. 49, no. 10, pp. 17881800, Oct 2001.
[16] G. J. M. Janssen, P. A. Stigter, and R. Prasad, Wideband indoor channel meas-
urements and BER analysis of frequency selective multipath channels at 2.4, 4.75
and 11.5 GHz, IEEE Trans. on Commun., vol. 44, no. 10, pp. 12721288, Oct.
1996.
[17] A. A. M. Saleh and R. A. Valenzuela, A statistical model for indoor multipath
propagation, IEEE J. Select. Areas Commun., vol. 5, no. 2, pp. 128137, Feb.
1987.
[18] S. J. Howard and K. Pahlavan, Autoregressive modeling of wide-band indoor
radio propagation, IEEE Trans. Commun., vol. 40, no. 9, pp. 15401552, Sep.
1992.
[19] G. J. M. Janssen, Robust receiver techniques for interference-limited radio chan-
nels. Ph.D. Thesis, Delft University of Technology, Delft, The Netherlands,
1998.
[20] K. Witrisal and A. Bohdanowicz, Influence of Noise on a Novel RMS Delay
th
Spread Estimation Method, in Proc. PIMRC 2000 (11 International Sympo-
sium on Personal Indoor Mobile Radio Communications), London, Sept. 2000,
pp. 560566.
[21] G. Landman, Frequency Domain Study of the Wide-Band Mobile Propagation
Channel, M.Sc. Thesis, Delft University of Technology (IRCTR), Aug. 2001.
Part II: OFDM System Proposal and
Evaluation

101
Chapter 4 OFDM Introduction and
System Modeling

4.1 Introduction
The aim of this chapter is to provide some theoretical background on the OFDM
transmission technique, which is the general topic of the rest of this thesis. A brief in-
troduction to OFDM is given in Section 4.2. We review the block diagram of a clas-
sic OFDM system, which employs a guard interval to mitigate the impairments of the
multipath radio channel. We also discuss several design considerations related to
hardware properties and derive the mathematical model for an idealized system, lead-
ing to the conclusion that data symbols can be transmitted independently of each other
(i.e., without inter-symbol-interference (ISI) and inter-carrier-interference (ICI).)
Moreover, the effects of synchronization imperfections are analyzed, like carrier fre-
quency and phase offsets, and timing errors.

Section 4.3 introduces a method of calculating uncoded BERs for this idealized
OFDM system model. This method is largely based on work presented in [1]. Differ-
ential and coherent detection schemes can be evaluated for Rayleigh and Ricean fading
channels. The results obtained are used in later chapters as a benchmark, in order to
evaluate the loss of implemented algorithms for the OFDM modems. We also show
that, for the system proposal under investigation, differential detection in time-direc-
tion is much preferable to differential detection in frequency direction. Imperfect syn-
chronization and channel estimation may be assessed by extending the system model
used and by incorporating the SNR degradations due to ICI and ISI. Basic aspects are
discussed in this chapter. Issues for a further refinement of the methods are addressed.

The rest of this chapter is organized as follows. The introduction to OFDM and the

103
104 Chapter 4 OFDM Introduction and System Modeling

derivation of the simplified system models are presented in Section 4.2. In Section 4.3,
the performance evaluation of the uncoded OFDM system is outlined, followed by
conclusions and recommendations in Section 4.4.

4.2 OFDM Introduction and System Model

Orthogonal frequency division multiplexing (OFDM) is a parallel transmission
scheme, where a high-rate serial data stream is split up into a set of low-rate sub-
streams, each of which is modulated on a separate sub-carrier (SC) (frequency division
multiplexing). Thereby, the bandwidth of the sub-carriers becomes small compared
with the coherence bandwidth of the channel, i.e., the individual sub-carriers experi-
ence flat fading, which allows for simple equalization. This implies that the symbol
period of the sub-streams is made long compared to the delay spread of the time-dis-
persive radio channel.
Selecting a special set of (orthogonal) carrier frequencies, high spectral efficiency is
obtained, because the spectra of the sub-carriers overlap, while mutual influence
among the sub-carriers can be avoided (see Figure 1-3 in Chapter 1). The derivation of
the system model shows that, by introducing a cyclic prefix (the so-called guard inter-
val (GI)), the orthogonality can be maintained over a dispersive channel (see Section
4.2.3).

This section starts with a brief introduction to the OFDM transmission technique,
based on the description of the systems block diagram. We then discuss some hard-
ware-related design considerations (Section 4.2.2) that become relevant if an OFDM
system is implemented in hardware. For instance the DC-subcarrier and sub-carriers
near the Nyquist-frequency must be avoided. Next, we derive the system model for a
perfectly synchronized system (Section 4.2.3), and we investigate the impact of the
most relevant synchronization errors (Section 4.2.4).
For a more elaborate introduction to OFDM, the reader may refer to the respective
chapters of [2], [3], and to [4][6]. An excellent overview over the effects of many
non-ideal transmission conditions is given in [7], wherein numerous further references
are found.

4.2.1 OFDM Introduction and Block Diagram

Figure 4-1 shows the block diagram of a simplex point-to-point transmission system
using OFDM and forward error correction coding. The three main principles incorpo-
rated are:
The inverse discrete Fourier transform (IDFT) and the discrete Fourier transform
(DFT) are used for, respectively, modulating and demodulating the data constella-
tions on the orthogonal sub-carriers [8]. These signal processing algorithms replace
the banks of I/Q-modulators and -demodulators that would otherwise be required.
4.2 OFDM Introduction and System Model 105

Channel Symbol OFDM I/Q Guard I/Q I/Q I/Q-modula- RF

coding / mapping modulation interval / DAC tion and up-
interleaving (modulation) (IDFT) windowing conversion s (t)
Data RF
source
N complex data transmitted base-
constellations {xi,k} band signal s(t) Multipath
radio
received data const. {yi,k} received signal r(t) channel

Decoding / symbol de- OFDM Guard down-conver- r

RF(t)
de- mapping demod. interval sion and I/Q-
ADC
Data interleaving (detection) (DFT) removal demodulation
sink I/Q I/Q I/Q RF
Channel est. Time sync. Carrier sync.

: digital signals : analog signals

Figure 4-1: Simplex, point-to-point transmission using OFDM.

The analysis of Section 4.2.3 will show this equivalence.

Note that at the input of the IDFT, N data constellation points {xi,k} are present,
where N is the number of DFT points. (i is an index on the sub-carrier; k is an
index on the OFDM symbol). These constellations can be taken according to any
phase-shift-keying (PSK) or quadrature-amplitude-modulation (QAM) signaling
set (symbol mapping). The N output samples of the IDFT being in time-domain
form the base-band signal carrying the data symbols on a set of N orthogonal sub-
carriers. In a real system, however, not all of these N possible sub-carriers can be
used for data, as elaborated in Section 4.2.2.3.
Usually, N is taken as an integer power of two, enabling the application of the
highly efficient (inverse) fast Fourier transform (IFFT; FFT) algorithms for modu-
lation and demodulation.
The second key principle is the introduction of a cyclic prefix as a guard interval
(GI), whose length should exceed the maximum excess delay of the multipath
propagation channel [9]. Due to the cyclic prefix, the transmitted signal becomes
periodic, and the effect of the time-dispersive multipath channel becomes
equivalent to a cyclic convolution, discarding the guard interval at the receiver.
Due to the properties of the cyclic convolution, the effect of the multipath channel
is limited to a point-wise multiplication of the transmitted data constellations by the
channel transfer function, the Fourier transform of the channel impulse response,
i.e., the sub-carriers remain orthogonal (see [4][7]). This conclusion will also fol-
low from the derivation of the system model in Section 4.2.3. The only drawback
of this principle is a slight loss of effective transmit power, as the redundant GI
must be transmitted. Usually, the GI is selected to have a length of one tenth to a
quarter of the symbol period, leading to an SNR loss of 0.51 dB. (See also Figure
4-2).
The equalization (symbol de-mapping) required for detecting the data constella-
106 Chapter 4 OFDM Introduction and System Modeling

tions is an element-wise multiplication of the DFT-output by the inverse of the es-

timated channel transfer function (channel estimation). For phase modulation
schemes, multiplication by the complex conjugate of the channel estimate can do
the equalization. Differential detection can be applied as well, where the symbol
constellations of adjacent sub-carriers or subsequent OFDM symbols are compared
to recover the data.
Forward error correction (FEC) coding and (frequency-domain) interleaving are
the third crucial idea applied. The frequency-selective radio channel may severely
attenuate the data symbols transmitted on one or several sub-carriers, leading to
bit-errors. Spreading the coded bits over the band-with of the transmitted system,
an efficient coding scheme can correct for the erroneous bits and thereby exploit
the wide-band channels frequency-diversity. OFDM systems utilizing error cor-
rection coding are often referred as coded OFDM (COFDM) systems. In Chapter 8,
the performance of coded OFDM systems is evaluated. The bit-error-rate (BER) of
the uncoded system is analyzed in Section 4.3.

The complex equivalent base-band signals generated by digital signal processing are
in-phase/quadrature (I/Q)-modulated and up-converted to be transmitted via an RF-
carrier. The reverse steps are performed by the receiver.

Synchronization is a key issue in the design of a robust OFDM receiver. Time- and
frequency-synchronization are paramount to respectively identify the start of the
OFDM symbol and to align the modulators and the demodulators local oscillator fre-
quencies. If any of these synchronization tasks is not performed with sufficient accu-
racy, then the orthogonality of the sub-carriers is (partly) lost. That is, inter-symbol-
interference (ISI) and inter-carrier-interference (ICI) are introduced. The effect of
small synchronization errors is analyzed in Section 4.2.4. Synchronization algorithms
are discussed in Chapter 6.

4.2.2 Design of the OFDM Signal

The proposal of a realistic OFDM-based communications system was one of the goals
of this research project. Therefore, we elaborate here on some hardware related design
considerations, which are often neglected in theoretical studies. Elements of the trans-
mission chain that have impact on the design of the transmitted OFDM signal are:
The time-dispersive nature of the mobile channel, which the transmission scheme
must be able to cope with.
The bandwidth limitation of the channel. The signal should occupy as little band-
width as possible and introduce a minimum amount of interference to systems on
adjacent channels.
The transfer function of the transmitter/receiver hardware. This transfer function
reduces the useable bandwidth compared to the theoretical one given by the sam-
4.2 OFDM Introduction and System Model 107

Transmitter pulse shape w(t)

T
Twin Tguard TFFT

GI
Prefix effective TX-time Postfix time
kT
Channel impulse response

excess delay time

max

Receiver filter (implemented by FFT)

TFFT time

Figure 4-2: Cyclic extension and windowing of the OFDM symbol.

pling theorem. I.e., some oversampling is required.

Phase-jitter and frequency offsets of the up- and down-converters, and Doppler
spreading of the channel.

4.2.2.1 Guard Interval

As mentioned above, a guard interval (GI) is introduced to preserve the orthogonality
of the sub-carriers and the independence of subsequent OFDM symbols, when the
OFDM signal is transmitted over a multipath radio channel. The guard interval, a cy-
clic prefix, is a copy of the last part of the OFDM symbol, which is transmitted before
the so-called effective part of the symbol (cf. Figure 4-2). Its duration Tguard is sim-
ply selected larger than the maximum excess delay of the (worst-case) radio channel.
Therefore, the effective part of the received signal can be seen as the cyclic convolu-
tion of the transmitted OFDM symbol by the channel impulse response.

4.2.2.2 Windowing
A rectangular pulse has a very large bandwidth due to the side-lobes of its Fourier
transform being a sinc-function. Windowing is a well-known technique to reduce the
level of these side-lobes and thereby reduce the signal power transmitted out of band.
In an OFDM system, the applied window must not influence the signal during its ef-
fective period. Therefore, cyclically extended parts of the symbol are pulse-shaped as
depicted in Figure 4-2 [3].
Note that this additional cyclic prefix extends the GI to some extent. I.e., the delay-
spread robustness is slightly enhanced. On the other hand, the efficiency is further re-
duced, as the window part is also discarded by the receiver. The orthogonality of the
108 Chapter 4 OFDM Introduction and System Modeling

timedomain frequencydomain
1.5 1.5

a (TD)

a (FD)
1 1
0.5 0.5
0 0
20 0 20 40 60 80 0 5 10
1.5 1.5
b (TD)

b (FD)
1 1
0.5 0.5
0 0
20 0 20 40 60 80 0 5 10
0.2 1.5
c (TD)

c (FD)
1
0.1
0.5
0 0
20 0 20 40 60 80 0 5 10
1.5 1.5
d (TD)

d (FD)
1 1
0.5 0.5
0 0
20 0 20 40 60 80 0 5 10
samples frequecy in subcarriers

(a) (d)
OFDM spectra for NFFT = 64, Nwin = 16, Nguard = 16
10
rectangular pulse
0 window function
windowed pulse

10

20
signal strength [dB]

30

40

50

60

70

80
0 5 10 15 20 25
frequency in subcarriers

(e)
Figure 4-3: (a): Shape and spectrum of the OFDM receive filter (realized by FFT); (b): rectan-
gular pulse of duration T and its spectrum; (c): sine-half-wave used for pulse-
shaping and its spectrum; (d): transmitter pulse prototype w(t) and its spectrum.
(e): Spectra of (b)(d) in logarithmic scale.

sub-carriers of the OFDM signal is restored by the rectangular receiver filter imple-
mented by the DFT (Figure 4-2), requiring the correct estimation of the DFT start time
kT, where T is the OFDM symbol period.
4.2 OFDM Introduction and System Model 109

The symbol periods in Figure 4-2 are given as times. Since the implementation is usu-
ally done on digital hardware, those periods are also often defined in terms of samples.
N, Nguard, and Nwin then define the number of samples in the effective part, guard-, and
windowing-interval, respectively. The effective part is also referred to as the FFT-
part, because this part of the OFDM symbol is applied to the FFT to recover the data
at the receiver.
Spectrum of the transmitter pulse shape
Windowing of the transmitter pulse using a raised-cosine function can be seen as a
convolution of the extended rectangular pulse of duration T with a sine-half-wave, as
shown in Figure 4-3. In the frequency-domain, this convolution means a multiplication
of the sinc-spectrum of the rectangular pulse with the spectrum of the sine-half-wave.
It is seen that this multiplication reduces the side-lobes of the transmitter pulse shape.
In Figure 4-3 (a), the zeros of the spectrum occur at positions iF = i/TFFT, i = {1, 2,
}, i.e., at those positions, where the adjacent sub-carriers are located. The extension
of the rectangular pulse to length T = TFFT + Tguard + Twin reduces the distance between
zeros to 1/T (Figure 4-3 (b)). The windowing function (Figure 4-3 (c)) has zeros at
positions 1/Twin{3/2, 5/2, 7/2, }.

4.2.2.3 System Transfer Function (ADCs, DACs, IF-Filters, RF Front-end, etc.)

Because of the low-pass filters required for the analog-to-digital and digital-to-analog
conversion (ADC and DAC) of the transmitted and received (baseband) signals, not all
N sub-carriers can be used, if an N-point IFFT is applied for modulation. The sub-car-
riers close to the Nyquist frequency fs/2 will be attenuated by these filters and thus
cannot be used for data transmission (see Figure 4-4). (fs = 1/Ts is the sampling fre-
quency.) Also the DC-sub-carrier might be heavily distorted by DC offsets of the
ADCs and DACs, by carrier feed-through, etc., and should thus be avoided for data.

Transfer function of
transmitter/receiver

frequency
fs/2 DC fs/2
useable sub-carriers useable sub-carriers
sub-carrier
N/2, , 1, 0, 1, , N/21
index i

Figure 4-4: Transfer function of the transmitter/receiver hardware and its impact on the de-
sign of an OFDM system.

4.2.3 OFDM System Model

The above-introduced features of the OFDM signal are defined mathematically in this
section. This will lead to the conclusion that, using the OFDM principle, data symbols
110 Chapter 4 OFDM Introduction and System Modeling

can be transmitted over multipath radio channels without influencing each other.

4.2.3.1 Signal Model and Definitions

Mathematically, the OFDM signal is expressed as a sum of the prototype pulses
shifted in the time- and frequency directions and multiplied by the data symbols. In
continuous-time notation, the k-th OFDM symbol is written

N / 21

j 2 f c + (t kT )
i

Re w(t kT ) xi ,k e TFFT


i = N / 2

s RF ,k (t kT ) = kT Twin Tguard t kT + TFFT + Twin . (4-1)


0 otherwise

Most of the mathematical symbols have been defined in the previous figures already.
A complete list of symbols is given below:
T Symbol length; time between two consecutive OFDM symbols
TFFT FFT-time; effective part of the OFDM symbol
Tguard Guard-interval; duration of the cyclic prefix
Twin Window-interval; duration of windowed prefix/postfix for spectral
shaping
fc Center frequency of the occupied frequency spectrum
F = 1/TFFT frequency spacing between adjacent sub-carriers
N FFT-length; number of FFT points
k index on transmitted symbol
i index on sub-carrier; i {N/2, N/2+1, , 1, 0, 1, , N/21}
xi,k signal constellation point; complex {data, pilot, null} symbol
modulated on the i-th subcarrier of the k-th OFDM symbol

w(t) denotes the transmitter pulse shape defined as

12 [1 cos (t + Twin + Tguard ) / Twin ] Twin Tguard t < Tguard

w(t ) = 1 Tguard t TFFT (4-2)
1 [1 + cos (t T ) / T ] TFFT < t TFFT + Twin
2 FFT win

Finally, a continuous sequence of transmitted OFDM symbols is expressed as

s RF (t ) = s
k =
RF ,k (t kT ) (4-3)

The simulated spectrum of such an OFDM signal is depicted in Figure 4-5 for different
4.2 OFDM Introduction and System Model 111

OFDM spectrum for N = 64, N = 16, oversampling = 2

FFT guard
10
Nwin = 2
0 Nwin = 0
N = 16
power spectrum magnitude [dB]
10 win

20

30

40

50

60

70

80

90
60 40 20 0 20 40 60
frequency in subcarriers

Figure 4-5: Spectrum of an OFDM signal with 64 sub-carriers and different window lengths.
Two-fold oversampling has been applied in the time-domain; 48 sub-carriers are
used for data.

window lengths.

4.2.3.2 Lowpass Equivalent Transmitted Signal

From eqs. (4-1)(4-3), the complex equivalent lowpass signal transmitted can be di-
rectly given. The complex envelope of the OFDM signal is written

s (t ) = s
k =
k (t kT ) , (4-4)

with

N / 21
i
j 2 (t kT )
w(t kT ) xi ,k e TFFT
i = N / 2

s k t kT
( ) = kT Twin Tguard t kT + TFFT + Twin (4-5)


0 otherwise

Note the similarities of this expression to the equation of a Fourier series

v (t ) = c(nf
n =
0 )e j 2nf0t , (4-6)

where the complex-valued Fourier coefficients c(nf0) represent the complex-valued

signal constellation points xi,k, and the frequencies nf0 correspond to the sub-carrier
112 Chapter 4 OFDM Introduction and System Modeling

frequencies i/TFFT.
In a digital system, this modulated waveform can be generated by an inverse discrete
Fourier transform (IDFT) or by its computationally efficient implementation, the IFFT.
The data constellations xi,k are the input to this IFFT; the time-domain OFDM symbol
is its output.

4.2.3.3 Time-Dispersive Channel

The influence of the time-variant, multipath fading radio channel is expressed by its
(lowpass equivalent) impulse response h(,t) plus AWGN n(t):
max

r (t ) = h( , t ) s (t ) + n(t ) = h( , t ) s(t )d + n(t )

0
(4-7)

The range of integration in this convolutional integral (* denotes convolution) has

been limited to [0, max], because the channel impulse response is zero elsewhere. Ex-
cess delay = 0 of the channel is defined as the delay time at which the first wave ar-
rives at the receiver. Thus, transmit and receive time instants are mathematically de-
fined equal (compare Figure 4-2). max is the maximum excess delay of the channel.
Two assumptions are made to simplify the derivation of the received signal. The chan-
nel is considered quasi-static during the transmission of the k-th OFDM symbol, thus
h(,t) simplifies to hk(). Furthermore, we define the maximum excess delay
max < Tguard. Therefore, there is no interference of one OFDM symbol on the effective
period of the consecutive one (cf. Figure 4-2). I.e., inter-symbol-interference (ISI) is
suppressed in case of sufficiently accurate time synchronization.

4.2.3.4 OFDM Demodulation

The demodulation of the OFDM signal should be performed by a bank of filters,
which are matched to the effective part [kT, kT + TFFT] of the OFDM symbol (see
Figure 4-2). The reverse operation to eq. (4-6), i.e., the extraction of the Fourier coeffi-
cients c(nf0) (= xi,k) from the time-domain signal v(t) (= r(t)), exactly formulates such a
bank of matched filters. It is written
1
v (t )e
j 2nf 0t
c( nf 0 ) = dt , (4-8)
T0 T0

where T0 is the integration period being equivalent to TFFT. In a digital implementation,

a DFT or (preferably) a FFT is used to realize these filters.
Assuming knowledge of the exact time-instants kT at which the OFDM symbols start,
we try to extract the transmitted signal constellations xi,k from the received signal r(t).
The received signal constellations are denoted yi,k.
4.2 OFDM Introduction and System Model 113

kT +TFFT
1
r (t )e
j 2i ( t kT ) / TFFT
yi , k = dt =
TFFT t = kT
(4-9)
1
kT +TFFT
max j 2i (t kT ) / TFFT
hk ( ) s (t )d + n(t ) e dt
TFFT t = kT =0

Because of the integration ranges in eq. (4-9) and max < Tguard, there is no influence of
the adjacent OFDM symbols transmitted, and s(t) can be replaced by sk(t), eq. (4-5).
kT +TFFT max N / 2 1
i'
j 2 (t kT )
yi ,k =
1
hk ( ) xi ',k e TFFT d e j 2i (t kT ) / TFFT dt +
TFFT =0 i '= N / 2
t = kT
(4-10)
kT +TFFT
1
n( t ) e
j 2i ( t kT ) / TFFT
dt
TFFT t = kT

Note that w(t kT) = 1 in the range of integration. The window is thus omitted in this
equation. The second integral in eq. (4-10) leads to independent additive noise samples
ni,k since the complex exponential terms represent orthogonal functions. Substituting
u = t kT, for the ease of notation, and changing the order of integration and summa-
tion yields
N / 2 1
1
TFFT max
j 2iu / TFFT
yi , k = xi ',k =0
hk ( ) e j 2i '( u ) / TFFT
d e du + ni ,k =
i '= N / 2 TFFT u =0
(4-11)
N / 21
1
TFFT max

xi ',k
TFFT u =0 =0hk ( )e
j 2i ' / TFFT
d e j 2 (i i ')u / TFFT du + ni ,k

i '= N / 2
The inner integral of the second expression represents the Fourier transform of hk() at
the frequency instants i'/TFFT = i'F, which is the sampled channel transfer function at
time kT. It is expressed by the channel coefficients
max
hi ',k = FT {hk ( )} = h ( )e
k
j 2i ' / TFFT
d = H (i ' F , kT ) . (4-12)
=0

Using this notation, the output of the receiver filter bank simplifies to
N / 21 TFFT
1
yi ,k =
i '= N / 2
xi ', k hi ',k
TFFT e
j 2 ( i i ') u / TFFT
du + ni ,k (4-13)
u =0

The integral in this equation has the value 1, only if i = i'. For i i', i and i' being inte-
ger values, the integral is zero. Thus we finally obtain
yi ,k = xi ,k hi ,k + ni ,k . (4-14)

From this form it is seen that a perfectly synchronized OFDM system can be viewed as
a set of parallel Gaussian channels as depicted in Figure 4-6 [4][6]. The multipath
114 Chapter 4 OFDM Introduction and System Modeling

hi,k ni,k
xi,k yi,k

hi+1,k ni+1,k
xi+1,k yi+1,k

Figure 4-6: Idealized OFDM system model. The sub-channels of the OFDM system can be con-
sidered as parallel Gaussian channels under the assumptions of perfect time- and
carrier synchronization and perfect suppression of multipath by the guard inter-
val.

channel introduces an attenuation/amplification and phase rotation according to the

(complex-valued) channel coefficients {hi,k}.
Channel estimation is required in order to retrieve the data contained in these signal
constellations, because the receiver must have a phase (and amplitude) reference to
correctly detect the transmitted symbol. Differential detection can be used alterna-
tively, in which case the decision is made by comparing the phases (and amplitudes)
of symbols transmitted over adjacent sub-carriers or subsequent OFDM symbols.

Due to the attenuation/amplification, each sub-carrier typically has an individual sig-

nal-to-noise ratio (SNR). The SNR per sub-carrier (after the DFT) is defined as
2
( Ec / N 0 ) i ,k = E{| xi ,k |2 } hi ,k N2 , (4-15)

where N2 = E{| ni ,k |2 } is the noise variance. With the normalized received power being
written P0 = E{| hi ,k | 2 } , the average SNR becomes Ec / N 0 = E{| xi ,k |2 } P0 N2 . Usually,
the signal energy is normalized to unity, i.e., E{| xi ,k |2 } = 1 .

4.2.4 Synchronization Errors

As an introduction to the work on synchronization algorithms, the relevant effects of
synchronization errors are reviewed in this section. Original work on this topic is
found in numerous publications (see e.g. [10], [11]). A comprehensive overview is
given in [7].

4.2.4.1 FFT Time Synchronization Error

The impact of an FFT-timing offset at the receiver can be analyzed mathematically by
shifting the integration interval of the matched filter bank, eq. (4-9). For a timing error
of t, the ideal interval t [kT, kT + TFFT] becomes t [kT + t, kT + TFFT + t] and
(4-9) is written
4.2 OFDM Introduction and System Model 115

kT +TFFT +t
1
r (t )e
j 2i ( t kT t ) / TFFT
yi ,k = dt (4-16)
TFFT t = kT +t

t is assumed to be sufficiently small (typically t < Tguard) that no ISI arises due to the
timing error. In other words, the error is small enough for the channel impulse re-
sponse to remain within the guard interval. Therefore, the receiver window still does
not overlap with the preceding or consecutive OFDM symbol, i.e., no energy is col-
lected from these adjacent OFDM symbols, and the demodulated signal can be ex-
pressed from the transmitted symbol sk(t) again (compare eq. (4-10)). Following the
same steps as in Section 4.2.3 (eqs. (4-9)(4-14)), we obtain for the second part of eq.
(4-11) (with u = t kT t),
N / 21
1
TFFT max
j 2 [( i i ')u +i 't ] / T
yi ,k = xi ', k u =0 =0
h ( ) e
j 2i ' / TFFT
d e FFT
du + ni ,k (4-17)
i '= N / 2 TFFT
Moving the term e j 2i 't / T out of the integral yields the expression for the demodu-
FFT

lated signal constellations in case of a timing error,

yi , k = xi ,k hi ,k e j 2it / TFFT + ni ,k = xi ,k hi ,k e j 2it '/ N + ni ,k , (4-18)

where t' is the timing offset in samples. It is evident that a timing offset gives rise to a
progressive phase rotation of the signal constellations. The phase rotation is zero at the
center frequency and it linearly increases towards the edges of the frequency band. It is
easily verified from eq. (4-18) that a timing-offset of one sample introduces a phase
shift of to the outermost sub-carriers (having i N/2), regardless of the FFT-
length. In Figure 4-7, this effect is visualized for a 64-carrier OFDM system with zero
carriers at fc and at the edges of the frequency band.
If coherent detection is utilized, the induced progressive phase rotation is detected im-
plicitly by the channel estimation algorithm. The subsequent equalization (sub-carrier-
wise multiplication of the received symbols by the inverse of the estimated channel
coefficients) will thus automatically correct for small timing-offsets. No performance
degradation is thereby caused. However, if the timing offset is too large, ISI and ICI
are introduced because energy is also collected from one of the adjacent OFDM sym-
bols, leading to a partial loss of orthogonality [7].
Differential detection is also robust to small timing-offsets. If the differential detection
is applied in the frequency-direction, the progressive phase rotation may reduce the
distance between the compared constellation points, however, which can lead to a per-
formance degradation. Such performance results are given in Section 4.3.3.

A (small) sampling frequency offset leads to a (slowly) increasing timing offset, and
therefore to a progressive phase rotation at an increasing slope. Larger errors yield ICI,
because the SC-spacing at the receiver can no longer be assumed equal to the SC-
spacing at the transmitter. (The SC-spacing is defined as F = 1/(NTs), where Ts is the
116 Chapter 4 OFDM Introduction and System Modeling

effect of timeoffset t = 0.25 samples

0.2

0.1
Qchannel

0.1

0.2

0.2
0.1 20
0 0
0.1
0.2 20
Ichannel subcarrier index

Figure 4-7: Visualization of the influence of an FFT timing offset on the demodulated signal
constellations. A linearly increasing phase rotation is observed with increased fre-
quency distance to the center frequency. + indicate QPSK constellations without
the influence of a timing-offset; depict the rotated data symbols.

sampling period.)

4.2.4.2 Carrier Synchronization Error

Frequency offsets are typically introduced by a (small) frequency mismatch in the lo-
cal oscillators of the transmitter and the receiver. Doppler shifts can be neglected in
indoor environments.
The impact of a frequency error can be seen as an error in the frequency instants,
where the received signal is sampled during demodulation by the FFT. Figure 4-8 de-
picts this two-fold effect. The amplitude of the desired sub-carrier is reduced (+) and
inter-carrier-interference ICI arises from the adjacent sub-carriers ({).
Mathematically, a carrier offset can be accounted for by a frequency shift f and a
phase offset in the lowpass equivalent received signal

r ' (t ) = r (t )e j ( 2ft + ) . (4-19)

With eq. (4-9) we obtain
kT +TFFT
1
r (t )e
j ( 2ft + ) j 2i ( t kT ) / TFFT
yi , k = e dt =
TFFT t = kT
. (4-20)
1
kT +TFFT
max j 2ft j 2i (t kT ) / T
h( ) s (t )d + n(t ) e
j 2
e e FFT
dt
TFFT t = kT =0
4.2 OFDM Introduction and System Model 117

amplitude

frequency

f frequency offset

Figure 4-8: Inter-carrier-interference (ICI) arises in case of a carrier synchronization error.

The figure illustrates the spectra of three individual sub-carriers. These spectra
are superimposed in the OFDM signal spectrum.

Repeating the derivation leading to eq. (4-13), the received constellation points be-
come
N / 2 1 TFFT i i '
j 2 ( f ) u
1
yi , k = e j ( + 2fkT )

i '= N / 2
xi ',k hi ',k
TFFT e TFFT
du + ni ,k . (4-21)
u =0

Due to the frequency error, the integral is not equal zero for i i', neither it is one for
i = i', as in the idealized case above. I.e., the orthogonality between sub-carriers has
been partly lost. The evaluation of this expression yields two terms. The first term (for
i = i') accounts for equal phase rotation and attenuation of all sub-carriers, the second
one (for i i') describes the ICI.
TFFT
1
= e j ( +2fkT ) xi ,k hi ,k e
j 2fu
yi ,k du +
TFFT u =0

N / 2 1 TFFT j 2 (
i i '
f ) u
(4-22)

1

j ( + 2fkT )
e xi ',k hi ',k e TFFT
du + ni , k
i '= N / 2 TFFT u =0
i ' i

These expressions are valid for a frequency-offset f < 0.5 SC. For larger offsets, the
transmitted data symbols xi,k would get shifted by one or more positions in the fre-
quency-direction. I.e., the data symbol of the i-th transmitted SC would appear at the
(i + fi)-th SC at the receiver, where fi = round(f/F) is the integer part of the fre-
quency-error in sub-carriers.
The ICI term can be seen as an additional noise term and can thus be represented as a
degradation of SNR. The amount of degradation has been evaluated by Pollet et al.
[10] for AWGN channels and by Moose [11] for dispersive fading channels (see also
[7]). Frequency-offsets up to 2 % of the sub-carrier spacing F are negligible, according
to their results. Even 510 % can be tolerated in many situations.
118 Chapter 4 OFDM Introduction and System Modeling

distortion of signal constellations due to frequency offset f = F/16

correct states
0.3 erroneous states
cplx. spectrum

0.2

0.1
Qchannel

0.1

0.2

0.3

0.4 0.3 0.2 0.1 0 0.1 0.2 0.3 0.4

Ichannel

Figure 4-9: Phase rotation due to carrier offset of 1/16 of the sub-carrier spacing. The received
signal constellations distorted by ICI are shown.

Evaluation of the phase rotation and attenuation due to a frequency error yields
yi ,k = xi ,k hi , k sinc (f TFFT )exp{j[ + 2f (kT + TFFT / 2)]}+ n'i ,k , (4-23)

using

[ ] sin fTFFT
TFFT
1 1
e
j 2ft
dt = e j 2fTFFT 1 = e jfTFFT = e jfTFFT sinc fTFFT . (4-24)
TFFT t =0 j 2fTFFT fTFFT

The noise term n'i,k includes the additional noise due to ICI.
Figure 4-9 depicts the rotation and distortion of the received signal constellation points
for a carrier offset of f = F/16, = 0, and for QPSK modulation ({). The scattering
of the resulting complex valued signal constellations is caused by ICI. The figure also
shows the projection of the continuous Fourier spectrum of one OFDM symbol on the
complex plane, i.e., the spectrum in-between the sub-carrier frequencies. This line re-
sults from the superposition of the continuous sinc-spectra of individual sub-carriers of
one OFDM symbol. If a frequency-offset is present, the DFT samples this spectrum at
the wrong frequency-instants leading to ICI , which is indicated in the figure by
{. Without frequency-offset, the QPSK constellations are recovered perfectly, as
seen from the points marked by +.

4.2.4.3 Common Carrier and Timing Offset

Evaluating the above expressions for simultaneous timing (t), frequency (f,
fi = round(f/F)) and phase () offsets, the system model for the generalized case is
4.3 Performance of an Uncoded OFDM System 119

obtained. It is written as

yi +f i ,k = xi ,k hi ,k sinc [(f f i F )TFFT ]e

j i , k
+ n'i ,k , (4-25)

where the phase distortion due to the synchronization errors is expressed by

T i
i ,k = + 2f kT + FFT + t + 2t . (4-26)
2 TFFT

Note again that the noise variable n'i,k in (4-25) includes the noise caused by ICI and/or
ISI.
Often, the timing offset is expressed in samples, i.e., t' = t/Ts, and the frequency-off-
set is normalized to the sub-carrier spacing f ' = f F . Using these symbols, the phase
distortions are expressed by

1 N + N guard + N win t ' i

i ,k = + 2f ' + k + + 2t ' . (4-27)
2 N N N

4.3 Performance of an Uncoded OFDM System

In this section, analytical expressions are derived for the bit-error probabilities of un-
coded OFDM systems over Rayleigh and Ricean fading channels. The analysis is
based on the work by Proakis (Appendix B of [1]). The obtained expressions are very
general, allowing the evaluation of various modulation schemes, demodulation and
channel estimation techniques. The application of the formulas is shown for BPSK,
QPSK, 8-PSK, and 16-QAM modulation, with coherent detection (perfect channel es-
timation) and differential detection.
The idealized system model derived in Section 4.2.3 is employed in this study. By
incorporating in the system model the SNR degradation due to synchronization errors,
Doppler spread, or phase noise (which cause ICI and/or ISI), or by incorporating the
mean-square-error of channel estimation techniques, the effect of these impairments
on the BER can be analyzed. The systematic phase rotations induced by synchroniza-
tion errors must be considered as well. While we leave the evaluation of the SNR de-
gradations for future work, we briefly investigate the effects of phase rotations in pres-
ence of (small) synchronization offsets.
The results presented in this section will serve as benchmarks in the performance
evaluation of various signal-processing aspects of the OFDM air-interfaces, which are
treated in Chapter 6 and 7. In OFDM, differential detection can be employed in the
time- and frequency-directions. From the BER of differential detection it is evident
that the time-direction is preferable for the OFDM system parameters under investiga-
tion, since the channel variations versus frequency are larger.
Section 4.3.1 reviews the OFDM system model and channel model. The derivation of
120 Chapter 4 OFDM Introduction and System Modeling

the average BER is explained in Section 4.3.2. Performance results are given in Sec-
tion 4.3.3.

4.3.1 Mathematical Modeling

The OFDM system models derived in Section 4.2 are used in this analysis. For the
sake of simplicity we slightly change the indexing, however, as we only investigate the
time- or frequency direction at a time. The system model of eq. (4-14) becomes
y k = xk hk + nk , (4-28)
where {xk} and {yk} are the transmitted and received signal constellation points
(modulated data symbols), respectively, the {hk} account for the correlated, complex-
valued attenuation factors introduced by the time- and/or frequency-selective radio
channel, and the {nk} denote samples of an AWGN process with E{|nk|2} = N2 . The
index k can be used as a time- or as a frequency-index, depending on the system aspect
under investigation. The attenuation factors thereby constitute the time- or frequency-
transfer function of the channel, respectively:

H ( f , kT ) at given f
hk = , (4-29)
H (kF , t ) at given t
where T is the duration of an OFDM symbol including the guard and windowing inter-
vals, and F denotes the frequency spacing between adjacent OFDM sub-carriers.
The channel model is introduced in the analysis by considering respectively the
spaced-time and spaced-frequency correlation functions of the (wide-sense stationary
uncorrelated scattering WSSUS; see Chapter 2 and [1], [12], [13]) channel.
For describing the frequency-variability, the frequency-domain channel model is used
(see Section 2.3). Thereby, we confine ourselves to the case of the exponentially de-
caying delay power spectrum, where a direct relation can be given between the chan-
nel parameters {P0 average power, K Ricean factor, and rms RMS delay spread}
and the channel correlation function

P0 1
H (f ) = E{H * ( f ) H ( f + f )} = K + . (4-30)
K +1 1 + j 2f rms K1

In this equation, K1 = ( K + 1) 2 K + 1 , f is the frequency-lag, and * denotes the com-

plex conjugate. The normalized received power (average power) is defined as P0 =
E{|hk|2}.
To model the time-variability, the so-called Jakes Doppler spectrum can be used [14],
j ( 2f t + )
augmented by a line-of-sight (LOS) component e
at a given Doppler fre-
quency f. Such a Doppler spectrum corresponds to a spaced-time correlation function
4.3 Performance of an Uncoded OFDM System 121

H (t ) = E{H * (t ) H (t + t )} =
P0
K +1
(
j 2f t
K e + J 0 (2f m t ) , ) (4-31)

where J0() denotes the zero-th order Bessel function of the first kind, t is the time-
lag, and fm is the maximum Doppler frequency. ( f m = vm = vm f c c , where vm is the
mobiles velocity, is the wavelength, fc is the carrier frequency, and c is the speed-of-
light.)

4.3.2 Analytical Evaluation of the BER

Analytical expressions for the BER are derived in this section. Following [1], we start
our analysis with defining the symbol transmitted as xk,i, which is an element of the
symbol set {xk,m}, m = {1, 2, M}. (M is the order of the modulation scheme.) At the
receivers site, an optimum detector will choose the symbol xk,n {xk,m}, which mini-
mizes the distance metric

M d ( x k , n ) =| y k hk x k , n | 2 . (4-32)

This symbol is assumed to be most likely the transmitted symbol. The term hk xk ,n in
this equation accounts for the channel estimation. An error occurs when the metric cal-
culated for a symbol xk,n xk,i is smaller than the metric for the transmitted symbol xk,i.
The probability for this event is written as
Pe = Pr{M d ( xk ,n ) < M d ( xk ,i )} = Pr{D < 0} , (4-33)

where D = Md(xk,n) Md(xk,i) is called the decision variable. Using (4-32), D becomes

D = y k hk* ( xk*,i xk*,n ) + y k* hk ( xk ,i xk ,n )+ | hk |2 (| xk ,n |2 | xk ,i |2 ) . (4-34)

From the channel and system models, yk is known to be a complex Gaussian random
variable. The same holds for hk , which is an estimate of the transfer function H(f,t).
Thus, the decision variable D is a special case of the generic quadratic form (see Ap-
pendix B in [1])

( )
L
D = A | X l |2 + B | Yl |2 +CX l Yl* + C * X l*Yl , (4-35)
l =1

where Xl and Yl are complex-valued Gaussian random variables, and A, B, and C are
constants. In our case L = 1, considering one transmitted symbol over one (sub-) chan-
nel11.
The error probability is the probability that D < 0, which is evaluated in Appendix B of
[1]. This probability is denoted as the integral over the pdf of D

11
The equation for L > 1 allows for the evaluation of diversity schemes [1], [15][17].
122 Chapter 4 OFDM Introduction and System Modeling

0
Pe = Pr{D < 0} = p( D)dD .

(4-36)

For L = 1, the solution to this integral is written [1]

2 / 1 1 ( a 2 + b2 )
Pe = Q1 (a , b) I 0 ( ab)e 2 , (4-37)
1 + 2 / 1

where In(x) is the n-th order modified Bessel function of the first kind and Q1(a,b) is
the Marcums Q function, which can be expressed in terms of Bessel functions as

(a / b) I
1 ( a 2 +b 2 )
Q1 (a , b) = e
n
2
n (ab) , b > a > 0. (4-38)
n =0

The parameters a, b, 1, and 2 must be related to the moments of Xl and Yl, and to the
constants A, B, and C. As given in [1], these are obtained from
1/ 2
212 2 (1 2 2 )
a=
(1 + 2 ) 2
1/ 2
2 2 ( + )
b= 1 2 1 1 2 2
(1 + 2 )
1
1, 2 = w 2 + #w
4( xx yy | xy |2 )(| C | 2 AB)
(4-39)
A xx + B yy + C xy
*
+ C * xy
w=
4( xx yy | xy |2 )(| C |2 AB)

1 = 2(| C |2 AB)(| X 1 | 2 yy + | Y1 |2 xx X 1*Y1 xy X 1Y1* *xy )

2 = A | X 1 |2 + B | Y1 |2 +CX 1*Y1 + C * X 1Y1*

These equations are applied to our problem by comparing eqs. (4-34) and (4-35). Let-
ting Y1 = yk and X 1 = hk in (4-34), the constants A =| xk ,n |2 | xk ,i |2 , B = 0, and
C = x k ,n xk ,i are found, representing the properties of the modulation scheme. The be-
haviors of the channel and of the channel estimation technique will be expressed by
the first and second moments of the random variables X1 and Y1. These are

X 1 = E{hk }
Y1 = E{ y k }
xx = 12 [ E{| hk |2 } | X 1 |2 ] . (4-40)
yy = 12 [ E{| y k |2 } | Y1 |2 ]
xy = 12 [ E{hk y k* } X 1Y1* ]

The derivation of these parameters from the channel and system definitions is given in
4.3 Performance of an Uncoded OFDM System 123

the following sub-section. Coherent and differential detection are investigated.

4.3.2.1 Application of the Mathematical Models

Coherent detection with perfect channel estimation

The k-th symbol received is defined in eq. (4-28) as yk = xkhk + nk. Perfect channel esti-
mation means that the receiver has exact knowledge about the attenuation factor hk,
denoted by hk = hk . Considering the transmitted symbol xk,i as a constant yields
j
X 1 = E{hk } = e ,
j
Y1 = E{xk ,i hk + nk } = x k ,i E{hk } + E{nk } = xk ,i e ,
xx = 12 [ E{| hk |2 } | X 1 | 2 ] = 12 [ P0 2 ] , (4-41)
yy = 12 [ E{| x k ,i hk + nk |2 } | Y1 |2 ] = 12 [| xk ,i |2 ( P0 2 ) + N2 ] ,
xy = 12 [ E{hk ( xk ,i hk + nk )*} X 1Y1* ] = 12 xk*,i [ P0 2 ] ,
j
where e is the LOS-component, with arbitrary phase and with an amplitude de-
fined by 2 = P0 K ( K + 1) .

Differential detection
With differential detection, the decision for the received symbol yk is made based on
the adjacent symbol yk1 = xk1hk1 + nk1. For phase modulation schemes, this can be
seen as a detection based on the channel estimate
k 1 x k 1 = hk 1 + n ' k , where E{|n'k| } = N ' . Note that N ' = N , if
h = y / x = h + n 2 2 2 2
k k 1 k 1 k 1

the magnitude of xk1 is one. The parameters Y1 and yy are then equal as in eqs.
(4-41). The additional noise term n'k, the correlation between hk and hk1, and the Dop-
pler shift of the LOS-component are expressed in
j ( 2f T )
X 1 = E{hk 1 + n' k } = e ,
xx = 12 [ NRP + N2 ' 2 ] , and (4-42)
j 2f T
xy = 12 xk*,i [ E{hk 1hk* } 2 e ].
For evaluating differential detection in the frequency-direction, let T = 0. Using the
channel correlation functions given in Section 4.3.1, the correlation xy between the
attenuation factors at two adjacent symbols becomes

1
1 * P0 in frequency
xy = xk ,i 1 j 2F rms K1 . (4-43)
2 K +1
J 0 (2f mT ) in time

Note that the influence of the channel-variability is expressed by this correlation-value

only, which is defined by the parameter-products rmsF and fmT, for the two cases un-
124 Chapter 4 OFDM Introduction and System Modeling

der consideration.
Performance results are given in Section 4.3.3. It will be shown that differential detec-
tion in the time-direction is more robust than the frequency-domain variant, for the
OFDM system under consideration, i.e., for a wide-band indoor wireless LAN.

4.3.2.2 Application to Different Modulation Schemes

Assigning different constellation values to the variable xk,n xk,i, the probability can be
calculated that an erroneous symbol xk,n has been detected while the symbol xk,i was
transmitted. This allows for many modulation schemes an exact calculation of the
BER and for others the evaluation of close approximations.
In the following analysis we assume that all possible transmitted symbols xk,i {xk,m}
occur with equal probability.
BPSK and QPSK
Exact results can be obtained for BPSK and QPSK modulation. The signal constella-
tions for these techniques are depicted in Figure 4-10. For both schemes it is sufficient
to consider (any) one transmitted symbol, due to symmetries. This symbol will be the
+1, taken from the set {xk,m} = {1, 1} for BPSK, and from {xk,m} = {1, j, 1, j} for
QPSK. Note that |xk,m|2 = 1 for both modulation types.
BPSK is analyzed by evaluating the parameters A and C for xk,i = 1 and xk,n = 1. The
bit error probability is equal to the symbol error probability Pe (see eq. (4-37)).
Gray-coded QPSK transmits two bits per symbol on orthogonal carriers (I- and Q-
components). Thus, the error probabilities can be analyzed independently and the BER
equals their average. Suitable parameters for A and C are found by (e.g.) assigning
xk,i = 1 and xk,n = {j, j}.
Calculating the probability that xk,n has been detected, provided xk,i was transmitted,
the I/Q-plane is divided in two parts. An error occurs when the received symbol falls
within the half plane being closer to xk,n than to xk,i. No error occurs otherwise (see
Figure 4-10). Note that for the case of QPSK it is not necessary to evaluate the two-
error-event explicitly. The overlapping one-error-events account for one error each in

BPSK: Q QPSK: Q one error

(10) xk,n = j
(1) (0) (11) (00)
I I
xk,n = 1 xk,i = 1 xk,i = 1
xk,n = j
(01)
one error
Error region, when two errors
xk,i = 1 was transmitted

Figure 4-10: Selection of xk,i and xk,n for the performance evaluation of BPSK and QPSK.
4.3 Performance of an Uncoded OFDM System 125

(a): signal constellation (b): approximation

Q Q
(011) 1 less
(111) (001) 1 less xk,n = e+j/4
2 1
3 1 2 1
1
(101) (000)
2 more 3
I I
xk,i = 1 xk,i = 1
1 1 2 1
xk,n = e-j3/4
2 2 xk,n = e-j/4
(100) (010) 1 more
(110)

Figure 4-11: Error regions for 8-PSK, when xk,i = 1 was transmitted. (a) Signal constellations
and correct number of errors for each decision range. (b) Approximation by
evaluating error probabilities from the three error-states xk,n shown. In some error
regions, one extra error is considered, in other regions, one error is missed (indi-
cated as 1 more and 1 less).

this region, thus the two-error-event is included automatically. This may seem as an
advantage because computational complexity reduces, however, when evaluating
higher order modulation schemes, many of those half-planes will overlap and it is
sometimes not possible to obtain the exact number of errors for all decision regions.
This will be seen in the following case.
8-PSK
Upper and lower bounds on the BER can be calculated for 8-PSK. An exact calcula-
tion is not possible, because the eight signal states are not separable in the two or-
thogonal dimensions of the I/Q-plane.
Due to symmetries it is again sufficient to consider one transmitted symbol, xk,i = 1.
Figure 4-11 illustrates how errors occur in estimating error probabilities. The signal
constellations are shown in Figure 4-11a together with the exact numbers of errors for
each decision region. (Errors are denoted by .) Figure 4-11b shows the actual num-
bers of errors for each of these regions, when three different error states xk,n are evalu-
ated and averaged. Clearly, too few errors are considered in some of the decision re-
gions, while too many are considered in others. Thus the computational results are a
(close) approximation. The most likely errors, however, are appropriately treated.
16-Quadrature Amplitude Modulation (QAM)
16-QAM can be evaluated without any error. Four different transmitted symbols oc-
curring with equal probabilities and 24 error events must be considered. Some of them
must be subtracted in order to account for overlapping decision regions. A possible set
of symbols xk,i and xk,n to be used are listed in Table 4-1. The complex signal constella-
tions xk are denoted (Re{xk},Im{xk}). Error events whose probability must be sub-
tracted in the final result are written (Re{xk},Im{xk})1. All values must be divided by
10 to have an average power of one. Figure 4-12 illustrates the signal constellations
126 Chapter 4 OFDM Introduction and System Modeling

and error events for the symbol xk,i = (3,3), which carries the data symbol (0000).

Table 4-1: Transmitted symbols and error events for the evaluation of 16-QAM modulation.

Transmitted symbol Error symbols xk,n

xk,i
(3,3) (1,3), (3,3), (7,3)-1,
(3,1), (3,3), (3,7)-1
(1,3) (3,3), (1,3), (5,3),
(1,1), (1,3), (1,7)-1
(3,1) (1,1), (3,1), (7,1)-1,
(3,3), (3,1), (3,5)
(1,1) (3,1), (1,1), (5,1),
(1,3), (1,1), (1,5)

Q
(0000) (0001) (0011) (0010)
1 2 1
xk,i = (-3,3) xk,n = (-1,3) xk,n = (3,3) xk,n = (7,3)-1
error prob.
(0100) (0101) (0111) (0110) to be
subtracted!
1 2 3 2
xk,n = (-3,1)
I
(1100) (1101) (1111) (1110)
2 3 4 3

(1000) (1001) (1011) (1010)

1 2 3 2
xk,n = (-3,-3)

error prob. to
be subtracted!
xk,n = (-3,-7)-1

Figure 4-12: Illustration of the error events in 16-QAM, when the symbol xk,i = (3,3) was trans-
mitted.

16 Star-QAM
16 Star-QAM can be treated as a combination of 8-PSK and a binary amplitude
modulation. The binary AM is evaluated by transforming the I- and Q-variables to an
r2 = I2 + Q2 variable, resulting in similar expressions to the above defined ones. This is
described in [16] and [17].
4.3 Performance of an Uncoded OFDM System 127

BER for QPSK with coherent demod. (solid) and diff. demod. (dashed)
20
1e030
1e020
15

1e010
1e00
Rician Kfactor K [dB]

10 0.0 6
001
1e00
6
0.0
01
5

0.
00
01
0
0.01
0.1

0.001
0 5 10 15 20 25 30 35 40
Eb/N0 [dB]

Figure 4-13: Performance of QPSK for coherent detection (perfect channel estimation) ()
and for differential detection with F = 0, i.e., with perfect correlation between ad-
jacent sub-carriers (flat fading) ( ).

4.3.3 Performance Results

Some observations can be made from the mathematical expressions derived above
(eqs. (4-41)(4-43)):
(i) For coherent detection, the statistical parameters and thus the performance re-
sults only depend on P0, , and N2 . In other words, the performance depends
on the average signal-to-noise ratio (SNR) SNR P0 / N2 and on the Ricean K-
factor K = 2/(P0 2).
(ii) The same holds in the limits F 0 or T 0 (i.e., for flat fading) for differen-
tial detection.
(iii) The performance of differential detection degrades for F > 0 (or T > 0), because
of a systematic estimation error in hk = hk 1 + n' k , since hk1 hk. The parameter
products rmsF and fmT define the degradation, according to eq. (4-43).

Performance results (average BER) for (i) and (ii) and QPSK modulation are shown in
Figure 4-13, as a function of the average SNR per bit (denoted Eb/N0) 12 and as a func-

12
Several signal-to-noise ratio (SNR) parameters are used in this thesis: The SNR denoted Eb/N0 is
the average SNR per data bit. It thus depends on the order M of the modulation scheme. The average
SNR of the subcarrier symbols, being independent of the modulation scheme, is written as SNRSC (see
Section 6.2.7.2). In Section 6.2.3, the SNR of the time-domain OFDM signal is defined, written as
128 Chapter 4 OFDM Introduction and System Modeling

tion of K, where

Eb E{| x k , m |2 } P0
= . (4-44)
N0 log 2 (M ) N2

It is observed from the figure that the SNR required to achieve a certain BER-perform-
ance is dramatically increased for small K-factors (for Rayleigh channels). A 3-dB
disadvantage of SNR is evident for the differential detection method, since two noise
processes with equal variance are present , the noise of the channel estimate and the
noise of the data symbol to be detected.

A performance comparison of different modulation schemes is presented in Figure

4-14. Note that the result for 16-QAM (differential) is a theoretical one, because
differential demodulation for this scheme is hard to accomplish. Again the advantage
of a high K-factor is seen. With coherent detection, equivalent performance is obtained
for BPSK and QPSK. This is not the case for differential detection on AWGN or
Ricean channels, where BPSK has an additional advantage of 12 dB over QPSK (see
also [1]). It is important to note that twice the symbol energy is used with QPSK, be-
cause two bits are transmitted per symbol. The higher order modulation schemes (8-
PSK and 16-QAM) require approximately 34 dB more signal power than QPSK.

Taking into account the channel variability, irreducible error floors arise (see Figure
4-15). Both versions of differential detection have been evaluated for Rayleigh fading
channels, QPSK modulation, and for the following parameters. For detection in the
frequency-direction, the channels RMS delay spread rms was assumed to be three
samples, which corresponds to a maximum delay spread of about thirty samples, as-
suming an exponentially decaying channel delay profile (see Section 2.3.3.3). For 128
FFT-points, this value corresponds to about one quarter of the FFT-time, which is also
about the time duration that would be selected for the guard interval. It is seen that the
irreducible error floor associated with such quite realistic parameters (rmsF =
3/128) lies around 102 (curve {{).
The time-variability for differential detection in time-direction corresponds to a mobile
moving at 20 m/s, to a carrier frequency of 60 GHz, and to a symbol length of 1.3 s.
According to the system model, inter-carrier-interference (ICI) due to the Doppler
spreading has been neglected 13. While the symbol duration assumed is rather short,
the mobility considered is by an order of magnitude higher than the expected mobility

SNR. This value is different to the previous ones, because not all FFT-points are used for data sub-
carriers. In Chapter 8, the coding scheme is incorporated in some other SNR parameters.
13
That ICI truly is negligible for the system parameters selected is suggested from the comparison
of the maximum Doppler frequency and the OFDM sub-carrier spacing. The former, being 4 kHz is
just 0.4 % of the latter, which is 1 MHz. An approximate equation for the SNR-degradation due to mo-
bility can be found for instance in [7]. Its evaluation leads to the same conclusion (see Section 6.5.3.1).
4.3 Performance of an Uncoded OFDM System 129

Performance of OFDM with perfect channel estimation

0
10
BPSK, K =
BPSK, K = 4
BPSK, K = 0
QPSK
1
10 8PSK
16QAM
average BER

2
10

3
10

4
10
10 5 0 5 10 15 20 25 30
E /N
b 0

(a)
OFDM with differential detection and flat fading (F = 0)
0
10
BPSK, K =
BPSK, K = 4
BPSK, K = 0
QPSK
1
10 8PSK
16QAM
average BER

2
10

3
10

4
10
10 5 0 5 10 15 20 25 30
Eb/N0

(b)
Figure 4-14: Performance of different modulation schemes. (a): Coherent detection with perfect
channel estimation. (b): Differential detection with F = 0, i.e., perfect correlation
between adjacent sub-carriers (flat fading).

in an indoor wireless-LAN system. Despite this, the error-floor is much lower for this
method of differential detection (curve .
The other results depicted analyze the influence of synchronization errors. In the fre-
quency-domain results, ICI and ISI have been neglected, which is exact as long as the
130 Chapter 4 OFDM Introduction and System Modeling

Performance of differential detection; QPSK; Rayleigh channel

0
10

1
10
average biterrorrate (BER)

2
10

= 3 smp.; t = 0 smp.
rms
rms = 3 smp.; t = 3 smp.
3
10 rms = 3 smp.; t = 3 smp.
v = 20 m/s; f = 0
v = 20 m/s; f = 3% of F; w/o ICI
v = 20 m/s; f = 3% of F; incl. ICI
static channel; perfect sync.
4
10
0 5 10 15 20 25 30 35 40
average Eb/N0 [dB]

Figure 4-15: The channel variability is leading to irreducible error floors for the differential
modulation schemes. Differential QPSK is evaluated over Rayleigh fading chan-
nels.

channel impulse response remains within the guard interval. In other words, the per-
formance impact results from the systematic phase rotations only that are as seen
from the extended system model (Section 4.2.4) due to the time-synchronization er-
ror. Such phase rotations mean for the differential detection techniques that the signal
constellation points compared typically move closer together, which implies a per-
formance degradation.
Note however, that a negative timing offset equal to the channels rms slightly im-
proves the performance. This is because the channel as well induces some systematic
phase rotation, which is in the case discussed cancelled by the phase rotation due
to timing-offset 14. The impact on the performance is small, however, for the offsets
evaluated.
In the curves depicting the performance for the detection scheme in time-direction, the
impact of ICI due to a frequency-offset is shown as well. ICI has been included using
the approximation from [7]. It is seen that the impact of the phase distortion is evident
at all SNR-values, while ICI determines the error floor at high SNR.

In Figure 4-16, the performance of differential QPSK (in frequency-direction) is

14
The progressive phase rotation due to a timing-offset can be utilized for timing synchronization
(see Section 6.2.7). Thereby, the systematic phase rotation due to the channel leads to a bias in the
estimate. If this biased estimate is used for timing-synchronization, optimum performance is achieved,
because the systematic phase rotations due to the channel and due to the bias compensate each other.
4.4 Conclusions and Recommendations 131

BER for QPSK with diff. demod. T = {0 (dotted), .11 (solid), .27 (dashed)}
m
20
1e030
1e020
15
1e 1e010
0
0. 06

0.01
00
01
Rician Kfactor K [dB]

10 1e006
0.
0.1
00
1
0.0001
5
0.001

0.01
5
0.1

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

Figure 4-16: Performance of QPSK with differential detection in frequency-direction. The

maximum excess delay of the channel is related to the FFT-time, expressed by Tm.

shown as a function of Eb/N0 and K, where rms is a parameter. Since the maximum ex-
cess delay of the channel which should not exceed the guard interval is a function
of rms and K, all these parameters are interrelated. The following definitions are intro-
duced to get a set of general results. The FFT duration and GI-duration are connected
by a fixed factor, which is usually in the range of 4 10. The maximum excess delay
of the channel can be written max = 10rmsK1, according to the definition of Section
2.3.3.3 (see eq. (2-24)). This leads to the normalized excess delay, defined as
Tm = max TFFT = 10 rms K1 F . In Figure 4-16, the performance of differential QPSK is
shown for Tm = {0, 0.11, 0.27}. The curves for Tm = 0 allow a comparison with Figure
4-13. Especially for severely fading channels (low K-factors), the performance degra-
dation is significant for the delay spreads considered.

