CommunicationsLaboratory 20161206

A Digital Communication

Laboratory

Implementing a Software-Defined
Acoustic Modem

Lee C. Potter and Yang Yang
c
2015
c 2015 LC Potter

All rights reserved.
Terms of Use: This work is licensed under a Creative Commons

Attribution-ShareAlike 4.0 International License. It is attributed to
Lee Potter and Yang Yang.
https://creativecommons.org/licenses/by-sa/4.0/legalcode
This volume was typeset by the authors in LATEX.
First edition, December 2015
Matlab and Simulink are trademarks of MathWorks. LabVIEW, USRP, Uni-

versal Software Radio Peripheral, and National Instruments are trademarks of
National Instruments. Virtex is a trademark of Xilinx. Windows is a trademark
of Microsoft. OpenStax CNX is a trademark of Rice University. iPhone, iPad,
and iTunes are trademarks of Apple. Google Play and Android are trademarks
of Google. PicoZed is a trademark of Avnet. Max2830 is a trademark of Maxim
Integrated.
The routine plottf.m and Figures 1.5, 1.7, 2.1, 2.2, 4.2, and 4.4 are adapted from
[31] with permission from the author. Figure B.5 is provided courtesy of Aaron
Wise; Figure B.6 is used with permission.
This book is provided by the copyright holder as is. Any express or implied war-
ranties, including, but not limited to, the implied warranties of merchantability
and fitness for a particular purpose are disclaimed. In no event shall the authors
be liable for any direct, indirect, incidental, special, exemplary or consequential
damages (including, but not limited to, procurement of substitute goods or ser-
vices; loss of use, data, or profits; or business interruption) however caused and
on any theory of liability, whether in contract, strict liability, or tort (including
negligence or otherwise) arising in any way out of the use of this book or any
information, theories, or software contained or described in it, even if advised of
the possibility of such damage.
Neither the names of the authors nor the name of The Ohio State University may
be used to endorse or promote products derived from this book, or the software
contained in it, without specific prior written permission.
Contents
Preface vii
1 Introduction 1
1.1 A Physical Layer Model . . . . . . . . . . . . . . . . . 1
1.2 Software Radio . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Background . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3.1 Sampling and aliasing . . . . . . . . . . . . . . 6
1.3.2 Frequency upconversion . . . . . . . . . . . . . 7
1.3.3 Frequency downconversion . . . . . . . . . . . . 9
1.4 Explorations . . . . . . . . . . . . . . . . . . . . . . . 12
1.4.1 Getting started with audio I/O . . . . . . . . . 13
1.4.2 Spectral content of signals . . . . . . . . . . . . 14
1.4.3 Low-pass filtering . . . . . . . . . . . . . . . . . 16
1.4.4 Amplitude modulation . . . . . . . . . . . . . . 17
1.4.5 AM demodulation . . . . . . . . . . . . . . . . 18
1.5 Demonstration . . . . . . . . . . . . . . . . . . . . . . 19
1.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . 19
2 Quadrature Amplitude Modulation 23

2.1 Background . . . . . . . . . . . . . . . . . . . . . . . . 23
2.1.1 Quadrature amplitude modulation . . . . . . . 23
2.1.2 Complex-baseband representation . . . . . . . 25
2.1.3 Complex-baseband equivalent channel . . . . . 26
2.1.4 Coherent demodulation . . . . . . . . . . . . . 28
2.1.5 Linear phase filters . . . . . . . . . . . . . . . . 29
2.2 Digital Hardware [Optional] . . . . . . . . . . . . . . . 30
2.3 Explorations . . . . . . . . . . . . . . . . . . . . . . . 32
2.4 Demonstration . . . . . . . . . . . . . . . . . . . . . . 35
2.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . 36
iii
iv CONTENTS
3 Digital Modulation 37
3.1 Background . . . . . . . . . . . . . . . . . . . . . . . . 37
3.1.1 Digital modulation . . . . . . . . . . . . . . . . 37
3.1.2 Symbol detection . . . . . . . . . . . . . . . . . 40
3.1.3 Bit and symbol error rates . . . . . . . . . . . . 42
3.2 Explorations . . . . . . . . . . . . . . . . . . . . . . . 45
3.3 Demonstration . . . . . . . . . . . . . . . . . . . . . . 47
3.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . 48
4 Pulse Shaping & ISI 51

4.1 Background . . . . . . . . . . . . . . . . . . . . . . . . 51
4.1.1 Inter-symbol interference . . . . . . . . . . . . 53
4.1.2 Matched filter . . . . . . . . . . . . . . . . . . . 54
4.1.3 Eye diagram . . . . . . . . . . . . . . . . . . . 57
4.2 Explorations . . . . . . . . . . . . . . . . . . . . . . . 57
4.3 Demonstration . . . . . . . . . . . . . . . . . . . . . . 61
4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . 61
5 Synchronization 63
5.1 Background . . . . . . . . . . . . . . . . . . . . . . . . 64
5.1.1 Symbol timing . . . . . . . . . . . . . . . . . . 64
5.1.2 Frame timing for flat channels . . . . . . . . . 66
5.2 Explorations . . . . . . . . . . . . . . . . . . . . . . . 68
5.3 Demonstration . . . . . . . . . . . . . . . . . . . . . . 74
5.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . 75
6 Frequency Recovery 77
6.1 Background . . . . . . . . . . . . . . . . . . . . . . . . 77
6.1.1 Frequency recovery . . . . . . . . . . . . . . . . 77
6.1.2 Frequency-selective fading channel model . . . 81
6.1.3 Channel measurement . . . . . . . . . . . . . . 81
6.2 Explorations . . . . . . . . . . . . . . . . . . . . . . . 82
6.3 Demonstration . . . . . . . . . . . . . . . . . . . . . . 85
6.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . 85
7 Acoustic Modem 87
7.1 Background . . . . . . . . . . . . . . . . . . . . . . . . 87
7.1.1 Mobile app . . . . . . . . . . . . . . . . . . . . 87
7.1.2 Data packet . . . . . . . . . . . . . . . . . . . . 88
7.1.3 Spectral efficiency . . . . . . . . . . . . . . . . 88
CONTENTS v
7.1.4 Differential PSK modulation . . . . . . . . . . 90

7.2 Explorations . . . . . . . . . . . . . . . . . . . . . . . 91
7.3 Demonstration . . . . . . . . . . . . . . . . . . . . . . 93
7.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . 93
8 Frequency-Selective Fading 95
8.1 Background . . . . . . . . . . . . . . . . . . . . . . . . 97
8.1.1 Frame timing for ISI channels . . . . . . . . . . 97
8.1.2 Frequency recovery for ISI channels . . . . . . 97
8.1.3 Linear equalizer for ISI channels . . . . . . . . 99
8.2 Explorations . . . . . . . . . . . . . . . . . . . . . . . 100
8.3 Demonstration . . . . . . . . . . . . . . . . . . . . . . 102
8.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . 102
9 Channel Coding 103

9.1 Background . . . . . . . . . . . . . . . . . . . . . . . . 103
9.1.1 Channel capacity . . . . . . . . . . . . . . . . . 103
9.1.2 Channel coding . . . . . . . . . . . . . . . . . . 104
9.1.3 Syndrome decoding . . . . . . . . . . . . . . . 107
9.2 Explorations . . . . . . . . . . . . . . . . . . . . . . . 109
9.3 Demonstration . . . . . . . . . . . . . . . . . . . . . . 111
9.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . 112
10 OFDM Part 1 113

10.1 Background . . . . . . . . . . . . . . . . . . . . . . . . 114
10.1.1 Frequency-domain equalization . . . . . . . . . 115
10.1.2 Frequency-domain interpretation . . . . . . . . 118
10.2 Explorations . . . . . . . . . . . . . . . . . . . . . . . 119
10.3 Demonstration . . . . . . . . . . . . . . . . . . . . . . 120
10.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . 120
11 OFDM Part 2 121

11.1 Background . . . . . . . . . . . . . . . . . . . . . . . . 122
11.1.1 Orthogonal frequency division multiplexing . . 122
11.1.2 Channel estimation . . . . . . . . . . . . . . . . 126
11.2 Explorations . . . . . . . . . . . . . . . . . . . . . . . 127
11.3 Demonstration . . . . . . . . . . . . . . . . . . . . . . 128
11.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . 128
vi CONTENTS
12 Adaptive Processing 129

12.1 Background . . . . . . . . . . . . . . . . . . . . . . . . 130
12.1.1 Carrier phase errors . . . . . . . . . . . . . . . 130
12.1.2 Costas loop . . . . . . . . . . . . . . . . . . . . 131
12.1.3 Decision-directed phase tracking . . . . . . . . 133
12.2 Explorations . . . . . . . . . . . . . . . . . . . . . . . 134
12.3 Demonstration . . . . . . . . . . . . . . . . . . . . . . 137
12.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . 138
A Software Design Suggestions 139
B RF Experiments 141
B.1 Low-cost DIY Modulation . . . . . . . . . . . . . . . . 141
B.1.1 AM . . . . . . . . . . . . . . . . . . . . . . . . 141
B.1.2 FM . . . . . . . . . . . . . . . . . . . . . . . . 143
B.2 Commerical SDR Solutions . . . . . . . . . . . . . . . 145
C Functions 147
plottf . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
firlpf . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
char2psk template . . . . . . . . . . . . . . . . . . . . . . . 152
psk2char template . . . . . . . . . . . . . . . . . . . . . . . 153
srrc . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
eyediagram . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
makepilots . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
packetdetect . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
Glossary 159
Bibliography 163
Index 167
Preface
This text is intended as a student guide for hands-on exploration of

physical layer communication. Through a sequence of guided explo-
rations, students design and implement a baseband digital communi-
cation system with modulation to an acoustic carrier frequency. The
acoustic operation allows students to hear, see, and wirelessly trans-
mit signals using readily available, low-cost hardware, such as a PC
with sound card or a smartphone. Acoustic operation, while read-
ily accessible, nonetheless presents a student with the channel im-
pairments and synchronization issues encountered in radio frequency
systems.
Our approach is based on experiences, both successful and dis-
appointing, with several alternative pedagogical choices, including
simulation-based laboratory instruction, direct digital conversion to
a radio frequency carrier, and commercially available radio frequency
platforms. Our preference for an acoustic modem derives from four
goals. First, we seek a very low-cost pathway to provide labora-
tory experiences to students using ubiquitous equipment. An em-
phasis on cost is heightened by consideration of resource challenges
encountered world-wide and by Ohio State collaborations1 in Hon-
duras and Colombia. Second, in order to provide meaningful stu-
dent experiences in a short, seven-week format, we seek a balance
between investing time in tools versus concepts. We choose to tilt
the balance in favor of concepts, adopting familiar software tools for
a rapid learning curve. Third, we seek to require students to en-
gage in systems-level consideration of the full modem architecture,
rather than week-by-week operate in a black-box world of single sub-
steps. We are hopeful that we have achieved a proper balance in
this regard. Fourth, we seek to develop instructional materials that
are somewhat independent of specialized hardware platforms which
1
See https://osuhe.engineering.osu.edu/.
vii
viii PREFACE
inevitably experience rapid successions of updates or replacements.

