Академический Документы
Профессиональный Документы
Культура Документы
in
by
Steve C. Thompson
Committee in charge:
2005
Copyright
Co-Chair
Chair
2005
iii
“Before PhD,
I chopped wood and carried water;
After PhD,
I chopped wood and carried water.”
—[Slightly modified] Zen saying
iv
TABLE OF CONTENTS
Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 An Introduction to OFDM . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.1.1 ISI-Free Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.1.2 A Multicarrier Modulation . . . . . . . . . . . . . . . . . . . . . . 5
1.1.3 Discrete-Time Signal Processing . . . . . . . . . . . . . . . . . . . 8
1.2 Problems with OFDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3 Constant Envelope Waveforms . . . . . . . . . . . . . . . . . . . . . . . . 12
1.4 Constant Envelope OFDM . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.5 Thesis Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2 OFDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.1 More OFDM Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.1.1 The Cyclic Prefix . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.1.2 Discrete-Time Model . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.1.3 Block Modulation with FDE . . . . . . . . . . . . . . . . . . . . . 20
2.1.4 System Diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.2 PAPR Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3 Power Amplifier Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.4 Effects of Nonlinear Power Amplification . . . . . . . . . . . . . . . . . . . 30
v
2.4.1 Spectral Leakage . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.4.2 Performance Degradation . . . . . . . . . . . . . . . . . . . . . . . 32
2.4.3 System Range and PA Efficiency . . . . . . . . . . . . . . . . . . . 35
2.5 PAPR Mitigation Techniques . . . . . . . . . . . . . . . . . . . . . . . . . 37
vi
7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
A Generating Real-Valued OFDM Signals with the Discrete Fourier Transform . . 124
A.1 Signal Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
A.2 Spectral Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
Symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
vii
LIST OF FIGURES
viii
2.17 PAPR CCDF of clipped OFDM signal for various γ clip (dB). [N = 64] . . 40
2.18 PAPR of clipped signal as a function of the clipping ratio. (N = 64) . . . 40
2.19 A comparison of the total degradation curves of clipped and unclipped
M -PSK/OFDM systems. (N = 64) . . . . . . . . . . . . . . . . . . . . . . 41
ix
4.13 All unique ρm,n (K) for M = 2, N = 4 DCT modulation. . . . . . . . . . . 78
4.14 Phase demodulator receiver versus optimum. (N = 64) . . . . . . . . . . . 79
4.15 Noise samples PDF versus Gaussian PDF. (E b /N0 = 30 dB) . . . . . . . . 80
4.16 Performance of M -PAM CE-OFDM. (N = 64, †=leftmost curve, ‡=rightmost
curve) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.17 Spectral efficiency versus performance. . . . . . . . . . . . . . . . . . . . . 82
4.18 A comparison of CE-OFDM and conventional OFDM. (M = 2, N = 64) . 85
x
B.3 Running average of papers read per day. . . . . . . . . . . . . . . . . . . . 132
B.4 Year histogram. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
B.5 Projected year histogram? . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
xi
LIST OF TABLES
6.3 Data symbol contribution per tone for m n (t), n =1, 2, and 3. . . . . . . . 118
xii
ACKNOWLEDGEMENTS
I want to first thank my advisors, Professors Zeidler and Proakis, for giving me the
chance to do this work, for the encouragement, and for the guidance. I want to thank
Professor Milstein for the many helpful technical conversations and for his many sug-
gestions. Thanks to Professors Bitmead and Hodgkiss for taking the time to participate
as committee members. Also, thanks to Professor Proakis for carefully proofreading the
draft manuscripts of this thesis.
Thanks to UCSD’s Center for Wireless Communications for providing a good en-
vironment for conducting research; thanks to its industrial partners for the financial
support.
Thanks to my wife, Shannon, for the emotional and caloric support. Thanks to
Chaney the cat for waking me up in the morning. Thanks to my friends for fun support.
Thanks to my fellow graduate students in Professor Zeidler’s research group for the
camaraderie. Special thanks to Ahsen Ahmed for helpful collaboration over the past
couple years. Thanks to my family. Also, thanks to Karol Previte for her support early
in my graduate student existence.
Finally, I would like to thank the countless developers, documentation writers, bug
reporters, and users of the free software I’ve benefited from during the course of my PhD.
The text in this thesis, in part, was originally published in the following papers, of
which I was the primary researcher and author: S. C. Thompson, J. G. Proakis, and
J. R. Zeidler, “Constant Envelope Binary OFDM Phase Modulation,” in Proc. IEEE
Milcom, vol. 1, Boston, Oct. 2003, pp. 621–626; S. C. Thompson, A. U. Ahmed, J.
G. Proakis, and J. R. Zeidler, “Constant Envelope OFDM Phase Modulation: Spectral
Containment, Signal Space Properties and Performance,” in Proc. IEEE Milcom, vol. 2,
Monterey, Oct. 2004, pp. 1129–1135; S. C. Thompson, J. G. Proakis, and J. R. Zeidler,
“Noncoherent Reception of Constant Envelope OFDM in Flat Fading Channels,” in Proc.
IEEE PIMRC, Berlin, Sept. 2005; and S. C. Thompson, J. G. Proakis, and J. R. Zeidler,
“The Effectiveness of Signal Clipping for PAPR Reduction and Total Degradation in
OFDM Systems,” in Proc. IEEE Globecom, St. Louis, Dec. 2005.
xiii
VITA
PUBLICATIONS
xiv
S. C. Thompson, J. G. Proakis, and J. R. Zeidler, “The Effectiveness of Signal Clipping
for PAPR Reduction and Total Degradation in OFDM Systems,” in preparation.
xv
ABSTRACT OF THE DISSERTATION
by
Steve C. Thompson
xvi
this thesis, the fundamental aspects of the CE-OFDM modulation are studied, including
the signal spectrum, the signal space, optimum performance, and the performance of
a practical phase demodulator receiver. Performance is evaluated over a wide range of
multipath fading channel models. It is shown that CE-OFDM outperforms conventional
OFDM when taking into account the effects of the power amplifier.
This work was done at UCSD’s Center for Wireless Communication, under the “Mo-
bile OFDM Communications” project (CoRe research grant 00-10071).
xvii
Chapter 1
Introduction
Humans have always found ways to communicate, over space and over time. From
the messenger pigeon to the Pony Express, from the message in a bottle to cave drawings,
smoke signals and beacons, people have used inventive techniques, techniques derived
from their natural environment, to share information. A particularly good natural re-
source for communication is electricity for its speed and ability to be controlled with
devices like capacitors, microprocessors, electronic memory storage and batteries. Com-
munication was profoundly enhanced with Morse’s telegraph (1837), Bell’s telephone
(1876), Edison’s phonograph (1887), and Marconi’s radio (1896). From these early inven-
tions, communications technology has advanced with global telephone networks, satellite
communications, and magnetic storage systems; and with the rise of the internet and
digital computers, digital communications—the transfer of bits (1’s and 0’s) from one
point to another—has become important.
1
2
point B
point A
Propagation paths
An example profile of the channel in Figure 1.1 is shown in Figure 1.2(a). Each path
has its own associated delay and power. The first path arrives at the receiver 0.5 µs after
the signal is transmitted; the last path arrives with a 14 µs delay. The Fourier transform
of the profile yields the frequency-domain representation shown in Figure 1.2(b). The
channel is viewed over a 2 MHz range centered at the center frequency f c . Notice that
the channel power fluctuates by 30 dB (a factor of 1000) over the frequency range. The
dispersion in the time domain leads to frequency-selectivity in the frequency domain.
1 5
0
Channel power (dB)
-5
Path power
-10
0.1
-15
-20
-25
0.01
-30−1 −0.5
0 2 4 6 8 10 12 14 0 0.5 1
Time (µs) Frequency, f − fc (MHz)
where t is the time variable, {Ii } are the data symbols, Ts is the symbol period, and g(t)
is a transmit pulse shape. For time-dispersive channels, such as the 4-path example in
3
Figure 1.2, interference is caused from symbol to symbol. This intersymbol interference
(ISI) is illustrated in Figure 1.3. For simplicity, g(t) is rectangular. The channel is
represented by its time-variant impulse response h(τ, t), where τ is a propagation delay
variable. The received signal is expressed mathematically as [387, p. 97]
The severity of the ISI depends on the symbol period relative to the channel’s max-
imum propagation delay, τmax . Consider transmitting the signal in (1.1) over the 2
MHz channel in Figure 1.2. The signal bandwidth is roughly proportional to the sym-
bol rate 1/Ts Hz. Therefore making s(t) a 2 MHz signal, T s = (2 × 106 )−1 = 0.5 µs.
Since the maximum propagation delay of the channel is τ max = 14 µs, the ISI spans
τmax /Ts = (14 µs) / (0.5 µs) = 28 symbols. (For comparison, the ISI in Figure 1.3 spans
less than one symbol.) Such severe ISI must be corrected at the receiver in order to
provide reliable communication.
For scenarios like the example above with an ISI spanning 28 symbols, conventional
equalization becomes difficult. Training times become long and convergence of the chan-
nel estimator is problematic, especially for time-varying channels. In the example, 2×10 6
symbols/s are transmitted. Using a QPSK (quadrature phase-shift keying) signal con-
stellation, which maps kb = 2 bits per symbol, the bit rate is 4 Mb/s. Such a bit rate is
desired in current wireless systems, and in many cases demand for many tens of Mb/s
is common.
To meet the demanding data rate requirements, alternative techniques have been
considered. One approach, orthogonal frequency division multiplexing, has become ex-
ceedingly popular. OFDM has been implemented in wireline applications such as digital
subscriber lines (DSL) [95], in wireless broadcast applications such as digital audio and
video broadcasting (DAB and DVB) and in-band on-channel (IBOC) broadcasting [392].
It has been used in wireless local area networks (LANs) under the IEEE 802.11 and the
ETSI HYPERLAN/2 standards [552]. OFDM is being developed for ultra-wideband
(UWB) systems; cellular systems; wireless metropolitan area networks (MANs), under
the IEEE 802.16 (WiMax) standard; and for other wireline systems such as power line
communication (PLC) [119, 160, 264, 604].
OFDM’s main appeal is that it supports high data rate links without requiring
conventional equalization techniques. Instead of transmitting symbols serially, OFDM
5
sends N symbols as a block. The OFDM block period, T B , is thus N times longer than
the symbol period. Continuing the example above, and choosing N = 300, the block
period is TB = N Ts = 300 × 0.5 µs = 150 µs, which is more than 10 times the duration
of the channel’s impulse response. ISI is avoided by inserting a guard interval between
successive blocks during which a cyclic prefix is transmitted. The interval duration, T g ,
is designed such that Tg ≥ τmax so that the channel is absorbed in the guard interval
and the OFDM block is uncorrupted. This is illustrated in the figure below. Selecting a
guard interval Tg = 15 µs for the channel in Figure 1.2 results in a transmission efficiency
ηt = TB /(TB + Tg ) = 150/165 ≈ 0.91. Therefore, with a small reduction in efficiency, ISI
is eliminated.
s(t)
Tg TB
CP OFDM block
t
|h(τ, t)|
r(t)
ISI-free block
t
frequency of the kth subcarrier is f k = k/TB and the subcarrier spacing, 1/TB Hz, makes
the subcarriers orthogonal over the block interval, expressed mathematically as
Z TB Z TB
1 ∗ 1
ej2πfk1 t ej2πfk2 t dt = ej2π(fk2 −fk1 )t dt
TB 0 TB 0
1, k1 = k2 ,
(1.5)
=
0, k1 6= k2 ,
where (·)∗ represents the complex conjugate operation. The subcarrier orthogonality can
also be viewed in the frequency domain. Consider the 0th OFDM block:
N
X −1
s(t) = I0,k ej2πfk t , 0 ≤ t < TB . (1.6)
k=0
Figure 1.5 plots |S(f )/TB | for N = 16 subcarriers and data symbols with normalized
amplitudes. The individual subcarrier spectra are also plotted. Notice that at the kth
subcarrier frequency, k/TB , the kth subcarrier has a peak and all the other subcarriers
have zero-crossings. Therefore, the subcarriers, while tightly packed (which improves
spectral efficiency), are non-interfering (i.e. orthogonal).
1.2 Subcarrier
Overall
1
Spectrum magnitude, |S(f )/TB |
0.8
0.6
0.4
0.2
0
-2 0 2 4 6 8 10 12 14 16 18
Normalized frequency, f TB
Figure 1.5: Subcarrier and overall spectrum. (N = 16; |I 0,k | = 1, for all k)
For single carrier each symbol occupies the entire signal bandwidth, while for multicarrier
the bandwidth is split into many frequency bands (also referred to as frequency bins).
Notice that the multicarrier signal transmits the N data symbols in parallel over multiple
carriers each centered at (fc + k/TB ) Hz, k = 0, 1, . . . , N − 1.
By properly designing the subcarrier spacing, each frequency bin is made frequency-
nonselective. The wideband frequency-selective channel is converted into N contigu-
ous narrowband frequency-nonselective bins. Figure 1.6 shows 18 bins in the range
[−0.9, −0.78] MHz for the N = 300 OFDM system over the channel in Figure 1.2(b).
Notice that the channel gain per bin varies over a 15 dB range. The OFDM modulation
can be optimized for the channel by sending more bits in frequency bins with high gain
and fewer bits in frequency bins with low gain. This technique, known as bit loading,
requires a fairly stable channel, one that can be accurately measured. For this reason,
bit loading is more common in wireline systems and stationary wireless systems than in
wireless systems with high mobility.
-10
-15
-20
Frequency bins
-25
-30
−0.9 −0.85 −0.8
Frequency, f − fc (MHz)
nel is frequency non-selective and thus ISI is avoided. Therefore, Figure 1.6 illustrates a
frequency-domain interpretation of how OFDM avoids intersymbol interference.
Thus far, two of OFDM’s primary advantages have been discussed: the elimination
of ISI and the ability to optimize the modulation with bit loading. The third appeal of
OFDM is that the modulation and demodulation is done in the discrete-time domain with
the inverse fast Fourier transform (IFFT) and fast Fourier transform (FFT), respectively.
This is seen by sampling s(t) in (1.6) at N equally spaced time instances:
N
X −1
y[i] ≡ s(t)|t=iTB /N = I0,k ej2πki/N , i = 0, 1, . . . N − 1, (1.12)
k=0
which is the inverse discrete Fourier transform (IDFT) of the symbol vector I 0 =
[I0,0 , I0,1 , . . . , I0,N −1 ]. Therefore, s(t) is generated at the transmitter with an IDFT fol-
N −1
lowed by a digital-to-analog (D/A) converter. The frequency-domain symbols {I 0,k }k=0
can be expressed as
N −1
1 X
I0,k = y[i]e−j2πkn/N , k = 0, 1, . . . N − 1, (1.13)
N
i=0
which is the discrete Fourier transform (DFT) performed on the time-domain samples.
Consequently, the symbols are demodulated at the receiver with an analog-to-digital
(A/D) converter followed by a DFT.
9
OFDM has two primary drawbacks. The first is sensitivity to imperfect frequency
synchronization which is common for mobile applications. This sensitivity arises from
the close subcarrier spacing. Figure 1.5 shows that the subcarriers are properly orthog-
onal at f = k/TB , k = 0, 1, . . . , N − 1. However, if the frequency synthesizer at the
receiver is misaligned by, say, fo /TB Hz, where −0.5 < fo < 0.5, the subcarriers are
not orthogonal and therefore interfering with one another. This intercarrier interference
(ICI) is illustrated in Figure 1.7: assuming that the receiver is tuned to (k + fo )/TB
Hz rather than at the ideal k/TB Hz, the N − 1 neighboring subcarriers interfere with
the demodulation of the kth subcarrier. The intercarrier interference causes ISI—and
potentially high irreducible error floors.
The second problem with OFDM is that the signal has large amplitude fluctuations
caused by the summation of the complex sinusoids. The real and imaginary part of the
Spectrum magnitude, |S(f )/TB |
0.2
0.04
OFDM signal is
N
X −1
< {s(t)} = < {I0,k } cos (2πkt/TB ) − = {I0,k } sin (2πkt/TB ) , (1.14)
k=0
and
N
X −1
= {s(t)} = < {I0,k } sin (2πkt/TB ) + = {I0,k } cos (2πkt/TB ) , (1.15)
k=0
respectively. Figure 1.8(a) shows the real and imaginary parts of an example OFDM
signal with N = 16 subcarriers. Also plotted are the individually modulated sinusoids.
Notice that each sinusoids has a constant amplitude, but when summing the sinusoids
the resulting OFDM signal fluctuates over a large range. The instantaneous signal power,
|s(t)|2 = <2 {s(t)} + =2 {s(t)}, is plotted in Figure 1.8(b). The ratio between the peak
power and the average power is 144/16 = 9 (or in decibels, 10 log 10 9 ≈ 9.5 dB).
160
12
10 140
8 Subcarriers |s(t)|2
<{s(t)} 120 Peak power
={s(t)} Average power
6
Power magnitude
100
Signal amplitude
2
80
0
60
-2
-4 40
-6
20
-8
0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Normalized time, t/TB Normalized time, t/TB
Figure 1.8: A typical OFDM signal (N = 16). The PAPR is 9.5 dB.
OFDM’s high peak-to-average power ratio (PAPR) requires system components with
a large linear range capable of accommodating the signal. Otherwise, the circuitry
11
distorts the waveform nonlinearly, and nonlinear distortion results in a loss of subcarrier
orthogonality which degrades performance.
One such nonlinear device is the transmitter’s power amplifier (PA) which is respon-
sible for the system’s operational range [424]. Ideally the output of the PA is equal to
the input times a gain factor. In reality the PA has a limited linear region, beyond which
it saturates to a maximum output level. Figure 1.9 shows a representative input/output
curve, known as the AM/AM conversion. In the linear region the curve matches the
ideal, but as the input power increases the PA saturates. The most efficient operating
point is at the PA’s saturation point, but for signals with large PAPR the operating
point must shift to the left keeping the amplification linear. The average input power
is reduced and consequently this technique is called input power backoff (IBO). To keep
the peak power of the input signal less than or equal to the saturation input level, the
IBO must be at least equal to the PAPR. Thus the required IBO for the OFDM signal
in Figure 1.8 is 9.5 dB. At this backoff the efficiency of a Class A power amplifier is
less than 6%. Such an efficiency is detrimental to mobile battery-powered devices which
have limited power resources. Moreover, the operational range of the system is reduced
by a factor of nine2 .
Max output
Optimum
Actual
Backoff
AM/AM curve
Operating points
Ideal AM/AM
Input power
Constant envelope (CE) waveforms are appealing since the optimum operating point
in Figure 1.9 is attainable. The baseband CE signal representation is
where A is the signal amplitude and φ(t) is the information bearing phase signal. The
advantage of the CE waveform is that the instantaneous power is constant: |s(t)| 2 = A2 .
Consequently, the PAPR is 0 dB and the required backoff is 0 dB. The PA can therefore
operate at the optimum (saturation) point, maximizing average transmit power (good
for range) and maximizing PA efficiency (good for battery life). Also, since the linearity
requirement is reduced, nonlinear PAs can be used which are generally more efficient
and less expensive than linear PAs. For example, the maximum theoretical efficiency of
a linear Class A power amplifier is 50%, while for a nonlinear Class E PA the maximum
theoretical efficiency is 100% [424].
Constant envelope signals are thus ideal in terms of the practical considerations of the
power amplifier. The question is how to embed digital information into φ(t) providing
good performance, spectral economy, and high data rates over the wireless channel.
Notice that the single carrier signal in (1.1) is constant envelope when |I i | = 1 and g(t)
is rectangular. This type of modulation, however, has large spectral sidelobes which
cause adjacent channel interference. In practice, non-rectangular pulse shapes are used
which result in a non-CE signal.
Continuous phase modulation (CPM) is a class of signaling that has very low sidelobe
power while maintaining the constant envelope property [14,421]. CPM uses memory to
smooth φ(t). The memory, however, increases the complexity of the receiver, which is a
key disadvantage of CPM. Also CPM systems have difficulty operating over frequency-
selective channels [118].
13
Constant envelope OFDM (CE-OFDM) combines OFDM and constant envelope sig-
naling. The high peak-to-average power ratio OFDM signal is transformed into a CE
waveform. The CE-OFDM signal takes the form of (1.16) where the phase signal is an
OFDM waveform. For example, the phase signal can be the real part of the OFDM
signal:
N
X −1
φ(t) = < {sOFDM (t)} = < {I0,k } cos (2πkt/TB ) − = {I0,k } sin (2πkt/TB ) , (1.17)
k=0
where sOFDM (t) is the signal in (1.6). Figure 1.10 compares a conventional OFDM
bandpass signal with a bandpass CE-OFDM signal. Both are derived from the same
baseband OFDM message signal.
OFDM bandpass
OFDM message
R
CE-OFDM bandpass
The motivation for CE-OFDM is to eliminate the PAPR problem of the conventional
OFDM system. Certainly, this is accomplished since the CE-OFDM signal has the 0 dB
PAPR property. The question is: at what cost? What is the performance of CE-
OFDM? What is its bandwidth? Can the guard interval be used in CE-OFDM as it is
in conventional OFDM? This thesis aims to answering these questions by analyzing the
various aspects of the CE-OFDM modulation.
14
In Chapter 2 the basics of OFDM is further studied. The effect of the nonlinear
power amplification on OFDM is evaluated. In Chapter 3 the CE-OFDM modulation
format is defined and the spectral properties are studied. The performance aspects of
CE-OFDM in the presence of additive noise are analyzed in Chapter 4. Performance
analysis is extended to frequency-nonselective fading channels in Chapter 5, and multi-
path frequency-selective fading channels in Chapter 6.
Chapter 2
OFDM
In Sections 1.1 and 1.2 the basic properties of OFDM are identified. In this chapter,
OFDM is studied in more detail. Section 2.1 covers key properties of OFDM. In Section
2.1.1, the cyclic prefix is studied. In Section 2.1.2, the processing of the discrete-time
samples is described, and the equivalence of linear channel convolution and circular
channel convolution is explained. In light of this property, OFDM is considered a special
case of the more general block modulation with cyclic prefix scheme, as discussed in
Section 2.1.3. Finally, in Section 2.1.4 the main functional blocks of the OFDM system
are described.
The PAPR statistics are analyzed in Section 2.2 and power amplifier models used
to evaluated system performance are described in Section 2.3. Then in Section 2.4 the
effect of nonlinear power amplification on OFDM systems is studied in terms of spectral
leakage (Section 2.4.1), performance degradation (Section 2.4.2), and system range and
efficiency (Section 2.4.3). Lastly, the various PAPR mitigation techniques found in the
research literature are categorized in Section 2.5, and a technique called signal clipping
is evaluated in terms of its effectiveness to improve system performance.
15
16
In Section 1.1.1 it is claimed that the use of the guard interval results in ISI-free
operation. This is true so long as a cyclic prefix is transmitted during the interval. This
is demonstrated below and it is shown that ISI results if anything but the cyclic prefix
is transmitted.
where {Ik }N −1 N −1
k=0 are the data symbols, {exp(j2πf k t)}k=0 are the subcarriers, N is the
total number of subcarriers, fk = k/TB is the center frequency of the kth subcarrier and
TB is the block period. The guard interval is defined during −T g ≤ t < 0, where Tg is
the guard period. To transmit a cyclic prefix, the last T g s of the block is transmitted
during the guard interval:
N
X −1 N
X −1 N
X −1
s(t) = Ik ej2πfk (t+TB ) = Ik ej2πfk t ej2πk = Ik ej2πfk t , (2.2)
k=0 k=0 k=0
−Tg ≤ t < 0. Notice that the above simplification is made due to the periodicity of the
signal. Thus the OFDM signal having a guard interval with cyclic prefix is simply
N
X −1
s(t) = Ik ej2πfk t , −Tg ≤ t < TB . (2.3)
k=0
where h(τ ) is the time-invariant channel impulse response 1 and n(t) is additive noise.
The bounds of integration are simplified since the channel is assumed causal [h(τ ) = 0
1
In (1.2), the received signal is expressed in terms of the time-variant channel impulse response h(τ, t).
If the channel is assumed to be time invariant, the impulse response is referred to as simply h(τ ).