4.4 Conclusions and Recommendations

The derivation of the OFDM system model has confirmed that data symbols can be
transmitted independently over multipath fading radio channels. It has to be assumed,
however, that the channels maximum excess delay is shorter than the guard interval,
and that the system has been synchronized sufficiently. Small synchronization errors
lead to systematic phase rotations of the data constellation points a property which
can be exploited for estimating synchronization offsets. If the timing- or frequency-
synchronization error becomes too large, the orthogonality of the sub-carriers is partly
lost and the signal-to-noise ratio of the system is degraded. That is, inter-carrier-inter-
132 Chapter 4 OFDM Introduction and System Modeling

ference (ICI) and inter-symbol-interference arise. ICI can also result from very fast
channel variations (Doppler spreads) or from carrier phase jitters.
The system models presented can be utilized in analytical studies of various aspects of
the OFDM technique, as, for instance, in the performance evaluation. The basic model
introduced assumes perfect synchronization, while an extended model considers the
phase rotations due to small synchronization-offsets.

The performance analysis of an uncoded OFDM scheme is based on the classic for-
mulas given by Proakis ([1]: Appendix B). Expressions are derived for the evaluation
of different modulation schemes and for coherent and differential detection. The fre-
quency-domain channel model (see Chapter 2) for Ricean fading channels has been
applied. It allows to show performance results as a function of the channel parameters
{P0, K, rms} the normalized received power, the Ricean K-factor, and the RMS de-
lay spread.
Assuming perfect channel estimation, or if differential schemes are applied com-
plete channel correlation, the performance is determined by P0 and K. These parame-
ters specify the average signal power and the depth of the fades. Better performance is
thus achieved over channels having a higher K-factor, because the fades are shallower.
Performing differential detection in the frequency-domain, a degradation of the results
is seen, due to the small differences of the channel transfer function at adjacent sub-
carriers (whose data symbols are compared). Since, for small frequency-lags, there is a
very strict relationship between this correlation function and the RMS delay spread,
rms, of the channel (in particular for Rayleigh fading channels see Appendix A), it is
concluded that the performance degradation is well characterized by rms. (To be exact,
the performance is defined by the product rmsF, where F is the sub-carrier spacing.)
Imperfect timing-synchronization also has an impact, because systematic phase offsets
are introduced between adjacent sub-carriers.
For the low-mobility OFDM based WLAN system under investigation, the correlation
of subsequent symbols in time-direction is much higher than the correlation of sym-
bols on adjacent sub-carriers. Therefore, it is recommended to apply differential de-
tection in time-direction, not in frequency direction. In this case, systematic phase off-
sets are induced by imperfect carrier frequency-synchronization.
By extending the OFDM system model, it becomes possible to analyze imperfections
of OFDM systems. Frequency synchronization-offsets, for example, give rise to inter-
carrier-interference, which can be accounted for by an additional noise term [11]. In a
similar fashion, the impact of Doppler spreads, phase noise, or channel estimation er-
rors can be incorporated. The evaluation of such imperfections is a topic for future
work. Using the original equations of [1], it is also possible to investigate diversity
techniques (see e.g. [15][17]).
4.5 References 133

4.5 References
[1] J. G. Proakis, Digital Communications, 3rd edition. New York: McGraw Hill,
1995.
[2] R. Prasad, Universal Personal Communications. Boston: Artech house, 1998, ch.
10.
[3] R. van Nee and R. Prasad, OFDM for Wireless Multimedia Communications.
Boston: Artech House, 2000.
[4] O. Edfors, M. Sandell, J. J. van de Beek, D. Landstrm, F. Sjberg, An Intro-
duction to Orthogonal Frequency-Division Multiplexing, Research Report
TULEA 1996:16, Division of Signal Processing, Lule University of Technology,
http://www.sm.luth.se/ csee/sp/publications.html.
[5] O. Edfors, Low-complexity algorithms in digital receivers, Ph.D. Thesis, Lule
University of Technology, Sept. 1996.
[6] M. Sandell, Design and Analysis of Estimators for Multicarrier Modulation and
Ultrasonic Imaging, Ph.D. Thesis, Lule University of Technology, Sept. 1996.
[7] M. Speth, S. A. Fechtel, G. Fock, and H. Meyr, Optimum Receiver Design for
Wireless Broad-Band Systems Using OFDMPart I, IEEE Trans. Commun.,
vol. 47, no. 11, pp. 16681677, Nov. 1999.
[8] S. B. Weinstein and P. M. Ebert, Data Transmission by Frequency-Division
Multiplexing Using the Discrete Fourier Transform, IEEE Trans. Commun.
Techn., vol. COM-19, no. 5, pp. 628634, Oct. 1971.
[9] A. Peled and A. Ruiz, Frequency Domain Data Transmission Using Reduced
Computational Complexity Algorithms, in Proc. IEEE Int. Conf. Acoust.,
Speech, Signal Processing, Denver, CO, 1980, pp. 964967.
[10] P. Pollet, M. van Bladel, and M. Moenclaey, BER Sensitivity of OFDM Sys-
tems to Carrier Frequency Offset and Wiener Phase Noise, IEEE Trans. on
Commun., vol. 43, no. 2/3/4, pp. 191193, Feb./March/April 1995.
[11] P. H. Moose, A technique for orthogonal frequency division multiplexing fre-
quency offset correction, IEEE Trans. Commun., vol. 42, no. 10, pp. 2908
2914, Oct. 1994.
[12] P. A. Bello, Characterization of randomly time-variant linear channels, IEEE
Trans. on Commun. Systems, vol. CS-11, pp. 360393, Dec. 1963.
[13] R. Steele, Mobile Radio Communications. New York: John Wiley & Sons, 1992.
[14] W. C. Jakes Jr., Microwave Mobile Communications. New York: John Wiley &
Sons, 1974.
134 Chapter 4 OFDM Introduction and System Modeling

[15] J. Lu, T. T. Tjhung, F. Adachi, and C. L. Huang, BER performance of OFDM-

MDPSK systems in frequency-selective Ricean fading with diversity reception,
IEEE Trans. Veh. Technol., vol. 49, no. 4, pp. 12161225, July 2000.
[16] X. Dong, T. T. Tjhung, and F. Adachi, Error Probability Analysis for 16 STAR-
QAM in Frequency-Selective Ricean Fading with Diversity Reception, IEEE
Trans. on Veh. Techol., vol. 47, no. 3, pp. 924935, Aug. 1998.
[17] T. T. Tjhung, F. Adachi, K. H. Tan, X. D. Dong, and S. S. Ng, BER Perform-
ance of 16 STAR-QAM in Ricean Fading with Diversity Reception, in Proc.
PIMRC94, The Hague 1994, pp. 8084.
Chapter 5 OFDM System Proposal and
Emulation System

5.1 Introduction
Main parts of this Ph.D. research have been performed under the framework of a co-
operative research project between Korea Telecom and Delft University of Technol-
ogy, as mentioned in the introduction. One of the goals of this project was the devel-
opment of an OFDM based wideband communications system operating at up to 155
Mbit/s in 60 GHz radio channels. Wireless asynchronous transfer mode (ATM) trans-
mission was considered in indoor and low-range (and low mobility) outdoor environ-
ments.
The first part of this chapter (Section 5.2) gives an overview of the proposed OFDM
system. Most of the parameters given and choices made are motivated in the text; oth-
ers have been selected rather intuitively. One purpose of the system proposal was to
specify an OFDM system that could be used as a benchmark system in analytical
studies and simulation studies of several aspects of the OFDM technique. These stud-
ies are described in the remaining chapters of this thesis. They mostly concern imple-
mentation issues of OFDM, like synchronization and channel estimation, and their per-
formance evaluation. Some of the conclusions have been incorporated in the system
proposal given here. That is, although some parameters may have been selected ad-
hoc, their suitability was investigated and confirmed afterwards.

The proposed system consists of a central base station (BS) and several mobile termi-
nals (MT). The base station acts as an interface between the physical transmission me-
dia of the wired ATM backbone network and the broadband radio air-interface. There-
fore, it must be optimized for operating robustly and efficiently in these two very dis-
similar communications media , the highly reliable optical fiber network and the ran-

135
136 Chapter 5 OFDM System Proposal and Emulation System

domly fading radio channel.

The target bit-rate of 155 Mbit/s is reached under ideal conditions, and when the up-
link and down-link data rates are added up. Channel bandwidths in the order of 100
MHz are required; in multi-cell systems, several of these channels must be available
for adjacent base stations. Such large bandwidths may be available in the 60 GHz fre-
quency-band.
The air-interface is the main research topic of this Ph.D. thesis, hence the system pro-
posal is focused on the physical layer. Orthogonal frequency division multiplexing
(OFDM) is considered to be the key enabling technology for such high data rates in
multipath fading, mobile channels. One ATM cell was assumed to be the smallest in-
formation entity to be transmitted at a time. A multiple access scheme having both
time- and frequency division multiple access components (TDMA and FDMA), maps
the ATM cells on the OFDM symbols, after interleaving and error correction coding.
Periodic training symbols are transmitted on the down-link for synchronization and
channel estimation. These training symbols are followed by signaling symbols, which
are used by the medium access control (MAC) protocol to negotiate transmission ca-
pacity.
On the up-link, pre-equalization and time-division duplexing (TDD) are proposed for
minimizing the signal processing required and for enabling asymmetric data rates. Re-
ciprocity of the channel is assumed. Pre-equalization allows the efficient transmission
of information entities as short as a single ATM cell on the up-link, without adding
vast amounts of overhead for synchronization and channel estimation. It also reduces
the signal processing needed at the base station for serving multiple mobile terminals.
Slow time-variability of the radio channel is a necessity, however.

Another primary goal of the cooperative research project was the demonstration of the
proposed air-interface on a hardware platform to be developed. Unfortunately, the
technical specifications of the investigated systems are so demanding that an imple-
mentation is almost impossible for a small research team, as real-time DSP of two data
streams is required at sampling frequencies of ~100 MHz. A drastically downscaled
hardware platform is therefore used, which is presented in Section 5.3.

5.2 OFDM Based System Proposal

A highly efficient multiple access scheme is one of the main requirements of the pro-
posed, OFDM-based, multimedia communications system. In order to minimize delays
in low-rate applications and in automatic repeat request (ARQ)-schemes, which are
needed to meet defined quality-of-service (QoS) requirements, the system should be
able to transmit small data entities (single ATM cells) without the need for excessive
signaling and synchronization overhead. This design target becomes critical because of
the high data rate of up to 155 Mbit/s.
5.2 OFDM Based System Proposal 137

Another important aspect was to consider an option for simultaneously accessing the
system with low-complexity terminals that can only cope with one quarter of the sys-
tem bandwidth. Such transceivers have simplified RF front-ends and baseband proc-
essing units, which both lead to reduced power consumption. Moreover, the limited
bandwidth implies a 6-dB advantage in their link-budgets.
Adaptability to current channel conditions is provided by two mechanisms. Two
transmission modes (modes I and II) with different delay spread robustness can be
used according to the typical channel properties of the environment. Their main differ-
ence is the length of the guard interval and the number of FFT-points and sub-carriers.
A set of coding and modulation techniques allows for a trade-off between the range
and the bit-rate. I.e., Using more efficient coding techniques and more robust modula-
tion schemes, the range can be extended at the cost of a decreased data rate.
The key specifications and design paradigms considered are:
up to 155 Mbit/s ATM data rate
operation in the 60 GHz (mm-wave) frequency-band
indoor pico-cells and short range outdoor environments (< 100 m diameter)
limited mobility (max. 25 m/s)
fulfill defined QoS and delay requirements
provide high spectral efficiency
low implementation complexity (to cope with the high data rates)
low power consumption
high robustness against delay spread and interference
provide adaptability to changing channel conditions (e.g. fall-back modes)

This section introduces the OFDM system proposal, which takes the above listed re-
quirements into consideration. It is organized as follows. In Section 5.2.1, the main
transmission modes are explained, and the selected time-division duplex (TDD) frame
structure and OFDM symbol configuration are introduced. The multiple access scheme
having time and frequency division multiple access (TDMA, FDMA) components is
described in Section 5.2.2. Section 5.2.3 presents the hardware architecture of base sta-
tions and mobile terminals in the form of block diagrams. Interleaving and coding
schemes are outlined in Section 5.2.4, followed by link budget considerations in Sec-
tion 5.2.5. In Section 5.4, the system proposal is summarized and preliminary conclu-
sions are drawn.

5.2.1 Air Interface Physical Layer

The OFDM modulation technique is considered a strong candidate for the air-interface
of high-speed wireless LANs and similar systems. The main advantages of OFDM in
this context are its high spectral efficiency and robustness against multipath delay
138 Chapter 5 OFDM System Proposal and Emulation System

spread. The latter is obtained through the introduction of a guard interval. The well-
known, and extensively studied disadvantages of this modulation technique include
[1]:
Sensitivity against frequency offsets and phase jitters
Non-constant power envelope (requiring linear power amplifiers to avoid distor-
tion)
These issues have to be solved sufficiently, otherwise the orthogonality among data
symbols is partly lost (ICI and ISI arise). To overcome, or at least ease these impair-
ments, some of the design considerations of the proposed OFDM scheme were:
Transmit sufficient overhead (training-symbols) on the down-link to allow the MTs
to synchronize to the BS with low computational complexity and to allow for accu-
rate and efficient channel estimation.
Use as little carriers as possible, which is of advantage for the frequency synchroni-
zation algorithm, for the peak-to-average power ratio (PAPR) problem, and for the
robustness against carrier phase jitter and Doppler spreads.

5.2.1.1 Modes of Operation

For different environments, antenna configurations and link qualities, several modes of
operation are proposed. There are two incompatible main modes (mode I and mode
II) that will be permanently assigned to a base station, depending on the properties of
the environment. Mode II has a higher robustness against long delay spreads, which
may be required in very large rooms (cf. Section 2.5.2). This higher robustness is ob-
tained by increasing the guard interval duration, the symbol length, and the number of
FFT points and sub-carriers. Mode I allows for maximum excess delays of about 200
ns (~25 ns RMS delay spread, rms), mode II permits about 500 ns (rms 65 ns). Ex-
ceeding those maximum values leads to a gradual performance degradation that may
be tolerable if the signal-to-noise ratio is sufficient.
Mode I has advantages due to the smaller number of FFT points, allowing for relaxed
hardware requirements. For instance, the robustness to phase noise and Doppler shifts
is enhanced, the PAPR problem is reduced, and the efficiency is slightly higher, due to
a more efficient frame format. These advantages may enable the design of cheaper ac-
cess points for small environments, where mode I is typically appropriate and suffi-
cient.

Another design goal was to provide the possibility to simultaneously access the net-
work with full and reduced (one quarter) bandwidth mobile terminals, as mentioned
above. These two options are denoted as the full-rate (-fr) and the quarter-rate (-qr)
bandwidth modes. Using four sub-symbols per OFDM symbol, the coded ATM
cells are mapped on the transmission medium in a flexible TDMA/FDMA scheme (see
Section 5.2.2). Table 5-1 lists the characteristic parameters of modes I and II,
5.2 OFDM Based System Proposal 139

Table 5-1: Characteristic parameters of the transmission modes I and II for full and quarter
bandwidth terminals

mode: I II
parameter full quarter full quarter
Number of FFT points 128 32 256 64
Nr. of data + pilot sub-carriers 76 + 12 19 + 3 152 + 24 38 + 6
FFT time [s] 1.000 2.000
in [samples] 128 32 256 64
Guard + windowing interval [s] 0.187 + 0.094 0.47 + 0.094
in [samples] 24 + 12 6+3 60 + 12 15 + 3
Bandwidth (-20 dB) [MHz] ~ 100 ~ 28 ~ 100 ~ 28
Sampling frequency [MHz] 128 32 128 32

operating at full and quarter bandwidth.

Quarter rate terminals cannot use the full transmission rate, however, they have ad-
vantages regarding the complexity of the base-band processing unit and the RF front-
ends, regarding the link budget (see Section 5.2.5), and regarding the power consump-
tion. The quarter-rate mode can also serve as a fallback mode for the full-rate users,
when shortcomings are experienced in their link-budgets. (This applies particularly for
the up-link). But the reduced bandwidth also implies some disadvantages. Less band-
width means less frequency-diversity, thus considering forward error correction
coding the performance is expected to be worse. The bit- and frame-error-rate analy-
sis of a coded and interleaved OFDM system confirms this expectation (see Section
8.2). Frequency hopping across the four quarter-rate sub-bands would reduce this per-
formance gap. Another possible remedy is the application of diversity schemes, as
proposed in Section 8.3.

The transmission speed is determined by the modulation technique and coding scheme
used. OFDM allows for high flexibility in using different modulation and coding tech-
niques simultaneously, depending on the users channel qualities. The main options
considered are QPSK and 16-QAM modulation with coding rates of approximately
and . However, many other combinations can be applied as well. These modulation
modes are referred to as sub-modes -H, -L, and -LL. Their description and the actual
transmission rates achieved are presented in Section 5.2.1.4, Table 5-3.

5.2.1.2 Frame Format and Modulation Schemes

A time division duplexing (TDD) frame format has been proposed, which has a fixed
total frame length and a flexible boundary between the down-link and the up-link in
order to support asymmetric data rates. To simplify the switching of the transceivers
between transmission and reception and to be able to compensate for propagation de-
140 Chapter 5 OFDM System Proposal and Emulation System

Mode I: flexible boundary

down-link up-link

time
data symbols data symbols free data 69T = 88.4 s
training TDD guard- symbols TDD guard-
down-link up-link signaling
sequence symbol symbol
signaling

Mode II:
down-link up-link
time

t=0 372T = 94.8 s

Figure 5-1: Frame structure of the time division duplex scheme in modes I and II. Each rec-
tangle indicates one OFDM symbol.

lays, null-symbols are inserted. An illustration is given in Figure 5-1; the frame related
parameters for the two transmission modes are listed in Table 5-2.
Using a TDD scheme, most synchronization can be done by the mobile terminals on
the down-link. For this purpose, the base station transmits a training symbol at the be-
ginning of each frame, being a unique OFDM symbol. This symbol is used for (frame)
timing and frequency synchronization, as well as for the derivation of a channel esti-
mate for the down-link channel. The required signal-processing algorithms are de-
scribed and evaluated in Chapter 6.
There is no training symbol on the up-link, assuming reciprocity of the channel. In or-
der to facilitate coherent detection on the up-link, pre-equalization techniques are sug-
gested. I.e., the up-link data constellation points are multiplied by the inverse of the

Table 5-2: Frame related parameters of the OFDM system proposal

parameter mode I mode II mode II

OFDM symbol duration T [s] 1.28 2.56
OFDM symbols per frame 69 37 35
frame duration [s] 88.4 94.8 89.7
frame rate [kFrames/s] 11.31 10.54 11.15
TDD guard symbols per Frame (overhead) 2 (2.8 %) 2 (5.4 %) 20.5 (2.8 %)
synchronization symbols per Frame 1 (1.4 %) 1 (2.7 %) 1 (2.8 %)
signaling symbols per Frame 2 (2.8 %) 2 (5.4 %) 1 (2.8 %)
signaling channels per Frame; up + down- 4+4 4+4 4 up-link or
link (4 per OFDM symbol) 4 down-link
modulation for signaling BPSK
coding rate for signaling ~
data bits per signaling channel per Frame 9.5 19
signaling rate per signaling ch. [kbit/s] 107.5 200.4 105.9
5.2 OFDM Based System Proposal 141

channel transfer function that has been estimated on the down-link. Passing the pre-
equalized OFDM signal through the multipath channel, the overall transfer function
becomes flat or phase-linear. Channel estimation and synchronization on the up-link
thereby reduce to an estimation of the magnitude of the received constellation values,
and the estimation of timing-offsets and carrier phase-offsets. Chapter 7 explains those
techniques.
In mode II, the number of OFDM data symbols per frame was halved in order to
maintain the rate of synchronization symbols. (Below, a discussion on the frame length
and channel time-variability is given.) For the sake of simplicity, the overhead
symbols (TDD guard, synchronization and signaling) were adopted from mode I,
resulting in an increased overhead and a doubled data rate on the signaling channel15.
Doubling the period of the synchronization symbol is reasonable because the
requirements on the frequency synchronization are also doubled.
Mode II' introduced in the table above is a slightly more efficient proposal compared
to mode II. The efficiency is increased by alternatingly transmitting signaling channels
on the up- and down-link, and by halving the TDD guard intervals. According to mode
I, the transceivers have to be able to switch between up and down-link within the du-
ration of one mode I OFDM symbol. Thus extending the TDD guard symbols is a
waste of efficiency. It has to be evaluated, however, if the propagation delays and filter
group delays can be fit within this guard period. The up-link transmission must start
slightly in advance to ensure that the up-link symbols FFT-periods arrive at the ideal
time-instants.
Channel variability and mobility
In these paragraphs, the speed and character of the time-variability of the channel are
briefly analyzed. Assuming the maximum speed in an indoor environment to be vm 2
m/s, the maximum Doppler frequency is fm = vm/ = 400 Hz at 60 GHz. Defining the
coherence time (t)c as the time-separation for which the spaced-time correlation
function of the channel H(t) = 0.9, we obtain for the above fm and Jakes fading
model [2] (t)c = 0.25 ms (see eq. (4-31) in Section 4.3.1 for K = 0). For the above
defined OFDM system, 0.25 ms are equal to ~200 (mode II: ~100) OFDM Symbols or
approximately 3 frames, during which the channel is almost constant.
Alternatively, the suitability of the proposed frame duration can be evaluated based on
the sampling theorem. In order to use the training sequence transmitted in the begin-
ning of each frame for estimating the channel, the Nyquist criterion has to be fulfilled.
To track the variations of a channel with maximum Doppler frequency fm, the chan-
nels transfer function has to be sampled at time-instants separated by at most

15
A doubled data rate on the signaling channels in mode II can be useful, because mode II may
operate in bigger environments with higher numbers of potential users. Thus, a higher signaling data
rate might increase the total throughput.
142 Chapter 5 OFDM System Proposal and Emulation System

t pilot < 1 (2 f m ) . This yields tpilot < 1.25 ms for fm = 400 Hz. Thus the frame duration
of ~90 s guarantees sufficient oversampling of the channel estimates. We anticipate
that the proposed OFDM system is suitable for such slowly time-variant radio chan-
nels, but not for (much) faster mobility. Performance results for the air-interface over
time-variant channels are given in Section 6.5 of this thesis.

5.2.1.3 OFDM Symbol Configuration

Sub-bands and OFDM sub-symbols

Figure 5-2 shows the OFDM symbol configurations in modes I and II. Several zero-
carriers are introduced. The zeros at fc,0 + {24, 0, +24} MHz are used to separate the
four quarter rate sub-bands, where fc,0 is the center frequency of the full-rate OFDM
channel.
The zero at fc,0 is also required to avoid problems with carrier feed-through and
AD/DA converter offsets, as explained in Section 4.2.2.3. For the same reason, zero-
carriers are used at the center frequencies of the quarter rate channels, which are lo-
cated at fc,0 + {36, 12, 12, 36} MHz.
In mode II, the number of (data and pilot) sub-carriers is doubled. Three zero sub-car-
riers are separating the OFDM sub-bands in order to keep the sub-band center fre-
quencies at fixed positions. These extended guard-bands may be of benefit to avoid
inter-sub-band-interference, because the requirements on carrier stability are higher in
mode II. Alternatively, two of the zero sub-carriers may be used as additional pilots.
Discussion of pilot assignment and purpose of pilots
The pilots present in the system cannot be used for channel estimation, because their
frequency spacing is too large. In stead, they are used for synchronization purposes
and for coherent demodulation as follows. On the down-link and up-link, the pilots are
needed for the estimation of carrier phase-offsets due to residual carrier frequency-er-
rors, which lead to a common phase rotation of all sub-carrier constellation values (see
Section 4.2.4.2). On the up-link, they are also used for the tracking of up-link time-de-
lays, which cause a progressive phase rotation of the sub-carriers (see Section 4.2.4.1).

Mode I; full rate Mode I; quarter rate (sub-band 1)

pilots
sub-band 3 sub-band 4 sub-band 1 sub-band 2
frequency f
fc,3 fc,4 fc,0 fc,1 fc,2 fc,0 + 48 MHz fc,1 fc,1 + 12 MHz

Mode II; full rate Mode II; quarter rate (sub-band 1)

sub-band 3 sub-band 4 sub-band 1 sub-band 2
frequency f
fc,3 fc,4 fc,0 fc,1 fc,2 fc,1

Figure 5-2: OFDM symbol configurations in the different operation modes

5.2 OFDM Based System Proposal 143

Those delays must be known very accurately, when coherent demodulation is applied.
On the one hand, pilots should be distributed evenly across the signal bandwidth in or-
der to exploit frequency diversity when some of them are in deep fades. On the other
hand, a constant spacing between adjacent pilots can be problematic in a multipath
channel with two main paths. For a two-path channel with relative delay time , the
channel transfer function has minima separated by 1/. Thus, if the pilot spacing fpilot
= 1/, all pilots might be in a deep fade at the same time. A short calculation is given
to show that the latter problem can probably be neglected in the system under investi-
gation.
Distributing the number of pilots proposed for the system evenly across the signal
bandwidth leads to a constant spacing of 8 and 4 MHz among adjacent pilots, in modes
I and II, respectively. Such a fpilot corresponds to a of 125 and 250 ns, respectively,
or equivalently, to path-length differences of 37.5 and 75 m. Due to the very limited
sizes of the radio cells in the proposed system and due to the low transmission powers,
it is rather unlikely that rays with such large path-length difference can significantly
interfere with each other. Therefore, a constant pilot spacing can be used, which also
simplifies the algorithm for extracting the up-link timing-offset (see Section 7.5.2).
A pilot assignment with a non-uniform spacing is presented in Section 5.2.4.
Spectral shape
The spectral shape of the proposed OFDM operation modes is shown in Figure 5-3,
assuming a perfect power amplifier. That means, out-of-band radiation because of
non-linear distortion of the signal in non-linear amplifiers is not included. Note that the

spectra of full rate users

10

0
power spectrum magnitude [dB]

10

20

30

40

50

60
mode I
mode II
70
60 40 20 0 20 40 60
f in MHz

Figure 5-3: Spectra of the OFDM signals in different operation modes. Null-carriers are used
to separate the four sub-bands and to avoid data transmission at the DC-carrier(s).
144 Chapter 5 OFDM System Proposal and Emulation System

spectra of mode II are shifted by 10 dB in that figure. It is seen that the spectral shape
is very similar for modes I and II. The steep decay at the band edges is due to the time-
domain windowing applied (see Figure 4-2, Section 4.2.2.2). The purpose of this type
of windowing is the reduction of out-of-band radiation in order to allow a close spac-
ing of adjacent frequency-bands (channels) used by the system in different radio cells.
In the current system proposal, the length of the window is quite significant, being
~10 % and ~5 % of TFFT for mode I and II, respectively. While such a long windowing
interval can be seen as an extension of the guard interval, it also means a loss of effec-
tive transmission power, additional to the loss due to the guard interval. Considering a
real system, filters are required for channel-selection, aliasing suppression (ADC), and
signal reconstruction (DAC). The specifications of those filters can be slightly relaxed,
due to the windowing.

5.2.1.4 Data Transmission Rates

Table 5-3 lists the data transmission rates that can be achieved in the different opera-
tion modes with various modulation and coding schemes and for full and quarter
bandwidth terminals. It is a topic for further research, whether the proposed modula-
tion and coding techniques are the ideal combinations.
The gross data rates listed in Table 5-3 are the numbers of data bits per OFDM symbol
divided by the duration of an OFDM symbol. Thus, the overheads introduced for sig-
naling (signaling channels and signaling overheads appended to ATM cells), synchro-

Table 5-3: Transmission rates using different modulation and coding techniques (LL: Very
low speed; L: Low speed; H: High speed). Values in brackets are for mode II'.

mode band- modu- coding OFDM sym- data bits gross data ATM bit-rate
width lation rate bols per / OFDM rate in in Mbit/s
mode ATM cell symbol Mbit/s
I-LL full QPSK 6 76 59.3 51.2
quarter 24 19 14.8 12.8
I-L full QPSK 4 114 89.0 76.7
quarter 16 28.5 22.2 19.2
I-H full 16- 2 228 178.0 153.5
quarter QAM 8 57 44.5 38.4
II-LL full QPSK 3 152 59.3 47.7 (50.4)
quarter 12 38 14.8 11.9 (12.6)
II-L full QPSK 2 228 89.0 71.6 (75.6)
quarter 8 57 22.2 17.9 (18.9)
II-H full 16- 1 456 178.0 143.1 (151.3)
quarter QAM 4 114 44.5 35.8 (37.8)
5.2 OFDM Based System Proposal 145

Table 5-4: Summary of all overheads required for wireless access and ATM. The total val-
ues are relative to the gross data rate, i.e., the data rate including coding.

description mode I mode II mode II'

TDD guard symbols 2.8 % 5.4 % 2.8 %
synchronization training sequence 1.4 % 2.7 % 2.8 %
signaling in signaling channels 2.8 % 5.4 % 2.8 %
total overhead per frame 7.2 % 13.5 % 8.6 %
signaling appended to ATM cells (4/57 bytes) 7% 7% 7%
total overhead for wireless access 13.7 % 19.6 % 15.0 %
overhead of ATM headers (5/53 bytes) 9.4 % 9.4 % 9.4 %
total overhead relative to gross data rate 21.9 % 27.2 % 23 %

nization, and time division duplexing have not been subtracted. Pilots are excluded.
The sub-modes -H, -L and -LL use different modulation and FEC-techniques, namely,
QPSK and 16-QAM modulation with rate - and rate -coding, as seen from the ta-
ble.
Depending on the modulation and coding scheme, a varying number of OFDM sym-
bols is required for the transmission of one ATM cell (see also Section 5.2.2). Each
ATM cell consisting of 48 data bytes and 5 header bytes is augmented by 4 bytes for
the signaling required by the MAC-protocol. I.e., a data entity comprises 456 informa-
tion bits, which make 912 and 684 coded bits, respectively, with rate - and rate -
coding. The amount of signaling information was estimated based on the work pre-
sented in [3].
The ATM bit-rates listed in the last column still include the overhead of the ATM
headers, but none of the overheads introduced for the wireless access. Defining the
gross data rate as 100 %, the overheads are summarized in Table 5-4.

5.2.2 Up- and Down-link Multiple Access Scheme

This section gives an overview of the above-introduced frame structure and OFDM
symbol configurations, which make up the multiple access scheme. TDMA and
FDMA components are combined, as indicated in Figure 5-4. TDD is proposed as a
duplexing scheme. A fixed frame structure of 69 (in mode I; mode II: 37) OFDM sym-
bols results, with a flexible boundary between the up- and the down-link. The FDMA
component (sub-bands) is suggested in order to simultaneously allow mobile terminals
that are using only one quarter of the full system bandwidth the quarter-rate users.
The BS operates in mode I or II, depending on the present environment (see Section
5.2.1.1).
In the beginning of each frame, a training symbol (TS) (a unique OFDM symbol) is
broadcast on the down-link, which is used for time- and frequency synchronization
146 Chapter 5 OFDM System Proposal and Emulation System

TDMA-TDD/FDMA-OFDM Multiple Access Scheme:

Mode I: Mode II:
OFDM Symbol Nr. Symbol Nr.
68 36
TDD - guard symbol 69 37
TS: Synchronization and channel est. 1 1
SIG 1 SIG 2 SIG 3 SIG 4 2 Downlink signalling 2
3 3
Mode I-H (or II-L); full rate; 1 ATM cell
4 4
5 5
6 6
Mode I-L; full rate; 1 ATM cell
7 7
8 8
Downlink

9 9
Mode I-LL; 10 10
Mode I-H
Mode I-L; quarter 11 11
(or II-L);
quarter rate; 1/4 12 12
quarter
rate; 1/2 ATM cell 13 13
rate; 1
ATM cell 14 14
ATM cell
M I-H; qr; 15 15
1/4 cell 16 16
17 17
18
19
20
TDD - guard symbol variable position 21
SIG 1 SIG 2 SIG 3 SIG 4 Uplink signalling 22
23
Mode I-H (or II-L); full rate; 1 ATM cell
24
Mode I-H Mode I-L; 25
(or II-L); quarter 26
Mode I-H
Mode I-L; quarter r.; rate; 1/4 27
(or II-L);
Uplink

quarter 1/2 cell ATM cell 28

quarter
rate; 1/2 29
rate; 1
ATM cell 30
ATM cell
31
32
frequenc 33
34
67 35
68 36
TDD - guard symbol 69 37
TS: Synchronization and channel est. 1 1
SIG 1 SIG 2 SIG 3 SIG 4 2 2
3 3
time
Figure 5-4: The proposed TDMA-TDD/FDMA-OFDM multiple access scheme. Each row
represents one OFDM symbol, which is divided into four sub-bands. Note: Fre-
quency hopping is not shown.

and for channel estimation (see Chapter 6). The TS is followed by four parallel
signaling channels, each in a separate OFDM sub-band. To ensure highest possible
reliability, BPSK modulation and rate -coding is used for all signaling (see Table
5-2).
The consecutive symbols carry the user data on the down-link. The sub-modes -H, -L
and -LL use different modulation and FEC-techniques, namely, QPSK and 16-QAM
modulation, with rate - and rate -coding. Depending on the modulation and coding
schemes, a varying number of OFDM symbols is required for the transmission of one
5.2 OFDM Based System Proposal 147

ATM cell (see Table 5-3). To efficiently map ATM cells on this fixed frame structure,
it is considered to allow the transmission of half or quarter ATM cells per frame as
well.
On the up-link, another four dedicated signaling channels are available. No training
symbols are used there to maximize the spectral efficiency. In stead, the application of
pre-equalization is suggested in order to facilitate coherent detection, which requires a
reciprocal channel. This principle is discussed in detail in Chapter 7.
Super-frame structures should be defined to efficiently implement the mapping of
ATM cells on the given frame-structure using the multiple access schemes described,
using a centralized, scheduled MAC protocol.

5.2.3 Architecture of the Transceivers

5.2.3.1 Base Station (Access Point)

The architecture of the base-band system and RF front-end of the base station is rather
simple. However, the MAC and control sub-layers have rather high complexity, since
the data streams of multiple MTs have to be multiplexed in the BS, requiring ARQ,
power control, signaling, and other functions. And they have to operate at very high
speed. Using pilots and/or blind techniques, transmission delays and carrier phase off-
sets have to be estimated for each user (see Section 7.5.2).
A block diagram of the base-band and RF parts is given in Figure 5-5. Note that none
of the required filter stages are shown for the sake of simplicity.

Base station: pilot symbols, guard interval,

training sequence windowing

coding / symbol multi- IFFT DSP I/Q up-

D/A
decoding, mapping plexer mod conv.
interleav./
deinterl.
du-
LO LO
W-MAC, pilot plexer
W-control, extraction timing
ATM cell
multi- symbol FFT I/Q down-
A/D
plexing detection dmod conv.

Figure 5-5: Architecture of the base-band system and RF front-end of the base stations.

5.2.3.2 Mobile Terminal

The base-band system of the mobile terminals has a somewhat higher complexity, be-
cause synchronization, channel estimation, and pre-equalization algorithms have to be
implemented. The block diagram is depicted in Figure 5-6.
Note the hardware feedback signal for frequency-synchronization. Interesting simplifi-
cations of the synchronization tasks required (timing, frequency, sampling frequency)
148 Chapter 5 OFDM System Proposal and Emulation System

Mobile terminal:
down- I/Q FFT symbol coding /
A/D
conv. dmod detection decoding,
interleav./
timing pilot ext./ channel deinterl.
synchron.
du- sync. chan. est. estimate
LO LO
plexer freq. sync. W-MAC,
W-control,
ATM cell
up- I/Q DSP IFFT s-mapping multi- multi-
D/A
conv. mod pre-equal. plexer plexing

guard interval, pilot symbols

windowing

Figure 5-6: Architecture of the base-band system and RF front-end of the mobile terminals

can be achieved by locking all local oscillators (LO) and the sampling clocks on one
adjustable frequency source. Thereby, synchronizing for one of the frequency-offsets
in hardware (using the feedback signal), all other frequency-offsets are cancelled si-
multaneously, which simplifies the implementation of the proposed OFDM system.
Note that this principle also requires the oscillators of the base station to be locked on
one another to provide fixed ratios among their frequencies.

5.2.4 Forward Error Correction Coding

5.2.4.1 Forward Error Correction Coding

Standard convolutional coding schemes with soft decision Viterbi decoding can be ap-
plied. Possible parameters for the encoder and decoder are a coding rate of and a
constraint length of 7. The rate coding used in modes -L and -H (see Table 5-3) can
be derived from the rate code by puncturing.
Note that the same codes are used in the IEEE 802.11 and in the HIPERLAN/2 wire-
less LAN standards [1], [4]. This was the only reason for selecting these codes for our
proposal. No research towards optimization of the coding scheme was performed.
Error rate results for the coding and interleaving schemes are presented in Chapter 8.

5.2.4.2 Interleaving Schemes

Bit-level interleaving is used to break-up burst errors into, ideally independent, well-
distributed errors, which can be corrected more efficiently by conventional coding
schemes. Because of the frequency-selective fading channel experienced by the
OFDM system, errors usually result from the sub-carriers being attenuated by the
channel. This often results in burst errors the application of an interleaver is thus
necessary.
Block interleavers are proposed for the investigated transmission system. The convo-
lutionally encoded data bits to be transmitted are written column-wise into a rectangu-
lar array of k columns and n rows, where k denotes the degree (or depth) of the inter-
5.2 OFDM Based System Proposal 149

Interleaver 1 (degree k = 4): interleaver 2 (degree k = 3):

block structure: bit- sub-carrier

number number
read matrix read matrix
1 2 3 4
write 1 6 P 16 write 11 92 P3
matrix 25 76 117 178 matrix 24 10 5
166
39 810 1211 Z12 n 37 11 8
179
413 P14 1315 1816 410 12 11
Z12
517 918 1419 1920 513 P 14
1817
P21 1022 1523 P: pilot SC 616 13 17
1918
Z: zero SC 719 1420 P21
k 822 1523

spreading of coded bits over sub-carriers (numbers indicate bit-numbers of coded bits):

IL 1: SC 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
bit/sym
b 1, s 1 1 6 16 2 7 11 17 3 8 12 4 13 18 5 9 14 19 10 15
pilot

pilot

pilot
20 25 35 21 26 30 36 22 27 31 zero 23 32 37 24 28 33 38 29 34
b 2, s 1
b 1, s 2 39 44 40 41 42 43

f
k=4
IL 2: SC 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
bit/sym
b 1, s 1 1 9 2 10 16 3 11 17 4 12 5 18 6 13 19 7 14 8 15
pilot

pilot

pilot
zero

b 2, s 1 20 28 21 29 35 22 30 36 23 31 24 37 25 32 38 26 33 27 34
b 1, s 2 39 40 41 42 43 44

f
k=3

Figure 5-7: Interleaving schemes for transmission mode I-LL, quarter rate.

leaver. The bits to be modulated on the OFDM sub-carriers are read from this array
row-wise. Two consecutive bits of the coded sequence are therefore separated by k
sub-carriers in an OFDM symbol. Note that only one bit of the QPSK or 16-QAM
constellations is defined at a time, because the whole symbol undergoes the same fad-
ing, implying that all its bits have the same error probability. Thus defining a whole
data symbol (sub-carrier) at once would counteract the goal of breaking up error
bursts.
Not only the two or four bits transmitted (in one QPSK or 16-QAM symbol) over one
sub-carrier undergo the same fading, but, because we consider slow time variability of
the channel, sub-carriers of consecutive OFDM symbols are also affected in the same
way. Therefore, the interleaving must be done in the frequency-domain, i.e., across the
sub-carriers of the OFDM symbols. The block-size of the interleaver is determined by
the number of sub-carriers of the OFDM scheme. Error correction coding is done over
a whole ATM cell, which is carried by a number of OFDM symbols as indicated in
Table 5-3. This requires that the bits are periodically written into the interleaving ma-
trix and read from it in order to be modulated on different bits of the (QPSK or 16-
QAM) data constellations and on consecutive OFDM symbols. For instance, in trans-
mission mode I-LL quarter rate, 24 OFDM symbols are required to transmit the 912
150 Chapter 5 OFDM System Proposal and Emulation System

interleaver for mode II, quarter rate:

read matrix (SC. nr.)
write 11 92 P3 244 305
matrix 26 107 P8
(bit nr.)
311 1112
416 1217 P19
521 1322 23
Z
P26 1427
631 1532
736 : P39 3740 P: pilot SC
841 : P44 3845 Z: zero SC

interleaver for mode I, full rate: interleaver for mode II, full rate:
read matrix (SC. nr.) read matrix (SC. nr.)
write P write P P
matrix Z P matrix P Z P
(bit nr.) P Z P (bit nr.) P P Z Z Z P
Z P P
P P Z P
Z P n= P P Z Z Z P n=
11
Z P P P Z 12
P Z P P
P P Z Z Z P P
Z P Z PP
P P
P
k=9
k = 17

Figure 5-8: Interleaving schemes for transmission modes I; full rate, and II; full and quarter
rate.

bits of one whole, rate- coded ATM cell.

Time-interleaving is not considered in order to minimize throughput delays and mem-
ory buffers. However, when frequency hopping is implemented in the quarter rate
mode, time-interleaving over all OFDM symbols on which an ATM cell is mapped
should be added in order to distribute the coded bits more evenly over the frequency-
band.

For the quarter rate option of transmission mode I (using QPSK and rate- coding),
two possible interleaving schemes are depicted in Figure 5-7. One of degree 4 (inter-
leaver (IL) 1), the other one of degree 3 (IL 2). Next to the rectangular block-inter-
leaving matrix, the translation of coded bits to sub-carriers and OFDM symbols is il-
lustrated. Pilot and zero sub-carriers are included in the interleaving scheme, so that,
using IL 2, subsequent coded bits are modulated alternately on odd and even sub-carri-
ers. (To obtain this, the degree of the interleaver must be odd.) This property is desir-
able with the transmitter diversity schemes presented in Section 8.3.
For the other transmission modes, possible interleaver structures and pilot/zero carrier
assignments are depicted in Figure 5-8. For mode II quarter rate (-qr), mode I full rate
5.2 OFDM Based System Proposal 151

(-fr), and mode II-fr, interleavers of depth 5, 9, and 17, respectively, are suggested.
Obviously, adjacent coded bits are spread more in the full rate modes, thus we expect
better performance from these schemes. Theoretical and simulation results are given in
Section 8.2. To improve the performance of the quarter rate modes, it is suggested to
implement frequency hopping across the four quarter-rate sub-bands or antenna diver-
sity techniques. Some proposals for computationally efficient diversity schemes are
described in Section 8.3.

5.2.5 Link Budget

A link budget for the 60 GHz frequency band is given in Table 5-5, assuming that
50 mW transmit power are available. Full and quarter rate terminals are considered in
the table. Figures in Italics identify parameter values used in calculations, when pa-
rameter ranges are specified.
It is evident from the table that this link budget is rather critical. No fading margins are
considered for the shadowing, for instance, and a path loss exponent of two was as-
sumed. Several investigations have shown that shadowing is a very critical problem in
60 GHz communications. The leaves of a tree, for example, or the user himself may
completely block a 60 GHz link [5]. The conclusion is that a line of sight (LOS) is
needed in most cases.
Note that a LOS reduces or eliminates shadowing. And it results in fading distributions
with higher Ricean K-factors, which means that the fades are shallow compared with
Rayleigh fading channels. Both effects do allow for operation lower average SNR. A
loss of the LOS most likely results in dropouts, however.
Multiple transmit/receive antennas within a room could counteract this problem, just
like multiple light sources are placed within a room to achieve a more uniform light
distribution and to reduce shadows (see e.g. [6]). Radio-over-fiber links were sug-
gested to connect the RF front-ends to the base station over considerable distances.
A slightly different (and somewhat more promising) situation is indicated by the chan-
nel measurements conducted by Smulders [7]. With and without LOS, Smulders re-
ports values of normalized received power (transmission loss including the antenna
gains) between 85 and 70 dB for rooms of different sizes, and for distances up to
about 30 m. The path loss exponents observed are below one, due to the antenna de-
sign used (cf. Section 2.5). This would allow for a good coverage within the relatively
large rooms investigated.
It is concluded that the system is feasible, however, only for limited ranges (maximum
30100 m).
152 Chapter 5 OFDM System Proposal and Emulation System

Table 5-5: Link budget at 60 GHz for full and quarter rate users

parameter quarter rate full rate comments

bandwidth ~28 MHz ~100 MHz spectrum down by 20 dB
carrier frequency 60 GHz
FFT points 32 (64) 128 (256)
in TX mode I (II)
used sub-carriers 19 (38) 88 (176)
Power budget:
transmit power Ptx 17 dBm i.e. 50 mW
TX power back-off 6 10 dB Because of signal dynamics
HPA power 23 27 dBm TX power plus back-off
antenna gains gtx,rx 3 6 dB omnidirectional antennas
free space prop. loss calculated as 20log(4d/) at 60
88, 98, 108 dB at 10, 30, 100 m
ap at distance d GHz ( = 5 mm)
other losses atx ~2 dB connectors, cables, etc.
received power Prx -67, -77, -87 dBm at 10, 30, 100 m Ptx + gtx + grx atx ap
Noise budget:
noise bandwidth BN 22.5 MHz 90 MHz 88 (176) sub-carriers of 1 (0.5)
MHz
noise figure F 5 dB assumed value
equiv. noise temp. T 627 K T = (F 1)T0
noise density N0 -170 dBm/Hz N0 = Tk
noise power PN -97 dBm -91 dBm PN = N0BN
SNR 30, 20, 10 dB 24, 14, 4 dB Prx/PN (at 10, 30, 100 m)
Requirement:
frame error rate max. 10-2 with Rc = , = 5 coding,
required Eb/N0 min. 12 dB QPSK modulation; in Rayleigh
fading channels (see Chapter 8)
required Ec/N0 min. 9 dB
guard interval aGI 1 dB
implementation loss 3 5 dB phase noise, non-linearity, chan-
aimp nel estim., etc. (estimated)
required SNR approx. 15 dB Ec/No aGI aimp

5.3 The Emulation System

A main topic of this Ph.D. thesis is the development of digital signal processing (DSP)
algorithms for OFDM-based wide-band air-interfaces. The final goal of this research
and development effort is the demonstration of the proposed air-interface techniques
on a realistic hardware platform.
5.3 The Emulation System 153

OFDM air-interface emulator: host PC

RS232 frequ.
f0 f 0
sync.
audio audio
host PC transmitter I/Q channel I/Q I/Q receiver host PC
RS232 mod. dmod. RS232
(data source) DSP board 1 (analog) DSP board 2 DSP board 3 (data sink)
I/Q IF I/Q

Provide coded Interleaving;

channel sim. AGC, frequency Decoding;
data blocks symbol mapping; Time-variant trans- I/Q modulation and and time synch.; bit and packet
pilot insertion; versal filter for demodulation to FFT; pilot extrac- error rate
IFFT; guard multipath channel simulate frequency tion; channel est.; estimation
interv.; windowing; simulation; channel synchronization; symbol de-map-
training sequence noise phase noise; etc. ping; de-interleav.

Figure 5-9: Architecture of the emulation system.

Unfortunately, the technical specifications of the investigated systems are so demand-

ing that an implementation is almost impossible for a small research team, as real-time
DSP of two data streams (the in-phase and quadrature (I/Q) components of the com-
plex base-band signal) is required at sampling frequencies of ~100 MHz.
A drastically downscaled hardware platform is therefore presented (Figure 5-9) that
enables the assessment of real-time DSP for such applications at low cost. Moreover,
(software) implementation difficulties are largely avoided due to the reduced speed.
The transmitter and receiver are each implemented on separate audio frequency (sam-

Figure 5-10: Photo of the emulation system. The oscilloscope shows the time-domain OFDM
signal and its spectrum with a simulated channel. Note that the spectrum is calcu-
lated in real-time by the receiver-DSP (DSP board 3). The Korea-Telecom (KT)-
version is depicted.
154 Chapter 5 OFDM System Proposal and Emulation System

pling at 48 kHz) DSP boards. A third DSP board is used for simulating the multipath-
fading channel, which would behave very differently at the largely reduced bandwidth.
Finally, I/Q-modulation and -demodulation are emulated in analog hardware, for the
realistic evaluation of frequency synchronization algorithms, a main issue in OFDM.
A hardware feedback signal enables the tuning of the demodulators local oscillator, to
achieve frequency-synchronization. DSP software development was partly performed
in the framework of student projects [8][11].
In this section, the hardware set-up of the emulation system is presented, which has
been integrated in a 19 rack, as depicted in Figure 5-10. (Note that two copies of the
emulation system have been built. One copy (the KT-version) has been delivered to
Korea Telecom, while the second one (the TUD-version) is at Delft University of
Technology.)
The main components of the emulator and their key specifications are described in
Section 5.3.1. Some relevant measurement results on the hardware characteristics of
the system are presented in Section 5.3.2. The linearity of amplifiers and other compo-
nents in the signal path, the phase noise of mixer-oscillators, and filter characteristics
are of major importance for OFDM systems. Finally, the signal processing for the
time-variant channel simulator is outlined in Section 5.3.3.

5.3.1 Description of the Emulation System

5.3.1.1 Hardware Architecture

The basic architecture of the emulation system for the evaluation of simplex air-inter-
faces is shown in Figure 5-9. This figure also indicates the functionality implemented
on the different hardware sub-systems.
The DSP platforms used are SHARC EZ-KIT Lite evaluation kits with the Analog De-
vices ADSP-21061 floating-point processor [12], which is a 32-bit DSP with 1 Mbit of
internal RAM, running at 40 MHz. The boards are connected to a Windows based PC,
which is used for software development and for program control at the time of execu-
tion. Development tools and an extensive documentation library are provided with the
DSP boards. The tools include C-language and assembler compilers and applications
for downloading and uploading programs and data from the DSP.
The ADSP evaluation boards contain 16-bit stereo analogue input and output ports
with a maximum sampling frequency of 48 kHz. Further specifications of the DSP
boards are given in Table 5-6.

Two MAXIM MAX2450 Evaluation kits are used for I/Q-modulation and -demodula-
tion. Each of these boards contains an I/Q-modulator, an I/Q-demodulator, and a volt-
age controlled oscillator (VCO). The intermediate frequency (IF) is approx. 70 MHz.
Crystal oscillators (XO) have been integrated for stabilizing the drifting VCOs of the
MAXIM boards, and for reducing phase noise. The XOs are based on 16.934 MHz
5.3 The Emulation System 155

Table 5-6: Main specifications of the DSP boards

Parameter Value
CPU clock rate 40 MHz
Maximum sampling frequency fs 48 kHz
DAC/ADC channels 2 + 2 (stereo, used for I/Q)
Full-scale input voltage 2.8 Vpp
Full-scale output voltage 2 Vpp
Pass-band of ADCs and DACs 0.4fs
Pass-band ripple 0.1 dB
Stop-band 0.6fs
Stop-band rejection 74 dB
DC offset (max.) 55 LSB

crystals driven by standard 74HCU04 CMOS inverters. (We note that 18 MHz crystals
were used in the TUD-version of the emulation system.) The VCOs on the MAXIM
boards, operating at twice the IF-frequency, are injection locked on the eighth har-
monic of the XOs.
The XO of modulator 2 can be voltage controlled (VCXO) in order to allow the
evaluation of frequency-synchronization algorithms. A control voltage is generated by
DSP 3 the receiver (see Figure 5-9) via the mechanism outlined in the following
paragraph. A block diagram of the I/Q-modulator and -demodulator boards is shown in

I/Q Modulator 1 I/Q Modulator 2

BNC BNC
I/Q- Atten. Atten. I/Q-
I/Q Input BNC BNC I/Q Input
mod. 20 dB 20 dB mod.

BNC BNC
I/Q- BPF BPF I/Q-
I/Q Output BNC 70 MHz 70 MHz BNC I/Q Output
dmod. dmod.

fIF/4 out BNC BNC

fIF/4 out
~20 mVpp 8 VCO VCO 8 ~20 mVpp
~140 MHz ~140 MHz
Manual Poti = 2f IF = 2f IF Poti Manual tuning
tuning * * BNC External tuning
VCXO XO VCXO Switch VCXO on/off
Switch
on/off 16.934 MHz 16.934 MHz Poti Manual tuning
= f IF/4 = fIF/4
Switch Manual/Auto
FLAG 2 RS
on front panel from
DSP 3 Flip-Flop
TIMEXP
on front panel
*: The VCO locks on the 8th
harmonic of the (VC)XO internal connections

Figure 5-11: Block diagram of the internal and external wiring of the I/Q-modulator and -de-
modulator boards. (KT-version)
156 Chapter 5 OFDM System Proposal and Emulation System

Table 5-7: Key specifications of the I/Q-modulators and -demodulators

Parameter Value
KT-version TUD-version
intermediate frequency fIF ~6580 MHz
intermediate frequency with (VC)XOs 67.74 MHz 72 MHz
(VC)XO frequency fXO 16.934 MHz 18 MHz
VCXO tuning range fXO 282 Hz fXO 198 Hz
VCXO tuning sensitivity (at fXO) ~0.7 Hz/LSB = ~0.48 Hz/LSB =
~230 Hz/V ~160 Hz/V
max. input voltage swing 1.35 Vpp
max. output voltage swing 1.35 Vpp
base-band attenuation input
output ~3.5 dB
modulator input bandwidth (max.) 15 MHz (~20 kHz used)
demodulator output bandwidth (max.) 9 MHz (~20 kHz used)

Figure 5-11. Specifications are given in Table 5-7.

Generation of the Feedback Signal for Frequency Synchronization
DSP 3 can control the VCXO of the I/Q-demodulator by generating a pulse duration
modulated (PDM) signal. For this purpose, the DSPs output pins FLAG 2 and
TIMEXP are connected to an RS-flip-flop. Periodically, e.g. at the beginning of the
interrupt service routine used for acquiring the analog input samples, a short pulse is
generated on FLAG 2 for setting this flip-flop. At the same time, the DSPs timer is
initialized with a variable value corresponding to the desired tuning voltage (a value
between 0 and 833 at 48 kHz sampling frequency). The timer now counts down from
that value at the CPU clock rate, causing a short pulse on TIMEXP when zero is
reached. This pulse is used for re-setting the flip-flop. The outlined procedure gener-
ates a rectangular signal with a frequency equal to the sampling frequency, and with a
duty cycle between 0 and 100 %. A low-pass filter converts this signal to the VCXO
control voltage in the range of 2.55 V. Note that the CPU is not occupied by this task,
except for setting FLAG 2 and writing the tuning value to a register.

5.3.1.2 Pros and Cons of the Emulation System

Advantages of the proposed emulation system compared to pure computer simulation
are primarily due to the independence of the transmitters and receivers hardware.
This introduces arbitrary (fractional) sampling instant offsets and carrier (and sam-
pling) frequency offsets, just like in a real system. It is rather cumbersome to include
such effects in e.g. a MATLAB simulation. Moreover, the assessment of real-time al-
gorithms is possible, however, at the price of more complex software development. A
pro and con for the emulator is that generally all system components must be present
5.3 The Emulation System 157

in order to allow operation, which makes it more difficult to evaluate system parts
independently but allows for overall assessment. Certain blocks are easily removed, as
for instance the I/Q-(de)modulation hardware, or the channel simulator.
Compared with a real-time (demonstration) system, software development is drasti-
cally simplified because of the low sampling frequency used. So far no serious timing
problems were encountered in the real-time software written for the emulator. Draw-
backs are that the specifications of the analog-to-digital converters (ADC) and digital-
to-analog converters (DAC) (bit resolution, filter characteristics, etc.) are superior to
ADCs and DACs available for real wide-band systems (with sampling frequencies
above 100 MHz). The same applies for filter stages and amplifiers present in the ana-
log front-ends. Therefore, the impact of these system components is not emulated re-
alistically. Some of those effects can be modeled on the DSPs, however, if their influ-
ence is of interest. It is impossible as well to demonstrate the investigated transmission
scheme over real wide-band radio channels.
Finally, the system can be useful in the education of electrical engineers, e.g. for dem-
onstrating modulation techniques. It can be also used in M.Sc. graduation projects, but
the time to get familiar with the system and the skills required for handling real-time
DSP must not be underestimated by the supervisor.

5.3.1.3 Demonstration of the Time-Division Duplex Scheme

The time-division duplex (TDD) scheme can be emulated by applying the alternating
down- and up-link signals to the same channel simulator, as depicted in Figure 5-12.
The two transmitted signals are added up before feeding them into the channel, just
like they are added up on the (linear) radio channel in the real system. Note that during
the transmission time of one of the transceivers, the other one is receiving and not gen-
erating (transmitting) a signal on its own. Therefore, the adder shown in the figure
rather acts as a multiplexer. The AUX 1 (auxiliary) input of the codec of DSP board 1
is used to add the mobiles transmitted signal. This input has a programmable attenua-
tion/gain block for matching signal levels.

AUX in
IQ- IQ-
TRX 1 Channel dmod. mod. TRX 2
(Base, (DSP 2) (Mobile,
DSP 1) IQ- IQ- DSP 3)
mod. dmod.
PDM signal
DSP board 1 generated via
f-sync. timer
LO VCO
: I/Q-signals feedback
I/Q mod. 1 I/Q mod. 2

Figure 5-12: Architecture of the emulation system for evaluating the TDD techniques.
158 Chapter 5 OFDM System Proposal and Emulation System

5.3.2 Hardware Characteristics

The main hardware issues in OFDM transmission systems are the linearity of amplifi-
ers, mixers, and filters, the characteristics of ADCs, DACs, and the phase noise of os-
cillators used for up- and down-conversion of signals and for I/Q-modulation and -de-
modulation. Some of these topics are briefly addressed in this section.

5.3.2.1 Linearity and Intermodulation

For testing the possible impact of the emulators hardware, the following set-up is
suggested (see Figure 5-13). DSP-board 1 is used to generate a two-tone complex ex-
ponential signal. This signal is I/Q-modulated and -demodulated, using the I/Q-
modulator and -demodulator of only one of the MAXIM boards, which are both con-
nected to the same local oscillator (LO). There is thus no frequency offset introduced.
Spectrum analysis of the demodulated I/Q-signals is performed using DSP board 3.
(256 points FFT; 32 times averaging of the power spectrum estimates; no windowing.)
FFT-windowing is not required here, because the generated tones are exactly put on
frequency-instants of the analyzing FFT.
One of the observed spectra is depicted in Figure 5-14. Symmetrically to the two
generated tones at 8.25 and 9.75 kHz, a mirror image is seen at about 32 dBc. By cor-
recting for gain mismatch at the receiver it is possible to decrease these side-bands to
about 55 dBc. The gain mismatch is mainly introduced by the I/Q-(de)modulator.
This is confirmed by directly connecting the two DSP boards, where the mirror image
lies at 47 dBc without mismatch correction. DC offsets are the source of the line seen
at 0 Hz, 27 dBc. The other tones are intermodulation products from the mixers. Being
at an excellent 55 dBc and more below the test tones, we conclude that the hardware
is well suited for the implementation of an OFDM system, regarding linearity.

I/Q IF (~70 MHz) I/Q

DSP I/Q -20dB, I/Q DSP
(board 1) modulator BPF demod. (board 3)

LO 1 LO 2 (for phase-noise
measurement)

Figure 5-13: Test set-up for evaluating the non-linearity, harmonic distortions, and phase noise
of the I/Q-modulator/demodulator board.

5.3.2.2 Phase Noise

Another critical parameter for OFDM modems is the phase noise performance of all
mixer oscillators. Phase noise means that a spectral line, like an OFDM sub-carrier,
undergoes some broadening (see Figure 5-15), resulting in inter-carrier-interference
5.3 The Emulation System 159

Power spectrum test9a

0 dB @ 8.25 kHz
20 0.00732 dB @ 9.75 kHz

10

0
signal magnitude [dB]

27.1 dB @ 0 kHz
10 32.4 dB @ 8.25 kHz
32.4 dB @ 9.75 kHz

20

30 55.5 dB @ 6.75 kHz

56 dB @ 11.25 kHz
56.3 dB @ 1.5 kHz
57.3 dB @ 1.5 kHz
40

50
20 15 10 5 0 5 10 15 20
frequency [kHz]

Figure 5-14: Observed spectrum at receiving DSP board. A two-tone signal was generated, IQ-
modulated, and -demodulated.