At The Ohio State University, this text accompanies a laboratory
course, ECE 5007, which meets three hours weekly during the second
half of each semester. The laboratory course was introduced in con-
junction with our university’s transition from quarters to semesters.
Undergraduate students enter the laboratory course with a semester-
long introduction to signals and systems and co-requisite enrollment
in a digital communication lecture course. The laboratory course
uses Chapters 1 through 7 as guided learning exercises, reinforc-
ing and exploring concepts abstractly presented in a typical digital
communication lecture course. The second term schedule facilitates
coordination of the laboratory topics with a lecture syllabus based
on any of the many excellent digital communication textbooks, e.g.,
[10, 22, 31, 27, 18, 13]. The scope of each chapter has been purpose-
fully selected with the aim that students successfully implement a
portion of an acoustic digital modem during a three hour session.
The text supports other learning experiences beyond the Ohio
State course offering. First, the chapters are written with the in-
tent of allowing self-study. A brief review of concepts is presented
in each chapter, with emphasis on intuition and algorithm imple-
mentation; more detailed and rigorous treatments of topics can be
found in referenced sources. Second, the approach can be adopted for
use in high school STEM education. Pilot high school projects have
been taught in Annandale-on-Hudson, New York and Seattle, Wash-
ington by Taylor Williams, a Knowles Science Teaching Foundation
Fellow. Third, additional chapters are included in the text to accom-
modate a full semester laboratory course; additional topics include
forward error correction coding, equalization of frequency-selective
fading channels, multi-carrier modulation, and adaptive processing.
Further, the acoustic signal can provide an intermediate frequency
signal for radio frequency modulation; several radio frequency hard-
ware options are discussed in Appendix B.
The guided explorations challenge students to design and imple-
ment a digital modem operating at audio frequencies. The compu-
tations for the software-defined modem are performed in a high-level
language, such as Matlab; the exercises presented in the text have
been run in Matlab .
The basis for our acoustic modem approach to a digital commu-
nication laboratory comes from a senior design project conducted
some years ago at Ohio State. Undergraduate teams were judged
ix
on communication system performance measured as bits per second

per Hertz per dollar per decameter; transmission was at 918 MHz, an
acceptable error rate was specified, and a laptop or smartphone was
considered a zero-cost option for baseband processing. The open-
ended design project elicited solutions ranging from high-gain di-
rectional antennas to merely radiating with paper clips, and stu-
dents typically chose PSK or OFDM designs. The combination of
systems-level design and detailed subroutine development was judged
a strength of the experience.
The authors thankfully acknowledge the insights and contribu-
tions we have gathered from a variety of sources. The approach is
heavily influenced by the digital communication textbook by Rick
Johnson and Bill Setheras [18] and course notes by Phil Schniter
[31]. We have benefited from the laboratory courses taught by Steven
Tretter [37] and Robert Heath [15]. We are grateful to the teaching
assistants and students who embraced the senior design challenge of
building a 918 MHz modem; particular thanks go to Adam Margetts,
Aditi Kothiyal, Gene Whipps, Marshall Haker, and Aaron Wise.
We have learned much from Rohit Aggarwal’s development of exer-
cises for the National Instruments RF-RIO board. We thank Abigail
Gillett, Adam Rich, and Philip Chen for development of mobile apps
for acoustic transmission. Finally, we are grateful to The Mathworks,
Inc. for hosting course materials at Matlab Courseware.
This text and associated course materials are freely available in
electronic format at Mathworks Matlab Courseware, iTunesU, and
OpenStax CNX. Additionally, this text is available in hardcopy from
Lulu Press. An acoustic transmitter is available as a free mobile app
at the iTunes App Store and Google Play.
Lee C. Potter
Riverlea, Ohio
Yang Yang
San Diego, California
x PREFACE
CHAPTER 1
Introduction
The goals of this introductory exercise are to review basic concepts

from signals and systems and to introduce the hardware and soft-
ware tools to be used during the course. Concepts reviewed include
sampling, aliasing, sinc reconstruction, low-pass filtering, frequency
upconversion and frequency downconversion. The chapter begins
with a survey of physical layer communication and discussion of the
software-defined radio concept.
1.1 A Physical Layer Model

Figure 1.1 displays a digital communication processing chain for both
the transmitter and receiver. In the seven layer Open Systems In-
terconnection model, the physical layer (PHY), or Layer 1, defines
the means of transmitting raw bits over a physical link. For digital
communication, the chain begins and ends with a binary message;
this binary message may represent, for example, speech, audio, an
image, video, or high-definition television. At the transmitter, bits
are encoded for error protection and converted to a sequence of dis-
crete symbols. Symbols must be converted to a waveform suitable
for transmission. In a process known as “pulse shaping,” symbols are
converted to a sampled-data baseband message; then, the sampled
data are converted to an analog voltage signal through a digital-
to-analog converter. This conversion often also entails a frequency
shift to an intermediate frequency. The intermediate frequency sig-
nal is then modulated to a frequency that is matched to the desired
transmission channel, such as a licensed band for commercial radio
1
2 CHAPTER 1. INTRODUCTION
frequency transmission.
The receiver processing chain mirrors the transmission chain, but
with added complexity to handle synchronization of the receiver to
the transmitter and to accommodate signal distortions, or impair-
ments, due to transmission or resulting from thermal noise at the
receiver. Receiver design to mitigate these impairments is consid-
ered in Chapter 5 and Chapter 6; additional techniques to overcome
channel impairments are presented in Chapters 8 through 12.
Symbols are illustrated in Figure 1.2(a) for the case of the quad-
rature phase shift keying (QPSK) symbol constellation, also known
as 4-ary quadrature amplitude modulation (4-QAM). In QPSK, pairs
of bits are mapped to one of four equal energy symbols. The map-
ping of bits to symbols and the decoding of noisy symbols back to
bits are considered in Chapter 3. A typical impulse response used
for pulse shaping is illustrated in Figure 1.2(b). The role of the pulse
shaping filter and the associated design trade-offs are considered in
Chapter 4. Moving down the transmission chain in Figure 1.1, the in-
phase channel of a 47-bit baseband message is shown in Figure 1.2(c),
and the corresponding bandpass signal with acoustic carrier fre-
quency 1800 Hz is shown in Figure 1.2(d).
In this laboratory course, students are guided through a sequence
of directed explorations, culminating in the design and implementa-
tion of an acoustic modulator/demodulator (modem). The acoustic
modem implemented in this course works much like a fax modem.
1.2 Software Radio

A radio is any kind of device that wirelessly transmits or receives
signals at the radio frequency (RF) wavelengths of the electromag-
netic spectrum. Radios are found in many products, such as cell
phones, tablet computers, automobiles, televisions, spectrometers,
and magnetic resonance imagers. A software-defined radio (SDR)
can be defined as a radio in which some or all of the physical layer
functions are implemented through modifiable software operating on
programmable processing technologies [32, 12].
Conventional hardware-based radio devices provide operational
flexibility only by physically changing hardware components. In con-
trast, a SDR provides an inexpensive multi-mode, multi-functional,
multi-protocol device that can be modified using software upgrades
ã[n] Pulse Shaping m̃[k] m̃(t) QAM Modulation s(t)
Channel Symbol Add Pilot Digital Baseband Analog Baseband
Bits Bits Symbols Symbols L gtx [k] DAC Re RF Signal
Coding Mapping Sequence Signal Signal
j2πfct
e Tx Amp
1.2. SOFTWARE RADIO
Matched Filtering
ỹ[n] ṽ[k]
and Symbol Sync ṽ(t) QAM Demodulation r(t)
Channel Symbol Digital Baseband Analog Baseband
Bits Bits Symbols Frame Sync L grx [k] Signal ADC LPF RF Signal
Decoding Mapping Signal
Pilot sequence -j2π(fc+fΔ)t+φ Rx Amp

Symbol Timing e
Detection Recovery
Frequency Sync
Channel
Symbols Equalization
Figure 1.1: Physical layer communication processing chain for transmission and reception.
3
to provide new features to existing systems. The steady advancement

of digital technologies has provided digital sampling rates and pro-
cessing speeds to process both baseband and intermediate frequency
signals. SDR may soon provide interoperability of communication
devices in public safety and emergency response scenarios.
A typical SDR architecture is shown in Figure 1.3. For opera-
tion at frequencies above the UHF band, current SDRs use mixers at
the front end to perform analog upconversion and downconversion to
and from the desired frequency band. On receive, the wireless signal
induces on the antenna a current, which is boosted by a low-noise am-
Im Pulse Shaping Filter

(01) (00) 0.15
0.1
Re
amplitude
0.05
(11) (10)
4-QAM −0.05
−3 −2 −1 0 1 2 3
Symbol intervals
(a) (b)
Baseband Signal, I Channel Bandpass Signal
0.25 0.4
amplitude
0.2 0.2
0.15 0
0.1 −0.2
amplitude
0.05 −0.4
0 0.05 0.1 0.15 0.2 0.25
0 time(s)
−3
x 10
−0.05 6
magnitude
−0.1
4
−0.15
2
−0.2
−0.25 0
0 0.05 0.1 0.15 0.2 0.25 −4000 −3000 −2000 −1000 0 1000 2000 3000 4000
time(s) frequency (Hz)
(c) (d)
Figure 1.2: Illustrations of (a) an example symbol constellation; (b)

pulse shaping filter impulse response; (c) baseband signal, I channel;
and, (d) bandpass signal for fc = 1800 Hz.
1.2. SOFTWARE RADIO 5
plifier and filtered before downconversion to in-phase (I) and quad-

rature (Q) channels at an intermediate frequency (IF). The IF signal
is then directly sampled by an analog-to-digital converter (ADC).
Digital downconverters (DDCs) then further downconvert the signal
via multiplication by a signal generated by a numerically controlled
oscillator (NCO). Multiplication results in signals at the sum and dif-
ference of frequencies. The sum-of-frequencies signal is filtered out,
and the difference-of-frequencies signal provides the baseband wave-
form, which is decimated by a factor of L, meaning only every Lth
sample is retained. The decimated signal lowers the sample rate to
the baseband bandwidth, simplifying data processing yet retaining
the information carried in the signal.
A similar set of steps occurs on transmission. Digital upconver-

sion (DUC) numerically mixes a baseband signal to an IF carrier
frequency. The DUC, like the DDC at the receiver, is commonly
packaged as a single integrated circuit. The IF signal signals in the I
and Q channels feed an analog quadrature mixer that produces the
RF signal, which is amplified and fed to the antenna.
Antenna
ADC DDC
LNA LPF
Data
out
ADC DDC
LPF
Local oscillator Baseband

synthesizer processor
(DSP, FPGA, etc.)
DAC DUC
PA LPF
Data
in
DAC DUC
LPF
Figure 1.3: A typical SDR architecture.