17
for τ < 0] and to have a maximum propagation delay τ max [h(τ ) = 0 for τ > τmax ]. The
received signal during the guard interval, which has interference from the previous block
(see Figure 1.4), is ignored and r(t) during 0 ≤ t < T B is processed. An estimate of the
k0 th data symbol is made by correlating r(t) with the k 0 th subcarrier:
Z TB
1 h i∗
Iˆk0 = r(t) ej2πfk0 t dt, (2.5)
TB 0
which expands to
TB
1
Z
Iˆk0 = r(t)e−j2πfk0 t dt
TB 0
Z TB "Z τmax N −1
#
1 X
j2πfk (t−τ )
= h(τ ) Ik e dτ e−j2πfk0 t dt + Nk0 (2.6)
TB 0 0 k=0
N −1 τ Z TB
1 X
Z max
= Ik h(τ )e−j2πfk τ dτ ej2πt(fk −fk0 ) dt + Nk0 ,
TB 0 0
k=0
where
TB
1
Z
N k0 = n(t)e−j2πfk0 t dt. (2.7)
TB 0
But since
TB 1, k = k0 ,
1
Z
ej2πt(fk −fk0 ) dt = (2.8)
TB 0 0,
k 6= k0 ,
(2.6) simplifies to
Iˆk0 = Ik0 H[k0 ] + Nk0 , (2.9)
where Z τmax
H[k0 ] = h(τ )e−j2πfk0 τ dτ, (2.10)
0
which is the Fourier transform of h(τ ) evaluated at f = f k0 .
Now it is shown that by transmitting a signal other than the cyclic prefix during the
guard interval causes ISI. Suppose that the transmitted signal is
b(t),
−Tg ≤ t < 0,
s(t) = P (2.11)
N −1 Ik ej2πfk t , 0 ≤ t < TB ,
k=0
18
PN −1
where b(t) 6= k=0 Ik ej2πfk t . The estimate of the k0 th data symbols is
Z TB
ˆ 1
I k0 = r(t)e−j2πfk0 t dt
TB 0
Z TB Z τmax
1 (2.12)
= h(τ )s(t − τ )dτ e−j2πfk0 t dt + Nk0
TB 0 0
= A k0 + B k0 + N k0 .
The bounds of integration are separated into two segments, [0, T g ] and [Tg , TB ]:
Tg τmax
1
Z Z
A k0 = h(τ )s(t − τ )e−j2πfk0 t dτ dt, (2.13)
TB 0 0
and
TB τmax
1
Z Z
B k0 = h(τ )s(t − τ )e−j2πfk0 t dτ dt. (2.14)
TB Tg 0
Ak0 is a non-zero offset term which is a function of b(t). For the second term, t − τ > 0,
thus
N −1
"Z #
TB τmax
1
Z X
j2πfk (t−τ )
B k0 = h(τ ) Ik e dτ e−j2πfk0 t dt
TB Tg 0 k=0
(2.15)
N −1 τmax TB
1 X Z Z
= Ik h(τ )e−j2πfk τ dτ ej2πt(fk −fk0 ) dt.
TB 0 Tg
k=0
Due to the integration bounds for t, the orthogonality condition in (2.8) can’t be applied
to (2.15), and this results in ISI. The estimated data symbol is expressed as
The interference is denoted as ICI, intercarrier interference, since the subcarriers are no
longer orthogonal and interfere with one another. This phenomenon was described in
Section 1.2 in the context of imperfect frequency synchronization. Therefore, ICI can
manifest itself in more than one way, and when it does the data symbols interfere with
one another resulting in ISI.
In [358], cyclic prefixed OFDM is compared to zero-padded OFDM [b(t) = 0]. The
zero-padding causes ISI, but has the advantage of being able to recover data symbols
19
located at channel zeros. This is in contrast with cyclic prefixed OFDM since, as shown
in (2.9), a channel zeros at the kth subcarrier, that is, H[k] = 0, results in an estimated
data symbol that consists entirely of noise. The zero-padded system avoids this problem
at the cost of increased receiver complexity due to equalization requirements.
··· ··· t
−Ng Tsa −Tsa 0 Tsa (NB − 1)Tsa
−Ng Tsa ≥ −Tg
··· τ
0 Tsa (Nc − 1)Tsa
(Nc − 1)Tsa ≤ τmax
where {n[i]} are samples of the noise signal n(t). The guard interval samples are ignored
and the samples
NX
c −1
are processed.
and
NDFT
X−1
H[k] = h[i]e−j2πik/NDFT , k = 0, . . . , NDFT − 1 (2.26)
i=0
are the NDFT -point DFTs of the signal and channel samples, respectively. The DFT
size is, in general, NDFT ≥ NB . If NDFT > NB , the signal vector is zero-padded. Since
NDFT > Ng , the channel samples are zero-padded: h[i] = 0 for i = N c , . . . , NDFT − 1.
Figure 2.2 shows a block diagram representing the calculation of (2.24). The effect of
the channel is simply a DFT followed by a multiplier bank (H[k]), which is then followed
by an IDFT. Also shown is the inverse channel which is a DFT followed by a multiplier
bank (1/H[k]) followed by an IDFT. Thus the transmit samples s[i] can be reconstructed
by passing the receive samples r[i] through the inverse channel.
The inverse channel structure in Figure 2.2 corrects the distortion caused by the
channel in the frequency domain, and is therefore called a frequency-domain equalizer
21
Channel
Inverse channel
1
r[i] DFT IDFT s[i]
H[k]
Figure 2.2: Circular convolution with channel and the inverse channel.
Frequency-domain equalizer
Frequency-domain equalizer
(FDE). Such an equalizer can be used only when the effect of the channel is a circular
convolution. This is the case for OFDM, but isn’t unique to OFDM since any modulation
can use a cyclic prefix. This observation was first identified by Sari et al. [462] and
suggests a more general modulation approach: block modulation with cyclic prefix and
frequency-domain equalization. Figure 2.3 shows a simplified block diagram of such
a system. (The insertion of the cyclic prefix at the transmitter and removal at the
receiver is implied but not included in the diagram for simplicity.) For the special case
of OFDM, the modulation is a IDFT and the demodulation is a DFT as shown in Figure
2.4. Notice that the DFT and IDFT cancel each other and the resulting diagram depicts
the conventional OFDM system.
The multiplier bank at the output of the DFT is often referred to as a one-tap
equalizer, one complex multiplication per frequency bin. This operation is required for
data symbols that rely on coherent demodulation, such as M -ary phase-shift keying
(M -PSK) and M -ary quadrature-amplitude modulation (M -QAM).
As Sari et al. pointed out, OFDM doesn’t eliminate the equalization problem (asso-
ciated with conventional single carrier modulation); rather, OFDM converts the problem
to the frequency domain. Since Sari’s original paper, there has been a considerable num-
ber of publications focused on the block modulation technique using conventional single
carrier modulations [8, 30, 54, 107, 116, 132, 142, 153, 154, 196, 197, 245, 388, 460, 461, 463,
481, 533, 565, 574].
The block diagram in Figure 2.4 conceptually illustrates the OFDM system. Figure
2.5 shows a more detailed description of OFDM’s functional blocks.
The encoder adds redundancy to the bit stream for error control. The encoded bits
are then mapped to the data symbols I k . In general, the data symbols are complex
numbers which result from mapping the bits to points on the complex plane. Next, the
symbols are serial-to-parallel (S/P) converted and processed by the IDFT. The cyclic
prefix is added and the signal samples, s[i], are passed through the digital-to-analog
(D/A) converter to obtain the continuous-time OFDM signal s(t). Finally, the signal is
amplified and transmitted.
23
Transmitter
Receiver
At the receiver, the inverse operations are performed. First, the received signal,
r(t), is sampled to obtain the discrete-time sequence r[i]. The guard interval samples
are removed, the DFT is performed and each frequency bin is equalized by a complex
multiplication. The estimated data symbols, Iˆk , are processed by the detector which
outputs a stream of estimated receive bits, and the decoder attempts to correct any bit
errors that may have occurred.
As discussed in Section 1.2, one of OFDM’s key drawbacks is the high peak-to-
average power ratio. Nonlinearities in the power amplifier distort the transmitted signal
and large input power backoff is required which results in low amplifier efficiency. In the
next sections the impact of the PA is studied. But first, the statistical properties of the
PAPR are discussed.
24
The peak-to-average power ratio of the OFDM signal is best viewed statistically. For
any given block interval, the PAPR is a random quantity since it depends on the data
symbols {Ik }N −1
k=0 . Assuming that they’re selected randomly from a set of M complex
numbers, there are M N unique symbol sequences, and thus M N unique OFDM wave-
forms per block. Of these waveforms, some have a high PAPR, while others have a
relatively low PAPR. Therefore, it is desirable to understand the statistical distribution
of this quantity.
The signal during the guard interval is ignored since it has no impact on the PAPR
distribution. M -PSK data symbols are assumed, therefore |I k | = 1 for all k. The
average power of s(t) is
TB
1
Z
Ps = |s(t)|2 dt = N. (2.28)
TB 0
The peak-to-average power ratio is defined as
.
PAPRs = max |s(t)|2 Ps . (2.29)
t∈[0,TB )
Notice that the absolute maximum signal power is N 2 , so the PAPR can be as high as
N . However, the likelihood that all the subcarriers align in phase is extremely low. For
example, as pointed out in [381], a N = 32 subcarrier system having 4-ary data symbols
and a block period of TB = 100 µs obtains the theoretical maximum PAPR once every
3.7 million years. Thus it is more meaningful to describe the PAPR statistically rather
than in absolute terms.
Since the average signal power is a constant, the randomness of the PAPR depends on
the randomness of the instantaneous power |s(t)| 2 , and more specifically, the maximum
instantaneous power over 0 ≤ t < TB . For large N , the real and imaginary parts of
s(t) are accurately modeled as Gaussian random processes (due to the application of
the central limit theorem [394, 421]). Consequently, the instantaneous signal power is
chi-squared distributed with two degrees of freedom [421, p. 41], and the complementary
25
100 100
Simulation Simulation
Approximation (2.30) Lower bound (2.31)
10−1 10−1
´
CCDF, P |s(t)|2 /Ps > x
10−2 10−2
`
10−3 10−3
10−4 0 2 4 6 8 10 10−4 4 6 8 10 12 14
x (dB) x (dB)
Figure 2.6(a) compares a simulated instantaneous power CCDF with the approxi-
mation in (2.30). This demonstrates the accuracy of the Gaussian approximation to the
real and imaginary part of s(t). Figure lower bound in (2.31). 2.6(b) compares PAPR
simulation results to the The bound is shown to be within 1 dB of the simulated result
for lower values of x. The 0.0001 PAPR is shown to be at around 11.25 dB, and at this
26
level the bound is tight. Notice that essentially all OFDM blocks have a PAPR greater
than 6 dB, 10% have a PAPR greater than 8.5 dB, and 0.5% have a PAPR greater than
10 dB.
For the results in Figure 2.6, the number of subcarriers is N = 64 and QPSK data
symbols (4-ary PSK) are used, that is, I k ∈ {±1, ±j}. While the symbols constellation
has little impact on the PAPR statistics, the number of subcarriers does. Figure 2.7
shows the lower bound (2.31) over a range N = 32 to N = 1024. Notice that the 0.001
PAPR is 1 dB larger for N = 512 than for N = 32. For the N = 64 system, the PAPR
is greater than 8 dB for roughly 10% of the time. For the N = 1024 system, however,
the PAPR is greater than 8 dB nearly all of the time.
100
k 5 6 7 8 9 10
10−1
CCDF, P (PAPRs > x)
10−2
10−3
10−4 4 5 6 7 8 9 10 11 12
x (dB)
To determine the impact of the PAPR on system performance, power amplifier mod-
els must be defined. Two models commonly used in the research literature are the
solid-state power amplifier (SSPA) model and the Saleh traveling-wave tube amplifier
(TWTA) model [454]. They are described here and then used in Section 2.4 for perfor-
mance evaluation.
27
In general, modeling nonlinear power amplifiers is complicated (see [233, chap. 5]).
A common simplification is to assume that the PA is a memoryless nonlinearity, and
therefore has a frequency-nonselective response. For example, if the PA input is
the output is
sout (t) = G[A(t)] exp j{φ(t) + Φ[A(t)]} , (2.33)
where G(·) and Φ(·) are known as the AM/AM and AM/PM conversions, respectively.
The SSPA model is expressed as
g0 A
G(A) = h i1/2p , and Φ(A) = 0, (2.34)
2p
1 + (A/Asat )
where g0 is the amplifier gain, Asat is the input saturation level, and p controls the
AM/AM sharpness of the saturation region. For this model the AM/PM conversion is
assumed to be negligibly small.
Though widely known as the Rapp model [426], (2.34) should be credited to the
original work by A. J. Cann, published a decade earlier in the IEEE literature [71].
Cann’s formula is obtained with the simple manipulation:
A
G(A) =g0 h i1/2p
1 + (A/Asat )2p
A [(Asat /A)2p ]1/2p
=g0 h i1/2p × (2.35)
1 + (A/Asat )2p [(Asat /A)2p ]1/2p
Asat
=g0 h i1/2p ,
1 + (Asat /A)2p
g0 A αφ A2
G(A) = , and Φ(A) = . (2.36)
1 + (A/Asat )2 1 + β φ A2
Notice that the AM/PM conversion, determined by the constants α φ and βφ , is non-zero.
The TWTA model is therefore more nonlinear than the SSPA model.
28
To reduce nonlinear distortion in the amplified OFDM signal, input power backoff
(IBO) is required. It is defined as [375]
A2sat
IBO = , (2.37)
Pin
where Pin = E{|sin (t)|2 } = E{A2 (t)} is the average power of the input signal. Equiva-
lently, (2.37) can be written as
A2sat
Pin = ; (2.38)
IBO
thus, given Asat and IBO, the input signal power can be scaled accordingly to satisfy
(2.38).
Assuming that the PAPR of the input signal is PAPR in , the peak power can be
written as
PAPRin 2 A2
Pmax = PAPRin · Pin = Asat = sat , (2.39)
IBO K
where
IBO
K= (2.40)
PAPRin
is defined as the backoff ratio. Notice that for K > 1 the backoff is greater than the input
signal’s PAPR; for K < 1 the backoff is less than the input PAPR. Now, the maximum
value of the input, Amax = max |A(t)|, can be written in terms of the backoff ratio and
the input saturation level:
p Asat
Amax = Pmax = √ . (2.41)
K
Figure 2.8 shows the AM/AM (solid lines) and AM/PM (dashed lines) conversions
for the SSPA (thick lines) and TWTA (thin lines) models for various backoff ratios K.
For the SSPA model, p = 2; for the TWTA model, α φ = π/12 and βφ = 1/4. The x-axis
is normalized to the maximum input level A max , and the y-axis is normalized to the
maximum output level g0 Asat . For K = −10 dB the IBO is one-tenth the input signal
PAPR, and thus the nonlinearity is severe. One the other hand, for K = 10 dB the
IBO is ten times the input signal PAPR and the PA response is nearly linear. As stated
above, the non-zero AM/PM conversion of the TWTA model makes it more nonlinear
than the SSPA model.
Insight can be gained by comparing Figure 2.6(b) and Figure 2.8. For example,
assuming that the backoff is IBO = 6 dB, the conversions are never as linear as the
K = 3 dB curves (the PAPR is a always greater than 3 dB) and are more nonlinear
29
1 1
0.5 0.5
0 0
−0.5 −0.5
−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
Normalized input value, A/Amax Normalized input value, A/Amax
(a) K = 10 dB (b) K = 3 dB
1 1
Normalized output value, G(A)/g0 Asat
Normalized output value, G(A)/g0 Asat
0.5 0.5
0 0
−0.5 −0.5
−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
Normalized input value, A/Amax Normalized input value, A/Amax
than the K = −3 dB curves for about 5% of the OFDM blocks (the 0.05 PAPR is 9 dB).
Therefore, even with a large IBO of 6 dB, the PA can impose high nonlinear distortion
on the transmitted signal. Also, the degree of distortion for a given OFDM block is
random (given a fixed IBO) since the PAPR for a given block is random.
The first problem considered is spectral leakage. By using the Welch method [422, pp.
911–913], the power density spectrum at the output of the power amplifier can be quickly
estimated. The result is used to calculate estimated fractional out-of-band power curves,
defined as Rf
ˆ 0 Φ̂s (x)dx
FOBP(f )= , f > 0, (2.42)
0.5P̂s
R∞
where Φ̂s (f ) is the estimated power density spectrum of the signal and P̂s = −∞ Φ̂s (f )df
is the signal power. Figure 2.9 shows the curves for an N = 64 subcarrier OFDM signal
amplified by the TWTA power amplifier according to (2.36) at various backoff levels.
Also plotted is the FOBP curve for ideal linear amplification. These results show that
at least 6 dB backoff is required by the TWTA to avoid spectral broadening.
Figure 2.10 shows the 99.5% bandwidth as a function of IBO. The bandwidth of the
undistorted OFDM signal is f = 1.07W . For sufficient backoff, the bandwidth of the
nonlinearly amplified signal is the same. However, for IBO < 6 dB, the bandwidth is
shown to grow roughly linearly with IBO. For IBO = 1 dB, the 99.5% bandwidth is 73%
larger than the undistorted signal. Notice that the spectral leakage is roughly the same
for the two amplifier models.
31
100
OFDM amplified with: TWTA PA
ideal PA
10−1
Fractional out-of-band power
IBO
0
10−2
2
10−3
0 0.25 0.5 0.75 1 1.25 1.5
Normalized frequency, f /W
Figure 2.9: Fractional out-of-band power of OFDM with ideal PA and with TWTA
model at various input power backoff. (N = 64, IBO in dB)
1.6
1.4
1.2
1.0
0.8
0 2 4 6 8 10
Input power backoff, IBO (dB)
where sout (t) is the output of the PA from (2.33) and n(t) is a complex-valued Gaussian
additive noise signal having a power density spectrum [421, p. 158]
N0 , |f | ≤ Bn /2,
Φn (f ) = (2.44)
0,
|f | > Bn /2,
where Bn is the bandwidth of the noise signal. The noise spectrum is assumed to be
constant over the effective bandwidth of the information bearing signal and is thus called
“white”. The transmitted data symbols are estimated by the correlation in (2.5) then
passed to the detector which makes the final decision. This decision is based on the
maximum-likelihood (ML) criterion assuming a linear PA; that is, the nearest point in
the symbol constellation [421, pp. 242–247].
The autocorrelation function of n(t) [the inverse Fourier transform of (2.44)] has zero-
crossings at τ = 1/Bn . Thus assuming Bn = fs , (2.45) is satisfied and the noise sample
variance is σn2 = fsa N0 .
Figure 2.11 shows bit error rate (BER) performance as a function of E b /N0 , where
R TB
0|sout (t)|2 dt
Eb = (2.46)
Number of bits per block
33
10−1
Nonlinear PA Nonlinear PA
Ideal PA Ideal PA
10−1
10−2
10−2
Bit error rate
10−4
10−4
10−5 0 2 4 6 8 10 12 14 10−5 0 5 10 15 20 25 30
Signal-to-noise ratio per bit, Eb /N0 (dB) Signal-to-noise ratio per bit, Eb /N0 (dB)
(a) SSPA model, IBO = 0, 1, 2, 3, 4, 6, 8 dB; (b) TWTA model, IBO = 0, 1, . . . , 10, 16 dB;
0 = worst, 8 = best. 0 = worst, 16 = best.
Figure 2.11: Performance of QPSK/OFDM with nonlinear power amplifier with various
input power backoff levels. (N = 64)
is the energy per bit. The quantity E b /N0 is referred to as the signal-to-noise ratio (SNR)
per bit, or simply the SNR. QPSK data symbols are used, and the oversampling factor
is J = 4. For the SSPA results in Figure 2.11(a), the IBO ranges from 0 to 8 dB. At
the 0.0001 BER level, the IBO = 0 dB case suffers a 3 dB performance loss compared to
ideal AWGN performance, which is [421, pp. 268].
r !
Eb
BER = Q 2 , (2.47)
N0
R∞ 2 /2 √
where Q(x) = x e−y dy/ 2π is the Gaussian Q-function. To avoid degradation, 8
dB of backoff is required. The TWTA results in Figure 2.11(b) use IBO ranging from 0
to 16 dB. Notice the irreducible error floors for IBO ≤ 7 dB. To avoid degradation, 16
dB of backoff is required—8 dB more than for the SSPA case. The greater nonlinearity
of the TWTA model is evident from the results in this figure.
Figure 2.12 compares performance for higher-order PSK modulations. For M -PSK
34
10
Ideal PA
10−1 SSPA: IBO = 3 dB
IBO = 6 dB M = 16
8
M = 16
6
M =8
10−3
4 Ideal PA
SSPA
M = 2, 4
10−4 M = 2, 4
2
10−5 0 5 10 15 20 25 30
0
0 2 4 6 8 10
Signal-to-noise ratio per bit, Eb /N0 (dB) Input power backoff, IBO (dB)
The number of bits per data symbols is log 2 M , therefore the bit energy is
R TB
|sout (t)|2 dt
Eb = 0 . (2.49)
N log2 M
Higher-order constellations are used for increased spectral efficiency at the price of BER
performance2 . In Figure 2.12(a) BER results for the SSPA model are shown. (The
results for M = 2 and M = 4 are very similar so only M = 2 is plotted.) The higher-
order modulations are shown to be more sensitive to the PA nonlinearity. For example,
the M = 16 result for IBO = 3 dB has an irreducible error floor at 5 × 10 −3 , while the
M = 2, 4 result at the same backoff shows only a 1 dB degradation. When increasing
the backoff to IBO = 6 dB, the error floor for M = 16 drops to 2 × 10 −5 and the 0.001
BER is about 2 dB worse than AWGN. Using IBO = 6 dB for M = 8 results in 2 dB
less degradation at the 0.001 bit error rate when compared to using IBO = 3 dB.
2
This is the case for linear modulation formats. This isn’t necessarily the case for nonlinear modulation
formats as discussed in Section 4.4.
35
where SNRAWGN is the required signal-to-noise ratio per bit to achieve a target bit
error rate in AWGN; SNRPA (IBO) is the required SNR when taking into account the
distortion caused by the power amplifier at a given backoff. The “optimum” IBO, denote
as IBOopt , minimizes the total degradation, that is,
The target BER for the curves in Figure 2.12(b) is 0.001. Clearly the modulation order
influences the degradation. The minimum TD for M = 16 is 7.7 dB at IBO opt = 6.5
dB; for M = 8, TDmin = 5 dB at IBOopt = 3 dB. This can be interpreted as follows:
M = 8, while having lower spectral efficiency than M = 16 (3 b/s/Hz vs. 4 b/s/Hz),
suffers less degradation and can operate with less backoff, resulting in improved range
and higher PA efficiency. The M = 2 and M = 4 examples are shown to have the lowest
degradation and are thus the more robust against nonlinear distortion.
The total degradation is directly related to the system’s operational range. Consider
a transmitter operating at maximum transmit power. The range is represented by the
outermost ring in Figure 2.13. Now assume that the system requires a 3 dB backoff:
the range is reduced by one-half, as represented by the middle ring. Any degradation
caused by the PA further reduces range, as represented by the innermost circle. Thus
the actual range of the system is far less than the potential range of the transmitter.
The true capability of the power amplifier is greatly underutilized.
To quantify the relationship between the PA efficiency and the power backoff, the
theoretical efficiency of a Class A power amplifier is used [374]:
1 1
ηA = × 100%, IBO ≥ 1. (2.52)
2 IBO
The efficiency is thus inversely proportional to IBO and the maximum efficiency, 50%,
occurs at IBO = 1 (0 dB). The efficiency curve, shown in Figure 2.14, can be used
36
Potential range
Potential range w/ IBO
Actual range
Figure 2.13: The potential range of system is reduced with input backoff; the range is
reduced further from nonlinear amplifier distortion.
in conjunction with Figures 2.10 and 2.12(b) to gain insight to the various tradeoffs
between PA efficiency, spectral containment, and performance/range. For example, the
optimum IBO in terms of total degradation for the 8-PSK SSPA example is IBO opt = 3
dB [Figure 2.12(b)]: however, the bandwidth expansion is 42% (Figure 2.10) and the
PA efficiency is ηA = 25% (Figure 2.14). The optimum IBO for the 16-PSK example,
6.5 dB, results in no bandwidth expansion but the PA efficiency is reduced to 11%. The
M = 2, 4 systems required minimal IBO for the SSPA, thus maximizing efficiency, but
the bandwidth expands by 87%.
50
45
Class-A PA efficiency, ηA (%)
40
35
30
25
20
15
10
5
00 1 2 3 4 5 6 7 8 9 10
Input power backoff, IBO (dB)
There have been many schemes proposed in the research literature aimed at reducing
the impact of the PAPR problem. The goal of any scheme is to reduce the minimum
total degradation (for increased range) and the IBO opt (for increased PA efficiency). The
various schemes can be placed in one the following three categories:
The PA linearization schemes attempt to predistort the OFDM signal such that the
overall response of the predistorter followed by the PA is linear—essentially equalizing
the amplifier. In [230], an LMS algorithm is applied for adaptive predistortion; in [395]
a neural network learning technique is used. Parametric techniques, which design a
predistorter based on a PA model, have been proposed. In [85, 122, 250, 567] nonlinear
polynomial models are used, and in [86] a Volterra-based model is suggested.