(ICI).
As in the previous sub-section, this problem was assessed by spectral analysis on the
demodulated I/Q-signals, using DSP board 3, where, this time, Blackman windowing
was applied. The one-tone test signal was generated by DSP board 1, to be I/Q-modu-
lated and -demodulated using different modulator boards with independent mixer os-

reduction of phase noise using chrystal oscillators

0 dB @ 12 kHz 0 dB @ 15 kHz
0 original MAXIM VCOs
with chrystal stabilization
10
power spectrum magnitude [dBc]

20

30 31.1 dB @ 12 kHz

37.9 dB @ 0 kHz
40 39.2 dB @ 15 kHz 41.9 dB @ 0 kHz

50

60

70

80
20 15 10 5 0 5 10 15 20
frequency [kHz]

Figure 5-15: Broadening of a spectral line due to phase noise.

160 Chapter 5 OFDM System Proposal and Emulation System

cillators. Two measurements are shown. The first one is for the original VCOs in-
cluded on the MAXIM boards, where a 15 kHz test-signal was used. The noise floor
rising towards this tone is caused by the phase noise. The two smaller tones, being
again caused by DC-offsets and I/Q-mismatch, are not of interest for this test. A 50 Hz
frequency modulation (interference from the mains) with a frequency deviation of
30 Hz was identified as one main component of the phase noise. Comparing these
30 Hz to the sub-carrier spacing of 375 Hz, it was decided to stabilize the VCOs us-
ing crystal oscillators, as mentioned above.
The second measurement, showing a test-tone at 12 kHz, includes these crystal os-
cillators. The noise floor observed in this case is due to the quantization noise of the
ADCs and DACs, and due to other noise sources in the signal path, but not due to the
phase-jitter of the local oscillators. Another major advantage of the crystal oscillators
is their low frequency drift, which drastically simplifies the implementation of real-
time synchronization algorithms.

Performance evaluations of OFDM system components and bit error rates are dis-
cussed in Chapters 6 and 7.

5.3.3 Implementation of the Channel Simulator

This section describes the implementation of the real-time, time-variant channel
simulator for the emulation system.
The real-time channel simulator is based on a transversal filter (FIR-filter) whose coef-
ficients are gradually changed in order to realize time-variability. The channel filter is
written
L 1
rk = hIR ,i ,k s ( k i ) , (5-1)
i =0

where {sk} and {rk} are the input and output samples, respectively, and {hIR,i,k} are the
tap weights at delay tap i and time index k. The number of taps is written L.
The FIR filter has complex-valued tap-weights and complex in- and outputs, since the
emulation system operates with complex baseband (lowpass equivalent) signals. This
requires four real multiplications and four additions/subtractions per FIR-tap. The
whole filter is calculated on each (complex) input sample leading to a rather high
computational load for the processor.
An L = 30-tap FIR-filter is used in the channel simulator, requiring approximately
19 % of the processors capacity at 48 kHz sampling frequency. Another 26 % of its
processing power are required for generating a complex Gaussian noise sample (addi-
tive channel noise), scaling and adding it to the (filtered and scaled) signal, and for
serving the interrupt service routine (ISR), in which all these tasks are performed. The
coefficients of the FIR-filter are generated and updated in the main-loop of the DSP
program, employing the remaining computational power.
5.3 The Emulation System 161

5.3.3.1 Calculation of the Complex Tap-Weights

Power Delay Profile

The complex tap-weights of the FIR-filter are generated according to complex Gaus-
sian random processes. They have means of zero, except at delay time zero, where a
constant is added in order to enable the simulation of Ricean fading channels. The
variances {i2 = var(hIR,i,k)} of the Gaussian random variables (RV) is defined by an
exponentially decaying, average power delay profile (PDP). They relate to the delay
power spectrum (DPS) which is the mathematical description of the frequency do-
main (FD) channel model proposed in Chapter 2 as shown in the following deriva-
tion.
The parameters {, , } define the DPS for the case of an exponentially decaying
channel response, written as

0 <0

h ( ) = 2 ( ) = 0, (5-2)
e >0

where is the excess delay-time variable. The relations of these parameters to the
channel parameters (the normalized received power, RMS delay spread, and Ricean K-
factor) can be found in Section 2.3.3.
Integration of the DPS over the delay bins corresponding to the sampling interval
yields a good approximation for the desired variances {i2} needed to implement the
FD-model by means of a transversal filter. They are obtained as
( i +1 / 2 ) Ts ( i +1 / 2 )Ts
[
1 e s /2
T
] i =0
h ( )d = e

i2 = d =
= ( i 1 / 2 )Ts = ( i 1 / 2 )Ts

[
e (i 1/ 2 )Ts e (i +1 / 2)Ts ] i >0
, (5-3)
>0 >0

where Ts is the spacing of the delay taps that is equal to the sampling period.
Time-Variant Channel Simulation
To simulate the time-variability, the complex FIR-coefficients need to be updated
continuously. It was mentioned in the previous section that each coefficient of the FIR-
filter can be considered as a sample of a complex Gaussian process with variance
{i2}. Appropriate time-variability is yielded when the power spectrum of these com-
plex Gaussian processes are shaped (colored) according to the Doppler spectrum of the
radio channel to be simulated.
A very popular model of a Doppler spectrum is the U-shaped spectrum resulting from
a uniform angular distribution of incident waves at a receiver moving with constant
velocity. It is often called Jakes fading spectrum [2]. A well known and efficient
method of generating such colored Gaussian processes is based on Rices sum of sinu-
soids [13]. The stochastic process is thereby approximated by a finite sum of appropri-
162 Chapter 5 OFDM System Proposal and Emulation System

ately designed sinusoids with constant frequencies. A block diagram of such a model
is found in [14]. Adapted to our application it is written

hIR,i ,k = i ci [cos(2f i ,l ,1kTs + i ,l ,1 ) + j cos(2f i ,l , 2 kTs + i ,l , 2 )]

Mi
i >0
l =1
, (5-4)
hIR, 0,k = 0 c0 [cos(2f 0,l ,1kTs + 0,l ,1 ) + j cos(2f 0,l , 2 kTs + 0,l , 2 )]+ e
M0
j ( 2f + )

l =1

where ci are normalizing constants in order to obtain variance of one from the sums of
sinusoids, and Mi are the number of sinusoids added up at each delay tap. The sinu-
soids parameters are their frequencies {fi,l,n} and their initial phases {i,l,n}, with i be-
ing the tap index, l being the index of the sinusoid, and n = {1,2} standing for the real
or imaginary component. f and are the frequency and initial phase of the LOS path,
respectively.
Methods of finding parameters for the sinusoids are discussed in [14] and in the refer-
ences therein. Also Jakes [2] describes a fading simulator based on this approach.
In our implementation, the taps are updated (one at a time) as quickly as the DSP can
calculate the new sums-of-sinusoids for the current time-instant. Mi = 10 pairs of sinu-
soids are added up per delay tap. Updating one tap lasts for about 80 s; the whole IR
(with 30 FIR-taps) is re-calculated within approx. 2.4 ms (rate of ~420 Hz), i.e. ap-
proximately 120 sample periods at 48 kHz. This is sufficiently frequent since the fad-
ing is rather slow compared to this time period. The maximum Doppler frequency of a
typical indoor channel is 400 Hz (2 m/s movement) at 60 GHz, which scales down to
only 0.15 Hz in the emulation system. Even a movement at 100 kph (28 m/s) can be
easily realized, which corresponds to a maximum (downscaled) Doppler frequency of
approx. 2 Hz.
Calculation of the Sinusoids Parameters
A pure Monte Carlo approach was followed to determine the frequencies and initial
phases of the summed sinusoids. No negative frequencies are considered for {fi,l,n},
because the real- and imaginary sums-of-sinusoids are independently generated (see
eq. (5-4)). Uniform random variables {i,l,n} [/2, /2] describe the angles of arri-
val of the l-th path of the i-th delay tap. Applying these variables to fi,l,n = fmcos(i,l,n)
yields appropriately distributed frequencies, where fm is the maximum Doppler fre-
quency. (fm = fcvm/c, where fc is the carrier frequency, vm is the mobiles velocity, and
c is the speed of light).
The initial phases of the sinusoids are selected from a uniform distribution between
and .
5.4 Summary and Conclusions 163

5.4 Summary and Conclusions

In the first part of this chapter, a novel OFDM based air-interface and multiple access
scheme are proposed for a wireless ATM communications system. Transmission rates
up to 155 Mbit/s can be supported in slowly time-variant (indoor) channels at 60 GHz.
The maximum cell radius is in the order of 30 m, under optimum conditions 100 m. A
line-of-sight (LOS) to the base station is needed to obtain the maximum range, while
more robust coding and modulation techniques have to be used, reducing the transmis-
sion rate.
A relatively low number of active carriers (88 in mode I) and a short FFT block-length
(128) are selected in order to keep the complexity of the system low and to limit the
carrier synchronization and peak-to-average power problems. Training symbols are pe-
riodically broadcast by the BS to allow the MTs to synchronize efficiently and accu-
rately and to calculate a channel estimate. A flexible time division duplexing scheme is
proposed to switch between the up- and the down-link, allowing for asymmetric data
rates. Pre-equalization is utilized on the up-link to minimize the required overhead for
training symbols. This enables the efficient transmission of short data packets (single
ATM cells).
Several transmission modes, which apply different modulation and coding schemes,
are suggested for different channel conditions, leading to various data rates. This
adaptability also allows fall-back modes when shortcomings occur in the link budget.
Currently, combinations of QPSK and 16-QAM modulation schemes and rate- and -
coding rates are considered. It remains to be evaluated, however, if the proposed
combinations are the optimum ones. This is a topic for further research.
Another feature of the proposed system is the simultaneous support of two different
terminal bandwidths (full bandwidth equal ~ 100 MHz and quarter bandwidth) by the
BS through the use of a combined TDMA/FDMA-OFDM multiple access scheme.
These so-called quarter-rate terminals have largely simplified hardware specifica-
tions.
The algorithm development of the main functions of the proposed air-interface is de-
tailed in consecutive chapters of this thesis. The performance analysis of many aspects
is given there. Purpose of these investigations is the evaluation of the suitability of the
proposed techniques for the intended application, i.e., for wireless indoor LANs and
for short-range, low-mobility outdoor systems.

The main design targets for the proposed OFDM air-interface were efficiency, sim-
plicity, and flexibility in order to realize the extreme data rates considered. The main
drawback of the proposed design is its limited range and the low data rate at long
range. Methods for enhancing the performance and extending the range include:
Diversity techniques
Adaptive antennas for beam forming and/or multi-user detection
164 Chapter 5 OFDM System Proposal and Emulation System

Turbo coding
Space-time coding
Note that some of these techniques have to be embedded in a system proposal (for in-
stance coding schemes), while others can be implemented on top of an existing pro-
posal/standard (e.g., beam forming and diversity techniques), although adaptations
may be desirable.
The purpose of advanced techniques like multi-user detection is to enhance the system
capacity by increasing the possible number of users and by mitigating interference ef-
fects from adjacent cells.

The strictly hierarchical structure of the proposed system, with base stations and syn-
chronized terminals, may be a significant disadvantage in the implementation of con-
cepts like ad-hoc networking. On the other hand, exactly this feature could make the
system ideal for applications where multiple deterministic high-rate data streams have
to be managed. A wireless TV-studio is a good example for such a scenario.

The second part of this chapter presents the hardware platform designed for the dem-
onstration of the OFDM air-interface and its signal-processing algorithms. A largely
downscaled DSP-based system, the so-called emulator, was built, where the OFDM
transmitter and receiver are realized on separate DSP-boards. A third board is used for
channel simulation. Analog hardware implements I/Q-modulation and -demodulation,
leading to carrier frequency-offsets and enabling thereby the demonstration of real-
time synchronization algorithms. System components are connected by analog (I/Q)
baseband signals.
Main advantages of the emulator are its low hardware cost and the largely simplified
software development compared with the real-system, due to the reduced speed. Sam-
pling-instant and sampling- and carrier-frequency offsets are present between the
transmitter and the receiver, which are hard to implement in a pure computer simula-
tion. Moreover, the interaction of several system components is realistically modeled.
Note that this may also be a disadvantage as it makes isolating the influence of a single
effect more difficult.
The hardware behavior is similar to a real wide-band system, although the specifica-
tions of certain hardware components are superior in the emulation system, because of
the reduced bandwidth of ~ 40 kHz (compared to ~ 100 MHz in the proposed OFDM
system). It was concluded from brief studies of inter-modulation distortion, linearity,
and phase noise, that the realized system is suitable for the demonstration of OFDM
systems. This is confirmed by performance results given in Section 6.5.
5.5 References 165

5.5 References
[1] R. van Nee and R. Prasad, OFDM for Wireless Multimedia Communications:
Artech House, 2000.
[2] W. C. Jakes Jr., Microwave Mobile Communications. New York: Wiley-Inter-
science, 1974.
[3] D. Petras, A. Krmling, and A. Hettich, Design Principles for a MAC Protocol
of an ATM Air Interface, in Proc. ACTS Mobile Communications Summit, Gra-
nada, Spain, Nov. 1996, pp. 639644.
[4] R. van Nee, G. Awater, M. Morikura, H. Takanashi, M. Webster, and K. W. Hal-
ford, New High-Rate Wireless LAN Standards, IEEE Commun. Mag., pp. 82
88, Dec. 1999.
[5] M. Flament, On 60 GHz Wireless Communication Systems. PhD Thesis,
Chalmers Univ. of Techn., Gteborg, Sweden, 2000.
[6] S. Okamura, M. Okada, and S. Komaki, Impact of Ubiquitous Antennas to the
th
Interference Cancellation of COFDM Systems, in Proc. 6 international
OFDM-Workshop (InOWo) 2001, Hamburg, Germany, Sept. 2001, pp. 2-12-4.
[7] P. F. M. Smulders, Broadband Wireless LANs: A Feasibility Study. PhD Thesis,
Eindhoven University of Technology, Eindhoven, The Netherlands, 1995.
[8] P. Teneva, Pseudo Real-time Simulation of an OFDM System for Wireless
Broadband Communications: OFDM Transmitter, Research Report, IRCTR,
Delft Univ. of Tech., Jan. 1999.
[9] K. Bke , DSP Manual, Research Report, IRCTR S-025-99, Delft Univ. of
Techn., July 1999.
[10] K. Bke, Assessment of OFDM Based Air-interface Techniques Using an Emu-
lation Platform: Investigation and Implementation of OFDM Synchronization
Algorithms, M.Sc. Thesis, IRCTR S-001-00, Delft Univ. of Tech., Jan. 2000.
[11] I. Gultekin, DSP Software Implementation for a Broadband Air-interface Emu-
lation Platform: Onderzoek naar de serile interace tussen de PC an de DSP
board en het maken van een GUI voor de DSP applicaties, Graduation Thesis,
Haagse Hogeschool, June 2000 (in Dutch).
TM
[12] ADSP-21061 EZ-KIT Lite Reference Manual. Norwood, MA: Analog Devices
Inc., 1997, http://www.analog.com/support/dsp/documentation/Ezklrm.pdf.
[13] S. O. Rice, Mathematical Analysis of Random Noise, Bell Syst. Tech. J., vol.
23, pp. 282332, July 1944; vol. 24, pp. 46156, Jan. 1945.
166 Chapter 5 OFDM System Proposal and Emulation System

[14] M. Ptzold, U. Killat, F. Laue, and Y. Li, On the statistical properties of

deterministic simulation models for mobile fading channels, IEEE Trans. Veh.
Technol., vol. 47, no. 1, pp. 254269, Feb. 1998.
Chapter 6 DSP Algorithm Development
for the Down-Link

6.1 Introduction
The purpose of this chapter is to describe the key algorithms for down-link data trans-
mission, which are required for the OFDM-based communications system proposed in
Section 5.2.
Two main tasks have to be performed: synchronization and channel estimation. Syn-
chronization steps include, respectively, timing and frequency synchronization. That is
the correct start-time of the OFDM symbols has to be found in order to perform de-
modulation using the FFT, and the transmitters and receivers local-oscillator-fre-
quencies have to be matched so that the sub-carriers data symbols are obtained at the
FFTs output. Estimation algorithms are proposed and evaluated for these synchroni-
zation tasks. Channel estimation is performed to determine the transfer function of the
frequency-selective radio channel, which is required to equalize and detect the data
symbols. Most of the techniques described are based on a special training symbol,
which is periodically transmitted on the down-link (see Section 5.2).

In Section 6.2, we give an overview of the synchronization steps required, and we

evaluate their performance for the proposed OFDM system. We also introduce signal-
processing steps that must be utilized for coherently detecting the transmitted data.
The general conclusion is that the investigated methods enable sufficiently accurate
synchronization, with only one OFDM symbol used as a training symbol per frame.
The overhead of this symbol is therefore well invested.
An additional synchronization step is applied for enhancing the initial timing-synchro-
nization. With the training symbol proposed, the computational effort for this estima-
tion step becomes extremely low, while its accuracy is very high. Novel work on this
167
168 Chapter 6 DSP Algorithm Development for the Down-Link

estimation technique is presented in Appendix C of this thesis, where the estimation

bias and the standard deviation are analyzed. The bias is derived for Ricean multipath
channels (including Rayleigh and AWGN). It is given in terms of channel parameters,
namely, the RMS delay spread and the Ricean K-factor. Moreover, the standard devia-
tion is studied for AWGN and Rayleigh channels. The results allow for some optimi-
zation of the estimation technique.

For the final validation of the synchronization algorithms, OFDM modems have been
implemented on a digital signal processor (DSP) based experimental platform, the so-
called emulation system (see Section 5.3). The interaction of all synchronization
tasks and the impact of remaining offsets on the bit-error-rate were demonstrated. It
was observed from these results that a very popular class of algorithms for frequency-
synchronization in OFDM systems suffers severely from DC-offsets and from carrier
feed-through hardware impairments that are hard to suppress. The work of Section
6.3 provides new insights into these issues. The influence of these factors is evaluated,
leading to a novel, enhanced estimation method capable of eliminating most of the de-
gradations.

Using the known training symbol, which has been utilized for synchronization pur-
poses as well, channel estimation is a straightforward task. The channel transfer func-
tion is determined in the beginning of each frame by removing the known data from
the received signal constellations. The obtained channel estimate is then employed to
detect the data of the whole down-link frame. Therefore, a slowly time-variant channel
has to be assumed, i.e., the channel should be quasi-static during the transmission of
one frame.
In Section 6.4, we investigate if a short linear filter with fixed coefficients can reduce
the mean-square-error of the channel estimate. The filter is to be applied in the fre-
quency-domain in order to smoothen the estimated channel transfer function. Wie-
ner filters designed for a given set of channel parameters are evaluated over a number
of different environments. It is seen that a fixed filter designed for the worst-case
channel can enhance the estimate on a quite wide range of channels. The gain
achieved is reduced, however, in cases where the filter design does not match to the
channel. Novel contributions in this section are the description of the channel estima-
tion technique, which efficiently exploits the training symbol, and the selection of a
fixed filter for a representative set of actual channels.

In Section 6.5, experimental performance results are presented, which were measured
with the emulation system described in Section 5.3. The influence of the channels
time-variability is seen from these results. Acceptable performance is achieved at ve-
locities up to about 2 m/s are possible, which is appropriate for indoor environments.
A simple linear extrapolation using the last two channel estimates is briefly assessed
(using simulation results). It proves to enhance the maximum velocity allowed up to
6.2 Overview of Synchronization Steps 169

about 8 m/s.

The sections are summarized separately with conclusions and recommendations.

6.2 Overview of Synchronization Steps

The synchronization scheme developed and analyzed is based on the work by Schmidl
and Cox [1], with a number of modifications and extensions (see e.g. [2], [3]). Accord-
ing to the system proposal, the mobile terminals synchronize to a known training sym-
bol (TS), which is continuously broadcast by the base station. One single training
symbol is sufficient to perform full synchronization (and to calculate the channel esti-
mate). Multiple training symbols can be used in low-SNR situations, in order to en-
hance the robustness.
Synchronization is acquired in the following steps:
Frame and rough timing synchronization (from time-domain signal)
Fractional frequency synchronization (from time-domain signal)
Correction for the fractional frequency offset (mathematically)
Application of the FFT to demodulate the TS
Integer frequency-synchronization (based on the TS data)
Remaining timing-offset synchronization (based on the phase rotation of the TSs
data constellations)
Estimation of and correction for sampling frequency-offsets (based on the above
timing-offset estimates)
Estimation of and correction for carrier phase offsets (based on pilot sub-carriers)

This section starts with a brief classification of synchronization algorithms and gives a
number of references, which describe typical examples (Section 6.2.1). The descrip-
tion of the design of a training symbol is given in Section 6.2.2. We then outline in
Sections 6.2.36.2.9 the synchronization steps and show performance results for the
OFDM system under investigation (see Section 5.2). Although the proposed methods
are studied in the framework of that system proposal, wed like to emphasize that most
of the principles can be applied much more generally.

6.2.1 Classification of Synchronization Techniques

Synchronization principles for OFDM can be characterized in several respects. Certain
algorithms are suitable for initially acquiring coarse synchronization, while others can
only track small errors in a system that has been roughly synchronized before. These
classes are referred as acquisition and tracking algorithms, respectively.
Acquisition algorithms often work directly on the received, time-domain OFDM sig-
170 Chapter 6 DSP Algorithm Development for the Down-Link

nal, exploiting its specific properties. For instance, the correlation between duplicated
parts of this signal (the guard interval (GI) or a special training symbol) can be used
for initial timing- and frequency-synchronization [1], [4], [5]. (The estimation steps of
Sections 6.2.3 and 6.2.4 are in this class of algorithms.) Tracking algorithms typically
exploit the systematic phase rotations of the data constellations, which are induced by
synchronization offsets (cf. Section 4.2.4) [6][11]. This implies that these algorithms
work on the output of the FFT, which requires that coarse synchronization has been
acquired before, because otherwise the signal constellation points cannot be recovered.
(Examples for such techniques are described in Sections 6.2.7 and 6.2.9.1.) Pre- and
post-FFT algorithms can be distinguished.
Although most acquisition algorithms can be used in the tracking phase as well, their
performance may be inferior compared with specific tracking algorithms. A combina-
tion of the above mentioned classes is therefore favorable for solving the synchroniza-
tion problem in OFDM.

6.2.2 Design of the Training Symbol

The design of the training-symbol originates from the paper by Schmidl and Cox [1]
on frequency and time-synchronization for OFDM. The first of the two training sym-
bols (TS) used in [1] is a unique OFDM symbol because every second sub-carrier (SC)
is zero. This implies that the symbol has identical halves in the time-domain, due to
properties of the Fourier transform (see Figure 6-1a).
In the original scheme [1], the odd-numbered sub-carriers are zero, while the even
sub-carriers contain a known, binary pseudo noise (PN)-sequence. Analyzing the sys-
tem proposal of Section 5.2, it is seen that many of the even-numbered sub-carriers
must be zero, because they are used to separate sub-bands and to avoid problems with
DC-offsets and carrier feed-through. In order to apply the synchronization scheme to
this OFDM system, we therefore use the odd-numbered sub-carriers for the PN-se-

GI training symbol GI data symbol (TS 2)

(a)
GI training symbol GI data symbol

(b)
Figure 6-1: Illustration of the training symbol in the time-domain. (a): Training symbol of
Schmidls original method, where the PN-sequence is carried by the even sub-car-
riers. The symbol has identical halves in the time-domain. (b): Modified training
symbol. The PN sequence is modulated on the odd sub-carriers, leading to a time-
domain symbol with identical halves but opposite signs.
6.2 Overview of Synchronization Steps 171

Training symbol
frequency
-47 -45 -43 -3 -1 0 1 3 43 45 47 sub-carrier number
differential demod.
1 1 1 1 1 1 1 47 BPSK symbols

8 sym. length 31 m-sequence 8 sym. cyclically ext. m-seq.

Figure 6-2: Frequency-domain diagram of the training symbol in transmission mode I-fr.

quence (see Figure 6-2). This is a novel modification of the well-known technique.
Still the corresponding time-domain sequence has identical halves, but the samples of
its second halve have opposite signs, which is easily taken into consideration in the es-
timation steps. The training symbols of the original and of the modified schemes are
illustrated in Figure 6-1 in time-domain.
The TS has thus 48 active sub-carriers (SC) in transmission mode I-fr. (And, respec-
tively, {12, 96, 24} active sub-carriers in modes {I-qr, II-fr, II-qr}. The transmission
modes are described in Section 5.2.) Note that all the sub-bands edges are occupied
by pilot-carriers, which is beneficial to channel estimation since all sub-carriers can
be estimated by interpolation, and no extrapolation is needed.
Across the sub-carriers, an m-sequence is modulated using differential BPSK; i.e., the
data (m-sequence) is contained in the phase difference among each pair of sub-carriers,
being {+1, 1}. 47 bits are thus carried by the TS in case of TX-mode I-fr on 48 SCs.
Those 47 binary symbols are a length 31 m-sequence (m = 5; shift-register stages 1 and
4 connected to the modulo 2 adder [12]), which is cyclically extended by 8 bits to-
wards both ends (see Figure 6-2).
In [1], two OFDM symbols comprise the training sequence. Known data is modulated
there differentially between the two OFDM symbols, for the purpose of integer fre-
quency-synchronization. The modification used here therefore reduces the overhead
introduced by one OFDM symbol per frame [2], [3].

6.2.3 Frame Timing Synchronization

In order to detect the training symbol and thereby the frame-start , the first step of
the synchronization scheme calculates a sliding correlation sum over the length and
time-lag of the periodic part (identical halves) of the training symbol in time-domain
[1]. The mathematical formulation of this so-called correlation algorithm is given in
the following equations. A metric

M (d ) = P(d )
2
(R(d ) )2 (6-1)
172 Chapter 6 DSP Algorithm Development for the Down-Link

is used, where
L 1
P(d ) = (rd*+i rd +i + L ) = P(d 1) + rd*+ L 1rd +2 L 1 rd*1rd + L 1
i =0
L 1
(6-2)
R(d ) = rd +i + L = R(d 1) + rd +2 L 1 rd + L 1
2 2 2

i =0

are a correlation sum with lag L and a power metric, respectively. The variable ri ex-
presses the sequence of received complex-valued samples of the OFDM signal. L is
the separation of equal samples in the training sequence, i.e., L = N/2, where N is the
number of FFT-points. The expressions given in (6-2) indicate an iterative way of cal-
culating R(d) and P(d), which requires just one complex multiplication and two addi-
tions/subtractions per input sample. Each of the samples ri is the sum of a signal and a
noise component, ri = si + ni (16). Their variances and the SNR are defined as

{ } { }
E Re[si ]2 = E Im[si ]2 = s2
E{Re[n ] }= E{Im[n ] }=
i
2
i
2 2
n . (6-3)
SNR = s2 n2

Detection of the TS is based on comparing M(d) with a fixed threshold th [0,1]. In

the absence of noise, M(d) takes a value of one when calculated over the TS, because
samples separated by L positions are then equal (with opposite signs); thus |P(d)| and
R(d) become equivalent. Outside the TS, the value of M(d) is close to zero, as the
sample-pairs spaced by L samples are (largely) uncorrelated. Figure 6-3 depicts the
metric M(d) as the correlation window slides across the TS.
Observe the flat top of the function M(d), due to the guard interval, because the cyclic
prefix also has equal samples at the required correlation-lag (in a non-dispersive chan-
nel to be exact). Therefore, the maximum is obtained for all positions where the
sliding correlation window overlaps the training symbol including its guard interval
(cf. Figure 6-1 and Figure 6-3). The optimum timing is at the end of the flat top. The
maximum of M(d) decreases in the presence of noise, i.e., its expected value is a func-
tion of SNR [1].

6.2.3.1 Received Signals Power Metric R(d)

R(d) is a measure for the received signals strength (plus noise) during the second half
of the FFT period. A small modification to the derivation of R(d) is suggested here.
The original method has an undesirable behavior at the end of an OFDM burst, be-
cause R(d) takes very small values when there is no signal at the input. Being the de-
nominator of the expression yielding M(d), this metric may show very large values in

16
Note that according to this definition the signal samples si are convolved with the channel impulse
response.
6.2 Overview of Synchronization Steps 173

Timing metrics after Schmiedl; mode Ifr; SNR = 30 dB

1.4

1.2

1
metric value

0.8

0.6

0.4

0.2

0
0 50 100 150 200 250 300 350 400
sample index

Figure 6-3: Ten simulations of the metric M(d) during the training sequence. The optimum
frame start corresponds to the end of the plateau.

that case, which may lead to false-detections. Extending the summation range solves
this problem, i.e., R(d) should be calculated over the full FFT-period and divided by
two

R(d ) =
1 N 1

2 i =0
1
(
rd +i = R(d 1) + rd + N 1 rd 1 .
2

2
2 2
) (6-4)

Independently, this modification was also proposed in [13].

6.2.3.2 Selecting the Threshold Value

Detecting the peak (plateau) in the metric M(d) is not a trivial task. For instance, the
correct threshold setting is a parameter to be optimized. Schmidl and Cox [1] have
studied the distribution of M(d) at the optimum position, M(dopt), and at any position
outside the TS, M(doutside). As previously mentioned, at the optimum position, E{M(d-
opt)} decreases with low SNR. Therefore, selecting the threshold too high will cause a
missing of the peak at low SNR. Outside the TS, the expected value of the metric
M(d), E{M(doutside)}, is not a function of the SNR, but it is a function of the length of
the TS in samples. The shorter the TS, the larger is this mean. Thus, selecting the
threshold too low may lead to miss-detections as the threshold may occasionally be
exceeded.
The variance of the metric (at the optimum position and outside the TS) increases with
decreased length of the TS.

Using the stochastic description of the metric M(d) as given in [1], it is possible to cal-
174 Chapter 6 DSP Algorithm Development for the Down-Link

Probabilities of missing the training sequence or of false detection

0
10

Prmiss: mode Iqr; th = 0.43

1 Prfalse: "
10
Prmiss: mode Iqr; th = 0.6
Pr(missed peak), Pr(false det.)

Prfalse: "
2
10 Pr : mode Ifr; th = 0.43
miss
Pr : " (below 106)
false
3
10

4
10

5
10

6
10
0 5 10 15 20 25 30
SNR [dB]

Figure 6-4: Probability of missing the training-sequence at the optimum position and probabil-
ity of false detection at one particular position outside the training sequence.

culate
(i): the probability that the metric M(dopt) at the optimum position is smaller than a
specific threshold th, thus the peak would not be detected (at that position), and
(ii): the probability that the metric M(doutside) is larger than the threshold, meaning
that a peak would be detected where actually no peak exists.
Figure 6-4 shows the analysis of these probabilities as a function of the SNR for the
full- and quarter-rate options of TX-mode I, respectively, denoted I-fr and I-qr (see
Section 5.2). In case of the full-rate mode, the false detection probability is very low
(below 10-6, thus it is not shown in the figure). Therefore, a threshold value th = 0.2
can be used for mode I-fr, where the missing probability becomes even smaller com-
pared to the results shown (for th = 0.43).
For the quarter rate mode, it is more difficult to find a trade-off between these two
probabilities, since the missing as well as the false-detection probabilities are rather
high at low SNR. A threshold value th = 0.43 is selected for this mode.
The probability of false detection might seem very low. However, since a miss-detec-
tion can occur at any position doutside, the event that M(d) exceeds the threshold be-
comes much more likely. Some attention is given to this problem in [1] and below.

6.2.3.3 Peak Detection

A procedure to detect the correct frame start should involve the following filtering
step. Searching for the longest peak during the frame duration rejects most false-de-
6.2 Overview of Synchronization Steps 175

tections, because the plateau at the optimum position always has considerable length
(usually it is longer than the guard interval), while outside the TS, the metric raises for
just a few samples above the threshold. (Figure 6-9 on page 181 depicts simulation re-
sults for this procedure. The probability of missing the TS is shown there as a function
of the SNR.)

DC-offsets may cause additional trouble. It is easily verified that the metric M(d) also
takes values around one if the input signal is just a constant value. Such input may be
seen in the absence of an OFDM signal (when no data or training symbols are trans-
mitted on the down-link), if the ADCs used on the I/Q-signal components have DC
offsets. Such a situation may lead to miss-detections. By monitoring the value of R(d),
which is a measure for the input signals power, it is possible to identify these false
detections as the signal strength is then very small.
Note as well that the periodic frame structure can be exploited to further enhance the
robustness of the system. Once synchronization has been acquired, there is no need to
search for the next frames start using this technique. Tracking the residual timing-off-
sets is then sufficient.
The real-time implementation of the described algorithms on the emulation platform
has proven their suitability (see Section 6.5).

6.2.4 Fractional Frequency-Offset Synchronization

The algorithm for frequency-synchronization is based on the fact that the phase-angle
of the above-defined correlation sum at its optimum position, P(dopt), is proportional to
the carrier frequency-offset, as seen from the following derivation. We first introduce
the frequency-offset to the (lowpass equivalent) signal model, written as

ri = si e j ( 2f 'i / N + ) + ni , (6-5)

where f ' is the frequency-offset normalized to the sub-carrier spacing F, and is the
carrier phase offset. Considering that during the TS si = si+L, we obtain from eq. (6-2)
in the absence of noise
L 1
P(d opt ) = si + dopt e j 2f ' L / N .
2
(6-6)
i =0

The estimate of the frequency-offset is obtained from the phase-angle of this expres-
sion,

f ' =
1

[
P(d opt ) . ] (6-7)

Due to the 2 ambiguity of the phase term, only the fractional frequency-offset in the
range of 1 sub-carriers can be determined with this method.
176 Chapter 6 DSP Algorithm Development for the Down-Link

Theoretical standard deviation of the frequencyerror estimate

1
10
mode Ifr

normalized standard deviation of ferror estimate

mode Iqr
mode IIfr
mode IIqr

2
10

3
10
0 5 10 15 20 25 30
SNR [dB]

Figure 6-5: Standard deviation (normalized to the sub-carrier spacing) of the frequency esti-
mation error vs. SNR. Theoretical results for all transmission modes.

6.2.4.1 Performance Analysis

The standard deviation of the estimate (6-7) is quantified by [4], [1]
1
f ' = . (6-8)
L SNR
The derivation of this expression is also given in Section 6.3.2. Note that the multipath
radio channel has no influence on this or on the previous equations.
Computational results are shown in Figure 6-5 for all transmission modes of the sys-
tem proposal. The standard deviation is given normalized to the sub-carrier spacing F
= 1/TFFT. It is evident from the figure and from (6-8) that the estimates get better as
more samples are available for the correlation processing introduced in eq. (6-2). The
worst results are thus seen for mode I-qr, where the standard deviation of the fre-
quency-offset estimates is respectively 2.5 and 8% of the SC-spacing at 10 and 0 dB
SNR. These values are still acceptable, as the performance wont degrade down to
about 10 dB and synchronization wont be lost down to at least 0 dB.

6.2.4.2 DC-Offset and Carrier Feed-Through

Both, DC-offsets, which may be induced for example by the ADCs of the receiver, and
carrier feed-through, which is often caused by non-perfect carrier suppression in the
transmitters I/Q-modulators, degrade the performance of the frequency-offset estima-
tion technique. A DC-offset, for instance, results in an additive constant term in
P(dopt), thus the estimate becomes biased. In Section 6.3, this problem is thoroughly
6.2 Overview of Synchronization Steps 177

performance of fractional frequencyoffset estim.; AWGN ch.; no I/Q(de)mod.

2
10
mode Ifr; experimental
mode Iqr; experimental
mode Ifr; theoretical
mode Iqr; theoretical
standard dev. of festimate [Hz]

1
10

0
10
5 0 5 10 15 20 25 30 35
signaltonoise ratio SNR [dB]

Figure 6-6: Experimental performance results of the frequency-synchronization algorithm.

Compared to the analytical curve, degradation is evident due to the impact of DC-
offsets. (The sub-carrier spacing F = 375 Hz in the emulated air-interface.)

analyzed. An extension to the algorithm is presented, which can cancel the bias in the
most relevant cases.
Experimental performance results showing the degradation due to DC-offsets are de-
picted in Figure 6-6. These data were obtained from the DSP-based emulation plat-
form running the proposed OFDM air-interface (in mode I-qr and I-fr) 17. Analytical
results are shown for comparison. The experimental results were acquired in a system
set-up without (simulated) radio channel and without I/Q-modulation and -demodu-
lation hardware. In this way, the performance of the estimation algorithm is obtained,
i.e., eventual frequency-jitters of the local oscillators used for the I/Q-(de)modulation
are not seen. The degradation of the experimental results compared with the theoretical
curves (at high SNR) is due to DC-offsets.

6.2.5 Application of the FFT

In the previous steps, the position of the TS has been roughly located, and the fre-
quency-offset has been estimated in the range of 1 sub-carrier spacing. These steps
were performed using the received time-domain signal. As further synchronization pa-

17
I wish to appreciate the contributions by my students and colleagues, P. Teneva, K. Bke, I.
GulteNLQ $ 6QLMGHUV DQG 9 -RYL WR WKH VRIW- and hardware implementation of the emulation system,
and for gathering the experimental data shown in this and in the next chapter. Detailed descriptions of
their work can be found in [14][17].
178 Chapter 6 DSP Algorithm Development for the Down-Link

rameters will be derived from the (known) data-constellation points of the training
symbol, it is required to demodulate the TS using the FFT.
In order to thereby avoid inter-carrier-interference, frequency-synchronization must be
acquired at an accuracy better than approx. 5 % of the sub-carrier spacing (see Section
4.2.4.2). As the initial (fractional) frequency-offset may be larger than that value, a fre-
quency-offset correction will be applied to the received signal, prior to the execution
of the FFT. This is achieved by multiplying the (time-domain) samples of the TS by a
complex exponential sequence according to the estimated fractional frequency-offset
f ' . The correction is written as

rdopt +i = rdopt +i e j 2f 'i / N ,

(6-9)

where i = {0, 1, , N 1}. Computing the FFT on the sequence rd opt + i
yields the data-
symbols of the TS, {yl,TS}, where l = {N/2, N/2+1, , N/21}.

6.2.6 Integer Frequency-Synchronization

The integer frequency-synchronization is based on the known data carried in the TS.
The even-indexed SCs are removed from this sequence (because they dont carry data;
cf. Figure 6-2) and differential demodulation is applied in order to recover the known
PN-sequence. Neglecting noise, the demodulated symbols are written as

y n, 2 = y 2*n 2,TS y 2n ,TS

{1}h2*n 2,TS h2 n,TS e j2 if y 2 n-2 and y 2n are data symbols , (6-10)

0 otherwise

where n = {N/4+1, N/4+2, , N/41}. The phase term

2 = 2n 2,TS 2 n,TS = 4t / TFFT (6-11)

results from the remaining timing-offset t (see eq. (4-26) in Section 4.2.4). The chan-
nel response is assumed to be highly correlated at each neighboring pair of sub-carri-
ers, therefore, the products of the channel coefficients are approximately real-valued
factors. The unknown signs express the data symbols.
The position of the PN-sequence in the demodulated training symbols is an indication
of the integer frequency-offset. For instance, in transmission mode I-fr of the proposed
OFDM system, a (cyclically extended) length-31 m-sequence mn = {1}, n =
{15,14,,15} is carried in the TS (see Figure 6-2). The integer frequency-offset is
found by searching for the magnitude-maximum of the metric
15
M int (e) = m y
n = 15
*
n n + e, 2 , (6-12)
6.2 Overview of Synchronization Steps 179

Metrics for fsync acquisition

1
hard decision metric
soft decision metric
0.8

0.6
Metric value

0.4

0.2

0.2

0.4
40 30 20 10 0 10 20 30 40
foffset in carriers

Figure 6-7: Soft- and hard decision metrics used for integer frequency-offset correction.

for e = {16,15,,16}. We obtain fi = 2 arg max ( M int (e) ) , where fi is the integer
e

frequency-error normalized to the sub-carrier spacing F. Note that this offset must be a
multiple of two sub-carriers, since fractional frequency-offset correction up to 1 SC
has been performed before. The sum (6-12) can be seen as a coherent accumulation of
all the differentially demodulated constellation points yn,2, because the multiplication
with the original m-sequence perfectly removes the signs of the data symbols at the
optimum e.
Making a decision on the data symbols prior to the calculation of Mint(e) would allow
for simplified implementation on an ASIC, as (6-12) is then easily realized in binary
arithmetic. With the latter option, Mint(e) is referred to as the hard-decision metric,
while in the former case it is called the soft-decision metric. Figure 6-7 illustrates
these metrics as a function of 2e (after normalization to the maximum magnitude).

6.2.6.1 Integer Frequency-Synchronization at the Quarter-Rate Receiver

Only a subset (about one quarter) of the data symbols yn,2 are available for determin-
ing the integer frequency offset, when the quarter rate mobile terminal has to synchro-
nize to a full-rate base-station. Similarly to the above introduced full-rate case, the po-
sition of the maximum of the metric Mint(e) gives an indication of this offset. To
maximize the acquisition range, it is recommended to cyclically extend the known m-
sequence by the number of demodulated data symbols minus one.
The performance of this method will depend on the Hamming distance the number
of bits with opposite sign between the demodulated sequence and the known PN-se-
180 Chapter 6 DSP Algorithm Development for the Down-Link

Hamming distances between l=31 msequence and l=11 subsequence

8

5 7

6
10
shift of msequence

15
4

20 3

2
25

30
0
5 10 15 20 25 30
start of subsequence

Figure 6-8: Hamming distances between any length 11 sub-sequence of the length 31 m-se-
quence and any length 11 sub-sequence of the cyclically extended sub-sequence.

quence. The Hamming distance of any length 11 sub-sequence of the cyclically ex-
tended l = 31 m-sequence to the cyclically extended l = 31 m-sequence is illustrated in
Figure 6-8. The minimum distance is three; the maximum is eight.

6.2.6.2 Performance
Results of a performance simulation are given in Figure 6-9 for the full- and quarter-
rate options of transmission mode I. Shown is the probability that the integer fre-
quency-offset has been detected falsely as a function of the SNR, for the AWGN
channel. The soft- and hard decision cases are considered. Moreover, the probability is
depicted that the training symbol is missed, using the frame timing-synchronization
algorithm described in Section 6.2.3.
It is evident from the results that the hard-decision algorithm fails with a probability up
to 20% at low SNR. In these situations, the soft-decision algorithm is clearly better; its
false detection rate is usually well below 1 %, even at 0 dB SNR. Generally, the full-
rate receiver is more robust than the quarter-rate receiver. Performance results over
Rayleigh fading channels are essentially equivalent.
As the integer frequency-offset is zero after its initial acquisition and correction, there
is no need to re-estimate this error for each frame. The estimates can be used, however,
to check if the synchronization is still correct. If a number of consecutive estimates
show arbitrary values ( 0), then synchronization has probably been lost, or the signal
is no longer strong enough. Not relying on a single erroneous estimate to identify such
a situation enhances the robustness.
6.2 Overview of Synchronization Steps 181

probability of missed peaks and false detections of integer foffset probability of missed peaks and false detections of integer foffset
0.06 0.5
missed peaks missed peaks
false int. foffset: soft dec. 0.45 false int. foffset: soft dec.
0.05 false int. foffset: hard dec. false int. foffset: hard dec.
0.4

0.35
0.04
0.3
probability

probability
0.03 0.25

0.2
0.02
0.15

0.1
0.01
0.05

0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
SNR [dB] SNR [dB]

(a) (b)
Figure 6-9: False detection probability of the integer frequency-offset. AWGN channel. (a):
transmission mode I-fr; (b): mode I-qr.

Note also the relatively high probability that the peak of the frame-synchronization
metric is missed with the quarter-rate receiver at very low SNR.

6.2.7 Remaining Timing-Offset Synchronization

A timing-offset causes a progressive phase rotation of the data constellation points
with increasing distance to the center frequency, as elaborated in Section 4.2.4. This
phase rotation is explained by the time-delay property of the Fourier transform, which
says that a time-shift in the input-waveform is seen as a linearly increasing phase-shift
in the frequency-domain.
Between equally spaced sub-carriers, i.e., for instance between adjacent sub-carriers, a
constant differential phase-shift is therefore evident due to this progressive phase rota-
tion. Differential demodulation in frequency-direction reveals this phase-shift, as seen
in eqs. (6-10) and (6-11).
From this point it is a small step to observe that a (highly accurate) estimate of the re-
maining timing-offset is obtained from the phase-angle of the soft-decision metric
Mint(e) at the optimum position eopt = fi / 2 . At eopt, the known data is removed from
the symbols yn,2, therefore, the timing-offset is conveniently detected from M int (eopt ) ,
where
TFFT N
t = M int (eopt ) [s] and t' = M int (eopt ) [samples]. (6-13)
4 4
(TFFT is the FFT-period in seconds and N is the number of FFT-points.) Note that a
positive t suggests late timing, i.e., the FFT window has started too late.

The progressive phase rotation has been previously suggested for timing-offset syn-
182 Chapter 6 DSP Algorithm Development for the Down-Link

chronization, e.g., in [7][11]. Independently of the current work, the authors of [2]
have proposed the modification of Schmidls training symbol [1], which yields the es-
timate of the timing-offset in the highly efficient way indicated above by using differ-
ential modulation in the frequency-domain. Novel work in this thesis is the perform-
ance analysis of this estimation method, which is given in Appendix C. The main re-
sults are summarized here, applied to the proposed OFDM system.

6.2.7.1 Estimation Bias on the Dispersive Channel

On multipath channels, the timing-offset is biased, since a systematic phase rotation is
caused by the product of the channel coefficient on adjacent sub-carriers h2*( n1),TS h2 n,TS
(see eq. (6-10)). This bias in terms of channel parameters is seen from the expected
value
rms
E{t} t , (6-14)
2K + 1

where rms is the RMS delay spread of the channel, and K is the Ricean K-factor.

6.2.7.2 Estimation Variance

Analytical results for the standard deviation of the estimate have been derived in
Appendix C. For AWGN channels, this performance measure is obtained from

N SNRSC + N m / 2
t ' = [samples], (6-15)
4 SNRSC N m

where Nm is the number of differentially demodulated symbols summed up to obtain

Mint (for example in eq. (6-12), Nm = 31), and SNRSC is the SNR of the used SCs of the
TS. SNRSC is related to the time domain SNR (see eq. (6-3)) by
N
SNRSC = SNR , (6-16)
N SC

where NSC is the number of active SCs in the TS.

In Appendix C, the estimates variance is derived for Rayleigh fading multipath chan-
nels as well (see eq. (C-23)). The obtained expression relates the spaced-frequency
correlation function of the frequency-selective channel to the variance of the timing-
offset estimates. The expressions for Ricean channels were not fully developed, be-
cause the Rayleigh case can be seen as a worst-case, which is sufficient to analyze in a
system design. Computational results are given below.

6.2.7.3 Performance Results

Figure 6-10 shows simulation results for the fine timing-offset estimation algorithm
6.2 Overview of Synchronization Steps 183

standard deviation of fine timingoffset estimation error; mode Ifr bias (mean) of fine timingoffset estimation error; mode Ifr; simulation res.
6
0
10 AWGN
5 rms = 3; Rayleigh
rms = 1; Rayleigh
rms = 5; Rayleigh
4
rms = 1; K = 4
1
error [samples]

error [samples]
10
3

2
AWGN; simulation
2
10 theoretical results
1
rms = 3; Rayleigh; sim.
rms = 1; Rayleigh; sim.
rms = 5; Rayleigh; sim. 0
rms = 1; K = 4; sim.
3
10 1
0 5 10 15 20 25 30 0 5 10 15 20 25 30
SNR [dB] SNR [dB]

(a) (b)
standard deviation of fine timingoffset estimation error; mode Iqr bias (mean) of fine timingoffset estimation error; mode Iqr; simulation res.
0
10
1.4
AWGN
rms = 3/4; Rayleigh
1.2
rms = 1/4; Rayleigh
rms = 5/4; Rayleigh
standard deviation [samples]

1
1
10 rms = 1/4; K = 4
error [samples]

0.8

0.6

2 0.4
10 AWGN; simulation
rms = 3/4; Rayleigh; sim.
0.2
rms = 1/4; Rayleigh; sim.
rms = 5/4; Rayleigh; sim.
rms = 1; K = 4; sim. 0
analytical results
3
10 0.2
0 5 10 15 20 25 30 0 5 10 15 20 25 30
SNR [dB] SNR [dB]

(c) (d)
Figure 6-10: Performance of the fine timing-offset estimator. (a), (b): Mode I-fr; (c), (d): Mode
I-qr; (a), (c): Standard deviations for several channels; (b), (d): Mean values (bi-
ases) obtained from computer simulations.

discussed in this section. The mean and standard deviation of the estimation error are
given as a function of the SNR for modes I-fr and I-qr of the proposed OFDM system,
over AWGN and different Rayleigh/Ricean channels. The simulation results of stan-
dard deviation are compared to analytical results, for the AWGN and Rayleigh cases.
The simulated mean values correspond to the theoretical bias given by (6-14).
In the AWGN case, the standard deviation of these estimates is as low as 0.5 samples
at an SNR of 0 dB (mode I-fr). It is much higher over the multipath channels, which is
due to the variations of the spaced-frequency correlation function (correlogram) for
individual channel simulations. It is seen that even fractional sample timing offsets can
be estimated and tracked with the proposed algorithm. This can be used to estimate the
sampling frequency offset between the transmitter and receiver (see Section 6.2.8).
Theoretical and simulation results agree well for mode I-fr. For mode I-qr, however,
184 Chapter 6 DSP Algorithm Development for the Down-Link

the disagreement is slightly bigger for some parameter values. This is probably due to
the simulation set-up implemented for the quarter rate mode, where the transmitted
signal was a full-rate signal, from which the quarter rate signal was derived by fre-
quency-shifting, low-pass filtering, and down sampling. The filter, being a rather short
FIR filter (21 taps), did not perfectly cut out the desired quarter of the frequency-band:
i.e., the amplitude response was not perfectly flat, and some aliasing was present at the
band-edges. It is assumed that the main part of the discrepancies is caused by this sig-
nal-processing step, which was not considered in the analysis.

6.2.7.4 Correction for Timing-Offset and Compensation for the Estimation Bias
The integer part (in samples) of the estimated timing-offset is used to adjust the start
sample of the FFT-interval. It is not necessary to correct explicitly for the still-re-
maining fractional part of this timing-offset, because the progressive phase rotation
caused by it is included in the channel estimate; thus the correction is done implicitly
during the equalization step performed prior to the data detection. It is noted that
equalization in an OFDM system is a multiplication of the received data constellation
points by the inverse of the estimated channel coefficients. For phase modulation
schemes (like QPSK), a multiplication by the complex conjugate is sufficient to cancel
for the phase rotations. Channel estimation is described in Section 6.4.

As the start-time of the FFT-period is estimated too late by the remaining timing-offset
estimation algorithm, the maximum bias should be considered in the OFDM system
design in order to prevent inter-symbol-interference. This is achieved by cyclically
shifting the FFT-period (of data symbols and the training symbol) so that it actually
starts a few samples before the end of the guard interval. The shift-period should cor-
respond to the bias for the Rayleigh fading channel with maximum RMS delay spread,
which is the worst-case channel in this respect. Figure 6-11 illustrates this principle.
Alternatively, this correction could be applied at the receiver, where the maximum bias
is subtracted from the estimated symbol start time. This method leads to a significant
progressive phase rotation, however, which must be considered in the channel estima-
tion algorithm (see Section 6.4.2.4).

Transmitted OFDM symbol (time-domain)

T
Twin Tguard TFFT

IFFT output

Prefix effective TX-time Postfix time

kT Nshift

Figure 6-11: Shifting the transmitted OFDM symbol to compensate for the biased timing-offset
estimates
6.2 Overview of Synchronization Steps 185

Two more aspects have to be considered in order to perform coherent detection in

OFDM systems: namely, sampling frequency offsets and carrier phase offsets. These
impairments introduce phase rotations to the data constellations, which have to be cor-
rected. In the sub-section below, we briefly discuss these issues.

6.2.8 Sampling Frequency-Offsets

The sampling clocks of ADCs and DACs are usually derived from crystal oscillators.
Typically, such oscillators have a frequency inaccuracy up to fs = 20 50 ppm,
which leads to a timing-offset of one sample after every 20,000 to 50,000 processed
samples. Although these numbers seem very large, sampling frequency-errors cannot
be neglected, as seen from the following analysis.
A sampling frequency-offset leads to a gradually increasing timing-offset. Using the
OFDM-symbol-index k as a time-variable, the resulting timing-offset in samples is
written
t ' k = t '0 + k f s ( N + N guard + N win ) , (6-17)

where t '0 is the time-offset at symbol zero, and N, Nguard, and Nwin are the numbers of
samples in the FFT-, guard-, and windowing-intervals, respectively. To assess the im-
pact of the sampling frequency-offset, we compare the phase rotation at the l-th sub-
carrier of the k-th OFDM symbol to the phase of the same sub-carrier of the zero-th
symbol. Recall that a timing-offset induces a progressive phase-rotation of the data
constellation points at the receiver. The zero-th symbol could be the training symbol,
where timing-offset correction and channel estimation are performed. That is, at the
zero-th symbol, all phase rotations are corrected. However, k OFDM symbols later, the
following phase rotation is evident, according to the system model derived in Section
4.2.4 (see eq. (4-26)) and assuming that the carrier frequency-offset is zero, f = 0.

l N + N guard + N win
= l , k l , 0 = 2 (t ' k t '0 )Ts = 2 k l f s , (6-18)
TFFT N

where Ts is the sampling interval.

It is easily verified that a sampling frequency-offset of 50 ppm leads to a phase rota-
tion of ~52 at the 47th sub-carrier (the outmost one) of the 48th OFDM symbol (l = 47,
k = 48), for system parameters according to the mode I-fr of the system proposal of
Section 5.2.

As such a phase error yields certain bit errors (with QPSK modulation), the sampling
frequency-error has to be estimated and corrected. One possibility for performing this
correction in hardware requires that the sampling clock and the oscillators used for
I/Q-(de)modulation and up/down-conversion are all locked to the same frequency-
source. Synchronizing for the carrier-frequency-offset would then automatically re-
186 Chapter 6 DSP Algorithm Development for the Down-Link

move the sampling offset as well. Below, the signal processing is outlined, which is
required to correct mathematically for the sampling frequency-offset.

6.2.8.1 Estimation of the Sampling Frequency-Offset

The sampling frequency-offset can be obtained from the timing-offset estimates of two
subsequent training symbols indexed by n 1 and n. This estimate is written as

t ' nN F t '( n1) N F

fs ,n = , (6-19)
N F ( N + N guard + N win )

where NF is the number of OFDM symbols per frame, and t'k is the estimated timing-
offset at OFDM symbol number k. Considering that each of the estimates t'k has a
standard deviation t ' (cf. Section 6.2.7.2), the standard deviation of fs , n becomes

1
f = 2 t ' . (6-20)
s
N F ( N + N guard + N win )

Evaluation with the system parameters shows that the accuracy obtained from one
such estimate is not sufficiently high, particularly, if time-variant multipath channels
are present 18. In such cases, multiple estimates (6-19) can be averaged, yielding

1 N av
t N' av N F t 0'
fs =
N av
fs ,n =
n =1 N av N F ( N + N guard + N win )
. (6-21)

It is seen that the standard deviation is reduced by a factor of Nav, which quantifies the
number of averaged estimates. This averaging can be conducted during the initial log-
in phase, i.e., when the mobile terminal initially communicates with a base station.
During this phase, communication is done via the signaling symbols, which are trans-
mitted immediately after the training symbol (k = 1). These symbols are therefore
much less affected by the sampling frequency-offset, making a correction unnecessary.

6.2.8.2 Correction for the Sampling Frequency-Offset

Once the sampling frequency-offset has been determined, a multiplication of the re-
ceived signal constellations with a complex exponential term with inverse phase of eq.
(6-18) can cancel for the phase distortions. Thereby, k can be seen as the symbol index
within a down-link frame, where the training symbol is indexed by zero.

18
The simulation and analytical results presented in Section 6.2.7 assume that the time-offset is
estimated over totally uncorrelated realizations of multipath radio channels. Since the channel transfer
functions at subsequent training symbols are correlated, the standard deviation due to the changing
channel is much smaller, when only pairs of subsequent estimates are investigated. Thus the sampling
frequency-offset estimation over Rayleigh channels performs better than suggested by the results in
Figure 6-10 and eq. (6-20). This is confirmed by experimental results presented in Section 6.5.
6.2 Overview of Synchronization Steps 187

6.2.9 Carrier Phase-Offset

A carrier phase-offset leads to a common phase rotation of the signal constellation
points, as seen from the system model derived in Section 4.2.4 (see eq. (4-26)). At the
training symbol, this common phase rotation is also incorporated in the channel esti-
mate. Therefore it is implicitly corrected by the (one-tap) equalization. However, if
residual frequency-offsets exist, the phase-offset is not constant on consecutive OFDM
symbols, but it gradually increases, just like the timing-offset increases due to a sam-
pling frequency-offset 19. The pilots can be used to estimate this common phase rota-
tion. Coherently adding the pilots (after equalizing for the channel and removing the
data) increases the SNR of the sum by the number of pilots NP compared with the SNR
of a single sub-carrier. Moreover, spreading the pilots over the whole frequency-band
exploits the frequency-diversity of the radio channel. If a sub-carrier is in a deep fade,
its contribution to the sum is small and has therefore little impact on the phase-esti-
mate.
A consecutive multiplication of all data constellations by a complex exponential with
inverse phase angle yields the correction of such offsets.

6.2.9.1 Estimation of Carrier Frequency-Offset

The progressing phase-offset due to a frequency error can be used for estimating the
carrier frequency-offset in an alternative way. The phase offset at the k-th OFDM
symbol is given by (see eq. (4-26), in Section 4.2.4)
N + N guard + N win
k = 0 + 2fkT = 0 + 2f ' k , (6-22)
N
where f is the frequency-offset in Hz, f ' is the frequency-offset normalized to the
sub-carrier spacing F, and T [s] is the total symbol duration (FFT plus guard plus win-
dowing intervals). The frequency-offset can thus be estimated from the difference of
the phase estimates of subsequent OFDM symbols,
1 1 N
f = ( k k 1 ) , or f ' = (k k 1 ) . (6-23)
2T 2 N + N guard + N win

As the phase difference can take values between , the unambiguous range of this
estimate is N /( N + N guard + N win ) F / 2 , which is less than half of the sub-carrier spac-
ing. This small acquisition range and the fact that the method requires prior execution
of the FFT (because the pilot constellations have to be recovered), limit the applicabil-

19
Note that a frequency-offset of 12 or 24 sub-carriers is required to demodulate one of the four
sub-bands of the full-rate signal using a quarter-rate receiver (in mode I). (In mode II, the required
offset is 24 or 48 sub-carriers.) Therefore, in this operation mode, a systematically progressing
common phase rotation is evident.
188 Chapter 6 DSP Algorithm Development for the Down-Link

comparison of frequency offset estimates; experimental results

2
10
estimate from TS
estimate from pilots
estimate is mean from pilots
AWGN
Rayleigh; v = 2 m/s
standard deviation of ferror [Hz]
1
10 Rayleigh; v = 5 m/s

0
10

1
10

2
10
5 0 5 10 15 20 25 30 35 40
SNR [dB]

Figure 6-12: Standard deviation of frequency-offset estimates obtained with several estimation
techniques. Experimental results without I/Q-(de)modulation hardware. SC-spac-
ing F = 375 Hz.

ity of the method to the tracking phase. Assuming that the phase estimates are obtained
from pilot tones, the variance of the frequency estimate depends primarily on the num-
ber of pilots and on the accuracy of the channel estimate. Again, the estimates of sub-
sequent symbol-pairs can be averaged in order to enhance the accuracy, which results
in a highly robust method of frequency-offset estimation.

A set of experimental results is depicted in Figure 6-12. The standard deviation of dif-
ferent estimators is shown as a function of the SNR. In order to obtain these results,
the I/Q-(de)modulation hardware was excluded from the signal path. Two types of
results are given for the method presented in this section, namely, the standard
deviation of one single estimate, and the standard deviation of the averaged estimate
using the data symbols of a whole down-link frame (48 symbols). Experiments were
performed over an AWGN channel and over two Rayleigh channels with different
mobility. For comparison, we also show the performance of the correlation-based
estimation technique, which has been described in Section 6.2.4 (estimate from TS).
Again the impact of DC-offsets is evident in the results of the correlation-based tech-
nique, leading to an error floor. Since the average signal powers have been selected
equivalently, these results are about equal for the three different channels investigated.
Equivalent results are also seen for the standard deviation of single estimates derived
from the pilots. DC-offsets and carrier feed-through do not deteriorate the performance
of this estimation scheme. Averaging can enhance the performance by a factor up to
the number of averaged estimates, which is in our case 47. Here, some impact of the
6.3 Impact of DC-Offsets and Carrier Feed-Through on Fractional Frequency-Synchronization 189

time-variant channel is observed in irreducible error floors, because the channel influ-
ences the phase of the received signal constellations as well.
Note the excellent performance for the averaged estimate. The standard deviation is
(far) below 1 Hz for any SNR above 0 dB. This is below 1 % of the sub-carrier spac-
ing of the implemented system, being F = 375 Hz.