1.3 Background
This section provides a brief review of concepts explored in the labo-
ratory exercises. First, sampling and aliasing are reviewed. Second,
upconversion and downconversion are defined for the real-valued sig-
nal case. Complex-valued sampling and quadrature modulation are
explored in Chapter 2.
1.3.1 Sampling and aliasing

The basic intuition of sampling and aliasing is evident in Figure 1.4:
given only samples of a sinusoidal waveform, there exist tones of
infinitely many possible frequencies that pass through the samples.
The frequencies of these possible interpolating waveforms differ by a
multiple of the sampling frequency. In the figure, the solid line shows
m(t) = sin(2π20t), a sine wave at 20 Hz. Samples, shown by dots-on-
sticks, are taken every 0.01 s, for a sampling rate of fs = 100 samples
per second (sps). The top panel shows that sin(2π(20 + fs )t) =
sin(2π120t) also is consistent with the same set of samples; likewise,
the bottom figure illustrates that sin(2π(20−fs )t) = sin(2π(−80)t) =
sin(2π80t + π) is another tone yielding the same samples.
The Fourier spectrum, X(f ), of the sampled signal contains the
P
sum of these possibilities. Let x(t) = m(t) ∗ n δ(t − nT ) be the
sampled signal; then,
∞
1 X
X(f ) = M (f − kfs ), (1.1)
T k=−∞
where M (f ) is the spectrum of the original analog signal, m(t). Each

shifted replica of the spectrum M (f ) in Equation 1.1 is called an
image. To eliminate confusion among the possible signals yielding a
sampled waveform, the Nyquist criterion stipulates that M (f ) = 0
for all |f | ≥ fs /2; that is, the Nyquist criterion stipulates that the
sampling rate yield at least two samples per cycle for any frequency
present in m(t).
More generally, bandpass sampling of a real-valued signal m(t)
requires that the two-sided bandwidth of the spectrum M (f ) be con-
fined to a single Nyquist zone,
(k − 1)fs kfs
≤ |f | ≤ , for some positive integer k. (1.2)
2 2
1.3. BACKGROUND 7
0.5
amplitude
−0.5
−1
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
time (seconds)
0.5
amplitude
−0.5
−1
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
time (seconds)
Figure 1.4: Illustration of sampling and aliasing.
In this manner, an alias of M (f ) appears with frequency support

confined to f ∈ [−fs /2, fs /2).
1.3.2 Frequency upconversion
Upconversion (amplitude modulation or mixing) of a real-valued

message m(t) is given by
s(t) = m(t) cos(2πfc t) (1.3)
and serves to translate the message to a desired frequency band cen-

tered on the carrier frequency, fc . Upconversion is graphically de-
picted in Figure 1.5. From the Euler identity, ejx = cos x + j sin x;
therefore,
1 n j2πfc t o
cos(2πfc t) = e + e−j2πfc t . (1.4)
2
The spectrum, S(f ), is found via the Fourier transform

Z ∞
S(f ) = m(t) cos(2πfc t)e−j2πf t dt
−∞
Z Z
1 ∞ 1 ∞
= m(t)e−j2π(f −fc )t dt + m(t)e−j2π(f +fc )t dt
2 −∞ 2 −∞
1 1
= M (f − fc ) + M (f + fc ). (1.5)
2 2
Because m(t) is real-valued, M (f ) is conjugate symmetric about f =

0, implying that the upconverted spectrum for f > 0 is conjugate
symmetric about fc and that the spectrum lower sideband (below fc )
is redundant with the upper sideband (above fc ). The magnitude
spectrum is illustrated in Figure 1.6.
m(t) × s(t)
∼ cos(2πfc t)
Figure 1.5: Amplitude modulation performs a translation in fre-

quency.
1 |M(f)| 1 |S(f)|
2W
1/2
f f
-W W -fc fc
Figure 1.6: Magnitude spectrum resulting from amplitude modula-

tion.
1.3. BACKGROUND 9
1.3.3 Frequency downconversion

With fc known, downconversion, or amplitude demodulation, can be
accomplished by mixing and low-pass filtering
v(t) = LPF{s(t) × 2 cos(2πfc t)}

= LPF{m(t) 2 cos 2 (2πfc t)}
| {z }
1+cos(4πfc t)
= LPF {m(t) + m(t) cos (2π{2fc }t)}
= m(t) (1.6)
This process in depicted in Figure 1.7 and Figure 1.8.
r(t) × LPF v(t)
∼ 2 cos(2πfct)
Figure 1.7: Amplitude demodulation.
In Equation 1.6 and Figure 1.7, the low-pass filter (LPF) has a
passband cutoff Bp ≥ W Hz and stopband cutoff Bs ≤ (2fc − W ) Hz,
where W is the one-sided bandwidth of the baseband signal, m(t).
The filter’s magnitude response is illustrated in Figure 1.8.
1
1/2
f
−2fc − W −2fc −2fc + W −Bs −Bp −W W Bp Bs 2fc − W 2fc 2fc + W
Figure 1.8: Low-pass filter (dashed line) used in amplitude demodu-

lation.
When the receiver oscillator has frequency offset, f∆ , or phase

offset, φ, relative to the upconversion, the demodulated signal is
distorted:

v(t) = LPF m(t) cos(2πfc t) × 2 cos(2π(fc + f∆ )t + φ)
| {z }
cos(2πf∆ t+φ) + cos(2π(2fc +f∆ )t+φ)
= m(t) cos(2πf∆ t + φ) . (1.7)
| {z }
time-varying attenuation
The time-varying attenuation can be viewed as an upconversion of

m(t) to a frequency equal to the offset, f∆ . Estimation and correction
of this offset will be considered in Chapter 2 and Chapter 5.
Interpolation
Interpolation is a digital filtering procedure that increases the sam-
pling rate by an integer factor, L:
∞
X sin(π(k − nL)/L)
y[k] = x[n] . (1.8)
n=−∞
π(k − nL)/L
The processing to produce the interpolated samples can be viewed

as two steps: upsampling (i.e., interlacing L − 1 zeros between con-
secutive samples of x[n]) followed by low-pass filtering with gain L
π
and cutoff frequency L radians per sample. See Figure 1.9.
x[n] y[k]
L
LPF
Figure 1.9: Block diagram for ideal interpolation.
DAC: sinc interpolation

The ideal digital-to-analog conversion (DAC) is implemented by sinc
interpolation, which is illustrated in Figure 1.10:
∞
X sin(π(t − kT )/T )
x(t) = x[k] , (1.9)
k=−∞
π(t − kT )/T
where T is the sampling period for x[k]. The summation is the

convolution of a low-pass filter impulse response (i.e., a scaled sinc
1.3. BACKGROUND 11
function) with the impulse train of samples, x[k]δ(t − kT ). This

summation is depicted in the figure, where the solid circles denote
the sample values {−2, 2, 4, 4, 2}, the dashed lines denote each shifted
and scaled sinc function, as in Equation 1.9, and the summation is
shown by the solid line. The zero crossings of the sinc function,
which occur every T seconds, ensure the faithful interpolation of the
original samples.
2
amplitude
−1
−2
−3
−0.5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
time (seconds)
Figure 1.10: Illustration of sinc interpolation.
Decimation
Decimation is a digital filtering procedure that decreases the sam-

pling rate. Given an original sampling rate of fs sps, decimation by
a positive integer L produces a new sampling rate of fs /L. The signal
is first low-pass filtered, typically in stages, with a cutoff frequency
fs
of 2L ( Lπ radians per sample). The filtering prevents aliasing from
resulting when the signal is downsampled. Only every Lth sample of
the filtered output is computed and retained. Refer to Figure 1.11.
Note that the low-pass filtering is an averaging that can reduce noise,
particularly noise due to quantization error.
L
LPF
Figure 1.11: Block diagram for decimation by L; the anti-aliasing

low-pass filter has cutoff frequency π/L radians per sample.
1.4 Explorations
Here, as in each chapter, a guided set of explorations is presented.
Useful commands are introduced and used in example snippets of
code to reinforce and visualize key concepts. Chapter by chapter,
students build towards a fully functional acoustic modem.
Each chapter is designed for a single laboratory session with a
brief report and a demonstration. In subsequent chapters, exercises
are increasingly less scripted, with freedom for creative detours and
experimentation. Students are encouraged to be brief and direct
in reporting; to this end, questions to be answered in the report
are enumerated in the margins. The recommended demonstration is
detailed in a separate subsection.
Microphone input
For Matlab to recognize a sound recording device in some Windows
configurations, an external microphone must be inserted into the mi-
crophone jack of the PC tower before starting Matlab. Right-click
on the volume controls in the system tray (a speaker icon typically
located at the bottom right of the Windows screen), select “Audio
Devices” and select the “Recording” tab. Ensure that there is some
reaction from the microphone as shown by the volume bars beside the
microphone entry. If there is no reaction, then go to the properties
of the microphone (double click) and adjust the microphone volume
and/or boost until an input reaction is observed. More generally,
from right-click on the speaker icon, you can select playback devices,
recording devices, and volume control options. From these command
windows, you can select a playback device and a recording device.
By clicking on the icon for a device, you can enable the device, check
device properties, and set sound levels.
1.4. EXPLORATIONS 13
Workflow suggestions
For numerical and graphical explorations, several practices may be
helpful. First, students are advised to create a memory folder for lab-
oratory work; download files and select the folder using the “browse
for folder” icon in the Matlab tool bar. For example, the explo-
rations in Chapter 1 use the function plottf.m, which is given in
Appendix C. A useful practice is to create a new script file (available
at the top left of the menu bar) for each chapter. An editor window
can then be used to copy and execute code snippets, write new com-
mands, or edit and re-run codes. For each chapter, the code snippets
appearing in the chapter are available as a stand-alone .m script, such
as Chapter1snippets.m. All scripts and functions are available for
download at OpenStax CNX and Mathworks Matlab Courseware.
Students are encouraged to embrace the development of engi-
neering skills, and software design skills in particular, that are a
part of this laboratory experience. To this end, we highlight several
suggestions in Appendix A. Subsections of code developed in each
laboratory section will eventually be integrated into a single work-
ing modem; thus, purposeful attention to a few simple practices will
greatly simplify work as the lessons progress.
1.4.1 Getting started with audio I/O

In this experiment, you will familiarize yourself with recording and
playing audio signals.
Recording
Four commands allow you to access the sound card hardware from
within Matlab to record audio signals.
>> recObj = audiorecorder(Fs,Nbits,nChannels) sets the

sample rate Fs (in samples per second), the number of bits per
sample Nbits, and the number of channels nChannels. The de-
faults are Fs=8000, Nbits=8, and nChannels=1.
>> recordblocking(recObj,length) records sound for length
seconds and does not return control until recording completes.
>> getaudiodata(recObj) creates an array to store the recorded
signal values.
>> get(recObj) queries the properties of the audiorecorder ob-
ject recObj.
Valid values of the sampling rate depend on the specific audio hard-
ware installed. Typical values supported by most sound cards are
8000, 11025, 22050, 44100, 48000, and 96000 samples per second.
The code snippet in Table 1.1 provides an example. Typing your
own code as you do example exercises will help in building and re-
taining familiarity with simple commands. In addition, code snippets
are available for download; see links given in Appendix C. If, in exe-
cuting the function audiorecorder, you encounter the error message
no audio device found, then exit Matlab, unplug and re-plug the
microphone into the PC input jack, and re-start Matlab.
Playing audio
Similarly, two commands allow a user to play audio through a sound
card from within Matlab.
>> playObj = audioplayer(Y,Fs,Nbits) creates an audio-

player object for the signal Y, using sample rate Fs and Nbits bits
per sample. The function returns a handle to the audioplayer ob-
ject, playObj. The audio signal is represented by a vector (mono)
or two-dimensional array (for stereo). The default is Nbits=16.
>> play(playObj) plays audio from beginning to end.
The following code snippet provides two examples.