The second category, receiver enhancement techniques, have been suggested in [513],
[376] (maximum-likelihood decoding); in [259, 453] (signal reconstruction), and in [87]
(interference cancellation). Finally, the third category includes techniques that are based
on transforming the OFDM signal prior to the PA, and applying the inverse transform at
the receiver prior to demodulation. This category includes constant envelope OFDM (as
studied in the second half of this thesis) which uses a phase modulator as the transformer.
In [215, 329, 569–571] a companding transform is suggested.
38
Signal Clipping
The remainder of this section focuses on the effectiveness of signal clipping, which has
been claimed to be the “simplest” and “most effective” PAPR reduction scheme [27, 87,
290, 375, 377, 380, 382, 391]. The impact of “clipping noise”—the intercarrier interference
caused by the clipping process—on system performance has been extensively analyzed
[39, 124, 382]. However, a common assumption is that the PA is linear [27, 39, 138,
290, 371, 380, 382, 391]. It is argued here that the effectiveness of a PAPR reduction
scheme must be measured not only by PAPR reduction, but by the more meaningful
measures of TDmin and IBOopt reduction. It is shown that clipping, while an effective
PAPR reduction scheme, does not reduce TD min nor does clipping reduce IBOopt for
an OFDM system. This result brings into question the usefulness of non-distortionless
PAPR reduction techniques in general.
The system under consideration is shown in Figure 2.15. When the switch is “on”
the PAPR reducing signal clipper is used. When “off” the system is identical to the one
studied in Section 2.4.2. Therefore, the earlier unclipped results serve as a performance
benchmark in which to compare the clipped results. The channel, as before, has an
impulse response h(τ ) = δ(τ ).
Figure 2.15: Block diagram. The system is evaluated with and without PAPR reduction.
The input to the clipping block is the OFDM signal s(t) from (2.27), the output is
the clipped OFDM signal:
s(t),
if |s(t)| ≤ Amax ,
sclip (t) = (2.53)
Amax ejψ(t) ,
if |s(t)| > Amax ,
where ψ(t) = arg[s(t)]. Therefore, the magnitude of the clipped signal does not exceed
Amax and the phase of s(t) is preserved. (This has been called “polar clipping” in the
literature [276].) The clipping severity is measured by the clipping ratio, defined as [375]
Amax
γclip = √ . (2.54)
Ps
39
2
10
0
Imaginary axis
−10
−20
−20 −10 0 10 20
Real axis
Figure 2.16: Unclipped OFDM signal (9.25 dB PAPR). The rings have radius A max
which correspond to various clipping ratios γ clip (dB).
Figure 2.16 shows a typical OFDM signal on the complex plane. The dark rings have
radius Amax which correspond to clipping ratios γ clip = 0, 2, and 4 dB.
Clipping’s effectiveness at reducing PAPR is shown in Figure 2.17. For clipping ratio
γclip = 5 dB, the peak-to-average power ratio of the clipped signal is PAPR clip ≤ 10
dB; for γclip = 4 dB, PAPRclip ≤ 8 dB, and so forth. The 0.0001 PAPR improvement,
compared to the unclipped signal, is 1.2 dB for γ clip = 5 dB and by 3.2 dB for the
γclip = 4 dB.
Figure 2.18 shows PAPRclip as a function of the clipping ratio. The PAPR of
the unclipped signal is 13 dB3 . Notice that for large γclip , sclip (t) is unclipped, there-
3
This figure is made by generating 2 × 104 consecutive OFDM blocks. The PAPR of the overall block
is 13 dB.
40
100
10−1 Clipped
Unclipped
P (PAPRclip > x)
10−2
10−3
γclip 3 4 5
10−4 0 2 4 6 8 10 12
x (dB)
Figure 2.17: PAPR CCDF of clipped OFDM signal for various γ clip (dB). [N = 64]
16
14
PAPRs
12
Peak-to-average power ratio (dB)
10
8
PAPRclip
6
0 PAPRclip as γclip → 0
−2
2
γclip
−4
−8 −6 −4 −2 0 2 4 6 8 10
Clipping ratio, γclip (dB)
Figure 2.18: PAPR of clipped signal as a function of the clipping ratio. (N = 64)
41
fore PAPRclip = PAPRs . As γclip → 0, the peak and average powers converge, thus
PAPRclip → 0 dB. For the region 3 dB < γclip < 6.5 dB, sclip (t) is clipped so the
peak power is A2max = γclip
2 P . However, the clipping is mild so the average power is
s
2 P /P = γ 2 .
approximately the same as s(t); therefore, PAPR clip ≈ γclip s s clip
Clipping is clearly an effective technique at reducing the PAPR. The question is, does
the PAPR reduction translate into reduced total degradation? Figure 2.19 compares the
total degradation curves of the unclipped system [from Figure 2.12(b)] with the clipped
system. Interestingly, the unclipped results are shown to provide a lower bound for
the clipped, reduced PAPR, system results. The clipper is shown to increase both the
minimum total degradation and the optimum backoff. For example, using the clipping
ratio γclip = 3 dB for the M = 8 case increases the TD min by 0.2 dB; using γclip = 2 dB
increases TDmin by 1.2 dB. For M = 16, the γclip = 4 dB result is nearly identical to the
unclipped result; γclip = 3 dB increases the degradation by 1.2 dB, and the TD curve
associated with γclip = 2 dB is beyond the viewing range of the figure. For M = 2, 4 the
PAPR reducing clipping yields nearly identical results as the unclipped system.
10
M = 16
8
Total degradation (dB)
6
M =8
4 Ideal PA
Unclipped
M = 2, 4 Clipped: γclip = 4 dB
3 dB
2 2 dB
0
0 2 4 6 8 10
Input power backoff, IBO (dB)
Figure 2.19: A comparison of the total degradation curves of clipped and unclipped
M -PSK/OFDM systems. (N = 64)
Thus the effectiveness of a PAPR reduction scheme should be measured not only
by its PAPR reducing capabilities but by its effectiveness in reducing total degradation
(which increases range) and reducing the optimum IBO (which increases power amplifier
42
efficiency). The distortion caused by non-distortionless schemes can outweigh the benefit
of the reduced PAPR. This is clearly shown to be the case for the clipped N = 64 M -
PSK/OFDM systems studied in this section. The clipping is shown to reduce the 0.0001
PAPR by > 1 dB, but this reduction does not translate into increased PA efficiency.
This result brings into question the validity of the claims that clipping is an effec-
tive scheme. In fact, the effectiveness of non-distortionless PAPR reduction schemes in
general is suspect. For these types of techniques it is important to take into account the
effect of the nonlinear power amplifier.
Conventional OFDM systems, even with the use of effective PAPR reduction and/or
power amplifier linearization techniques, typically require more input power backoff than
convention single carrier systems. Therefore, OFDM is considered power inefficient,
which is undesirable particularly for battery-powered wireless systems.
The technique described in the remainder of the thesis takes a different approach to
the PAPR problem. CE-OFDM can be thought of as a mapping of the OFDM signal to
the unit circle, as depicted in Figure 3.1. The instantaneous power of the resulting signal
is constant. Figure 3.2 compares the instantaneous power of the OFDM signal and the
mapped CE-OFDM signal. For the CE-OFDM signal the peak and average powers are
the same, thus the PAPR is 0 dB.
Signal
Unit circle
⇒
CE-OFDM
OFDM
43
44
5
OFDM
CE-OFDM
0
0 0.2 0.4 0.6 0.8 1
Normalized time
where {Ii,k } are the data symbols and {qk (t)} are the orthogonal subcarriers. For con-
ventional OFDM the baseband signal is up-converted to bandpass as
n o
y(t) = < m(t)ej2πfc t
(3.2)
= Am (t) cos [2πfc t + φm (t)] ,
where Am (t) = |m(t)| and φm (t) = arg[m(t)]. For real-valued m(t), φ m (t) = 0 and y(t)
is simply an amplitude modulated signal. (For complex-valued m(t), y(t) can be viewed
as an amplitude single-sideband modulation.) For CE-OFDM, m(t) is passed through a
phase modulator prior to up-conversion. The baseband signal is
= e−αAm (t) sin φm (t) cos [2πfc t + αAm (t) cos φm (t)] .
Transform
Receiver
Inverse
transform
The idea of transmitting OFDM by way of angle modulation isn’t entirely new.
In fact, Harmuth’s 1960 paper suggest transmitting information by orthogonal time
functions with “amplitude or frequency modulation, or any other type of modulation
suitable for the transmission of continuously varying [waveforms]” [202]. Using existing
FM infrastructure for OFDM transmission has been suggested in [76, 77, 575]. These
papers don’t consider the PAPR implications, however. Two conference papers, [101]
and [506], on the other hand, suggest using a phase modulator prior to the power amplifier
for PAPR mitigation—though intriguing, these papers lack a solid theoretical foundation
and ignore fundamental signal properties such as the signal’s power density spectrum.
The origin of this work, which is independent of the previous references, stems from
work done at the US Navy’s spawar Systems Center, (San Diego, CA). Mike Geile,
a principle engineer at Nova Engineering, (Cincinnati, OH), which is the contractor of
the OFDM component for JTRS (Joint Tactical Radio System), suggested a low PAPR
enhancement to OFDM by phase modulation. The motivation is to reduce the 6 dB
backoff used in the JTRS radio.
where Eq = TB /2.
where A is the signal amplitude. The phase signal during the ith block is written as
N
X
φ(t) = θi + 2πhCN Ii,k qk (t − iTB ), iTB ≤ t < (i + 1)TB , (3.12)
k=1
where h is referred to as the modulation index, and θ i is a memory term (to be described
below). The normalizing constant, C N , is set to
s
2
CN ≡ , (3.13)
N σI2
48
and
N
X
φ(iT + ) = K Ii,k Ab (k), (3.20)
k=1
Notice that θi depends on θi−1 ; the OFDM signal at the beginning of the ith block,
PN PN
k=1 Ii,k Ab (k); and the OFDM signal at the end of the (i−1)th block, k=1 Ii−1,k Ae (k).
The recursive relationship can be written as
∞ X
X N
θi = K [Ii−l,k Ab (k) − Ii−1−l,k Ae (k)] . (3.23)
l=0 k=1
Thus, the memory term is a function of all data symbols during and prior to the ith
block.
Figure 3.4 plots the phase discontinuities {c i } at the boundary times t = iTB , i =
0, 1, . . . , 49. In Figure 3.4(a), ci is plotted for memoryless modulation, that is, θ i = 0,
for all i; therefore, ci = K N
P
k=1 [Ii−1,k Ae (k) − Ii,k Ab (k)]. Figure 3.4(b) shows that the
phase discontinuities are eliminated with the use of memory as defined in (3.22).
1.5 1.5
1 1
Phase discontinuity, ci
Phase discontinuity, ci
0.5 0.5
0 0
−0.5 −0.5
−1 −1
−1.5 −1.5
0 10 20 30 40 50 0 10 20 30 40 50
Normalized time, t/TB Normalized time, t/TB
The benefit of continuous phase CE-OFDM is a more compact signal spectrum. This
property is studied further in Section 3.2. A second consequence of the memory terms
is the entire unit circle is used for the CE-OFDM phase modulation. This is illustrated
in Figure 3.5 which plots continuous phase CE-OFDM signal samples on the complex
50
Unit circle
Starting point
Figure 3.5: Continuous phase CE-OFDM signal samples, over L blocks, on the complex
plane. (2πh = 0.7)
plane. The modulation index is 2πh = 0.7. Figure 3.5(a) shows signal samples over
L = 1 block, where the phase signal occupies about one-half the unit circle. Viewing
samples over L = 100 blocks, Figure 3.5(b) shows that the phase signal occupies the
entire unit circle.
3.2 Spectrum
Since the Fourier transform approach isn’t computationally feasible, other techniques
are required to understand the CE-OFDM spectrum. The simplest is with the Taylor
51
P∞
expansion ex = n=0 x
n /n!. The CE-OFDM signal, with θi = 0, can be written as
where
N
XX
m(t) = CN Ii,k qk (t − iTB ) (3.25)
i k=1
is the normalized OFDM message signal. The effective double-sided bandwidth, defined
as the twice the highest frequency subcarrier, of m(t) is
N N
W =2× = . (3.26)
2TB TB
The bandwidth of s(t) is at least W : in (3.24), the n = 0 term contains no information
and thus has zero bandwidth; the n = 1 term is information bearing and has bandwidth
W ; the n = 2 term has a bandwidth 2W ; and so on. Thus, due to the n = 1 term, the
bandwidth of s(t) is at least W , and depending on the modulation index the effective
bandwidth can be greater than W .
The power density spectrum, Φs (f ), can be easily estimated by the Welch method
of periodogram averaging [526]. The result, Φ̂s (f ) ≈ Φs (f ), is used to calculate the
fractional out-of-band power,
Rf Rf
0 Φs (x)dx 0 Φ̂s (x)dx ˆ
FOBP(f ) = ≈ = FOBP(f ), (3.27)
0.5Ps 0.5Ps
R∞
where Ps = −∞ Φs (f )df = Es /TB = A2 is the signal power. Figure 3.6 shows estimated
fractional out-of-band power curves for N = 64 and various 2πh. Due to the normalizing
constant CN these curves are valid for any M . The dashed lines represent the RMS
bandwidth,
Brms = σφ W = 2πhN/TB . (3.28)
As defined in (3.28), the RMS bandwidth can be less than W , but, as shown by the
Taylor expansion in (3.24), the CE-OFDM bandwidth is at least W . A more suitable
52
100
ˆ
FOBP(f )
Brms
10−1
2πh
10−2
2.0
Fractional out-of-band power
1.8
1.6
10−3
1.4
1.2
10−4
1.0
0.8
10−5
0.6
0.4
10−6
0.2
bandwidth is thus
Bs = max(2πh, 1)W. (3.29)
Figure 3.7 plots Bs versus 2πh, and compares it with the 90–99% bandwidths as de-
termined by the Welch method. Notice that (3.29) is an accurate 90–92% bandwidth
measure for 2πh ≥ 1.0. For small modulation index, B s is a conservative bandwidth.
With 2πh = 0.4, for example, (3.29) accounts for 99.8% of the signal power (from Figure
3.6).
Figure 3.8 compares spectral estimates for CE-OFDM signals with the three sub-
carrier modulations from (3.7), (3.8) and (3.9). The modulation index is 2πh = 0.6.
Memoryless, non-continuous phase CE-OFDM is compared to continuous phase CE-
OFDM (the continuous phase examples are prefixed with “CP”). The estimates are also
53
2.8
Bs
Welch: 90%
2.6 92%
95%
99%
Normalized double-sided bandwidth, B/W 2.4
2.2
1.8
1.6
1.4
1.2
0.8
0.6
0.4
0.2
0
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Modulation index, 2πh
where 2
e−σφ σφ2n
an = , (3.31)
n!
and
δ(f ), n = 0,
Un (f ) = Φm (f ), n = 1, (3.32)
Φm (f ) ∗n Φm (f ),
n > 1.
P∞ n
The weighting factors {an } are Poisson distributed, and n=0 an = 1; ∗ denotes the
3
n-fold convolution, for example x(t) ∗ x(t) = x(t) ∗ x(t) ∗ x(t); and Φm (f ) is the power
54
Welch estimate
0 terms from (3.30)
Φ̂Ab (f )
−10
n=2
−20
n=3
Power spectrum (dB)
−30 n=1
n=4
DCT
DFT
−40
−50 DST
CP-DFT
−60
CP-DCT
−70
−80−3 −2 −1 0 1 2 3
Normalized frequency, f /W
where
1,
x = 0,
sinc(x) = (3.34)
sin πx ,
otherwise.
πx
R∞
The functions {Un (f )} have the property: −∞ Un (f )df = 1, for all n [1]. Therefore
the nth term in (3.30) has an an × 100% contribution to the overall spectrum. For
example, the carrier component, represented as δ(f ), has a fractional contribution of
2 2 2
e−σφ ; Φm (f ) ∗ Φm (f ) has a fractional contribution (e −σφ σφ4 )/2; and so on. Notice that
2
for 2πh = 0.2, the carrier component accounts for e −0.2 ×100 ≈ 96% of the signal power.
(This explains why the 90–92% curves at 2πh = 0.2 in Figure 3.7 are equal zero.)
55
Figure 3.8 plots the n = 1, 2, 3, 4 terms in (3.30), and the resulting sum
4
X
2
Φ̂Ab (f ) = A an Un (f ) ≈ ΦAb (f ). (3.35)
n=0
The Abramson spectrum is shown to match all estimates over the range |f /W | ≤ 1. For
|f /W | > 1, the spectral height depends on the overall smoothness of the phase signal. For
example, DST has a continuous phase [with or without memory since A b (k) = Ae (k) = 0,
for all k] and has a lower out-of-band power than memoryless DFT, which isn’t phase-
continuous. Memoryless DFT results in a slightly smoother phase than memoryless DCT
since one-half of the subcarriers have zero-crossings at the signal boundaries [A b (k) =
Ae (k) = 0, for k = N/2 + 1, . . . , N , and Ab (k) = Ae (k) = 1, otherwise] while DCT
doesn’t [Ab (k) = Ae (k) = 1, for all k]. The smoothest phase results from CP-DCT
which, unlike DST and CP-DFT, has a first derivative equal to zero at the boundary
times t = iTB . Consequently, the CP-DCT is the most spectrally contained.
Figure 3.9 shows estimated fractional out-of-band power curves that correspond to
the signals in Figure 3.8. For reference, conventional OFDM is also plotted. Notice that
the 99% spectral containment at f /W = 0.5 is the same for each signal. The continu-
ous phase CE-OFDM signals are the most spectrally contained and are shown to have
better than 99.99% containment at f /W = 1.25. Over the range 0.5 ≤ f /W ≤ 0.8, This
100
CE-OFDM
OFDM
10−1 Brms
10−2
Fractional out-of-band power
10−3
10−4
DCT
DFT
10−5
DST
10−6 CP-DFT
10−7
CP-DCT
figure shows that the CE-OFDM spectrum has more out-of-band power than conven-
tional OFDM. Since the modulation index controls the CE-OFDM spectral containment,
smaller h can be used if a tighter spectrum is required. The tradeoff is that smaller h
results in worse performance, as will be discussed in the next chapter. Therefore, the
system designer can trade performance for spectral containment, and visa versa.
Figure 3.10 compares CE-OFDM, with CP-DFT modulation over a large range of
modulation index, to conventional OFDM. For 2πh ≤ 0.4 the fractional out-of-band
power of CE-OFDM is always better than OFDM; otherwise CE-OFDM has more out-
of-band power for at least some frequencies f /W > 0.5. The 2πh = 2.0 example has a
broad spectrum, greater than OFDM over all frequencies. Notice that the shape of the
spectrum appears Gaussian shaped. This is due to the fact that for a large modulation
index, the higher-order terms in (3.32) dominate. They are Gaussian shaped due to the
multiple convolutions of (3.33). The shape of “wideband FM” signals is well covered in
the classical works of [1, 341, 437, 472].
100
CE-OFDM
OFDM
10−1
2πh
10−2
2.0
Fractional out-of-band power
1.8
1.6
10−3
1.4
1.2
10−4
1.0
0.8
10−5
0.6
0.4
10−6
0.2
Finally, Figure 3.11 compares CE-OFDM and OFDM with nonlinear power ampli-
fication. The OFDM curves (from Figure 2.9) require > 6 dB backoff to avoid spectral
broadening. The CE-OFDM signals have a bandwidth that depends only on the modu-
lation index and are not effected by the PA nonlinearity.
100
OFDM, TWTA
OFDM, Ideal
CE-OFDM
10−1
IBO (dB)
0
2
Fractional out-of-band power
4
6
10−2
10−3
10−4
2πh 0.4 0.5 0.6 0.7
Performance of Constant
Envelope OFDM in AWGN
In this chapter the basic performance properties of CE-OFDM are studied. The
baseband signal, represented by (3.11) and (3.12), is up-converted and transmitted as
the bandpass signal
n o
sbp (t) = < s(t)ej2πfc t = A cos [2πfc t + φ(t)] , (4.1)
where nw (t) denotes a sample function of the additive white Gaussian noise (AWGN)
process with power density spectrum Φ nw (f ) = N0 /2 W/Hz. The primary focus of the
chapter is to analyze the phase demodulator receiver, depicted by the block diagram
below. An expression for the bit error rate (BER) is derived by making certain high
carrier-to-noise ratio (CNR) approximations. The analytical result is then compared
against computer simulation and it is shown to be accurate for BER < 0.01. It is also
58
59
demonstrated that with the use of a phase unwrapper, the receiver is insensitive to phase
offsets caused by the channel and/or by the memory terms {θ i }.
cos(x − y) − cos(x + y)
sin(x) sin(y) = , (4.3)
2
sin(x + y) + sin(x − y)
sin(x) cos(y) = , (4.4)
2
cos(x + y) + cos(x − y)
cos(x) cos(y) = , (4.5)
2
sin(x + y) − sin(x − y)
cos(x) sin(y) = . (4.6)
2
1
This is the standard model used for representing received baseband signals, and more discussion of
the model can be found in [421, sec. 4.1], [624, sec. 5.5], among other places.
60
Lowpass
filter
2 cos(2πfc t)
where
nbp (t) = nc (t) cos(2πfc t) − ns (t) sin(2πfc t) (4.8)
is the result of passing nw (t) through the bandpass filter. The terms n c (t) and ns (t)
are referred to as the in-phase and quadrature components of the narrowband noise,
respectively, and have the power density spectrum
N0 , |f | ≤ Bbpf /2,
Φnc (f ) = Φns (f ) = (4.9)
0,
|f | > Bbpf /2,
where Bbpf is the bandwidth of the bandpass filter. Note that B bpf is assumed to be
sufficiently large so sbp (t) is passed through the front-end filter with negligible distortion
[421, pp. 157–158]. Writing sbp (t) in the form
where sc (t) = A cos[φ(t)] and ss (t) = A sin[φ(t)], the filter output can then be written as
u(t) = [sc (t) + nc (t)] cos(2πfc t) − [ss (t) + ns (t)] sin(2πfc t). (4.11)
61
where LP{·} denotes the lowpass component of its argument (i.e., double-frequency terms
are rejected) [624, p. 364]. Likewise, the output of the bottom (quadrature) branch is
where s(t) is the lowpass equivalent CE-OFDM signal from (3.11), and
is the lowpass equivalent representation of the bandpass white noise, n bp (t) [421, p. 158].
The power density spectrum of n(t) is [421, p. 158]
N0 , |f | ≤ Bn /2,
Φn (f ) = (4.16)
0,
|f | > Bn /2,
The continuous-time receive signal is then sampled at the rate f sa = 1/Tsa samp/s
to obtain the discrete-time signal 3
Phase demodulator
where s[i] = s(t)|t=iTsa and n[i] = n(t)|t=iTsa . As discussed in Section 2.4.2, the noise
samples {n[i]} are assumed independent:
σ 2 ,
i1 = i2 ,
n
E {n[i1 ]n[i2 ]} = (4.19)
0,
i1 6= i2 ;
The discrete-time phase demodulator studied in this thesis is shown in Figure 4.3.
The finite impulse response (FIR) filter is optional, but has been found effective at
improving performance; arg(·) simply calculates the arctangent of its argument; and the
phase unwrapper is used to minimize the effect of phase ambiguities. As will be shown,
the unwrapper makes the receiver insensitive to phase offsets caused by the channel
and/or by the memory terms.
The output of the phase demodulator is processed by the OFDM demodulator which
consists of the N correlators, one corresponding to each subcarrier. This correlator bank
is implemented in practice with the fast Fourier transform.
In this section a bit error rate approximation is derived for the phase demodulator
receiver. Although the receiver operates in the discrete-time domain, it is convenient to
analyze it in the continuous-time domain. The angle of the received signal is
N
X
arg[r(t)] = θi + 2πhCN Ii,k qk (t − iTB ) + ξ(t), (4.20)
k=1
for DCT and DFT modulations [(3.7), (3.9)], Ψ i,k = 0 and therefore has no effect on sys-
tem performance. This highlights an important observation: DST subcarrier modulation
(3.8) is inferior to DCT and DFT since Ψ i,k = 0 isn’t guaranteed.