6.2.10 Summary and Conclusions

In this section, the key algorithms have been described required for synchronizing an
OFDM receiver for the system proposal of Section 5.2. The following synchronization
steps are needed to conduct coherent demodulation of the data symbols. These steps
are all based on either the special (known) training symbol, or on pilot sub-carriers.
In order to recover the data constellation points using the FFT, time- and frequency-
synchronization must be achieved. Otherwise, the orthogonality between sub-carriers
is (partly) lost, i.e., inter-carrier-interference and inter-symbol-interference are intro-
duced. These synchronization steps are initially achieved by processing the time-do-
main input signal of the receiver. The correlation properties of the special training
symbol yield estimates for the (frame) timing and carrier frequency-offset.
Secondly, (progressing) phase rotations due to sampling frequency and remaining car-
rier frequency-offsets have to be corrected. Pilots and the phase rotation of the known
data symbols of the training symbol are used to estimate these offsets. After initial
synchronization has been acquired, these estimates also enable the tracking of residual
time- and frequency-offsets with very high accuracy.
One training symbol is employed in each frame of 69 symbols (in mode I), corre-
sponding to an overhead of 1.4 %. (37 and 35 symbols comprise a frame in mode II
and II, respectively, i.e. the overhead is about 2.8 %.) This overhead is well invested,
as it enables the efficient estimation of all synchronization parameters, with sufficient
accuracy. Performance results are shown in this section for mode I. The synchroniza-
tion performance of mode II is superior, because more signal energy is available for
the signal processing, due to the doubled symbol period.
It is noted that the training symbol is also employed for computationally efficient
channel estimation schemes, which are presented and discussed in Sections 6.4 and
6.5.3.

6.3 Impact of DC-Offsets and Carrier Feed-Through on

Fractional Frequency-Synchronization
This section further analyzes the fractional frequency-offset estimation technique,
which was described in Section 6.2.4 as step two of the synchronization procedure
proposed. Specifically, the impact of DC-offsets and carrier feed-through (CFT) is in-
vestigated. Such impairments are introduced by the receivers and transmitters hard-
190 Chapter 6 DSP Algorithm Development for the Down-Link

ware, respectively.
The analysis presented is applicable to a whole class of frequency-synchronization al-
gorithms, which utilize duplicated sequences in the OFDM signal, as e.g. the guard
interval (cyclic prefix) or a dedicated, periodic training symbol. The phase angle of the
correlation sum over the duplicated parts is an accurate measure for the carrier fre-
quency offset (cf. Section 6.2.4).
The mathematical analysis of the impairments leads to an extended frequency-offset
estimation algorithm that can be used to cancel the undesirable effects.

Orthogonal frequency division multiplexing (OFDM) is a broad-band transmission

scheme splitting the high-rate data stream into several parallel sub-channels in order to
mitigate the effects of multipath fading (see Section 4.2). Classic OFDM systems are
assumed, which use a so-called guard interval (GI) (cyclic prefix) in order to maintain
the orthogonality among sub-channels over time dispersive channels. Accurate fre-
quency synchronization is a prerequisite for these schemes, because even small inaccu-
racies (down to 2 % of the frequency spacing among sub-carriers) lead to inter-carrier-
interference (ICI) [4], [6], [18].
A popular algorithm for estimating the frequency-offset is based on the correlation
sum over duplicated parts within the OFDM signal, as for instance the GI (see [5]) or a
specific training symbol ([1], [4]). The phase angle of the correlation sum is a measure
for the frequency-offset. The algorithm is computationally highly efficient because it is
calculated directly from the time-domain signal, not requiring a prior (fast) Fourier
transform (FFT). Its accuracy is sufficient for most practical applications.
From experimental results obtained from the implementation of two OFDM systems
([19], [20]) on the DSP-based emulation platform presented in Section 5.3, a perform-
ance degradation of the frequency synchronization schemes was observed, when com-
pared to theoretical results from [1], [4], [5]. (See the experimental performance re-
sults given in Figure 6-6 and Figure 6-12.) DC-offsets introduced by the analog-to-
digital converters (ADCs) of the receiver and carrier feed-through from the I/Q-
modulators of the transmitter were identified as the main sources for this degradation.
This section presents the mathematical analysis of the influence of those hardware im-
perfections, which cause biased estimates and increased estimation error variances. An
extended algorithm is presented that perfectly cancels the impact of the DC-offset, and
for most cases, the impact of CFT as well. The computational cost of the extension is
extremely low, amounting to 2M additional complex additions and one additional
complex multiplication, where M is the length of the duplicated sequence in samples.
The analysis is presented in the following sub-sections. Section 6.3.1 presents the sig-
nal model and a number of definitions used in the analysis. The correlation algorithm
is reviewed in Section 6.3.2, and its performance is analyzed in the presence of noise,
DC-offset, and CFT. In Section 6.3.3, the proposed extension is described and ana-
lyzed. Performance results are given in Section 6.3.4, which have been obtained from
6.3 Impact of DC-Offsets and Carrier Feed-Through on Fractional Frequency-Synchronization 191

computer simulations, from the experimental platform (see Section 5.3), and from the
analytical expressions derived. Conclusions and recommendations are summarized in
Section 6.3.5.

6.3.1 Mathematical Modeling and Definitions

The sampled OFDM signal at the receiver in complex lowpass-equivalent form is
written as

ri = ( s i + VCFT )e j ( 2fi / N + ) + VDC + ni , (6-24)

where i is the sample index, N is the FFT-length, si is the transmitted OFDM signal
convolved by the channel impulse response, {ni} are independent Gaussian noise sam-
ples, f is the carrier frequency offset normalized to the sub-carrier (SC) frequency-
spacing, is a carrier phase offset, and VDC and VCFT are the (complex-valued) DC-
and CFT-components, respectively.
Note that, in this section, f denotes the normalized frequency-offset, not the frequen-
cy-offset in Hz. This variable change has been introduced for the sake of notational
convenience.
The signal and noise powers are respectively defined as 2 s2 = E{| si | 2 } and
2 n2 = E{| ni |2 } , leading to the signal-to-noise ratio SNR = s2 n2 (cf. eq. (6-3)). Fur-
thermore, it is appropriate to relate the magnitude of the DC- and CFT-components to
the OFDM signal power, written DC = | VDC | 2 2 s2 and CFT = | VCFT | 2 2 s2 . The
attenuation of the CFT by the fading channel is not considered in this study.

6.3.2 Analysis of the Synchronization Algorithm

6.3.2.1 Review of the Algorithm

The correlation sum used for frequency (and coarse time) synchronization is given by
M 1
Popt = r
i=0
i
*
ri + L , (6-25)

where L is the correlation lag (the distance in samples between the two identical se-
quences of the OFDM signal) and M is the length of the sequences. For notational
convenience, we let the index of the first sample of the first sequence be i = 0. The
magnitude of the sum (6-25) has a maximum at this position, where the correlation
window matches the duplicated sequences (cf. Section 6.2.3). Therefore, looking for
this maximum is a means for initial time-synchronization. The phase angle of (6-25) is
a measure for the frequency-offset, because the multiplicative complex exponential
sequence in (6-24) leads to a constant phase offset between each pair of samples in the
correlation sum.
192 Chapter 6 DSP Algorithm Development for the Down-Link

Original Schmidl method (PN-sequence on even sub-carriers):

GI training symbol

A A

Modified Schmidl method (PN-sequence on odd sub-carriers):

GI training symbol

A A

Guard interval-based method:

GI data symbol

Figure 6-13: Illustration of the OFDM signal-parts used for synchronization in different meth-
ods. The arrows indicate the duplicate sequences, which are correlated by eq.
(6-25).

This principle has been employed in multiple variations. In [5], the correlation sum be-
tween the GI and its equivalent at the end of the OFDM symbol is used. Therefore,
L = N, and M = Nguard, the number of samples in the GI. Moose [4] and Schmidl [1]
introduce specific training symbols (TS), which have periodic parts in the time-do-
main. We will focus on Schmidls method, who suggests an OFDM symbol as TS,
where only the even SCs are modulated with data, while the odd SCs are zero. The
FFT properties demand that such a symbol consists of two identical halves in the time-
domain, i.e., L = M = N/2. The third method considered here is a novel modification
of Schmidls method. Modulating the odd SCs in stead of the even ones and setting the
even SCs to zero leads to a symbol with identical halves but opposite signs (cf. Section
6.2.2). The modified symmetry has advantages in certain OFDM symbol configura-
tions, as, for instance, in the OFDM system proposed in Section 5.2. Moreover, the DC
component of such a symbol is zero. An illustration of these schemes is given in
Figure 6-13.
The following expression yields the estimate of the frequency-offset, where the nega-
tive sign is required for the modified Schmidl method.

( Popt )
N
f = (6-26)
2L
In the following sub-sections, the performance of the estimation technique is sepa-
rately evaluated in the presence of noise, DC-offset, and CFT.

6.3.2.2 Performance Analysis in the Presence of Noise

The analysis of the signal model (6-24) reduced to frequency (and phase) offsets and
6.3 Impact of DC-Offsets and Carrier Feed-Through on Fractional Frequency-Synchronization 193

additive noise leads to an expression for the estimates standard deviation as a function
of the SNR. The derivation outlined below can be also found in [1], [4], [5].
Without DC-offset and CFT, the correlation sum Popt (eq. (6-25)) becomes
M 1 * j ( 2f Ni + ) j ( 2f
i+ L
+ )
Popt = i
i =0
s e + n i i+ L
*
s e N
+ ni + L . (6-27)

During the correlation window, si = si+L (20), which leads to

M 1 M 1 * j ( 2f Ni + ) j ( 2f
i+ L
+ )
Popt = e j 2fL / N
s + si e ni + L + s i + L e ni* + ni*ni + L .
2 N
i (6-28)
i =0 i =0
The negative sign applies for the modified Schmidl method. With the definition of the
signal power, the first sum can be approximated by 2M s2 . Because the second sum is
a zero mean complex Gaussian random variable (RV) (with approximate variance
2M 2 s2 2 n2 ), the expectation of Popt has a phase angle proportional to the frequency-
offset. (That is, the estimate of the frequency-offset is unbiased.) The component of
this RV being perpendicular to this phase angle causes the estimation inaccuracies.
Since for small arguments x << /2, atan( x ) x , the standard deviation of the estimates
becomes
j 2fL / N
N std[Im( Popt e )] N 2 M s n N 1
f = . (6-29)
2L E[| Popt |] 2L 2 M s2
2L M SNR

This result complies with the results from [1], [4], [5]. It is seen that the standard de-
viation decreases as the number of samples accumulated in the correlation sum in-
creases, and as the correlation lag increases. An increase of the latter decreases the un-
ambiguous range of the estimate, however (see (6-26)).

6.3.2.3 Impact of DC-Offsets

To keep the involved expressions simple and compact, we next concentrate on the DC-
offset only noise and CFT are assumed to be zero. This yields

Re '*
= 2 M s2 e j 2fL / N + M VDC + 2
Popt ~
2
(
s VDC e jfL / N e jfL / N , ) (6-30)
j Im
where the positive sign and Re() apply for the conventional methods, and the negative
sign and jIm() apply for the modified Schmidl method. 's = i =0 si e j ( 2fi / N + ) is the
M 1

20
Note that in case of the guard interval based method, inter-symbol-interference is neglected to
assume si = si+L. In case of the training symbol method, the channel impulse response must be shorter
than the guard interval.
194 Chapter 6 DSP Algorithm Development for the Down-Link

sum of M (almost independent) complex Gaussian RVs si, therefore, 's is also com-
plex Gaussian with variance of ~ 2M s2 . This assumes random data being modulated
on the OFDM symbol under investigation. For a fixed TS, 's is a complex constant
with a phase angle varying with . Since, in presence of a frequency-offset, is uni-
formly distributed in [0,2), the expectation of the last term of (6-30) is zero, in both
cases.
Assuming a random TS, the errors standard deviation is given by the standard devia-
tion of the last term of (6-30) in a direction orthogonal to e , related to the ex-
j2 fL/N

pected magnitude of Popt (compare with the derivation of (6-29)).

Re sin
( )
2 std '*s VDC e jfL / N (fL / N )
f
N Im cos
2L 2 M s 2
(6-31)
N 2 M s VDC sin N 2 DC sin
= (fL / N ) = (fL / N )
2L 2M s2 cos 2L M cos
While for the standard methods (where sin() and Re() apply) this standard deviation
is zero at f = 0, the degradation is at a maximum for the modified Schmidl method.
The second term of (6-30) introduces a bias to the estimated frequency-offset, if f 0.
For small relative DC-components, the magnitude of this term in a direction orthogo-
nal to e leads to the bias
j2 fL/N

j 2fL / N
N M VDC sin( 2fL / N )
2
N E[Im(Popt e )] N L
f =# DC sin( 2f ) .
2L E[| Popt |] 2L 2 M s
2
2L N
(6-32)
The solution with the negative sign results for Schmidls original method and for the
guard interval method. The modified Schmidl method shows a positive bias. Compu-
tational results are presented in Section 6.3.4 and compared with simulation results.

6.3.2.4 Impact of Carrier Feed-Through (CFT)

In an equivalent way, bias and error-variance are derived for the simplified case that
the performance is degraded by CFT only. The correlation sum then becomes

Re *
Popt = 2 M s2 + M | VCFT | 2 +2 ( sVCFT ) e j 2fL / N , (6-33)
j Im

with s = i =0 si . The negative sign and jIm() apply for the modified Schmidl
M 1

method. It is readily seen from (6-33) that only for that method some degradation has
to be expected. For the other schemes, CFT just introduces an additional component in
the direction of e . The following results are thus relevant for the modified
j2 fL/N
6.3 Impact of DC-Offsets and Carrier Feed-Through on Fractional Frequency-Synchronization 195

Schmidl method, only.

If the TS is modulated with random data, then s is a complex Gaussian RV with ap-
proximate variance 2M s2 , just like 's . This results in an error deviation of about

N 2 M s | VCFT | N 2 CFT
f = . (6-34)
2L 2 M s 2
2L M

If, however, the TS carries a constant PN sequence, then s is a complex constant and
the modified Schmidl method yields a biased estimate with an error of

N 2 Im( sVCFT )
*
f . (6-35)
2L 2M s2

Note that this bias depends on the phase angle of the CFT-component. If this phase
changes for a set of realizations (which is the case in a realistic system set-up in lack
of perfect frequency-synchronization), the bias may appear to be a noise component.
Computational results are given in Section 6.3.4.

6.3.2.5 Common DC-offset and CFT

In the common presence of a DC-component and CFT, an additional term is present in
Popt, compared with eqs. (6-30) and (6-33). This term in direction e jfL / N is similar to
the term in (6-30) in this direction, but with 's replaced by
VCFT MI e = VCFT i =0 e j ( 2fi / N ) . The additional term in Popt is written as
M 1 +

Popt =  + 2 Re(V DC
*
VCFT )
MI e*e jfL / N e jfL / N . (6-36)

For any method, this term leads to an additional bias and to further reduced accuracy,
as seen from simulation results in Section 6.3.4. The total expression for Popt becomes

Re *
Popt = 2 M s2 + M | VCFT | 2 +2 ( sVCFT ) e j 2fL / N +
j Im
Re '*
(
+ 2 Re VDCVCFT
*
MI e*e jfL / N + 2 ) ( )
s VDC e jfL / N e jfL / N + . (6-37)
j Im
+ M VDC
2

An illustration of the estimates obtained with the modified Schmidl technique in a

system corrupted by DC-offsets and CFT is given in Figure 6-14, by the solid line.
These experimental results were acquired from the DSP-based emulation system de-
scribed in Section 5.3. The analog I/Q-modulators and -demodulators have been util-
ized to introduce some constant frequency-offset. A fixed TS has been used; therefore,
the randomness of the estimates is caused by the carrier phase offset, which changes
randomly due to the frequency-offset. No noise has been added; no fading channel
196 Chapter 6 DSP Algorithm Development for the Down-Link

experimental results; modified Schmidl technique; no channel; SNR > 30 dB

0.14

0.12
estimated frequency offset [SC]
0.1

0.08

0.06

0.04

standard algorithm
0.02 enhanced algorithm
enhanced algorithm; synchronized

0.02
20 40 60 80 100 120
frame number

Figure 6-14: Illustration of the degradation of the frequency-offset estimation algorithm in

presence of DC-offsets and carrier feed-through (CFT). A significant improvement
is achieved by using the enhanced algorithm proposed in Section 6.3.3.

was included. The other two traces are described below.

6.3.3 Extension of the Algorithm

The above analysis has shown that in particular the modified Schmidl method is sus-
ceptible to DC-offsets and CFT. The other techniques suffer less from these impacts,
especially for small frequency-offsets. However, in the acquisition phase, where the
frequency-error is large, a significant degradation is possible. Degradation is also un-
avoidable, if the frequency-error correction is done mathematically (by multiplying the
received signal with a complex harmonic; see eq. (6-9)) and not by hardware. This sec-
tion proposes an extension to the correlation algorithm, which is capable of canceling
the major impairments at low computational cost.
Subtraction of
M 1 M 1
1
Pcorr =
M
r r
i =0
i
*

i =0
i+L (6-38)

from Popt gives the improvement mentioned, as seen from the analysis of

1 ' 2 Re '* j 2f N
( )
L

= s + M VCFT I e + 2 s VCFT I e e + M VDC +

2 2 2
Pcorr
M j Im
(6-39)
Re L
jf NL
N
L
jf jf
VDC s e N
2 + 2 Re MVDC VCFT I e e e
'* * *

j Im
6.3 Impact of DC-Offsets and Carrier Feed-Through on Fractional Frequency-Synchronization 197

The negative sign and jIm() apply for the modified Schmidl method. After the sub-
traction, the interfering terms from (6-30) and (6-36) are perfectly cancelled. For the
modified Schmidl method, a small impact remains from the CFT-component in (6-33),
as seen from the jIm() term in

Popt Pcorr = 2M s2 #

1 ' 2
(
s + M VCFT 1 I e
2 2
)+ 2V
CFT
Re *
( )
s s'* I e e j 2fL / N . (6-40)
M j Im

It is noted that this term disappears for small f, because *s then becomes equal to
'*s I e . The performance is analyzed below by computer simulations. In the sole pres-
ence of noise, the standard deviation is given by (6-29), with M replaced by (M 1) in
order to acount for the subtractive (second) term in (6-40).

The improvement achieved by this enhanced algorithm for the modified Schmidl
method is seen from the experimental results depicted in Figure 6-14. The estimates
variance is drastically reduced, particularly in the synchronized case, where the fre-
quency-offset is zero. The estimates bias and standard deviation are analyzed below.

6.3.4 Analytical and Simulation Results

Simulation results are presented in this section to support the above analysis. The
OFDM system proposed in Section 5.2 is investigated in mode I, full-rate. It has an
FFT-length of N = 128 points and a guard interval of Nguard = 24 samples. The modi-
fied Schmidl scheme and the GI-based scheme are evaluated. Within the TS used with
the former, 48 of the odd-indexed SCs are modulated with random data (see Section
6.2.2). 88 SCs are modulated during the regular OFDM symbols, which are used with
the latter technique. A Rayleigh fading channel with an exponentially decaying delay
power profile at an RMS delay spread of 3 samples has been included.

6.3.4.1 Estimation Bias

In Figure 6-15, simulation results are presented for the estimation bias as a function of
the relative strength of the DC-offset and/or the CFT-component, DC and CFT. Equal
strength was assumed for the case that both impairments are present. A frequency-off-
set of 25 % of the SC-spacing was introduced (denoted in the figure as 0.25 SC). For
comparison, analytical results are given for the bias due to DC-offset.
This DC-induced systematic error is seen to raise to significant values, when DC >
10 dB, a value that may be reached particularly when the received signal is attenu-
ated by the fading channel.
A CFT-component alone does not cause any systematic offset, since random data is
assumed on the TS. But in combination with a DC-offset, the bias is influenced. Both
schemes show similar characteristics.
The extended algorithm presented in Section 6.3.3 can fully remove the degradation in
198 Chapter 6 DSP Algorithm Development for the Down-Link

foffset = 0.25 SC; no noise; Rayleigh ch. 3 samples

0.1
mod. Schmidl; DC & CFT
0.08 guard interv.; DC & CFT
mod. Schmidl; DC
mean error of estimated foffset [SC]

guard interv.; DC
0.06 mod. Schmidl; CFT
guard interv.; CFT
0.04 theory; mod. Schmidl; DC
theory; guard interv.; DC

0.02

0.02

0.04

0.06
30 25 20 15 10 5
relative power of DC and CFT component [dB]

Figure 6-15: Estimation bias of frequency-synchronization in the presence of a DC-offset and/or

a carrier feed through (CFT) component. Simulated offset 0.25F.

the cases depicted.

6.3.4.2 Standard Deviation of the Estimates

Different behavior of the two schemes is evident from the estimates standard devia-
tions shown in Figure 6-16. The performance of the guard interval (GI) based tech-
nique is indicated in Figure 6-16a. Clearly seen is an error floor at about 0.5% of the
SC-spacing, which is caused by inter-symbol-interference during the guard interval
due to the time-dispersive channel. This effect was not considered in the analysis,
therefore, it is not seen from the analytical result, which is depicted for the case of a
DC-offset. Simulation and analytical results agree at large DC (relative power of the
DC-component). Again, CFT alone does not lead to performance degradation.
Both, DC-offsets and CFT degrade the modified Schmidl scheme (Figure 6-16b),
where the analytical results are seen to appropriately describe the performance. The
small discrepancy is due to the assumption that s and 's are the sums of independent,
complex Gaussian random variables, which is not perfectly true in a practical OFDM
scheme as the one simulated.
The degradation becomes significant when the relative power of the DC- and CFT-
components is above 10 dB.

The performance as a function of the SNR is given in Figure 6-17, for a relative DC-
offset and CFT power of 10 dB, and for a frequency-offset of 10 % of the SC-spac-
ing. The figure shows the potential improvement by using the enhanced technique pro-
6.3 Impact of DC-Offsets and Carrier Feed-Through on Fractional Frequency-Synchronization 199

guard interval m.; foffset = 0.25 SC; no noise; Rayleigh ch. 3 samples
1
10
DCoffset and CFT; sim.
DCoffset only; sim.
CFT only; sim.
DCoffset only; theory
stdv. of estimated foffset [SC]

2
10

3
10
30 25 20 15 10 5
relative power of DC and CFT component [dB]

(a)
modified Schmidl m.; foffset = 0.25 SC; no noise; Rayleigh ch. 3 samples

DCoffset and CFT; sim.

DCoffset only; sim.
1 CFT only; sim.
10 DCoffset only; theory
stdv. of estimated foffset [SC]

CFT only; theory

2
10

3
10
30 25 20 15 10 5
relative power of DC and CFT component [dB]

(b)
Figure 6-16: Performance of frequency-synchronization in the presence of a DC-offset and/or a
carrier feed through (CFT) component. Simulated offset 0.25F. (a): Estimation
bias. (b): Standard deviation of the estimates.

posed in Section 6.3.3.

For the GI-based method (solid curves), the performance limit is set by the time-dis-
persive channel, being at about 0.5% SC. The theoretical result (++) shows the
performance that would be reached over an AWGN channel. There is a clear im-
provement for the enhanced technique. It is noted that the original algorithms error
200 Chapter 6 DSP Algorithm Development for the Down-Link

rel. power of DC & CFT = 0.1; foffset = 0.1 SC; Rayleigh ch. 3 samples
1
10

stdv. of estimated foffset [SC]

2
10

modified Schmidl method

guard interval (GI) method
enhanced modified Schmidl
enhanced GI method
theory: m. Schmidl, noise only
theory: GI, noise only
3
10
0 5 10 15 20 25 30
signaltonoise ratio SNR [dB]

Figure 6-17: Estimation accuracy as a function of SNR, for DC-offset and CFT of 0.1. The im-
provement by the enhance estimation technique is seen.

floor of 0.8% SC () increases for larger frequency-offsets f, while the en-

hanced method () is not influenced by f.
The time-dispersive channel does not affect the Schmidl methods (dashed lines). How-
ever, for the modified variant and the standard algorithm, the DC- and CFT-compo-
nents lead to an error floor (which is independent of f) (see ). This error
floor is drastically reduced by the extended algorithm proposed in Section 6.3.3
( ). The remaining error floor increases, however, as the frequency-offset f
increases. The performance at f = 0 approaches the theoretical result for the noise-in-
fluenced case, which is indicated by { {. Since the degradation disappears for
f = 0, the modified method utilizing the extended algorithm is as well suited for fre-
quency tracking as the original method.

6.3.4.3 Experimental Results

A set of experimental results is given in Figure 6-18 for the modified Schmidl method,
and for system parameters equal to the parameters used in the simulations. The esti-
mates standard deviation is shown as a function of the SNR over the AWGN channel.
These results have been derived from the DSP-based emulation platform described in
Section 5.3. I/Q-modulation and -demodulation are implemented in this system in ana-
log hardware, in order to realistically evaluate frequency-synchronization algorithms.
However, this hardware was removed for acquiring some of the results shown, which
gives an indication of the impact of the mixer oscillators, and allows analyzing the per-
formance of the estimation algorithm itself.
6.3 Impact of DC-Offsets and Carrier Feed-Through on Fractional Frequency-Synchronization 201

experimental results of foffset estim.; modified Schmidl method; AWGN ch.

1
10
standard algorithm; no I/Q(de)modulation
enhanced algorithm; no I/Q(de)modulation
standard deviation of ferror estimates [SC]
enhanced algorithm; with I/Q(de)modulation
theoretical result

2
10

3
10
0 5 10 15 20 25 30
signaltonoise ratio SNR [dB]

Figure 6-18: Experimental results (standard deviation of the frequency-error estimates vs.
SNR) over AWGN channels.

Comparing the standard algorithm to the theoretical result, degradation is seen at high
SNR. As the I/Q-(de)modulators were not present in the experimental system when
this result was acquired, the curve shows the variance of the estimator influenced by
DC-offsets, only. CFT did not exist, because there was no frequency-offset. Although
the absolute performance results are still promising even in the degraded case , re-
duced signal strength (for instance within a fade) would lead to loss of synchronization
and outages.
This degradation at high SNR is not evident in the results for the enhanced algorithm
proposed in Section 6.3.3. The performance of the estimation algorithm under influ-
ence of DC-offsets is seen from the curve, where the I/Q-(de)modulators were ex-
cluded. This curve agrees almost perfectly with the theoretical one. The second result
was obtained, while frequency-synchronization was performed in real-time (I/Q-
(de)modulation included). Some implementation loss is observed there.

6.3.5 Conclusions and Recommendations

Correlation-based algorithms are very efficient means of frequency-synchronization
for OFDM, and therefore very popular. This section addresses the issue of DC-offsets
and carrier feed-through components, which degrade the performance of those tech-
niques. From the analysis it is seen that estimation biases are introduced in the pres-
ence of those impairments, and that the estimates standard deviation is enlarged. The
analysis has also led to an extension of the algorithm, capable of correcting for the
most significant interference due to DC-offset and carrier feed-through.
A novel variation of Schmidls synchronization scheme [1] is discussed, where the odd
202 Chapter 6 DSP Algorithm Development for the Down-Link

sub-carriers of the training symbol are modulated with the known PN-sequence in
stead of the even ones. This modified training symbol was proposed for the OFDM
system described in Section 5.2, where it allows decreasing the complexity and opti-
mizing the performance of the synchronization and channel estimation schemes. Un-
fortunately, the modification makes the technique more susceptible to the impairments
due to DC-offsets and CFT. The extended algorithm is capable of removing the esti-
mation bias and it cancels the performance degradation in the tracking phase, i.e.,
when the frequency-offset is zero. Therefore, the application of the extended algorithm
is strongly recommended for the modified synchronization method.
When the above-mentioned modification is not applied, then the impairments disap-
pear naturally in the synchronized state, i.e., when the receivers oscillators have been
tuned to the transmitters. The original algorithms are thus well suited for frequency
tracking, even without the enhancement proposed. Certain system proposals assume,
however, that the correction for frequency-offsets is done mathematically, by multi-
plying the received signal with a complex exponential sequence, prior to the execution
of the FFT. This allows compensating for the frequency shift introduced by the oscil-
lator offsets in order to restore the orthogonality of the OFDM sub-carriers. The actual
frequency offset, however, is not cancelled; thus the degradation of the frequency-off-
set estimator is present. The enhanced algorithm can improve the performance in such
systems as well.

6.4 Channel Estimation

Channel estimation is performed using the same training symbol as for synchroniza-
tion. A known PN-sequence is modulated on the odd (pilot) sub-carriers (SC) of this
symbol (see Figure 6-2). The number of pilot SCs in this symbol is thus 48 for trans-
mission mode I-fr. (And, respectively, 12, 96, and 24 for modes I-qr, II-fr, and II-qr.)
Channel estimation can be done with largely different levels of complexity. The most
simple scheme removes the modulation from the pilots yielding the channel estimate
on the odd SCs. Linear interpolation (averaging) can be used to obtain the channel es-
timates on the even SCs. These estimates (the channel transfer function (TF) at the
data sub-carriers) will be used to demodulate the data on all consecutive down-link
symbols. Enhancements to this method include the following techniques.
Taking linear combinations of the channel attenuations at the pilots can reduce addi-
tive channel noise. This is actually a filtering operation across the SCs in order to
smoothen the estimated channel TF. Different approaches for optimizing the filtering
(Wiener filters; MMSE principle) and/or minimizing the computational effort (e.g.:
transform domain processing) have been investigated in the literature (see e.g. [21],
[22]). Section 6.4.1 presents the Wiener solution and elaborates on reducing the num-
ber of operations and memory required for its application. A disadvantage of the opti-
mum Wiener solution is the dependency of the result on the current propagation envi-
6.4 Channel Estimation 203

ronment; namely the spaced-frequency correlation function of the channel and the
noise level. In Section 6.4.2, the tradeoffs are investigated when one fixed set of filter
coefficients is used for different actual channels. It will be seen that little is gained by
such a solution in terms of bit-error-rates.

The second type of enhancements concerns the time-variability of the channel. During
the reception of the data symbols of the current frame, the channel (slowly) changes.
This will lead to increasing bit-error rates the older the channel estimate gets , if the
channel estimate is not up-dated accordingly. Channel tracking can be done in differ-
ent ways. For instance, the channel coefficients can be predicted based on previous
training symbols. The most basic scheme would linearly extrapolate to get the channel
estimate for the current symbol, using the estimates from the last two training symbols.
The substantial performance improvement enabled by this principle is seen from
simulation results given in Section 6.5. More advanced techniques for predicting chan-
nel coefficients could use (adaptive of fixed) linear filters. Decision feedback princi-
ples, where data symbols are used to track the channels changes without introducing
further training sequences, are another alternative (see e.g. [23]).

6.4.1 Wiener Filtering for Noise Reduction

An enhanced channel estimate with reduced error is obtained from linear combi-
nations of the rough estimate at the pilot SCs. This process is written in matrix nota-
tion as

h = Wp , (6-41)

where h = h + is the enhanced estimate of the channel attenuation vector (channel

 
TF) h = [h0 , h1 , hM 1 ]T , p = [ p0 , p1 , p L1 ]T is the vector of channel attenuations (plus
noise) at the pilot-SCs positions, and W is the Wiener filter, an M by L matrix. (M is
the number of channel coefficients to be estimated, which is equal to the number of
data sub-carriers, and L is the number of pilots, which is in our case the number of odd
SCs in the TS.) The remaining error is minimized in the mean square error sense
(MMSE-solution), applying the orthogonality principle E{p*T} = 0. This leads to the
well-known result for the filter W
1
W = R hp R pp , (6-42)

where R hp = E{h p*T } is the M by L cross-correlation matrix between the channel TF at

all sub-carriers and the pilot SCs, and R pp = E{p p*T } is the L by L autocorrelation ma-
trix of the pilots. Note that p is corrupted by AWGN, p = hp + n , thus
R pp = E{(h p + n) (h p + n)*T } = R hphp + N2 I L , where hp is the length L vector of channel
attenuations at the pilots, n is a vector of noise samples of variance N2, and IL is an L
204 Chapter 6 DSP Algorithm Development for the Down-Link

by L identity matrix. (see e.g. [24], [25] for more details on Wiener filtering)

6.4.1.1 Mean square error

The average mean square error (MSE) of the channel estimates h = h + is obtained
from the main diagonal of the error terms autocorrelation matrix R, i.e.,
1
2 = trace{R } . (6-43)
M
(The trace operator computes the sum over the main diagonal). Assuming that the
transformation of the pilots to the channel estimate (filtering) is done with a general
linear filter G ( h = G p ), the following expression is obtained for R

R = G R pp G *T G R*T
hp R hp G
*T
+ R hh , (6-44)

where Rhh is the M by M autocorrelation matrix of the channel vector h.

This result simplifies to

R = R hh W R*T
hp , (6-45)

if the filter G is the Wiener filter G = W, due to the orthogonality principle applied.

6.4.1.2 Reduction of computations and memory

For instance, in case of TX-mode I-fr, the length of the vector h is 88 according to the
number of data plus pilot SCs. The length of the vector p is 48. Assuming that the fil-
ter matrix W has been pre-calculated, then the calculation of one channel estimate
would require 48 88 = 4224 complex multiplications, and the same amount of com-
plex filter coefficients. This section describes how to reduce the resources needed for
this calculation by linearly combining the most relevant pilots, only.
The most significant noise reduction is obtained by linearly combining the channel es-
timates of the current and the directly neighboring SCs, since they have the largest cor-
relation among each other. The following scheme is proposed with the design con-
straint that each estimated channel coefficient should be calculated from L or L + 1
pilots, where L is an odd number. Two FIR Wiener filters (of respective lengths L and
L + 1) are designed for channel estimation on the odd and even SCs. They are applied
when there are respectively (L 1)/2 or (L + 1)/2 pilots available on both sides of the
SC to be estimated (see Figure 6-19). It is easily shown that the Wiener filter for any
of those odd or even SCs is the same (because the channel is assumed to be wide-
sense-stationary in frequency-domain), thus two short FIR filters can be utilized for
calculating most of the estimates.
Special treatment is given to the SCs on the band edges, which do not have the re-
quired number of pilots on both sides available. The L lowest (and uppermost) pilots
are used to estimate the L 1 lowest (and uppermost) SCs. An L by L 1 Wiener filter
6.4 Channel Estimation 205

L=5 zero SCs pilots

frequency

odd SCs:
use L-point symmetric
FIR Wiener filter
even SCs:
use (L+1)-point symmetric
FIR Wiener filter
(L1) lowest/upper-
most SCs:
use (L1)L Wiener filter
matrix

Figure 6-19: Principle of channel estimation at reduced complexity for L = 5.

is used for this.

This principle is visualized in Figure 6-19. The required resources reduce to approx.
(L + 0.5) 88 complex multiplications and approx. L(L + 1) complex memory loca-
tions for coefficients. E.g., if L = 5, this makes 484 multiplications and 30 memory
registers, which is a very significant reduction compared to the original numbers. The
performance evaluation is presented in Section 6.4.2.

Channel estimation for seperate parts of the TF

2

1
real

2
40 30 20 10 0 10 20 channel
30 40
SCindex pilots
odd
2 even
lowest
1 uppest
imag

2
40 30 20 10 0 10 20 30 40
SCindex

Figure 6-20: Illustration of the enhanced channel estimation scheme at reduced complexity (for
L = 7). Indicated by {, , , are the estimated channel coefficients. The
original channel without noise is shown by the solid line; the noise-corrupted pilots
are marked by +.
206 Chapter 6 DSP Algorithm Development for the Down-Link

The FIR Wiener filter is designed by replacing the vector h in eq. (6-41) by a single
channel coefficient denoted h . Accordingly, the matrices W and Rhp reduce to row
vectors w and rhp.
Figure 6-20 illustrates how a complete estimate of the channel transfer function is ob-
tained, using the four filters mentioned above.

6.4.1.3 Introducing the Channel Model

The correlation matrices used in designing the Wiener filters must be calculated from
the channel model. For the frequency-domain channel model introduced in Chapter 2,
the spaced-frequency correlation function is given by (cf. eq. (2-22))

P0 1
H (f ) = K + , (6-46)
K +1 1 + j 2f rms K 1

where the channel is defined by its average power P0, its RMS delay spread rms, and
its Ricean K-factor K, and f is the frequency spacing for which the autocorrelation
function is evaluated. The constant K1 relates to the factor K as K1 = ( K + 1) 2 K + 1 .
An exponentially decaying delay spectrum is assumed, with a direct path at excess
delay time = 0.
To evaluate the effect of time-synchronization errors, possible timing-offsets are con-
veniently incorporated in this correlation function. A time-shift t in the excess delay
time-domain the domain of the Fourier transform of H(f) corresponds to a pro-
gressive phase rotation in H(f)

~ P 1 j 2tf
H (f ) = H (f )e j 2tf = 0 K + e . (6-47)
K +1 1 + j 2f rms K1

These continuous time correlation functions can be used to calculate the auto- and
cross-correlation matrices needed in the filter design. As an example, the derivation of
the vector rhp = E{h p*T } is demonstrated, which is needed for the calculation of the
MMSE filter for the odd SCs. The frequency-spacing of the OFDM sub-carriers is de-
noted F. Using (6-46) or (6-47) allows to evaluate

rhp = E{h p*T } = E{h (h p + n)*T } = E{h h *pT } =

E{hi [hi L+1 hi L+3  hi + L1 ]* } = . (6-48)
[ H (( L 1) F ) H (( L 3) F )  H (( L 1) F )]*

The other correlation vectors and matrices are obtained in an equivalent way.

6.4.2 Computational Results

Performance results are presented in this section in terms of the mean square error
6.4 Channel Estimation 207

MSE vs. SNR; no design mismatch; rms = 3 samp.; K = 0 MSE vs. SNR; no design mismatch; rms = 1 samp.; K = 0
0 0
10 10
average MSE sim. average MSE sim.
average MSE theory average MSE theory
MSE of pilots 1 MSE of pilots
L=3 10 L=3
1
10 L=7 L=7
mean square error (MSE)

mean square error (MSE)

L = 15 L = 15
2
10
2
10
3
10

3
10
4
10

4 5
10 10
0 5 10 15 20 25 30 0 5 10 15 20 25 30
SNR [dB] SNR [dB]

(a) (b)
Figure 6-21: Comparison of MSE results vs. SNR for different channel estimation filter lengths.
Rayleigh channels; (a): rms = 3 samples; (b): rms = 1 sample.

(MSE) and bit-error-rate (BER), as a function of the signal-to-noise ratio. Particularly,

we investigate the tradeoff between a solution optimized to the current channel/noise
parameters and a fixed solution, where a single set of filter parameters is used for a va-
riety of channels. Moreover, the influence of timing-offsets is evaluated with an em-
phasis on the systematic timing-offset introduced by the timing-synchronization
scheme.

6.4.2.1 Number of Multiplications

Firstly, we investigate the influence of the number L of (complex) multiplications per
estimated channel coefficient on the mean square error performance. Figure 6-21 com-
pares theoretical and simulation results for different L. Two Rayleigh fading channels
are considered with RMS delay spreads of rms = {1, 3} samples. The channel estima-
tion filters were optimized (adapted) to these channel parameters and to the current
signal-to-noise ratio (SNR). I.e., new filters were designed for each measurement point
shown. As a reference, the MSE of the pilot SCs is shown prior to any post-processing.
It is seen that using 15 multiplications per channel coefficients, only very small im-
provements in terms of MSE occur compared to using 3 or 7 multiplications. The gain
is slightly more significant for the rms = 1 sample channel, since higher correlation
between pilots can be exploited. We conclude that a filter length of 3 to 7 is sufficient.
Furthermore, in the figure, analytical MSE results are compared to simulation results.
The excellent agreement suggests that both types of results are appropriate.

6.4.2.2 Filter Design for Fixed SNR

Figure 6-22 shows MSE results for the case that the channel estimation filters were de-
signed for a fixed SNR, while the channel is varied over a whole range of SNR-values.
208 Chapter 6 DSP Algorithm Development for the Down-Link

MSE vs. SNR; SNR design mism.; = 3 samp.; K = 0; L = 7 MSE vs. SNR; SNR design mism.; = 1 samp.; K = 0; L = 7
rms rms
0 0
10 10
variable SNR design variable SNR design
design for SNR = 10 dB design for SNR = 10 dB
design for SNR = 30 dB design for SNR = 30 dB
MSE of pilots MSE of pilots
1 1
10 10
mean square error (MSE)

mean square error (MSE)

2 2
10 10

3 3
10 10

4 4
10 10
0 5 10 15 20 25 30 0 5 10 15 20 25 30
SNR [dB] SNR [dB]

(a) (b)
Figure 6-22: Performance of filter designs for fixed SNR in terms of MSE vs. actual SNR.
Rayleigh channels; (a): rms = 3 samples; (b): rms = 1 sample.

The SNR values mentioned relate to the SNR of the pilot SCs. These are approx. 4.3
dB above the time-domain SNR shown on the abscissas of the performance plots. The
design for low SNR shows an error floor at high SNR, where the filtered channel esti-
mate is becoming worse than the rough estimate without any filtering. Designing the
filter for high SNR means loosing some accuracy in the low-SNR range, but the per-
formance stays superior to the unfiltered case. Again it is seen that higher gains can be
achieved over less dispersive channels (when the filters are adapted to the channels
RMS delay spread).

6.4.2.3 Filter Design for a Fixed Channel

The next step is the design of a filter for fixed channel parameters (and for fixed
SNR = 30 dB). As seen from Figure 6-23, the best common design results from con-
sidering the channel with most frequency selectivity (i.e., highest rms). No noise floors
are observed, when the filters are designed for the rms = 5 samples channel (see Figure
6-23c). It is also seen, however, that less is gained over any channel when this design
is used. We conclude that a fixed filter design should be done for the channel with the
largest rms allowed by the guard interval of the OFDM system, and for high SNR.

6.4.2.4 Impact of Timing-Offsets

As seen from eq. (6-47), timing-offsets do have influence on the channel correlation
function as well. Therefore, an adaptive Wiener filter has to be optimized for this pa-
rameter too. In particular, the bias of the fine timing-offset estimation algorithm pre-
sented in Section 6.2.7 should be considered in the design of the channel estimator.
The bias is given by eq. (6-14).
Computational results of the MSE estimator performance are shown in Figure 6-24 in
6.4 Channel Estimation 209

fixed estimator design: = 1 samp.; K = 0; L = 7; SNR = 30 dB fixed estimator design: = 3 samp.; K = 0; L = 7; SNR = 30 dB
rms rms
0 0
10 10

ch: rms = 1; K = 0 ch: rms = 1; K = 0

ch: = 3; K = 0 ch: = 3; K = 0
rms rms
ch: = 5; K = 0 ch: = 5; K = 0
1 rms 1 rms
10 ch: rms = 1; K = 4 10 ch: rms = 1; K = 4
mean square error (MSE)

mean square error (MSE)

MSE of pilots MSE of pilots

2 2
10 10

3 3
10 10

4 4
10 10
0 5 10 15 20 25 30 0 5 10 15 20 25 30
SNR [dB] SNR [dB]

(a) (b)
fixed estimator design: = 5 samp.; K = 0; L = 7; SNR = 30 dB fixed estimator design: = 1 samp.; K = 4; L = 7; SNR = 30 dB
rms rms
0 0
10 10

ch: = 1; K = 0 ch: = 1; K = 0
rms rms
ch: rms = 3; K = 0 ch: rms = 3; K = 0
ch: rms = 5; K = 0 ch: rms = 5; K = 0
1 1
10 ch: rms = 1; K = 4 10 ch: rms = 1; K = 4
mean square error (MSE)

mean square error (MSE)

MSE of pilots MSE of pilots

2 2
10 10

3 3
10 10

4 4
10 10
0 5 10 15 20 25 30 0 5 10 15 20 25 30
SNR [dB] SNR [dB]

(c) (d)
Figure 6-23: MSE performance of channel estimation vs. SNR for fixed estimation filters. The
four sub-figures depict results for several channel parameters used for the filter
design. (All channel estimators have been designed for SNR = 30 dB). The MSE is
evaluated over a range of actual channel parameters.

order to evaluate the importance of this issue. One fixed filter (designed for a rms = 3
samples Rayleigh channel at SNR = 30 dB and with L = 7) is evaluated over a rms = 3
and a rms = 1 sample channel, for timing offsets of {3, 0, 3} samples. Quite a signifi-
cant degradation is evident at high SNR, especially, when the delays are positive.
Positive delays mean that the phase of the transfer function increases more rapidly vs.
frequency; more rapidly than the phase increase accounted for by the correlation func-
tion. This leads to errors. Negative delays, on the other hand, (up to a certain extent)
mean slower increase of the phase than the design target, which is less critical for the
estimators performance.
Since the fine timing-offset estimator described in Section 6.2.7 is biased, the channel
estimation filter should be designed for such offsets. MSE results for this case are de-
picted in Figure 6-25, where the filter was designed for a rms = 3 samples Rayleigh
210 Chapter 6 DSP Algorithm Development for the Down-Link

MSE vs. SNR; filter design: t = 0 samp., rms = 3 samp., K = 0, L = 7 MSE vs. SNR; filter design: t = 0 samp., rms = 3 samp., K = 0, L = 7
0 0
10 10

t = 0 smp.; rms = 3 t = 0 smp.; rms = 1

t = 3 smp.; rms = 3 t = 3 smp.; rms = 1
t = 3 smp.; rms = 3 t = 3 smp.; rms = 1
1 1
10 MSE of pilots 10 MSE of pilots
mean square error (MSE)

mean square error (MSE)

2 2
10 10

3 3
10 10

4 4
10 10
0 5 10 15 20 25 30 0 5 10 15 20 25 30
SNR [dB] SNR [dB]

(a) (b)
Figure 6-24: MSE results of a fixed channel estimator in presence of timing offset. Filter de-
signed for zero offset, rms = 3 samples Rayleigh channel at SNR = 30 dB. (a): Ac-
tual channel has rms = 3 samples; (b): Actual channel has rms = 1 sample.

channel at SNR = 30 dB, and for the corresponding expected timing-offset of 3 sam-
ples (see eq. (6-14)). In Figure 6-25a, an actual channel with matching rms = 3 samples
is investigated for time offsets of {3, 0, 6} samples, corresponding to the expected bias
plus {0, 3, +3} samples. The results are almost equivalent to the results shown above
in Figure 6-24a. In Figure 6-25b, a channel with rms = 1 samples is considered, for
time offsets of {1, 2, 4} samples, which is again the expected offset plus {0, 3, +3}
samples. Less degradation is seen in this case, which is because the relative positive
offset is now smaller than in the case shown in Figure 6-24b.

MSE vs. SNR; filter design: t = 3 samp., rms = 3 samp., K = 0, L = 7 MSE vs. SNR; filter design: t = 3 samp., rms = 3 samp., K = 0, L = 7
0 0
10 10

t = 3 smp.; rms = 3 t = 1 smp.; rms = 1

t = 0 smp.; rms = 3 t = 2 smp.; rms = 1
t = 6 smp.; rms = 3 t = 4 smp.; rms = 1
1 1
10 MSE of pilots 10 MSE of pilots
mean square error (MSE)

mean square error (MSE)

2 2
10 10

3 3
10 10

4 4
10 10
0 5 10 15 20 25 30 0 5 10 15 20 25 30
SNR [dB] SNR [dB]

(a) (b)
Figure 6-25: MSE results of a fixed channel estimator in presence of timing offset. Filter de-
signed for offset of 3 samples, rms = 3 samples Rayleigh channel at SNR = 30 dB.
(a): Actual channel has rms = 3 samples; (b): Actual channel has rms = 1 sample.
6.4 Channel Estimation 211

QPSK; channel estimator: L = 7, SNR = 30 dB, t = 3 s., rms = 3 s., K = 0 16QAM; channel estimator: L = 7, SNR = 30 dB, t = 3 s., rms = 3 s., K = 0
0 0
10 10

1 1
10 10

2 2

average BER
average BER

10 10

3 3
10 10
K = 0; rms = 1 sample K = 0; rms = 1 sample
K = 0; rms = 3 samples K = 0; rms = 3 samples
K = 0; rms = 5 samples K = 0; rms = 5 samples
4 4
10 K = 4; rms = 1 sample 10 K = 4; rms = 1 sample
theory; K = 4 theory; K = 4
theory; K = 0 theory; K = 0
5 5
10 10
5 0 5 10 15 20 25 30 5 0 5 10 15 20 25 30
average Eb/N0 [dB] average E /N [dB]
b 0

(a) (b)
Figure 6-26: BER results vs. SNR for various radio channels. The parameters in the figures
headers specify the design parameters of the channel estimator. (a): QPSK modu-
lation; (b): 16-QAM.

We conclude that the bias of the timing-offset estimation technique might degrade the
performance of channel estimation schemes that conduct filtering across the fre-
quency-domain. It is recommended to consider this bias in the design of the channel
estimator.

6.4.2.5 Bit-Error-Rate Results

Using the same fixed estimation filter design investigated above (L = 7, SNR = 30 dB,
rms = 3 samples, K = 0, t' = 3 samples), bit-error rate (BER) simulations were per-
formed for various sets of channel parameters. Results are depicted in Figure 6-26 as a
function of Eb/N0. QPSK and 16-QAM modulation schemes were investigated. Ana-
lytical plots are included in the figures as a benchmark, for which perfect channel
knowledge was assumed.
A small degradation of approx. 1 dB is evident, compared to the theoretical plots. Only
for the rms = 5 samples channel, an error floor is seen to arise for high SNR (at the
SNR-range depicted). The timing-offset estimation bias was considered in all simula-
tions. Except for the 3-dB difference in Eb/N0, results are equivalent for QPSK and 16-
QAM modulation. Note that the 3-dB difference in Eb/N0 is equal to a 6-dB difference
in time-domain SNR and in the SNR of the data symbols (sub-carriers), because re-
spectively two and four bits are transmitted per QPSK or 16-QAM symbol.

6.4.2.6 Performance Using the Unfiltered Estimate

In this sub-section, performance results are presented for the basic estimation tech-
nique, which just removes the pilot data to obtain the estimates at the odd sub-carriers.
To determine the channel coefficients of the even sub-carriers, the estimates of the two
212 Chapter 6 DSP Algorithm Development for the Down-Link

no interpolation on odd subcarriers; linear interpolation on even subcarriers QPSK; channel estimation without filtering; linear interpolation on even SCs
0
0
10 10

ch: rms = 1; K = 0
ch: rms = 3; K = 0 1
ch: rms = 5; K = 0 10
1
10 ch: rms = 1; K = 4
mean square error (MSE)

MSE of pilots
2

average BER
10

2
10
3
10
K = 0; rms = 1 sample
K = 0; rms = 3 samples
3 K = 0; rms = 5 samples
10 4
10 K = 4; rms = 1 sample
theory; K = 4
theory; K = 0
5
4
10 10
0 5 10 15 20 25 30 5 0 5 10 15 20 25 30
SNR [dB] average Eb/N0 [dB]

(a) (b)
Figure 6-27: Performance of the basic channel estimation method using interpolation at the
even sub-carriers and no processing on the odd ones. (a): MSE results; (b): BER
results for QPSK.

adjacent carriers are averaged (linear interpolation). Figure 6-27 depicts the perform-
ance of this technique in terms of the MSE and the BER for QPSK. It is remarkable to
observe that the BER results are hardly worse than the results for the complex scheme
evaluated above, although error floors are observed in the MSE results for the channels
with more frequency-selectivity (= longer RMS delay spread).
The discrepancy between the average MSE results of the estimate and the MSE of the
pilot SCs is due to the averaging used on the even SCs. Averaging two noisy estimates
halves the noise variance. Therefore, the average MSE (odd and even SCs) is by about
a factor of 1.5 better than the MSE of the pilots, which is also the MSE of the odd SCs.
This applies at low SNR, where noise is dominant. At high SNR, the averaging intro-
duces an error compared to the actual channel transfer function, particularly on chan-
nels with more frequency-selectivity, i.e., on channels with higher RMS delay spread.
In this range, the Wiener solution may show clear advantages, if its design parameters
are well selected (see Figure 6-23).
The results of this simple technique are that promising, because the SNR at the pilot
SCs is almost double the SNR of data SCs when the average transmitted power is kept
constant, as the number of pilots in the TS is about halve the number of data SCs. If
there were a pilot on each sub-carrier, the average MSE would be a factor of three
higher at low SNR. On the other hand, the error floors were not present.

6.4.3 Concluding Remarks

The channel estimation scheme presented in this section is based on periodically trans-
mitted training symbols (TS). The mean-square-error (MSE) of the estimated channel
transfer function is reduced by calculating linear combinations of the rough estimates,
6.5 Experimental Results 213

which are obtained from the known TS. (Filtering is performed in the frequency-do-
main.) It has been shown by analysis that with only a few complex multiplications per
estimated channel coefficient, the MSE performance of the estimator can be
optimized. Even a fixed set of parameters can yield good performance on a wide
variety of channels, if selected appropriately. Thereby, the bias of the timing-offset
estimator should be considered, as timing-offsets introduce progressive phase-rotations
to the data constellations. These phase rotations will be reflected in the channel
estimate. It is assumed that an adaptive filter can yield optimum results on a variety of
practical channels. However, the computational effort and the complexity increase
with every additional processing step.
BER results show that little is gained with the proposed filtering techniques, compared
to the most basic estimation method. This basic method just removes the data from the
(odd-indexed) sub-carriers to get the channel coefficients. On the even sub-carriers,
where no pilot data is transmitted, averaging of two adjacent estimates is performed.
The second advantage next to the surprisingly good performance is the fact that no
adaptability is needed to optimize this estimator. It is therefore recommended to utilize
this basic estimation method, in practice.
Note that the time-variability of the radio channel was not considered in the above
analysis. This impairment introduces an irreducible error floor to the BER results, as
seen from the experimental results below. It will be shown that simple channel predic-
tion techniques are capable of significant improvements.

6.5 Experimental Results

In this section, experimental performance results are presented and discussed, which
have been acquired from the emulation system presented in Section 5.3. The main as-
pects of the implemented OFDM air-interface comply with the system proposal given
in Section 5.2. Transmission mode I has been realized, which utilizes a 128-point FFT
in the full rate option. The number of OFDM symbols transmitted on the down-link
had been fixed to 48 symbols per frame. Pseudo random data was modulated on those
symbols using QPSK, in order to determine the bit error rate (BER) at the receiver. 69
symbols comprise one frame, which includes the training symbol and guard symbols
for time-division-duplexing. The signal processing steps of the realized OFDM re-
ceiver agree with the procedures presented in Section 6.2.
BER performance results and the standard deviation of the fine timing-offset estima-
tion technique (see Section 6.2.7) are presented in Section 6.5.1 for the full-rate re-
ceiver. In Section 6.5.2, the BER performance of the quarter-rate receiver is discussed,
which demodulates one of the four sub-bands of the transmitted full-rate signal. A
general discussion of the results is given in Section 6.5.3. Particularly, we address the
performance limitation due to the channels time-variability, and we analyze a tech-
nique to predict these time-variations.
214 Chapter 6 DSP Algorithm Development for the Down-Link

6.5.1 Performance Results for the Full-Rate Receiver

6.5.1.1 AWGN Channel

This sub-section presents performance results for the AWGN channel as a function of
the signal-to-noise-ratio (SNR). Several hardware set-ups are compared in order to
analyze the impact of various system components. Particularly, the I/Q-modulator and
-demodulator were modified in order to study their influence and to evaluate the influ-
ence of real-time frequency-synchronization. The following configurations were in-
vestigated:
(1) I/Q-modulator/demodulator excluded
(2) I/Q-modulator/demodulator included with free-running VCOs
(3) I/Q-modulator/demodulator included with crystal oscillator (XO) stabilized
VCOs (note that two measurements are shown for this set-up)

Figure 6-28 compares bit error rate (BER) results for these cases. Ten measurements
were conducted at each SNR-value. About one million of bits were transmitted for
each measurement. The BER-curves versus the SNR show the averages of these meas-
urement sets; the error-bars indicate the ranges between the minimum and maximum
values observed. An analytical result is shown as a benchmark, where perfect channel
estimation is assumed. Compared to this curve marked by , approx. 2 dB of loss are
evident for the implemented system. This loss is made up of the imperfect channel es-

BER results vs. SNR over AWGN channel; influence of I/Qmodulator

0
10
no I/Qmodulator
w/ VCXOs
1
10 w/ VCXOs
w/ VCOs
analytical
2
10
biterrorrate (BER)

3
10

4
10

5
10

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

Figure 6-28: Experimental results of bit error rate (BER) as a function of the signal-to-noise
ratio (SNR) for AWGN channels. The influence of the I/Q-(de)modulation block
can be observed.
6.5 Experimental Results 215

performance of fine timingoffset estimation from the training sequence; AWGN

1
10
no I/Qmodulator
w/ VCXOs

stdv. of fine timingoffset estimates [samples]

w/ VCXOs
0
w/ VCOs
10 analytical

1
10

2
10

3
10
5 0 5 10 15 20 25 30 35 40
SNR [dB]

Figure 6-29: Standard deviation of the remaining timing-offset estimation algorithm vs. SNR.
AWGN channel. Experimental results.

timation and of hardware-induced implementation losses (non-linearity; intermodula-

tion; etc.). There is no difference between results from set-up (1) no I/Q-
(de)modulation, and set-up (3) I/Q-(de)modulation using the XOs. This indicates that
the real-time frequency-synchronization algorithm is sufficiently robust. An error floor
at a BER of approx. 210-4 results from the phase noise and frequency jitter of the free-
running VCOs for set-up (3), where free-running LC-oscillators were applied.

Standard deviation results for the timing-offset estimator (see Section 6.2.7) indicate
that the hardwares influence on this algorithm is negligible (see Figure 6-29). The
flattening of the curves observed in the high SNR-region occurs at standard deviation
values as low as 0.01 samples and is therefore not investigated further.

6.5.1.2 Performance over Time-variant Rayleigh Channels

The following figures depict performance results over time variant Rayleigh fading
channels. Again, the impact of I/Q-(de)modulation using the XOs is investigated by
comparison to results without I/Q-(de)modulation. Results are given for two different
simulated radio channels. Their common parameter is the RMS delay spread of 3 sam-
ples, which corresponds to an impulse response length in the order of the guard inter-
val. The channel-variance was selected according to a mobile moving at 2 and 5 m/s,
considering a carrier frequency of 60 GHz
BER results are shown in Figure 6-30. Compared to the analytical curve, for which
perfect channel knowledge was assumed, the experimental results show error floors
due to the channels time-variability, because the channel estimate obtained at the be-
216 Chapter 6 DSP Algorithm Development for the Down-Link

BER results vs. SNR; Rayleigh, rms = 3 samples; influence of I/Qmodulator

0
10
w/ I/Q; slow tvar. (v = 2 m/s)
w/ I/Q; fast tvar. (v = 5 m/s)
no I/Q; slow tvar.
1
no I/Q; fast tvar.
10 analytical
average biterrorrate (BER)

2
10

3
10

4
10
5 0 5 10 15 20 25 30 35 40
average Eb/N0 [dB]

Figure 6-30: Performance results for Rayleigh channels with rms = 3 samples and different
time-variability. The impact of the real-time frequency-synchronization and of the
I/Q-(de)modulation hardware is evident.

ginning of each frame is used to demodulate all data symbols of the current frame.
Therefore, the performance is worse for the channel with higher velocity. This topic is
further discussed in Section 6.5.3.
Inserting the I/Q-(de)modulators with the crystal oscillators leads to some additional
degradation. The source of these implementation losses was not fully determined. Rea-
sons for these degradations may include the slightly reduced accuracy of the frequency
synchronization particularly when the received signal is attenuated by a fade , and
inter-carrier interference due to intermodulation distortion.