Suggested explorations:
(a) Follow the example in Table 1.1 to record and plot 2 seconds
of speech.
(b) Follow the examples in Table 1.2 to play a tone and the recorded
speech.
1.4.2 Spectral content of signals
Visualization can be a powerful tool for understanding signals and

the effects of systems on signals. Here we consider use of the routine
plottf.m to view a sampled signal in both the time and frequency
domains. The custom routine is given in Appendix C.
>> plottf(x,Ts) plots the time-domain samples in vector x, as-

suming that Ts is the sampling interval in seconds, and also plots
the Riemann-sum approximation of the Fourier transform between
the frequencies of -1/(2Ts) and 1/(2Ts) Hertz.
>> plottf(x,Ts,’t’) plots only the time-domain signal.
>> plottf(x,Ts,’f’) plots only the frequency-domain signal.
In all cases, plottf returns handles to the graphical objects; as

with any Matlab command, display to the command window is
suppressed by a semi-colon.
Suggested explorations:
(c) Use plottf to view the speech signal recorded in Experiment

1.4.1 and stored as myRecording. Note that plottf asks for
the sampling interval, which is the reciprocal of the sampling
rate. To start a new figure window, rather than overwrite the
most recently accessed figure, use the figure command. The
error statement “Undefined function ‘plottf’” indicates that the
routine plottf.m is neither in the current working directory
nor specified in the current search path (see help path).
(d) Use the zoom tool in the figure to look more closely at speech in Q1.1
the time domain. What qualitative differences do you observe
between vowels and consonants? Is a DC bias present in your
recorded speech signal?
(e) Look more closely at speech in the frequency domain. What Q1.2
are the lowest and highest frequencies displayed in the graph,
and why? What band of frequencies is non-negligibly occupied
by the speech signal? How would a DC bias be evident in
the frequency domain? (Hint: use the zoom option with the
frequency domain plot to observe X(f ) for f = 0.)
(f) With sampling rate fixed at Fs=16000 sps, plot 40 milliseconds

of the sinusoidal signal, cos(2πfc t), for frequencies 1000, 7000,
9000, 16000 and 17000. For an example, refer to Table 1.2.
View the signal in both the time and frequency domains using
plottf. What do you observe, and why?
(g) With sampling rate fixed at Fs=8000 sps, specify four values
Q1.3 of f0, 0 < f0 < 16000, so that cos(2*pi*f0*(0:Fs/2)/Fs)
sounds like one-half second of a 440 Hz tone.
(h) Experiment with the audioplayer command to discover what
sampling rates are supported by your sound card hardware; an
unsupported rate will return an error message when the play
command is invoked.
1.4.3 Low-pass filtering

For design of a low-pass filter, the custom function firlpf.m is pro-
vided in Appendix C. The command-line function requires the syn-
tax
h = firlpf(Lh,Fpb,Fsb,Fs);
where h is the impulse response of length Lh, Fpb is the passband
edge (in Hertz), Fsb is the stopband edge (in Hertz), and Fs is the
sampling frequency (in samples per second). The routine computes
the least-squares fit to the ideal magnitude response illustrated in
Figure 1.12. The frequency-selective filtering using the impulse re-
sponse h can be implemented via the convolution operation by the
command conv or filter.
(i) Create a low-pass noise signal to experiment with frequency
translation. The following code snippet illustrates use of a
Gaussian random number generator and a low-pass filter.
Fs=16000;% 16000 sps

t=0:1/Fs:0.25;% 250 milliseconds segment
whitenoise=randn(size(t));% Gaussian random vbls
figure;plottf(whitenoise,1/Fs);% view signal
title(’White Noise’)
%filter design; see custom firlpf.m in Appendix
h = firlpf(49,1200,1800,Fs);
pinknoise=filter(h,1,whitenoise);% convolution
figure;plottf(pinknoise,1/Fs);% view signal
title(’Bandlimited Noise’)
(j) View both your whitenoise and pinknoise signals in the time
and frequency domains. Comparing the two signals, what do
you observe in each domain?
(k) [Optional] Repeat step (j) for several smaller values of filter
length small than the Lh=49 used above. What do you observe?
(l) [Optional] Add a tone at 2100 Hz to your whitenoise signal.

Observe the resulting signal before and after the low-pass filter.
1.4.4 Amplitude modulation

In this experiment, you will explore amplitude modulation illustrated
in Figure 1.5 and gain familiarity with several basic instructions.
(m) Modulate your signal from step (i) to a carrier frequency fc =

4 kHz and view in time and frequency. Note that to multiply
sampled signals point-by-point, the Matlab syntax “.*” should
be used. From step (i) the sampling rate is 16000 sps.
LPF: Lh=39; Fpb=1000; Fsb=2000; Fs=8000

50
0
Magnitude (dB)
−50 One−half
Sampling rate
Passband Stopband
Edge, Fpb Edge, Fsb
−100
−150
0 500 1000 1500 2000 2500 3000 3500 4000
Frequency (Hz)
Linear Phase
Phase (degrees)
−500
−1000
−1500
−2000
0 500 1000 1500 2000 2500 3000 3500 4000
Frequency (Hz)
Figure 1.12: Example of FIR low-pass filter design using firlpf.

fc=4000;%4000 Hz
s = cos(2*pi*fc*t) .* pinknoise;
figure;plottf(s,1/Fs);
title(’Modulated Bandpass Noise’)
Q1.4 (n) Keep Fs=16000. Experiment with various values of the carrier
frequency, fc, from the previous step. What do you observe
for carrier frequencies 6, 500 ≤ fc ≤ 16, 000 Hz?
Q1.5 (o) For the bandlimited noise pinknoise above, suppose you wish
to modulate to a carrier frequency of 19, 000 Hz. Suggest an
appropriate sampling rate.
1.4.5 AM demodulation
For frequency downconversion, follow the diagram in Figure 1.7. Re-
call 2 cos(α) × cos(β) = cos(α − β) + cos(α + β). The role of the
low-pass filter is to remove this second term, the so-called “double
frequency” term.
(p) Demodulate the signal s constructed in step (m) above. View

the demodulated signal in both the time and frequency do-
mains. For the low-pass filter, select a filter length Lh=49.
What are appropriate values for the passband and stopband
edge frequencies, Fpb and Fsb (refer to Figure 1.8)?
(q) Omit the filter convolution step in the demodulator, and again
view the result. What do you observe?
(r) [Optional] For demonstration to the instructor, implement the

following amplitude modulation and demodulation steps. Use
plottf to view each step. The aim is to reinforce the principles
and techniques explored above.
• Record 900 milliseconds of speech at a sampling rate of

your choice. (Hint: read all steps below before making
your choice).
• Amplitude modulate your recorded speech signal using a
carrier frequency of fc = 12, 000 Hz. Use play to listen to
the modulated speech signal.
1.5. DEMONSTRATION 19
• Open simultaneously a second Matlab command win-

dow. Use one window to play the modulated speech sig-
nal, and the other to record the audio transmission.
• Demodulate the received AM signal and play the recov-
ered audio.
1.5 Demonstration
For demonstration, use the Communication Laboratory mobile app
to transmit an acoustic signal using the app’s default settings. The
app is freely available at iTunes and Google Play; a description of
the app is found in Section 7.1.1.
Record the transmission with sampling frequency Fs=44100. View
the received signal using the plotttf.m function. Observe and re-
port the following parameters from your graph.
1. What is the approximate time duration of the transmission, in

milliseconds?
2. What, approximately, is the carrier frequency?
3. What, approximately, is the two-sided bandwidth of this band-

pass signal?
1.6 Summary
In Chapter 1, basic signals and systems concepts have been reviewed:
sampling, aliasing, interpolation, low-pass filtering, frequency upcon-
version and frequency downconversion. The exercises introduced the
recording and playing of audio-rate signals. And, the explorations
served to introduce the commands and conventions summarized in
Table 1.3.
Table 1.1: Code snippet for audio recording
%% Create an audiorecorder object with 8000 sps and

% a single channel (mono); view its properties
Fs=8000;
recObj = audiorecorder(Fs, 16, 1);
get(recObj)
%% Collect a sample of your speech with a microphone;

% and, plot the signal data
% Record your voice for 2 seconds. Use display and

% pause as aids to control the start of the recording.
Trec = 2; %2 second record time
disp(’Press Enter to start recording.’)
pause;%wait for keystroke
disp(’Recording.’)
recordblocking(recObj, Trec);
disp(’End of Recording.’)
% Store data in double-precision array.

myRecording = getaudiodata(recObj);
%% Plot the waveform.

t = (0:length(myRecording)-1)/Fs; %sample times (sec)
plot(t,myRecording);
1.6. SUMMARY 21
Table 1.2: Code snippet for audio play
%% Example: play a one second tone at 440Hz

Fs=8000; % sampling rate of 8000 sps
t = 0:1/Fs:1; %list of sampling times
signal = cos(2*pi*440*t);
% Create audioplayer object

ToneObj = audioplayer(signal,Fs);
% Play the sound

play(ToneObj)
%% Example: play back the speech recorded previously

play(recObj);
Table 1.3: Summary of commands introduced in Chapter 1.

Command Description
audiorecorder create audio recording object
audioplayer create audio player object
recordblocking record sound
getaudiodata store recorded sound samples
play play an audio object
get display object properties
disp display text to command window
pause pause execution awaiting key stroke
plot make a line plot
figure initiate new figure window
title place title on figure
: construct lists
; suppress output
* matrix multiply
.* array multiply
/ division
size returns size of a data array
cos cosine function
plottf convenient plotting routine
randn Gaussian random number generator
firlpf least-squares design of a low-pass filter
conv convolution
% remainder of line is a comment
%% creates an executable subsection of code
path get or set the search path
help display help text in command window
CHAPTER 2
Quadrature Amplitude Modulation
Chapter 1 reviewed basic signals and systems concepts; the exercises

also introduced the recording and playing of audio-rate signals. In
this lesson, quadrature amplitude modulation (QAM) and coherent
quadrature demodulation are introduced. Complex baseband is pre-
sented as a convenient mathematical notation for representing QAM
signals. Here we build upon the introduction in the previous chapter
to implement a QAM transmitter and receiver. The exercise also
introduces the practical impairments of frequency and phase offsets,
which are compensated in the receiver design to be implemented in
Chapter 6.
2.1 Background
2.1.1 Quadrature amplitude modulation
In quadrature amplitude modulation two real-valued messages are
modulated simultaneously, resulting in a bandpass spectrum that
is not conjugate symmetric about the carrier frequency, fc . The
modulation exploits the orthogonality of the sine and cosine functions
Z 1/(2fc )
sin(2πfc t) cos(2πfc t)dt = 0. (2.1)
−1/(2fc )
The quadrature modulator is shown in Figure 2.1. The modulator

has two “branches,” one for an in-phase signal, mI (t), that modulates
the cosine carrier and a second for a quadrature signal, mQ (t), that
modulates the sine carrier. The sine wave is 90 degrees (2π/4 radians)
phase delayed from the cosine, giving rise to the name “quadrature.”
23
24 CHAPTER 2. QUADRATURE AMPLITUDE MODULATION
The two modulated waveforms, sometimes called “I” and “Q” for
short, are combined to produce the modulated waveform, s(t), given
by
s(t) = mI (t) cos(2πfc t) − mQ (t) sin(2πfc t). (2.2)
mI (t) × “in-phase”
cos(2πfc t)
∼ + s(t)
sin(2πfc t) −
90◦
mQ (t) × “quadrature”
Figure 2.1: Quadrature amplitude modulator.
Ideal QAM demodulation is accomplished as shown in Figure 2.2.

For baseband signals {mI (t), mQ (t)} with one-sided bandwidth W Hz,
the lowpass filter in the demodulator has passband edge frequency
Bp ≥ W Hz and stopband edge frequency Bs ≤ (2fc − W ) Hz. (See
Figure 1.8.)
× LPF vI (t)
2 cos(2πfc t)
r(t) ∼
2 sin(2πfc t)
−
× LPF vQ (t)
Figure 2.2: Quadrature amplitude demodulator.