The symbol error rate is computed by determining the probability of error for each
signal point in the M -PAM constellation. For the M − 2 inner points, the probability of
error is
Pinner = P (|Ni,k | > d) = 2P (Ni,k > d), (4.30)
where s
1
d = 2πh . (4.31)
2N σI2
[Notice that (4.30) is not averaged over i nor k since var{N i,k }, as approximated by (4.27),
is a constant.] Due to the Gaussian approximation applied to the random variable N i,k ,
Z ∞
1
exp −x2 / 2N0 /(2A2 TB ) dx
Pinner ≈ 2 p
d 2πN0 /(2A2 TB )
Z ∞
1
√ exp −x2 /2 dx
=2 (4.32)
d[N0 /(2A2 TB )]−0.5 2π
s ! r !
A2 TB 6 log 2 M Eb
= 2Q 2πh = 2Q 2πh .
N0 N σI2 M 2 − 1 N0
1
Pouter = P (Ni,k > d) = Pinner . (4.33)
2
Notice that for 2πh = 1, (4.34) is equivalent to the SER for conventional M -PAM [483,
pp. 194–195]. For high SNR, the only significant symbol errors are those that occur in
adjacent signal levels, in which case the bit error rate is approximated as [483, p. 195]
r !
SER M −1 6 log 2 M Eb
BER ≈ ≈2 Q 2πh . (4.35)
log2 M M log2 M M 2 − 1 N0
65
Suppose the channel imposes a phase offset of φ 0 . The received signal is then
iTB ≤ t < (i + 1)TB . Which is identical to (4.20) with the addition of the channel offset
term. The kth correlator is the same as (4.22), except the third term is
(i+1)TB
1
Z
Ψi,k = [θi + φ0 ]qk (t − iTB )dt = 0. (4.38)
TB iTB
Therefore, the phase offset due to the channel has no impact on performance, and the
analytical approximation in (4.35) is applicable.
Figure 4.4 compares the performance of N = 64, M = 2 CE-OFDM with phase offset
{(θi + φ0 ) ∈ [0, 2π)}, and without (θi + φ0 = 0). The former is referred to as System
1 (S1), the later as System 2 (S2). The system is computer simulated with a sampling
rate fsa = JN/TB , where J = 8 is the oversampling factor 4 . For Eb /N0 ≥ 10 dB and
2πh ≤ 0.5, S1 and S2 are shown to have identical performance. For these cases the
analytical approximation (4.35) closely matches the simulation results for BER < 0.01.
With the 2πh = 0.7 example, S1 is shown to have a 1 dB performance loss compared
to S2. In this case, the analytical approximation is shown to be overly optimistic. This
demonstrates a limitation of the phase demodulator receiver: for a large modulation
index and low signal-to-noise ratio, the phase demodulator has difficulty demodulating
the noisy samples. The performance of S1 is slightly worse than S2 since the output
of the phase demodulator, the arg(·) block in Figure 4.3, has more phase jumps since
the received phase crosses the π boundary more frequently. Proper phase unwrapping
is therefore required. However, phase unwrapping a noisy signal is a difficult problem
and the unwrapper makes mistakes. As a result the performance degrades slightly. For
a smaller modulation index, the unwrapper works perfectly and the performance of S1
isn’t degraded.
4
Also, the FIR filter (see Figure 4.3) has length Lfir = 11 and normalized cutoff frequency fcut /W =
0.2. See Section 4.1.4 for more on the filter design.
66
100
System 1
System 2
Approx (4.35)
10−1
10−4
10−5 0 5 10 15 20 25 30
Signal-to-noise ratio per bit, Eb /N0 (dB)
Figure 4.4: Performance with and without phase offsets. System 1 (S1) has phase offsets
{(θi + φ0 ) ∈ [0, 2π)}, and System 2 (S2) doesn’t (θ i + φ0 = 0). [M = 2, N = 64, J = 8]
The high-CNR approximation made in (4.26), which leads to the BER approximation
(4.35), is a standard technique for analyzing phase demodulator receivers [423, 624]. A
well-known characteristic of such receivers is: at low CNR, below a threshold value, the
approximation is invalid and system performance degrades drastically. In this section, the
CNR is defined and the threshold effect for CE-OFDM is observed by way of computer
simulation.
is the noise power. From (3.17), the carrier power can be written in the form
Eb N log 2 M
A2 = ; (4.41)
TB
67
thus
(Eb /N0 )N log2 M
CNR = . (4.42)
TB Bn
Since the noise samples are assumed independent [see (4.19)],
Simulation
Approx (4.35)
10−1
Simulation
Approx (4.35)
−2 0 2 4 6 8 10 10−310 11 12 13 14 15 16
Carrier-to-noise ratio (dB) Carrier-to-noise ratio (dB)
Simulation
2πh Approx (4.35)
10−1
0.2
0.4
Bit error rate
10−1
0.6 10−2
Simulation
Approx (4.35) 0.8 0.6 0.4 2πh 0.2
0.8
10−2
−2 0 2 4 6 8 10 10−310 12 14 16 18 20 22 24
Carrier-to-noise ratio (dB) Carrier-to-noise ratio (dB)
The FIR filter preceding the phase demodulator (see Figure 4.3) can improve per-
formance. Figure 4.7 shows BER simulation results of an M = 2, N = 64, J = 8,
2πh = 0.5 system. The SNR is held constant at E b /N0 = 10 dB. The filter, designed
using the window technique described in [422, pp. 623–630], has a length 3 ≤ L fir ≤ 101
and a normalized cutoff frequency 0 < f cut /W ≤ 1. Hamming windows are used5 . The
performance without a filter is shown to be BER = 0.05, while the analytical approxi-
mation (4.35) is BER = 0.012. For fcut /W ≥ 0.4 all the filtered results are shown to be
better than the unfiltered result. The filters with L fir > 5 and fcut /W > 0.5 are shown
to have roughly the same performance. The higher-order filters, which have a narrower
transition bands, require fcut /W > 0.5 to yield good performance. This is explained
by noting that the (single-sided) signal bandwidth is at least W/2 Hz. Therefore, the
higher-order filters with fcut /W < 0.5 distort the signal. Notice that the L fir = 11 filter
has equally good performance so long as f cut /W ≥ 0.1. This is due to the wide transition
band of the lower-order filter.
Lfir = 3
5
7
9
11
21
31
61
101
10−1
Bit error rate
No filter
Approx (4.35)
5
It has been observed that the window type has negligible impact on performance.
70
Lfir , fcut /W
−20
9, 0.7
3, 0.1
Magnitude response (dB)
−40
101, 0.7 9, 0.1
−60
31, 0.1
−80
−100
0 0.5 1 1.5 2 2.5 3
Normalized frequency, f /W
The figure above shows the magnitude response of the various Hamming FIR filters.
The filters with relatively flat response over |f /W | ≤ 0.5 result in good performance.
The Lfir = 31, fcut /W = 0.1 example is shown to not have this property, and, as shown
in Figure 4.7, has worse BER performance than the other filters.
Figure 4.9 compares the performance of binary (M = 2) CE-OFDM with and without
the FIR filter. The Lfir = 11, fcut /W = 0.2 filter is used. These results show that the
filter becomes important for larger modulation index: for 2πh = 0.1 the filtered and
unfiltered results are the same; for 2πh = 0.3 the filtered performance is a fraction of
a dB better than the unfiltered; for 2πh = 0.7 there is a 2 dB improvement in the
range 10−3 < BER < 10−5 . Notice the error floor developing below 10 −5 . This is a
consequence of imperfect phase demodulation. The filter lowers the error floor resulting
in a 9 dB improvement at BER = 10−6 .
71
100
Without FIR filter
With FIR filter
Approx (4.35)
10−1
10−2
Bit error rate
10−4
10−5
10−6 0 5 10 15 20 25 30
Signal-to-noise ratio per bit, Eb /N0 (dB)
During each block one of M N CE-OFDM signals is transmitted. Consider the mth
bandpass signal
N
" #
(m)
X
sm (t) = A cos 2πfc t + θ0 + K Ik qk (t) , 0 ≤ t < TB , (4.45)
k=1
N
where K = 2πhCN . The set of all possible signals, {s m (t)}M
m=1 , is determined by the
(m) (m) (m) N
set of all possible data symbol vectors {I (m) = [I1 , I2 , . . . , IN ]}M
m=1 . The optimum
72
receiver, as shown in Figure 4.10, correlates the received signal, r bp (t) = sm (t) + nw (t),
with each potentially transmitted signal. The detector then selects the largest result [421,
pp. 242–247].
R TB
0
(·)dt
s1 (t)
R TB
0 (·)dt
s2 (t)
Select
Received the Output
signal rbp (t) largest decision
.. ..
. .
R TB
0
(·)dt
sM N (t) Sample
at t = TB
It is desired to obtain an analytical expression for the bit error probability 6 , P (bit error).
However, there are two other probabilities to consider:
The first is the probability that the output of the optimum receiver is in error—that is,
the receiver selects a different signal than the one transmitted. The second is the data
symbol error probability. Determining exact expressions for the above probabilities is
intractable for large N . However, upperbounds and approximations can be derived in a
straightforward way, as described below.
6
The bit error probability is used interchangeably with the bit error rate. Likewise for the symbol
error probability and symbol error rate.
73
The above expression is the probability of detection error for M N signals with equal
correlation −1 ≤ λ ≤ 1. Therefore, it provides an upperbound given that
where Kd2 is the number of neighboring signal points having the minimum squared
min
Euclidean distance
d2min = min d2m,n , (4.50)
m,n; m6=n
where Z TB
d2m,n = [sm (t) − sn (t)]2 dt (4.51)
0
is the squared Euclidean distance between s m (t) and sn (t). This quantity is related to
the signal correlation as
d2m,n = 2Es (1 − ρm,n ), (4.52)
thus
d2min = 2Es (1 − ρmax ). (4.53)
Therefore to obtain the performance bound (4.46) and the approximation (4.49) the
signal correlation properties must be studied, and in particular ρ max must be determined.
The normalized correlation between the mth and nth signal, as a function of the phase
74
constant K = 2πhCN , is
TB
1
Z
ρm,n (K) = sm (t)sn (t)dt
Es 0
N
" #
TB
A2
Z
(m)
X
= cos 2πfc t + θ0 + K Ik qk (t) ×
Es 0 k=1
" N
# (4.54)
(n)
X
cos 2πfc t + θ0 + K Ik qk (t) dt
k=1
N
" #
TB
A2
Z X
= cos 2K ∆m,n (k)qk (t) dt,
2Es 0 k=1
(m) (n)
where ∆m,n (k) = 0.5[Ik − Ik ]. The double frequency term is ignored since f c 1/TB
is assumed. Notice that for k where ∆ m,n (k) = 0, the data symbols are the same, and
these indices don’t contribute to the correlation. Therefore
D
" #
A 2 TB
Z X
ρm,n (K) = cos 2K ∆m,n (kd )qk (t) dt, (4.55)
2Es 0
d=1
where {kd }D
d=1 are the indices where the data symbols differ, that is, ∆ m,n (kd ) 6= 0, and
D is the total number of differences. Writing (4.55) in exponential form yields
D
( " #)
A 2 TB
Z X
ρm,n (K) = < exp j2K ∆m,n (kd )qk (t) dt
2Es 0
d=1
(D ) (4.56)
A 2 TB
Z Y
= < exp [j2K∆m,n (kd )qk (t)] dt.
2Es 0
d=1
To proceed, the DCT modulation (3.7) is assumed. Making use of the Jacobi-Anger
expansion [580],
∞
X
ja cos b
e = Ji (a)eji(b+π/2) , (4.57)
i=−∞
where Ji (a) is the ith-order Bessel function of the first kind, (4.56) is written as
" ∞ ∞
A 2 TB
Z X X
ρm,n (K) = < ···
2Es 0
i1 =−∞ iD =−∞
#
Ji1 [2K∆m,n (k1 )] × · · · × JiD [2K∆m,n (kD )]ejσ(i) dt
(4.58)
TB ∞ ∞
A2
Z X X
= ··· Ji1 [2K∆m,n (k1 )]×
2Es 0 i1 =−∞ iD =−∞
πt PD π PD
where σ(i) = ω(i) + ψ(i), ω(i) ≡ TB d=1 id kd and ψ(i) ≡ 2 d=1 id . Index values that
result in ω(i) 6= 0 have no contribution, so (4.58) simplifies to
D
XY
ρi,j (K) = Ji0i,d [2K∆m,n (kd )] cos[ψ(i0i )], (4.59)
i d=1
where i0i ≡ [i0i,1 , . . . , i0i,D ], i = 1, 2, . . ., represent the vectors whereby ω(i 0i ) = 0. This
result is the same for DST modulation except ψ(i 0i ) = 0. For DFT modulation, (4.59) is
slightly different since both sinusoids and cosinusoids are used as subcarriers.
For D = 1,
ρm,n (K) = J0 [2K∆m,n (k1 )]. (4.60)
Therefore the correlation is simply the 0th-order Bessel function. Figure 4.11(a) plots
(4.60) for |∆m,n (k1 ) = 1|. Also plotted is the envelope of the 0th-order Bessel func-
tion [580, p. 121]. Note that ρm,n (K) doesn’t depend on the subcarrier frequency
fk1 = k1 /TB , k1 ∈ {1, 2, . . . , N }, just on the magnitude of the difference |∆ m,n (k1 )| ∈
{1, 2, . . . , (M − 1)}.
Figure 4.11(b) plots all unique ρm,n (K) for M = 2, N = 8 DCT subcarrier modulation.
Notice that the largest correlation function is associated with D = 1. For any given
signal, there are N other signals with D = 1: therefore, K d2 = N , and from (4.49),
min
A minimum distance signal error results in one data symbols error. Therefore, the symbol
error probability is approximated as
P (signal error) p
P (symbol error) ≈ ≈Q Es [1 − J0 (2K)]/N0 . (4.63)
N
For M = 2, one symbol error corresponds to one bit error. For M > 2, a symbol error
can result in 1 to log 2 M bit errors. Assuming each outcome is equally likely, a symbol
76
0.5
p
1/πK
ρm,n (K)
−0.5
0 1 2 3 4 5
K
(a) D = 1.
J0 (2K)
0.8
ρm,n (K)
0.6
1 Plog2 M
error results in log2 M i i = 0.5(log 2 M + 1) bit errors. Thus
0.5(log 2 M + 1)
P (bit error) ≈ P (symbol error)
log2 M
(4.64)
0.5(log 2 M + 1) p
≈ Q Es [1 − J0 (2K)]/N0 .
log2 M
The bit error probability is bounded by noting that P (bit error) ≤ P (signal error),
and using (4.46) with λ = ρmax = J0 (2K):
Z ∞h
1 N
i
P (bit error) ≤ √ 1 − [1 − Q(y)]M −1 ×
2π −∞
2
(4.65)
s
1 2Es [1 − J0 (2K)]
exp − y − dy.
2 N0
Figure 4.12 shows simulation results of the optimum receiver for M = 2 and N = 8.
The number of correlators at the receiver is therefore 2 8 = 256. Two values of modulation
index are plotted: 2πh = 0.3 and 2πh = 0.7 which corresponds to K = 0.15 and
K = 0.35. The upperbound (4.65) is shown to be within 3 dB of the simulated results
for high SNR. The analytical approximation (4.64) is shown to be very accurate.
100
10−1
10−2
Bit error rate
10−3
Approx (4.64)
Bound (4.65)
Simulation
10−4
0.7 2πh 0.3
10−5
10−6 0 3 6 9 12 15 18 21
Signal-to-noise ratio per bit, Eb /N0 (dB)
0.5
ρm,n (K)
p
1/πK
−0.5
0 1 2 3 4 5
K
Figure 4.14 shows simulation results for the phase demodulator receiver with N = 64
and for various modulation index values 2πh and modulation order M . The simulation
79
100
Approx (4.35)
Approx (4.64)
Simulation
10−1
10−2
10−5 5 10 15 20 25 30 35 40
Signal-to-noise ratio per bit, Eb /N0 (dB)
results are compared to the analytical approximation (4.35) and the optimum receiver
approximation (4.64). All curves are shown to be essentially identical for BER < 0.01.
This implies that the phase demodulator receiver is nearly optimum. For this to be
true, the phase demodulator must perfectly invert the phase modulation done at the
transmitter, and the noise at the output of the phase demodulator must be “white” and
Gaussian. That is, the OFDM demodulator is optimum given that the input, φ(t) + ξ(t),
is comprised of the transmitted message signal plus an AWGN corrupting signal. As
shown by (4.26), ξ(t) is approximately “white”. The probability density function of ξ(t)
samples is represented by the well-known form [421, p. 268]
Z ∞ 2
y + A2 − 2yA cos x
y
pξ (x) = exp − dy, (4.68)
0 2πσn2 2σn2
where σn2 = Bn N0 is the power of the noise signal n(t). Figure 4.15 compares (4.68) to
the Gaussian probability density function. The SNR per bit is E b /N0 = 30 dB. This
shows that ξ(t) is well approximated as Gaussian, and near optimum performance of the
phase demodulator receiver is expected.
80
pξ (x)
100 Gaussian
10−10
10−15
10−20
−1.5 −1 −0.5 0 0.5 1 1.5
x
Figure 4.15: Noise samples PDF versus Gaussian PDF. (E b /N0 = 30 dB)
It is first demonstrated that CE-OFDM with modulation index 2πh > 1 can out-
perform the underlying M -PAM subcarrier modulation. Figure 4.16 shows simulation
results7 for M = 2, 4, 8 and 16. The bit error rate is plotted against the SNR per bit on
the bottom x-axis and the carrier-to-noise ratio on the top x-axis. The viewable range
is such that CNR ≥ 5 dB. Notice that for M ≥ 4 and 2πh > 1, CE-OFDM outperforms
M -PAM. This is predicted by (4.35), since for 2πh = 1.0, the expression is equal to the
performance of M -PAM, and for 2πh > 1.0, it is better than M -PAM. For CE-OFDM
to operate in the region 2πh > 1, the carrier-to-noise ratio must be above threshold.
7
The oversampling factor is J = 8 for M = 2, 4 and 8, and J = 16 for M = 16. The FIR filter has
length Lfir = 11 and a normalized cutoff frequency 0.2 cycles per sample for M = 2, 4 and 16, and 0.3
cycles per sample for M = 8.
81
10−2 10−2
Bit error rate
10−4 2πh ∈ {0.5† , 0.4, . . . , 0.1‡ } 10−4 2πh ∈ {1.0† , 0.9, . . . , 0.1‡ }
10−5 16 18 20 22 24 26 28 30 10−5 15 20 25 30 35
Signal-to-noise ratio per bit, Eb /N0 (dB) Signal-to-noise ratio per bit, Eb /N0 (dB)
(a) M = 2. (b) M = 4.
10−2 10−2
Bit error rate
10−3 10−3
10−4 2πh ∈ {1.5† , 1.2, 1.0, 0.9, . . . , 0.1‡ } 10−4 2πh ∈ {2.0† , 1.5, 1.2, 1.1, . . . , 0.1‡ }
10−5 10 15 20 25 30 35 40 10−5 15 20 25 30 35 40 45
Signal-to-noise ratio per bit, Eb /N0 (dB) Signal-to-noise ratio per bit, Eb /N0 (dB)
To plot the spectral efficiency versus performance, the data rate must be defined,
which for uncoded CE-OFDM is
N log 2 M
R= b/s. (4.69)
TB
Using (3.29) as the effective signal bandwidth, the spectral efficiency is
R log2 M
R/B = = b/s/Hz. (4.70)
Bs max(2πh, 1)
Figure 4.17 shows result for M = 2, 4, 8 and 16. The target bit error rate is 0.0001. For
reference the channel capacity is also plotted, which is expressed as [421, p. 387]
C Eb
C = B log 2 1 + , (4.71)
B N0
or equivalently,
Eb 2C/B − 1
= . (4.72)
N0 C/B
10
M =2
M =4 M = 16: 2πh = 2.0, 1.8, . . . , 0.6
7 M =8
6 M = 16
Capacity M = 8: 2πh = 1.4, 1.2, . . . , 0.4
5
4
Spectral efficiency (b/s/Hz)
0.5-1.6 0 5 10 15 20 25
Performance: Eb /N0 (dB) to achieve 0.0001 bit error rate
There are two main observations to be made. First, for a fixed modulation index,
CE-OFDM has improved spectral efficiency with increase modulation order M at the cost
of performance degradation. For example consider 2πh = 0.4. The spectral efficiency
83
The second observation is that CE-OFDM can have both improvements in spectral
efficiency and in performance. Compare M = 2, 2πh = 0.5 with M = 4, 2πh = 1.0,
for example. The spectral efficiency doubles in the later case while also having a 2 dB
performance gain. Conventional CPM systems also have the property of increase spectral
efficiency and performance [14]. However, with CPM the receiver complexity increases
drastically with M (due to phase trellis decoding), which isn’t the case for CE-OFDM.
where SNRAWGN is the required signal-to-noise ratio required to achieve a target bit error
rate, SNRPA (IBO) is the required SNR when taking into account the nonlinear power
amplifier at a given backoff. Applying the PA model from Section 2.3 to CE-OFDM, the
input signal is
sin (t) = A exp[jφ(t)], (4.73)
where SNRsub is the required SNR for the underlying subcarrier modulation and SNR PM
is the required SNR for the phase modulated CE-OFDM system. By this definition,
the total degradation can be negative since, as observed in Figure 4.16, CE-OFDM can
outperform the underlying subcarrier modulation at the price of lower spectral efficiency.
84
50
(a) PA efficiency.
16
OFDM, TWTA
14 OFDM, SSPA
Total degradetion (dB)
OFDM, ideal
12 CE-OFDM: 2πh = 0.4
10 0.5
0.6
8
6
4
2
00 2 4 6 8 10
Input power backoff, IBO (dB)
2 OFDM, TWTA
OFDM, SSPA
99.5% bandwidth, f /W
1.2
0.8
0 2 4 6 8 10
Input power backoff, IBO (dB)
Performance of CE-OFDM in
Frequency-Nonselective Fading
Channels
where s(t) is the CE-OFDM signal according to (3.11), α and φ 0 is the channel amplitude
and phase, respectively, and n(t) is the complex Gaussian noise term represented in
R∞
(4.15). The received signal can be written as r(t) = −∞ h(τ )s(t − τ )dτ + n(t) [see (1.2),
(2.4)], where the channel impulse response is h(τ ) = αe jφ0 δ(τ ). In the frequency domain,
the channel is H(f ) = F{h(τ )}(f ) = αe jφ0 , and is thus constant at all frequencies—that
is, the channel is frequency nonselective.
In the previous chapter only the simple case of α = 1 (i.e. no fading) was considered.
In this chapter the channel amplitude is treated as a random quantity. Such a channel
model, since it’s frequency nonselective, is commonly referred to as flat fading. The
signal-to-noise ratio per bit for a given α is
Eb
γ = α2 , (5.2)
N0
86
87
−K
" s #
(1 + KR )e R (1 + KR )x KR (1 + KR )x
pγ (x) = exp − I0 2 , x ≥ 0, (5.4)
γ̄ γ̄ γ̄
where I0 (·) is the 0th-order modified Bessel function of the first kind, and
ρ2
KR = (5.5)
2σ02
is the Rice factor: ρ2 and 2σ02 represent the power of the LOS and scatter component,
respectively [401, p. 40]. For channels without a line-of-sight, ρ → 0 and γ is Rayleigh
distributed [483, p. 101]:
1 x
pγ (x) = exp − , x ≥ 0. (5.6)
γ̄ γ̄
To obtain BER(γ̄), the conditional BER is averaged over the distribution of γ [421, p.
817]: Z ∞
BER(γ̄) = BER(x)pγ (x)dx. (5.7)
0
If this were true, the bit error rate for the Ricean channel, described by (5.4), is [483, p.
102]
π/2
c1 (1 + KR ) sin2 θ
Z
BERRice (γ̄) = ×
π 0 (1 + KR ) sin2 θ + c22 γ̄/2
(5.10)
KR c22 γ̄/2
exp − dθ,
(1 + KR ) sin2 θ + c22 γ̄/2
88
However, as discussed in Section 4.1.3, the bit error rate of CE-OFDM, as a result of
the threshold effect, isn’t simply expressed by the Q-function for all values of SNR.
Consequently (5.10) and (5.11) are not generally accurate.
100 100
Simulation Simulation
Approx (5.10) Approx (5.11)
10−1 10−1
Bit error rate
10−2 10−2
10−4 10−4
10−5 0 5 10 15 20 25 30 10−5 0 10 20 30 40 50
Average signal-to-noise ratio per bit, γ̄ (dB) Average signal-to-noise ratio per bit, γ̄ (dB)
A Semi-Analytical Approach
The problem with (5.10) and (5.11) is that the conditional bit error rate, BER(x),
is not accurately described by the Q-function at low SNR and/or for large modulation
index. For a limited range of 2πh (for example, the values shown in Figure 4.16) the
following observation can be made: above a certain SNR, say x 0 , the conditional bit
error rate closely matches the Q-function, that is, (5.8) holds. Therefore (5.7) can be
approximated as
Z x0 Z ∞
BER(γ̄) = BER(x)pγ (x)dx + BER(x)pγ (x)dx
0
x0 Zx0∞ (5.12)
√
Z
≈ BER(x)pγ (x)dx + c1 Q(c2 x)pγ (x)dx.