Experimental results of the standard deviation of the timing-offset estimation algo-

rithm show a behavior different to the simulation results and analytical results shown
in Section 6.2.7 (see Figure 6-31 and Figure 6-10). According to these previous
results, the multipath-fading channel causes a standard deviation-floor of about one
sample for a channel with an RMS delay spread of three samples. The experimental
results in Figure 6-31 show this error floor at below one sample. The reason for this
behavior is that the experimental result is obtained from the time-difference of two
subsequent timing estimates, because no absolute time-reference is available at the
receiver. That is, the error floor of the standard deviation curves is caused by the
change of the channel transfer function in-between two training symbols. Although
this function does change, consecutive realizations are well-correlated, which leads to
smaller standard deviation values compared to the theoretical results, where the stan-
6.5 Experimental Results 217

performance of fine timingoffset estimation from TS; Rayleigh, = 3 samples

rms
1
10
w/ I/Q; slow tvar. (v = 2 m/s)
w/ I/Q; fast tvar. (v = 5 m/s)
stdv. of fine timingoffset estimates [samples] no I/Q; slow tvar.
no I/Q; fast tvar.
analytical

0
10

1
10
5 0 5 10 15 20 25 30 35 40
SNR [dB]

Figure 6-31: Standard deviation of fine timing-offset estimates vs. SNR over Rayleigh channels
with rms = 3 samples. The influence of the frequency-synchronization is evident.

dard deviation is investigated in case of uncorrelated channel realizations. This also

explains the dependency of the results on the mobiles velocity, as the correlation de-
creases when the speed increases.
The I/Q-(de)modulation hardware has hardly any influence on this performance meas-
ure.

The influence of channel parameters on BER-results is further evaluated in Figure

Rayleigh channel; no IQ; various rms; fm = 2105, i.e., v = 2 m/s at 60 GHz Frameo. sch.; Rayleigh ch.; no IQ;rms = 3 samples; various fm (mobility)
0 0
10 10

rms = 1 sample fm = 2e5 (v = 2 m/s @ 60 GHz)

rms = 3 samples f = 2e6 (v = 0.2 m/s)
m
rms = 5 samples fm = 5e5 (v = 5 m/s)
1 1
10 theory; perfect channel est. 10 theory; perfect channel est.
average BER

average BER

2 2
10 10

3 3
10 10

4 4
10 10
5 0 5 10 15 20 25 30 35 40 5 0 5 10 15 20 25 30 35 40
average Eb/N0 [dB] average E /N [dB]
b 0

(a) (b)
Figure 6-32: BER vs. SNR for Rayleigh fading channels. (a): Dependency of the BER on the
RMS delay spread of the channel. (b): Dependency on the mobiles velocity.
218 Chapter 6 DSP Algorithm Development for the Down-Link

6-32, where the BER performance vs. SNR is depicted for Rayleigh channels with dif-
ferent mobility and different RMS delay spread rms. No I/Q-(de)modulation was pre-
sent. While the impact of rms at a given velocity is small (see Figure 6-32a), a clear
influence of the mobiles velocity is observed again (see Figure 6-32b). Note that the
individual measurement results span a larger range on the slower channels, because the
signal may be in a good or bad fading situation for considerable time while a
measurement point is acquired (recall that ten measurements are performed per SNR-
value).

6.5.2 Performance Results for the Quarter-Rate Receiver

In this section, experimental results are presented for the down-link from a full-rate
(mode I) transmitter (base-station) to a quarter-rate receiver (mobile). The idea of the
quarter rate option is to enable mobile-terminals that can access the system with base-
band processing algorithms running at one quarter of the sampling frequency. In the
system implemented on the emulator, the transmitter operates on 44.1 kHz sampling
rate, while the receiver runs on 11.025 kHz (see also below). The anti-alias filters of
the receivers ADCs thereby suppress all signal components that lie outside of the de-
sired frequency band.

6.5.2.1 Results over the AWGN Channel

The BER for the quarter rate receiver vs. SNR is depicted in Figure 6-33. Results for
different hardware set-ups are compared to the analytical result for perfect channel es-
timation. The performance is slightly inferior to the performance of the full-rate re-
ceiver given in Figure 6-28.
The degradation with respect to the idealized analytical result has increased to ap-
proximately 2.5 3 dB, compared to 2 dB for the full-rate mode. Also the irreducible
BER, when the free-running VCOs are used, has increased slightly from 2e-4 to about
4e-4. These additional losses are most likely due to the smaller accuracy of common
phase rotation estimates obtained from the pilots (cf. Section 6.2.9), since the full rate
mode has 12 pilots to exploit, while the quarter rate mode has only three (see Section
5.2).
To enable the demodulation of any of the four sub-bands, the transmitter can shift the
generated signal in frequency by means of multiplication with a complex exponential
sequence. In this way, the results were obtained without I/Q-(de)modulators, and with
the crystal oscillators stabilized VCOs. Unfortunately, the tuning range of the crystal
oscillators is not sufficient to perform this frequency shifting. With the free-running
VCOs, however, the shifting can be performed by I/Q-modulation and -demodulation,
using local oscillators with an appropriate frequency-offset.
It is quite difficult to accurately measure the SNRs in this transmission mode. The val-
ues shown in the figures of these sections are obtained from estimates of the signal
plus noise power and of the noise power. Both powers are measured from the time-
6.5 Experimental Results 219

BER results vs. SNR over AWGN channel; influence of I/Qmodulator

0
10
no I/Qmodulator
w/ VCXOs
1
10 w/ VCOs
analytical

2
10
biterrorrate (BER)

3
10

4
10

5
10

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

Figure 6-33: BER for the quarter rate receiver vs. SNR. Results for different hardware set-ups
are compared to the analytical result for perfect channel estimation.

domain received signal, which is limited in bandwidth by the ADCs anti-aliasing fil-
ters. However, these filters also pass parts of the adjacent sub-bands, therefore it is dif-
ficult to estimate which ratio of the signal power belongs to the demodulated sub-band.
The 22/29-th part of the incoming signal was assumed to belong to the desired signal,
where 22 is the number of data sub-carriers, and 29 is the number of FFT-points minus
three zero-carriers.

6.5.2.2 Results over Fading Channels

The potential limitation for the QR-receiver over fading channels is the possible lack
of sufficient frequency diversity, in particular, when the channels RMS delay spread
is low. In such cases, the whole OFDM signal may be affected similarly by fades, re-
sulting in significant loss of received signal strength. Apart from extensive error bursts
introduced in such situations, the challenge is to stay synchronized. The results shown
indicate that our receiver is sufficiently robust for such situations.
Figure 6-34a shows BER-results for a fixed value of RMS delay spread (RDS) rms and
for several velocities v = {0.2, 2, 5} m/s (at 60 GHz). Figure 6-34b depicts the BER for
a fixed velocity of 2 m/s and various RDS rms = {0.23, 0.7, 1.1}. (These values relate
to the receivers sampling rate of 11.025 kHz. At the channel simulator, which is run-
ning at 48 kHz sampling rate, {1, 3, 5} samples were selected.)
Again a significant dependency on the channels time-variability is observed, for the
same reasons as given above. The average results are not significantly worse than the
results for the FR-receiver (compare Figure 6-30), however, the deviation of single
220 Chapter 6 DSP Algorithm Development for the Down-Link

BER results vs. SNR; QRreceiver; Rayleigh, = 0.7 samples; different mobility BER results vs. SNR; QRreceiver; Rayleigh, f = 2e5; w/ I/Q (VCXO)
m
rms
0 0
10 10
RDS = 0.7 samples (at RX)
RDS = 0.23 samples
RDS = 1.1 samples
analytical
1 1
10 10

average biterrorrate (BER)

average biterrorrate (BER)

2 2
10 10

3 3
10 10
w/ I/Q; v = 2 m/s
w/ I/Q; v = 5 m/s
w/ I/Q; v = 0.2 m/s
analytical
4 4
10 10
5 0 5 10 15 20 25 30 35 40 5 0 5 10 15 20 25 30 35 40
average E /N [dB] average E /N [dB]
b 0 b 0

(a) (b)
Figure 6-34: BER results vs. SNR for the QR-receiver over Rayleigh fading channels. I/Q-
(de)modulators are present, using the VCXOs. (a): Varied mobility. (b): Varied
RMS delay spread.

measurement points from the means is increased, especially, for the slowly varying
channels and for low RMS delay spread. The reason is the decreased frequency-diver-
sity for the QR-system, which leads to a rather flat fading channel over the reduced
bandwidth.

6.5.3 Discussion of the Channels Time-Variability

For both modes it is seen that the time-variability of the channel leads to irreducible
error floors, because the channel estimate obtained at the beginning of a frame is used
to demodulate all data symbols of the current frame. The accuracy of this estimate thus
decreases for data symbols that are transmitted later within the frame, i.e., with more
distance to the training symbol. Assuming that a BER of about 1 % can be still cor-
rected by the coding scheme, it is concluded that the mobility of the system is limited
to a maximum speed of about 2 m/s, which is just sufficient for indoor applications (cf.
Figure 6-30, Figure 6-32, and Figure 6-34). It is desirable to increase the robustness of
the system in this respect, in order to reduce the probability of outages and to enhance
the applicability of the system.

Computationally simple techniques for tracking/predicting the channels changes were

investigated in the graduation work by David Murargi [26].
The most basic method considered performs linear extrapolation to predict the channel
coefficient of the current data symbol, using the channel estimates from the last two
training symbols. Denoting the second last and the last estimates at carrier i as
{hi ,TS , hi ,TS } , respectively, the predicted channel coefficient is calculated from
1 0
6.5 Experimental Results 221

timevariant Rayleigh channel with velocity v at 60 GHz; linear extrapolation

0
10
v = 2 m/s
v = 5 m/s
v = 10 m/s
1
v = 15 m/s
10 theory; perfect channel est.
average biterrorrate (BER)

2
10

3
10

4
10
0 5 10 15 20 25 30 35 40
signaltonoise ratio E /N [dB]
b 0

Figure 6-35: Simulated average BER for time-variant Rayleigh fading channels with different
mobility. The channel coefficients used for coherent detection of the QPSK sym-
bols were predicted by linear extrapolation using the last two channel estimates
derived from training symbols.

hi ,k = hi ,TS0 +
k
NF
( )
hi ,TS0 hi ,TS1 , (6-49)

where k = {1, 2, 3, } is the index on the data symbol within a frame (the distance in
OFDM symbols of the current data symbol to the training symbol), and NF is the frame
length (the number of OFDM symbols in-between two training symbols). Pre-calcu-
lating the (constant) coefficients, it is seen that only two multiplications and one addi-
tion are required per predicted channel coefficient. Despite the simplicity of the
method, the performance improvement is significant, as seen from the simulation re-
sults presented below.
The relevant parameters in the computer simulation were selected according to the
OFDM system implemented on the emulation system (and thus according to the
OFDM system proposal of Section 5.2), in order to be able to compare the simulation
results to the above-presented experimental results. Only one sub-carrier was simu-
lated, having the system model of Section 4.2.3 in mind, which shows that the OFDM
sub-carriers can be considered independent under certain conditions. A time-variant
Rayleigh channel with a Jakes Doppler spectrum [27] has been implemented (cf. Sec-
tion 5.3.3).
The performance results presented in Figure 6-35 suggest that the maximum velocities
allowed have increased by a factor of about four. The significant improvement ob-
tained justifies the small additional complexity.
222 Chapter 6 DSP Algorithm Development for the Down-Link

Even bigger improvements are possible with optimized prediction techniques. For in-
stance, Wiener filters can be used to calculate the predicted channel coefficients from a
linear combination of the last L channel estimates. Simulation results from [26] show
that thereby a further enhancement of the maximum speed is possible, by a factor of
about two. The big disadvantage of this technique is that the Wiener filter is only ap-
propriate for one particular velocity. Adaptive filters are required to obtain optimum
results. Otherwise, sub-optimum performance is achieved, comparable to the estima-
tion problem discussed in Section 6.4. Moreover, the computational complexity of
these techniques is significantly higher.

6.5.3.1 Inter-Carrier-Interference due to Doppler Spread

The inter-carrier-interference (ICI) (interference among sub-carriers of the OFDM
scheme) induced by the Doppler spread of the radio channel is another issue related to
the channels time-variability. The analysis below shows that this problem is negligi-
ble for the system proposal under investigation.
In [6], [28] two approximations are given for the carrier-to-interference ratio (CIR)
due to Doppler spread, assuming the classical Jakes Doppler spectrum. The equa-
tions presented are equivalent, except for a small constant factor. We therefore restrict
the analysis to the more conservative one, which is presented in [6]. It is written as

6 F2
CIR = , (6-50)
2 f m2

where F is the frequency-spacing among sub-carriers and fm is the maximum Doppler

frequency, fm = vmfc/c (vm is the mobiles velocity, fc is the carrier frequency, and c is
the speed-of-light). This yields respectively CIR = {46, 40 dB} for transmission modes
I and II where F = {1, 0.5 MHz}, at vm = 20 m/s. Even for this velocity, being about an
order of magnitude higher than the design target, no considerable degradation is ex-
pected due to ICI, due to the relatively large sub-carrier spacing.

6.5.4 Conclusions and Recommendations

Its implementation on an experimental platform (the so-called emulation system pre-
sented in Section 5.3) strongly supports the claim that the proposed OFDM system is
suitable for the application considered, i.e., for indoor wireless LANs. Particularly the
synchronization scheme developed has proven robust and reliable over the channels
that are expected in such environments. Good performance was observed even for the
worst-case scenario, a time-variant, frequency-selective Rayleigh channel with a
delay profile about 20 % longer than the guard interval. More critical is the perform-
ance of the quarter rate-receiver. Its small bandwidth (only 22 sub-carriers are used in
mode I-qr) leads to situations where the whole signal is attenuated by a fade, because
the frequency-diversity of the wide-band signal cannot be sufficiently exploited. It is
recommended to implement frequency hopping to reduce this issue.
6.6 References 223

The small implementation losses also indicate the suitability of the emulation systems
hardware for the demonstration of OFDM systems. It has to be assumed, however, that
the emulators hardware with its downscaled bandwidth of about 40 kHz is far supe-
rior to the hardware of a real system at 60 GHz, which has a bandwidth of about 100
MHz.
One of the main impairments seen from the BER-results is due to the time-variability
of the radio channel. In the implemented scheme, the channel estimate derived at the
beginning of a frame from the known training symbol is used to demodulate the data
of 48 consecutive OFDM symbols. As the channel changes continuously, the inaccu-
racy of this estimate increases, which leads to an irreducible error floor. The maximum
velocity allowed by such an estimation scheme is in the order of 2 m/s, according to
the experimental results. This is just sufficient for an indoor system. But it may be in-
sufficient, if there are fast moving objects in the room, like machinery or automated
vehicles in a production facility. Simulation results have indicated that a basic channel
prediction scheme can increase the allowed mobility up to about 8 m/s, at a small in-
crease in complexity.

6.6 References
[1] T. M. Schmidl and D. C. Cox, Robust frequency and timing synchronization for
OFDM, IEEE Trans. Commun., vol. 45, no. 12, pp. 16131621, Dec. 1997.
[2] Y. H. Kim, Y. K. Hahm, H. J. Jung, and I. Song, An efficient frequency offset
estimator for timing and frequency synchronization in OFDM systems, in Proc.
IEEE 1999 Pacific Rim Conf. on Commun., Computers and Signal Proc., pp.
580583.
[3] D. Matic, N. Petrochilos, A.J.R.M. Coenen, F. Schoute, R. Prasad, Acquisition
of synchronisation parameters for OFDM using a single training symbol, in
Proc. Second International Workshop on Multi-Carrier Spread-Spectrum and
Related Topics, Oberpfaffenhofen (Germany), Sept. 1999.
[4] P. H. Moose, A technique for orthogonal frequency division multiplexing fre-
quency offset correction, IEEE Trans. Commun., vol. 42, no. 10, pp. 2908
2914, Oct. 1994.
[5] J.-J. van de Beek, M. Sandell and P. O. Brjesson, ML estimation of time and
frequency offset in OFDM systems, IEEE Trans. Signal Processing, vol. 45, no.
7, pp. 18001805, July 1997.
[6] M. Speth, S. A. Fechtel, G. Fock, and H. Meyr, Optimum Receiver Design for
Wireless Broad-Band Systems Using OFDMPart I, IEEE Trans. Commun.,
vol. 47, no. 11, pp. 16681677, November 1999.
224 Chapter 6 DSP Algorithm Development for the Down-Link

[7] B. McNair, L. J. Cimini, Jr., N. Sollenberger, A Robust Timing and Frequency

Offset Estimation Scheme for Orthogonal Frequency Division Multiplexing
th
(OFDM) Systems, in Proc. IEEE 49 Vehicular Techn. Conf. (VTC99-spring),
Houston, USA, May 1999, pp. 690694.
[8] N. Petrochilos, A. Trindade, D. Matic, and R. Prasad, Blind Channel Synchroni-
zation and Equalisation Algorithm for OFDM Wireless Communications, in
Proc. WPMC 2001, Aalborg, Denmark, Sept. 2001.
[9] D. Lee and K. Cheun, A new Symbol Timing Recovery Algorithm for OFDM
Systems, IEEE Trans. Consumer Elec., vol. 43, no. 3, pp. 767775, Aug. 1997.
[10] D. K. Kim, S. H. Do, H. B. Cho, H. J. Choi, and K. B. Kim, A New Joint Algo-
rithm of Symbol Timing Recovery and Sampling Clock Adjustment for OFDM
Systems, IEEE Trans. Consumer Elec., vol. 44, no. 3, pp. 11421149, Aug.
1998.
[11] D. Matic, N. Petrochilos, A. Trindade, F. Schoute, P. Comon, and R. Prasad,
OFDM Synchronization Based on the Phase-Rotation of Sub-Carriers, in Proc.
IEEE 51st Vehicular Techn. Conf. (VTC 2000-spring), Tokyo, Japan, May 2000,
pp. 12601264.
[12] J. G. Proakis, Digital Communications, 3rd Edition. New York: McGraw Hill,
1995.
[13] H. Minn and V. K. Bhargava, A simple and efficient timing offset estimation for
st
OFDM systems, in Proc. IEEE 51 Vehicular Techn. Conf. (VTC 2000-spring),
Tokyo, Japan, May 2000, pp. 5155.
[14] P. Teneva, Pseudo Real-time Simulation of an OFDM System for Wireless
Broadband Communications: OFDM Transmitter, Research Report, IRCTR,
Delft University of Technology, Jan. 1999.
[15] K. Bke, Assessment of OFDM Based Air-interface Techniques Using an
Emulation Platform: Investigation and Implementation of OFDM Synchroniza-
tion Algorithms, M.Sc. Thesis, IRCTR S-001-00, Delft University of Technolo-
gy, Jan. 2000.
[16] I. Gultekin, DSP Software Implementation for a Broadband Air-interface
Emulation Platform: Onderzoek naar de serile interace tussen de PC an de DSP
board en het maken van een GUI voor de DSP applicaties, Graduation Thesis,
Haagse Hogeschool, June 2000 (in Dutch).
[17] A. Snijders, Emulator: De testopstelling voor OFDM, Graduation Thesis,
IRCTR S-030-00, Technische Hogeschool Rijswijk, Nov. 2000 (in Dutch).
[18] P. Pollet, M. van Bladel, and M. Moeneclaey, BER sensitivity of OFDM sys-
tems to carrier frequency offset and Wiener phase noise, IEEE Trans. Commun.,
vol. 43, no. 2/3/4, pp. 191193, Feb./March/April 1995.
6.6 References 225

[19] K. Witrisal, K. Bke, Y.-H. Kim, R. Prasad, and L. P. Ligthart, Air-interface

Emulation for Wireless Broadband Communications Applied to OFDM, in
Proc. IEEE Symposium on Personal Indoor and Mobile Radio Communications
(PIMRC'2000), London, Sept. 2000, pp. 12511255.
[20] K. Witrisal, Y.-H. Kim, R. Prasad, and L. P. Ligthart, Experimental Study and
th
Comparison of OFDM Transmission Techniques, in Proc. 5 international
OFDM-Workshop, Hamburg (Germany), Sept. 2000, pp. 5-15-5.
[21] O. Edfors, Low-complexity algorithms in digital receivers. Ph.D. Thesis, Lule
University of Technology, Sept. 1996.
[22] M. Sandell, Design and Analysis of Estimators for Multicarrier Modulation and
Ultrasonic Imaging. Ph.D. Thesis, Lule University of Technology, Sept. 1996.
[23] P. K. Frenger, N. Arne, and B. Svensson, Decision-directed coherent detection
in multicarrier systems on Rayleigh fading channels, IEEE Trans. Veh. Tech-
nol., vol. 48, no. 2, pp. 490498, March 1999.
[24] L. L. Scharf, Statistical Signal Processing: Detection, Estimation, and Time Se-
ries Analysis, Addison-Wesley, 1991.
[25] C.W. Therrien, Discrete Random Signals and Signal Processing. Englewood
Cliffs: Prentice Hall, 1992.
[26] D. Murargi, Channel Estimation Enhancement in OFDM Systems for Wireless
Multimedia Communications, Graduation Thesis, Delft Univ. of Technol., Rep.
Nr. IRCTR: S-023-01, July 2001.
[27] W. C. Jakes Jr., Microwave Mobile Communications. New York: Wiley-Inter-
science, 1974.
[28] Y. Zhao and S. Haggman, Sensitivity to Doppler Shift and Carrier Frequency
Errors in OFDM Systems The Consequences and Solutions, in Proc. IEEE
46th Vehicular Techn. Conf. (VTC 1996), Atlanta, GA, USA, April/May 1996, pp.
15641568.
Chapter 7 DSP Algorithm Development
for the Up-Link

7.1 Introduction
In this chapter, the potential of a pre-equalization scheme for the up-link of a time di-
vision duplex (TDD) OFDM system (see Section 5.2) is investigated. The application
of this principle is suggested to enhance the spectral efficiency by omitting the usage
of training sequences on the up-link. This is desired in order to enable the efficient
transmission of very short data packets, like single ATM cells, without adding the
overhead of a training symbols to each packet. Moreover, at the base station, the com-
plexity of synchronization is kept low, which is of advantage considering that the base
station has to serve multiple mobile users at extremely high data rates.
The main purpose of applying pre-equalization is to enable coherent detection without
channel estimation. With pre-equalization, the frequency-selective channel is turned
flat or phase linear, depending on the pre-equalization technique implemented and on
the modulation scheme. Basic strategies are described to thereby keep the transmission
power constant. In one of the schemes described, for instance, this is achieved by at-
tenuating strong sub-carriers while weak ones are boosted. Sub-carriers that are very
severely attenuated by fades are set to zero in order to limit the overall attenuation.
Synchronization steps for retrieving the data at the base-station are explained and
evaluated. Those include the estimation of the timing-offset, the carrier phase offset,
and the received magnitude.

In Section 7.2, the concept of pre-equalization is introduced and discussed, and the
mathematical models of OFDM are reviewed. Section 7.3 studies the power-limiting
techniques, followed by a number of performance results in Section 7.4. The synchro-
nization steps required for the coherent detection of the up-link data streams are out-

227
228 Chapter 7 DSP Algorithm Development for the Up-Link

lined and coarsely evaluated in Section 7.5. Section 7.6 shows experimental perform-
ance results derived with the emulation platform of Section 5.3. Conclusions are drawn
in Section 7.7.

7.2 Pre-Equalization in OFDM

7.2.1 OFDM System Model

Assuming an OFDM system using a guard interval (GI), a dispersive radio channel
with a maximum excess delay being shorter than this GI, and perfect synchronization,
the OFDM system can be modeled as a set of parallel Gaussian channels (see Section
4.2.3). On the input of this system model, which is written as
yi ,k = xi ,k hi ,k + ni ,k , (7-1)

are the symbol constellation points {xi,k} that are applied to the IFFT block of the
transmitter; at its output are the constellation points {yi,k} after the receivers FFT. i is
the index on the subcarrier (SC), being proportional to the SCs distance to the center
frequency, and k is the OFDM symbol index, a time-variable. The attenuation factors
{hi,k} and the AWGN samples {ni,k} denote the channels influence on the transmitted
symbol constellation points. Since these factors are complex-valued, each constellation
point typically suffers attenuation according to the magnitude of {hi,k}, and phase rota-
tion according to its phase.
Note that the attenuation factors seen as a function of frequency and time are equiva-
lent to the transfer function of the channel hi,k = H(iF,kT). (F is the frequency-spacing
of adjacent SCs; T is the total OFDM symbol period.) In a real environment, this is
usually a time-variant function. Within the transmission of a few subsequent OFDM
symbols however, the channel can be considered time-invariant. Therefore, the time-
index k will be neglected in the mathematical expressions describing the pre-equaliza-
tion scheme.

7.2.2 Impact of Synchronization Errors

Small synchronization errors, i.e., timing errors t in the order of a few samples (<
length of the GI), frequency offsets f of a few percent of F, and carrier phase offsets
introduce additional phase rotations (and inter-carrier-interference) (see Section 4.2.4).
Incorporating these offsets, the system model becomes
ji , k
yi ,k = xi ,k hi , k sinc(f TFFT ) e + n'i ,k , (7-2)

T i
with i ,k = + 2f kT + FFT + t + 2t . (7-3)
2 TFFT
7.2 Pre-Equalization in OFDM 229

TFFT is the effective FFT period, the part of the (received) OFDM symbol applied to
the FFT for demodulation, and n'i,k is the noise term incorporating inter-carrier-inter-
ference in case of a frequency-offset.

7.2.3 Channel Reciprocity and Pre-Equalization

The assumption of reciprocity in the context of an OFDM system means that a SC
transmitted on the up-link at a certain frequency suffers the same attenuation and
phase rotation as a SC transmitted on the down-link at the same frequency. In other
words, the channel transfer functions on the up-link and down-link are equivalent.
The idea of pre-equalization seeks to exploit this property in the context of a time-divi-
sion duplex (TDD) scheme. The receiver performs channel estimation and synchroni-
zation on the down-link, obtaining an estimate of the channel attenuation factors {hi } .
(Note that the time-index k is being omitted from here on.) Multiplying the constella-
tion points of the up-link data symbols with the inverse of these factors {1 / hi } prior to
applying them to the transmitters IFFT will cancel the signal distortion by the up-link
channel, thereby simplifying the detection of the data at the base-station.
In such a system, down-link and up-link data packets will be transmitted at the same
frequency band in an alternating pattern (see Section 5.2). It is obvious that the chan-
nel should not change after its estimation on the down-link until the estimate is used
on the up-link for pre-equalization. Small variation leads to error floors in uncoded
BER-results, as seen from the experimental data presented in Section 7.6.
Another problem of pre-equalization is the large power required for faded SCs. As-
sume for instance, a SC is in a deep fade, i.e., it is attenuated by say 30 dB compared
to the average SC-energy. In order to compensate for this attenuation, the SC would
have to be boosted by 30 dB, requiring a large amount of transmission power and a
high DAC-resolution at the mobile21. Section 7.3 describes approaches of limiting the
additional transmit power needed.

7.2.4 Is the Channel Reciprocal?

Reciprocity is a property of the radio channel, which is commonly accepted by radio
engineers. It has been used in several system proposals and theoretical studies, because
it leads to a number of potential advantages [1][6]. In TDD-code division multiple
access (CDMA) systems, for instance, pre-rake filtering has been proposed for the
transmitter, which can eliminate the need for a rake receiver at the mobile (if applied
for the down-link), or it can simplify multi-user detection at the base station [2][4].
Other advantages of TDD schemes include antenna diversity or beam-forming tech-

21
Another reason for why it is problematic to do perfect pre-equalization for such deeply faded SCs
is that their phases typically change quickly during deep fades. Therefore, even if the transmit power
can be provided, the phase may be wrong, leading to a wrong decision (cf. [1]).
230 Chapter 7 DSP Algorithm Development for the Up-Link

niques, where, at the base-station, a combining scheme optimized for the up-link can
be used on the down-link as well. TDD schemes can also apply closed loop power
control [2], [5].
Some studies, however, report on non-reciprocal behavior of the mobile radio channel
[7], [8], which would eliminate the above-mentioned advantages. In [7], electromagne-
tic coupling effects are suggested as the source for asymmetries. Concerning interfer-
ence effects, there is no doubt that differences exist between the up- and down-links.
Asymmetries may also exist in the RF front-ends between reception and transmission.
These fixed differences can be compensated by means of calibration. It is possible that
such asymmetries and the required calibration may be affected by electromagnetic
coupling.
At the time being, we cannot draw final conclusions on the possibility of exploiting
channel reciprocity in the proposed transmission scheme. The potential advantages of-
fered by such a scheme, however, have motivated the study presented in this chapter.

7.3 Power Limiting Strategies for Pre-equalization

This section describes strategies of limiting the power amplification due to pre-equali-
zation. We will seek constant output power after pre-equalization.

7.3.1 Phase Pre-Equalization for Phase-Modulation Schemes

For pure phase modulation schemes, like QPSK, it is sufficient to perform the pre-
equalization for the phase of the constellation values, turning the channel phase-linear.
This means multiplying the data constellation points d i {e j 2m / M } (m = 0, 1, M 1
for M-ary modulation) with a phase factor equal to the inverse of the channels phase
rotation:

xi = d i e j hi (7-4)
Applying these symbols to the parallel Gaussian channels of the OFDM system model
(7-1) yields at the receiver
j hi hi hi ]
yi = d i e hi + ni = d i hi e + ni .
j[
(7-5)

Assuming that the channel estimates phase has been correct, the phase distortion term
becomes zero. Thus the received symbols suffer from attenuation according to the
magnitudes {|hi|} and from additive noise. Figure 7-1 depicts such distorted symbol
constellations for QPSK-modulated SCs. Note that there will be systematic phase rota-
tions in the presence of synchronization errors, which can be estimated from pilot SCs
or from the limited symbol alphabet properties of the data symbols. These systematic
phase distortions must be corrected.
Performance results are shown and discussed in Section 7.4.1.
7.3 Power Limiting Strategies for Pre-equalization 231

signal constellations with QPSK and phase preequalization

imaginaryvalued component 1

0.5

0.5

1 0.5 0 0.5 1 1.5

realvalued component

Figure 7-1: Received signal constellation points with phase pre-equalization for QPSK. Addi-
tive noise is present but no synchronization-offsets on the up-link.

7.3.2 Phase and Magnitude Pre-Equalization with Power Limiting

Higher order modulation schemes, as 16-QAM, need phase and magnitude estimates
for simplifying their coherent detection. Accordingly, pre-equalization must also com-
pensate for the different attenuation factors at the OFDM-SCs.
Ideal pre-equalization is written mathematically as

xi = d i ei , ei = hi1 , (7-6)
where the {ei} denote the pre-equalization function being simply the inverse of the
channel estimate, and the {di} stand for the constellation points of the modulation
scheme. As mentioned above, this technique might significantly boost the average
power of the constellation points {xi}, causing the output power to vary largely de-
pending on the current channel. Since such effects are not desirable in digital commu-
nications systems, methods are proposed to limit or eliminate them.

7.3.2.1 Normalization
The method presented in this section normalizes the pre-equalization function, result-
ing in constant transmitted power. This normalization is achieved by deviding the {ei}
by their RMS value, written
232 Chapter 7 DSP Algorithm Development for the Up-Link

N SC N SC
e~i = ei = hi1 , (7-7)
| ei |
iiSC
2
| hi1 |2
iiSC

where NSC is the total number of used (data & pilot) SCs, and iSC are their indices. This
normalization will actually lead to an attenuation of the average received power. While
SCs being largely attenuated by the channel are boosted by the pre-equalization tech-
nique, the power of strong sub-carriers must be reduced to keep the transmitted power
constant. It is seen from eq. (7-7) that very small channel estimates {hi } will have a
large contribution to the RMS value of the normalization term, leading to greater at-
tenuation on the average. We will seek to limit the maximum attenuation.
The attenuation factor due to pre-equalization is defined as the ratio of the normalized
received powers with and without pre-equalization,
P0
Apre = . (7-8)
P0, pre

Without pre-equalization, the normalized received power is calculated from the mean
of the power transfer function (TF) at the used sub-carriers (cf. eq. (7-1)),
Prx 1
P0 = = h
2
i , (7-9)
Ptx N SC iiSC

where Prx and Ptx are the received and transmitted powers. Note that xi = di without
pre-equalization.

Substituting (7-6) with (7-7) for xi in the OFDM system model (7-1) yields the constel-
lation points at the receiver with pre-equalization as

N SC
yi = d i hi e~i + ni = d i hi hi1 + ni . (7-10)
hi1
2

iiSC

Assuming that the channel estimate is equal to the actual channel, hi = hi , the normal-
ized received power with pre-equalization becomes
Prx , pre N SC
P0, pre = = . (7-11)
hi1
2
Ptx , pre
iiSC

Note that the transmitted power with pre-equalization, Ptx,pre is equal to Ptx, due to the
normalization. Therefore, the ratio of the two normalized received powers is the aver-
age attenuation or power loss due to pre-equalization,
7.3 Power Limiting Strategies for Pre-equalization 233

h h e h
2
1 2 2 2
i i i i
P0 ii iiSC iiSC ii SC
Apre = = SC 2
= 2
. (7-12)
P0, pre N SC N SC

This factor is generally larger than one, due to the amplification effect of the 1/x op-
eration on small arguments. (It is equal to one if the channel is flat). It can be calcu-
lated by the mobile terminal, using the down-link channel estimate and assuming
hi = hi .

7.3.2.2 Limiting the Attenuation

The second strategy of the proposed pre-equalization scheme is the limitation of this
factor. It is accomplished by zeroing the pre-equalization function {ei} for the most
attenuated SC, until the attenuation falls below the threshold Amax. The following steps
are executed repeatedly:
Calculate the attenuation factor Apre using (7-12). Stop the procedure if Apre Amax.
Find the index of the most attenuated sub-carrier imin = arg min{| hi |} and set the pre-
i

equalization function for this sub-carrier to zero, ei = 0 . min

Continue with the first step.

I.e., no power at all is put on those SCs being most severely attenuated by the channel.
The data transmitted on these SCs will be lost, which the coding scheme has to correct
for. Figure 7-2 illustrates this principle. The power normalization using eq. (7-7) is
done after limiting the attenuation factor. Performance results are shown in Section

preequalization for uplink; method 0; attenuation factor = 4.8 dB

3
preeq. TF, normalized
channel estimate
2.5 received carrier magnitudes
TF magnitude [linear]

1.5

0.5

0
0 10 20 30 40 50 60 70 80 90
subcarrier index

Figure 7-2: Attenuation limiting for pre-distortion by setting the most attenuated SCs to zero.
234 Chapter 7 DSP Algorithm Development for the Up-Link

7.4.2.
Keller and Hanzo [1] have suggested a similar method of pre-equalization. However,
they do not consider power-normalization. Instead, they do allow the output power to
vary to a certain extent. Our approach has the advantage that the requirements on the
power amplifier, which is a critical element of OFDM systems, are not enforced.

7.4 Performance Results

The performance results discussed in this section were obtained from computer simu-
lations. (Experimental results are presented in Section 7.6.) Transmission mode I-f (see
Section 5.2) was investigated for all results shown, where a 128-point FFT is em-
ployed, and 76 data sub-carriers and 12 pilot sub-carriers are used. The following ele-
ments of the transmission system have been considered or neglected.
- Channel estimation on the down-link
Has been considered. The channel estimate is required in order to calculate the pre-
equalization function {ei}. The channel estimation scheme implemented is the one
described in Section 6.4. Seven-tap filters have been designed for a fixed set of
channel parameters, namely: Rice-factor K = 0 (Rayleigh channel); RMS delay
spread rms = 3 samples; and SNRSC = 30 dB (SNR at the SCs of the TS). The bias
of the down-link timing-offset estimation was considered as well, being t' = 3
samples for these channel parameters.
- Synchronization on the down-link
Not considered. Small timing-offsets on the down-link will be included in the
channel estimate and thereby corrected. Frequency-offsets larger than ~2 % would
result in performance degradations as mentioned and evaluated in Section 4.2.4.
- Time-variability of the channel
Not considered. Time-variations in between the channel estimation and the trans-
mission of the up-link OFDM symbols will lead to non-ideal pre-equalization and
further performance degradation [1]. In these simulations we will concentrate on
the performance achievable in a static environment. Time-variability was present in
the performance results derived from the experimental platform (see Section 7.6).
- Estimation of the constellation points magnitude on the up-link
Is considered in some of the simulations for the 16-QAM modulation scheme. The
performance loss resulting from estimating this parameter from a single data sym-
bol is evaluated in Section 7.5.1.
- Up-link timing-offset estimation
Is considered in some of the simulations. The performance loss resulting from es-
timating this timing offset using the pilots of only one OFDM symbol is evaluated
in Section 7.5.2.
7.4 Performance Results 235

QPSK; phase preequalization; perfect correction for timingoffsets

0
10

1
10

2
average BER

10

3
10
K = 0; rms = 1 sample
K = 0; rms = 3 samples
K = 0; rms = 5 samples
4
10 K = 4; rms = 1 sample
theory; K = 4
theory; K = 0
5
10
5 0 5 10 15 20 25 30
average Eb/N0 [dB]

Figure 7-3: Performance of phase pre-equalization with QPSK under the assumptions of ideal
up-link timing-offset correction and a time-invariant channel.

7.4.1 Application of Phase Pre-Equalization for QPSK

For pure phase modulation schemes, phase pre-equalization can be applied on the up-
link, as described in Section 7.3.1.
Assuming that the up-link timing-offset can be estimated and compensated perfectly,
and assuming a time-invariant channel, the performance on the up- and down-link is
equivalent. Simulation results for several channels are shown in Figure 7-3. Note that
up- and down-link results cannot be distinguished since they lie exactly on top of each
other. As a benchmark, the theoretical performance is depicted that could be achieved
with perfect channel estimation and synchronization. The small degradation of the
simulation results is caused by the channel estimation scheme that has been taken into
consideration (see Section 6.4).
A certain performance improvement for the up-link can be achieved by optimizing the
power assignment to individual sub-carriers. In [9], the optimum assignment is studied
for minimizing the average bit-error rate of an OFDM system. If applied with phase-
modulation schemes, no modifications at the receiver are needed. (See also [10].)

7.4.2 Application of Pre-Equalization with Power Limiting

The pre-equalization scheme with limitation of the output power (see Section 7.3.2) is
evaluated in this section. Up to a certain extent, namely as long as the power limitation
criterion is not violated, also the attenuation magnitudes are pre-equalized. This is re-
quired for efficient detection of higher order modulation schemes, like 16-QAM.
236 Chapter 7 DSP Algorithm Development for the Up-Link

Note that the power attenuation due to pre-equalization is considered in the SNR-val-
ues of all simulation results to allow a fair comparison with down-link results. If the
SNR would be measured at the receiver, the BER-curves would shift to the left, sug-
gesting much better performance vs. Eb/N0.

7.4.2.1 Results for QPSK Modulation

Considering the QPSK modulation scheme, it is of interest whether the average BER
can be improved by applying phase and magnitude pre-equalization. Such a behavior
may be anticipated, since more energy can be put on SCs that are attenuated by the
channel, while less is needed on reliable SCs. The impact of the maximum attenuation
factor allowed is investigated. BER results for several channels and several maximum
attenuation factors are depicted in Figure 7-4.

Rayleigh ch., = 3 samples; preequalization method 0 Rayleigh ch., = 1 sample; preequalization method 0
rms rms
0 0
10 10

1 1
10 10
average BER

average BER

2 2
10 10

uplink; atten. = 3 dB uplink; atten. = 3 dB

3 3
10 uplink; atten. = 7 dB 10 uplink; atten. = 7 dB
uplink; atten. = 10 dB uplink; atten. = 10 dB
uplink; atten. = 13 dB uplink; atten. = 13 dB
downlink downlink
theory; perf. channel est. theory; perf. channel est.
4 4
10 10
5 0 5 10 15 20 25 30 5 0 5 10 15 20 25 30
average Eb/N0 [dB] average Eb/N0 [dB]

(a) (b)
Rayleigh ch., rms = 5 samples; preequalization method 0 Ricean ch., rms = 1 sample, K = 4; preequalization method 0
0 0
10 10

1
10
1
10

2
average BER

average BER

10
2
10
3
10

uplink; atten. = 3 dB uplink; atten. = 3 dB

3
10 uplink; atten. = 7 dB uplink; atten. = 7 dB
uplink; atten. = 10 dB 4 uplink; atten. = 10 dB
10
uplink; atten. = 13 dB uplink; atten. = 13 dB
downlink downlink
theory; perf. channel est. theory; perf. channel est.
4 5
10 10
5 0 5 10 15 20 25 30 5 0 5 10 15 20 25 30
average Eb/N0 [dB] average Eb/N0 [dB]

(c) (d)
Figure 7-4: Performance results for up-link pre-equalization for QPSK. Deletion of SCs is
used for power limitation. (a), (b), (c): Rayleigh channel with rms = {3, 1, 5} sam-
ples, respectively. (d): Ricean channel with K = 4 and rms = 1 sample.
7.5 Synchronization Parameter Estimation on the Up-link 237

The general trend of all these plots is similar. At low SNR, the results are worse than
the down-link results due to the attenuation factors considered in the Eb/N0 values. At
medium SNRs, the curves typically show a steeper slope than the down-link curves,
because the channel including pre-equalization approximates an AWGN channel. The
noise floors seen at the high SNR-region are due to the power limitation method set-
ting some SCs to zeros.
Higher maximum attenuation factors result in slightly larger loss at low SNR but also
in lower noise floors at high SNR. At certain points, the up-link outperforms the down-
link. To find an optimum value for the maximum attenuation factor, error correction
coding should be considered too. If the FEC-scheme can perfectly correct an average
channel BER of 10-2, there is no need to select the attenuation factor larger than 7 dB,
following Figure 7-4. Note however that channel variability is not considered.
The behavior is largely independent of the channel parameters as seen from comparing
the plots among each other.

7.4.2.2 Results for 16-QAM

Pure phase pre-equalization cannot be applied for higher order modulation schemes
having amplitude and phase modulation components. The proposed pre-equalization
with attenuation limiting and constant, normalized transmit power is evaluated in this
section for 16-QAM. Figure 7-5 shows computer simulation results for various
channel parameters.
Again, the influence of the maximum attenuation factor is depicted. The results and
conclusions are quite consistent to the results for QPSK. Around the typical working
point of the FEC-decoder (BER ~ 10-2), the up-link performance is similar to the
down-link. (Note again that the attenuation factor is considered in the average Eb/N0
values shown.) The Eb/N0 required is approx. 3 dB higher than for QPSK, translating
to a required increase in transmission power of approx. 6 dB, because four bits are
transmitted per sub-carrier in stead of two in the QPSK case. (Eb is the symbol energy
per bit see Section 4.3.3, eq. (4-44)).

7.5 Synchronization Parameter Estimation on the Up-link

Pre-equalization can transform a frequency-selective fading channel into a flat one
(with some restrictions). This can simplify the detection of the data on the receivers
site. Carrier- and sampling-frequency synchronization can also be done by the mobile,
based on estimates from the down-link. A few parameters still need to be estimated by
the receiver (the base station) in order to perform coherent detection. Those are:
timing offset,
carrier phase offset, and
(the common) magnitude of the constellation points
238 Chapter 7 DSP Algorithm Development for the Up-Link

Rayleigh ch., rms = 3 samples; preequalization method 0 Rayleigh ch., rms = 1 sample; preequalization method 0
0 0
10 10

1
10
1
10
average BER

average BER
2
10

2
10
uplink; atten. = 3 dB
3
10 uplink; atten. = 7 dB uplink; atten. = 3 dB
uplink; atten. = 10 dB uplink; atten. = 7 dB
uplink; atten. = 13 dB uplink; atten. = 10 dB
downlink uplink; atten. = 13 dB
theoretical; perfect ch. est. downlink
4 3
10 10
5 0 5 10 15 20 25 5 0 5 10 15 20 25 30
E /N average E /N [dB]
b 0 b 0

(a) (b)
Rayleigh ch., rms = 5 samples; preequalization method 0 Ricean ch., rms = 1 sample, K = 4; preequalization method 0
0 0
10 10

1
10

1
10
2
average BER
average BER

10

3
10
2
10
uplink; atten. = 3 dB
uplink; atten. = 3 dB uplink; atten. = 7 dB
4 uplink; atten. = 10 dB
uplink; atten. = 7 dB 10
uplink; atten. = 10 dB uplink; atten. = 13 dB
uplink; atten. = 13 dB downlink
downlink theoretical; perf. ch. est.
3 5
10 10
5 0 5 10 15 20 25 30 5 0 5 10 15 20 25
average Eb/N0 [dB] average Eb/N0 [dB]

(c) (d)
Figure 7-5: Performance results for up-link pre-equalization with 16-QAM. Deletion of SCs is
used for power limitation. (a), (b), (c): Rayleigh channel with rms = {3, 1, 5} sam-
ples, respectively. (d): Ricean channel with K = 4 and rms = 1 sample.

These parameters depend on the distance between the two transceivers and on the
propagation environment. Therefore, they are different for each mobile accessing the
base station and they are time-variant to various degrees. This section describes meth-
ods to estimate them.

7.5.1 Magnitude of the Constellation Values

Pre-equalization can make the channel flat, however, it can not correct for the (loca-
tion-dependent) path loss. Higher order modulation schemes consist of phase and am-
plitude modulation components, therefore, a magnitude reference is needed for their
coherent detection.
The magnitude reference can be estimated from the average power of the data and
7.5 Synchronization Parameter Estimation on the Up-link 239

Rayleigh ch., = 3 samples; atten. = 7 dB; Method 0; 16QAM

rms
0
10

1
10
average BER

2
10

estimation of atten. from data; perfect timing

perfect attenuation factor; perfect timing
perfect atten.; timing estd. from one data symb
downlink
3
10
5 0 5 10 15 20 25
average E /N [dB]
b 0

Figure 7-6: BER-result for the up-link and for 16-QAM with and without magnitude and tim-
ing-offset estimation (solid lines). The dashed lines indicate the down-link per-
formance, being equivalent for the three cases compared.

pilot SCs. A problem arises, however, when the channel noise is large, because the
power average will then include the noise power. The estimates are thus biased. The
training symbol may be used to estimate the SNR (see [11]) in order to correct the
power estimate.
However, even if the biased estimate obtained from the power average is used, it is
found that the BER-performance results are sufficiently close to the reference results,
where perfect knowledge of the received signals magnitude is assumed. An example
result for 16-QAM is depicted in Figure 7-6. Details on the computer simulation can
be found in Section 7.4.

7.5.2 Estimation of the Up-link Timing-Offset

Timing errors introduce a so called progressive phase rotation to the signal constel-
lation points, i.e., a phase rotation proportional to the frequency distance of the SCs to
the center frequency (see Section 7.2.2). The receiver can estimate the timing-offset
and then correct for it by multiplying the constellations with a complex exponential
sequence corresponding to the inverse of this systematic phase rotation. The informa-
tion on the timing-offset must be fed back to the transmitter in order to optimize the
arrival time of the up-link OFDM symbols of different users, and to avoid interference
among users and inter-symbol-interference.
240 Chapter 7 DSP Algorithm Development for the Up-Link

7.5.2.1 Up-link Timing Synchronization Using the Signaling Symbol

The first data symbol transmitted by any (mobile) user on the up-link must be a user
registration or request for transmission capacity message, sent during the OFDM
symbol for up-link signaling. This is a reserved OFDM symbol right in the beginning
of the up-link portion of the frame format (see Section 5.2). The mobile may only
transmit in this time-slot after it has achieved synchronization to the base station.
The limited symbol alphabet of the BPSK modulation (M = 2) used on this OFDM
symbol enables extracting the timing offset without knowing the data transmitted. The
principle is similar to the principle of fine timing-offset estimation proposed in Section
6.2.7. Differential demodulation between adjacent SCs is performed, yielding in the
noiseless case (with (7-2) and (7-3), where the time-index k is omitted, and with
(7-10))

yi ,1 = yi yi*1 = {1}hi ~
ei hi*1~
ei*1 sinc 2 (f TFFT )e j ( i i 1 ) , (7-13)

This expression can be simplified to

j
yi ,1 = {1} A e ( i i 1 ) , (7-14)

assuming small frequency-offset f and negligible channel-variability. The (real-val-

ued) factor A accounts for the attenuation by the channel and by the pre-equalization; it
has a value 1. The angle-shift of these differentially demodulated constellation val-
ues from the real axis (i.e., from {A}) is a measure for the timing offset, being
t[ s ] t '[ samples ]
1 = i i 1 = 2 = 2 . (7-15)
TFFT N

In order to enhance the SNR of this measure, the modulation of {1} is removed from
the {yi,1} by making a decision on the unknown data, and subsequently adding up all
the resulting complex numbers. The sum has the same phase rotation, however much
bigger amplitude. A performance evaluation is shown in Section 7.5.2.3.
The unambiguous estimation range of this technique is theoretically N/2 samples,
where N is the number of FFT-points. However, at such large offsets the uncertain de-
cisions degrade the performance since the differentially demodulated constellation
points {yi,1} get rotated closely towards the decision threshold.
Note that the decision feedback step can be replaced by squaring the {yi,1}, which
may be preferable at large time-offsets. The performance of the two methods should be
compared.

7.5.2.2 Up-Link Timing-Offset Estimation Using the Pilot Sub-Carriers

The principle of timing-offset estimation using the pilots is similar to the principle of
the above described method and the fine timing-offset estimation method using the
training symbol (see Section 6.2.7). The advantage of taking the pilots lies in the fact
7.5 Synchronization Parameter Estimation on the Up-link 241

that their constellation values are known, in contrast to the previous method, where the
data is first removed based on an uncertain decision. Again, the first step is the calcu-
lation of phase differences between pilot-pairs by means of complex multiplication
with the complex conjugate. The purpose is the removal of common phase rotations,
which may be due to carrier phase or frequency-offsets. The phase-difference between
SCs spaced in frequency by l SC-intervals is written
t[ s ] t '[ samples ]
l = i i l = 2 l = 2 l, (7-16)
TFFT N

yielding the time-offset t [s] or t' [samples].

It is seen that the unambiguous range depends on the spacing l between two pilot SCs.
It is t unamb
'
. [ samples ] = N ( 2l ) , which is less than above, limiting the applicability of

the method to tracking.

7.5.2.3 Performance of the Two Estimation Schemes

We try to compare computer simulation results to theoretical results. An analytical ex-
pression is available, quantifying the standard deviation of the timing-offset estimation
error [samples] over an AWGN channel and for known pilot data (see Appendix C-3).
This performance measure is related to the SNR and OFDM parameters as

N SNRSC + N m / 2 N
t ' = , with SNRSC = SNR , (7-17)
2 l N m SNRSC N SC

where SNR is the signal power to noise power ratio of the time-domain signals, SNRSC
is the SNR of the data/pilot SCs after the FFT, N, NSC, and Nm are respectively the
number of FFT points, the number of active (data + pilot) SCs, and the number of dif-
ferentially demodulated pilot-pairs being added up.
Unfortunately, the assumptions of an ideal AWGN channel and of perfectly known
data (in the decision feedback case) are not fully met. The channel including pre-dis-
tortion attenuates the signal by a certain factor, which may vary according to the cur-
rent channel conditions. Moreover, some of the SCs may have been set to zero in order
not to exceed the maximum attenuation factor allowed, and there are also channel es-
timation errors on the down-link.
Computer simulation results are presented for mode I-fr of the proposed OFDM sys-
tem (see Section 5.2), and compared to results derived from eq. (7-17). To obtain the
analytical results for the timing estimation technique using the pilots, a constant pilot
spacing of l = 8 is chosen, which is not equal but very close to the non-uniform pilot-
separations in the proposed OFDM system (see Section 5.2.4.2).
The performance is shown in Figure 7-7 for maximum attenuation factors of 3 and 7
dB. In the 3-dB case, simulation results are mostly worse than the theory, because
quite many SCs have been set to zero in order to stay below the low maximum at-
242 Chapter 7 DSP Algorithm Development for the Up-Link

analysis of timing offset estimation; max. atten. factor = 3 dB analysis of timing offset estimation; max. atten. factor = 7 dB
1 2
10 10
standard dev. of timingoffset estimate [samples]

standard dev. of timingoffset estimate [samples]

estimated from signaling symbol estimated from signaling symbol
estimated from pilots estimated from pilots
theory; signaling symbol theory; signaling symbol
theory; pilots theory; pilots
1
10
0
10

0
10

1
10
1
10

2 2
10 10
0 5 10 15 20 25 30 0 5 10 15 20 25 30
SNR [dB] SNR [dB]

(a) (b)
Figure 7-7: Standard deviation of timing-offset estimation error in samples. Computer simula-
tions and theoretical results. (a): Maximum attenuation factor Amax = 3 dB. (b):
Amax = 7 dB.

tenuation factor. From the results shown for the pilot-based method in the 7-dB plot,
just the opposite situation is seen. The simulation results are better than the theoretical
ones, because the actual attenuation factor is likely to be smaller than the maximum
one considered in the theoretical plot.
Generally, the results suggest that at an SNR of about 10 dB, timing offsets can be es-
timated with an accuracy better than 1 sample. This is sufficient for the optimization
of the up-link transmission time-instants. But is such accuracy also sufficient for co-
herent demodulation on the up-link?
Including up-link timing-offset estimation in the BER simulation leads to some per-
formance degradation at low SNR, where the estimation errors become significant.
Figure 7-8 shows results for QPSK, using pure phase pre-equalization. The pilots of
only one OFDM symbol were used to estimate the timing-offset. The down-link
results can be used for comparison, since they dont change (cf. Figure 7-3). Note that
the degradation is limited to the area of high BER, which is less suitable for communi-
cations, anyway.
A result for 16-QAM modulation with phase- and magnitude pre-equalization has been
shown in Figure 7-6. The impact of timing-offset estimation from the pilots of a single
data symbol is depicted and compared to the ideal case without any offset. The per-
formance is degraded by 12 dB.
Considering that the timing-offset does not vary rapidly, its estimate can be enhanced
by averaging a few subsequent measurements. L-times averaging reduces the variance
by a factor of approx. L. Note that, to apply this idea, eventual sampling frequency off-
sets must be taken into account.
7.5 Synchronization Parameter Estimation on the Up-link 243

QPSK; phase preequalization; correction for timingoffs. estd. from pilots

0
10

1
10

2
average BER

10

3
10

K = 0; rms = 1 sample
K = 0; rms = 3 samples
4
10 K = 0; rms = 5 samples
K = 4; = 1 sample
rms
downlink
5
10
5 0 5 10 15 20 25 30
average E /N [dB]
b 0

Figure 7-8: Performance of phase pre-equalization with QPSK. Time-invariant channel. Up-
link timing-offset estimation from the pilots of one single OFDM symbol.

7.5.2.4 Feedback of Time-Offset

In order to optimize the performance of the up-link TDMA scheme, where each trans-
mitted OFDM symbol may originate from a different user, the transmission times
should be adjusted such that all the up-link symbols arrive at optimum time instants
according to the frame format. Excess delays of the channel, which can be (slowly)
time-variant as the users move around, and also group delays of filters of the transmis-
sion scheme (being time-invariant) must be taken into account.
This optimization requires feedback information from the base station to the mobile,
since transmission delays cannot be estimated on a simplex link. I.e., the delay error
for the up-link, which is estimated by the base-station, must be signaled back to the
mobile.
As a reply to the signaling symbol from which the initial delay was estimated on the
up-link, the base-station will send back an acknowledgement message to that user,
via the down-link signaling symbol. The delay information can be fed back using this
symbol. To allow the detection of the delay by the signal-processing algorithm, the
delay is not digitally coded and appended to the acknowledgement message, but it is
applied on the signal constellations of the down-link signal by means of an artificial
progressive phase rotation. Since this signaling OFDM symbol is preceded by the
training sequence used for channel estimation, this phase rotation can be estimated in-
dependently from down-link delays. This can be achieved by applying the (decision
feedback) estimation method introduced in Section 7.5.2.1. The advantage of the ana-
log feedback is that neither additional digital data needs to be transmitted, nor (FEC)
244 Chapter 7 DSP Algorithm Development for the Up-Link

coding/decoding and modulation/demodulation are needed, which reduces delays and

simplifies the implementation.

7.6 Experimental Results

Experimental performance results are presented in this section, which have been ob-
tained from the emulation system presented in Section 5.3 (see Figure 5-12).

7.6.1 Performance over Different Radio Channels

The degradation (error floor) due to the channel variability is evident in the experi-
mental results depicted in Figure 7-9. The phase pre-equalization scheme for QPSK
has been implemented on the emulator, with system parameters according to transmis-
sion mode I-fr of the proposed OFDM system (see Section 5.2). Up-link results are in-
ferior to down-link results because of the larger time-distance between the training
symbol and the up-link time-slots. For the AWGN channel, both results are equivalent.
The impact of the speed of the channels variability is seen from Figure 7-10. Both up-
and down-link show increased irreducible error floors for higher speeds of movement.
The bars indicating the range between minimum and maximum results measured (10
measurements were performed at each SNR-value) suggest that this range is larger on
the up-link. In particular the channel with slowest variability shows a very large span
between measurements. This is most likely caused by the inaccuracy of the synchroni-

BER results vs. SNR for the TDDsystem; no I/Qmodulator; different channels

0
uplink
10 downlink
analytical results
AWGN
1
10 K = 4; RDS = 1
(average) biterrorrate (BER)

Rayleigh; RDS = 1

2
10

3
10

4
10

5
10

6
10
5 0 5 10 15 20 25 30 35 40
(average) Eb/N0 [dB]

Figure 7-9: Experimental performance results for the phase pre-equalization scheme using
QPSK.
7.6 Experimental Results 245

TDDsystem; Rayleigh channel, rms = 3 samples; no I/Q(de)modulator

0
10

1
10
average biterrorrate (BER)

2
10

uplink
3
10 downlink
v = 2 m/s (at 60 GHz)
v = 5 m/s
v = 0.2 m/s
analytical result
4
10
5 0 5 10 15 20 25 30 35 40
average Eb/N0 [dB]

Figure 7-10: BER of the TDD system vs. SNR for Rayleigh fading channels with fixed RMS de-
lay spread (rms = 3 samples); for varied velocities, vm = {0.2, 2, 5} m/s (at 60 GHz).

zation steps needed for coherent detection at the base station. The progressive phase
shifts caused by timing-errors are estimated from the pilots and corrected for. The per-
formance of these estimation and correction steps seems to deteriorate under bad (in-
stantaneous) channel conditions. Enhancing the robustness for such situations is an is-
sue for further research.
As the error floor for Rayleigh channels exceeds bit error rates of 102 at pedestrians
speed, a simple form of channel tracking/prediction should be considered. In Section
6.5.3, linear extrapolation was briefly evaluated to tackle this problem.

The impact of the channels excess delay is evident from Figure 7-11, where the BER
is shown as a function of SNR, for Rayleigh fading channels with various RMS delay
spreads (rms = {1, 3, 5} samples) and for a fixed velocity vm = 2 m/s (at 60 GHz).
While the performance on the down-link is relatively constant (thin lines), more sig-
nificant variations are observed from the up-link results. Note that the range of meas-
urement results increases with rms. From this observation we conclude that the addi-
tional performance degradation is again due to the inaccuracy of timing-offset esti-
mates on the up-link. The robustness of the applied technique to this problem should
be enhanced.

7.6.2 Impact of the I/Q-Modulator and -Demodulator

Figure 7-12 illustrates the influence of the I/Q-(de)modulation hardware, which was
not present in the system when the above presented measurements were done. The fig-
246 Chapter 7 DSP Algorithm Development for the Up-Link

BER results vs. SNR for the TDDsystem; fm = 2e5 (2 m/s); no I/Qmodulator
0
10

1
10
average biterrorrate (BER)

2
10

uplink
3
10 downlink
RDS = 3 samples
RDS = 1 sample
RDS = 5 samples
analytical result
4
10
5 0 5 10 15 20 25 30 35 40
average Eb/N0 [dB]

Figure 7-11: BER of the TDD system vs. SNR for Rayleigh fading channels with variable RMS
delay spread (rms = {1, 3, 5} samples) and for fixed velocities, vm = 2 m/s (at 60
GHz).

ure depicts the BER vs. SNR for a Rayleigh fading channel with rms = 3 samples and
for vm = 2 m/s. The similarity of the curves suggests that the I/Q-(de)modulation hard-
ware has no significant influence on the performance, as also concluded from the
down-link results presented in Section 6.5.