2.1. BACKGROUND 25
Simple trigonometric identities can be used to verify that, for

r(t) = s(t) (i.e, a noiseless and distortionless channel), the message
signals are perfectly recovered.
vI (t) = LPF{r(t) × 2 cos(2πfc t)}

= LPF mI (t) 2 cos2 (2πfc t) −mQ (t) 2 sin(2πfc t) cos(2πfc t)
| {z } | {z }
1+cos(4πfc t) sin(4πfc t)
= mI (t) (2.3)
vQ (t) = LPF{−r(t) × 2 sin(2πfc t)}

= LPF −mI (t) 2 cos(2πfc t) sin(2πfc t) +mQ (t) 2 sin2 (2πfc t)
| {z } | {z }
sin(4πfc t) 1−cos(4πfc t)
= mQ (t). (2.4)
Low-pass filtering removes the double frequency terms, cos(4πfc t)

and sin(4πfc t). For the coherent receiver described by Equation 2.3
and Equation 2.4, we require that the receiver oscillator is perfectly
synchronized to the transmitter oscillator: that is, the frequency, fc ,
and phase of the sine and cosine terms at the demodulator must
exactly match the frequency and phase at the modulator. When
the oscillators are not synchronized, there results a coupling at the
receiver between the I and Q components, as well as attenuation of
each.
2.1.2 Complex-baseband representation

The use of a complex-baseband signal representation provides sim-
plification in writing, analyzing, and programming quadrature am-
plitude modulation waveforms. In the complex-baseband form, the
in-phase and quadrature message signals are conveniently packaged
together as the real and imaginary components of a single, complex-
valued signal,
m̃(t) = mI (t) + jmQ (t),

ṽ(t) = vI (t) + jvQ (t). (2.5)
The superscript tilde is used to denote a complex-valued signal. Em-

ploying the complex-baseband notation, the modulation and demod-
ulation are very simply described, as seen in Figure 2.3.
Euler’s identity is invoked to verify the complex-baseband model

shown in Figure 2.3. For the modulator, observe
Re{m̃(t)ej2πfc t } (2.6)

= Re mI (t) + jmQ (t) cos(2πfc t) + j sin(2πfc t)
= mI (t) cos(2πfc t) − mQ (t) sin(2πfc t) = s(t). (2.7)
The demodulation structure is verified in a similar manner:
ṽ(t) = LPF{s(t) · 2e−j2πfc t }

= LPF{ mI (t) cos(2πfc t) − mQ (t) sin(2πfc t) · 2e−j2πfc t }

= LPF mI (t) ej2πfc t + e−j2πfc t e−j2πfc t

− mQ (t) je−j2πfc t − jej2πfc t e−j2πfc t

= LPF mI (t) 1 + e−j4πfc t − mQ (t) je−j4πfc t − j
= mI (t) + jmQ (t) = m̃(t). (2.8)
Frequency-domain interpretations of the steps in the modulation

and demodulation are given by the frequency spectra in Figure 2.3.
Observe the asymmetry of the passband signals about the carrier,
fc , and the corresponding asymmetry of the baseband signal’s mag-
nitude spectrum. Thus, the upper and lower sidebands are not re-
dundant, in contrast to AM. The Re operation at the modulator
denotes taking the real part of the complex-valued signal and can be
interpreted using the Fourier transform relation
1 F 1
Re{u(t)} = u(t) + u∗ (t) ←→ U (f ) + U ∗ (−f ) . (2.9)
2 2
The convenience of the complex-baseband representation results
in widespread use of complex-valued signal notation for passband
signals encountered, for example, in communication, radar, spec-
troscopy, and medical imaging.
2.1.3 Complex-baseband equivalent channel

Next, we augment the quadrature modulation model to include a
channel described by an impulse response, h(t), and additive noise,
w(t). This is depicted in the top panel of Figure 2.4. All signals
present in the top panel are real-valued, consistent with the physical
I and Q voltages present in a hardware implementation of QAM. The
2.1. BACKGROUND 27
QAM modulation QAM demodulation
m̃(t) s(t) r(t) ṽ(t)

Re
LPF
ej2πfc t 2e−j2πfc t
|M (f )| |R(f )|
|M (f − fc )| |2R(f + fc )|
|S(f )| |V (f )|
Figure 2.3: Quadrature amplitude modulation and demodulation are

simplified using a complex-baseband notation. The steps are illus-
trated by the frequency spectra shown beneath the block diagrams.
behavior of h(t) is of consequence only on the frequency band occu-

pied by the modulated signal, s(t). Thus, let hbp (t) be the bandpass
equivalent channel
(
H(f ), fc − W ≤ |f | ≤ fc + W
Hbp (f ) = (2.10)
0, else.
This equivalence is depicted in the middle panel of Figure 2.4. With

proper modulation and demodulation and commuting the low-pass
and channel filters, we arrive to the bottom panel of Figure 2.4 which
is known as the complex-baseband equivalent model.
The complex-baseband channel model is given by

(
Hbp (f + fc ), −W ≤ f ≤ W
H̃(f ) = (2.11)
0, else.
This model simplifies representation, implementation, simulation,

and analysis of the communication system.
I/Q modulation wideband I/Q demodulation

channel
mI (t) LPF vI (t)
cos(2πfc t) s(t) r(t) cos(2πfc t)
h(t)
_
sin(2πfc t) _ sin(2πfc t)
mQ (t) w(t) LPF vQ (t)
complex baseband bandpass complex baseband

modulation equivalent demodulation
channel
s(t) r(t)
m̃(t) Re hbp (t) LPF ṽ(t)
ej2πfc t w(t) 2e−j2πfc t
complex
baseband
equivalent
channel
m̃(t) h̃(t) ṽ(t)
w̃(t)
Figure 2.4: Complex-baseband equivalent channel model.
2.1.4 Coherent demodulation

For coherent demodulation, the local oscillator must have frequency
and phase synchronized with the transmitter. The impairment is in-
troduced here, and implementation of a frequency recovery algorithm
is the topic of Chapter 6.
2.1. BACKGROUND 29
Consider a receiver oscillator given by ej(2π(fc +f∆ )t+φ) with a fre-

quency offset, f∆ Hz, and a phase offset, φ radians. Using the con-
venient complex-baseband representation of Figure 2.3 we have that
the received message is
ṽ(t) = m̃(t)e−j(2πf∆ t+φ) . (2.12)
For sampled values indexed by k with sampling period Ts seconds
per sample, we have
ṽ[k] = m̃[k]e−j(2πf∆ Ts k+φ) . (2.13)
Thus, if there is a frequency offset between transmitter and re-
ceiver, then a time-varying phase error results. Define the phase er-
ror as the true transmitted phase minus the received signal’s phase;
then, the phase error, (2πf∆ t + φ), at the receiver is a linear function
of time, with slope proportional to the frequency offset. Figure 2.5
gives an example graph of this time-varying phase offset, which is
due to both the phase mismatch (intercept) and frequency mismatch
(slope).
True minus Received Phase

14
12
10
slope = 2 π f∆ Ts
Phase (rad)
2 Intercept = φ
0
0 500 1000 1500 2000
Time sample, k
Figure 2.5: A time-varying phase error φ + 2πf∆ Ts k, appears at the

receiver output due to a frequency offset, f∆ , and phase offset, φ.
2.1.5 Linear phase filters

The filters used at the transmitter and receiver will have linear phase,
meaning the filter imparts the same time delay, in samples, to every
frequency component of the discrete-time input signal. The linear

phase is characterized in the filter impulse response as symmetry
about the midpoint of the filter’s impulse response. The delay in the
output of a linear phase filter is one-half the filter order, and the or-
der is one less than the length of the filter impulse response. Thus, in
the notation of Section 1.4.3, the delay is (Lh-1)/2 samples. When
indexing into a filter output signal, this delay must be taken into ac-
count. For input s and impulse response h, the convolution output,
conv(s,h) exhibits transients at the beginning and end of the out-
put, due to zero boundary conditions used in the convolution sum.
A communication packet signal is typically measured with noise, and
the noise-only samples provide the boundary conditions.
Delay in a linear phase filter is illustrated in the following ex-
ample. The noiseless signal is a truncated square-root raised cosine
pulse, where the center of the pulse sits at index 100. The noise has
variance 5 × 10−5 . The noisy signal is then low-pass filtered with
cutoff of 1000 Hz. Here, the delay is 24 samples, which is one-half
the filter order of 48.
Fs=8000;
k =0:200;
varw=0.00005;
s = sqrt(varw)*randn(size(k)); %create noise
s(51:151)=s(51:151)+srrc(2,0.7,25); %signal+noise
%LPF order 48, passband edge 1000 Hz
h = firlpf(49,1000,2000,Fs);
y = conv(s,h); %convolution
figure;plot(k,s);
xlabel(’sample index’,’fontsize’,13)
axis([0 250 -0.1 0.3]);
title(’Filtering example’,’fontsize’,13)
k2=0:length(y)-1;
hold on;plot(k2,y,’r--’);hold off
xlabel(’sample index’,’fontsize’,13);
legend(’input’,’output’)
2.2 Digital Hardware [Optional]

In this section, optional background material is presented regarding
digital hardware implementation of quadrature amplitude modula-
2.2. DIGITAL HARDWARE [OPTIONAL] 31
Filtering example
0.3
0.25 Filter Delay
0.2
input
0.15 output
0.1
0.05
−0.05
−0.1
0 50 100 150 200 250
sample index
Figure 2.6: Illustration of delay imparted by a linear phase filter.
tion in radio frequency systems.

Advantages of digital implementation of upconversion and down-
conversion, versus analog mixers, are: (a) the ability to closely match
the in-phase and quadrature branches in amplitude, frequency, and
90 degree phase offset; and, (b) the ability to avoid local oscillator
leakage in the analog circuit. Thus, in many communication systems,
digital processing is used to produce an intermediate frequency (IF)
signal via quadrature modulation; the real-valued IF signal is then
mixed and filtered, via analog hardware, to create the RF signal.
Digital upconversion
To illustrate digital implementation, consider, for example, the Ana-

log Devices AD9857. The digital-to-analog converter (DAC) is com-
bined with a digital upconverter (DUC) on a single integrated cir-
cuit. The AD9857 integrates a high-speed direct-digital synthesizer
(DDS), a high-performance, high-speed 14-bit digital-to-analog con-
verter (DAC), clock multiplier circuitry, digital filters, and other DSP
functions onto a single chip, to form a complete quadrature digital
upconverter device. The AD9857 is intended to function as a univer-
sal quadrature modulator and agile upconverter, single-tone DDS,
or interpolating DAC for communication applications. The basic
operation is abstracted in Figure 2.7.
A direct digital synthesizer (DDS) is the core of the AD9857
and is a numerically controlled oscillator (NCO). A phase accumula-