0 x0
R x0
Determining x0 for a given M and 2πh, and dealing with 0 BER(x)pγ (x)dx in (5.12) are
the problems that remain to obtain an accurate approximation of BER(γ̄). As observed
in Section 4.1.3 [see Figure 4.6(a)], at low SNR the bit error rate is roughly 1/2. Assume
for the moment that BER(x) = 1/2 for x ≤ x 0 ; then
1 x0
Z ∞
√
Z
BER(γ̄) ≈ pγ (x)dx + c1 Q(c2 x)pγ (x)dx. (5.13)
2 0 x0
This simplified model, referred to as a two-region model since the conditional BER is split
into two regions, is illustrated in Figure 5.2: below x 0 the BER is 1/2, otherwise the BER
is equal to the Q-function. Also shown is the observed simulation result. Notice that
the two-region model doesn’t account for the transition region in which BER(x) ≈ 1/2
√
to where BER(x) ≈ c1 Q(c2 x). [For more examples of the transition region, see Figure
4.6.] Consequently, (5.13) is not generally accurate, and a more elaborate approach is
required which accounts for the transition region.
90
Transition region
1
0.5
Conditional bit error rate, BER(x)
0.1
Two-region model
Observed (simulation)
Q-function (4.35)
0.01
x0
Signal-to-noise ratio per bit, x (dB)
For SNR in the range 0 ≤ x ≤ γ1 the bit error rate is assumed to be BER 0 = 1/2. Figure
5.3 illustrates the n + 1 regions of (5.15). Notice that for n = 1, (5.15) is equivalent to
(5.13). In other words, (5.15), a (n+1)-region model, is a generalization of the two-region
model (5.13).
1
BER0 = 1/2
BER1
BER3
BER2
Conditional bit error rate, BER(x)
BER4
..
.
BERn−2
(n + 1)-region model
Observed (simulation)
Q-function (4.35)
BERn−1
0.01
for M = 4, 2πh ≤ 1.2; for M = 8, 2πh ≤ 1.8; and for M = 16, 2πh ≤ 2.4. The
results are shown in Figure 5.4: the circles represent Rayleigh results; the squares and
triangles represent the Ricean results for K R = 3 dB and KR = 10 dB, respectively.
The solid lines are the results of the semi-analytical approach, (5.15). The transition
region is sampled every 0.5 dB, that is, γ i+1 − γi = 0.5 dB, i = 1, 2, . . . , n − 1; the
starting point is γ1 = −5 dB. Therefore γi = 0.5(i − 1) − 5 dB, i = 1, 2, . . . , n. The
sampling continues until BERn < 0.01. For SNR x ≥ γn the conditional bit error rate
is approximated with the Q-function (5.8). This criteria used for γ n is based on the
observation that, for the modulation index values under consideration, the Q-function is
accurate for BER < 0.01. As shown in the figure, this semi-analytical approach yields
curves for BER(γ̄) that closely match simulation.
Figure 5.5 shows the improvement of (5.15) over (5.10) and (5.11). The semi-
analytical approach closely matches the simulation results, even at low SNR, while (5.10)
and (5.11) are overly optimistic by several dB.
10−1 10−1
Bit error rate
10−3 10−3
10−4 10−4
10−5 0 10 20 30 40 50 10−5 0 10 20 30 40 50
Average SNR per bit, γ̄ (dB) Average SNR per bit, γ̄ (dB)
10−1 10−1
Bit error rate
10−3 10−3
10−4 10−4
10−5 0 10 20 30 40 50 10−5 0 10 20 30 40 50
Average SNR per bit, γ̄ (dB) Average SNR per bit, γ̄ (dB)
10−1 10−1
Bit error rate
10−2 10−2
10−3 10−3
10−4 10−4
10−5 0 10 20 30 40 50 10−5 0 10 20 30 40 50
Average SNR per bit, γ̄ (dB) Average SNR per bit, γ̄ (dB)
10−1 10−1
Bit error rate
10−2 10−2
10−3 10−3
10−4 10−4
10−5 0 10 20 30 40 50 10−5 0 10 20 30 40 50
Average SNR per bit, γ̄ (dB) Average SNR per bit, γ̄ (dB)
100
Rayleigh simulation
Rayleigh approximation (5.11)
Ricean (KR = 3 dB) simulation
10−1 Ricean (KR = 3 dB) approximation (5.10)
Ricean (KR = 10 dB) simulation
Ricean (KR = 10 dB) approximation (5.10)
Bit error rate Semi-analytical technique (5.15)
10−2
10−3
10−4
10−5 0 10 20 30 40 50
Average signal-to-noise ratio per bit, γ̄ (dB)
Figure 5.5: Comparison of semi-analytical technique (5.15) with (5.10) and (5.11).
(M = 4, N = 64, 2πh = 1.2)
simulated Rayleigh result in Figure 5.5 requires about 6 hours of computer time (on a
workstation with 1 gigabytes of memory and a single 3 gigahertz microprocessor). The
semi-analytical result, on the other hand, requires less than 7 s (to obtain {BER i }, and
perform numerical integration): a speed improvement of 4 orders of magnitude.
The disadvantage, however, is that this technique doesn’t yield a closed-form expres-
sion. As of the time of this writing, such a solution, that is general and accurate, doesn’t
seem possible.
Chapter 6
Performance of CE-OFDM in
Frequency-Selective Channels
CE-OFDM has the same block structure as conventional OFDM, with a block period,
TB , designed to be much longer than τmax . A guard interval of duration Tg ≥ τmax is
inserted between successive CE-OFDM blocks to avoid interblock interference. At the
receiver, r(t) is sampled at the rate f sa = 1/Tsa samp/s, the guard time samples are
discarded and the block time samples are processed. Using the discrete-time model
outlined in Section 2.1.2, the processed samples are
NX
c −1
94
95
where {H[k]} is the DFT of {h[i]} and {S[k]} is the DFT of {s[i]}. The effect of the
channel can be reversed with the frequency-domain equalizer: a DFT followed by a
multiplier bank, followed by an IDFT. The FDE output is
NDFT
X−1
1
ŝ[i] = Rp [k]C[k]ej2πik/NDFT , i = 0, . . . , NB − 1, (6.4)
NDFT
k=0
where {Rp [k]} is the DFT of the processed samples and {C[k]} are the equalizer correc-
tion terms, which are computed as [463]
1
C[k] = (6.5)
H[k]
H ∗ [k]
C[k] = (6.6)
|H[k]|2 + (Eb /N0 )−1
Ignoring noise (n[i] = 0), the output of the frequency-domain equalizer using (6.5) is
NDFT
X−1
1
ŝ[i] = H[k]S[k]C[k]ej2πik/NDFT
NDFT
k=0
NDFT
X−1
1 1 j2πik/NDFT
= H[k]S[k] e
NDFT H[k] (6.7)
k=0
NDFT
X−1
1
= S[k]ej2πik/NDFT
NDFT
k=0
= s[i], i = 0, . . . , NB − 1.
Therefore, the ZF frequency-domain equalizer perfectly reverses the effect of the channel.
When noise can’t be ignored, the ZF suffers from noise enhancement. For example, a
fade of −30 dB results in a correction term with gain +30 dB, which corrects the channel
but amplifies the noise by a factor of 1000. The MMSE criterion (6.6) takes into account
96
the signal-to-noise ratio, making an optimum trade between channel inversion and noise
enhancement. Notice that the MMSE and ZF are equivalent at high SNR:
H ∗ [k] 1
lim C[k]|MMSE = 2
= = C[k]|ZF . (6.8)
Eb /N0 →∞ |H[k]| H[k]
n(t)
The study is separated into two parts. In Section 6.1, the performance of the MMSE
and ZF equalizers are compared over various frequency-selective channels. In Section 6.2,
performance is evaluated for frequency-selective fading channels, in which case {h[i]} is
described statistically. In both sections an N = 64 CE-OFDM system is considered,
with a block period of TB = 128 µs. The subcarrier spacing is 1/T B = 7812.5 Hz and
the mainlobe bandwidth is W = N/TB = 500 kHz. The guard period is Tg = 10 µs,
resulting in a transmission efficiency η t = 128/138 ≈ 0.93. The simulation uses an
oversampling factor J = 8; therefore the sampling rate is f sa = JN/TB = 4 Msamp/s,
and the sampling period is Tsa = 1/fsa = 0.25 µs.
In this section, the performance of CE-OFDM using the MMSE and ZF frequency-
domain equalizers is compared over six frequency-selective channels.
The channel samples {h[i]}, over the corresponding guard interval [0, 10 µs], are
shown in Table 6.1. For Channels A–C the maximum propagation delay is τ max = 0.75
97
|h[i]|2 = 1. (6.9)
i=0
Figure 6.2 shows Channel D in the time and frequency domains. In subfigure (a),
|h[i]|2 , that is, the power of the time samples, is plotted. In subfigure (b), |H(f 0 )|2 is
plotted, where [422, p. 256]
NX
c −1
0 0
H(f ) = h[i]e−j2πf i , (6.10)
i=0
where f 0 is the normalized frequency variable having units cycles/samp [422, p. 16].
Notice that over the signal’s mainlobe frequency range, −250 kHz ≤ f ≤ 250 kHz, the
channel is frequency selective. The magnitude response fluctuates over a 8.5 dB range,
−2.5 dB ≤ |H(f 0 )|2 ≤ 6 dB.
as
H[k] = H(fk0 ), k = 0, 1, . . . , NDFT − 1, (6.13)
1
Stochastic models are discussed in the next section.
99
0.4
0.35
0.3
0.25
|h[i]|2
0.2
0.15
0.1
0.05
0 0 1 2 3 4 5 6 7 8 9
Propagation delay, iTsa (µs)
−5
Using a DFT size NDFT = JN = NB and noting (6.11), the frequency samples {H[k]}
correspond to the frequencies
k, k = 0, 1, . . . , NDFT
2 ,
0 0 JN TB
fk = fk fsa = fk = (6.15)
TB k − fsa ,
k= NDFT
+ 1, . . . , . . . , NDFT − 1.
TB 2
Included in Figure 6.2(b) is the response of the MMSE and ZF equalizers. The ZF
response, (6.5), is simply the inverse of the channel. The MMSE response, (6.6), is shown
for Eb /N0 = 0, 10, and 20 dB. Notice that at high SNR the MMSE approaches the ZF
equalizer, which is to be expected from (6.8). For this particular channel the MMSE and
ZF are shown to be equivalent for Eb /N0 ≥ 20 dB.
The results for Channel A are shown in Figure 6.3. Of the six test channels, Channel
A is the most mild in terms of its frequency-domain response. The magnitude response
|H(f 0 )|2 spans a 3 dB region in a nearly linearly manner. The equalizers are shown to
effectively correct the channel: the BER curves in Figure 6.3(c) are nearly indistinguish-
able from the simple AWGN curves. Results are plotted for 2πh = 0.1, 0.3 and 0.6. For
the 2πh = 0.6 example at the lower SNR values E b /N0 < 10 dB, the ZF result is shown
to be slightly worse than the MMSE result; for higher values of SNR the performance of
the two equalizers becomes nearly identical. This is to be expected since, as illustrated
in Figure 6.3(b), their frequency response become the same at high E b /N0 .
Results for 2πh = 0.1, 0.2, 0.4 and 0.6 over Channel B are shown in Figure 6.4.
The frequency response of this channel is more severely varying than Channel A. Over
the signal bandwidth, |H(f 0 )|2 spans a 6 dB range. As with the previous example, the
MMSE is shown to slightly outperform the ZF at low SNR (i.e., the 2πh = 0.6 example
for Eb /N0 < 10 dB), but the two equalizers have essentially the same performance at the
101
0.7
2
0.6
0.4
|h[i]|2
−2
0.3
Channel A
0.2 −4 ZF
MMSE: Eb /N0 = 0 dB
10 dB
0.1 20 dB
−6
0 −200 −100
0 1 2 3 4 5 6 7 8 9 0 100 200
Propagation delay, iTsa (µs) Frequency, f = f 0 fsa (kHz)
10−1
ZF
MMSE
AWGN sim
AWGN approx (4.35)
10−2
Bit error rate
10−4 5 10 15 20 25 30
Signal-to-noise ratio per bit, Eb /N0 (dB)
0.9
4
0.8
0.7
0.4
−2
Channel B
0.3 ZF
MMSE: Eb /N0 = 0 dB
−4 10 dB
0.2
20 dB
30 dB
0.1
−6
0 −200 −100
0 1 2 3 4 5 6 7 8 9 0 100 200
Propagation delay, iTsa (µs) Frequency, f = f 0 fsa (kHz)
10−1
ZF
MMSE
AWGN sim
AWGN approx (4.35)
10−2
Bit error rate
10−4 5 10 15 20 25 30
Signal-to-noise ratio per bit, Eb /N0 (dB)
higher SNR values. For BER ≤ 0.001 the degradation caused by the frequency selective
channel, when compared to the simple AWGN result, is slightly less than 1 dB.
Channel C has the most frequency-selective response of the three maritime channel
realizations. As shown in Figure 6.5(b), the magnitude response varies over a 20 dB
range. It is also shown that very high SNR is required for the MMSE response to
approach the ZF response. Over the frequency range −250 kHz ≤ f 0 fsa ≤ −200 kHz,
for example, the two are equivalent only for E b /N0 > 35 dB. This equivalence is also
demonstrated in Figure 6.5(c): for the 2πh = 0.1 example, the ZF performance gradually
approaches the MMSE performance at these high SNR values. Clearly, the large amount
of frequency selectivity of this channel results in a large performance degradation when
compared to the AWGN results. At the bit error rate 0.001, the degradation is 10 dB
for the 2πh = 0.1 case. The improvement of the MMSE is pronounced for 2πh = 0.5.
At the bit error rate 0.001, the MMSE outperforms the ZF by 7 dB, and is only 2 dB
worse than the performance over the simple AWGN channel.
Figures 6.6–6.8 show the results for Channels D–F. As stated earlier, the three chan-
nels are three different realizations of a stochastic model with an exponential delay power
density spectrum. The degree that the each channel varies over the signal bandwidth
progresses from Channel D to Channel F. Channel F, having a 50 dB attenuation at
185 kHz, is the most harsh of the test channels. The results in Figure 6.8(c) show the
dramatic performance degradation as a consequence of the severe frequency selectivity.
An 18 dB loss, compared to the AWGN performance, is experienced for the 2πh = 0.6,
MMSE example at the bit error rate 0.001; the ZF case degrades more than 20 dB further.
A 40 dB loss is suffered for the 2πh = 0.1 and 0.3 cases. These results show that fre-
quency selective channels having deep fades in the signal bandwidth impact performance
greatly.
At this point, several observations can be made. First, the performance of the
equalized CE-OFDM systems studied depends on the amount of frequency selectivity
over the signal bandwidth. For channels with a relatively mild frequency response—
Channels A, B and D, for example—the performance degradation is minor. The noise
enhancement that results from equalizing channels with severe frequency responses—
104
0.6
Channel C
ZF
20 MMSE: Eb /N0 = 0 dB
0.5 10 dB
20 dB
0.3 0
0.2 −10
0.1
−20
0 −200 −100
0 1 2 3 4 5 6 7 8 9 0 100 200
Propagation delay, iTsa (µs) Frequency, f = f 0 fsa (kHz)
10−1
ZF
MMSE
AWGN sim
AWGN approx (4.35)
10−2
Bit error rate
10−4 5 10 15 20 25 30 35 40
Signal-to-noise ratio per bit, Eb /N0 (dB)
0.4 6
0.35 4
0.2 10 dB
−2 20 dB
0.15
−4
0.1
0.05 −6
0 −8 −200 −100
0 1 2 3 4 5 6 7 8 9 0 100 200
Propagation delay, iTsa (µs) Frequency, f = f 0 fsa (kHz)
10−1
ZF
MMSE
AWGN sim
AWGN approx (4.35)
10−2
Bit error rate
10−4 5 10 15 20 25 30 35
Signal-to-noise ratio per bit, Eb /N0 (dB)
0.35 10
Channel E
ZF
0.3 MMSE: Eb /N0 = 0 dB
10 dB
5
0.2
|h[i]|2
0.15
0.1 −5
0.05
−10
0 −200 −100
0 1 2 3 4 5 6 7 8 9 0 100 200
Propagation delay, iTsa (µs) Frequency, f = f 0 fsa (kHz)
10−1
ZF
MMSE
AWGN sim
AWGN approx (4.35)
10−2
Bit error rate
10−4 5 10 15 20 25 30 35 40
Signal-to-noise ratio per bit, Eb /N0 (dB)
0.4
40 Channel F
0.35 ZF
MMSE: Eb /N0 = 0 dB
0.2 0
0.15
−20
0.1
0.05 −40
0 −200 −100
0 1 2 3 4 5 6 7 8 9 0 100 200
Propagation delay, iTsa (µs) Frequency, f = f 0 fsa (kHz)
10−1
ZF
MMSE
AWGN sim
AWGN approx (4.35)
10−2
Bit error rate
10−4 5 10 15 20 25 30 35 40 45 50 55 60 65 70
Signal-to-noise ratio per bit, Eb /N0 (dB)
Last, the MMSE equalizer is more complicated than the ZF equalizer since the SNR
per bit, Eb /N0 , must be estimated at the receiver. The results of this study show that
this added complexity doesn’t always translate into improved performance. That is, the
ZF performance is the same as the MMSE performance for many cases—the 2πh ≤ 0.4
cases in Channel B, for example. In other cases, the MMSE performs much better, and
thus estimating Eb /N0 pays substantial dividends—the 2πh = 0.5 case for Channel C
illustrates this point.
As demonstrated in the following section, the MMSE equalizer offers significant im-
provement over the ZF equalizer when averaging performance over many channel real-
izations of a stochastic channel model.
In contrast to the test channels used in the previous section, which were deterministic
as defined in Table 6.1, the channels used in this section are described statistically.
The mathematical foundation for stochastic time-variant linear channels was pioneered
by Bello [50]; more recently Pätzold’s text, Mobile Fading Channels [401], provides a
excellent treatment of the topic, with a focus on the various aspects of simulation. In the
study here, the widely used assumption of WSSUS (wide-sense stationary uncorrelated
scattering) is applied. Also, it is assumed that the channel is composed of discrete paths,
each having an associated gain and discrete propagation delay. This assumption is based
on the Parsons and Bajwa ellipse model for describing multipath channel geometry [401,
p. 244]. The channel’s impulse response is
L−1
X
h(τ ) = al δ(τ − τl ), (6.16)
l=0
109
where al is the complex channel gain and τl is discrete propagation delay of the lth path;
the total number of paths is represented by L. The propagation delay differences are
That is, they are set equal to the sampling period of the simulation [401, p. 269]. The
delay of the 0th path is defined as τ0 ≡ 0, thus
τl = lTsa , l = 0, 1, . . . , L − 1. (6.18)
σa2l = E |al |2 ,
l = 0, 1, . . . , L − 1. (6.19)
Both the real and imaginary parts of the path gains are Gaussian distributed [401, p.
267]; thus the envelope |al |2 is Rayleigh distributed. Also, the channels are normalized
such that
L−1
X
σa2l = 1. (6.20)
l=0
As outlined in Pätzold’s text (pp. 276–279) the parameters σ a2l , τl and L determine
the fundamental characteristic functions and quantities of the channel models, such as the
delay power spectral density and the delay spread 2 . The relevant formulas are expressed
below.
• Average delay:
L−1
X
Bτ(1)
τ = σa2l τl . (6.22)
l=0
2
The phrase “delay power spectral density” is also commonly referred to as “power delay profile”
(PDP) or “multipath intensity profile” (MIP). For the sake of being consistent with [401], “delay power
spectral density” is used here. In Pätzold’s text, a clear distinction is made between stochastic channel
models, which provide the theoretical and mathematical foundations, and “deterministic” channel models
which are generated in software or hardware for simulation purposes. For the sake of simplicity, this
distinction isn’t stressed here (which results in a slightly different notation for the expressed formulas
in his text). Also, since only time-invariant channels are considered in this thesis, the Doppler power
spectral density, time correlation function and coherence time (see [401, pp. 277–279]) are not discussed.
110
• Delay spread: v
uL−1
(1) 2
uX
Bτ(2)
τ =t (σal τl )2 − Bτ τ (6.23)
l=0
where
35
X
C Cf = 1 exp(−τl /2e-6) = 0.1188 . . . (6.27)
l=0
is the normalizing constant used to guarantee (6.20). Note that the maximum propaga-
tion delay is 8.75 µs.
3
To avoid notational ambiguities, the channel model labels in this section have the subscript “f”
(“fading”).
111
The last model, Channel Df , has a uniform delay power density spectrum:
CD , 0 ≤ τl ≤ 8.75 µs,
2 f
σal ,D = (6.28)
0,
otherwise,
In Figure 6.9 the delay power density spectrum (6.21) and the frequency correlation
function (6.24) are plotted for each of the four models. The corresponding average
delay (6.22), delay spread (6.23) and coherence bandwidth (6.25) for each model is
labeled. Notice that Channel Df has the smallest coherence bandwidth, B C = 67 kHz.
For Channel Af the coherence bandwidth isn’t finite since, as shown in subfigure (b),
|rτ τ (v 0 )| > −3 dB for all frequency separation values 4 .
The average performance of various CE-OFDM systems is evaluated over the four
stochastic channel models. This is done by randomly generating {a l }—which, as stated
above, are complex-valued quantities, drawn from the Gaussian distribution, with zero
mean and variance {σa2l }—computing the received samples (6.2), then processing the
samples with the frequency-domain equalizer and the CE-OFDM demodulator. At each
average Eb /N0 considered, the simulation runs for at least 20,000 bit errors, or until
100,000,000 bits are transmitted, whichever happens first. This corresponds to many
thousands of channel realizations5 . Some channel realizations result in very poor per-
formance (for example, see Figure 6.8), while others result in a bit error rates not much
worse than that of the simple AWGN channel. This performance difference is attributed
to the severity of the channel’s frequency response, as observed with the several examples
in Section 6.1.
The performance also depends on the gain of the channel realization. Due to (6.20)
the channel gain, on average, is normalized to unity; however, for a given trial, the
channel may be fading such that the gain is less than unity, resulting in degraded per-
formance. The likelihood of a deep channel fade depends on the number of independent
4
˛ 10 1 9 1
For Channel Af , min |rτ τ (v 0 )| = min ˛ 11 exp(−j2πv 0 5 µs)˛ =
˛
+ 11 11
> 2
≈ −3 dB.
5
Example simulation code can be found in Appendix C.
113
(a) Delay power spectral density, Channel Af (b) Frequency correlation function, Channel Af
0 0
(1)
−5 Bτ τ = 0.45 µs −3
10 log 10 [rτ τ (v 0 )]
10 log10 [Sτ τ (τ )]
(2) BC → ∞
Bτ τ = 1.44 µs
−10
−6
−15
−9
−20
−25 −12
−30 −15
−450 −300 −150
0 1 2 3 4 5 6 7 8 9 10 0 150 300 450
Propagation delay, τ (µs) Frequency separation, v 0 (kHz)
(c) Delay power spectral density, Channel Bf (d) Frequency correlation function, Channel Bf
0 0
(1)
−5 Bτ τ = 1.67 µs −3
10 log 10 [rτ τ (v 0 )]
10 log10 [Sτ τ (τ )]
(2) BC
Bτ τ = 2.36 µs
−10
−6
−15
−9
−20
−25 −12
−30 −15
−450 −300 −150
0 1 2 3 4 5 6 7 8 9 10 0 74 150 300 450
Propagation delay, τ (µs) Frequency separation, v 0 (kHz)
(e) Delay power spectral density, Channel Cf (f) Frequency correlation function, Channel Cf
0 0
(1)
−5 Bτ τ = 1.78 µs −3
10 log 10 [rτ τ (v 0 )]
10 log10 [Sτ τ (τ )]
(2) BC
Bτ τ = 1.75 µs
−10
−6
−15
−9
−20
−25 −12
−30 −15
−450 −300 −150
0 1 2 3 4 5 6 7 8 9 10 0 140 300 450
Propagation delay, τ (µs) Frequency separation, v 0 (kHz)
(g) Delay power spectral density, Channel Df (h) Frequency correlation function, Channel Df
0 0
(1)
−5 Bτ τ = 4.38 µs −3
10 log 10 [rτ τ (v 0 )]
10 log10 [Sτ τ (τ )]
(2) BC
Bτ τ = 2.60 µs
−10
−6
−15
−9
−20
−25 −12
−30 −15
0 1 2 3 4 5 6 7 8 9 10 −450 −300 −150 0 67 150 300 450
Propagation delay, τ (µs) Frequency separation, v 0 (kHz)
propagation paths [the WSSUS assumption makes each path in (6.16) independent]. It is
unlikely that multiple paths fade simultaneously. For this reason, channels characterized
by multiple propagation paths possess a type of diversity known at multipath diversity—
which can be exploited by the receiver. Of the four models considered in this study,
Channel Df can be said to have the most multipath diversity: the gain of a given realiza-
tion depends on 36 independent paths, each having, on average, an equal contribution.