7.7 Conclusions
Pre-equalization techniques are proposed for the up-link of the OFDM based wireless
communications system. They are computationally quite efficient: the mobile has to
calculate the pre-equalization function once per frame (which unfortunately involves
the inversion of complex numbers) and multiply the data constellation points by it
(prior to the IFFT used at the OFDM transmitter). The base station only has to estimate
(precisely) the timing-offset for each mobile, a reference magnitude for the received
signal constellation points, and the carrier phase offset. The computational effort for
these operations is relatively low. No frequency-synchronization is needed on the up-
link due to the time-division duplex scheme. Neither is channel estimation required, or
any feedback of channel state information.
Keeping the complexity of the base station low is somewhat contrary to the commonly
applied strategy of decreasing the complexity of the mobile terminals in order to en-
able power efficient terminals. The additional signal processing required at the mobile
terminals is not very complex, however. And the additional complexity is very well
7.7 Conclusions 247

TDD; Rayleigh, = 3 smp., v = 2 m/s (at 60 GHz); impact of I/Q(de)mod.

rms
0
10

1
10
average biterrorrate (BER)

2
10

3
10 uplink
downlink
no I/Q(de)mod.
with I/Q(de)mod. (VCXO)
analytical result
4
10
5 0 5 10 15 20 25 30 35 40
average Eb/N0 [dB]

Figure 7-12: Impact of the I/Q-(de)modulation hardware on the BER vs. SNR for the TDD sys-
tem. The results are for a Rayleigh fading channel with rms = 3 samples and vm = 2
m/s.

invested, considering that the base station is accessed by many users at very high data
rates, which requires lots of processing power for managing these multiple data
streams with low delays. MAC and ARQ schemes have to be executed, for instance.
Moreover, the spectrum and the transmission power for the training symbols is saved
on the up-link, and very tight multiple access schemes are enabled, because the
overhead for training symbols, guard-periods and guard frequency bands for separat-
ing multiple users is very low.

Assuming a working point of the FEC-coding scheme at an average channel BER of

approx. 10-2, the performance on the up-link is very similar to the down-link. A realis-
tic channel estimation scheme has been considered in all computer simulations. Using
QPSK modulation, an Ec/N0 (energy per coded bit over noise power density) of about
15 dB is required for such performance, equivalent to a SNR (time-domain) of ~16 dB.
The same BER is reached with 16-QAM for an Ec/N0 of ~18 dB, equivalent to an SNR
of ~22 dB. Note, however, that usually the performance of the coding scheme is
largely dependent on the frequency diversity available over the channel bandwidth
(see Chapter 8). This may reduce the above-mentioned SNRs. On the other hand, the
time-variability of the channel was not included in the simulations leading to those re-
sults.
It is recommended for further work to perform computer simulation studies for the up-
link that incorporate error correction coding and also the time-variant channel.
248 Chapter 7 DSP Algorithm Development for the Up-Link

7.8 References
[1] T. Keller and L. Hanzo, Sub-band adaptive Pre-equalised OFDM Transmis-
sion, in Proc. IEEE VTC99 (fall) conference, Amsterdam (Netherlands), Sept.
1999, pp. 334338.
[2] G. J. R. Povey and M. Nakagawa, A Review of Time Division Duplex CDMA
th
Techniques, in Proc. IEEE 5 Intern. Symp. on Spread Spectrum Techniques
and Applications, 1998, pp. 630633.
[3] R. Esmailzadeh, M. Nakagawa, and E. A. Sourour, Time-Division Duplex
CDMA Communications, IEEE Personal Communications, vol. 4, no. 2, pp.
5156, April 1997.
[4] R. Esmailzadeh, E. A. Sourour, and M. Nakagawa, Pre-RAKE diversity com-
bining in time-division duplex CDMA mobile communications, IEEE Trans.
Veh. Technol., vol. 48, no. 3, pp. 795801, May 1999.
[5] R. Esmailzadeh and M. Nakagawa, Time-Division Duplex Method of Transmis-
sion of Direct Sequence Spread Spectrum Signals for Power Control Implemen-
tation, IEICE Trans. Commun., vol. E76-B, no. 8, Aug. 1993, pp. 10301038.
[6] H. Koorapaty, A. A. Hassan, and S. Chennakeshu, Secure Information Trans-
mission for Mobile Radio, IEEE Commun. Letters, vol. 4, no. 2, pp. 5255, Feb.
2000.
[7] G. J. M. Janssen, Robust receiver techniques for interference-limited radio chan-
nels, Ph.D. Thesis, Delft Univ. of Techn., Delft, The Netherlands, June 1998.
[8] M. Hmlinen, J. Talvitie, V. Hovinen, and P. Leppnen, Wideband Radio
th
Channel Measurement in a Mine, in Proc. IEEE 5 Intern. Symp. on Spread
Spectrum Techniques and Applications, 1998, pp. 522526.
[9] M. Ditzel and W. A. Serdijn, Optimum Energy Assignment for Frequency Se-
th
lective Fading Channels, in Proc. PIMRC 2001 (12 International Symposium
on Personal Indoor Mobile Radio Communications), San Diego, Oct. 2001, pp.
D-104D-108.
[10] J. Jang, K. B. Lee, and Y.-H. Lee, Frequency-Time Domain Power Adaptation
th
for a Multicarrier System in Fading Channels, in Proc. PIMRC 2001 (12 Inter-
national Symposium on Personal Indoor Mobile Radio Communications), San
Diego, Oct. 2001, pp. D-100D-103.
[11] T. M. Schmidl and D. C. Cox, Robust frequency and timing synchronization for
OFDM, IEEE Trans. Commun., vol. 45, no. 12, pp. 16131621, Dec. 1997.
Chapter 8 Performance Evaluation and
Enhancement of COFDM

8.1 Introduction
A coded OFDM (COFDM) system exploits the frequency-diversity of the wide-band
(frequency-selective) radio channel for ensuring good performance even though some
sub-carriers (SC) may be largely attenuated by fades. The performance evaluation of a
coded system is quite tedious. Analytical methods are difficult to apply in particular
for Ricean fading channels , and computer simulations require incredible computing
power. In Section 8.2, a semi-analytical approach is utilized, the so-called Concept of
Effective Eb/N0 [1]. This method calculates based on channel simulations effec-
tive Eb/N0-values, which, on an AWGN channel, would lead to equivalent error prob-
abilities. A set of channel simulations therefore yields a set of BER-estimates.
In this thesis, a novel extension to this method is proposed. We try to derive from the
frequency-domain (FD) channel model (see Chapter 2) a PDF for the effective Eb/N0.
Knowing this PDF, we can calculate average BERs and also outage probabilities.
Thereby, the performance analysis technique becomes fully analytical, because the
channel simulations are now replaced by the known PDF of the effective Eb/N0. It will
become evident, however, that an uncertainty remains (an offset depending on channel
and OFDM system parameters), which so far still requires computer simulations to be
determined.

Longer channel delay spreads imply higher frequency-diversity, and therefore better
performance for the coded OFDM system. (Provided that the delay spread does not
exceed the guard interval duration.) A (relatively) flat fading channel is thus the worst
situation, because the channel may attenuate all (or most) sub-carriers simultaneously,
leading to long error bursts, which are difficult to correct. In order to enhance the ro-

249
250 Chapter 8 Performance Evaluation and Enhancement of COFDM

bustness of the system in particular in these situations , antenna diversity techniques

are discussed in Section 8.3.
A novel class of diversity techniques for OFDM is introduced, which employs cyclic
delays (and weighing factors). The methods can be applied at the transmitter and at the
receiver. They are computationally highly efficient, because the (simple) signal proc-
essing steps are performed on the time-domain OFDM signals. That is, the IFFT and
FFT blocks are only executed once. Basic performance results show the potential im-
provement by the proposed techniques. Topics for further work are addressed.
Conclusions and recommendations are presented in Section 8.4.

8.2 Performance of a Coded OFDM System

Applying the concept of effective signal-to-noise ratio (Eb/N0) proposed by Nanda and
Rege [1] to an OFDM system model, estimates of frame error rates (FER) are obtained
for convolutional coding and bit-level interleaving22. The effective Eb/N0 for a fading
channel is defined as the Eb/N0 on an equivalent AWGN channel that results in the
same FER. The problem of calculating a FER is thus changed to the problem of cal-
culating the effective Eb/N0. In [1], two methods are proposed to translate a vector of
Eb/N0 values that describe the fading of subsequent channel bits to the effective Eb/N0.
The procedure of performance evaluation is therefore a semi-analytical one. Channel
simulations are performed and transformed to (instantaneous) error probabilities. A set
of channel simulations (for given channel parameters like average SNR) thereby leads
to a set of effective Eb/N0-values and furthermore to a set of respective FER-values. In
a second step, the average FER or the outage probability (the probability that the FER
is above a given threshold) can be calculated. These methods are briefly reviewed in
this section; their application to OFDM is elaborated.
Using the FD-channel model of the frequency-selective radio channel (see Chapter 2),
the probability density function (PDF) of the effective Eb/N0 is studied for OFDM
systems in quasi-static channels. Having appropriate PDFs for this parameter enables
the calculation of average FERs and outage probabilities for given channel and system
parameters, and a comparison and assessment of these. The performance analysis
thereby becomes fully analytical.

Other analytical work on the performance of coded transmission over fading channels
is mostly based on the calculation of pair-wise error probabilities (see Section 8.2.1.3)
over Rayleigh fading channels [2]. The influence of non-perfect interleaving was in-

22
A data frame is a data packet transmitted via the air-interface. In the system proposed in Section
5.2, ATM cells constitute such data packets. Each packet is coded individually. Note that the packets
(data frames) are not equal to OFDM symbols nor to the frames of the multiple access scheme (see
Section 5.2).
8.2 Performance of a Coded OFDM System 251

vestigated in [3], [4]. Sandell has applied this concept to OFDM for a Rayleigh fading
channel with a rectangular power delay profile [5]. This approach was not further pur-
sued in this Ph.D. research program. One reason was that, because the analysis devel-
oped in these references is based on zero-mean Gaussian processes, it appears difficult
to apply it to Ricean fading channels.

The current section proceeds with a review, analysis, and evaluation of the concept of
effective Eb/N0 and its application to the OFDM based wireless ATM transmission
system under investigation (Sections 8.2.1 and 8.2.2). In this context, the FER can be
seen as an ATM cell-loss probability. Finding appropriate PDFs for the statistical de-
scription of the effective Eb/N0 is elaborated in Section 8.2.3. Performance results are
presented in Section 8.2.4, followed by a brief summary in Section 8.2.5.

8.2.1 Review of the Concept of Effective Eb/N0

The idea behind the concept of effective Eb/N0 is to find an analytical relationship be-
tween the signal-to-noise ratios of the coded bits of a data frame and the frame error
probability. The discussed concept applies for frame-oriented, convolutionally coded
transmission over fading channels, with bit-level interleaving. Here, the method is ap-
plied to an OFDM based system for wide-band data communications, where the par-
allel OFDM sub-channels are modeled by a set of parallel Gaussian channels (see
Section 4.2.3). A quasi-static channel is assumed, with respect to the transmission time
of a block of OFDM symbols carrying one ATM cell the data frame under consid-
eration. That is, each OFDM symbol of a data-frame is affected equally by the chan-
nel. Therefore, interleaving is performed in the frequency-direction only (see Section
5.2.4.2).

8.2.1.1 Calculation of the Effective Eb/N0

Due to the frequency selective fading channel, each OFDM sub-carrier has an individ-
ual SNR, which is given as (cf. eq. (4-14), Section 4.2.3)

( Ec / N 0 ) l = E{| xl | 2 } hl N2 ,
2
(8-1)

where {hl} are the channel coefficients, expressing the channel transfer function, {xl}
are the signal constellations, and N2 is the variance of the additive channel noise per
sub-carrier. The index l = {1, 2, , L} designates the sub-carriers, just like the index i
used in Section 4.2.3. Note that a different indexing system has been introduced for
notational convenience; the time index k has been dropped. Considering the coding
rate and the number of bits per sub-carrier nm, the SNR per data bit on the l-th sub-car-
rier is written
( Eb / N 0 ) l = ( Ec / N 0 ) l ( Rc nm ) . (8-2)
In this equation, Rc is the rate of the convolutional code, defined as the ratio of infor-
252 Chapter 8 Performance Evaluation and Enhancement of COFDM

mation bits and coded bits. Note that the loss of signal power due to the discarded
guard interval is not accounted for.
Following [1], the L (Eb/N0)l values of the data-bearing sub-carriers are written as an
L-dimensional vector Eb / N 0 , which completely specifies the channel conditions:

Eb / N 0 = [( Eb / N 0 )1 ( Eb / N 0 ) 2  ( Eb / N 0 ) L ] (8-3)

The objective is to determine the relationship between the frame error probability PFE
and the vector Eb / N 0 , written as

PFE = I ( Eb / N 0 ) (8-4)

Assuming an AWGN channel, a relationship between PFE and Eb/N0 can be determined
based on analysis or simulation, for a given convolutional code and frame length Nc
(number of coded bits per data packet). It is assumed that this relationship, expressed
by
PFE , AWGN = K ( Eb / N 0 ) (8-5)

is known. For a given L-dimensional vector Eb / N 0 , the scalar (Eb/N0)eff, called the ef-
fective signal-to-noise ratio, is defined as the Eb/N0 in an equivalent AWGN channel,
which yields the same frame error probability. This is written as
PFE = I ( Eb / N 0 ) = K (( Eb / N 0 ) eff ) , (8-6)

where
( Eb / N 0 ) eff = J ( Eb / N 0 ) . (8-7)

Note that up to this point only the formulation of the problem is changed, no approxi-
mation has been introduced. Approximations are used for estimating (Eb/N0)eff from a
vector Eb / N 0 .

8.2.1.2 Estimation of (Eb/N0)eff

Two estimation methods are described in [1], for both of which the evaluation of the
following equations is required

( Eb / N 0 ) (effk ) = {( Eb / N 0 ) J ( k ) + ( Eb / N 0 ) J ( k +1) +  + (E / N )
b 0 J ( k + D 1) } D. (8-8)

In this expression, the symbol J(k + i) denotes the index of the Eb/N0 value corre-
sponding to the coded bit in position k + i, where k = {1, 2, , Nc D + 1} and i = {0,
1, , D 1}. Nc is the length of a data frame in (coded) bits, and D is the sequence-
length over which the dfree error bits in a minimum distance error-event are spread.
Thus (8-8) represents the average Eb/N0 over any D subsequent coded bits.
These D Eb/N0 values are taken from the L different Eb/N0 values of the OFDM sub-
8.2 Performance of a Coded OFDM System 253

carriers in an order described by the interleaving scheme. The symbol J(k + i) intro-
duces the interleaving scheme. It maps the Nc coded bits of a data frame on the L data-
bearing sub-carriers of a set of m = Nc/(Lnm) subsequent OFDM symbols. Note that it
is sufficient to calculate the first L averages (8-8), assuming the application of a peri-
odic interleaving scheme (cf. [1]), because the (L + 1)-st average will be equal to the
first one, and so on.
Estimate 1
(k)
Estimation method one simply selects the smallest of the ( Eb / N 0 ) eff values, based on
the reasoning that the most likely error event will occur at the sequence of D bits being
received with the smallest total signal energy. Thus, estimate 1 is expressed as

k
{
( Eb / N 0 ) eff = min ( Eb / N 0 ) eff
(k )
. } (8-9)

Estimate 2
Estimate two is calculated from the average of the FERs corresponding to all values
(k)
{ ( Eb / N 0 ) eff }. This is written as
N c D +1
PFE = K (( E
k =1
b / N 0 ) (effk ) ) ( N c D + 1) . (8-10)

The effective Eb/N0 is then

( Eb / N 0 ) eff = K 1 (PFE ). (8-11)

Note that with estimate 1, a change of the average SNR Eb/N0 which corresponds to
a common change of all Eb/N0-values in the vector Eb / N 0 , leads to an equivalent
change of (Eb/N0)eff. Such a linear relationship is not evident in case of estimate 2. Es-
timate 1 therefore drastically simplifies the analysis of error probabilities as a function
of average SNR. For this reason, and because of the good results obtained (see below),
most of the further analysis concentrates on estimate 1.

8.2.1.3 Derivation and Assessment of the FER for the AWGN Channel
Upper bounds on error probabilities for convolutional codes and AWGN channels are
given in most books on coding or digital communications (see e.g. [6], [7]). These
bounds use information on the weight structure of the codes, i.e., the number of code
sequences must be known that have a certain Hamming distance (HD) to the transmit-
ted sequence. Since convolutional codes have, per definition, no start or ending, so
called first-event error probabilities are evaluated, being the probabilities that an erro-
neous path merges at a certain node under consideration in stead of the correct path.
The numbers of erroneous paths merging in any node B and having Hamming dis-
tances d = dfree + {0, 1, 2, } from the correct path are denoted {ad}.
254 Chapter 8 Performance Evaluation and Enhancement of COFDM

Table 8-1: Parameters of rate maximum free distance codes (see [7], Appendix A)

Constraint Generators dfree (ad, d = dfree, dfree + 1, ) [cd, d = dfree, dfree + 1, ]

length in octal
3 5, 7 5 (1, 2, 4, 8, 16) [1, 4, 12, 32, 80]
5 23, 35 7 (1, 3, 5, 11, 25) [2, 7, 18, 49, 130]
7 133, 171 10 (11, 0, 38, 0, 193) [36, 0, 211, 0, 1404]

The probability of deciding in favor of a path with HD of d to the correct path is called
the pair-wise error probability. It is upper bounded for a binary symmetric channel
(BSC) with error probability p by (see [6], p. 490)

(
P2 (d ) = 2 p(1 p ) )
d
(8-12)
For coherent BPSK (or QPSK) modulation over AWGN channels and maximum-like-
lihood soft-decision decoding, it can be shown that

( )
P2 (d ) = Q 2dRc Eb / N 0 < exp( dRc Eb / N 0 ) , (8-13)

where Q is the integral over the Gaussian PDF (see [6], p. 40). The Q-function can be
upper bounded by the exponential function given in eq. (8-13). (Results calculated by
the Q-function will be referred as bound 1, results from the exponential approxima-
tion as bound 2.)
Finally, an upper bound (union bound) on the first-event error probability is written as

Pe < a
d = d free
d P2 (d ) , (8-14)

where dfree is the HD of the minimum distance error event.

The frame error probability PFE is determined from its complementary event, the prob-
ability that no error event occurs during the transmission of V information bits be-
longing to one frame. This event is well approximated by

PFE = 1 PFE (1 Pe )V , (8-15)

being the probability that correct decisions were made at V nodes (or decoding steps).
The frame error rate is found as

PFE 1 (1 Pe )V (8-16)

Similarly, a bound on the bit error probability can be obtained by including in the
summation in (8-14) the number of information bit errors resulting from each error
event, expressed by cd = ad f(d) (see [6], p. 488). Thus we can write an upper bound on
the bit error probability
8.2 Performance of a Coded OFDM System 255

convolutional coding; = 3, R = 1/2 convolutional coding; = 5, R = 1/2

0 0
10 10

1 1
10 10

2 2
10 10
Error prob.

Error prob.
3 3
10 10

BER of channel bits BER of channel bits

4 4
10 BER of information bits 10 BER of information bits
FER (simulated) FER (simulated)
Pb, bound 1 Pb, bound 1
5
10 Pb, bound 2 5
10 Pb, bound 2
P(E), bound 1 P(E), bound 1
P(E), bound 2 P(E), bound 2
6 6
10 10
2 0 2 4 6 8 10 4 2 0 2 4 6 8
Eb/N0 [dB] Eb/N0 [dB]

(a) (b)
Figure 8-1: Performance of rate- convolutional codes over AWGN channels. (a): constraint
length (= shift register stages + 1) = 3; (b): = 5

1
Pb <
kc c
d = d free
d P2 (d ) , (8-17)

where kc is the number of input bits of the encoder. kc = 1 in case of a binary encoder.
{ad} and {cd} are listed in Appendix A of [7] for a number of useful codes. Table 8-1
gives a few examples.
Performance Results
In this section, results are presented for the above equations and compared to computer
simulations. Two different rate- convolutional codes with constraint lengths of = 3
and = 5 were evaluated (see Table 8-1).
In the simulations, 8-level soft-decision Viterbi decoding23 was implemented with
memory lengths of 20 and 25 for the = 3 and = 5 code, respectively. QPSK modu-
lation was used.
It is seen from the results depicted in Figure 8-1 that the simulated and theoretical er-

23
Prior to the 8-level quantization, the received signal constellations yl were multiplied by
( Ec / N 0 ) l . This yields almost optimum performance (for QPSK and BPSK) when constant quantiza-
tion levels and metric values (optimized for (Ec/N0)l = 0 dB) are used for quantized soft-decision deco-
ding. This conclusion was drawn based on the evaluation of optimum metrics following [8]. With the
above-mentioned pre-multiplication, the optimum metrics largely remain constant over a large range
of (Ec/N0)l values. Note that, except for a constant factor related to N0, this multiplication is equivalent
to multiplying the FFT-output by the complex conjugates of the channel estimate, hl* , which is also an
efficient means for compensating for the phase distortion introduced by the channel. This solution is
specific for QPSK (or BPSK) modulation, however.
256 Chapter 8 Performance Evaluation and Enhancement of COFDM

ror probabilities match very closely for bound 1 (using the Q-function), especially for
the = 3 code. For the = 5 code, the simulation results are slightly worse than their
upper bound, for two reasons. Firstly, Pe is calculated using only the first five summa-
tion terms in eq. (8-14). And secondly, 8-level quantization is used in the simulation,
in stead of unquantized soft-decision decoding.
The mismatch observed at high error rates is due to the definition of the union bound.
All possible sequence error probabilities are simply added, thus, at high channel error
rates, the bound accounts for more than one error on one single node. This may even
lead to error probabilities larger than one.
Bound 2 yields results that are approx. 1 dB worse.

8.2.2 Assessment of the Concept of Effective Eb/N0

In this section, the concept of effective Eb/N0 is evaluated by computer simulations,
considering the OFDM system proposal introduced in Section 5.2.

8.2.2.1 OFDM System Configuration

The OFDM system parameters in this assessment were selected according to the quar-
ter rate transmission mode I-LL (see Section 5.2.1). Rate- convolutional coding with
QPSK modulation is used, where the data stream is transmitted over 19 data bearing
sub-carriers. Thus 24 OFDM symbols are required for carrying one complete ATM
cell (19 sub-carriers 2 bits per sub-carrier 24 = 912 bits = 1 coded ATM cell plus
signaling overhead). The simulation and estimation of (Eb/N0)eff assumes the mapping
of an ATM cell to subsequent OFDM symbols and time-invariance of the channel
during transmission of the whole ATM cell. Interleaving is performed in the fre-
quency-domain (FD) only (across the 19 active sub-carriers), which is appropriate be-
cause of the assumption of time-invariance. Two possible interleaving schemes have
been introduced in Section 5.2.4. A constraint length = 5 code was used, having
dfree = 7 and D = 10.

8.2.2.2 Computer Simulation Results

For one particular value of average Eb/N0 = 2 dB, the simulation program generates a
number of different frequency-selective channel realizations (approx. 250). The chan-
nel model is characterized by an exponentially decaying delay power spectrum. Figure
8-2a and d show typical TFs for different RMS delay spread, rms = {10, 55 ns}, re-
spectively. The sub-carrier spacing of the OFDM system is F = 1 MHz.
For each channel realization, a number of ATM cells (max. 100) is transmitted in order
to obtain the FER and BER. Additionally, the theoretical performance result is calcu-
lated, using the concept of effective Eb/N0. The results are compared based on scatter
plots (Figure 8-2b and e) and cumulative plots (Figure 8-2c and f).
8.2 Performance of a Coded OFDM System 257

E /N of the subcarriers, average E /N = 2 dB E /N of the subcarriers, average E /N = 2 dB

b 0 b 0 b 0 b 0
10 10

5 5
E /N [dB]

E /N [dB]
0 0
0

0
b

b
5 5

10 10
0 0.005 0.01 0.015 0.02 0.025 0 0.005 0.01 0.015 0.02 0.025
frequency [GHz] frequency [GHz]
BER of the subcarriers BER of the subcarriers
0 0
10 10

1
10
BER

BER
1
10
2
10

2 3
10 10
0 0.005 0.01 0.015 0.02 0.025 0 0.005 0.01 0.015 0.02 0.025
frequency [GHz] frequency [GHz]

(a) (d)
effective Eb/N0, estimation method 1 effective Eb/N0, estimation method 1
0 0
10 10

1 1
10 10

2 2
10 10
error probability

error probability

3 3
10 10

4 4
10 10
BER, bnd. 1 BER, bnd. 1
FER, bnd. 1 FER, bnd. 1
BER, bnd. 2 BER, bnd. 2
5 5
10 FER, bnd. 2 10 FER, bnd. 2
BER, sim. BER, sim.
FER, sim. FER, sim.
6 6
10 10
5 0 5 10 5 0 5 10
(Eb/N0)eff (Eb/N0)eff

(b) (e)
cumulative plot of frame error rates cumulative plot of frame error rates
0 0
10 10

1 1
10 10
FER

FER

2 2
10 10

3 3
10 simulation 10 simulation
est. 1, bnd. 1 est. 1, bnd. 1
est. 2, bnd. 1 est. 2, bnd. 1
est. 1, bnd. 2 est. 1, bnd. 2
est. 2, bnd. 2 est. 2, bnd. 2
4 4
10 10
50 100 150 200 50 100 150 200 250
simulation index simulation index

(c) (f)
Figure 8-2: Assessment of the concept of effective Eb/N0 applied to OFDM. (a)(c): Channel
has rms = 10 ns; (d)(f): rms = 55 ns. (a), (d): Typical realizations of the channel
transfer function with estimates 1 and 2 of (Eb/N0)eff (>) and the BER per sub-
carrier; (b), (e): Scatter plots of FER and BER vs. effective Eb/N0; (c), (f): Cumu-
lative plots of simulated and predicted FER.
258 Chapter 8 Performance Evaluation and Enhancement of COFDM

The matching of the cumulative plots shows the suitability of the investigated method.
In particular, estimation method 1 in combination with bound 1 (evaluating Q-func-
tions for obtaining PFE = K[(Eb/N0)eff]) yields an excellent approximation for the inves-
tigated system environment. It should be noted, however, that the matching is better
for the channel having rms = 10 ns than for the second channel. This is explained by
the fact that in the first case the fading can be characterized as gentle, while it is
much more severe in the second case (compare Figure 8-2a and d). The gently fading
channel leads to more accurate estimates of (Eb/N0)eff, evident in the smaller variance
of the performance results around the theoretical curves (see Figure 8-2b and e). It is
concluded that, in case of a more severely fading channel, using bound 2 in combi-
nation with estimation method 1 may be preferable. This option generally yields an
upper bound on the actual performance, while the estimation obtained with bound 1
may be too optimistic in some cases.

8.2.3 PDF of the Effective Eb/N0

8.2.3.1 Average Error Rate and Outage Probability

Using the concept of effective Eb/N0, frame and bit error probabilities can be predicted
for specific (simulated) fading channels. It is suggested in [1] to use this method in hy-
brid simulation models, i.e., simulating the fading channels and analytically evaluating
their error rate performance.
A more sophisticated way of performance evaluation can be derived from this concept
by finding the PDF of (Eb/N0)eff denoted p(~b | b ) , where ~b = ( Eb / N 0 ) eff and b is the
average Eb/N0, i.e. b = E{Eb / N 0 } . From this PDF and (8-6), the average FER is writ-
ten

PFE ( b ) = K (~b ) p(~b | b )d~b . (8-18)
0

The outage probability is defined as the probability that the FER is greater than some
threshold value PFE,th.
Pout ( b ) = Pr( PFE > PFE ,th ) = Pr(~b < ~b ,th ) =
~b , th , (8-19)
0
p(~b | b )d~b

where ~b,th is the effective Eb/N0 resulting in PFE = PFE,th, i.e., PFE ,th = K (~b,th ) . ~b,th can
be derived from the results given in Section 8.2.1.3.

8.2.3.2 Derivation of the PDF of (Eb/N0)eff

Basically, two different methods can be used to obtain the PDF p(~b | b ) :
8.2 Performance of a Coded OFDM System 259

Curve fitting to simulation results

Theoretical approaches
One theoretical approach investigates the PDF of the wide-band-average (normalized
received) power of the frequency selective channel. Assuming that the interleaved bits
are uniformly distributed over a certain bandwidth, the expressions (8-8) which are
used in the calculation of (Eb/N0)eff correspond to such average powers. The deriva-
tion of the PDF of the wide-band-average power, written as p P (x) , is elaborated in
a

Appendix D for Rayleigh and Ricean fading channels.

8.2.3.3 Application to (Eb/N0)eff

Two methods have been investigated to obtain the PDF of (Eb/N0)eff from the PDF of
the wide-band average power. Method 1 takes p P (x) for the minimum bandwidth
a

over which a minimum length error-sequence of D consecutive coded bits is spread

due to the interleaving. Method 2 takes p P (x) for the total bandwidth occupied by the
a

OFDM signal. It was found that method 1 yields better results for estimate 1 of
(Eb/N0)eff, on which we concentrate in this thesis. The bandwidths taken for the deriva-
tion of the PDFs p P (x) are listed in Table 8-2.
a

Results are shown in Figure 8-3 for Rayleigh fading channels and in Figure 8-4 for the
Ricean case. Different channel parameters and the full rate (FR) and quarter rate (QR)
transmission (TX) modes (see Section 5.2) are investigated. Shown in each figure are:
simulation results for the cumulative distribution function (CDF) of (Eb/N0)eff,
a log-normal CDF fitted to the simulation outcomes,
the theoretical exact CDF (see Appendix D), and
another log-normal CDF for analytically derived parameters.
CDFs are depicted in stead of PDFs, because the fitting in the lower tail is of great im-
portance for the accuracy of outage-rate results. This is observed much better from
CDFs.
The fitted (fitted to the simulated PDF) and theoretical (parameters derived from the
channel model) log-normal PDFs can be compared well using their parameters (see
Table 8-3, and Table 8-4). Observe that the standard deviations agree almost perfectly.
However, the means of the theoretical PDFs are too large due to the min{} operation
or non-linear averaging applied in estimating (Eb/N0)eff, which was not taken into
consideration (cf. eqs. (8-9)(8-11)). The mismatch of the means can be corrected for

Table 8-2: Bandwidths used to calculate the PDFs of (Eb/N0)eff .

estimate 1 estimate 2
full-rate (FR) mode 62 MHz 94 MHz
quarter-rate (QR) mode 17 MHz 22 MHz
260 Chapter 8 Performance Evaluation and Enhancement of COFDM

CDF of (E /N ) (estimate 1) and theoretical CDFs CDF of (Eb/N0)eff (estimate 1) and theoretical CDFs
b 0 eff 0
10
0 10

< abscissa)
Pr((Eb/N0)eff < abscissa)

1 1
10 10

0 eff
Pr((E /N )
b
2 2
10 10

CDF for simulation CDF for simulation

fitted lognormal CDF fitted lognormal CDF
exact CDF; corrected by 1.0 dB exact CDF; corrected by 2.4 dB
theoretical lognormal; corrected theoretical lognormal; corrected
3 3
10 10
12 10 8 6 4 2 0 2 4 2 4 6 8 10 12
(E /N ) [dB] (Eb/N0)eff [dB]
b 0 eff

(a) (b)
CDF of (Eb/N0)eff (estimate 1) and theoretical CDFs CDF of (Eb/N0)eff (estimate 1) and theoretical CDFs
0 0
10 10
Pr((Eb/N0)eff < abscissa)

< abscissa)

1 1
10 10
0 eff
Pr((E /N )
b

2 2
10 10

CDF for simulation CDF for simulation

fitted lognormal CDF fitted lognormal CDF
exact CDF; corrected by 0.44 dB exact CDF; corrected by 1.8 dB
theoretical lognormal; corrected theoretical lognormal; corrected
3 3
10 10
20 15 10 5 0 5 15 10 5 0 5
(E /N ) [dB] (Eb/N0)eff [dB]
b 0 eff

(c) (d)
Figure 8-3: Comparison of simulated distributions of (Eb/N0)eff, estimate 1 with a fitted log-
normal CDF, the shifted exact CDF of wide-band average power, and its approxi-
mation by the lognormal CDF. Rayleigh fading channels are shown. The parame-
ters are listed in Table 8-3 (boldface rows). (a): rms = 55 ns, QR TX-mode; (b):
rms = 55 ns, FR TX-mode (SNR = 10 dB); (c): rms = 10 ns, QR TX-mode; (d): rms =
10 ns, FR TX-mode.

Table 8-3: Parameters of the lognormal PDFs (in dB) fitted to simulation results and theoreti-
cal parameters obtained from the channel model. For Rayleigh fading channels.

Parameters Estimate 1 Estimate 2

TX- rms SNR theoretical fitted corr. theoretical fitted corr.
mode [ns] [dB] mean stdv. mean stdv. factor mean stdv. mean stdv. factor
QR 55 0 0.77 2.72 1.80 2.59 1.03 0.64 2.48 0.91 2.37 0.27
QR 55 10 9.23 2.72 8.20 2.59 1.03 9.36 2.48 8.64 2.40 0.72
FR 55 10 9.72 1.68 7.35 1.64 2.37 9.81 1.44 8.11 1.41 1.69
QR 10 0 1.90 4.48 2.34 4.34 0.44 1.73 4.22 1.68 4.10 0.05
FR 10 0 1.02 3.14 2.82 3.11 1.80 0.77 2.72 1.05 2.68 0.28
8.2 Performance of a Coded OFDM System 261

CDF of (Eb/N0)eff (estimate 1) and theoretical CDFs

CDF of (Eb/N0)eff (estimate 1) and theoretical CDFs 0
0 10
10

< abscissa)
1
Pr((Eb/N0)eff < abscissa)

10
1 10

0 eff
Pr((E /N )
b
2 2
10 10

CDF for simulation CDF for simulation

fitted lognormal CDF fitted lognormal CDF
exact CDF; corrected by 0.51 dB exact CDF; corrected by 0.98 dB
theoretical lognormal; corrected theoretical lognormal; corrected
3 3
10 10
6 4 2 0 2 4 5 4 3 2 1 0 1
(E /N ) [dB] (E /N ) [dB]
b 0 eff b 0 eff

(a) (b)
CDF of (Eb/N0)eff (estimate 1) and theoretical CDFs CDF of (Eb/N0)eff (estimate 1) and theoretical CDFs
0 0
10 10
Pr((Eb/N0)eff < abscissa)

Pr((Eb/N0)eff < abscissa)

1 1
10 10

2 2
10 10

CDF for simulation CDF for simulation

fitted lognormal CDF fitted lognormal CDF
exact CDF; corrected by 0.2 dB exact CDF; corrected by 0.47 dB
theoretical lognormal; corrected theoretical lognormal; corrected
3 3
10 10
3 2 1 0 1 2 2.5 2 1.5 1 0.5 0 0.5 1
(Eb/N0)eff [dB] (Eb/N0)eff [dB]

(c) (d)
Figure 8-4: Comparison of simulated distributions of (Eb/N0)eff, estimate 1 with a fitted log-
normal CDF, the shifted exact CDF of wide-band average power, and its approxi-
mation by the lognormal CDF. Ricean fading channels are evaluated. The parame-
ters are listed in Table 8-4 (boldface rows). (a), (b): rms = 12.5 ns, K = 8.3 dB; (a):
QR TX-mode; (b): FR TX-mode; (c), (d): rms = 6.4 ns, K = 14.8 dB; (c): QR TX-
mode; (d): FR TX-mode.

Table 8-4: Parameters of the lognormal PDFs (in dB) fitted to simulation results and theoreti-
cal parameters obtained from the channel model. For Ricean fading channels.
SNR = 0 dB.

Parameters Estimate 1 Estimate 2

TX- rms K theoretical fitted corr. theoretical fitted corr.
mode [ns] [dB] mean stdv. mean stdv. factor mean stdv. mean stdv. factor
QR 12.5 8.3 -0.33 1.75 -0.84 1.69 -0.51 -0.29 1.61 -0.39 1.60 -0.10
FR 12.5 8.3 -0.14 1.08 -1.11 1.02 -0.98 -0.10 0.90 -0.21 0.89 -0.11
QR 22.2 0 -1.01 3.12 -1.89 3.19 -0.88 -0.86 2.86 -1.07 2.94 -0.21
FR 22.2 0 -0.40 1.91 -2.17 1.82 -1.77 -0.29 1.59 -0.58 1.49 -0.29
QR 6.4 14.8 -0.09 0.86 -0.29 0.81 -0.20 -0.08 0.80 -0.07 0.79 0.00
FR 6.4 14.8 -0.04 0.55 -0.51 0.49 -0.47 -0.03 0.46 -0.04 0.45 -0.02
262 Chapter 8 Performance Evaluation and Enhancement of COFDM

by shifting the theoretical PDFs. This was done in the figure, leading to an excellent
fitting of the theoretical PDFs and the simulation results. For estimate 1 of (Eb/N0)eff,
the mismatch depends on rms, K, the BW of the OFDM signal, and on the interleaving
scheme; for estimate 2 (not investigated here), the average Eb/N0 also influences this
correction factor. Typical values for this offset, denoted as cm, are between 0 and
2.5 dB (see Table 8-3 and Table 8-4). The exact numbers can be determined from
computer simulations. The correction is formulated as
p' Pa ( x) = (1 / cm ) p Pa (x / cm ) . (8-20)

8.2.3.4 PDF of Estimate 1 of (Eb/N0)eff

Using the correction factor cm, the PDF of the average power p P (x) is matched to the
a

PDF of estimate 1 of (Eb/N0)eff at b = 1 . This leads to the PDF of estimate 1 as a func-

tion of the average power, being

1 ~
p(~b | b ) = p Pa b . (8-21)
b cm b cm
From (8-19) and (8-21), the outage probability can be directly related to the CDF of
(Eb/N0)eff and to the CDF of the wide-band average power.

~b ,th
Pout ( b ) = FPa (8-22)
b cm
The results as a function of b would thus have the same shape as the CDFs shown in
Figure 8-3 and Figure 8-4, however, mirrored at the (Eb/N0)eff = 0 dB-axis and shifted
by ~b,th .
Since, for estimate 2, the correction factor is a function of the average SNR, cm ( b ) ,
the above equations are not exact for that case.

8.2.4 Performance Results and Discussion

Further evaluation of the proposed method of performance evaluation can be achieved
by analyzing special (theoretical) cases of fading channels. Namely, the flat fading
channel as one extreme, where all sub-carriers experience the same attenuation (and
phase distortion), and independent fading of each transmitted coded symbol as the
other extreme. Results for practical channels with correlated fading will be located in-
between these cases.
The special cases are analyzed and evaluated in the first part of this section, for
Rayleigh distributed fading. Secondly, performance results are presented for practical
Rayleigh and Ricean fading channels.
8.2 Performance of a Coded OFDM System 263

8.2.4.1 Special Cases: Flat and Independent Rayleigh Fading

Considering flat fading, both estimates of the effective Eb/N0 are equivalent and equal
to the (Eb/N0)l of any sub-carrier. Thus, the curves shown in Figure 8-1 exactly de-
scribe the frame and bit error probabilities for one particular channel realization. By
averaging over the PDF of Eb/N0, which is defined by the fading statistics, average er-
ror rates and outage rates can be calculated. For the Rayleigh fading channels, the PDF
of Eb/N0 is a chi-square distribution with two degrees of freedom, i.e., an exponential
distribution (see [6], p. 45).
Flat fading results in the worst possible performance in Rayleigh fading environments
(for a given OFDM system and error correction coding scheme), because all carriers
undergo the same fading and might thus be simultaneously attenuated by a deep fade.
No frequency-diversity is present in this case.

The best possible performance on Rayleigh fading channels can be expected in the
second special case, assuming completely independent fading of subsequent coded
bits. In this case, the frequency-diversity is maximized.
Evaluation of the second special case is done using the concept of effective (Eb/N0),
where a chi-square distribution is used to model the PDF of (Eb/N0)eff. Appropriate pa-
rameters have to be found for this PDF to obtain performance results. The following
considerations are made.
In the most probable error event, the decoder selects an erroneous sequence having the
minimum Hamming distance of dfree to the correct, transmitted sequence. The PDF of
the sum of dfree squared, independent, Rayleigh distributed random variables is de-
scribed by a chi-square distribution with 2dfree degrees of freedom.
d free
1 d free
~b free exp ( d free~b b )
d 1
p(~b | b ) (8-23)
(d free ) b

Bound 2 is used for evaluating the performance for this special case, to be on the save
side, since no correction is applied for the min{} operation in eq. (8-9).
Computational Results
In Figure 8-5, theoretical results of average frame error rate (FER) are compared to
computer simulation results for the two special cases. The same (rate Rc = , con-
straint length = 3 (Figure 8-5a) and = 5 (Figure 8-5b)) codes are analyzed, which
were evaluated over AWGN channels in Section 8.2.1.3 (Figure 8-1). Up to 20000
transmitted ATM cells were simulated, at each value of average SNR. Alternatively,
the simulation for a specific SNR was interrupted after detecting 100 erroneously re-
ceived frames. (One frame is equivalent to an ATM cell). A good matching of the re-
sults is observed from the figure.
The comparison of different coding schemes in terms of the average FER and outage
probability is illustrated in Figure 8-6. Additionally, results are given for a rate Rc = ,
264 Chapter 8 Performance Evaluation and Enhancement of COFDM

convolutional coding; constraint length = 3; special cases convolutional coding; constraint length = 5; special cases
0 0
10 10
flat fading; simulation flat fading; simulation
analytical (dfree = 1; bd. 1) analytical (d = 1; bd. 1)
free
indep. fading; simulation indep. fading; simulation
analytical (dfree = 5; bd. 2) analytical (d = 7; bd. 2)

average frameerrorrate (FER)

average frameerrorrate (FER)

1 1 free
10 10

2 2
10 10

3 3
10 10

4 4
10 10
0 5 10 15 20 25 30 0 5 10 15 20 25 30
average E /N [dB] average E /N [dB]
b 0 b 0

(a) (b)
Figure 8-5: Average FER for the two special cases: Flat fading and independent Rayleigh fad-
ing. Comparison of theoretical results to computer simulations for two different
convolutional coding schemes; rate Rc = , constraint length = {3, 5}.

constraint length = 7 code. It was not possible to assess the latter code with computer
simulations because of the excessive simulation time required. (One ATM cell takes
approx. 20 minutes).
The figure clearly indicates that higher performance gains are obtained with more
complex coding schemes. This holds especially for the case of independently fading
sub-carriers.

Outage probabilities may give a better indication of the performance in fading chan-
nels. In the scenario of a broad-band computer communications system, the outage
convol. coding; rate = 1/2; flat fading (FF) and independent fading (IF) convol. coding; rate = 1/2; flat fading (FF) and independent fading (IF)
0 0
10 10
FF; = 3 FF; = 3
IF; = 3 IF; = 3
FF; = 5 FF; = 5
IF; = 5 IF; = 5
outage probability; Pr(FER > 102
average frameerrorrate (FER)

1 1
10 FF; = 7 10 FF; = 7
IF; = 7 IF; = 7

2 2
10 10

3 3
10 10

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

(a) (b)
Figure 8-6: Comparison of different rate Rc = coding schemes for the two special channels.
(a): Average FER; (b): Outage probability
8.2 Performance of a Coded OFDM System 265

probability Pout = Pr(PFE > 10-2) quantifies the probability that the FER exceeds one
percent at a certain location, with a certain average Eb/N0. In other words, if the chan-
nel is considered time-invariant, the probability is quantified that the current channel
leads to an FER exceeding a certain level. In all other cases (for the given local-area-
average SNR), the system performs better. The outage rate is thus usually a stricter
performance bound than the average BER or FER.
It is evident from Figure 8-6 that outage rate results (for FER > 10-2) are almost
equivalent to results of average FER, except for an SNR gap of about 1 2 dB. I.e.,
about 1 2 dB more SNR is required to reach a certain level of outage probability
(FER > 10-2) than to reach the same level of average FER. Due to the way these results
were obtained (by a convolution of the error probability curve of the AWGN channel
with the PDF of the effective Eb/N0), there is reason to assume that this similarity of
the result-curves is a general property. Other cases studied below confirm this conclu-
sion. The rather steep decay of the error probability curve for the AWGN case (see
Figure 8-1) compared with the flat shapes of the (Eb/N0)eff-PDFs (see Figure 8-3 and
Figure 8-4) explain this property.
The uncorrelated channel shows more than 12 dB gain (at FER = 102 and at
Pout = 102) compared with the fully correlated, flat fading channel.

8.2.4.2 Results for Rayleigh Channels

Results of average FER as a function of b are depicted in Figure 8-7a and b, for the
QR and FR TX-modes, respectively. Two Rayleigh fading channels with rms = 10 ns
and 55 ns were evaluated. For comparison, theoretical results are also given for the
two special cases discussed above.
It is clearly seen that the FR mode outperforms the QR mode. This result is not sur-
prising, because in the FR-mode, the interleaver can spread the coded bits over a much
larger bandwidth, thus the correlation of the fading of consecutive coded bits is re-
duced. In other words, more frequency-diversity can be exploited by the cod-
ing/interleaving scheme. The less correlation, the closer the performance approaches
the bound for the special case of independent fading.
More frequency-diversity also explains the superior performance on channels with
higher delay spread. Remember that it was assumed in the OFDM system model that
the delay spread of the channel is shorter than the guard interval, i.e., there is no ISI.
To mitigate the shortcoming of the QR-scheme, frequency-hopping in combination
with time-interleaving during one transmitted packet can be applied.
Simulation results match remarkably well with the theoretical curves. Note that simu-
lations for higher b are less representative because too few error events (sometimes <
10) were observed. Extremely long simulation times made it impracticable to increase
the accuracy. Generally, the simulation results are between the two theoretical bounds
(see eq. (8-13)). Bound 2 is rather conservative, while bound 1 is sometimes too opti-
266 Chapter 8 Performance Evaluation and Enhancement of COFDM

QRmode; R = 1/2, = 5 coding; IL 2; est. 1 of (E /N )

c b 0 eff
0
10

1
10
average FER

2
10
= 10 ns; bd. 1
rms
" ; bd. 2
" ; sim.
3 = 55 ns; bd. 1
10 rms
" ; bd. 2
" ; sim.
flat fading
indep. fading
4
10
5 0 5 10 15 20 25
average E /N [dB]
b 0

(a)
FRmode; R = 1/2, = 5 coding; IL 3; est. 1 of (E /N )
c b 0 eff
0
10

1
10
average FER

2
10
= 10 ns; bd. 1
rms
" ; bd. 2
" ; sim.
3 rms = 55 ns; bd. 1
10
" ; bd. 2
" ; sim.
flat fading
indep. fading
4
10
5 0 5 10 15 20 25
average Eb/N0 [dB]

(b)
Figure 8-7: Average frame error rate as a function of the average signal-to-noise ratio for two
Rayleigh fading channels having rms = 10 ns, and rms = 55 ns. (a) QR TX-mode
with approx. 25 MHz bandwidth; (b) FR mode with approx. 100 MHz. Note: A fit-
ted log-normal PDF of (Eb/N0)eff has been used here to depict the result for the spe-
cial case of independent fading.

mistic. The average FER has been obtained with an accuracy in the order of 1 dB.
Similar conclusions can be drawn from the outage-probability results, which are de-
8.2 Performance of a Coded OFDM System 267

2
FERth = 10 ; SNRth = 4.3 dB; exact PDFs; Rc = 1/2, = 5
0
10

Outage probability Pr(FER > FER )

th
1
10

2
10

3
10 QR, rms = 10 ns
QR, rms = 55 ns
FR, = 10 ns
rms
FR, = 55 ns
rms
4
10
5 0 5 10 15 20 25
average Eb/N0 [dB]

Figure 8-8: Analytical outage probabilities for different Rayleigh channels and for full- and
quarter-rate terminals.

picted in Figure 8-8 for the same system and channel parameters. No simulation results
were available to assess these analytical results. Note again, the resemblance of these
results with the average FER plots.

8.2.4.3 Results for Ricean fading channels

Results for two Ricean fading channels and for the two transmission modes are de-
picted in Figure 8-9. Generally, the performance over Ricean channels is superior
compared with Rayleigh channels.
A comparison to simulation results is shown for the rms = 12.5 ns, K = 8.3 dB channel.
The analytical error rates are by approx. 0.5 dB worse than these simulation results, in
case of the FR transmission mode. This is different for the QR-mode.
Again a clear advantage of the FR mode compared to the QR mode is observed, which
is due to the increased frequency-diversity implied by the larger bandwidth. However,
the difference is much smaller (< 2 dB at FER = 102) than over the Rayleigh fading
channels (~ 5 dB for the rms = 10 ns channel at FER = 102). In practice, the higher
transmission power enabled by the smaller bandwidth (6 dB in the link budget) can be
used to compensate for this slightly higher SNR requirement.
The performance difference of the two transmission modes is hardly observed for the
second Ricean channel investigated. The high Ricean K-factor of K = 14.8 dB (rms =
6.4 ns) yields almost equivalent, excellent performance for both transmission modes.
Results are only about 1 dB worse than results for the AWGN channel.
Outage probability results are given in Figure 8-10.
268 Chapter 8 Performance Evaluation and Enhancement of COFDM

rms = 12.5 ns; K = 8.3 dB; Rc = 1/2, = 5 coding; est. 1 of (Eb/N0)eff rms = 6.4 ns; K = 14.8 dB; Rc = 1/2, = 5 coding; est. 1 of (Eb/N0)eff
0 0
10 10

1 1
10 10

average FER
average FER

2 2
10 10

QRmode (IL 2); bd. 1

3 3
10 " ; bd. 2 10
" ; sim. QRmode (IL 2); bd. 1
FRmode; bd. 1 " ; bd. 2
" ; bd. 2 FRmode; bd. 1
" ; sim. " ; bd. 2
4 4
10 10
0 2 4 6 8 10 0 2 4 6 8 10
average Eb/N0 [dB] average Eb/N0 [dB]

(a) (b)
Figure 8-9: Average frame error rate as a function of the average signal-to-noise ratio for two
Ricean fading channels. (a) rms = 12.5 ns, K = 8.3 dB; (b) rms = 6.4 ns, K = 14.8 dB

2
FERth = 10 ; SNRth = 4.3 dB; lognormal PDFs; Rc = 1/2, = 5
0
10
QR; K = 8.3 dB, = 12.5 ns
rms
QR; K = 14.8 dB, rms = 6.4 ns
FR; K = 8.3 dB, rms = 12.5 ns
Outage probability Pr(FER > FER )

FR; K = 14.8 dB,

th

1 = 6.4 ns
10 rms

2
10

3
10

4
10
2 4 6 8 10 12 14
average Eb/N0 [dB]

Figure 8-10: Analytical outage probability results for Ricean channels and for full- and quarter-
rate terminals.

8.2.5 Summary
In this section, the estimation of frame error rates (FER) is demonstrated, for packet
oriented OFDM transmission systems including convolutional coding and bit-level in-
terleaving. The concept of effective Eb/N0 [1] is used to achieve this goal without the
need for time-intensive computer simulations. This procedure requires the following
steps:
8.2 Performance of a Coded OFDM System 269

For a given channel realization (a channel transfer function (TF)), the Eb/N0-values
at individual coded symbols are transformed into one scalar, the so-called effective
Eb/N0 (see eq. (8-9)).
This effective Eb/N0 is then translated to a bit or frame error probability (see Sec-
tion 8.2.1.3).
A set of channel realizations thus leads to a set of performance results, at given aver-
age Eb/N0 and channel parameters. These results may be averaged consecutively. The
original method [1] proposes the simulation of channel realizations to derive the per-
formance results.
An extension/modification has been introduced in this section. The idea is to
determine the PDF of the effective Eb/N0 for a given set of channel parameters {rms,
K, average Eb/N0} and for given OFDM system parameters, which allows for the
analytical evaluation of average error probabilities and outage probabilities (Section
8.2.3.1). The derivation of this PDF requires the following steps:
Calculate analytically the PDF of the wide-band average power (for the bandwidth
over which the coded symbols are spread in the most likely error event; see Section
8.2.3.3)
Correct for the bias between the average-power PDF, and the PDF of the effective
Eb/N0. This bias (approx. 0 2.5 dB) can be determined exactly by computer
simulations.
Except for the last step, the method is thus fully analytical. Based on the following
rules-of-thumb, the correction factor may be roughly estimated, in order to skip the
simulation step.
More frequency-diversity usually implies higher correction factors. I.e., larger
bandwidth and larger rms require a larger correction factor.
If the interleaving leads to almost independently fading coded bits, then 2.5 dB cor-
rection are appropriate.
A channel that introduces only about one significant fade to the whole signal re-
quires a correction of about 0.5 dB.
It should be possible to empirically find a relation between for instance the coher-
ence bandwidth or frequency-domain level crossing rate of the channel and the correc-
tion factor, considering the system bandwidth. The investigation and formulation of
such a relation is subject for further work.
Conclusions on performance results are summarized in Section 8.4.
270 Chapter 8 Performance Evaluation and Enhancement of COFDM

8.3 Performance Enhancement using Antenna Diversity

Coded OFDM systems exploit the frequency-diversity of wide-band radio channels to
enable robust data communication over frequency-selective fading channels. The
coded data stream is therefore well spread over the bandwidth of the signal spectrum,
using an interleaving scheme in the frequency-direction. The channels frequency-se-
lectivity, and thus the potential frequency-diversity, depends on the delay spread of the
channel. The longer the delay spread, the more fades are present per unit of bandwidth,
which is of advantage for the coded and interleaved transmission scheme, because er-
rors occur more independently. (See the performance results in Section 8.2.4).
The proposed antenna diversity schemes aim to improve the performance in the oppo-
site situation. If the channel delay spread is very short, then the whole OFDM signal is
faded equally (i.e., the channel is flat), which may induce long error bursts that are
hard to correct. There is not sufficient frequency-selectivity in this case to be exploited
by the coding and interleaving schemes. Using a set of transmit or receive antennas,
the presented diversity schemes can randomize the channel response and thereby in-
crease the frequency diversity. Note that the discussed methods can be used at the
transmitter and/or at the receiver.
The first strategy to achieve such gains employs cyclic delays applied to the effective
parts of the OFDM symbols (Section 8.3.1). Another method transmits the even and
odd sub-carriers via separate antennas (Section 8.3.2). A common property of these
novel diversity techniques is their high computational efficiency. In the first scheme,
the signal processing is done on the time-domain OFDM signal, thus the Fourier trans-
form does not need to be duplicated. In the second scheme, the FFT is divided in two
halves.
To employ a conventional diversity technique, all blocks of an OFDM receiver in-
cluding the FFTs must be present M-times (M is the number of diversity branches).
Such a scheme for the receiver is shown in Figure 8-11. There, for each sub-carrier the

branch 1:
down- I/Q FFT
A/D
conv. dmod
M-branch
diversity symbol
combining detection
on each
branch M: sub-carrier
down- I/Q FFT
A/D
conv. dmod

L.O. L.O.

Figure 8-11: Block diagram of an OFDM receiver allowing conventional diversity techniques on
each sub-carrier.
8.3 Performance Enhancement using Antenna Diversity 271

ins. cyclic
prefix (GI) branch 1

fwd. error cycl. delay ins. cyclic

symbol IFFT branch 2
correction n,2, 2 [s] prefix (GI)
mapping
encoding
input data
stream

cycl. delay ins. cyclic

n,M, M [s] prefix (GI) branch M

Figure 8-12: Diversity technique using cyclic delays for the transmitter.

output values of the M FFT blocks are combined using the well-known methods for
space diversity, as, for instance, selection diversity, maximal ratio combining, equal
gain combining, etc..

8.3.1 Antenna Diversity for OFDM Using Cyclic Delays

The core idea of this novel scheme concerns the introduction of cyclic delays to the
effective parts (FFT-parts) of the OFDM signals transmitted/received via several an-
tennas.
For the transmitter, the novel diversity technique is depicted in Figure 8-12. Up to the
IFFT, which is used to modulate data constellations on the OFDM sub-carriers and
whose output is the time-domain OFDM signal, a conventional OFDM system is pre-
sent. In order to generate signals for a number of transmit antennas, cyclic delays (of
n,i samples or i seconds, i = {1, 2, , M}; 1 = 0) are introduced to the (effective)
FFT-parts of the OFDM symbols. A cyclic delay means that the n,i samples shifted
beyond the effective part are transmitted in the beginning of that part of the symbol
(see Figure 8-13). In the conventional manner, the cyclic prefix (guard interval) is

FFT-output:

n,i, i
cyclic delay:

guard interval:

Figure 8-13: Applying a cyclic delay to the effective part of the OFDM symbol.
272 Chapter 8 Performance Evaluation and Enhancement of COFDM

channel NSCM calc. combined

estimation channel TF

remove factor
branch 1: guard w1

remove cycl. delay factor

branch 2: FFT symbol fwd. error
guard n,2, 2 [s] w2
detection correction

remove cycl. delay factor

branch M: guard n,M, M [s] wM

VCO tuning time- and

for frequency frequency
synch. synch.

Figure 8-14: Diversity technique using cyclic delays and weighting factors for the receiver.

transmitted prior to the effective part.

The method presented is similar to the delay diversity scheme described in [9]. The
cyclic delays allow for much longer delays, however, which are otherwise limited to
fractions of the guard interval period to avoid inter-symbol-interference. This fact is
paramount for applying the proposed technique at the transmitter, where no informa-
tion is available on the length of the current channels impulse response.

Similarly, cyclic delays can be applied to the OFDM signals received via multiple an-
tennas, in order to perform diversity combining at the OFDM receiver prior to the FFT
(see Figure 8-14). Utilized at the receiver, the delay times can be adapted (optimized)
based on individual channel estimates for each diversity branch. Moreover, weighing
factors {wi} can be applied to allow for more flexibility. (A pre-FFT diversity scheme
based on such weighing factors, but without delays, is analyzed in [10].) The optimiza-
tion of the parameters is subject for further research. Since the channel transfer func-
tions are (usually) not available at the transmitter, such an optimization is not possible
there. Therefore, the application of weighing factors at the transmitter is less promis-
ing, although it is generally possible.
The operation of the diversity schemes, and some design considerations for the delay
times (for a non-adaptive scheme) are discussed below.

8.3.1.1 Illustration of the Method

As seen from the idealized system model derived in Section 4.2.3, the OFDM sub-
carriers are attenuated and phase distorted according to the channel transfer function
(TF). The diversity scheme randomizes the TF of the composite channel as follows.
The TFs Hi(f) (time-variability is neglected) of a Rayleigh fading channel are corre-
lated, zero-mean, complex Gaussian random processes. The superposition of the chan-
nel TFs of individual antennas is therefore also a Rayleigh channel, as the sum of zero-
8.3 Performance Enhancement using Antenna Diversity 273

channel transfer function channel impulse response (real part)

magnitude [dB]
10 0.5

magnitude
0
0
10
20 0.5
40 20 0 20 40 0 2 4 6 8
10 0.5
0
0
10
20 0.5
40 20 0 20 40 0 2 4 6 8
10 0.5
0
0
10
20 0.5
40 20 0 20 40 0 2 4 6 8
10 0.5
0
0
10
20 0.5
40 20 0 20 40 0 2 4 6 8
subcarrier index delay tap index

Figure 8-15: Four (independent) channel realizations. The left-hand side depicts the channels
magnitude transfer functions; the right-hand side illustrates the real-parts of the
impulse responses.

mean Gaussian processes is yet another zero-mean Gaussian process. The correlation
properties are altered, however, by the cyclic delays introduced. The following equa-
tion gives the TF for the composite channel. It is seen that the cyclic delays introduce
progressive phase rotations to the TFs.
M
1
H ( f ) =
M
H ( f )e
i =1
i
j 2 i f
. (8-24)

The normalization by 1 / M is introduced to keep the transmission power constant,

when the method is used at the transmitter. At the receiver, this factor can account for
the channel noise that adds up incoherently. The noise for the combined channel can
then be considered equal to the noise of a single channel. Note that the weighing fac-
tors are not incorporated in this brief analysis.
The correlation properties of the composite channel are investigated in the following
sub-section. Here, we firstly illustrate the principle of the diversity technique, and we
discuss the necessity that signals of multiple antennas are combined and that delays are
introduced.