tor provides a selectable modulus counter that increments with each
clock pulse; the increment is determined by a digital word, and the
value in the accumulator is used to index a sine-wave look-up table
to produce accurate, stable, and phase-continuous frequency tuning.
When used as a quadrature synthesizer, both cosine and sine out-
puts are produced, yielding excellent matching of in-phase (I) and
quadrature (Q) outputs.
Quadrature Modulator
I
Interpolator _ Inv Sinc DAC
Q
sin cos
DDS
Figure 2.7: Basic operation of a quadrature digital upconverter.
Digital downconversion
The AD6654 is a mixed-signal IF-to-baseband receiver consisting of
a 14-bit, 92.16M sps analog-to-digital converter (ADC) and a multi-
channel digital downconverter (DDC). It has been optimized for
wideband standards such as CDMA2000, UMTS, and TD-SCDMA.
The AD6654 is used as part of a radio system that digitally demodu-
lates and filters IF sampled signals. The basic operation is abstracted
in Figure 2.8.
2.3 Explorations
In this chapter, students explore quadrature modulation and the
complex-baseband signal representation. The quadrature modula-
tion and demodulation steps are illustrated on the right-hand side of
Figure 1.1. For the laboratory exploration, work through the guided
steps below using an acoustic carrier frequency. Recommendations
for preparing a brief, descriptive laboratory report are included in
the steps marked in the margins.
LPF L I
DAC
LPF L Q
-2sin 2cos
DDS
Figure 2.8: Basic operation of a quadrature digital downconverter.
(a) Verify Equation 2.12. Referring to Equation 2.8, derive the I Q2.1
and Q received signals, in terms of mI (t) and mQ (t), when the
receiver operates with reference signal
2 exp {−j(2π(fc + f∆ )t + φ)}
instead of the ideal 2 exp{−j2πfc t}. Here, f∆ is called the fre-

quency offset, and φ is the phase offset. For reporting, compose
a brief derivation relating ṽ(t) to m̃(t).
(b) Create and plot a complex exponential signal s(t) = exp(j2πf t).
Here is an example code snippet:
Fs=16000;%sampling rate, sps

t=0:1/Fs:0.005; %5 milliseconds
s = exp(1j*2*pi*1000*t);%1j is sqrt(-1)
figure; %starts a new figure window
plottf(s,1/Fs);
Notice the use of colons to construct a list. Given three num-

bers separated by colons, a uniformly sampled list is generated
by starting at the first entry, incrementing by the middle en-
try, and stopping at the greatest number not exceeding the
last entry. Note also that the semi-colon is used at the end
of a command to suppress display of output to the command
screen.
(c) Use the rotate tool from the figure menu bar to adjust the view
of the time-domain plot of the complex-valued signal. What
do you see? Likewise, explain the visual appearance of the
frequency domain plot.
(d) Implement, in software, quadrature amplitude modulation and

demodulation for sampled message signals, mI (t) and mQ (t).
Use the following system parameters: sampling rate of 8000
sps; 200 time samples; and, carrier frequency of 2.1 kHz. For
the low-pass filter, use length 35, passband edge 1 kHz, and
stopband edge 2 kHz. For input signal m̃(t) = mI (t) + jmQ (t),
use the following values for m̃(t):
√ √
• first 50 samples, 1/ 2 + j/ 2
√ √
• second 50 samples, −1/ 2 + j/ 2
√ √
• third 50 samples, −1/ 2 − j/ 2
√ √
• last 50 samples, 1/ 2 − j/ 2.
Hint: a constant list of 50 complex-valued samples with mod-
ulus 1 and angle 45 degrees can be generated as
(1/sqrt(2) + 1j/sqrt(2))*ones(1,50).
From the output of your quadrature amplitude demodulator,

make three plots.
• time plot of received “I” signal
• time plot of received “Q” signal
• time plot of the phase of the received complex-baseband
signal
Note that the phase of a complex number v, in degrees, can
be obtained as angle(v)*180/pi. It is recommended that the
I and Q plots be combined as subpanels within a single figure
using subplot, for easy side-by-side comparison.
figure;%create new figure window
subplot(2,1,1) %2 rows, 1 column, first plot
plot(real(v));title(’In-phase signal’)
subplot(2,1,2)
plot(imag(v));title(’Quadrature signal’);
xlabel(’sample number’)
Q2.2 (e) View, using plottf.m, the demodulated signal in step (d) be-
fore and after applying the low-pass filter that removes the
double-frequency terms. For reporting, present the graphs;
briefly compare and explain the two frequency domain plots.
(f) For the received signal generated in step (d), create an IQ plot Q2.3
by plotting the quadrature signal (vertical axis) versus the in-
phase signal (horizontal axis). Mark a red circle for every Lth
point, as in the following code snippet for some signal, v, and
L = 10.
figure %create new figure window
plot(real(v),imag(v));
L=10;
hold on; %keep existing plot and overlay
plot(real(v(1:L:end)),imag(v(1:L:end)),’ro’);
%option ’ro’ uses red circles as markers
hold off;title(’IQ Plot’)
xlabel(’In-phase’);ylabel(’Quadrature’)
Additionally, plot the phase difference between the message
samples, m̃(t), and the received signal, ṽ(t). Take note to ac-
commodate the filter delay, which is one-half the filter order.
Further, truncate the list of received delayed samples to be
the same length as the list of message samples. For reporting,
present the IQ plot of the received signal.
(g) Recreate the IQ plot and phase difference plot from step (f), but
add a phase offset, φ, at the demodulator. That is, implement
the demodulator as in Figure 2.3 but with receiver oscillator
2e−j(2πfc t+φ) for φ 6= 0. What do you observe in your plots?
Why?
(h) Recreate the IQ plot and phase difference plot from step (f), Q2.4
but insert both a phase offset, φ, and a frequency offset, f∆ ,
at the demodulator. Try two or three different offsets, both
negative and positive. What do you observe in your plots?
Why?
For reporting, present the IQ plot and the phase difference
plot for one of your parameter selections. Briefly interpret the
graphs.
2.4 Demonstration
For demonstration, use the Communication Laboratory mobile app
to transmit a carrier tone. Descriptions of the app and how to down-
load it are found in Section 7.1.1.

Record the transmission with sampling frequency Fs=44100. View
the received signal using the plotttf.m function.
Demodulate the received signal, and view graphs of the I and
Q components of the demodulated signal. What do you observe?
Provide an explanation for your observations.
2.5 Summary
In Chapter 2, quadrature amplitude modulation and coherent quad-
rature demodulation have been explored. Complex baseband was
presented as a convenient mathematical notation for representing
QAM signals. The exercise also introduced the practical impair-
ments of frequency and phase offsets, which are compensated in the
receiver design to be implemented in Chapter 6. The explorations
served to introduce the commands and conventions summarized in
Table 2.1.

Command Description
angle compute phase, in radians, of a complex number
exp exponential function
real real part
imag imaginary part
1j imaginary unit
length returns length of vector
sqrt square root
xlabel create a label on the horizontal axis of a plot
ylabel create a label on the vertical axis of a plot
subplot create array of plots within a figure
hold hold current graph (to create overlay)
CHAPTER 3
Digital Modulation
In this lesson we review digital modulation, with emphasis on the spe-

cial cases of binary and quadrature phase-shift keying. The chapter
equips students to implement the initial steps of the digital modula-
tion chain depicted in Figure 1.1, mapping text to bits to symbols.
Likewise, the chapter develops the inversion of these steps in the
receiver chain. Through exercises and a demonstration, students vi-
sualize digital modulation via signal space plots and conduct both
analytical and empirical characterizations of bit error rate.
3.1 Background
3.1.1 Digital modulation
The transmission chain in Figure 1.1 depicts a digital modulator that

maps digital information to an analog waveform suitable for the com-
munication channel. The mapping is typically performed on blocks
of log2 M binary digits (bits) at a time. When the mapping of bits to
waveforms is performed without dependence on previously transmit-
ted waveforms, then the modulator is memoryless. In addition, if the
superposition principle holds for the mapping of the digital sequence
to an analog waveform, then the modulator is linear. Throughout
these chapters, we consider a class of memoryless linear modulators
known as complex pulse amplitude modulation, or quadrature am-
plitude modulation (QAM). In QAM, bits are mapped to complex
symbols, which then modulate pulse shapes. The complex baseband
37
38 CHAPTER 3. DIGITAL MODULATION
signal, or message, is given by

X
m̃(t) = ã[n]gtx (t − nT ) (3.1)
n
where ã[n] is the sequence of complex-valued symbols, gtx (t) is a

pulse shape waveform, and T is the symbol period. The symbol rate,
or “baud rate,” of m̃(t) is 1/T symbols per second, and the bit rate
is (log2 M )/T bits per second for a constellation of M symbols. The
pulse shape, gtx (t), is normalized to unit energy,
Z ∞
|gtx (t))|2 dt = 1. (3.2)
−∞
Design of the pulse-shaping function is considered in Chapter 4.

In this chapter, we consider the mapping of a sequence of bits to
a sequence of symbols and the detection of a symbol from a noisy
received data sample. Each symbol is a complex number selected
from a constellation of M symbols. Four example symbol constel-
lations are given in Figure 3.1. The three constellations shown in
Figure 3.1(a-c) have points that differ only in their phase; this spe-
cial case of digital quadrature amplitude modulation is known as
digital phase modulation, or phase-shift keying (PSK). The name
“keying” comes from physical operation of the electrical telegraph
demonstrated by Samuel Morse in 1837 [27]. The Morse code is a
variable-length binary code in which letters of the English alphabet
are mapped to dots and dashes.
The constellation for M = 2 in Figure 3.1(a) is binary phase-shift
keying (BPSK). The graph depicts modulated signals using axes cor-
responding to the two orthogonal signals in QAM: gtx (t) cos(2πfc t)
and gtx (t) sin(2πfc t). BPSK is a real-valued constellation and does
not utilize the quadrature carrier, sin(2πfc t); however, bandwidth
efficiency is improved by simultaneously placing separate values on
both orthogonal carriers, cos(2πfc t) and sin(2πfc t). For M = 4,
quadrature phase-shift keying (QPSK) is the resulting constellation
and is shown in Figure 3.1(b). QPSK is essentially BPSK for both
the I (cosine) and Q (sine) carriers. Figure 3.1(c) shows an 8-PSK
constellation and an associated mapping of bits to symbols. Finally,
Figure 3.1(d) shows a 16-QAM constellation; here, the symbols dif-
fer in both amplitude and phase. 64-QAM and 256-QAM constel-
lations are often used in digital cable television and cable modem
applications for high spectral efficiency over low-noise channels. The
3.1. BACKGROUND 39
Im Im
01 11
0 1 Re Re
00 10
(a) BPSK (b) QPSK
Im Im
011 0000 0100 1100 1000
010 001
0001 0101 1101 1001
110 000 Re Re
0011 0111 1111 1011
111 100
0010 0110 1110 1010

101
(c) 8PSK (d) 16QAM
Figure 3.1: Illustrations of four symbol constellations: (a) BPSK;

(b) QPSK; (d) 8-PSK; and (d) 16-QAM.
16-QAM constellation is used, for example, in the ITU-T V.22bis

standard for digital communication over telephone networks.
It is convenient to normalize the constellation so that the M
symbols, {ã1 , · · · , ãM }, have unit energy on average, i.e.,
M
1 X
|ãi |2 = 1. (3.3)
M i=1
With an M -ary alphabet for a constellation, there are log2 M

bits per symbol; therefore, a single symbol error can result in up to
log2 M bit errors. And, there are M factorial possible assignments of
bits to the M symbols. In general, a preferred assignment is to map
bits to symbols so that neighboring symbols differ by only a single
bit, thereby limiting the number of bit errors resulting from a symbol
error. Such a mapping is called a Gray code or unit-distance code.
Table 3.1: Bit assignments for (a) BPSK and (b) QPSK digital mod-
ulation.
Input bits Symbol
MSB, LSB
Input bit Symbol √
11 (1 + j)/ √2
1 1
01 (−1 + j)/√2
0 -1
00 (−1 − j)/√ 2
10 (1 − j)/ 2
(a) (b)
Suggested assignments of bits to symbols for BPSK and QPSK are

given in Figure 3.1(a)-(b) and are repeated in Table 3.1. In QPSK
it is customary that the most significant bit (MSB) is the first taken
from the bit stream. With this convention, for each pair of bits
QPSK modulates the first bit onto the I channel and the second bit
onto the Q channel.
3.1.2 Symbol detection