Channel Af can be said to have the least amount of multipath diversity: over 90% of the
channel gain depends on a single path. Channel B f has more multipath diversity than
Channel Af since the gain is distributed more equally between the two paths. That is,
the multipath diversity depends not only on the number of independent paths but also
on the way in which the power is distributed over the paths, as determined by {σ a2l }. [It
is worth noting that the frequency-nonselective channel models considered in Chapter 5
have L = 1 path of which 100% of the channel gain depends (σ a21 = 1), and thus these
channels have no multipath diversity.] In the results that follow, the impact of multipath
diversity—and its frequency-domain dual frequency diversity—on CE-OFDM systems is
studied.
The simulation results of this study are presented over three figures: Figure 6.10
compares the performance of a CE-OFDM system, with fixed modulation order M and
modulation index h, over the four channel models; Figure 6.11 compares the performance
of a CE-OFDM system with fixed M but varying h over Channel C f ; and Figure 6.12
compares the performance of constant envelope and conventional OFDM systems, in the
presence of power amplifier nonlinearities, over Channel C f . For each case, the number
of subcarriers is N = 64.
MMSE: Channel Af
Bf
10−1 Cf
Df
ZF: Channel Af
Bf
Cf
Df
Rayleigh, L = 1
AWGN
AWGN approx (4.35)
10−2
Bit error rate
10−3
10−4 5 10 15 20 25 30 35 40
Average signal-to-noise ratio per bit, Eb /N0 (dB)
Figure 6.10: Performance results. (Multipath results are labeled with circle and triangle
points; the Rayleigh, L = 1 result is that of the frequency-nonselective channel model.
M = 4, N = 64, 2πh = 1.0)
that is to be had by using the MMSE equalizer. At the bit error rate 0.001, for example,
MMSE outperforms ZF by 10 dB for Channel D f . These results also show the impact of
multipath diversity. Consider the MMSE results. For E b /N0 > 15 dB, the performance
over Channels Af –Df is better than the performance over the frequency-nonselective
Rayleigh (L = 1 path) channel. For BER ≤ 0.001, the performance over the multipath
channels is at least 5 dB better than the performance over the single path channel.
Notice that Channel Df , which has the most multipath diversity, results in a better
performance that all the other channels. The performance over Channel B f , which has
more multipath diversity than Channel A f , is in fact better than the performance over
Channel Af . These results indicate that the CE-OFDM receiver exploits the multipath
diversity of the channel.
The fact that constant envelope OFDM exploits multipath diversity is an interesting
result since conventional OFDM doesn’t. This was shown in Section 2.1.1; specifically,
116
by (2.9). So long as the duration of the guard interval is greater than or equal to
the channel’s maximum propagation delay, that is, T g ≥ τmax , and a cyclic prefix is
transmitted during the guard interval, the performance of OFDM in a time-dispersive
channel is equivalent to flat fading performance. In other words, the multipath fading
performance is the same as single path fading performance. In the context of Section
2.1.1, this property was considered beneficial since ISI is avoided. In the context here,
however, this property is considered a weakness since the multipath diversity of the
channel isn’t leveraged6 .
CE-OFDM, in contrast, has the ability to exploit the frequency diversity of the
channel since the phase modulator, in effect, spreads the data symbol energy in the
frequency domain. This can be seen by viewing the CE-OFDM waveform by the Taylor
series expansion [see Section 3.2, (3.24)]:
" #
σφ2 2
σφ3 3
s(t) = A 1 + jσφ m(t) − m (t) − j m (t) + . . . , (6.30)
2 6
0 ≤ t < TB , where A is the signal amplitude, σφ2 = (2πh)2 is the phase signal variance,
and m(t) = CN N
p
6/N (M 2 − 1), is the normalized
P
k=1 Ik qk (t), 0 ≤ t < TB , CN =
OFDM message signal. The higher-order terms m n (t), n ≥ 2, results in a frequency
spreading of the data symbols. This property is best demonstrated by way of a simple
example.
Example 6.2.1
Consider a CE-OFDM waveform with an OFDM message signal composed of N = 2 orthogonal
6
Note that OFDM systems typically employ channel coding and frequency-domain interleaving, which
offers diversity. However, since this thesis only deals with uncoded systems, these topics are beyond its
scope—and are topics for further research.
117
where Ik ∈ {±1}, k = 1, 2. Assume that the modulation index, h, is such that the higher-order
terms m2 (t) and m3 (t) contribute to the make up of s(t) according to (6.30). It is desired to
write m2 (t) and m3 (t) in terms of I1 , I2 and {cos 2πkt/TB }. This task requires some algebra,
but is simply done. For notational simplicity, let’s define
The expansions above are represented in Table 6.3. The data symbol contribution at each
tone cos 2πkt/TB , k = 0, 1, . . . , 6, for m(t), m2 (t) and m3 (t) is shown. Referring to the tones
as frequency bins, it can be said that for m(t) the two data symbols are simply contained in
the k = 1 and k = 2 frequency bins. For the second-order term, m2 (t), the data symbols mix
across the k = 0, 1, 2, 3, and 4 frequency bins. For m3 (t), the data symbols mix across the
k = 0, 1, . . . , 6 frequency bins.
The simple example above shows how the data symbols spread across multiple fre-
quency bins. In general, it can be said that the N data symbols that constitute the
constant envelope OFDM signal are not simply confined to N frequency bins—as is the
case with conventional OFDM. The phase modulator mixes and spreads—albeit in a
nonlinear and exceedingly complicated manner—the data symbols in frequency, which
gives the CE-OFDM system the potential to exploit the frequency diversity in the chan-
nel. This isn’t necessarily the case, however. For small values of modulation index,
118
Table 6.3: Data symbol contribution per tone for m n (t), n =1, 2, and 3.
kth tone, cos 2πkt/TB
0 1 2 3 4 5 6
m(t) – I1 I2 – – – –
0.5I12 ,
m2 (t) I1 I2 0.5I12 I1 I2 0.5I22 – –
0.5I22
0.75I13 , 1.25I12 I2 , 0.25I13 ,
m3 (t) 0.75I12 I2 0.75I12 I2 0.75I1 I22 0.25I23
1.5I1 I22 0.5I23 0.75I1 I22
where only the first two terms in (6.30) contribute, that is,
the CE-OFDM signal doesn’t have the frequency spreading given by the higher-order
terms. In this case, the CE-OFDM signal is essentially equivalent to a conventional
OFDM signal, jσφ m(t), (plus a relatively large DC term, A) and therefore doesn’t have
the ability to exploit the frequency diversity of the channel. Simply put, CE-OFDM
has frequency diversity when the modulation index is large and doesn’t have frequency
diversity when the modulation index is small.
In the final figure, Figure 6.12, the performance of constant envelope OFDM is
compared to conventional OFDM in the presence of power amplifier nonlinearities. The
SSPA model (see Section 2.3) is used at various input backoff levels. The x-axis is
adjusted to account for the negative impact of input power backoff. The systems are
simulated over Channel Cf . For the OFDM system, QPSK data symbols are used. Three
different CE-OFDM systems are tested: M = 4, 2πh = 0.9; M = 8, 2πh = 2.0; and
M = 16, 2πh = 3.0. The advantage of the CE-OFDM systems is twofold. First, the
CE-OFDM systems operate with IBO = 0 dB. Second, the CE-OFDM systems exploit
119
100
Multipath
Single path
10−1
Bit error rate
10−2
10−4 5 10 15 20 25 30 35 40 45 50
Average signal-to-noise ratio per bit, Eb /N0 (dB)
At the bit error rate 0.001 the CE-OFDM systems outperform the OFDM system
by at least 10 dB. At this bit error rate, the OFDM system has essentially the same
performance with backoff levels of 6 and 10 dB; therefore, IBO = 6 dB is preferred since
the performance is the same but the power efficiency is higher (see Figure 2.14). Even
so, the 6 dB backoff required by the OFDM system is still far less desirable as the 0 dB
backoff used by the CE-OFDM system. Notice that the OFDM system with IBO = 0
dB results in an irreducible error floor just below the bit error rate 0.1.
The results in Figure 6.12 also highlight the poor performance of CE-OFDM at
low SNR due to the threshold effect (as studied in Section 4.1.3). Over the region 0 dB
≤ Eb /N0 ≤ 10 dB, the OFDM system performs better than the CE-OFDM system. Also,
it should be noted that the M = 8 and M = 16 CE-OFDM systems shown have large
modulation index values (2πh = 2.0 and 2πh = 3.0 respectively) which results in spectral
broadening. Roughly speaking, the spectral efficiency of the QPSK/OFDM system is 2
b/s/Hz, which, according to (4.70), is about the same as the M = 4, 2πh = 0.9 CE-
OFDM system. The M = 8 and M = 16 systems have spectral efficiencies of 1.5 and
1.3 b/s/Hz, respectively.
10−1
10−2
OFDM: IBO = 0 dB
10−3 3 dB
6 dB
10 dB
CE-OFDM: M = 4, 2πh = 0.9
M = 8, 2πh = 2.0
M = 16, 2πh = 3.0
10−4 0 5 10 15 20 25 30 35 40
Eb /N0 + IBO (dB)
due to the various parameters involved (M , 2πh, IBO, etc.), and due to the fact that
system requirements vary from system to system. For example, if power amplifier effi-
ciency is the most important requirement, then the input power backoff of 0 dB should
be chosen. At this backoff level, the OFDM system has a very high irreducible error
floor due to the power amplifier distortion, while the CE-OFDM system is relatively
unaffected. Alternatively, if operation at low SNR is important, then CE-OFDM may
not be well suited due to the threshold effect.
The results in this chapter show that CE-OFDM can perform quite well in multipath
fading channels—so long as the channel information (i.e., {H[k]}) is known at the receiver
and so long as the added complexity of the frequency-domain equalizer (i.e., two extra
FFTs) is acceptable. Further work is needed to study the effects of channel coding,
time-varying channels, phase noise, and so forth. Also, a thorough study comparing CE-
OFDM, OFDM and single carrier frequency-domain equalizer (SC-FDE) systems could
provide for interesting results.
Chapter 7
Conclusions
In this thesis the peak-to-average power ratio problem associated with orthogonal
frequency division multiplexing is evaluated. The PAPR statistics are studied and the
effect of power amplifier nonlinearities as a function of power backoff is evaluated by
computer simulation. It is shown that the amount of backoff required to reduce spectral
growth and performance degradation is significant: 6–10 dB depending on the subcarrier
modulation used. Large backoff is an unsatisfactory solution for battery-powered systems
since PA efficiency is low.
A signal transformation method for solving the PAPR problem is presented and
analyzed. The high PAPR OFDM signal is transformed to a 0 dB PAPR constant enve-
lope waveform. At the receiver, the inverse transform is performed prior to the OFDM
demodulator. For the CE-OFDM technique described, phase modulation is used. The
effect of the phase modulator on the transmitted signal’s spectrum is studied. It is shown
that the modulation index controls the spectral containment. The modulation index also
controls the system performance. The optimum receiver is analyzed and a performance
bound and approximation is derived. For a large modulation index, the CE-OFDM sig-
nals become less correlated which improves detection performance. The approximation
of the optimum receiver closely matches simulation results. It also closely matches a
derived bit error rate approximation for a practical phase demodulator receiver. For a
small modulation index and high signal-to-noise ratio, the phase demodulator receiver is
nearly optimum. For a larger modulation index the phase demodulator receiver becomes
sub-optimum due to the limitations of the phase demodulator and phase unwrapper.
121
122
This problem can be suppressed with the use of a properly designed finite impulse re-
sponse lowpass filter which precedes the phase demodulator.
For example, a channel characterized by a two-path model with a weak secondary path,
CE-OFDM might provide acceptable performance without equalization. CPM systems
in the other hand require high quality coherent channels.
Additional future work includes comparing CE-OFDM with other block modulation
technique in terms of PAPR, spectral efficiency, power amplifier efficiency, performance
and complexity. There has been an increasing amount of attention given to conventional
single carrier modulation with the addition of a cyclic prefix which allows for frequency-
domain equalization [107, 154, 460, 463, 574]. However, most single carrier modulations
have a non-constant envelope due to pulse shaping and multilevel QAM symbol con-
stellations. A study is needed to compare these modulation techniques to CE-OFDM
taking into account the effects of the PA at various backoff levels. Also, using CPM with
a cyclic prefix is an interesting idea. Comparing the complexity and spectral efficiency of
such a technique with CE-OFDM would be interesting. Such research will help provide
insight into good designs for future wireless digital communication systems that require
power efficiency and high data rates.
Appendix A
For some applications, a real-valued OFDM signal is required. This can be done by
taking a DFT of a conjugate symmetric vector. The spectral efficiency of the real-valued
OFDM signal is the same as the spectral efficiency of the complex-valued OFDM signal.
X = [X0 , X1 , . . . , XN −1 ]. (A.3)
124
125
∗
XN/2+k = XN/2−k (A.4)
and
X0 = XN/2 = 0. (A.5)
i = 0, 1, . . . , N − 1. But since
= e−j2π(N/2−k)i/N ,
N/2−1
X
x[i] = XN/2−k ej2π(N/2−k)i/N + XN/2−k
∗
e−j2π(N/2−k)i/N , (A.8)
k=1
Now, suppose the data symbols are derived from a M 2 -QAM (quadrature-amplitude
modulation) constellation; that is,
where
<{Xk }, ={Xk } ∈ {±1, ±3, ±(M − 1)}, for all k. (A.13)
In other words, the real and imaginary components are derived from M -PAM (pulse-
amplitude modulation) constellations. Therefore, processing M 2 -QAM data with the
IDFT, (A.11) is a real-valued M -PAM OFDM signal.
In the frequency domain, x(t) is shifted to the right by f c Hz, and the subcarriers are
centered at fc , fc + 1/TB , fc + 2/TB , . . . , fc + (N − 1)/TB Hz. The effective bandwidth of
the signal is therefore N/TB Hz. Each data symbol represents log 2 M bits (i.e., they are
assumed to be selected from a M -ary constellation), therefore the spectral efficiency is
Bits per second (b/s) N log2 M/TB
S1 = = = log2 M b/s/Hz. (A.15)
Bandwidth (Hz) N/TB
The real-valued OFDM signal in (A.11) has the same spectral efficiency as the
complex-valued signal, so long as it is transmitted at baseband. Transmitting the
signal as-is, <{Xk }, k = 1, 2, . . . , (N/2) − 1, modulate cosine subcarriers centered at
1/TB , 2/TB , . . . , [(N/2) − 1]/TB Hz; and likewise, ={Xk }, k = 1, 2, . . . , (N/2) − 1, mod-
ulate sine subcarriers at the same frequencies. The effective bandwidth of the signal is
127
(N/2)/TB Hz1 , and since the real and imaginary parts of X k represent 0.5 log 2 M bits,
the spectral efficiency of the real-valued OFDM signal is
Therefore the spectral efficiency is the same as for the complex case.
However, the spectral efficiency of the real-valued signal is 1/2 that of the complex-
valued signal if the real-valued signal is translated up to a carrier frequency. This is due to
the fact that the cosine and sine subcarriers in (A.11) have a double sideband spectrum:
that is, cos(2πkt/TB ) [or sin(2πkt/TB )] has a spectral components at ±k/T B Hz. [This
isn’t the case for the complex-valued signal, which has complex sinusoids: exp(j2πkt/T B )
has a spectral component only at k/T B Hz and is thus considered single sideband.] The
carrier frequency is typically much larger than the signal bandwidth, so the frequency
translation brings all the negative frequencies to the positive side: −(N/2)/T B + fc 0.
Consequently, the passband transmission of (A.11) results in a signal with double the
bandwidth and 1/2 the spectral efficiency.
1
Only the positive frequencies, f ≥ 0, count.
Appendix B
The first OFDM-like radio to be found in the research literature is the Kineplex sys-
tem presented by Mosier et. al in 1958 [354]. Developed at the Collins Radio Company,
Burbank, CA, the radio used 20 tones separated by 110 Hz, each differentially phase
modulated. This paper caused some interest and some controversy as indicated by E.
D. Sunde’s (Bell Laboratories) comments found at the end of the journal paper.
Weinstein and Ebert, in 1971, where to first to suggest using a DFT for OFDM
modulation [579]. This observation was made six years after Cooley and Tukey published
details of the fast Fourier transform; these developments were significant since all modern
OFDM systems are based on the FFT.
A decade passed with little mention of OFDM in the literature. Then, in the early
80’s researchers from IBM’s Watson Research Center suggests OFDM for a wireline DSL-
type application [408]. They were the first to suggest bit loading. Around this time,
Japanese researcher suggest OFDM for wireless communications [207–209] (also see [6]).
L. J. Cimini’s 1985 paper [102] generated interest when he suggested applying OFDM
128
129
to mobile systems.
In the late 80’s and early 90’s OFDM received wide interest for the applications
of DSL and for wireless digital broadcasting. Kalet and Zervos compare OFDM to
single carrier with decision feedback equalization [248, 614]. The acceptance of OFDM
into xDSL standards was lead primarily by Stanford University’s J. M. Cioffi et al.
[9, 61, 95–97, 105, 446]. Now, OFDM is widely deployed for this consumer electronics
application. In terms of digital broadcasting, OFDM has been accepted for the European
DAB and DVB standards [162, 477, 552]. In the US, OFDM is being used for IBOC
broadcasting [221, 392].
OFDM is being applied to indoor wireless local area networks under the IEEE 802.11
and the ETSI HYPERLAN/2 standards [552]. And as mentioned in Chapter 1, OFDM
is being developed for ultra-wideband systems; cellular systems; wireless metropolitan
area networks; and for power line communication [119, 160, 264, 604].
Active OFDM research continues. The major focus in the OFDM literature includes
OFDM’s sensitivity to Doppler, phase noise, carrier frequency offsets, and nonlinearities.
Channel estimation and synchronization techniques are of interest, along with techniques
to address the PAPR problem.
Conducting a 100% thorough literature review in this field, over the course of a PhD,
is a formidable, if impossible, task. Some statistics of the current author’s attempt are
displayed below.
130
First, to get an idea of the size of the literature, Figure B.1 shows the result of
searching for “OFDM” in the IEEE online literature database. As of the year 2004,
there are over 800 OFDM-specific IEEE journal papers and over 4300 papers when
including papers presented at IEEE conferences.
1400 4500
Journal papers Journal papers
Journal plus conference papers Journal plus conference papers
4000
1200
3500
1000
3000
800
2500
Papers
Papers
2000
600
1500
400
1000
200
500
0 0
1988 1992 1996 2000 2004 1988 1992 1996 2000 2004
So, there are many papers to read and to learn from. Besides the OFDM-specific
papers, there are many interesting and fundamental papers dealing with the general
area of digital communications and information theory. Being familiar with the relevant
literature, which may include several thousands of papers published over many decades,
is the goal, however long-term it may be.
131
600
550
500
450
Papers
400
350
300
250
200
150
Oct 2004 Jan 2005 Apr 2005 Jul 2005 Oct 2005
Filed Piled
This figure shows the number of filed and the number of piled papers as a function
of time, spanning my final year as a PhD student. A filed paper has been printed out,
read, added to a citation list (using BibTeX), and briefly summarized in one or two
paragraphs. A piled paper is in queue waiting to be filed. As the figure shows, the pile
is in good health. In late Spring 2005, a concerted effort was made to “kill the pile”. It
briefly dipped below 150 papers, but the literature is too large—and the battle continues.
132
8
Papers read per day (log scale)
Oct 2004 Jan 2005 Apr 2005 Jul 2005 Oct 2005
Running average Daily points
Figure B.3 shows the running average of papers read per day, and Figure B.4 shows
a histogram of the filed papers’ publication year. One unknown is the true papers-of-
interest count. A simple model might be: 20 papers per year from 1920–1960; 50 papers
per year from 1960–1980; and 100 papers per year from 1980 to present. A histogram of
this projected goal in relation to the current progress is shown in Figure B.5. According
to the model, 4300 papers are of interest, of which roughly 3700 have yet to be filed. Say
350 papers are read per year (which, according to Figure B.3, isn’t entirely unreasonable).
Of these 350 papers, assume that 100 are current-year, leaving the remaining 250 papers
to be from the past. It would therefore take 3700/250 = 14.8 years to “kill the pile”.
133
70
60
50
40
Papers
30
20
10
Desired?
100
80
Papers
60
40
20 Current
Sample Code
The simulations were performed using GNU Octave [188] and the figures were gen-
erated with Gnuplot [189]. In this appendix sample code is provided.
Below is GNU Octave code used to obtain the results for the Channel C f , MMSE
curve in Figure 6.10. The code can easily be adapted to obtain other results, as outlined
below.
134
135
%% Subcarrier matrix
t=0:Tsa:(TB-Tsa); % time vector
W=zeros(NB,N); % initialize unitary matrix
for k=1:N/2 % W is a set of orth. sines and cosines
W(:,k)=cos(2*pi*k*t/TB)’;
end
for k=(N/2+1):N
W(:,k)=sin(2*pi*(k-N/2)*t/TB)’;
end
%% Generate L blocks
in=ceil(M*rand(N,L)); % random symbol index
I=SymMap(in); % data symbols
m=CN*W*I; % OFDM message signal
theta0=2*pi*rand(1,L)-pi; % memory terms (assume uniform)
phi=zeros(NF,L); % initialize CE-OFDM phase signal
for i=1:L % cyclic prefix
phi(:,i)=[2*pi*modh*m(NB-Ng+1:NB,i)+theta0(i);...
2*pi*modh*m(:,i)+theta0(i)];
end
s=A*exp(j*phi); % CE-OFDM signal
%% Channel
tmp=sqrt(1/2)*(randn(Nc,1)+j*randn(Nc,1)); % Gaussian vector
Ch=sqrt(p/sum(p)).*tmp; % channel (normalize average power)
%% Frequency-domain equalizer
H=fft(Ch,Ndft); % channel gains
C=conj(H)./(abs(H).^2+EbN0^(-1)); % correction term (MMSE)
X=fft(rp,Ndft); % to frequency domain
hatS=X.*(C*ones(1,L)); % equalize
x=ifft(hatS,Ndft); % to time domain
%% Filter signal
hats=zeros(NB,L); % initialize
for i=1:L
tmp=(conv(hf,x(:,i))).’; % filtered signal
hats(:,i)=tmp(n2); % filtered signal, desired indices
138
end
%% Display (optional)
if rem(Trans_num,10*L*N*log2(M))==0 % print-frequency
clc
printf([’MMSE, fading ChC, EQ, M=%d, 2pih=%1.1f, J=%d, ’...
’fc=%1.1f, EbN0=%2.1f, Trans_num=%d, ’...
’Error_num=%d, BER=%1.1e’], M, 2*pi*modh, J, fc,...
EbN0_dB(end), Trans_num, Error_num, Error_num/Trans_num)
end
end % end this SNR
%% Plot
semilogy(EbN0_dB,BER)
%% Save
tmp=[EbN0_dB’ BER’];
save -ascii data tmp
To get other results, the above code is used with different values of M , 2πh, equalizer
settings, and/or channel definitions. The ZF equalizer is simulated by changing the
equalizer to
The other fading channels are generated by changing the code that defines the channel.
For Channel Af :
139
For Channel Bf :
For Channel Df :
Additionally, the above template can be used for conventional OFDM with some
minor alterations.
The majority of the figures in this thesis were generated with Gnuplot. Below is
sample code which generates Figure 6.10.
# Define labels.
set xlabel ’[t]{Average signal-to-noise ratio per bit,\
$\mathcal{E}_\text{b}/N_0$ (dB)}’
set ylabel ’Bit error rate’
141
142
Set Theory
∈ is an element of
∈
/ is not an element of
[·] closed interval
[·) open interval
{xn }N
n=1 set of elements x1 , x2 , . . . , xN
143
144
x(t) x as a function of t
x[i] discrete-time samples of x at the ith index
δ(·) delta function
π 3.14159265358979. . .