Figure 8-15 depicts the transfer functions and impulse responses of four independent
frequency-selective channels. Having short impulse responses, their frequency-selec-
tivity is limited. It is seen that the channels have similar fading characteristics, par-
ticularly, their impulse responses have similar length. Superimposing these channels
without delays, the sum of channels is just another channel with a similar IR and fre-
274 Chapter 8 Performance Evaluation and Enhancement of COFDM

channel transfer function sum of impulse responses

10 0.5

magnitude [dB]

magnitude
0
0
10

20 0.5
40 20 0 20 40 0 2 4 6 8

channel transfer function composite IR with cyclic delays

10 0.5
magnitude [dB]

magnitude
0
0
10

20 0.5
40 20 0 20 40 0 10 20 30
subcarrier index delay tap index

Figure 8-16: Transfer functions (left-hand side) and impulse responses (right-hand side) of the
composite channels. The figures in the first row show the channel of the super-
posed channel without the introduction of cyclic delays. The characteristics of this
channel are equivalent to the characteristics of the component channels. In the
second row, cyclic delays were applied, leading to a clear randomization of the
channel transfer function.

quency-selectivity, as seen from the top-row of figures in Figure 8-16. Nothing is lost
or gained in this case.
Introducing the cyclic delays, the composite channel consists of all those IRs, shifted
by the respective delay times. This leads to a much-extended overall IR, corresponding
to a more random channel TF, as seen from the second row of graphs in Figure 8-16.
Thereby, inter-symbol-interference is avoided due to the cyclic way of applying the
delays.
It is not possible to obtain a similar randomization by combining (with delays) multi-
ple copies of the received signal of a single antenna. This would be equivalent to the
application of a transversal filter to the received signal. If the filter can be adapted, it is
well possible to equalize the transfer function and to get a flat channel response. Un-
fortunately, the noise level is thereby modified accordingly; therefore nothing is
gained. Combining (cyclically) delayed copies of the signal without weighing factors,
for instance, means a multiplication of the channel TF by a filter TF that has zeros at
certain positions. Clearly, such filtering cannot enhance the performance of an OFDM
system.

8.3.1.2 Analysis and Selection of the Delay Times

Let us briefly analyze the correlation properties of the combined channel. It is assumed
8.3 Performance Enhancement using Antenna Diversity 275

that the (independent) component channels have similar stochastic properties, ex-
pressed by a common spaced-frequency correlation function H (f ) =
E{H i* ( f ) H i ( f + f )} . The sum (8-24) is a sum of zero-mean complex Gaussian ran-
dom processes, which gives another zero-mean complex Gaussian process, as men-
tioned above. Accounting for the phase rotations, the spaced-frequency correlation of
the composite channel is written as
M
1
H ( f ) = H ( f )
M
e
i =1
j 2 i f
. (8-25)

The correlation is reduced, because the magnitude of the sum term is less or equal to
one [9]. Appropriately selecting i, it is possible to force a zero in this correlation func-
tion for the frequency separation corresponding to the separation of subsequent coded
bits. Unfortunately, for double the separation, the normalized sum in (8-25) becomes
one again (if two branch diversity is used). The bit at triple that distance is then in a
zero again. Additional diversity branches enable the nulling of more subsequent bits.
A large reduction of the correlation function is obtained, for instance, when zeros are
forced to be on directly adjacent sub-carriers, i.e., on sub-carriers separated by f = {F,
2F, , (M 1)F}, where F is the sub-carrier spacing. This makes the correlation
function at frequency-separations of integer multiples of F

H (kF ) if k = lM
H (kF ) = , (8-26)
0 otherwise

where {k,l} are integer variables. Such a result is obtained for delays
i 1 N
i = or n ,i = (i 1) , (8-27)
MF M
where N is the number of FFT-points. Note that, when using this parameter set, it is
important that the interleaving depth is selected differently to M, otherwise the corre-
lation of the fading on subsequent coded bits is not reduced at all.

8.3.2 Even/odd Sub-carrier Transmitter Diversity

This section describes another computationally highly efficient method to achieve an-
tenna diversity gain using multiple transmitter antennas.
Figure 8-17 shows the block diagram for the case of two transmitter antennas. As in a
conventional OFDM transmitter, the input data stream is forward error correction en-
coded and mapped on constellation points of any (typically low-order) modulation
scheme. The de-multiplexer separates the constellation points to be transmitted as odd
and even sub-carriers. Two N/2-point IFFTs are executed on these constellations. To
produce the N samples of a (full bandwidth) OFDM signal, the (complex-valued) out-
put samples of the FFTs are essentially duplicated. The resulting signals now occupy
276 Chapter 8 Performance Evaluation and Enhancement of COFDM

N/2-point ins. cyclic

even sub- IFFT N-point prefix (GI)
carriers parallel-
to-serial
converter branch 1
fwd. error symbol de-multi-
correction mapping plexing
encoding
input data
stream odd sub- N/2-point rotate by
carriers ins. cyclic
IFFT [WN0, WN-1, N-point
WN-N/2+1] prefix (GI)
parallel-
-1
to-serial
-1
-1 converter branch 2
-1

Figure 8-17: Block diagram of the even/odd sub-carrier transmitter diversity scheme.

exactly every second sub-carrier. The signal of branch one is directly applied to the
first RF transmitter, after adding the cyclic prefix. This signal occupies the even sub-
carriers; the odd sub-carriers are zero.
To fill in the odd sub-carriers, a multiplication of the second signal with the complex
exponential sequence [WN0, WN1, WNN+1] is required, where WN = exp(j2/N). This
shifts the second signal in the frequency domain by one sub-carrier. The signal thus
occupies the odd sub-carriers after conversion to RF.
A more efficient way of calculating the N/2 complex multiplications is shown in
Figure 8-17. The N/2-output samples of the IFFT of branch two are multiplied by
[WN0, WN-1, WN-N/2+1]. This yields the first N/2 (complex-valued) time-samples for
transmission. The second N/2 time-samples are obtained by flipping the signs of the
first N/2 samples, since WNi = WN ( N / 2+i ) . Thereby, N/2 complex multiplications are
saved. Note the equivalence of the proposed technique to the last step of a decimation-
in-time FFT algorithm [11].
The two signals are modulated on equal frequency RF carriers for transmission via two
antennas. On the (linear) channel, the signals are superimposed and then received with
a conventional OFDM receiver. To enhance the channel estimation, it will be of bene-
fit for the OFDM receiver to know about the implementation of this diversity tech-
nique. Then the receiver can estimate separately the two channels being present on the
even and odd sub-carriers.
The scheme can be extended to higher numbers of diversity branches M. Efficient im-
plementation as presented in Figure 8-17 is possible if M is an integer power of two.
Note that the correlation function for this technique is equivalent to the correlation
function given by eq. (8-26) (assuming independent component channels). The per-
formance is thus expected to be equal to the performance of the delay diversity tech-
nique, with delays according to eq. (8-27). The simulation results given below confirm
the anticipated behavior.
8.3 Performance Enhancement using Antenna Diversity 277

8.3.3 Performance
Simulation results of frame-error-rates (FER) are given in Figure 8-18a, for the diver-
sity scheme using cyclic delays and for Rayleigh fading channels. The same simula-
tion parameters have been selected as in Section 8.2.4.2. Since the greatest gain is ex-
pected for the quarter-rate transmission mode (due to its small bandwidth) and for the
rms = 10 ns channel (due to the low frequency-selectivity), this case is considered
here. The OFDM system model assumes perfect synchronization and channel estima-
tion.
We investigate the transmitter diversity scheme, where the transmitted power per di-
versity branch was divided by the number of branches, M, in order to maintain a con-
stant total transmit power (see eq. (8-24)). Equivalent results would be obtained for the
receiver diversity schemes without optimizing the delays {i} and weighing factors
{wi}.
Delays of {0,4} and {0,2,4,6} samples were introduced for the two and four-branch
diversity schemes, respectively. Such delays yield zeros in the correlation function at
adjacent coded bits, when applying a depth four interleaver (Interleaver 1, IL 1, from
Section 5.2.4.2). For comparison, the FER is also shown without diversity and for the
special case that all sub-carriers are faded independently.
It is seen that, at a FER of 1 %, a gain of almost 10 dB is possible for the special case
(independent fading) over the single antenna result. About 5 and 7 dB are obtained
with two- and 4-branch diversity, respectively.

Simulation results for the even/odd sub-carrier scheme (two branch) and for com-
parison for a two-branch selection diversity technique are depicted in Figure 8-18b.
The performance of the former is similar to the delay diversity scheme for two
branches (see Figure 8-18a), as expected.
Selection diversity, which is utilized at the receiver (see Figure 8-11), adaptively
chooses on each sub-carrier the signal constellation from the strongest branch. Due to
this adaptivity, about 3 dB gain is obtained, compared to the transmitter diversity
scheme. The slopes of the FER-curves are about equal, however.

8.3.3.1 Performance Evaluation using the Concept of Effective (Eb/N0)

The PDF of the effective (Eb/N0) has been studied by modeling the distribution of the
wide-band average received power of the frequency-selective channel over the band-
width over which the minimum distance error event is spread by the interleaver (Sec-
tion 8.2.3). In this section, we briefly discuss the impact of antenna diversity on this
PDF.
The composite channel for the non-adaptive diversity schemes is the superposition of
M independent wide-band channels. Therefore, we assume that for M-branch antenna
diversity, the wide-band average power is described by the sum of M independent ran-
278 Chapter 8 Performance Evaluation and Enhancement of COFDM

QR; rms = 10 ns; Rc = 1/2 coding, = 5

0
10
no diversity
2branch diversity
4branch diversity
uncorrelated fading
average frame error rate (FER)

1
10

2
10

3
10
2 4 6 8 10 12 14 16 18 20
average E /N [dB]
b 0

(a)
QR; rms = 10 ns; Rc = 1/2 coding, = 5
0
10
no diversity
even/odd SC diversity (2branch)
selection diversity (2branch)
average frame error rate (FER)

1
10

2
10

3
10
2 4 6 8 10 12 14 16 18 20
average Eb/N0 [dB]

(b)
Figure 8-18: Frame error rate results for a rate- convolutional code with a constraint length of
five. QR transmission mode; rms = 10 ns Rayleigh channel. (a): Delay diversity
with nulling of the correlation function at adjacent coded bits. (b): Even/odd sub-
carrier diversity scheme and two-branch selection diversity.

dom variables (RV) that describe the wide-band average power of the component
channels. The analysis of the Eigenvalues for the combined channel correlation func-
tion (8-26) confirms this assumption. Remember that the Eigenvalues specify the vari-
8.3 Performance Enhancement using Antenna Diversity 279

Table 8-5: Parameters of log-normal PDFs (mean and standard deviation in dB) fitted to the
distributions of the effective (Eb/N0) (estimate 1) for several channels and delay-di-
versity settings. IL: Interleaving scheme: IL 1 has depth 4; IL 2 has depth 3; a
depth 9 interleaver is used for the FR-mode (see Section 5.2.4.2). DIV 1: Delays are
chosen such to force zeros in the correlation function at adjacent sub-carriers.
DIV2: Zeros are forced at adjacent coded bits according to the interleaver. DIV 1
is used unless otherwise specified.

Nb. M rms [ns] TX-mode mean stdv. comment

1 1 2.3 4.3 one branch; for reference
2 2 2.0 3.0 IL 1; DIV 2
3 2 10 QR 1.8 2.9 IL 2; DIV 1
4 4 1.8 2.2 IL 1; DIV 2
5 4 1.7 2.2 IL 2; DIV 1
6 1 1.8 2.6 one branch; for reference
7 2 55 QR 1.9 2.0 IL 2
8 4 1.7 1.6 IL 2; stdv. approaches limit
9 - - QR 2.0 1.3 independent fading; for reference
10 1 2.8 3.1 one branch; for reference
11 2 10 FR 2.5 2.2
12 4 2.9 1.8 stdv. approaches limit
13 1 2.7 1.6 one branch; for reference
14 2 55 FR 2.9 1.4 stdv. approaches limit
15 4 3.2 1.2 stdv. approaches limit
16 - - FR 3.3 1.0 independent fading; for reference

ances of independent Gaussian RVs that are added up to model the distribution of the
wide-band average received power (see Appendix D, eq. (D-4)). Investigating the in-
fluence of antenna diversity, it is seen that the original Eigenvalues are replaced by
groups of M equal Eigenvalues of 1/M the original value. (Considering the power-
normalization by 1 / M ). This leads, compared to the non-diversity case, to a PDF
having an equal mean and a variance divided by M (see eqs. (D-8)).
Unfortunately, the exact PDF as defined by eq. (D-7) cannot be used in this case for
modeling the PDF of the composite channels average power, because this equation
requires all Eigenvalues to be different. We do not give here the exact PDF for this
case. In stead, we investigate the influence of M-branch diversity-combining on the
parameters of the approximated PDFs. In particular, we focus on the mean and stan-
dard deviation of the log-normal PDF. Computer simulations of channel realizations
have been used to fit this PDF to the distribution of effective (Eb/N0).
Some general behavior is evident from these parameters, which are listed in Table 8-5.
280 Chapter 8 Performance Evaluation and Enhancement of COFDM

QR; rms = 10 ns; Rc = 1/2, = 5 coding; IL 1; est. 1 of (Eb/N0)eff, bd. 2

0
10
no diversity
2branch; IL 1; DIV 2
2branch; IL 2; DIV 1
4branch; IL 1; DIV 2
4branch; IL 2; DIV 1
1 uncorrelated fading
10
average FER

2
10

3
10
2 4 6 8 10 12 14 16 18 20
average Eb/N0 [dB]

Figure 8-19: Theoretical performance results in terms of average FER, obtained by the concept
of effective Eb/N0.

For instance, the standard deviation for M-branch diversity is well predicted by divid-
ing its dB-value without diversity (M = 1) by M . This empirical rule only works
however, until the standard deviation for the special case that all sub-carriers fade in-
dependently is approached (see lines 9, 16 in the table). This limit can not be exceeded
(lines 8, 12, 14, 15). The mean values (in dB) remain almost constant.
Figure 8-19 depicts performance results for the quarter-rate receiver over the
rms = 10 ns channel. The results were derived by using the concept of effective (Eb/N0)
with the log-normal PDFs for the parameters listed in Table 8-5, lines 15, and 9.
These curves are to be compared to Figure 8-18a, where good agreement is evident.
Moreover, two delay settings and interleaving schemes are compared, which are seen
to result in largely equivalent performance.

8.3.4 Discussion of the Diversity Schemes

A considerable performance improvement is the main advantage of the techniques in-
vestigated, as seen from the performance results. Basically, the transmitter diversity
schemes can be employed without modifying the receiver. However, care must be
taken with synchronization and channel estimation techniques and to some extent with
the interleavers. That is, the parameters of all system components should be carefully
chosen to minimize interference effects among them. The following points have to be
considered.
Channel estimation schemes sometimes exploit the correlation between the channel
coefficients of adjacent sub-carriers in order to reduce the noise floor of the channel
8.3 Performance Enhancement using Antenna Diversity 281

estimates. Eliminating this correlation by the suggested diversity techniques, it is clear

that such principles would perform sub-optimally or fail. For instance, in the channel
estimation scheme proposed for the OFDM system under investigation, the even-in-
dexed sub-carriers are determined by interpolating between the two adjacent (odd-in-
dexed) sub-carriers, which are estimated from pilots in the training symbol (see Sec-
tion 6.4). This interpolation must fail, if correlation among adjacent sub-carriers lacks.
Timing synchronization offsets introduce a progressive phase rotation, as shown in
Section 4.2.4. This progressive phase rotation can be accurately determined from a
known training symbol, yielding an accurate estimate of the residual timing offset (see
Section 6.2.7). Again, high correlation among the sub-carriers has to be assumed in
order to achieve good performance in presence of a frequency-selective channel. This
correlation is largely destroyed, however, by the diversity technique. A possible rem-
edy for this issue is to calculate the timing-offset estimate over the still-correlated sub-
carriers. (The ones spaced by M sub-carriers, if the {i} are selected according to the
criterion given in eq. (8-27).)
In Schmidls method [12] (see Section 6.2.3), (frame) timing-synchronization is done
by looking for a unique training symbol, which carries information only on the even-
indexed sub-carriers. The odd sub-carriers are zero. Such symbols do not occur in a
conventional OFDM signal, unless they are explicitly inserted. Using the even/odd
sub-carrier diversity technique, however, signals with very similar properties may be
evident at the receiver, whenever one of the signals is strong while the other one is
deeply attenuated. If known data is modulated on the training symbol, the frame start
can be confirmed by demodulation. Note that this problem may also occur for the two-
branch delay diversity technique, if the {i} are selected according to eq. (8-27), due to
the equivalent correlation properties.
The above issues only apply if the diversity schemes are utilized at the transmitter. At
the receiver, synchronization and channel estimation can be done prior to the diversity
combining, as shown in Figure 8-14.

The amplitude distribution of the combined channel is another point to be considered.

If the component channels are Rayleigh distributed then the sum of the channels will
be Rayleigh distributed as well, and performance gain is obtained due to the enhanced
frequency-diversity. The depth of the fades wont be influenced.
On channels with shallower fades for instance Ricean channels with considerably
high K-factors , the performance is generally improved due to the lower probability
of deep fades. Applying the proposed diversity schemes in this case, the combined
channel (see eq. (8-24)) would have deeper fades due to the randomization. A possible
performance gain by the non-adaptive techniques is therefore questionable.
282 Chapter 8 Performance Evaluation and Enhancement of COFDM

8.4 Conclusions and Recommendations

In Section 8.2, the concept of effective Eb/N0 [1] was successfully applied to OFDM
based wireless ATM transmission systems. It enables performance evaluation in terms
of bit and frame error rates for coded and interleaved data transmission over fading
channels.
A novel extension to this concept is proposed. By determining the PDF of the effective
Eb/N0, the calculation of outage probabilities and average error probabilities is en-
abled. Based on an analytical approach investigating the PDF of the wide-band-aver-
age received power of the frequency-selective Rayleigh and Ricean fading channels,
appropriate models have been found for these distribution functions. The log-normal
PDF can serve as an approximation, as seen from simulation results. The parameters of
the PDF of the effective Eb/N0 are derived analytically from the channel model except
for a correction factor in the range of 0 2.5 dB. This factor must be determined by
channel simulations, or it can be roughly estimated based on the observation of the
typical channel behavior. A description of this correction factor as a function of system
and channel parameters is subject for further research. This would yield a fully ana-
lytical method of performance evaluation.
The methods developed enable the fast and accurate comparison of coding and inter-
leaving techniques. The distribution of the wide-band-average power is also useful for
the analysis of other OFDM system components. For instance, the outage probability
of synchronization algorithms can be evaluated, as their performance depends on the
signal power received over the total bandwidth of the OFDM signal.
The performance results show clear improvements with increased frequency-selectiv-
ity of the channel (longer delay spreads), and with increased system bandwidth. (When
the fading distribution remains unchanged, e.g. Rayleigh.) These factors raise the fre-
quency-diversity that can be exploited by the coding scheme. Accordingly, worst per-
formance corresponds to the case where all OFDM sub-carriers fade equally (flat fad-
ing), and best performance is obtained for independently fading sub-carriers. These
special cases have been evaluated. About 12 dB of gain are evident between them, at
an average frame-error-rate of one percent. Results for practical channels lie in-be-
tween these borders.
Comparing the analytical results of outage probability (FER > 102) to the results of
average FER remarkable resemblance is evident. At the selected threshold FER and
coding scheme, an SNR gap of about 1 2 dB is the only difference between
equivalent levels of outage probability and average FER. The way these performance
measures are related to the PDF of the effective Eb/N0 leads to the conclusion that this
is a quite general property.

Novel antenna diversity schemes have been proposed to improve the frequency-diver-
sity in the undesirable case that all sub-carriers fade (about) equally (flat-fading). Con-
8.5 References 283

siderable diversity gain can be obtained by appropriately combining the signals of

multiple receiver antennas, or by transmitting appropriate signals via multiple transmit
antennas.
The first scheme discussed is a delay diversity technique employing cyclic delays to
avoid inter-symbol-interference. Applied at the transmitter (where no channel infor-
mation is available), the frequency-diversity of the channel is increased. Applied at the
receiver, channel estimates for individual diversity branches can be employed to opti-
mize the diversity combining. This optimization is a subject for further investigations.
The second scheme transmits sub-sets of the sub-carriers via different antennas, which
also improves the frequency-diversity. Both schemes require uncorrelated fading
channels, i.e., the antennas must be sufficiently separated in space, by at least about
one wavelength.
A common feature of the proposed techniques is their extremely low computational
complexity, compared with conventional diversity schemes that would required a
multiplication of main parts of the OFDM receiver including the FFTs. The novel
schemes work with the time-domain OFDM signals in one case, and in the other case,
partial FFTs are executed for each diversity branch.
Generally, the transmitter diversity techniques can be used with a conventional OFDM
receiver. Care must be taken, however, with synchronization and channel estimation
algorithms.
The results analyzed show about 5 and 7 dB gain for two and four diversity branches,
compared to the single antenna results, for the schemes that apply no optimization.
Such large gains are only feasible in almost flat channels, however. The gains are re-
duced in channels that already have more frequency-diversity.

8.5 References
[1] S. Nanda and K. M. Rege, Frame error rates for convolutional codes on fading
channels and the concept of effective Eb/N0, IEEE Trans. Veh. Technol., vol. 47,
no. 4, pp. 12451250, Nov. 1998.
[2] R. van Nobelen and D. P. Taylor, Analysis of the Pairwise Error Probability of
Noninterleaved Codes on the Rayleigh-Fading Channel, IEEE Trans. Commun.,
vol. 44, no. 4, pp. 456463, Apr. 1996.
[3] J. K. Cavers and P. Ho, Analysis of the Error Performance of Trellis-Coded
Modulations in Rayleigh-Fading Channels, IEEE Trans. Commun., vol. 40, no.
1, pp. 7483, Jan. 1992.
[4] P. Ho and D. Fung, Error Performance of Multiple-Symbol Differential Detec-
tion of PSK Signals Transmitted over Correlated Rayleigh Fading Channels,
IEEE Trans. Commun., vol. 40, no. 10, pp. 15661569, Oct. 1992.
284 Chapter 8 Performance Evaluation and Enhancement of COFDM

[5] M. Sandell, Design and Analysis of Estimators for Multicarrier Modulation and
Ultrasonic Imaging. Ph.D. Thesis, Lule University of Technology, Sept. 1996.
[6] J. G. Proakis, Digital Communications, 3rd edition. New York: McGraw Hill,
1995.
[7] C. Lee, Convolutional Coding: Fundamentals and Applications. Boston: Artech
House, 1997.
[8] G. C. Clark, J. B. Cain, Error-Correction Coding for Digital Communications.
New York: Plenum Press, 1981.
[9] Y. Li, J. C. Chuang, and N. R. Sollenberger, Transmitter Diversity for OFDM
Systems and Its Impact on High-Rate Data Wireless Networks, IEEE J. Select.
Areas Commun., vol. 17, no. 7, pp. 12331243, July 1999.
[10] M. Okada and S. Komaki, Pre-DFT Combining Diversity Assisted COFDM,
IEEE Trans. Veh. Technol., vol. 50, no. 2, pp. 487496, March 2001.
[11] A. V. Oppenheim and R. W. Schafer, Discrete-Time Signal Processing, 2nd Edi-
tion, Upper Saddle River, NJ: Prentice Hall, 1999.
[12] T. M. Schmidl and D. C. Cox, Robust frequency and timing synchronization for
OFDM, IEEE Trans. Commun., vol. 45, no. 12, pp. 16131621, Dec. 1997.
Chapter 9 Conclusions and
Recommendations

In this chapter, the main conclusions (C1C63) of this dissertation are summarized,
and recommendations for further research (R1R13) are given.

9.1 Part I: Channel Characterization

9.1.1 Modeling of the Frequency-Selective Radio Channel

A second order, wide-sense stationary stochastic model is presented to characterize the
frequency-selective transfer function of wide-band radio channels.
The model describes the small-scale behavior of the channel within a local area. It is
defined by a fixed set of model parameters, which can be related to physical channel
parameters. Different parameter-sets may apply at different local areas. A parameter
set relates to a constant power delay profile, which assumes that the ray-magnitudes do
not change abruptly for small displacements of the transmitter or receiver.
For a limited observation bandwidth, however, the channel parameters do change also
within a local area, because the channel impulse response is not fully resolvable. That
is, multiple impinging rays interfere at the resolvable delay time-bins, whose duration
relates to the inverse of the observation bandwidth.
C1 A fixed set of channel parameters can characterize the small-scale fading within a
local area, corresponding to a fixed set of parameters in the quasi wide sense sta-
tionary uncorrelated scattering (QWSSUS) -type of channel model used.
C2 Due to a limited observation bandwidth, channel parameters appear to vary
within a local area.

285
286 Chapter 9 Conclusions and Recommendations

An exponentially decaying delay power spectrum, the mathematical description of the

time-dispersive and frequency-selective radio channel, agrees well with the expected
physical behavior of a mm-wave indoor radio channel, where the transmitter and the
receiver are typically located within the same room.
C3 The actual channel impulse response has little impact on stochastic channel pro-
perties, like the level crossing rate of the channel transfer function, for a given set
of channel parameters (average power, Ricean K-factor, and RMS delay spread).
C4 Therefore, it is concluded that the proposed channel model can be applied in
other frequency-bands and environments as well.

9.1.2 Channel Measurement Technique using the FD-Level Crossing Rate

C5 The RMS delay spread (RDS) of a multipath radio channel is proportional to the
frequency-domain level crossing rate (LCRf) of the frequency-selective channel
transfer function.
For Rayleigh fading channels, the general proof of this relation is given. For Ricean
channels, the proof is based the channel model studied in Chapter 2.
R1 The derivation of the general proof of the relation given in C5 for Ricean fading
channels is a subject for further research.

The relation of C5 can be used to estimate the RDS from measurements of the power
transfer-function of the channel versus frequency, which is sufficient to determine the
LCRf. The Ricean K-factor and the average received power can be estimated from
such power measurements as well, leading to complete sets of parameters for the
channel model proposed in Chapter 2. That is, a novel channel measurement procedure
has been developed, using simplified non-coherent measurement equipment, because
only the power transfer function of the channel needs to be scanned.
C6 Complete sets of channel parameters can be estimated from non-coherent
measurements of the channels (power) transfer function.
C7 The relationship and the actual proportionality factor used for estimating the
RDS are insensitive to changes in the channel model or channel impulse re-
sponse.
C8 For Rayleigh fading channels, the relationship is completely independent of the
channel impulse response.

Measurement noise tends to increase the LCRf, leading to overestimation of the RDS.
This issue has to be taken into account when applying the proposed technique for
channel measurements.
R2 A basic analysis of the problem is given and some possible solutions are indi-
9.2 Part II: OFDM System Proposal and Evaluation 287

cated, but there is room for further work.

C9 To limit the influence of noise on the novel measurement technique, the sampling
interval in the frequency-domain has to be selected as large as possible.

R3 Another topic for further research is an elaborate comparison of the proposed

novel measurement technique to conventional techniques and to other ways of
data processing.

The level crossing rate in the frequency domain (LCRf) depends in very distinct ways
on the channel parameters average power, Ricean K-factor, and RMS delay spread.
This leads to the conclusion that those parameters comprise a most significant set of
parameters for characterizing the frequency-selective radio channel.
C10 The RMS delay spread and the Ricean K-factor are equally important for charac-
terizing the frequency-selective radio channel.

9.2 Part II: OFDM System Proposal and Evaluation

9.2.1 OFDM System Modeling

The derivation of the OFDM system model has confirmed that using the OFDM
modulation technique, data symbols can be transmitted independently over multipath
radio channels.
C11 There is no interference among subsequent OFDM symbols (inter-symbol-inter-
ference ISI) or among adjacent sub-carriers (inter-carrier-interference ICI), if
the following conditions are fulfilled: (Otherwise interference is introduced lea-
ding to some performance degradation.)
The channel impulse response must be shorter than the guard interval.
Time-synchronization must be sufficiently accurate: all multipath compo-
nents must remain within the guard interval.
No carrier frequency-offset, carrier phase-noise, Doppler spread, and samp-
ling frequency-offset.
All system components must be linear.

Using this system model and the channel model of Chapter 2, the bit-error-rate (BER)
of an uncoded OFDM system has been evaluated for coherent and differential detec-
tion and for various modulation schemes.
C12 The performance generally depends on the Ricean K-factor of the channel and on
the average received signal power. Increasing any of these parameters reduces
the average BER.
288 Chapter 9 Conclusions and Recommendations

The fading channel induces error floors, particularly if differential detection is applied.
C13 For the delay spread and time-variability encountered in a wide-band indoor
communications system, differential detection in the time-direction is much supe-
rior to differential detection in the frequency-direction, because the channels
time-variability between subsequent OFDM symbols is lower than the variation
among adjacent sub-carriers.
The equations introduced are a basis for evaluating other degrading effects as listed
above. However, more research is required for the implementation and comparison of
all these factors.

9.2.2 OFDM Air-Interface and Multiple Access Scheme Proposal

A novel OFDM based air-interface and multiple access scheme is proposed for a
wireless communications system. Transmission rates up to 155 Mbit/s can be sup-
ported in slowly time-variant channels at 60 GHz.
C14 Time-division duplexing enables asymmetric data rates.
C15 The application of various coding and modulation schemes provides adaptability
to varying channel conditions.

A so-called quarter-rate mode simultaneously supports simplified terminals operat-

ing at a quarter of the full system bandwidth.
C16 It is recommended to implement frequency hopping in the quarter-rate mode to
avoid performance degradation due to the decreased frequency-selectivity caused
by the reduced bandwidth.
C17 The quarter-rate mode can be used by terminals with largely reduced hardware
requirements, and as a fall-back mode when the link-budget becomes critical.

Training-symbols are periodically transmitted on the down-link for robust synchroni-

zation and channel estimation; pre-equalization is employed on the up-link.
C18 This makes a highly efficient multiple access scheme, where little overhead is
needed to separate multiple users in time or frequency. Small data entities like
single ATM cells can thus be supported.
C19 A periodically transmitted training symbol embedded in a fixed frame structure
allows for robust synchronization (on the down-link), introducing very little
overhead (1.42.8 %).
C20 Pre-equalization used on the up-link leads to a highly efficient multiple access
scheme and to simplified data detection algorithms at the base station with low
additional complexity at the mobile.
9.2 Part II: OFDM System Proposal and Evaluation 289

C21 The multiple access scheme described is optimized for a centralized scheduled
multiple access control (MAC) protocol, which makes the system attractive for
applications where multiple constant-rate wide-band data streams are present.
Examples are wireless video or audio recording studios.

The strictly hierarchical structure using base stations and mobile terminals may be a
disadvantage, as it will pose big difficulties when the proposed air-interface is utilized
for ad-hoc networks.
R4 The co-channel interference among cells with equal carrier frequencies is a topic
for further research, in this context.

A major drawback of the system is its limited range.

R5 Future work may focus on diversity techniques, adaptive antennas, or turbo cod-
ing in order to enhance the range.

9.2.3 Issues in utilizing the 60 GHz Frequency-Band

C22 Shadowing is a critical problem in a mobile system operating in the 60 GHz
band, because the radio link may drop when the line-of-sight (LOS) path gets ob-
structed by the user. Even the leaves of a tree may block the LOS (in outdoor en-
vironments), causing link failure.
C23 Multiple base-station antennas at well-separated locations within a room may re-
duce this problem.

R6 Additional research has to address the challenges of hardware technologies re-

quired for the 60 GHz RF-front-ends. Amplifier linearity and oscillator phase
noise are important issues in OFDM, where strict specifications have to be ful-
filled with low-power and low-cost devices.

9.2.4 Synchronization and Channel Estimation on the Down-Link

The key synchronization steps for the proposed OFDM system are described and
evaluated in Chapter 6. A training symbol (TS) periodically transmitted on the down-
link enables robust synchronization at low computational complexity. The overhead
introduced is small.
C24 Multiple, subsequently performed synchronization stages enable the estimation
of all relevant synchronization-offsets and their correction, using the training se-
quence and the pilot sub-carriers.

By processing the time-domain TS, coarse timing and frequency synchronization is

acquired. The TS is then transformed to the frequency-domain, where its known data
is used for refining the estimates of the synchronization offsets.
290 Chapter 9 Conclusions and Recommendations

C25 The performance of the synchronization steps is sufficient for the system pro-
posal considered, at average signal-to-noise ratios (SNR) down to about 0 dB.
I.e., the system can stay synchronized even if the SNR required for reliable data
communication is not available any more.
C26 Further enhancement is possible by appropriately combining offset estimates ob-
tained from multiple subsequent training symbols.

Synchronization may be lost in severe shadowing situations.

R7 Emphasis should be put on the development of algorithms that allow the trans-
mission system to quickly recover from outages in such cases.

C27 Large simplifications in the required processing steps are achieved by locking all
local oscillators and the sampling clock to a common (tunable) frequency source.
Doing so, all frequency-offsets are cancelled simultaneously by synchronizing
for just one of them.

The analysis of a fine timing-offset estimation scheme is presented, which is original

work. The estimates standard deviation is given for AWGN channels as a function of
the SNR and as a function of parameters of the TS, allowing for some optimization of
the TS. Unfortunately, a multipath radio channel deteriorates the estimation accuracy.
A bias is introduced, which can be expressed in terms of the channel parameters, em-
ploying the channel model of Chapter 2.
C28 In dispersive channels, time-synchronization schemes show a positive bias,
which is for Rayleigh channels roughly equal to the RMS delay spread of the
channel.
C29 In order to compensate for this bias, the FFT-period of the OFDM signal should
be cyclically shifted to start a few samples before the end of the guard interval.

C30 It has been discovered that DC-offsets and carrier feed-through may have a
significant impact on the performance of a popular class of frequency-synchroni-
zation algorithms, which evaluate correlation sums over periodic parts in the
(time-domain) OFDM signal.
C31 A simple extension to these algorithms can eliminate the induced performance
degradation in the most relevant cases.

The training symbol can be used to obtain a channel estimate, as known data is carried
on it.
C32 Efficient channel estimation can be performed using this training symbol, due to
the relatively slow time-variability of the channel.
9.2 Part II: OFDM System Proposal and Evaluation 291

C33 (Wiener) filtering in the frequency-domain can reduce the mean-square-error of

the estimated channel transfer functions.

However, the achieved BER performance improvement is small. Moreover, adaptabil-

ity becomes an issue because a Wiener filter has to be optimized for the current chan-
nel conditions.
C34 It has been realized that timing offsets must be considered in the design of the
Wiener filter, i.e., the bias of the timing-offset estimation scheme has to be taken
into account.
C35 In practice, non-adaptive filters can be used giving sub-optimum results.
C36 At small additional loss, filtering can be fully omitted.
C37 Due to the small gain achieved with enhanced channel estimation techniques at
drastically increased complexity, their application is not recommended.

9.2.5 Pre-Equalization for the Up-Link

Data transmission on the up-link is based on pre-equalization, where the frequency-
selectivity of the radio channel is (partly) compensated at the transmitters side using
the channel estimate from the down-link.
C38 Pre-equalization eliminates the need for training symbols (on the up-link), tur-
ning the communications system more efficient, particularly if short data packets
are transmitted.
C39 Coherent detection becomes possible at reduced complexity, because channel es-
timation is not required.
C40 Accurate timing-offset estimation and synchronization are needed on the up-link.
Furthermore, the carrier phase offset and the magnitude of the received symbols
have to be estimated. (The latter is only required if higher order quadrature am-
plitude modulation (QAM) schemes are applied.)

The radio channel is assumed quasi-static and reciprocal.

R8 The assumption of channel reciprocity in a practical hardware set-up (the equiva-
lence of the channel transfer functions on the up- and down-links) should be con-
firmed by an experimental study.

C41 Pure phase pre-equalization can be utilized for phase modulation techniques,
where just the phase distortion is compensated.
R9 Optimum power assignment can be added to minimize the average BER. This is
a topic for further research.
292 Chapter 9 Conclusions and Recommendations

C42 Full (phase and magnitude) pre-equalization is required for higher order QAM
schemes.
C43 The main objective of the technique described in this thesis is to maintain a con-
stant transmit power. Therefore, deeply attenuated sub-carriers are not used for
data, as they would require too much of the transmit power. Error correction cod-
ing can recover this data at the receiver.

C44 The average uncoded BER performance on the up-link is similar to the BER of
the down-link, assuming a static channel.
R10 In further research, error correction coding and the channels time-variability
should be included in the performance analysis.

R11 Pre-equalization can be combined efficiently with adaptive modulation. This is

another topic for further research.

9.2.6 Emulation System and OFDM Implementation

The emulation system developed is a downscaled hardware platform that can be used
for the demonstration of wide-band air-interface techniques like OFDM.
C45 Synchronization aspects can be implemented and studied in real-time, as the sys-
tem introduces realistic carrier and sampling frequency-offsets between the
transmitter and the receiver. Other hardware impairments are present as well, like
DC-offsets and carrier feed-through.
C46 The system specifications are suitable for the implementation of OFDM air-inter-
faces regarding the linearity and carrier phase noise.
C47 Software implementation difficulties are largely avoided due to the drastically re-
duced processing speed.

The key algorithms of the investigated OFDM air-interface have been implemented on
the emulation system.
C48 The implementation of synchronization algorithms has shown that time- and fre-
quency-synchronization are achieved at sufficient accuracy, resulting in excellent
performance results.
C49 This demonstrates that the orthogonality of data symbols can be maintained suffi-
ciently well, in a practical, implemented OFDM transmission system.

C50 The sampling frequency-offset between the transmitter and the receiver (of 20
50 ppm) has to be considered in the signal processing algorithms of the OFDM
receiver. Its estimation is rather time-consuming.
C51 Thus it is recommended to omit this critical step and implicitly correct for the
9.2 Part II: OFDM System Proposal and Evaluation 293

sampling frequency-offset by locking the sampling clock on the easily synchro-

nized mixer oscillator, as described above.

C52 The reduced bandwidth of the quarter rate terminals implies reduced frequency-
diversity, thus the robustness and the performance are degraded, compared with
full-rate receivers.

C53 The impact of DC-offsets and carrier feed-through on the correlation-based fre-
quency-synchronization algorithm, and the effectiveness of the enhanced, insen-
sitive algorithm, have been validated experimentally with the emulation system.

The performance limitation due to the time-variant channel has been studied.
C54 Error floors are sufficiently low to be corrected by forward error correction cod-
ing, under the low-mobility conditions assumed. However, the margins are small,
thus simple channel prediction (linear extrapolation) should be implemented, al-
lowing for significant enhancements at low additional complexity.

9.2.7 Performance Evaluation and Enhancement of COFDM

The concept of effective Eb/N0 can be applied for the performance evaluation of
convolutionally coded and interleaved OFDM schemes in frequency-selective chan-
nels. In this method, the simulated channel transfer function is converted to a scalar
value, the effective Eb/N0, which is then related to the (bit or frame) error probability.
C55 The PDF of the scalar effective Eb/N0 enables the analysis of average error
probabilities and outage probabilities. This PDF is a function of channel and
OFDM system parameters, including the average SNR, the signal bandwidth, and
the interleaving scheme of the OFDM system.
Analytical approaches have been developed for the derivation of this PDF, however, a
multiplicative factor remains uncertain, which should be determined by means of
computer simulations.
R12 Deriving an analytical expression for this parameter is a topic for further re-
search, which would turn the performance evaluation scheme to a fully analytical
one.
C56 From the way average error probabilities and outage probabilities are derived
from the PDF of the effective Eb/N0, it has been concluded that both performance
curves as a function of the SNR are very similar. A constant gap between these
curves depends on the error rate for which the outage probability is calculated.

C57 The performance of an OFDM system improves when the delay spread of the
channel increases, because more frequency-diversity can then be exploited. (Un-
der the condition that the maximum excess delay remains shorter than the guard
294 Chapter 9 Conclusions and Recommendations

interval.)
C58 Systems with larger bandwidth also show some performance advantage, for the
same reason.

Novel, computationally efficient antenna diversity schemes are proposed. Their main
idea is the introduction of cyclic delays to the effective parts of OFDM symbols
transmitted/received via separate antennas.
C59 The novel antenna diversity schemes increase the frequency-diversity in situa-
tions where the channel is relatively flat, or where the system bandwidth is small.
Significant performance gain is thereby obtained.
C60 The principle can be applied at the transmitter, at the receiver, or at both.
C61 The IFFT/FFT blocks are not duplicated; thus the complexity increase is low.

C62 Applied at the receiver, the delay times can be optimized based on estimates of
the channel transfer functions for each antenna.
C63 Complex-valued weight factors can be introduced.
R13 The development of optimum combining strategies and their performance com-
parison to conventional diversity techniques (which require multiple FFTs at the
receiver) are topics for further research.
Appendix A Correlation Coefficient for
the Discrete Impulse Response

In order to relate the RMS delay spread rms of a multipath channel described by its
discrete impulse response (IR) to the frequency-domain level crossing rate (LCRf), the
correlation coefficient c has to be calculated from the IR. This is the goal of this ap-
pendix. The derivation starts with a review of the definitions of the channels IR, trans-
fer function (TF), and RMS delay spread (cf. (2-1), Section 2.2.2). We show in this ap-
pendix that, in the limit F 0 , the correlation coefficient can be approximated by

c 1 (2 rms F ) 2 . (A-1)
It is remarkable that the channel IR has no influence on this equation, suggesting that
for the Rayleigh fading channel the proportionality factor relating the LCRf to rms is
independent of the channel model or channel IR. This is a very important observation
for the application of the relation to channel measurements.
This result can be generalized to any band-limited, Rayleigh distributed, WSS sto-
chastic process. It is proven that there exists a fixed function of r', relating the level
crossing rate to the second centralized moment of the normalized power spectrum (or
periodogram) of the (zero-mean) complex Gaussian process underlying the Rayleigh
process. This appendix focuses on the application of the relation to channel measure-
ments, without loosing generality.

A-1 Definitions
The static, time-dispersive (= frequency-selective) radio channel is described by its
complex, lowpass equivalent IR

295
296 Appendix A Correlation Coefficient for the Discrete Impulse Response

L 1
h( ) = i e
j i
( i ) , (A-2)
i =0

where {i}, {i}, and {i} are the path gains, phases, and delay times, respective, and L
is the number of paths. The time-invariant TF is obtained from the IR via the Fourier
transform.
L 1
H ( f ) = ie
j [ 2f 1 + i ]
(A-3)
i =0

The similarity of this equation to Rices sum of sinusoids suggests that H(f) could be
the underlying complex Gaussian process of any Rayleigh process R( f ) = H ( f )
(compare [1], (3.7-2) (3.7-4)).
The RMS delay spread rms is defined from the IR as

rms = 2 2 , (A-4)

where k = ( L 1 k
i2
i =0
) ( L 1
i =0
)
i2 , k = {1,2}. rms is seen to be the second centralized
moment of the normalized power delay profile. For simplified notation we introduce
pi = i2 (i=0
L 1
)
i2 , yielding k = i=0 k pi .
L 1

A-2 Calculation of the Correlation Coefficient

The correlation coefficient c = 1 02 is derived from the autocorrelation function
2

of the underlying complex Gaussian process, m = 12 E{Z n Z n*+ m }. In the case of the fre-
quency selective radio channel, this function is called the spaced-frequency correlation
function, being

m = 12 E{H ( f ), H * ( f + m F )}=
L 1 j 2 i m F
1
2 i =0
pi e , (A-5)

where F is the sampling interval in the frequency-domain.

To calculate m for one particular channel realization, the expectation operator should
be considered as the average over the frequency f. But this equation also holds for a set
of channels with common stochastic properties, or for any Rayleigh process, where the
right-hand side of the equation is the Fourier transform of the process (normalized)
power spectrum (compare [1], eq. (3.7-11)). The squared magnitude of m is
A-3 References 297

L 1 L 1
m = m m* = 14 pi e j 2 i m F pk e j 2 k m F
2

i =0 k =0
L 1
(
)
L 1 L 1
= 14 pi2 + pi pk e j 2 ( i k ) m F + e j 2 ( i k ) m F . (A-6)
i =0 i = 0 k =i +1
L 1 L 1 L 1

= 14 pi2 + 2 pi pk cos 2 ( i k )m F
i =0 i = 0 k =i +1

For F 0 (for 1 ) and F = 0 (for 02 ), the cos-term can be replaced by

2

cos x = 1 x 2 2 and cos 0 = 1 , respectively. This yields

1 14 pi2 + 2 pi pk (1 2 2 ( i k ) 2 f s2 )
2 L 1 L 1 L 1

i =0 i = 0 k =i +1
. (A-7)
1 2
L 1 L 1
= 4 4 0 4 f s pi pk ( i k )
2 2 2

i = 0 k =i +1

Noting that, due to the normalization of the {pi}, 02 = 14 , we obtain c 1

4 2 F 2 i =0 k =i+1 pi pk ( i k ) 2 . It remains to be shown that the double sum in this
L 1 L 1

expression is equal to rms

2
in order to get (A-1).
L 1
L 1 L 1
rms
2
= i2 pi i pi k pk
i =0 i =0 k =0
L 1 L 1 L 1 L 1
(A-8)
= i2 pi i2 pi2 2 i k pi pk
i =0 i =0 i =0 k =i +1

The double sum in c can be expanded to

( )
L 1 L1 L 1 L 1 L 1 L 1

pi pk ( i k ) = + k2 pi pk 2
2 2
i i k pi pk
i = 0 k =i +1 i =0 k =i +1 i = 0 k = i +1
L 1 L 1 L 1 L 1
= 2
i p i p k 2 i k pi p k , (A-9)
i =0 k =0 , k i i =0 k =i +1
L 1 L 1 L 1
= i2 pi (1 pi ) 2 i k pi p k
i =0 i = 0 k =i +1

L 1
where the last step follows from k =0
pk = 1 . The equivalence is readily seen from
these expressions.

A-3 References
[1] S. O. Rice, Mathematical Analysis of Random Noise, Bell Syst. Tech. J., vol.
23, pp. 282332, July 1944; vol. 24, pp. 46156, Jan. 1945
Appendix B FD-Level Crossing Rate in
the Presence of Noise

Goal of this appendix is to quantify the impact of noise on the frequency-domain (FD)
level crossing rate (LCRf). Moreover, it is desirable to separate the channels and the
noises contributions to the LCRf.

B-1 Derivation of the FD-Level Crossing Rate

The level crossing rate is related to second order statistical properties of the complex
Gaussian random process underlying the mathematical channel model. In the case of
the FD-channel model, this random process is the channel transfer function without the
line-of-sight component , written as H'(f) = rH(f) + jiH(f). The required parameters
can be calculated from the auto-covariance of the channels TF, H ' (f ) (eq. (3-34),
without 2):
~
H ' (f ) = H ' (f ) + PN sinc( f / F );
j 1f e j 21f , (B-1)
H ' (f ) = 1 sinc( 1f )e +
+ j 2f
where { ,1,,} are parameters of the FD-channel model (see Section 2.3.2), PN is
the variance of the additive noise process (see Section 3.4.1), and F is the FD sampling
interval. For notational convenience we introduce u = 1, specifying the shape of the
delay power spectrum of the channel, and u1 = u + 1 , u 2 = u 2 2 + u + 1 , and
u 3 = u 3 3 + u 2 + 2u + 2 . Furthermore, the normalized noise power N' = PN/(20) =
PN/(u1) is used.

In order to obtain the LCRf, we have to evaluate the first derivative of r H 'i H ' (f ) at

299
300 Appendix B FD-Level Crossing Rate in the Presence of Noise

f = 0, written 0 , and the second derivative of r H' (f ) , written 0 , where

H ' (f ) = 2[ r H ' (f ) + j r H 'i H ' (f )] ([1][3]; see Section 3.2.1). Due to the noises sym-
metric spectrum, 0 is not influenced by the additive noise ( 0 = u 2 2 ), and
0 becomes

2 N 2 u3 2 N ' u1
~

=
 = 2 . (B-2)
3 6 F 2
0 0
6 F2

It is evident from this equation that the sampling interval F has an important influence
on this expression. Increasing F by a certain factor has the same effect as decreasing
the noise power N' by the squared factor (while N' << 1, to be exact).
Using an expression from [3], the LCRf is written
~ r2 + 2
~ r 2 2~0 / 2
N R (r ) = 3 / 2 ~ e
0
r
{ ~
}
cosh ~ cos e ( sin ) + ~ sin( ) erf(~ sin ) d
2
, (B-3)
0 0

0 ~ ~ 0 , and ~ = 1 ~ (0) = 1 u (1 + N ' ) . This equation

2
where ~ = , = 
~ ~
0 0 2 H' 2 1
~0 2 0

can be evaluated straightforwardly. However, in order to separate the influence of

noise on the LCRf, two approximations are introduced.

B-2 Approximation 1
For simplifying the expressions involved, it is appropriate to assume that N' << 1, i.e.
~ ~
(1 + N') 1, yielding K K , P0 P0 , ~
r ' r ' , and ~0 0 . Analytical results (see [4])
have shown that this approximation causes errors below 1 % at practical values of r',
and for N' 0.01. As in Section 3.2.1, we define the variables {a', b, c, d} from the
elementary terms of (B-3). With the above approximations and the definitions from
Sections 2.3.3 and 3.4.1, these are
~
r 2 4 u3 u 22 1 N'
a' = 3 2 = r' K + 1 2 2 + (B-4a)
0 u1 u1 12 F 2
r2 + 2
b= = r ' 2 ( K + 1) + K (B-4b)
2 0
r
c= = 2r ' K ( K + 1) (B-4c)
0
N ' 2 2
d = ~ = K u 2 u1u3 u 22 + u1 . (B-4d)
12 F 2

The terms with the factor N ' F 2 in {a', d} account for the noises influence. These are
B-3 Approximation 2 301

also the expressions that depend on rms, since rms is proportional to 1/ (cf. Section
2.3.3)
2
1 1 u3 1 u2
rms = . (B-5)
K + 1 u1 ( K + 1) u1 2
2

Analytical evaluations of eq. (B-3) have shown rather small impact of the noise-term
on the integral. For 0 = N ' ( F rms ) 2 , the integral changes by about 50 % for K
= 10 and u = , and even less for lower K-factors and lower values of u.
Because of the small influence of the rest of the equation, we shall concentrate on the
term a' to investigate the relation of the LCRf and rms for the noise-afflicted case. Us-
ing (B-5), a' becomes

u1u3 u 22
r ' (K + 1)
4 N' 1
a' =
32
2 + , (B-6)
( K + 1)u1u3 u 2
2 rms
12 F ( K + 1) 2
2

leading to an expression of the following form for the LCRf

~ N'
N R( a1) (r ' ) = A rms
2
+ 2 B g ( r ' , K , u, ) . (B-7)
F
As mentioned above, the factor g(r',K,u,), which includes the integral expression, is
mainly determined by the parameters {r',K,u}; it hardly changes when is varied be-
tween 0 and .

B-3 Approximation 2
B N' B N'
Studying the asymptotic behavior of (B-3) in the limits rms
2
>> 2
and rms
2
<<
AF A F2
leads to the second simplification. In these limits, the sum-of-root expression of (B-7)
can be replaced by the root of only one of the two terms. The first case means that the
noise is negligible and yields the known proportionality between LCRf and rms (see
(3-5))
~
N R( a1) (r ' | N ' = 0 or rms ) =
. (B-8)
rms A g (r ' , K , u , = 0) = rms f ( K , u, r ' )

The second case is interpreted as the noise becoming dominant for the LCRf, thus no
relation between the LCRf and rms is evident in the following equation. As a result, the
estimation of rms must fail for such a measurement.
302 Appendix B FD-Level Crossing Rate in the Presence of Noise

~
N R( a1) (r ' | N ' or rms = 0) =
N'B N' (B-9)
g ( r ' , K , u , = ) = h( r ' , K )
F F
It can be shown that rms = 0 (or ) yields d = 0, which eliminates from
g(r',K,u, = ) the dependency on u (24). Thus the LCRf is solely determined by the
noise parameters and the K-factor.
Bringing g(r',K,u,) for these two special cases under the square root, i.e.

~ N'
N R( a 2) (r ' ) = A rms
2
g 2 (r ' , K , u , = 0) + B 2 g 2 (r ' , K , u, = ) =
F
(B-10)
N'
= rms
2
f 2 (r ' , K , u ) + 2 h 2 (r ' , K )
F
results in an expression having the same asymptotic behavior (for rms = 0 and
rms ) as (B-7). Comparing the two approximations numerically shows that, for
any rms and for most values of r' and u, the maximum relative error between them is
(well) below 1 %. Only if r' << 1 or r' < 0.5 and u > 1, it increases up to ~4 %.
This yields the compact relation given by (B-10) and by (3-35), between the LCRf, rms
and N ' F .

B-4 References
[1] S. O. Rice, Mathematical Analysis of Random Noise, Bell Syst. Tech. J., vol.
23, pp. 282332, July 1944; vol. 24, pp. 46156, Jan. 1945.
[2] S. O. Rice, Statistical Properties of a Sine Wave Plus Random Noise, Bell Syst.
Tech. J., vol. 27, pp. 109157, 1948.
[3] M. Ptzold, U. Killat, F. Laue and Y. Li, On the Statistical Properties of
Deterministic Simulation Models for Mobile Fading Channels, IEEE Trans.
Veh. Technol., Vol. 47, No. 1, pp. 254269, Feb. 1998.
[4] K. Witrisal and A. Bohdanowicz, Influence of Noise on a Novel RMS Delay
Spread Estimation Method, in Proc. PIMRC 2000 (Symposium on Personal In-
door Mobile Radio Communications), London, Sept. 2000, pp. 560566.

24
To prove this statement, note that 1 = 0 and = , for rms = 0.
Appendix C Analysis of Fine Timing-
Offset Estimation

In this appendix, the performance of the fine-timing offset estimation technique out-
lined in Section 6.2.7 is analyzed. After a brief review of the method and introduction
of the system model (Section C-1), we derive the estimation bias for Ricean (and
Rayleigh) channels (Section C-2). Only in case of non-dispersive (AWGN) channels,
the estimate is unbiased.
The standard deviation of the estimates is analyzed in Section C-3 for the AWGN
channel, and in Section C-4 for Rayleigh channels. The Ricean case is not elaborated.
However, the necessary modifications to the analysis are indicated.

C-1 Review of Estimation Technique and System Model

The OFDM system model in presence of a small timing synchronization error is given
in Section 4.2.4.1. It is written as

yi , k = xi ,k hi ,k e j 2it / TFFT + ni ,k = xi ,k hi ,k e j 2it '/ N + ni ,k , (C-1)

where t and t' are the timing offsets in seconds and in samples, respectively. {xi,k}
and {yi,k} are the transmitted and the received symbol constellations, {ni,k} are inde-
pendent Gaussian noise samples, and {hi,k} are the channel coefficients, where i and k
are the sub-carrier and symbol indices, respectively. The index i can take values in the
range i = {N/2, N/2+1, , N/21}. TFFT is the FFT-period [s] and N is the number
of FFT-points.
This system model assumes that no inter-symbol-interference is caused by the timing-
offset. That is, the channel impulse response including the timing-offset must be
shorter than the guard interval.
It is evident that a timing-offset gives rise to a progressive phase rotation of the signal

303
304 Appendix C Analysis of Fine Timing-Offset Estimation

constellations. The phase rotation is zero at the center frequency and it linearly in-
creases towards the edges of the frequency band. The method of estimating the re-
maining timing-offset (after coarse frame-synchronization) is based on detecting this
progressive phase rotation.
Known data is differentially modulated on the training symbol (TS) in frequency-di-
rection, as defined in Section 6.2.1. The first step towards the estimation of t is the
differential demodulation of the TS-data, written as

y n ,l = yl*( n 1),TS y l n ,TS = xn ,l hl*( n 1),TS hl n ,TS e j 2lt / TFFT +

. (C-2)
xl*( n 1),TS hl*( n 1),TS e j 2l ( n 1)t / TFFT nl n ,TS + xl n ,TS hl n ,TS e j 2l nt / TFFT nl*( n 1),TS + nl n ,TS nl*( n 1),TS

In this equation, l is the frequency-separation between differentially modulated sub-

carrier pairs. The data symbols carried on them are denoted xn ,l = xl*( n1),TS xl n ,TS . The
index n = {1, 2, , Nm} is another index on the sub-carriers carrying data25, where Nm
is the number of sub-carrier pairs. For the TS described in Section 6.2.1, l = 2, and
{xn,l} = {1} (cf. Section 6.2.7). TS indicates OFDM symbols, which are used as
training symbols.
It is seen that the phase of the data symbols is corrupted by noise, by the channel coef-
ficients, and by the systematic phase-rotation due to the timing-offset. Note that the
channel coefficients of adjacent sub-carriers are correlated, thus hl*( n1),TS hl n,TS | hl n,TS |2
| hl ( n 1),TS | 2 plus some (complex valued) noise.
The metric Mint(e), which is used for integer frequency-offset correction, provides an
efficient way of accurately estimating the timing-offset t. Assuming that the integer
frequency-offset has been corrected, the metric at its peak position is written
Nm
M int (eopt ) = mn y n ,l . (C-3)
n =1

As multiplying the symbols yn,l by the (conjugated) data sequence mn = xn*,l removes
the phase modulation, this metric is an accumulation of the energy of all differentially
demodulated symbols. Its phase corresponds to the phase offset induced by the timing-
offset. Therefore, the estimate of the timing offset is derived from
TFFT
t = M int (eopt ) [s]. (C-4)
2l
The channel also induces a systematic phase rotation leading to an estimation bias, as

25
Note that with the notation used in (C-2), the sub-carriers are indexed from 0 to lNm, not from
N/2 to N/21. This modified indexing largely simplifies the notation. It also (implicitly) introduces a
systematic common phase rotation, which, however, does not have any effect on the problem under
investigation.
C-2 Estimation Bias over Dispersive Channels 305

analyzed in the following sub-section. The noise terms are zero-mean, i.e., they do not
cause any systematic error.

C-2 Estimation Bias over Dispersive Channels

Unfortunately, the estimate of the timing offset (C-4) is corrupted by the channel TF,
expressed by the product hl*( n1),TS hl n,TS in eq. (C-2), which also causes a systematic
phase offset. The expectation of that term is seen to be the spaced-frequency correla-
tion function of the channel. For a wide sense stationary channel (in frequency-do-
main) it is written

{ }
E hl*( n 1),k hl n ,k = H (lF ) . (C-5)

Assuming a Ricean channel model with an exponentially decaying delay power spec-
trum and a line-of-sight (LOS) component at (excess) delay time = 0, it is possible to
relate this correlation function to channel parameters, namely, the RMS delay spread
(rms) and the Ricean K-factor (K). Using results from Chapter 2 (see eq. (2-22)), we
obtain

P0 1
H (lF ) = K + , (C-6)
K +1 1 + j 2lF rms K1

where K1 = ( K + 1) 2 K + 1 , and P0 is the normalized received power P0 = E{|hi,k|2}.

The phase angle of this term is
2lF rms K1 2l rms
{ H (lF )} = atan . (C-7)
K [1 + (2lF rms K 1 ) ] + 1
2
TFFT 2 K + 1

The approximation is based on the following relations.

The denominator in the atan-term can be simplified to (K+1), because (2lFrmsK1)2 <<
1 at small l. This relation is satisfied for OFDM systems whose design is based on a
guard interval length TGI TFFT/4. Remember that the guard interval itself is selected
in order to cope with the dispersive channel, thus the maximum excess delay of the
channel max TGI. For the exponentially decaying delay power spectrum, this maxi-
mum excess delay is given as (approx.) max = 10 rms K1 (see Section 2.3.3.3), hence,
10 rms K1 TFFT / 4 , which yields with F = 1/TFFT that FrmsK1 1/40.
Secondly, atan(x) x is used, which is valid for |x| << /2. In order to fulfill this con-
dition, 2l rms TFFT << 2 must be given, i.e., rms << TFFT 4l . This is again proven via
the above relations. Taking into consideration that K1 1 leads to rms TFFT / 40 . That
is, the approximation can be used for small l.