At the receiver, we must detect the transmitted symbol and recover
the associated bits. Referring to Figure 1.1, receiver processing re-
sults in a data stream modeled as a noisy sequence of complex-valued
samples,
ỹ[n] = ã[n] + w̃[n]. (3.4)
Design of receiver processing to justify this model is considered in
Chapter 4. The received noisy samples ỹ[n] can be visualized using
a scatter plot, as illustrated for QPSK in Figure 5.3. The following
code creates this example.
%% Scatter plot example for QPSK or BPSK

N = 160; %number of symbols
M = 4; %4=QPSK or 2=BPSK
if(M==2)
a = (2*round(rand(1,N))-1); %noiseless I
elseif(M==4)
a = (2*round(rand(1,N))-1) ... %noiseless I
+ 1j*(2*round(rand(1,N))-1); %and add Q
a = a/sqrt(2); %unit length QPSK symbols
3.1. BACKGROUND 41
end
EbperN0dB = 6;
EbperN0 = 10^(EbperN0dB/10);
varw = (EbperN0*log2(M))^(-1)/2; %per channel sample var
w = sqrt(varw)*(randn(1,N) + 1j*randn(1,N));
y = a + w; %add noise
% make scatter plot
figure;
plot(real(y),imag(y),’x’,’MarkerSize’,8,’LineWidth’,2);
title(’Scatter Plot’,’fontsize’,13)
xlabel(’In-phase’,’fontsize’,13);
ylabel(’Quadrature’,’fontsize’,13);
axis([-1.5 1.5 -1.5 1.5]);axis square% set axes extents
grid % draw grid lines
Scatter Plot
1.5
0.5
Quadrature
−0.5
−1
−1.5
−1.5 −1 −0.5 0 0.5 1 1.5
In−phase
Figure 3.2: Example scatter plot for 160 noisy QPSK symbols.
Here we consider symbol detection, whereby a decision is made on

a symbol-by-symbol basis; in contrast, for modulation schemes with
memory a receiver jointly performs detection of an entire sequence
of symbols. An intuitive symbol detection rule is to map a received,
noisy symbol to the nearest constellation point. Indeed, for equally

likely symbols and additive white Gaussian noise of equal variance
in the I and Q channels, the nearest neighbor criterion minimizes the
probability of symbol error.
The nearest neighbor rule, or minimum distance detector, defines
decision regions; each decision region contains those points closest
to a given constellation element. The decision regions are denoted
by the dashed lines in Figure 3.1. For BSPK, decision regions are
simply the half-planes separated by the imaginary axis; mapping to
logical 1 or 0 is easily implemented via the inequality test, (real(y)
> 0). In low-level software, this test is merely a test on the sign bit
of y. Similarly, for QPSK the decision regions are the four quadrants
of the complex plane.
3.1.3 Bit and symbol error rates

The probability of symbol error depends on the signal-to-noise ratio
in the received data stream, ỹ[n] = ã[n] + w̃[n]. The SNR of ỹ[k] is
given by
P Es /T Es
SNR = = = (3.5)
N0 × W N0 /T N0
where P is the received signal power experienced at ỹ, due to am-
plifier gains, channel loss, and receiver processing. The noise power
spectral density is N0 Watts per Hertz, to model primarily ther-
mal noise in the receiver amplifiers; with a receiver bandwidth W =
T −1 Hz, the noise power becomes the product, N0 W . And, Es is the
energy per symbol, while Eb = Es / log 2 (M ) is energy per bit. Thus,
we have the variance of ỹ conditioned on ã sent,
N0
var(w̃) = SNR−1 = . (3.6)
Es
BPSK
Considering the special case of BPSK, the energy per bit, Eb , is
simply Eb = Es . For ã = 1, an error occurs if the real part, yI =
Re(ỹ), is less than zero. The real part carries half the noise variance,
and hence the variance of Re(w̃) is N0 /(2Eb ). This distribution is
shown in Figure 3.3. The probability of bit error for BPSK when
transmitting ã = 1, then, is the area under the curve for yI ≤ 0.
3.1. BACKGROUND 43
Evaluating this integral we have

Z !
0 1 (yI − 1)2
PBP SK = p exp − dyI
−∞ πN0 /Eb N0 /Eb
s !
2Eb
= Q , (3.7)
N0
where the Q function, Q(x), is implemented via qfunc(x) and is

defined as Z ∞
1
Q(x) = √ exp −t2 /2 dt. (3.8)
2π x
By symmetry, the probability of error is the same when conditioned
on the symbol 0 transmitted. The bit error rate (BER) for BSPK is
shown versus Eb /N0 (SNR per bit) in Figure 3.4.
Es
Figure 3.3: Probability of error for BSPK modulation on a binary

Gaussian channel.
QPSK
Consider next the error probability for QPSK in two cases. First,
suppose the bandwidth, W = T −1 , is identical for BSPK and QPSK
and fix the energy per symbol, Es . Then, QPSK provides twice the
data rate as BPSK, but suffers a worse bit error rate. The worse
bit error rate is evident from Figure 3.1 where we observe that √ the
distance between nearest QPSK constellation points is only 2/2,
compared to a distance of 2 for BPSK.
p Thus, following Equation 3.7
the BER for QPSK is given by Q 2Eb /N0 ; but, for a fixed symbol
0
BPSK
10
−1
10
Probability of Bit Error
−2
10
theory
simulation
−3
10
−4
10
−5
10
−4 −2 0 2 4 6 8 10
E /N (dB)
b O
Figure 3.4: Bit error probability for BPSK versus Eb /N0 .
energy, the energy per bit, Eb , for QPSK is one-half the energy per
bit for BSPK.
Second, set the data rate equal for BPSK and QPSK. Then, the
signal bandwidth W = T −1 for BPSK must double, thereby doubling
the receiver noise power, N0 W . Thus, for equal data rate, QPSK
and BPSK have the same bit error rate, but QPSK uses only half
the spectral bandwidth. For this reason, QPSK is said to be more
spectrally efficient than BPSK.
In summary, in a bandwidth-limited application, QPSK is pre-
ferred over BPSK, because QPSK can offer the same rate of correct
bits using only half the bandwidth. On the other hand, in a power-
limited application BPSK may be preferred.
The detection error analysis above is a simple example demon-
strating that communication engineers can analytically predict per-
formance and evaluate system design trade-offs without expensive
build-and-test cycles. And, the quality of a prediction is limited
only by the validity of the attendant assumptions, such as channel
models (including AWGN) and amplifier linearity. Given that engi-
neers are able to design both the transmitter and receiver, modeling
assumptions can typically be very accurate for commercial systems.
Limiting bounds on performance provide additional insight to de-
sign trade-offs; Shannon’s channel capacity bound is considered in
Chapter 9.
3.2 Explorations
In this lesson students will implement the digital modulation and
demodulation steps shown in Figure 1.1. In particular, students
are given template functions in char2psk.m and psk2char.m found
in Appendix C. From these templates, two functions will be con-
structed: one function to map a character string to a sequence of
symbols, and a second function to map noisy received symbols to a
character string. The code templates provide examples of the for-
mat, syntax, and comment style for students to consider. Use of
well-commented modular code will facilitate construction of an end-
to-end acoustic digital modem.
The steps adopted in the code templates for mapping an ASCII
character string to bits are illustrated in Table 3.2. The command
double converts an ASCII character to a decimal label. Next, the
command dec2bin converts the decimal numbers to a binary repre-
sentation of specified length, which here is 8 bits. Then, .’ performs
a 90◦ rotation of the bit array, turning rows into columns, and the
command reshape reads the transposed bit array, column by col-
umn, into a single long row.
A lab report should provide answers to the four questions enu-
merated in the page margins and the bit error rate curve measured
in the demonstration.
(a) Create a function to convert a character string to a sequence

of symbols. The inputs should be a character string and an
identifier to select a symbol constellation of size 2 (BPSK) or
4 (QPSK); the outputs should be a list of bits and a list of
complex-valued symbols, each of unit magnitude. Study the
example template given in Appendix C and complete the por-
tions that are left blank.
(b) Create a function to implement a minimum distance decoder

for BPSK and QPSK modulation. The inputs should be a
Table 3.2: Example illustrating one possible path to convert an

ASCII character string to bits.
str=’abc’;
str_dec=double(str);
str_bin=dec2bin(str_dec,8);
bits=reshape(str_bin.’,1,8*length(str));
str_dec =
97 98 99
str_bin =
01100001
01100010
01100011
bits =
011000010110001001100011
complex-valued string of noisy symbols and an identifier to se-

lect BPSK or QPSK modulation; the outputs should be a list
of detected bits and a character string.
(c) Verify that your two functions from steps (a) and (b) work
together to properly map characters to bits to symbols and
back again. (Save these functions for re-use in subsequent lab-
oratory sessions as you continue to design and construct an
acoustic modem.)
Q3.1 (d) Suppose the pulse-shaping function, gtx (t), in Equation 3.1 has
two-sided bandwidth of (1 + α)/T Hz, where T is the symbol
period. What is the bandwidth efficiency of BPSK in bits per
second per Hertz ? What is the bandwidth efficiency of QPSK?
Of 64-QAM?
Q3.2 (e) Suppose a random bit stream is BPSK modulated such that
each symbol has energy Eb . In an AWGN channel, what is the
average probability of bit error for a minimum distance decoder
as a function of Eb /N0 , Hint: you may leave the answer in terms

of the Q-function. In your report, include a rough sketch that
accompanies your analysis by showing probability as an area
under a curve.
(f) Create a scatter plot of 1000 simulated received QPSK symbols. Q3.3
Refer to the example given in Section 3.1.2 and set the noise
power for the complex additive white Gaussian noise to be 5 dB
per symbol. From the analysis given in Section 3.1.3, what
BER do you expect?
(g) Qualitatively and intuitively, describe how BER and Es are Q3.4
evidenced in the signal space scatter plot.
(h) [Optional] Extend your functions from steps (a) and (b) to
implement 8-PSK modulation.
(i) [Optional] Extend your functions from steps (a) and (b) to
implement 16-QAM modulation.
(j) [Optional] Implement a function to estimate Eb /N0 , in deci-

bels, from symbol rate samples at the receiver; the inputs
should be M , ã[n] and ỹ[n].
(k) [Optional] Convert the code example in Section 3.1.2 into a

function, SimScatterPlot.m. The inputs are N,M,EbperN0dB.
The outputs should be an observed error rate and a scatter
plot.
3.3 Demonstration
Create a simulator for QAM symbols in additive white Gaussian
noise. Via simulation, create bit error (BER) curves for BPSK and
QPSK as a function of Eb /N0 over the range −4 dB to 10 dB in
0.25 dB steps. Use N=1e6 random trials. (Note that, in general, to
simulate an error rate of 10−d you should use at least N = 10d+1
random trials.) Compare your empirical result to your conclusion
from step (e). You can re-use your code from step (b) to implement
the minimum distance decoder, and the simulation was considered in
step (f) (and in step (k)). Hint: one method for counting bit errors
is
BER = length(find(Rxbits-Txbits))/N;
where Rxbits and Txbits are the length N bit streams at the receiver
and transmitter, respectively.
3.4 Summary
In Chapter 3, digital modulation has been explored with emphasis on
the special cases of binary and quadrature phase-shift keying. The
exercises prepare students to implement the initial steps and final
steps of the modem processing chain depicted in Figure 1.1: mapping
text to bits to symbols at the transmitter, and mapping receiver
samples to symbols to bits to text. The explorations introduced
several new commands and conventions summarized in Table 3.3.
3.4. SUMMARY 49