∞ infinity
Rb
Ra (·)dx definite integral
(·)dx indefinite integral
QN
multiple product
Pn=1
N
n=1 multiple sum
n! factorial
x→a x approaches a
x∗y x convolved with y
|·| absolute value
(·)∗ complex conjugate
d·e ceiling function
b·c floor function
= equal
≡ equal by definition
6= not equal
≈ approximately equal
≤ less than or equal to
≥ greater than or equal to
< strictly less than
> strictly greater than
much less than
much greater than
Power Amplifier
Amax maximum input level
Asat input saturation level
g0 gain
G(·) AM/AM conversions
p sharpness parameter for the SSPA model
αφ , β φ AM/PM parameters for the TWTA model
ηA efficiency of Class-A power amplifier
K backoff ratio
Φ(·) AM/PM conversions
145
Channel
2σ02 scatter component power of frequency-nonselective channel
al complex-valued gain of the lth path
BC coherence bandwidth
(1)
Bτ τ average delay
(2)
Bτ τ delay spread
C channel capacity
h(τ, t) time-variant channel impulse response
h(τ ) time-invariant channel impulse response
h[i] samples of the channel impulse response
H[k] discrete Fourier transform of h[i]
KR Rice factor
L number of discrete paths
rτ τ (v 0 ) frequency correlation function
Sτ τ (τ ) delay power spectral density
v0 frequency separation variable
∆τl propagation delay difference between τ l and τl−1 , that is, ∆τl = τl − τl−1
ρ line-of-sight component power of frequency-nonselective channel
σa2l average power of the lth path
τ continuous propagation delay
τl discrete propagation delay of the lth path
τmax maximum propagation delay
Signal
A signal amplitude
Ab (k) the value of the kth subcarrier at the beginning of the block interval
Ae (k) the value of the kth subcarrier at the end of the block interval
Amax clip level
Bbpf bandwidth of bandpass filter
Bn noise bandwidth
Brms root-mean-square bandwidth
Bs effective bandwidth of CE-OFDM signal
C[k] frequency-domain equalizer terms
CN normalizing constant
d2m,n squared Euclidean distance between mth and nth signal
d2m,n (K) squared Euclidean distance between mth and nth signal as a function of the
phase constant
d2min minimum squared Euclidean distance
D total number of data symbol differences
Eb energy per bit
Eb /N0 signal-to-noise ratio per bit
Eq subcarrier energy
Ex energy of signal x
f frequency variable (cycles/s)
146
TB block period
Tg guard period
Ts symbol period
Tsa sampling period
W effective bandwidth of OFDM signal, W = N/T B
∆m,n (k) data symbol difference between mth and nth signal at the kth subcarrier
γ signal-to-noise ratio per bit (used interchangeably with E b /N0 )
γ̄ average signal-to-noise ratio per bit
γclip clipping ratio
fo normalized carrier frequency offset
ηt transmission efficiency
θi memory term during ith CE-OFDM block interval
ξ(t) noise at the output of phase demodulator
σI2 data symbol variance
σn2 variance of noise samples, n[i]
σφ2 phase signal variance
ρm,n correlation between mth and nth signal
ρm,n (K) correlation between mth and nth signal as a function of the phase constant
ρmax maximum correlation among signals
φ(t) phase signal
φn (t) noise autocorrelation function
ΦAb (f ) Abramson spectrum
Φ̂Ab (f ) estimated Abramson spectrum
Φx (f ) power density spectrum of signal x.
Φ̂x (f ) estimated power density spectrum of signal x.
Bibliography
[5] J. Ahn and H. S. Lee, “Frequency Domain Equalisation of OFDM Signals Over
Frequency Nonselective Rayleigh Fading Channels,” IEE Electr. Lett., vol. 29,
no. 16, pp. 1476–1477, Aug. 1993.
148
149
[12] G. Almási and D. Padua, “MaJIC: Compiling MATLAB for Speed and
Responsiveness,” in Proc. AMC SIGPLAN Conf., vol. 37, no. 5, Berlin, May
2002, pp. 294–303.
[14] J. B. Anderson, T. Aulin, and C.-E. Sundberg, Digital Phase Modulation. New
York: Plenum Press, 1986.
[17] R. R. Anderson and J. Salz, “Spectra of Digital FM,” Bell Syst. Tech. J., vol. 44,
pp. 1165–1189, July–Aug. 1965.
[20] ——, “Adjacent Channel and Cochannel Interferences in CPFSK Systems with
Nonlinear Transmitters and Limiter-Discriminator Detection,” IEEE Trans.
Commun., vol. 36, no. 5, pp. 544–552, May 1988.
[21] O. Andrisano and N. Ladisa, “On the Spectral Efficiency of CPM Systems over
Real Channel in the Presence of Adjacent Channel and Cochannel Interference:
A Comparison Between Partial and Full Response Systems,” IEEE Trans. Veh.
Technol., vol. 39, no. 2, pp. 89–100, May 1990.
[26] S. Armour, A. Nix, and D. Bull, “The Impact of FFT Size on the Performance of
a Combined OFDM-Equalization Radio Modem,” in Proc. IEEE VTC, vol. 3,
Sept. 1999, pp. 1506–1510.
[29] ——, “New OFDM Peak-to-Average Power Reduction Scheme,” in Proc. IEEE
VTC, vol. 1, Rhodes, Greece, May 2001, pp. 756–760.
[33] ——, “Minimum Euclidean Distance and Power Spectrum for a Class of
Smoothed Phase Modulation Codes with Constant Envelope,” IEEE Trans.
Commun., vol. 30, no. 7, pp. 1721–1729, July 1982.
[34] ——, “An easy way to calculate power spectra of digital FM,” IEE Proc., vol.
130, no. 6, pp. 519–526, Oct. 1983.
[44] ——, “On the Rice Model of Noise in FM Receiver,” IEEE Trans. Inform.
Theory, vol. 34, no. 6, pp. 1406–1419, Nov. 1988.
[51] ——, “Selective Fading Limitations of the Kathryn Modem and Some System
Design Considerations,” IEEE Trans. Commun., pp. 320–333, Sept. 1965.
[52] P. A. Bello and B. D. Nelin, “The Effect of Frequency Selective Fading on
Intermodulation Distortion and Subcarrier Phase Stability in Frequency
Modulation Systems,” IEEE Trans. Commun., pp. 87–101, Mar. 1964.
[53] M.-G. D. Benedetto and P. Mandarini, “An Application of MMSE Predistortion
to OFDM System,” IEEE Trans. Commun., vol. 44, no. 11, pp. 1417–1420, Nov.
1996.
[54] N. Benvenuto and S. Tomasin, “On the Comparison Between OFDM and Single
Carrier Modulation With a DFE Using a Frequency-Domain Feedforward Filter,”
IEEE Trans. Commun., vol. 50, no. 6, pp. 947–955, June 2002.
[58] E. Biglieri, M. Elia, and L. Lopresti, “The Optimal Linear Receiving Filter for
Digital Transmission Over Nonlinear Channels,” IEEE Trans. Inform. Theory,
vol. 35, no. 3, pp. 620–625, May 1989.
[62] S. Boyd, “Multitone Signals with Low Crest Factor,” IEEE Trans. Circuits Syst.,
vol. 33, no. 10, pp. 1018–1022, Oct. 1986.
[63] A. Brajal and A. Chouly, “Compensation of nonlinear distortions for orthogonal
multicarrier schemes using predistortion,” in Proc. IEEE Globecom, vol. 3, San
Francisco, Nov. 1994, pp. 1909–1914.
153
[64] J. Brown and R. Churchill, Complex Variables and Applications, 6th ed.
McGraw-Hill, 1996.
[70] ——, “Some Results on the Spectral Efficiency Improvement in Partial Response
CPM Systems,” in Proceedings of the IEEE National Aerospace and Electronics
Conference, vol. 1, Dayton, OH, USA, May 1989, pp. 53–58.
[71] A. J. Cann, “Nonlinearity Model With Variable Knee Sharpness,” IEEE Trans.
Aerosp. Electron. Syst., vol. 16, no. 6, pp. 874–877, Nov. 1980.
[73] G. Cariolaro and F. C. Vagliani, “An OFDM Scheme with a Half Complexity,”
IEEE J. Select. Areas Commun., vol. 13, no. 9, pp. 1586–1599, Dec. 1995.
[76] E. F. Casas and C. Leung, “OFDM for Data Communication Over Mobile Radio
FM Channels–Part I: Analysis and Experimental Results,” IEEE Trans.
Commun., vol. 39, no. 5, pp. 783–793, May 1991.
[77] ——, “OFDM for Data Communication Over Mobile Radio FM Channels–Part
II: Performance Improvement,” IEEE Trans. Commun., vol. 40, no. 4, pp.
680–683, Apr. 1992.
154
[80] R. Cendrillion and M. Moonen, “Efficient Equalizers for Single and Multi-Carrier
Environments with Known Symbol Padding,” in Proc. IEEE ISSPA, vol. 2,
Kuala Lumpur, Aug. 2001, pp. 607–610.
[81] M.-X. Chang and Y. T. Su, “Performance Analysis of Equalized OFDM Systems
in Rayleigh Fading,” IEEE Trans. Wireless Commun., vol. 1, no. 4, pp. 721–732,
Oct. 2002.
[88] H. Chen and G. J. Pottie, “An Orthogonal Projection-Based Approach for PAR
Reduction in OFDM,” IEEE Commun. Lett., vol. 6, no. 5, pp. 169–171, May
2002.
[89] W. cheol Lee, H.-M. Park, K. jin Kang, and K. bae Kim, “Performance Analysis
of Viterbi Decoder Using Channel State Information in COFDM System,” IEEE
Trans. Broadcast., vol. 44, no. 4, pp. 488–496, Dec. 1998.
[106] S. Clanton, “Speeding Up the Scientific Process,” Linux Journal, no. 110, pp.
56–60, June 2003.
[113] ——, “Maximum Likelihood Synchronization for OFDM Using a Pilot Symbol:
Analysis,” IEEE J. Select. Areas Commun., vol. 19, no. 12, pp. 2495–2503, Dec.
2001.
[117] ——, “Low Overhead Pilot-Aided Synchronization for Single Carrier Modulation
With Frequency Domain Equalization,” in Proc. IEEE Globecom, vol. 4, Sydney,
Nov. 1998, pp. 2068–2073.
157
[119] H. Dai and H. V. Poor, “Advanced Signal Processing for Power Line
Communications,” IEEE Commun. Mag., vol. 41, no. 5, pp. 100–107, May 2003.
[125] J. A. Davis and J. Jedwab, “Peak-to-mean power control and error correction for
OFDM transmission using Golay sequences and Reed-Muller codes,” IEE Electr.
Lett., vol. 33, no. 4, pp. 267–268, Feb. 1997.
[133] P. Dent, G. E. Bottomley, and T. Croft, “Jakes Fading Model Revisited,” IEE
Electr. Lett., vol. 29, no. 13, pp. 1162–1163, June 1993.
[137] ——, “On the Performance Evaluation of OFDM Transmission Using Clipping
Techniques,” in Proc. IEEE VTC, vol. 5, Amsterdam, Sept. 1999, pp. 2923–2928.
[144] O. Edfors, M. Sandell, J.-J. van de. Beek, S. K. Wilson, and P. O. Börjesson,
“OFDM Channel Estimation by Singular Value Decomposition,” IEEE Trans.
Commun., vol. 46, no. 7, pp. 931–939, July 1998.
159
[145] P. V. Eetvelt, G. Wade, and M. Tomlinson, “Peak to average power reduction for
OFDM schemes by selective scrambling,” IEE Electr. Lett., vol. 32, no. 21, pp.
1963–1964, Oct. 1996.
[148] ——, “OFDM Uplink for Interactive Broadband Wireless: Analysis and
Simulation in the Presence of Carrier, Clock and Timing Errors,” IEEE Trans.
Broadcast., vol. 47, no. 1, pp. 3–19, Mar. 2001.
[163] P. Fortier, A. Ruiz, and J. M. Cioffi, “Multidimensional Signal Sets Through the
Shell Construction for Parallel Channels,” IEEE Trans. Commun., vol. 40, no. 3,
pp. 500–512, Mar. 1992.
[164] M. E. Fox and M. W. Marcellin, “Shaped BPSK and the 5 kHz UHF Satcom
Channels,” in Proc. IEEE Milcom, vol. 1, Nov. 1991, pp. 326–332.
[166] M. Freidhof and P. Havezov, “A Numerical Method for Spectral Shaping of CPM
Signals,” in Proc. IEEE Globecom, vol. 5, Sydney, Australia, Nov. 1998, pp.
2859–2863.
[169] M. Friese, “Multicarrier modulation with low peak-to-average power ratio,” IEE
Electr. Lett., vol. 32, no. 8, pp. 713–714, Apr. 1996.
[170] ——, “Multitone Signals with Low Crest Factor,” IEEE Trans. Commun.,
vol. 45, no. 10, pp. 1338–1344, Oct. 1997.
[171] ——, “OFDM Signals with Low Crest-Factor,” in Proc. IEEE Globecom,
Phoenix, Nov. 1997, pp. 290–294.
161
[173] ——, “On the Achievable Information Rate with Peak-Power-Limited Orthogonal
Frequency-Division Multiplexing,” IEEE Trans. Inform. Theory, vol. 46, no. 7,
pp. 2579–2587, Nov. 2000.
[175] I. Galton, “Analog-Input Digital Phase-Locked Loops for Precise Frequency and
Phase Demodulation,” IEEE Trans. Circuits Syst. II, vol. 42, no. 10, pp.
621–630, Oct. 1995.
[176] M. J. F.-G. Garcia, S. Zazo, and J. M. Páez-Barrello, “Pilot patters for channel
estimation in OFDM,” IEE Electr. Lett., vol. 36, no. 12, pp. 1049–1050, June
2000.
[179] W. A. Gardner, “On “The Optimal Linear Receiving Filter for Digital
Transmission Over Nonlinear Channels”,” IEEE Trans. Inform. Theory, vol. 37,
no. 1, p. 219, Jan. 1991.
[180] I. Ghareeb, “Bit Error Rate Performance and Power Spectral Density of a
Noncoherent Hybrid Frequency-Phase Modulation System,” IEEE J. Select.
Areas Commun., vol. 13, no. 2, pp. 276–284, Feb. 1995.
[182] M. Ghosh, “Analysis of the Effect of Impulse Noise on Multicarrier and Single
Carrier QAM Systems,” IEEE Trans. Commun., vol. 44, no. 2, pp. 145–147, Feb.
1996.
[184] S. Gifford, J. Kleider, and S. Chuprun, “OFDM Transmitter Power Amplifier and
PAR Reduction Performance: Measurement and Simulation,” in Proc. IEEE
Milcom, vol. 1, Oct. 2002, pp. 591–595.
162
[190] D. L. Goeckel, “Coded Modulation with Non-Standard Signal Sets for Wireless
OFDM Systems,” in Proc. IEEE ICC, vol. 2, Vancouver, BC Canada, June 1999,
pp. 791–795.
[191] L. Goldfeld, V. Lyandres, and D. Wulich, “Minimum BER Power Loading for
OFDM in Fading Channel,” IEEE Trans. Commun., vol. 50, no. 11, pp.
1729–1733, Nov. 2002.
[194] R. Gross and D. Veeneman, “SNR and Spectral Properties for a Clipped DMT
ADSL Signal,” in Proc. IEEE ICC, vol. 2, New Orleans, May 1994, pp. 843–847.
[197] A. Gusmão, R. Dinis, and N. Esteves, “On Frequency Domain Equalization and
Diversity Combining for Broadband Wireless Communications,” IEEE Trans.
Commun., vol. 51, no. 7, pp. 1029–1033, July 2003.
[206] W. Henkel and B. Wagner, “Another Application for Trellis Shaping: PAR
Reduction for DMT (OFDM),” IEEE Trans. Commun., vol. 48, no. 9, pp.
1471–1476, Sept. 2000.
[208] ——, “An Orthogonally Multiplexed QAM System using the Discrete Fourier
Transform,” IEEE Trans. Commun., vol. 29, no. 7, pp. 982–989, July 1981.
[213] H.-H. Hsieh and C.-H. Wei, “Channel Estimation for OFDM Systems Based on
Comb-Type Pilot Arrangement in Frequency Selective Fading Channels,” IEEE
Trans. Consumer Electron., vol. 44, no. 1, pp. 217–225, Feb. 1998.
[214] X. Huang, J. Lu, and J. Zheng, “Tradeoffs Between BER Performance and PAPR
Reduction for Broadband OFDM Transmission with Signal Processing,” in
Proceedings of the IEEE International Symposium on Wireless Personal
Multimedia Communications, vol. 1, Oct. 2002, pp. 154–157.
[216] X. Huang and Y. Li, “The PAM Decomposition of CPM Signals with Integer
Modulation Index,” IEEE Trans. Commun., vol. 51, no. 4, pp. 543–546, Apr.
2003.
[217] L. W. Hughes, “A Simple Upper Bound on the Error Probability for Orthogonal
Signals in White Noise,” IEEE Trans. Commun., vol. 40, no. 4, p. 670, Apr. 1992.
[218] C.-S. Hwang, “Peak power reduction method for multicarrier transmission,” IEE
Electr. Lett., vol. 37, no. 17, pp. 1075–1077, Aug. 2001.
[219] T. Hwang and Y. G. Li, “Iterative Cyclic Prefix Reconstruction for Coded
Single-Carrier Systems with Frequency-Domain Equalization (SC-FDE),” in
Proc. IEEE VTC, vol. 3, Apr. 2003, pp. 1841–1845.
[223] M. I. Irshid and I. S. Solous, “Bit Error Probability for Coherent M -ary PSK
Systems,” IEEE Trans. Commun., vol. 39, no. 3, pp. 349–352, Mar. 1991.
[224] Y. Iwanami and T. Ikeda, “A Numerical Method for Evaluating the Distortion of
Angle-Modulated Signals in a Time Domain,” IEEE Trans. Commun., vol. 34,
no. 11, pp. 1151–1156, Nov. 1986.
[243] P. Jung and P. W. Baier, “On the Representation of CPM Signals by Linear
Superposition of Impulses in the Bandpass Domain,” IEEE J. Select. Areas
Commun., vol. 10, no. 8, pp. 1236–1242, Oct. 1992.
[247] G. K. Kaleh, “Simple Coherent Receiver for Partial Response Continuous Phase
Modulation,” IEEE J. Select. Areas Commun., vol. 7, no. 9, pp. 1427–1436, Dec.
1989.
[248] I. Kalet, “The Multitone Channel,” IEEE Trans. Commun., vol. 37, no. 2, pp.
119–124, Feb. 1989.
[251] M.-S. Kang and W.-J. Song, “A Robust Channel Equalizer for OFDM TV
Receivers,” IEEE Trans. Consumer Electron., vol. 44, no. 3, pp. 1129–1133, Aug.
1998.
[254] ——, “A Data Predistortion Technique with Memory for QAM Radio Systems,”
IEEE Trans. Commun., vol. 39, no. 2, pp. 336–344, Feb. 1991.
167
[258] D. Kim and G. L. Stüber, “Residual ISI Cancellation for OFDM with
Applications to HDTV Broadcasting,” IEEE J. Select. Areas Commun., vol. 16,
no. 8, pp. 1590–1599, Oct. 1998.
[268] B. S. Krongold and D. L. Jones, “An Active-Set Approach for OFDM PAR
Reduction via Tone Reservation,” IEEE Trans. Signal Processing, vol. 52, no. 2,
pp. 495–509, Feb. 2004.
168
[269] ——, “PAR Reduction in OFDM via Active Constellation Extension,” IEEE
Trans. Broadcast., vol. 49, no. 3, pp. 258–268, Sept. 2003.
[271] H. M. Kwon and K. B. E. Lee, “A Novel Digital FM Receiver for Mobile and
Personal Communications,” IEEE Trans. Commun., vol. 44, no. 11, pp.
1466–1476, Nov. 1996.
[272] L. Lamb and A. Robbins, Learning the vi Editor, 6th ed. Beijing: O’Reilly, 1998.
[275] R. Landauer, “The physical nature of information,” Physics Letters A, vol. 217,
pp. 188–193, July 1996.
[282] J. C. Lee and C. K. Un, “Performance Analysis of Digital Tanlock Loop,” IEEE
Trans. Commun., vol. com-30, no. 10, pp. 2398–2411, Oct. 1982.
169
[285] S. H. Leung, S. M. Ju, and G. G. Bi, “Algorithm for repeated clipping and
filtering in peak-to-average power reduction for OFDM,” IEE Electr. Lett.,
vol. 38, no. 25, pp. 1726–1727, Dec. 2002.
[286] J. Li, G. Liu, and G. B. Giannakis, “Carrier Frequency Offset Estimation for
OFDM-Based WLANs,” IEEE Signal Processing Lett., vol. 8, no. 3, pp. 80–82,
Mar. 2001.
[289] X. Li and L. J. Cimini, Jr., “Effects of Clipping and Filtering on the Performance
of OFDM,” in Proc. IEEE VTC, vol. 3, May 1997, pp. 1634–1638.
[290] ——, “Effects of Clipping and Filtering on the Performance of OFDM,” IEEE
Commun. Lett., vol. 2, no. 5, pp. 131–133, May 1998.
[294] ——, “Simplified Channel Estimation for OFDM Systems with Multiple Transmit
Antennas,” IEEE Trans. Wireless Commun., vol. 1, no. 1, pp. 67–75, 2002.
[300] C.-P. Liang, J.-H. Jong, W. E. Shark, and J. R. East, “Nonlinear Amplifier
Effects in Communications Systems,” IEEE Trans. Microwave Theory Tech.,
vol. 47, no. 8, pp. 1461–1466, Aug. 1999.
[303] Y.-P. Lin and S.-M. Phoong, “Perfect Discrete Multitone Modulation with
Optimal Transceivers,” IEEE Trans. Signal Processing, vol. 48, no. 6, pp.
1702–1711, June 2000.
[304] ——, “MMSE OFDM and Prefixed Single Carrier Systems: BER Analysis,” in
Proc. IEEE ICASSP, vol. 4, Apr. 2003, pp. iv–229–iv–232.
[313] B. Lu, X. Wang, and Y. G. Li, “Iterative Receivers for Space–Time Block-Coded
OFDM Systems in Dispersive Fading Channels,” IEEE Trans. Wireless
Commun., vol. 1, no. 2, pp. 213–225, Apr. 2002.
[315] J. E. Ludman, “Gray Code Generation for MPSK Signals,” IEEE Trans.
Commun., vol. com-29, no. 19, pp. 1519–1522, Oct. 1981.
[316] M. Luise and R. Reggiannini, “Carrier Frequency Acquisition and Tracking for
OFDM Systems,” IEEE Trans. Commun., vol. 44, no. 11, pp. 1590–1598, Nov.
1996.
[321] D. Marelli and M. Fu, “Subband Methods for OFDM Equalization,” in Proc.
IEEE ICC, Anchorage, May 2003, pp. 2350–2354.
[322] G. Marsaglia, “Seeds for Random Number Generators,” Commun. ACM, vol. 46,
no. 5, pp. 90–93, 2003.
[323] G. Marsaglia and W. W. Tsang, “The Ziggurat Method for Generating Random
Variables,” Journal of Statistic Software, vol. 5, no. 8, pp. 1–7, 2000.
[326] T. Maseng, “Digitally Phase Modulated (DPM) Signal,” IEEE Trans. Commun.,
vol. 33, no. 9, pp. 911–918, Sept. 1985.
[327] E. Masry, “Distortionless Demodulation of Narrow-Band Single-Sideband Angle
Modulated Signals,” IEEE Trans. Inform. Theory, vol. 23, no. 5, pp. 582–591,
Sept. 1977.
[328] Mathematica. [Online]. Available: http://www.wolfram.com/
[330] T. May and H. Rohling, “Reducing the Peak-to-Average Power Ratio in OFDM
Radio Transmission Systems,” in Proc. IEEE VTC, vol. 3, Ottawa, May 1998,
pp. 2774–2478.
[331] T. May, H. Rohling, and V. Engels, “Performance Analysis of Viterbi Decoding
for 64-DAPSK and 64-QAM Modulated OFDM Signals,” IEEE Trans.
Commun., vol. 46, no. 2, pp. 182–190, Feb. 1998.
[332] J. E. Mazo and J. Salz, “Spectral Properties of Single-Sideband Angle
Modulation,” IEEE Trans. Commun. Technol., vol. 16, no. 1, pp. 52–62, Feb.
1968.
[343] S. L. Miller and R. J. O’Dea, “Peak Power and Bandwidth Efficient Linear
Modulation,” IEEE Trans. Commun., vol. 46, no. 12, pp. 1639–1648, Dec. 1998.
[346] H. Minn, C. Tellambura, and V. K. Bhargava, “On the Peak Factors of Sampled
and Continuous Signals,” IEEE Commun. Lett., vol. 5, no. 4, pp. 129–131, Apr.