Due to this systematic phase offset, the timing-offset is biased by

306 Appendix C Analysis of Fine Timing-Offset Estimation

rms
E{t} t , (C-8)
2K + 1
in presence of a multipath channel with parameters rms and K.

Note that correction for this bias minimizes the progressive phase-rotation induced by
the channel. This is desirable, for instance, if differential detection in frequency-direc-
tion is applied (see Section 4.3). We also use this result in the investigation of the
channel estimation scheme in Section 6.4.
As the start-time of the FFT-period is estimated too late by this algorithm, the maxi-
mum bias should be considered in the OFDM system design in order to prevent inter-
symbol-interference. This is achieved by cyclically shifting the FFT-period (of the data
symbols and the training symbol) so that it actually starts a few samples before the end
of the guard interval (see Figure 6-11 in Section 6.2.7).

C-3 Estimation Variance on the AWGN Channel

Goal of this section is the derivation of the variance (and standard deviation) of the es-
timated timing offset for the case of the AWGN channel, where the channel transfer
function is expressed as hi,k = 1. The noise variance on the sub-carriers is written
E{|ni,k|2} = 22. Furthermore, it is assumed that the magnitude of the data symbols is
unity, i.e., |xi,k| = 1. These definitions lead to the SNR per (modulated) sub-carrier be-
ing SNRSC = 1 /(2 2 ) .
The following considerations relate this SNR to the time-domain SNR, which is given
in eq. (6-3). The total signal power 2 s2 is shared among NSC active (modulated) sub-
carriers, while the total noise power 2 n2 is distributed over the N FFT-points. There-
fore,

2 s2 N 1 N
SNR = = SC 2 = SC SNRSC . (C-9)
2 n N 2
2
N

For deriving the standard deviation of the estimate t , we assume without loss of
generality that the timing-offset is zero. Noise is therefore the only cause of a devia-
tion of the metric Mint(eopt) from the real axis. This leads to the following approximate
expression for the desired standard deviation, where atan(x) x for |x| << /2 was
used.

TFFT TFFT var{Im[M int (eopt )]}

t = var{atan[M int (eopt )]} (C-10)
2l 2l E[ M int (eopt )]

The main problem is thus to calculate the variance of the imaginary part of Mint(eopt).
C-3 Estimation Variance on the AWGN Channel 307

Applying the above-introduced definitions, the metric Mint(eopt) is written

(x )( )
Nm
M int (eopt ) = *
l ( n 1),TS + nl*( n 1),TS xl n ,TS + nl n ,TS xl ( n 1),TS xl*n,TS =
n =1
(C-11)
(1 + x )(1 + x )
Nm
* *
l ( n 1),TS n
l ( n 1),TS l n ,TS nl n,TS
n =1

For simplified notation, we introduce nn = nl n ,TS xl*n ,TS . Remember that xl n,TS = 1 , there-
fore, the variance of nn is 22, equal to the variance of nln,TS.
The next step is to isolate the imaginary part of Mint(eopt), which follows from
Nm
M int (eopt ) = (1 + Re[ nn1 ] j Im[nn1 ])(1 + Re[nn ] + j Im[nn ]). (C-12)
n =1

Note that both, the real and imaginary parts of {nn} are independent, zero-mean Gaus-
sian random variables with variance 2. Therefore, the expectation of the above ex-
pression is E{Mint(eopt)} = Nm.
Separating the imaginary part yields

[ ]
Nm
Im M int (eopt ) = Im[n0 ] + Im[n N m ] + (Re[nn1 ] Im[nn ] Re[nn ] Im[nn 1 ]) . (C-13)
n =1

The variance of this random variable is 26

[ ]
var{Im M int (eopt ) } = 2 2 + 2 N m 4 . (C-14)

We finally obtain with (C-10)

T 2 2 + 2 N m 4 TFFT SNRSC + N m 2
t FFT = . (C-15)
2l Nm 2l N m SNRSC

Replacing TFFT by N (number of FFT-points) yields the standard deviation in samples.

A number of observations are made from (C-15). At high SNR (SNRSC >> Nm/2), the
standard deviation decreases with 1 SNRSC . At low SNR, it decreases with 1/SNRSC.
The number of demodulated sub-carrier pairs Nm (at high SNR) and their separation l
decrease the standard deviation proportionally. Note, however, that the SNRSC de-
creases (at fixed SNR) if the number of modulated sub-carriers NSC is increased (see
eq. (C-9)). In total the standard deviation (at high SNR) can be reduced by a factor of

26
The complete analysis has shown that the variance of the real part is not equal to the variance of
the imaginary part. The former is given as var{Re[Mint(opt)]} = (2Nm1)22 + 2Nm4. In the
expression for the imaginary part, many of the Im[nn]-terms are cancelled due to the negative sign in
(C-12).
308 Appendix C Analysis of Fine Timing-Offset Estimation

approximately q , when l is increased by the factor q, while Nm (and NSC) is de-

creased by 1/q, keeping the SNR constant. That is, the estimation performance is better
when a training symbol with less but farther spaced, sub-carriers is used. On the other
hand, the unambiguous estimation range is thereby reduced.
Computational results are presented in Section 6.2.7 (Figure 6-10) for the proposed
OFDM transmission scheme in modes I-qr and I-fr. It is seen from this figure that, for
the given design of the training symbol and detection technique, the performance over
the AWGN channel allows for an excellent standard deviation below 0.1 samples at an
SNR of 10 dB.
The performance degrades, however, over multipath channels. This behavior has been
anticipated, as it is generally hard to define an exact arrival time of a signal that has
been transmitted over a time-dispersive channel. The error variance over Rayleigh
fading channels is analyzed in the sub-section below.

C-4 Estimation Variance on Rayleigh Fading Channels

To simplify the analysis of the error variance due to the multipath channel, noise is ne-
glected. The metric then becomes
Nm
M int (eopt ) = hl*( n 1),TS hl n,TS e j 2lt / TFFT . (C-16)
n =1

The expectation of this term demonstrates that the estimate, which is obtained from the
phase of Mint(eopt), is biased, since a systematic phase offset is introduced by the chan-
nel. This offset corresponds to the phase of the channels spaced-frequency correlation
function at lag lF, H(lF) (see Section C-2). The phase is not equal to zero for an
asymmetric channel delay profile, as for instance for an exponentially decaying delay
spectrum (see Section 2.3.3.2). The expectation of Mint(eopt) is

{ }
E{M int (eopt )} = N m E hl*( n 1),TS hl n ,TS e j 2lt / TFFT = N m H (lF )e j 2lt / TFFT . (C-17)

For the analysis, we assume that the systematic phase rotation due to time-offset t ex-
actly cancels the phase rotation due to the channel. This is achieved by letting the
time-offset be t = TFFT (2l ) H (lF ) . The estimation variance induced by the chan-
nel is now evident from deviations of the metric from the real axis. Thus, eq. (C-10)
can be applied again to calculate the estimation variance.
For simplified notation, we introduce hn = hl n ,TS e jn H ( lF )
, which leads to
Nm
M int (eopt ) = hn*1hn , (C-18)
n =1

incorporating the above-mentioned time-offset. We furthermore separate the channel

coefficients hn into their real and imaginary parts, written hn = an + jbn. The correlation
C-4 Estimation Variance on Rayleigh Fading Channels 309

functions

Rh [k ] = E{hn*k hn } = H (lkF )e jkH (lF ) =

(C-19)
E{(an k jbn k )(an + jbn )} = 2 Raa [k ] + 2 jRab [k ]

will be required in the analysis below.

Note the similarity of the problem analyzed to the problem of calculating the variance
of a correlation function estimated from Nm samples of a discrete random process.
These computations are for instance found in [1]. In our case, however, the variance of
the imaginary part is of interest, which is very different to the variance of the complex
correlation function that is given in [1].
The variance of the imaginary part of the metric is obtained from

Nm Nm
{ [ ]}
E Im 2 M int (eopt ) = E Im hn*1hn Im hm* 1hm =
n=1 m =1
Nm Nm

E (a n1bn a nbn 1 ) (am 1bm am bm 1 ) = . (C-20)
n =1 m =1
Nm Nm

E{a
n =1 m =1
b a m1bm + an bn 1am bm1 an1bn am bm 1 a n bn1a m1bm }
n 1 n

Discussing Rayleigh fading channels, the variables an and bn are real-valued, zero-
mean Gaussian random variables. We can therefore apply [2]
E{wxyz} = E{wx}E{ yz} + E{wy}E{xz} + E{wz }E{xy} , (C-21)
which leads to
N m 1
{ [
E Im 2 M int (eopt ) = 2 ]} (N m | k |)
k = ( N m 1) , (C-22)
{2R 2
ab [1] Rab [k + 1]Rab [k 1] Raa [k + 1]Raa [k 1] + R [k ] + R [k ]
2
aa
2
ab }
where k = m n. Finally, the variance of the imaginary part of the metric is written as

[ ] { [ ]}
var{Im M int (eopt ) }= E Im 2 M int (eopt ) E 2 {Im M int (eopt ) }=[ ]
1 N m 1
{ (
( N m | k |) Rh [k ] Re Rh [k + 1]Rh* [k 1]
2 k = ( N m 1)
2
)}
. (C-23)

This expression leads to the standard deviation with eqs. (C-10), (C-17), (C-19), and
(C-6). Noise can be re-introduced by adding the error variances due to the channel and
due to the noise. Independence of the noise processes has to be assumed thereby.
Computational results are depicted and discussed in Section 6.2.7, Figure 6-10.
310 Appendix C Analysis of Fine Timing-Offset Estimation

C-5 References
[1] C.W. Therrien, Discrete Random Signals and Signal Processing. Englewood
Cliffs: Prentice Hall, 1992.
[2] A. Papoulis, Probability, Random Variables, and Stochastic Processes, 3rd edi-
tion. New York: McGraw-Hill, 1991.
Appendix D PDF of Wide-band Average
Received Power

D-1 PDF of Average Received Power for Rayleigh Fading

Channels
This section describes the derivation of the PDF of wide-band average received power
of a frequency-selective Rayleigh fading channel. Starting with a discrete-frequency,
complex lowpass equivalent TF, the normalized received power averaged over a given
bandwidth is written
1 N 1
Pa =
N i =0
| H (iF ) |2 , (D-1)

where F is the spacing between FD-samples, and N is the number of samples averaged.
Thus the bandwidth considered is approximately BW = NF. In matrix notation, we get

Pa = x *T x, x = 1 / N [H (0), H ( F ),..., H (( N 1) F )]
T
(D-2)
The elements of x are identically distributed, zero-mean, complex Gaussian random
variables (RV). They are correlated as specified by the spaced-frequency correlation
function of the channel H(f) (see Section 2.4.3 and (D-19)). It is assumed that the
channel is wide-sense stationary (WSS) in the frequency domain, corresponding to the
property of uncorrelated scattering in time domain. The complex-valued auto-correla-
tion matrix for x is defined as

H (0) H (F )  H (( N 1) F )

{ }
( F )
=
H (0)  H (( N 2) F )
R xx = E xx*T
   
H
. (D-3)


H (( N 1) F ) H (( N 2) F )  H (0)

311
312 Appendix D PDF of Wide-band Average Received Power

The eigenvalue decomposition can be applied to diagonalize the correlation matrix.

This yields a vector x' = E*Tx of independent, zero-mean, complex Gaussian RVs (be-
cause the elements of x are zero-mean Gaussian; see [1], pp. 50 ff.), where E is a ma-
trix whose columns are the eigenvectors. The eigenvalues i are the variances of the
elements of x' = [x'0 x'1 x'N 1]. They are real-valued because the correlation matrix
is hermitian symmetric (see e.g. [1], p. 50). With this decomposition, our original
problem (D-2) becomes
N 1
Pa = x' E Ex' = x' x' = | x'i | 2 ,
*T *T *T
(D-4)
i =0

because E*TE = I. For simplification, terms with (variance) i << max (k ) may be
k

omitted because their contribution to the sum is insignificant.

The squared-magnitude of a complex Gaussian RV pi = |x'i|2 with variance 2 i2 = i
has an exponential PDF with characteristic equation ([2], pp. 41 ff.)
1 1
pi ( ) = = . (D-5)
1 j 2 i 1 j i
2

The PDF of a sum of random variables is obtained from the product of their charac-
teristic equations ([2], p. 36), i.e.
N 1 N 1
1
Pa ( ) = pi ( ) = . (D-6)
i =0 i = 0 1 j i

Applying the inverse Fourier transform yields the PDF 27

N 1
exp ( x / i ),
1 Ai
d =
jx
p Pa ( x ) = ( )e (D-7)
2 i
Pa
i =0

N 1
where Ai = 1 k = 0 ,k i
(1 k / i ) .

The characteristic function is used to find the first and second moments of the PDF.
These are

d Pa ( ) N 1
E{Pa } = j = i
d =0 i =0

d 2 Pa ( )
2

{ }= ( j) N 1
N 1
EP 2 2
= + i .
2
(D-8)
d 2
a i
=0
i =0 i =0

{ }
N 1
= E P (E{Pa }) = 2i
2 2 2
Pa a
i =0

27
This result requires that all eigenvalues are different of another. It was obtained by replacing the
product of fractions in eq. (D-6) by a sum of fractions.
D-1 PDF of Average Received Power for Rayleigh Fading Channels 313

D-1.1 Approximations
Some commonly used standard PDFs were compared to the exact PDF (D-7) for de-
creasing the number of parameters from N (i.e. number of eigenvalues) down to two or
three. Another advantage of such approximations is that their parameters usually give
a much better indication of the shape of the distribution function than a vector of
eigenvalues (e.g. the mean and standard deviation of a log-normal PDF in dB). The fit-
ting of the following PDFs was tested:
The generalized chi-square distribution (having a non-integer degree of freedom)
The log-normal distribution
The Gaussian distribution
The Gaussian PDF (central limit theorem) is an appropriate approximation only in case
of averaging over extremely large bandwidths. One problem is that the Gaussian PDF
is not restricted to positive values, which must be the case for power values.
The Weibull distribution was tried as well, but it has a thicker tail than the genera-
lized chi-square distribution, whose tail is already thicker than the exact distributions.

D-1.1.1 Generalized Chi-square Distribution (Squared Nakagami)

Because the PDF of a sum of nf independent, identically distributed, squared Gaussian
RVs is described by a chi-square distribution of nf degrees of freedom, it is reasonable
to try to describe the sum of dependent squared Gaussian RVs by a chi-square PDF
with a generalized (non-integer) nf-parameter. The chi-square PDF is written as
1 n f / 2 1
pY ( y ) = y exp( y / 2 2 ), y 0. (D-9)
2 (n f / 2)
nf nf / 2

The parameter in this equation is the standard deviation of the underlying Gaussian
RVs.
It can be shown that this PDF is equivalent the PDF of a squared Nakagami random
variable: (It is noted that the Nakagami distribution is an amplitude PDF, thus squaring
makes sense to obtain a power PDF.) The Nakagami distribution is written [2]
m
2 m 2 m1 mr 2 /
p R (r ) = r e , (D-10)
( m)

where and m are defined as

{ }
= E R2
. (D-11)
m = E{( R2 2
) 2 }
The ratio of moments m is called the fading figure. The PDF of the squared amplitude
Y = R2 is thus
m
1 m m 1 my /
pY ( y ) = y e . (D-12)
(m)
314 Appendix D PDF of Wide-band Average Received Power

It can be seen that this distribution is equivalent to a chi-square distribution with 2m

degrees of freedom (i.e. m = nf/2), substituting for /m = 22.

The general method used for fitting two analytical PDFs compares their first v
moments, where v is the number of free parameters to be determined. In case of the
generalized chi-square PDF, the number of parameters is two, nf and . The first and
second moments are:

E{Y } = n f 2 (= )
E{Y 2 } = 2n f 2 + n 2f 4 (D-13)
Y2 = E{(Y E{Y }) 2 } = 2n f 2 (= 2 / m)

It is seen that these two moments correspond to the definition of and m of the Naka-
gami distribution (see expressions in brackets and (D-11)). The fitting is accomplished
by comparing these expressions to equations (D-8).

D-1.1.2 Log-normal PDF

Parameters for the log-normal PDF can be obtained in two different ways. Method
number one matches the first two moments of the exact PDF (see eqs. (D-8)) to the
first two moments of the log-normal PDF (moments on linear scale). The log-normal
PDF is written as28

f Z ( z) =
1
(
exp (ln z mln ) 2 2 ln2 , ) (D-14)
ln z 2

its first and second moments are

E{Z } = exp( mln + ln2 2)

. (D-15)
E{Z 2 } = exp( 2mln + 2 ln2 )

Comparing these expressions to eqs. (D-8) finally yields

(
N 1
mln = ln i =0 i ) + (
2 N 1
i =0
2
i
N 1
i =0
)2
i
. (D-16)

(
ln2 = ln i =0 i2
N 1
) ( ) + 1
N 1
i =0 i
2

Since power values are often written in [dB], it is convenient to convert the parameters
mln and ln to dB-values. This is accomplished by using

mln [dB] = mln 10 / ln(10)

. (D-17)
ln [dB] = ln 10 / ln(10)

28
Computing the logarithm of realizations of a log-normal random process results in a Gaussian
(normal) process. The parameters mln and ln are the mean and standard deviation of this Gaussian ran-
dom process.
D-1 PDF of Average Received Power for Rayleigh Fading Channels 315

It will be seen from the computational results that the log-normal PDF for these para-
meters does not match well at the lower tail of the PDF. Reason for this is the non-
symmetry (skewness) of the log-normal PDF when it is investigated over a linear
scale. It leads to a matching of the moments but does not take special care about values
in the lower tail of the PDF.

An alternative approach transforms the exact PDF to a logarithmic scale in order to fit
the resulting distribution to a Gaussian PDF. The parameters for this Gaussian PDF are
obtained from integrating over the exact PDF (D-7) yielding

mln = E{ln( Pa )} = + i =0 Ai ln( i )

N 1

. (D-18)
+ m = E{ln ( Pa )} = 2 + i =0 Ai [ln (i ) 2 ln(i )]
2 2 2 N 1 2
ln ln

The Euler constant = 0.5772157.

D-1.2 Analytical and Simulation Results

Since it was observed that the actual properties of the (small-scale) fading channel
model do not severely influence the fading characteristics in the frequency domain 29,
an exponentially decaying model for the DPS is used in all computer simulations and
analytical computations.
For such a channel, with parameters {rms, K, P0}, the correlation function required for
calculating the eigenvalues is given as (see Section 2.4.3, eq. (2-22))

P0 1
H (f ) = E{H * ( f ) H ( f + f )} = K + , (D-19)
K +1 1 + j 2f rms K1

where K1 = ( K + 1) 2 K + 1 , and f is the frequency-separation.

It is seen that the correlation function is uniquely defined by the rmsf product for a
particular set of parameters {K, P0}. The PDF of wide-band average power will thus
depend on the product of rms and the observed bandwidth (BW), and on the Ricean K-
factor. P0 just shifts the PDF along the power-axis. For Rayleigh fading channels, K =
0, hence we investigate here the PDF for different rmsBW products.

D-1.2.1 Computational Results

In Figure D-1, results are presented in form of CDFs with a normal probability scale
on the ordinate and with a logarithmic (dB) scale on the abscissa. Log-normal PDFs
thus appear as straight lines. Our analysis particularly concentrates on the lower tail of
the distribution functions, which is most important in communications applications
where one is usually interested in the probability that the average power is below a

29
The characteristics defined by the level crossing rate, coherence bandwidth, and similar properties
derived from second order stochastic parameters (see Section 3.2 and Appendix A), provided that the
channel parameters {rms, K, P0} are not changed.
316 Appendix D PDF of Wide-band Average Received Power

Table D-1: Errors and levels of the different approximations at probability values (CDF level)
of 10-2 and 10-3. All values in [dB].

CDF rms BW exact generalized log-normal log-normal Gaussian

prob. PDF chi-square (log fit) (linear fit)
level level error level error level error level error
0.027 -19.1 -19.8 -0.7 -15.2 3.9 -9.9 9.2
0.11 -15.7 -18.9 -3.2 -13.7 2.1 -9.7 6.0
10-2 0.43 -10.1 -13.2 -3.1 -9.4 0.7 -8.0 2.1
1.7 -5.6 -6.5 -1.0 -5.5 0.04 -5.1 0.5
7.0 -2.8 -3.1 -0.2 -3.2 -0.3 -2.7 0.1 -4.0 -1.2
0.027 -25.9 -28.7 -2.8 -19.3 6.6 -12.7 13.3
0.11 -21.2 -27.5 -6.3 -17.4 3.8 -12.4 8.8
10-3 0.43 -13.4 -19.1 -5.6 -12.1 1.3 -10.3 3.1
1.7 -7.4 -9.2 -1.8 -7.2 0.2 -6.6 0.8
7.0 -3.8 -4.2 -0.4 -4.2 -0.4 -3.6 0.2 -7.0 -3.2

certain threshold value. Shown are the computer simulation results (for 1000 channel
realizations), the exact CDF obtained from eq. (D-7), the generalized chi-square appro-
ximation, two log-normal approximations, and the Gaussian PDF. All results are de-
rived for an NRP of P0 = 0 dB.
Firstly it is observed that the exact PDF does not fail to model the simulated distribu-
tion curve. The tail of the chi-square PDF is generally too thick, whereas the tails of
the log-normal approximations are too thin.
It is seen that different approximations are more or less suitable for different rmsBW
products. For very small bandwidths (rmsBW < 0.05), the chi-square PDF for nf = 2
degrees of freedom is an excellent choice, which is a trivial result since the average
power will turn to a squared Rayleigh PDF in this case.
For larger bandwidths (0.05 < rmsBW < 4), the log-normal PDF with parameters ob-
tained using eq. (D-18) (designated logfit) yields the best results. This range of
rmsBW is most likely encountered in practical systems like indoor wireless LANs.
For even higher bandwidths (rmsBW > 4), the log-normal distribution with parameters
obtained using eq. (D-16) gives better approximations at probability levels in the range
of [10-3 1]. At lower probability values, the second log-normal PDF is still superior.
The Gaussian PDF becomes satisfactory only for very large rmsBW products >> 10.
Table D-1 lists the dB-values of the different PDFs at probability values of 10-2 and 10-
3
, and the errors of the approximations. The error is defined as the distance in dB
between the approximation and the exact curve at the specified probabilities.
From these results it is concluded that small approximation errors < 1 dB at 10-3 are
only possible if rmsBW > 1. Except for the exact PDF, there is no simple standard
distribution function that models the wide band average power well in any case of
rmsBW. The lognormal (logfit) PDF gives good results over a large (and practical)
range of parameters. Moreover, the parameters of the lognormal PDF are convenient to
D-1 PDF of Average Received Power for Rayleigh Fading Channels 317

CDF of wideband average power for Rayleigh fading and BWrms 0.027 CDF of wideband average power for Rayleigh fading and BWrms 0.11
0.99 0.99
CDF for data CDF for data
0.95 exact CDF 0.95 exact CDF
chisquare CDF chisquare CDF
0.9 0.9
lognormal lognormal
0.8 lognormal (logfit) 0.8 lognormal (logfit)
Gaussian Gaussian
Pr(P < abscissa)

Pr(P < abscissa)

0.5 0.5
a

a
0.2 0.2

0.1 0.1
0.05 0.05

0.01 0.01
0.005 0.005

0.001 0.001
25 20 15 10 5 0 5 25 20 15 10 5 0 5
average signal power, Pa [dB] average signal power, Pa [dB]

(a) (b)
CDF of wideband average power for Rayleigh fading and BWrms 0.43 CDF of wideband average power for Rayleigh fading and BWrms 1.7
0.99 0.99
CDF for data CDF for data
0.95 exact CDF 0.95 exact CDF
chisquare CDF chisquare CDF
0.9 0.9
lognormal lognormal
0.8 lognormal (logfit) 0.8 lognormal (logfit)
Gaussian Gaussian
Pr(Pa < abscissa)

Pr(P < abscissa)

0.5 0.5
a

0.2 0.2

0.1 0.1
0.05 0.05

0.01 0.01
0.005 0.005

0.001 0.001
15 10 5 0 5 8 6 4 2 0 2 4 6
average signal power, Pa [dB] average signal power, Pa [dB]

(c) (d)
CDF of wideband average power for Rayleigh fading and BWrms 7
0.99
CDF for data
0.95 exact CDF
chisquare CDF
0.9
lognormal
0.8 lognormal (logfit)
Gaussian
Pr(Pa < abscissa)

0.5

0.2

0.1
0.05

0.01
0.005

0.001
7 6 5 4 3 2 1 0 1 2 3
average signal power, Pa [dB]

(e)
Figure D-1: CDFs of the wide-band average power for Rayleigh fading channels and different
rmsBW (RMS delay spread times bandwidth) products

compare different results. If very accurate results are required, the exact PDF should
be used if rmsBW < 1. In this case, the number of eigenvalues remains small, which
allows for a reasonably simple evaluation of the expressions involved (see eq. (D-7)).
318 Appendix D PDF of Wide-band Average Received Power

D-2 PDF of Average Power for Ricean Fading Channels

The derivation of the PDF of wide-band average power for Ricean Channels follows
the same principle as applied for the Rayleigh channel. The problem is defined
~
equivalently, i.e., Pa = x*T x (compare eqs. (D-1), (D-2)). However, the elements of x
are now non-zero mean, identically distributed, correlated, complex Gaussian random
variables. Again, an eigenvalue decomposition is applied to transform the random
~
vector x with correlated elements to a random vector ~x = E *T x with uncorrelated and
thus statistically independent elements30. This is desirable for the derivation of the
~
PDF from the sum of random variables. The matrix of eigenvectors E must be derived
from the covariance matrix C xx = E{(x m)(x m)*T } so that the elements of the result-
ing vector ~x are uncorrelated (m is the vector of mean values of the elements of x).
~ *T
~ =E
The mean values of ~x are given as m m ; its elements variances are defined by
~
the eigenvalues i of E . Again our original problem becomes (compare to eq. (D-4))
N 1
x = | ~
~ ~ ~
Pa = ~
x *T E*T E~
x=~
x *T ~ xi | 2 . (D-20)
i =0

In case of Ricean channels, the summation terms are squared, non-zero mean, complex
Gaussian RVs ~pi =| ~xi |2 . The characteristic function of such a RV is found as the char-
acteristic equation of a non-central chi-square PDF of 2 degrees of freedom

1 j | m ~ |2 1 j | m~ |2
~pi ( ) = exp

i
= exp i
, (D-21)
1 j 2 i
2
1 j 2 i
2
1 j i 1 j i

~ is the complex-valued
where i2 is the variance of the underlying Gaussian RV and mi

mean of the i-th component of x. The PDF of the sum of RVs (the average power) is
obtained from the product of their characteristic equations,
N 1 1
N 1
j | m~ |2
P~ ( ) = p~i ( ) =
exp i
. (D-22)
a
i =0 i =0
1 j i 1 j i
To the knowledge of the author it is not possible to calculate the PDF for this charac-
teristic equation analytically. A numerical implementation was thus realized, which
applies the FFT algorithm for the required Fourier transform. A high number of FFT
points is needed to obtain the PDF over a sufficiently high dynamic range (215 points
correspond to ~45 dB of power range).
The first three centralized moments were derived analytically by evaluating the de-
rivatives of the characteristic equation.

30
Only in case of Gaussian random variables it is given that uncorrelated random variables are also
statistically independent (see [1], p. 58).
D-2 PDF of Average Power for Ricean Fading Channels 319

~
{ }
mP~ = E Pa = i =0 i + m
N 1 ~
i ( ) 2

= E {(P E{P }) }= ( + 2 m~ )
a

~ ~ 2 N 1
P2~ 2 2
a a i =0 i i i (D-23)
= E {(P E {P }) }= (2 + 6 m
~ )
a

~ ~ 3 N 1
1, P~ 3 2 2
a
a a i =0 i i i

These expressions will be used for fitting standard PDFs to the complex exact equa-
tion.

D-2.1 Approximations
The fitting of the following PDFs was tested for the Ricean case:
The generalized, non-central chi-square distribution
The log-normal distribution
The Gaussian distribution

D-2.1.1 Generalized, Non-central Chi-square Distribution

The reasoning for using this PDF as an approximation for the PDF of wide-band aver-
age power is equivalent to the reasoning given for the generalized (central) chi-square
PDF in the Rayleigh case. Because of averaging non-zero mean Gaussian random
variables, the non-central version of the chi-square distribution is selected, which
describes the PDF of a sum of nf magnitude squared Gaussian RVs with nonzero mean
i. It is written
( n f 2) / 4
1 y ( s 2 + y ) / 2 2 s
pY ( y ) = e I n f / 2 1 y 2 , y 0, (D-24)
2 2 s 2

where nf is the degree of freedom, s 2 = i =0 i2 , and I(x) is the -th order modified
n f 1

Bessel function of the first kind. We consider a generalized version of this PDF, where
nf can be non-integer. The first three (centralized) moments are

mY = E{Y } = n f 2 + s 2
{ }
Y2 = E (Y E{Y }) = 2n f 4 + 4 2 s 2
2
(D-25)
1,Y = E {(Y E{Y }) }= 8n + 24 4 s 2
3 6
f

Two methods were assessed for obtaining the three unknown parameters of this PDF.
The first one calculates s2 from the sum of the magnitude squared means of the
elements of ~x , or equivalently, the sum of the means of the elements of x (because the
eigenvalue decomposition is a unitary transformation); i.e., s 2 = i =0 m = i =0 mi .
N 1 ~ 2 N 1 2
i

Next, nf and 2 are calculated by comparing the first two moments, yielding

2 = i =0 i2 + 2i m
N 1 ~ (
i
2
) 2 ( + 2 m~ ).
N 1
i =0 i i
2

(D-26)
n f = i =0 (i ) 2
N 1
320 Appendix D PDF of Wide-band Average Received Power

Results for this method are designated chi-square CDF, M1, 2 in the results section.
The second method compares the first three moments, resulting in a quadratic equation
for the parameter 2.
2
P2~a P2~ ~
12, 2 = a 1, Pa (D-27)
2mP~a 2mP~ 8mP~
a a

Usually, only one of the solutions is positive valued and thus taken for 2. In some
cases, however, a complex conjugate pair of solutions is obtained, when the expression
under the root is negative. In these cases, only the real part, i.e., the term preceding the
root, is taken. Having 2, nf and s2 are obtained from
1, P~a mP~
s2 = a

16 4
2
. (D-28)
mP~ s 2

nf = a

2
Results are found in the next section, designated as chi-square CDF, M1-3.

D-2.1.2 Log-normal PDF

Exactly the same two methods as in the Rayleigh case are used. The first one compares
the first two moments yielding

mln = ln E Pa
~
(( { }) E{P~ }). 2 2

( { } (E{P~ }) )
a
(D-29)
~ 2
ln2 = ln E Pa2 a

Again the PDF for these parameters does not fit well in the important lower tail, which
can be seen from the results given in the following section (designated log-normal,
M1, 2).
As above, method two directly calculates the parameters for the log-normal PDF from
~
mln = E{ln( Pa )} = ln ( x ) p P~ ( x )dx
0a

. (D-30)
~
+ m = E{ln ( Pa )} = ln 2 ( x ) p P~a ( x )dx
2
ln
2
ln
2
0

Numerical methods were applied to solve these equations. Results are designated as
log-normal (log-fitting).

D-2.2 Analytical and simulation results

As in the previous section, we compare the simulated cumulative distribution function
to the exact theoretical one and various approximations (see Figure D-2 and Table D-
2). It was tried to select typical and representative parameter pairs of Ricean K-factor
and BWrms. The average power (normalized received power) was taken at 0 dB. The
smallest BWrms products correspond to a case very similar to flat fading, which could
D-2 PDF of Average Power for Ricean Fading Channels 321

be described exactly by the non-central chi-square distribution for nf = 2 degrees of

freedom. The bandwidth is multiplied by 4 and 16 yielding the other parameter sets
investigated. Results are given for K = {2, 4, 10} = {3, 6, 10 dB}.
It is seen from all plots, that the exact CDF appropriately describes the PDF of wide-
band average power. Depending on the BWrms product and the Ricean K-factor,
different approximations are more or less appropriate.
From the plots for rather small bandwidths (Figure D-2a, d, g) a bad fitting of the log-
normal distributions is seen, which appear as straight lines in these plots. However,
increasing the bandwidth, the errors decrease at all K-factors. Generally, the errors also
decrease when increasing K. As for the Rayleigh distribution, better results are
obtained when the parameters for the log-normal are calculated by the second method,
designated log-fitting.
In all cases, the best approximations are obtained by the chi-square PDF, when the first
three moments are used for deriving its parameters. Unfortunately, the quadratic
equation that has to be solved for 2 yields a complex conjugate pair of solutions in
some cases (however, with very small imaginary parts). Simply the real part was taken
in these cases. They are marked in Table D-2 by *.
The comparison of the exact CDFs for different bandwidths and K-factors is depicted
in Figure D-3.

Table D-2: Errors and levels of the different approximations at probability values (CDF level)
of 10-2 and 10-3. All values in [dB]. The * indicates cases, where the real part of the
complex conjugate pair of solutions for 2 was taken (see text).

CDF K rms exact chi-square, chi-square, log-normal log-normal Gaussian

prob. [dB] BW PDF M 1, 2 M 13 (M 1, 2) (log-fitting)
level level error level error level error level error level error
0.081 -13.3 -15.2 -1.9 -14.3 -1.1 -7.5 5.8 -10.4 2.8
3 0.32 -8.1 -10.0 -1.9 -8.5 -0.4 -6.1 2.0 -7.1 1.0
1.3 -4.3 -4.8 -0.5 -4.6 * -0.3 -3.8 0.5 -4.0 0.3 -7.8 -3.5
0.065 -10.2 -10.8 -0.6 -10.4 -0.2 -6.1 4.1 -7.9 2.4
10-2 6 0.26 -6.4 -7.3 -0.9 -6.5 -0.1 -4.9 1.4 -5.5 0.8
1.0 -3.4 -3.7 -0.3 -3.5 * -0.1 -3.0 0.3 -3.2 0.2 -4.8 -1.4
0.045 -5.8 -5.9 -0.1 -5.9 -0.02 -4.3 1.6 -4.9 1.0 -12.5 -6.6
10 0.18 -4.1 -4.3 -0.3 -4.1 -0.03 -3.4 0.6 -3.7 0.4 -6.1 -2.1
0.72 -2.3 -2.4 -0.1 -2.3 * -0.03 -2.1 0.2 -2.1 0.1 -2.7 -0.5
0.081 -18.2 -22.9 -4.7 -21.6 -3.4 -9.7 8.5 -13.4 4.8
3 0.32 -11.1 -15.5 -4.5 -12.2 -1.0 -7.9 3.2 -9.2 1.9
1.3 -5.8 -6.8 -1.1 -6.4 * -0.6 -4.9 0.9 -5.2 0.6
0.065 -15.0 -18.1 -3.1 -16.7 -1.7 -7.9 7.1 -10.2 4.8
10-3 6 0.26 -8.8 -11.2 -2.4 -9.2 -0.3 -6.4 2.4 -7.2 1.6
1.0 -4.6 -5.2 -0.6 -4.8 * -0.2 -4.0 0.6 -4.1 0.5 -9.6 -5.0
0.045 -8.7 -9.0 -0.3 -8.8 -0.07 -5.6 3.1 -6.3 2.4
10 0.18 -5.6 -6.3 -0.6 -5.7 -0.07 -4.5 1.1 -4.8 0.9
0.72 -3.1 -3.3 -0.2 -3.1 * -0.04 -2.8 0.3 -2.8 0.2 -4.2 -1.1
322 Appendix D PDF of Wide-band Average Received Power

CDF of wideband average power for K = 2 and BWrms 0.081 CDF of wideband average power for K = 2 and BWrms 0.32
0.99 0.99
CDF for simulation CDF for simulation
0.95 exact CDF 0.95 exact CDF
chisquare CDF, M1,2 chisquare CDF, M1,2
0.9 0.9
chisquare CDF, M13 chisquare CDF, M13
0.8 lognormal, M1,2 0.8 lognormal, M1,2
lognormal (logfitting) lognormal (logfitting)
Gaussian, M 1,2 Gaussian, M 1,2
Pr(P < abscissa)

Pr(Pa < abscissa)

0.5 0.5
a

0.2 0.2

0.1 0.1
0.05 0.05

0.01 0.01
0.005 0.005

0.001 0.001
20 15 10 5 0 5 15 10 5 0 5
wideband average power, Pa [dB] wideband average power, Pa [dB]

(a) (b)
CDF of wideband average power for K = 2 and BWrms 1.3 CDF of wideband average power for K = 4 and BWrms 0.065
0.99 0.99
CDF for simulation CDF for simulation
0.95 exact CDF 0.95 exact CDF
chisquare CDF, M1,2 chisquare CDF, M1,2
0.9 0.9
chisquare CDF, M13 chisquare CDF, M13
0.8 lognormal, M1,2 0.8 lognormal, M1,2
lognormal (logfitting) lognormal (logfitting)
Gaussian, M 1,2 Gaussian, M 1,2
Pr(P < abscissa)
Pr(Pa < abscissa)

0.5 0.5
a

0.2 0.2

0.1 0.1
0.05 0.05

0.01 0.01
0.005 0.005

0.001 0.001
6 4 2 0 2 4 15 10 5 0 5
wideband average power, Pa [dB] wideband average power, Pa [dB]

(c) (d)
CDF of wideband average power for K = 4 and BWrms 0.26 CDF of wideband average power for K = 4 and BWrms 1
0.99 0.99
CDF for simulation CDF for simulation
0.95 exact CDF 0.95 exact CDF
chisquare CDF, M1,2 chisquare CDF, M1,2
0.9 0.9
chisquare CDF, M13 chisquare CDF, M13
0.8 lognormal, M1,2 0.8 lognormal, M1,2
lognormal (logfitting) lognormal (logfitting)
Gaussian, M 1,2 Gaussian, M 1,2
Pr(Pa < abscissa)

Pr(Pa < abscissa)

0.5 0.5

0.2 0.2

0.1 0.1
0.05 0.05

0.01 0.01
0.005 0.005

0.001 0.001
10 8 6 4 2 0 2 4 5 4 3 2 1 0 1 2 3
wideband average power, Pa [dB] wideband average power, Pa [dB]

(e) (f)
Figure D-2: CDFs of the wide-band average power for the Ricean fading case and different
BWrms products. Comparison of approximations and assessment of the exact
equation.
D-2 PDF of Average Power for Ricean Fading Channels 323

CDF of wideband average power for K = 10 and BWrms 0.045 CDF of wideband average power for K = 10 and BWrms 0.18
0.99 0.99
CDF for simulation CDF for simulation
0.95 exact CDF 0.95 exact CDF
chisquare CDF, M1,2 chisquare CDF, M1,2
0.9 0.9
chisquare CDF, M13 chisquare CDF, M13
0.8 lognormal, M1,2 0.8 lognormal, M1,2
lognormal (logfitting) lognormal (logfitting)
Gaussian, M 1,2 Gaussian, M 1,2
Pr(Pa < abscissa)

Pr(Pa < abscissa)

0.5 0.5

0.2 0.2

0.1 0.1
0.05 0.05

0.01 0.01
0.005 0.005

0.001 0.001
8 6 4 2 0 2 4 6 5 4 3 2 1 0 1 2 3
wideband average power, Pa [dB] wideband average power, Pa [dB]

(g) (h)
CDF of wideband average power for K = 10 and BWrms 0.72
0.99
CDF for simulation
0.95 exact CDF
chisquare CDF, M1,2
0.9
chisquare CDF, M13
0.8 lognormal, M1,2
lognormal (logfitting)
Gaussian, M 1,2
Pr(P < abscissa)

0.5
a

0.2

0.1
0.05

0.01
0.005

0.001
3 2 1 0 1 2
wideband average power, Pa [dB]

(i)
Figure D-2: (continued) CDFs of the wide-band average power for the Ricean fading case and
different BWrms products. Comparison of approximations and assessment of the
exact equation.

CDF of wideband average power for K = {0, 4} and BWrms = {0.1, 0.4, 2} CDF of wideband average power for K = {2, 10} and BWrms = {0.1, 0.4, 2}
0.99 0.99

K = 0; BWrms = 0.1 K = 2; BWrms = 0.1

0.95 0.95
K = 0; BWrms = 0.4 K = 2; BWrms = 0.4
0.9 0.9
K = 0; BWrms = 2 K = 2; BWrms = 2
0.8 K = 4; BWrms = 0.1 0.8 K = 10; BWrms = 0.1
K = 4; BWrms = 0.4 K = 10; BWrms = 0.4
Pr(Pa < abscissa)

Pr(Pa < abscissa)

0.5 K = 4; BWrms = 2 0.5 K = 10; BWrms = 2

0.2 0.2

0.1 0.1
0.05 0.05

0.01 0.01
0.005 0.005

0.001 0.001
20 15 10 5 0 5 20 15 10 5 0 5
wideband average power, P [dB] wideband average power, P [dB]
a a

(a) (b)
Figure D-3: Wide-band average power for various parameters K and BWrms.
324 Appendix D PDF of Wide-band Average Received Power

D-3 References

[1] C. W. Therrien, Discrete Random Signals and Signal Processing. Englewood

Cliffs: Prentice Hall, 1992.

[2] J. G. Proakis, Digital Communications, 3rd Edition. New York: McGraw Hill,
1995.
List of Acronyms

3G; 4G 3rd; 4th generation mobile system

ABF average bandwidth of fades
ADC analog-to-digital converter
ARQ automatic repeat request
ASIC application specific integrated circuit
ATM asynchronous transfer mode
AWGN additive white Gaussian noise
BER bit error rate
BPSK binary phase shift keying
BS base station
CDF cumulative distribution function
CDMA code division multiple access
CEWPC center for wireless personal communications
CIR carrier-to-interference ratio
COFDM coded OFDM (orthogonal frequency division multiplexing)
COST European co-operation in science and technical research
CW continuous wave
DAC digital-to-analog converter
DC direct current
DFT discrete Fourier transform
DPS delay power spectrum
DSP digital signal processor (processing)
FD frequency-domain
FDMA frequency division multiple access

325
326 OFDM Air-Interface Design for Multimedia Communications

FEC forward error correction coding

FER frame error rate
FIR finite impulse response
FFT, FT (fast) Fourier transform
FR, -fr full rate
GI guard interval
GIS geographical information system
GSM global system for mobile communications
HD Hamming distance
HIPERLAN high performance wireless LAN (European WLAN standard)
I/Q in-phase/quadrature (channels of a complex baseband signal)
ICI inter-carrier-interference
IDFT inverse discrete Fourier transform
IF intermediate frequency
IFFT inverse fast Fourier transform
IMT-2000 international mobile telecommunications by 2000
IR impulse response
IRCTR international research center for telecommunications-transmission and
radar
ISI inter-symbol-interference
KT Korea Telecom
LAN local area network
LCR; LCRf level crossing rate; level crossing rate in the frequency domain
LO local oscillator
LOS line-of-sight
LSB least significant bit
MAC multiple access control
MBS mobile broadband systems
MMAC multimedia mobile access communications (Japanese WLAN standard)
MMSE minimum mean-square-error
MSE mean-square-error
MT mobile terminal
NRP normalized received power
OFDM orthogonal frequency division multiplexing
PAPR peak-to-average power ratio
PDF probability density function
List of Acronyms 327

PDM pulse duration modulation

PDP power delay profile
PN pseudo noise
PSD power spectral density
PSK phase shift keying
QAM quadrature amplitude modulation
QoS quality of service
QR quarter rate
QPSK quadrature phase shift keying
QWSSUS quasi wide-sense stationary uncorrelated scattering
RACE research into advanced communications systems in Europe
RAM random access memory
RDS RMS delay spread
RF radio frequency
RMS root mean square
RS reset/set (flip-flop)
RV random variable
RX receive(r)
SC sub-carrier
SNR signal-to-noise ratio
TD time-domain
TDD time division duplex
TDMA time division multiple access
TF transfer function
TRX transceiver
TU; TUD technical university; TU-Delft
TX transmit(ter)
UMTS universal mobile telecommunications system
US uncorrelated scattering
VCO voltage controlled oscillator
VCXO voltage controlled crystal oscillator
WLAN wireless local area network
WSS wide-sense stationary
WSSUS wide-sense stationary uncorrelated scattering
XO crystal oscillator
Publications by the Author

Journal Papers
[1] K. Witrisal, Y.-H. Kim, and R. Prasad, A new method to measure parameters of
frequency-selective radio channels using power measurements, IEEE Trans-
actions on Communications, vol. 49, no. 10, Oct. 2001, pp. 17881800.
[2] K. Witrisal, On estimating the RMS delay spread from the frequency-domain
level crossing rate, IEEE Communications Letters, vol. 5, no. 7, July 2001, pp.
287289.
[3] K. Witrisal, Y.-H. Kim, and R. Prasad, RMS delay spread estimation technique
using non-coherent channel measurements, IEE Electronics Letters, vol. 34, no.
20, Oct. 1998, pp. 19181919.
[4] K. Witrisal, Y.-H. Kim, L. P. Ligthart, and R. Prasad, Computationally efficient
antenna diversity techniques for OFDM using cyclic delays, to be published in
Korea Telecom Journal.

Conference Papers
[5] K. Witrisal, Y.-H. Kim, R. Prasad, L. P. Ligthart, Pre-equalization for the Up-
th
link of TDD OFDM Systems, in Proc. PIMRC 2001 (12 International Sympo-
sium on Personal Indoor Mobile Radio Communications), San Diego, CA, Oct.
2001, pp. E-93E-98.
[6] K. Witrisal, Y.-H. Kim, R. Prasad, L. P. Ligthart, Antenna diversity for OFDM
th
using cyclic delays, in Proc. SCVT-2001 (8 Symp. on Commun. and Vehic.
Technol. in the Benelux), Delft (Netherlands), Oct. 2001, pp. 1317.

329
330 OFDM Air-Interface Design for Multimedia Communications

[7] K. Witrisal, Impact of DC-offsets and Carrier Feed-through on Correlation-

th
based Frequency Synchronization for OFDM, in Proc. 6 international OFDM-
Workshop (InOWo), Hamburg (Germany), Sept. 2001, pp. 15-115-5.
[8] K. Witrisal, G. Landman, and A. Bohdanowicz, Practical Application of a
Novel Method for Estimating the RMS Delay Spread from Power Measure-
ments, in Proc. EPMCC 2001 (4th European Mobile Communications Confer-
ence), Vienna (Austria), Feb. 2001.
[9] K. Witrisal, K. Bke, Y.-H. Kim, R. Prasad, and L.P. Ligthart, Air-interface
Emulation for Wireless Broadband Communications Applied to OFDM, in
th
Proc. PIMRC 2000 (11 International Symposium on Personal Indoor Mobile
Radio Communications), London, Sept. 2000, pp. 12511255.
[10] K. Witrisal and A. Bohdanowicz, Influence of Noise on a Novel RMS Delay
th
Spread Estimation Method, in Proc. PIMRC 2000 (11 International Sympo-
sium on Personal Indoor Mobile Radio Communications), London, Sept. 2000,
pp. 560566.
[11] K. Witrisal, Y.-H. Kim, R. Prasad, and L.P. Ligthart, Experimental Study and
th
Comparison of OFDM Transmission Techniques, in Proc. 5 international
OFDM-Workshop, Hamburg (Germany), Sept. 2000, pp. 5-15-5.
[12] K. Witrisal, R. Prasad, L. P. Ligthart, OFDM Technology for Realizing Broad-
band Wireless Multimedia Communications, in Proc. Korea Telecom Interna-
tional Symposium (KTIS 99), Seoul (Korea), Oct. 1999, pp. 307312.
[13] K. Witrisal, Y.-H. Kim, and R. Prasad, A Novel Approach for Performance
Evaluation of OFDM with Error Correction Coding and Interleaving, in Proc.
VTC99-fall (IEEE Vehicular Technology Conference), Amsterdam, Sept. 1999,
pp. 294299.
[14] J. M. M. de Nijs, K. Witrisal and R. Prasad, Characterization and Simulation of
the 18 GHz Radio Channel, presented at the COST 259 7th Meeting of the Man-
agement Committee, COST-259 TD(99)020, Thessaloniki (GR), Jan. 1999.
[15] J. Purwaha, A. Mank, D. Matic, K. Witrisal and R. Prasad, Wide-band Channel
th
Measurements at 60 GHz in Indoor Environments, in Proc. IEEE Benelux 6
Symposium on Vehicular Technology and Communications, Brussels, Belgium,
Oct. 1998.
[16] R. El Hattachi, J. M. M. de Nijs, K. Witrisal and R. Prasad, Characterization and
th
Simulation of the 18 GHz Radio Channel, in Proc. IEEE Benelux 6 Symposium
on Vehicular Technology and Communications, Brussels, Belgium, Oct. 1998.
Publications by the Author 331

[17] K. Witrisal, Y.-H. Kim, and R. Prasad, Frequency-Domain Simulation and

Analysis of the Frequency-Selective Ricean Fading Radio Channel, in Proc.
PIMRC98 (9th International Symposium on Personal Indoor Mobile Radio
Communications), Boston, MA, Sept. 1998, pp. 11311135.
[18] K. Witrisal, Y.-H. Kim, and R. Prasad, Frequency-Domain Simulation and
Analysis of the Frequency-Selective Radio Channel for the Performance Analy-
sis of OFDM, in Proc. 3. OFDM Fachgesprch, Braunschweig (Germany),
Sept. 1998.
[19] K. Witrisal, Y.-H. Kim, H.-M. Pyo, and R. Prasad, Wireless Broadband Com-
munications in mm-Waves: Frequency-Domain Simulation of the Frequency Se-
rd
lective Radio Channel, in Proc. 3 ACTS Mobile Communications Summit,
Rhodes (Greece), 1998, pp. 885887.

Reports
[20] K. Witrisal, Emulator Implementation Software Update, Project KWATT,
Supplementary Report, IRCTR S-009-01, Delft Univ. of Tech., Feb. 2001.
[21] K. Witrisal, Algorithm Development and System Proposal, Project KWATT, 6th
(Final) Progress Report, IRCTR S-028-00 and IRCTR S-029-00, Delft Univ. of
Tech., Nov. 2000.
[22] K. Witrisal, Emulation System Implementation, Project KWATT, 5th Progress
Report, IRCTR S-015-00 and IRCTR S-016-00, Delft Univ. of Tech., May 2000.
[23] K. Witrisal, Demonstration and Emulation System Studies, Project KWATT, 4th
Progress Report, IRCTR S-035-99 and IRCTR S-036-99, Delft Univ. of Tech.,
Nov. 1999.
[24] K. Witrisal, OFDM Performance Evaluation, Project KWATT, 3rd Progress Re-
port, IRCTR S-015-99 and IRCTR S-016-99, Delft Univ. of Tech., May 1999.
[25] K. Witrisal, OFDM System Proposal, Project KWATT, 2nd Progress Report,
IRCTR S-033-98 and IRCTR S-034-98, Delft Univ. of Tech., Nov. 1998.
[26] K. Witrisal, Channel Modeling at 60 GHz, Project KWATT, 1st Progress Re-
port, IRCTR S-010-98 and IRCTR S-011-98, Delft Univ. of Tech., May 1998.

Relation to this Thesis

This Thesis is based, in part, on previously published work as listed above. The table
below indicates the minor and major relations between these publications and the
chapters of this thesis.
332 OFDM Air-Interface Design for Multimedia Communications

Relations between the publications and the chapters of this thesis. z: major relation; {: minor
relation

Publication Ch. 2 Ch. 3 Ch. 4 Ch. 5 Ch. 6 Ch. 7 Ch. 8

1 z z
2 { z
3 { z
4 z
5 { { z
6 z
7 z
8 z
9 z {
10 z
11 { z
12 { { { z {
13 z
14 { z
15 { z
16 { z
17 z z
18 { { z
19 z {
20 z z z
21 z z z
22 z {
23 z { z
24 { z
25 z z
26 z z
Samenvatting

Het doel van dit proefschrift: OFDM Air-interface Design for Multimedia Communi-
cations, vormt het onderzoek naar belangrijke aspecten voor de ontwikkeling van
breedband radioverbindingen. Orthogonal frequency division multiplexing (OFDM)
wordt beschouwd als een mogelijke oplossing voor de transmissie van data met ex-
treem hoge snelheid over tijd-dispersieve radiokanalen. In OFDM wordt de datastroom
opgedeeld in meerdere deelstromen welke met sterk verlaagde snelheid over parallelle
sub-kanalen verzonden worden.

Het eerste deel van dit proefschrift behandelt de modellering van het tijd-dispersieve
en frequentie-selectieve radio kanaal, op basis van Gaussische stochastische processen
van de tweede orde. Een nieuwe methode wordt ontwikkeld voor de karakterisering
van het radio kanaal. Hierbij wordt de level crossing rate van de frequentie-selec-
tieve overdrachtsfunctie van het radiokanaal gebruikt om de RMS delay spread te
schatten. Deze techniek maakt het mogelijk het radio kanaal te karakteriseren op basis
van eenvoudige niet-coherente metingen waarin het vermogen als functie van de fre-
quentie bepaald wordt.

De Air-interface en het multiple access schema van een op OFDM gebaseerd

communicatiesysteem worden gentroduceerd en onderzocht in deel twee van dit werk.
De maximale data snelheid van 155 Mbit/s wordt bereikt in een optimaal radiokanaal
bij een beperkte snelheid van de terminal (loopsnelheid). Draadloze computernetwer-
ken zijn een typische toepassing voor dit systeem.

Algoritmen voor synchronisatie en kanaalschatting worden ontwikkeld en geva-

lueerd. Hierbij wordt een bekend trainingssymbool gebruikt, welke periodiek in het
begin van een vaste framestructuur verzonden wordt. De conclusie is dat robuuste en
efficinte synchronisatie en kanaalschatting dit zijn kritische procedures in een
OFDM ontvanger mogelijk zijn, door gebruik te maken van een trainingssymbool,

333
334 OFDM Air-Interface Design for Multimedia Communications

waarbij slechts een kleine overhead gentroduceerd wordt.

Een diepgaande analyse is gemaakt van een aantal onderwerpen op het gebied van
synchronisatie: de schatting van de tijd-offset in dispersieve kanalen, en de invloed
van DC-offset en carrier feed-through op een veel gebruikte frequentie-synchroni-
satie techniek. Voor het laatste probleem is een oplossing ontwikkeld.

Voor de up-link kan pre-equalization gebruikt worden in een time-division du-

plexing (TDD) frameformaat. Deze methode compenseert voor het frequentie selec-
tieve gedrag van het radiokanaal, waarbij de detectie van data vereenvoudigd wordt.
De synchronisatie wordt hierbij overwegend in de mobiele stations uitgevoerd. Dit be-
tekent een distributie van de grote complexiteit die daarvoor nodig is. Er worden con-
cepten gentroduceerd, om een gelijkblijvend vermogen van het uitgezonden signaal te
bereiken, en om de resterende synchronisatiestappen op te lossen.

De belangrijkste signaal verwerkingsalgoritmen van de OFDM zender-ontvangers zijn

gemplementeerd en gevalueerd op een experimenteel systeem bestaande uit DSP
borden. Het systeem werkt in real-time maar met een sterk gereduceerde datasnelheid.

Foutcorrigerende codering vormt een belangrijk onderdeel van OFDM systemen.

Daarbij worden de datasymbolen goed over de grote bandbreedte van het OFDM
signaal verdeelt om de frequentie-diversiteit te exploiteren. De foutenkans van een
gecodeerd OFDM systeem is gevalueerd. Hieruit volgt dat een grotere bandbreedte en
een langere kanaal impulsresponsie (in een beperkte mate) tot betere resultaten leiden.
Een nieuwe methode voor antenne-diversiteit wordt gentroduceerd. Deze techniek kan
de foutenkans met weinig toegevoegde complexiteit aanzienlijk verlagen bij een kleine
bandbreedte van het radiokanaal of een korte impulsresponsie.

In het algemeen kan worden geconcludeerd dat OFDM een efficinte en robuuste
techniek is om data met hoge snelheid te verzenden. Enkele kritische implementatie-
aspecten moeten daarbij opgelost worden, zoals bijvoorbeeld de lineariteit van verster-
kers, en de faseruis van oscillatoren.

Klaus Witrisal
Acknowledgements

I owe my gratitude to my supervisors, Prof. Leo Ligthart, Prof. Ramjee Prasad, and Dr.
Gerard Janssen, who gave me all the support needed to finalize this thesis. Particularly,
I wish to acknowledge the confidence of Gerard Janssen and Prof. Prasad in my
abilities, when they initially encouraged me to come to Delft as a Ph.D. candidate,
after working with me on my graduation project. Moreover, I am most grateful to Prof.
Ligthart for taking over the task of the promotor after Prof. Prasad left the Univer-
sity, ensuring the continuity of my work. I wish to thank all my supervisors for their
permanent encouragement and for their comments and discussions on my thesis and on
my scientific papers.
Large parts of this work have been performed under a cooperative research project
between Network Access Lab of Korea Telecom, Seoul, and the IRCTR (International
Research Centre for Telecommunications-transmission and Radar) of TU Delft. This
cooperation was a unique experience for me, as I had the opportunity of working with
the friendly Korean people and discovering their culture. I am indebted to Korea Tele-
com and Delft University for providing the opportunity to carry out this research. I
want to thank all the colleagues and students who have contributed to this work,
particularly, Dr. Yong-Ho Kim of Korea Telecom, and Dr. Homayoun Nikookar,
Frans BrePHU 9ODGLPLU -RYL DQG +DQ 5HLMPHUV RI 'HOIW 8QLYHUsity. The im-
plementation of the experimental platform was made possible through the support by
the technical staff of the IRCTR.
I am thankful to the students who have made substantial contributions while working
on their graduation projects or on internships. I also wish to recognize the support of
the secretaries in organizational matters.
To my colleagues and friends at TU Delft I am grateful for numerous inspiring discus-
sions about the research work and about other topics and for the pleasant environ-
ment they created. I would like to mention personally: Nicolas Petrochilos, Antonio
7ULQGDGH 'XDQ 0DWL $GULDQ %RKGDQRZLF] )ULHGULFK 5RWK 0DUWHQ .ORN 7RQ

335
336 OFDM Air-Interface Design for Multimedia Communications

Coenen, Jos Weber, and Mohinder Jankiraman, with whom I shared many thoughts,
privately and professionally. I have experienced the Netherlands as a place where
foreigners are appreciated and well integrated in the society.
The foundation for this Ph.D. study was laid during my undergraduate studies. I am
most grateful to my parents for providing the opportunity to conduct the study of my
choice and for their personal and financial support.
Last but not least, I wish to thank my beloved wife Silke for moving with me to the
Netherlands four years ago. Thank you for sharing all the pleasant and difficult mo-
ments during this time. Without your mental support I had never been able to finalize
this Ph.D. project.
Curriculum Vitae

Klaus Witrisal was born in Graz, Austria on September 24, 1972. He obtained his
Dipl.-Ing. (Diplomingenieur; equivalent to M.Sc.) degree in Electrical Engineering
(Electronics and Telecommunications) from Graz University of Technology in 1997.
During his undergraduate studies, he had the opportunity of spending six months at the
Telecommunications and Traffic Control Systems Group (TVS) of Delft University of
Technology. This visit was enabled by the European Unions ERASMUS program for
students exchange.
In Nov. 1997, Klaus joined Delft University of Technology as a research engineer in
the CEWPC (Centre for Wireless Personal Communications), a sub-division of the
IRCTR (International Research Centre for Telecommunications-transmission and Ra-
dar). From Nov. 1997 till Dec. 2000 he was the main researcher on a collaborative re-
search project between the IRCTR and Korea Telecom in the field of wireless ATM
transmission in the 60 GHz frequency band. In 2001, he has been working on his
Ph.D. thesis, which mainly covers the topics treated in the Korea Telecom project.
Klaus was awarded with a best student paper award at the International Symposium on
Personal Indoor and Mobile Radio Communications (PIMRC) in September 1998.
In April 2001, he represented Delft University of Technology at the KIvI Telecom
Award, a prize given by the Royal Dutch Society of Engineers to a Ph.D. candidate
working in the field of telecommunications.
Since Jan. 2002, Klaus is a project leader at AVL-Ditest in Graz, Austria, working on
advanced diagnostics and test/measurement instruments for the automotive industry.

337