2001.
[347] H. Minn, M. Zeng, and V. K. Bhargava, “On Timing Offset Estimation for
OFDM Systems,” IEEE Commun. Lett., vol. 4, no. 7, pp. 242–244, July 2000.
[352] M. Morelli and U. Mengali, “An Improved Frequency Offset Estimator for
OFDM Applications,” IEEE Commun. Lett., vol. 3, no. 3, pp. 75–77, Mar. 1999.
174
[353] R. Morrison, L. J. Cimini, Jr., and S. K. Wilson, “On the Use of a Cyclic
Extension in OFDM,” in Proc. IEEE VTC, vol. 2, Atlantic City, Oct. 2001, pp.
664–668.
[355] S. H. Müller and J. B. Huber, “OFDM with reduced peak-to-average power ratio
by optimum combination of partial transmit sequences,” IEE Electr. Lett.,
vol. 33, no. 5, pp. 368–369, Feb. 1997.
[359] K. Murota and K. Hirade, “GMSK Modulation for Digital Mobile Radio
Telephony,” IEEE Trans. Commun., vol. com-29, no. 7, pp. 1044–1050, July
1981.
[363] Y. Nakamura, Y. Saito, and S. Aikawa, “256 QAM Modem for Multicarrier 400
Mbit/s Digital Radio,” IEEE J. Select. Areas Commun., vol. sac-5, no. 3, pp.
329–335, Apr. 1987.
[365] S. Narahashi and T. Nojima, “New phasing scheme of N -multiple carriers for
reducing peak-to-average power ratio,” IEE Electr. Lett., vol. 30, no. 17, pp.
1382–1383, Aug. 1994.
175
[366] R. Negi and J. Cioffi, “Pilot Tone Selection for Channel Estimation in a Mobile
OFDM System,” IEEE Trans. Consumer Electron., vol. 44, no. 3, pp. 1122–1128,
Aug. 1998.
[367] S. Neugebauer, G. Ford, and M. Ready, “Characterization of Multipath-Distorted
and Frequency-Discriminated CPM Signals,” in Conference Record of the
Asilomar Conference on Signals, Systems and Computers, vol. 2, Pacific Grove,
CA, Oct. 2000, pp. 819–822.
[368] W. T. Ng and V. K. Dubey, “Analysis of PCC-OFDM Systems for General
Time-Varying Channel,” IEEE Commun. Lett., vol. 9, no. 5, pp. 394–396, May
2005.
[369] Y. H. Ng, P. M. Grant, and R. A. Stirling-Gallacher, “Carrier tracking technique
for OFDM signal transmission,” IEE Electr. Lett., vol. 32, no. 22, pp. 2047–2048,
1996.
[370] H. Nikookar and K. S. Lidsheim, “Random Phase Updating Algorithm for
OFDM Transmission With Low PAPR,” IEEE Trans. Broadcast., vol. 48, no. 2,
pp. 123–128, June 2002.
[371] H. Nikopour and S. H. Jamali, “On the Performance of OFDM Systems Over a
Cartesian Clipping Channel: A Theoretical Approach,” IEEE Trans. Wireless
Commun., vol. 3, no. 6, pp. 2083–2096, Nov. 2004.
[372] A. J. Noga and T. K. Sarkar, “A Discrete-Time Method of Demodulating Large
Deviation FM Signals,” IEEE Trans. Commun., vol. 47, no. 8, pp. 1194–1200,
Aug. 1999.
[373] A. H. Nuttall, “Error Probabilities for Equicorrelated M -ary Signals Under
Phase-Coherent and Phase-Incoherent Reception,” IRE Trans. Inform. Theory,
vol. 8, pp. 305–314, July 1962.
[374] H. Ochiai, “Power Efficiency Comparison of OFDM and Single-Carrier Signals,”
in Proc. IEEE VTC, vol. 2, Sept. 2002, pp. 899–903.
[375] ——, “Performance Analysis of Peak Power and Band-Limited OFDM System
With Linear Scaling,” IEEE Trans. Wireless Commun., vol. 2, no. 5, pp.
1055–1065, Sept. 2003.
[376] ——, “Performance of Optimal and Suboptimal Detection for Uncoded OFDM
System with Deliberate Clipping and Filtering,” in Proc. IEEE Globecom, vol. 3,
Dec. 2003, pp. 1618–1622.
[377] ——, “A Novel Trellis-Shaping Design With Both Peak and Average Power
Reduction for OFDM Systems,” IEEE Trans. Commun., vol. 52, no. 11, pp.
1916–1926, Nov. 2004.
[378] H. Ochiai, M. P. C. Fossorier, and H. Imai, “On Decoding of Block Codes with
Peak-Power Reduction in OFDM Systems,” IEEE Commun. Lett., vol. 4, no. 7,
pp. 226–228, July 2000.
176
[379] H. Ochiai and H. Imai, “Channel Capacity of Clipped OFDM Systems,” in Proc.
IEEE Intern. Sympo. Inform. Theory, Sorrento, June 2000, p. 219.
[380] ——, “Performance of the Deliberate Clipping with Adaptive Symbol Selection
for Strictly Band-Limited OFDM Systems,” IEEE J. Select. Areas Commun.,
vol. 18, no. 11, pp. 2270–2277, Nov. 2000.
[381] ——, “On the Distribution of the Peak-to-Average Power Ratio in OFDM
Signals,” IEEE Trans. Commun., vol. 49, no. 2, pp. 282–289, Feb. 2001.
[387] A. V. Oppenheim and A. S. Willsky, Signals and Systems. Upper Saddle River,
New Jersey: Prentice Hall, 1997.
[389] R. Pandey, H. Leib, and S. Pasupathy, “Orthogonal Phase Functions for Full
Response Non-Coherent CPM Systems,” in Proceedings of the IEEE Pacific Rim
Conference on Communications, Computers and Signal Processing, Victoria, BC,
Canada, June 1989, pp. 320–323.
[394] A. Papoulis, Probability, Random Variables, and Stochastic Processes, 3rd ed.
Boston, MA: McGraw-Hill, 1991.
[396] J. Park, E. Joe, M.-J. Choe, and B.-S. Song, “A 5-MHz IF Digital FM
Demodulator,” IEEE J. Solid-State Circuits, vol. 34, no. 1, pp. 3–11, Jan. 1999.
[400] K. G. Paterson and V. Tarokh, “On the Existence and Construction of Good
Codes with Low Peak-to-Average Power Ratios,” IEEE Trans. Inform. Theory,
vol. 46, no. 6, pp. 1974–1987, Sept. 2000.
[401] M. Pätzold, Mobile Fading Channels. West Sussex, England: John Wiley &
Sons, 2002.
[402] M. Pätzold, 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. 254–269, Feb. 1998.
[406] ——, “On the Theory of Error Rates for Narrow-Band Digital FM,” IEEE Trans.
Commun., vol. com-29, no. 11, pp. 1634–1643, Nov. 1981.
[407] ——, “Refinements to the Theory of Error Rates for Narrow-Band Digital FM,”
IEEE Trans. Commun., vol. 36, no. 4, pp. 509–513, Apr. 1988.
178
[408] A. Peled and A. Ruiz, “Frequency Domain Data Transmission Using Reduced
Computational Complexity Algorithms,” in Proc. IEEE ICASSP, vol. 5, Apr.
1980, pp. 964–967.
[421] ——, Digital Communications, 4th ed. New York: McGraw-Hill, 2001.
179
[432] S. O. Rice, “Distributions for Noise Plus Several Sine Waves—The Problem of
Computation,” IEEE Trans. Commun., pp. 851–853, June 1974.
[440] H. Rohling and T. May, “Comparison of PSK and DPSK Modulation in a Coded
OFDM System,” in Proc. IEEE VTC, vol. 2, Phoenix, May 1997, pp. 870–874.
[441] C. Rößing and V. Tarokh, “A Construction of OFDM 16-QAM Sequences Having
Low Peak Powers,” IEEE Trans. Inform. Theory, vol. 47, no. 5, pp. 2091–2094,
July 2001.
[442] H. E. Rowe, Signals and Noise in Communication Systems. Princeton, N. J.: D.
Van Nostrand Company, Inc., 1965.
[446] A. Ruiz, J. M. Cioffi, and S. Kasturia, “Discrete Multiple Tone Modulation with
Coset Coding for the Spectrally Shaped Channel,” IEEE Trans. Commun.,
vol. 40, no. 6, pp. 1012–1029, June 1992.
[449] H.-G. Ryu, B.-I. Jin, and I.-B. Kim, “PAPR Reduction Using Soft Clipping and
ACI Rejection in OFDM System,” IEEE Trans. Consumer Electron., vol. 48,
no. 1, pp. 17–22, Feb. 2002.
[450] H.-G. Ryu, J.-E. Lee, and J.-S. Park, “Dummy Sequence Insertion (DSI) for
PAPR Reduction in the OFDM communication System,” IEEE Trans. Consumer
Electron., vol. 50, no. 1, pp. 89–94, Feb. 2004.
[451] H.-G. Ryu and K.-J. Youn, “A New PAPR Reduction Scheme: SPW (Subblock
Phase Weighting),” IEEE Trans. Consumer Electron., vol. 48, no. 1, pp. 81–89,
Feb. 2002.
[452] M. Sablatash, “Transmission of All-Digital Advanced Television: State of the Art
and Future Directions,” IEEE Trans. Broadcast., vol. 40, no. 2, pp. 102–121,
June 1994.
[453] H. Saeedi, M. Sharif, and F. Marvasti, “Clipping Noise Cancellation in OFDM
Systems Using Oversampled Signal Reconstruction,” IEEE Commun. Lett.,
vol. 6, no. 2, pp. 73–75, Feb. 2002.
[454] A. A. M. Saleh, “Frequency-Independent and Frequency-Dependent Nonlinear
Models of TWT Amplifiers,” IEEE Trans. Commun., vol. 29, no. 11, pp.
1715–1720, Nov. 1981.
[455] B. R. Saltzberg, “Performance of an Efficient Parallel Data Transmission
System,” IEEE Trans. Commun. Technol., vol. com-15, no. 6, pp. 805–811, Dec.
1967.
[456] A. A. Salvekar, C. Aldana, J. Tellado, and J. Cioffi, “Peak-to-average power ratio
reduction for block transmission systems in the presence of transmit filtering,” in
Proc. IEEE ICC, Helsinki, June 2001, pp. 175–178.
[457] S. D. Sandberg and M. A. Tzannes, “Overlapped Discrete Multitone Modulation
for High Speed Copper Wire Communications,” IEEE J. Select. Areas Commun.,
vol. 13, no. 9, pp. 1571–1585, Dec. 1995.
[458] G. Santella, “A Frequency and Symbol Synchronization System for OFDM
Signals: Architecture and Simulation Results,” IEEE Trans. Veh. Technol.,
vol. 49, no. 1, pp. 254–275, Jan. 2000.
[459] G. Santella and F. Mazzenga, “A Hybrid Analytical-Simulation Procedure for
Performance Evaluation in M-QAM-OFDM Schemes in Presence of Nonlinear
Distortions,” IEEE Trans. Veh. Technol., vol. 47, no. 1, pp. 142–151, Feb. 1998.
[460] H. Sari, G. Karam, and I. Jeanclaude, “An Analysis of Orthogonal
Frequency-Division Multiplexing for Mobile Radio Applications,” in Proc. IEEE
VTC, vol. 3, Stockholm, June 1994, pp. 1635–1639.
[461] ——, “Frequency-Domain Equalization of Mobile Radio and Terrestrial
Broadcast Channels,” in Proc. IEEE Globecom, vol. 1, San Francisco, Nov. 1994,
pp. 1–5.
182
[466] ——, “Coding to Reduce Both PAR and PICR of an OFDM Signal,” IEEE
Commun. Lett., vol. 6, no. 8, pp. 316–318, Aug. 2002.
[467] ——, “Partial Transmit Sequence and Selected Mapping Schemes to Reduce ICI
in OFDM Systems,” IEEE Commun. Lett., vol. 6, no. 8, pp. 313–315, Aug. 2002.
[468] T. M. Schmidl and D. C. Cox, “Blind synchronisation for OFDM,” IEE Electr.
Lett., vol. 33, no. 2, pp. 113–114, Jan. 1997.
[469] ——, “Robust Frequency and Timing Synchronization for OFDM,” IEEE Trans.
Commun., vol. 45, no. 12, pp. 1613–1621, Dec. 1997.
[475] M. Sharif and B. Hassibi, “On Multicarrier Signals Where the PMEPR of a
Random Codeword is Asymptotically log n,” IEEE Trans. Inform. Theory,
vol. 50, no. 5, pp. 895–903, May 2004.
[476] M. Shell, “How to Use the IEEEtran LATEXClass,” Journal of LATEXFiles, vol. 1,
no. 11, pp. 1–22, Nov. 2002.
183
[477] P. Shelswell, “The COFDM modulation system: the heart of digital audio
broadcasting,” IEEE Electron. Commun. Eng. J., pp. 127–136, June 1995.
[481] T. Shi, S. Zhou, and T. Tao, “Capacity of Single Carrier Systems With
Frequency-Domain Equalization,” in Proc. IEEE CAS Symp., vol. 2, Shanghai,
June 2004, pp. 429–432.
[483] M. K. Simon and M.-S. Alouini, Digital Communication over Fading Channels.
New York: John Wiley & Sons, Inc., 2000.
[488] ——, “Rayleigh Fading Channels in Mobile Digital Communication System, Part
II: Mitigation,” IEEE Commun. Mag., pp. 102–109, July 1997.
[489] H.-K. Song, Y.-H. You, J.-H. Paik, and Y.-S. Cho, “Frequency-Offset
Synchronization and Channel Estimation for OFDM-Based Transmission,” IEEE
Commun. Lett., vol. 4, no. 3, pp. 95–97, Mar. 2000.
[490] R. Song and S.-H. Leung, “A Novel OFDM Receiver with Second Order
Polynomial Nyquist Window Function,” IEEE Commun. Lett., vol. 9, no. 5, pp.
391–393, May 2005.
184
[492] M. Speth, S. A. Fechtel, G. Fock, and H. Meyr, “Optimum Receiver Design for
Wireless Broad-Band Systems Using OFDM–Part I,” IEEE Trans. Commun.,
vol. 47, no. 11, pp. 1668–1677, Nov. 1999.
[500] C.-E. Sundberg, “Continuous Phase Modulation,” IEEE Commun. Mag., vol. 24,
no. 4, pp. 25–37, Apr. 1986.
[501] E. D. Sunde, Communication Systems Engineering Theory. New York: John
Wiley & Sons, Inc., 1969.
[505] B. T. Tan and Y. K. Some, “Crest Factor Minimisation in FDM PSK Systems,”
IEE Electr. Lett., vol. 26, no. 13, pp. 859–861, 1990.
[508] V. Tarokh and H. Jafarkhani, “On the Computation and Reduction of the
Peak-to-Average Power Ratio in Multicarrier Communications,” IEEE Trans.
Commun., vol. 48, no. 1, pp. 37–44, Jan. 2000.
[509] I. A. Tasadduq and R. K. Rao, “OFDM-CPM Signals,” IEE Electr. Lett., vol. 38,
no. 2, pp. 80–81, Jan. 2002.
[510] ——, “PAPR Reduction of OFDM Signals using Multiamplitude CPM,” IEE
Electr. Lett., vol. 38, no. 16, pp. 915–917, Aug. 2002.
[514] C. Tellambura, “Upper bound on peak factor of n-multiple carriers,” IEE Electr.
Lett., vol. 33, no. 19, pp. 1608–1609, Sept. 1997.
[515] ——, “Use of m-sequences for OFDM peak-to-average power ratio reduction,”
IEE Electr. Lett., vol. 33, no. 15, pp. 1300–1301, July 1997.
[517] ——, “Phase optimisation criterion for reducing peak-to-average power ratio in
OFDM,” IEE Electr. Lett., vol. 34, no. 2, pp. 169–170, Jan. 1998.
[522] C. Tellambura, “Improved Phase Factor Computation for the PAR Reduction of
an OFDM Signal Using PTS,” IEEE Commun. Lett., vol. 5, no. 4, pp. 135–137,
Apr. 2001.
[526] ——, “Spectral Estimation of Digital Signaling Using The Welch Method.”
[Online]. Available: http://zeidler.ucsd.edu/∼sct/holdings/welch/
[530] ——, “The Effectiveness of Signal Clipping for PAPR and Total Degradation
Reduction in OFDM Systems,” in Proc. IEEE Globecom, St. Louis, Dec. 2005.
[531] T. T. Tjhung, “Power Spectra and Power Distributions of Random Binary F.M.
Signals with Premodulation Shaping,” IEE Electr. Lett., vol. 1, no. 6, pp.
176–178, Aug. 1965.
[532] T. T. Tjhung, K. K. Teo, and P. H. Wittke, “Effects of Pulse Shaping and Soft
Decisions on the Performance of Digital FM with Discriminator Detection,”
IEEE Trans. Commun., vol. 34, no. 11, pp. 1116–1122, Nov. 1986.
[533] S. Tomasin, “Overlap and Save Frequency Domain DFE for Throughput Efficient
Single Carrier Transmission,” in Proc. IEEE PIMRC, Berlin, Sept. 2005, in press.
187
[534] L. Tomba, “On the Effect of Wiener Phase Noise in OFDM Systems,” IEEE
Trans. Commun., vol. 46, no. 5, pp. 580–583, May 1998.
[542] U. Tureli, H. Liu, and M. D. Zoltowski, “OFDM Blind Carrier Offset Estimation:
ESPRIT,” IEEE Trans. Commun., vol. 48, no. 9, pp. 1459–1461, Sept. 2000.
[546] ——, “Optimal Finite Duration Pulses for OFDM,” IEEE Trans. Commun.,
vol. 44, no. 1, pp. 10–14, Jan. 1996.
[547] J.-J. van de Beek, M. Sandell, and P. O. Börjesson, “ML Estimation of Time and
Frequency Offset in OFDM Systems,” IEEE Trans. Signal Processing, vol. 45,
no. 7, pp. 1800–1805, July 1997.
188
[550] ——, “Peak Factor Minimization of Input and Output Signals of Linear
Systems,” IEEE Trans. Instrum. Meas., vol. 37, no. 2, pp. 207–212, June 1988.
[553] R. van Nee and A. de Wild, “Reducing the Peak-to-Average Power Ratio of
OFDM,” in Proc. IEEE VTC, vol. 3, May 1998, pp. 2072–2076.
[554] R. van Nee and R. Prasad, OFDM for Wireless Multimedia Communications.
Boston: Artech House, 2000, ch. 6.
[555] ——, OFDM for Wireless Multimedia Communications. Boston: Artech House,
2000.
[556] R. D. van Nee, “OFDM Codes for Peak-to-Average Power Reduction and Error
Correction,” in Proc. IEEE Globecom, London, Nov. 1996, pp. 740–744.
[562] E. Viterbo and K. Fazel, “How to combat long echos in OFDM transmission
schemes: Sub-channel equalization or more power channel coding,” in Proc.
IEEE Globecom, vol. 3, Nov. 1995, pp. 2069–2074.
[564] L. Wan and V. K. Dubey, “BER Performance of OFDM System Over Frequency
Nonselective Fast Ricean Fading Channels,” IEEE Commun. Lett., vol. 5, no. 1,
pp. 19–21, Jan. 2001.
[568] X. Wang, P. Ho, and Y. Wu, “Robust Channel Estimation and ISI Cancellation
for OFDM Systems With Suppressed Features,” IEEE J. Select. Areas Commun.,
vol. 23, no. 5, pp. 963–972, May 2005.
[570] ——, “Replay to the Comments on: “Reduction of peak-to-average power ratio of
OFDM system using a companding technique”,” IEEE Trans. Broadcast., vol. 45,
no. 4, pp. 420–423, Dec. 1999.
[571] X. Wang, T. T. Tjhung, and Y. Wu, “On the SER and Spectral Analyses of
A-Law Companding Multicarrier Modulation,” IEEE Trans. Veh. Technol.,
vol. 52, no. 5, pp. 1408–1412, Sept. 2003.
[573] ——, “Complex-Field Coding for OFDM Over Fading Wireless Channels,” IEEE
Trans. Inform. Theory, vol. 49, no. 3, pp. 707–720, Mar. 2003.
[591] ——, “Reduction of peak to mean ratio of multicarrier modulation using cyclic
coding,” IEE Electr. Lett., vol. 32, no. 5, pp. 432–433, Feb. 1996.
[592] ——, “Comments on the Peak Factor of Sampled and Continuous Signals,” IEEE
Commun. Lett., vol. 4, no. 7, pp. 213–214, July 2000.
[596] G. Wunder and H. Boche, “Upper Bounds on the Statistical Distribution of the
Crest-Factor in OFDM Transmission,” IEEE Trans. Inform. Theory, vol. 49,
no. 2, pp. 488–494, Feb. 2003.
[600] ——, “Statistical Simulation Models for Rayleigh and Rician Fading,” in Proc.
IEEE ICC, vol. 5, May 2003, pp. 3524–3529.
[601] F. Xiong, “M -ary Amplitude Shift Keying OFDM System,” IEEE Trans.
Commun., vol. 51, no. 10, pp. 1638–1642, Oct. 2003.
[603] B. Yang, K. B. Letaief, R. S. Cheng, and Z. Cao, “Timing Recovery for OFDM
Transmission,” IEEE J. Select. Areas Commun., vol. 18, no. 11, pp. 2278–2291,
Nov. 2000.
192
[606] K.-W. Yip and T.-S. Ng, “Efficient Simulation of Digital Transmission over
WSSUS Channels,” IEEE Trans. Commun., vol. 43, no. 12, pp. 2907–2913, Dec.
1995.
[607] ——, “Karhunen–Loéve Expansion of the WSSUS Channel and Its Application
to Efficient Simulation,” IEEE J. Select. Areas Commun., vol. 15, no. 4, pp.
640–646, May 1997.
[608] ——, “Matched Filter Bound for Multipath Rician-Fading Channels,” IEEE
Trans. Commun., vol. 46, no. 4, pp. 441–445, Apr. 1998.
[615] X. Zhang and M. P. Fitz, “Space–Time Code Design with Continuous Phase
Modulation,” IEEE J. Select. Areas Commun., vol. 21, no. 5, pp. 783–792, June
2003.
[616] Y. Zhao and S.-G. Häggman, “Sensitivity to Doppler Shift and Carrier Frequency
Errors in OFDM Systems—The Consequences and Solutions,” in Proc. IEEE
VTC, vol. 3, Apr. 1996, pp. 1564–1568.
193
[619] Y. Zhao and A. Huang, “A Novel Channel Estimation Method for OFDM Mobile
Communication Systems Based on Pilot Signals and Transform-Domain
Processing,” in Proc. IEEE VTC, vol. 3, Phoenix, May 1997, pp. 2089–2093.
[621] Y. R. Zheng and C. Xiao, “Improved Models for the Generation of Multiple
Uncorrelated Rayleigh Fading Waveforms,” IEEE Commun. Lett., vol. 6, no. 6,
pp. 256–258, June 2002.
[622] K. Zhong, T. T. Tjhung, and F. Adachi, “A General SER Formula for an OFDM
System With MDPSK in Frequency Domain over Rayleigh Fading Channels,”
IEEE Trans. Commun., vol. 52, no. 4, pp. 584–594, Apr. 2004.
[626] T. N. Zogakis and J. M. Cioffi, “The Effect of Timing Jitter on the Performance
of a Discrete Multitone System,” IEEE Trans. Commun., vol. 44, no. 7, pp.
799–808, July 1996.
This thesis was typeset using the LATEX document preparation system [348]. The
bibliography was managed using BibTeX (with help from bibtool). All numerical work,
including the computer simulations, was done with GNU Octave [188]. The block dia-
grams were drawn using Xfig [597] and all of the other figures were generated with Gnu-
plot [189] (using the pslatex driver). The source files were backed up and synchronized
among multiple computers using rsync. The LATEX output was converted to PostScript
using dvips; the PostScript was converted to PDF (portable document format) using
Ghostscript. The size of the PDF output is 1.9 megabytes.
The work was done at UCSD on a Dell Precision 370 workstation running the Debian
GNU/Linux operating system [131]. The X11 window system provided the graphical
user interface; the window manager used was IceWM. Typically the work was conducted
across several rxvt terminal emulators—arranged across multiple workspaces—running
bash. All work was done using the Vim (Vi improved) text editor [523]. PDF output
was viewed using xpdf. The work was also done at various locations throughout the San
Diego area on a Dell Inspiron 4000 laptop computer running the same software.
In terms of compilation time, this thesis takes roughly 10 s to compile on the work-
station which has a 3 gigahertz microprocessor and 1 gigabyte of memory.
194