Command Description
> returns 1 at indices where inequality is true
qfunc Q function gives tail probability of a Gaussian pdf
log2 base 2 logarithm
log10 base 10 logarithm
∧ exponentiation, e.g., 10∧6 is 106
find returns indices of nonzero entries
round nearest integer
axis control axis scaling and appearance
switch, switch among several cases based on expression
case,end
if,elseif, conditionally execute statements
else,end
end can also be used as pointer to end of an array
.’ turn rows into columns or vice versa;
array transpose
sign returns 1 if an element is greater than zero,
0 if zero, and −1 if less than zero
reshape reformat array by reading elements columnwise
bin2dec convert binary string to decimal integer
dec2bin convert decimal integer to a binary string
char convert ASCII integer codes to a character array
double convert to double precision
CHAPTER 4
Pulse Shaping and Matched Filtering
In Chapter 3 the digital modulation and detection parts of a base-

band modem were implemented. Here, development of the baseband
modem is continued by incorporating pulse shaping at the transmit-
ter and filtering at the receiver. The complete baseband processing
chain is illustrated in Figure 4.1.
This chapter begins with an abbreviated review of pulse shap-
ing. The review summarizes design issues associated with inter-
symbol interference and maximization of receiver signal-to-noise ra-
tio. Through exercises and a demonstration, students implement
transmitter and receiver filtering and visualize received baseband
BPSK signals using an eye diagram.
4.1 Background
A sequence of symbols, ã[n], is not suitable for direct transmission
over a communication channel. Symbols must be converted to a
waveform for transmission. In a process known as pulse shaping,
symbols are converted to a sampled-data baseband message; then,
as seen in Figure 1.1, the sampled data are converted to an ana-
log voltage signal through a digital-to-analog converter (DAC) and
mixed to a frequency band suitable for the communication channel.
From Equation 3.1, the sampled-data baseband message is given by
X
m̃[k] = ã[n]gtx (kTs − nT ) (4.1)
n
where ã[n] is the sequence of complex-valued symbols, gtx (t) is a

pulse shape waveform, T is the symbol period, and Ts = 1/fs is the
51
52
ã[n] Pulse Shaping m̃[k] m̃(t) QAM Modulation s(t)

Channel Symbol Add Pilot Digital Baseband Analog Baseband
Bits Bits Symbols Symbols L gtx [k] DAC Re RF Signal
Coding Mapping Sequence Signal Signal
j2πffct
e Tx Amp
h̃[k] Channel gain
w̃[k]
Additive White
Gaussian Noise
Matched Filtering
ỹ[n] ṽ[k]
Channel Symbol Detection Digital Baseband Analog Baseband
Pilot sequence -j2π(ffc+fΔ)t+φ Rx Amp

Symbol Timing e
Detection Recovery
Frequency Sync
Channel
Eye Diagram
Figure 4.1: Fractionally sampled baseband system model for physical layer digital communication processing.
CHAPTER 4. PULSE SHAPING & ISI
4.1. BACKGROUND 53
DAC sampling interval. The summation in Equation 4.1 creates a

superposition of scaled and time-shifted copies of the sampled pulse
waveform, gtx [k]. The summation is a convolution of the upsam-
pled symbol sequence, ã↑ [k], and the sampled pulse shape waveform,
gtx [k]. The upsampling operation merely inserts L − 1 zeros between
consecutive samples of ã[n], with L = T /Ts , in order to match the
baseband sampling interval, Ts < T . Note that here we have required
the symbol period, T , to be an integer multiple of the sampling pe-
riod, Ts . The processing is illustrated in Figure 4.2.
ã↑ [k] gtx [k]

−2L −L L 2L
ã[0] *
L=4 ã[1] ã[3]
k k
ã[2]
DAC
ã[n] ã↑ [k] m̃[k]
P
L gtx [k] sinc t m̃(t)
T
T
k
L
Figure 4.2: Illustration of pulse shaping for L = 4; upsampling by L

is followed by convolution with the pulse shape, gtx [k].
At the receiver, ideal performance would result in the nth out-

put, ỹ[n], equal to the nth input symbol, ã[n]. However, in practice
ỹ[n] is corrupted by both noise and interference from other symbols,
known as inter-symbol interference (ISI). The receiver filter seen
in Figure 4.3 has impulse response q[k] and is designed to meet two
goals: prevent ISI and suppress noise.
4.1.1 Inter-symbol interference

The first design goal for the receiver filter is to prevent inter-symbol
interference. In an idealized system with no noise and ideal channel
h̃[k] = δ[k], the pulse shaping filter at the transmitter, gtx [k], and
the linear filter, grx [k], at the receiver are in series; therefore, their
54 CHAPTER 4. PULSE SHAPING & ISI
w̃[k]
ã[n] ã↑ [k] m̃[k] ṽ[k] ỹ↑ [k]

L gtx [k] h̃[k] grx [k] L ỹ[n]
Figure 4.3: Illustration of the fractionally sampled complex-baseband

data model for additive noise, w̃[k], and a channel described by filter
impulse response h̃[k].
combined effect is a single filter with impulse response, p(t), given by

the convolution of gtx (t) and grx (t), and frequency response given by
the product of the two frequency responses, P (f ) = Gtx (f )Grx (f ).
Thus, the downsampled output sequence is
X X
ỹ[n] = ã[k]p(nT − kT ) = ã[k]p(T (n − k)). (4.2)
k k
To force ỹ[n] = ã[n], we therefore require that the pulse p(t) must
have gain 1 at the time origin and a value of zero at every integer
multiple of the symbol interval. Any waveform with this property is
called a Nyquist pulse, and the property is simply stated as p(nT ) =
δ[n], where δ[n] is the Kronecker delta sequence,
(
1 n=0
δ[n] = (4.3)
0 n=
6 0.
In the frequency domain, we equivalently have

∞
1 X
p(nT ) = δ[n] ←→ P (f − k/T ) = 1. (4.4)
T k=−∞
That is, the superposition of frequency-shifted spectra must sum to

T at all frequencies. These time and frequency domain characteriza-
tions of a Nyquist pulse are illustrated in Figure 4.4.
4.1.2 Matched filter

The second design goal for the receiver filter is to combat noise.
To maximize the signal to noise ratio at ỹ[n], application of the
4.1. BACKGROUND 55
p(t)
1
p(nT ) = δ[n]
t
−3T −2T −T 0 T 2T 3T
∞
1 X
P (f − k/T ) = 1. 1
T k=−∞
3
f
− 2T − T1 1
− 2T 0 1
2T
1
T
3
2T
Figure 4.4: Time-domain (top) and frequency-domain (bottom) il-

lustrations of the Nyquist pulse property.
Cauchy-Schwarz inequality leads to the conclusion that grx (t) must

be matched to the the pulse waveform:
∗
grx (t) = gtx (−t) ←→ Grx (f ) = G∗tx (f ). (4.5)
Thus, for SNR-maximizing ISI-free performance, the filter design
requires:
1. The convolution of the pulse waveform, gtx (t), and the receiver
filter impulse response, grx (t), must yield a Nyquist pulse. In
the frequency domain, we have Gtx (f )Grx (f ) = P (f ).
2. The receiver filter must be a matched filter for the pulse wave-
form; thus, Grx (f ) = G∗tx (f ).
Together, these two requirements imply that |Gtx (f )|2 must satisfy
the Nyquist pulse criterion.
A common design choice for the pulse shape, Gtx (f ), is the
square-root raised cosine (SRRC) pulse, which is given in the time-
domain as [18]

 √1 sin(π(1−α)t/T )+(4αt/T ) cos(π(1+α)t/T ) T
t 6= 0, ± 4α

 T (πt/T )(1−(4αt/T )2 )

gSRRC (t) = √1 (1 − α + 4α/π) t=0
 T
 n o

 √α 2 π 2 π T
2T
1 + π sin 4α + 1 − π cos 4α t = ± 4α .
(4.6)
SRRC Pulses: Time Domain SRRC Pulses: Frequency Domain

0.14
α=0 α=0
1
0.12 α=0.5 α=0.5
α=1 α=1
0.1
0.8
0.08
0.6
0.06
0.04
0.4
0.02
0.2
0
−0.02
0
−0.04
−5 0 5 −1 −0.5 0 0.5 1
time (in symbol periods) frequency (normalized to T=1)
Figure 4.5: Square-root raised cosine pulse in time domain (left) and
frequency domain (right).
Because gSRRC (t) is real-valued and symmetric about the origin, we

have
∗
grx (t) = gSRRC (−t) = gSRRC (t). (4.7)
The SRRC filter provides a parametrized family of filter designs. The
parameter α ∈ [0, 1] is called the excess bandwidth and provides a
trade-off. The two-sided bandwidth of gSRRC (t) is (1 + α)/T ; a larger
α spreads the frequency content of the pulse to a larger bandwidth,
but results in a more rapid decay of the time-domain signal. The
SRRC pulse is illustrated in Figure 4.5. Note that, for α > 0, the
square-root raised cosine pulse is not a Nyquist pulse; however, the
convolution of gSRRC (t) with itself does produce a Nyquist pulse.
For digital implementation, the filter gSRRC (t) is sampled and
truncated for implementation as a finite impulse response (FIR) fil-
ter, gSRRC [k], such that gSRRC [k] = 0 for |k| > DT . Setting the tails
of the filter to zero slightly violates the Nyquist pulse property and
therefore results in small amount of inter-symbol interference. How-
ever, for α ∈ [0, 1] large, the fast decay of the time signal limits the
truncation error and hence the ISI. This trade-off is explored in the
exercises.
Note that the impulse response gSRRC [k] is symmetric about its
midpoint and is therefore a linear phase filter. For an impulse re-
sponse with Ng = length(g tx) samples, the delay of the linear
phase filter is (Ng - 1)/2 samples. Thus, in the baseband process-
ing of Figure 4.3 the first symbol appears in the upsampled sequence
ỹ↑ [k] with a delay of (Ng -1) due to the combined action of the
transmitter and receiver filters.
4.1.3 Eye diagram

An eye diagram may be used to visualize received symbols and better
understand error behavior. For real-valued symbols, the eye diagram
is a plot that superimposes many T -second segments of Re{ỹ(t)}. For
sampled baseband data, the eye diagram is made by eyediagram.m
by superimposing L-sample segments from the real part of the up-
sampled signal, ỹ↑ [k]. For visualization, L ≥ 8 samples per symbol is
preferred. The traces on the plot depict the voltage paths taken by
the received signal as it progresses from one symbol to the next. Ow-
ing to the filtering caused by the transmitter, the channel, and the
receiver, the path taken can depend on many neighboring symbols.
When there is neither noise nor inter-symbol interference, then the
diagram passes through the ideal symbol values, creating the open-
ing seen in Figure 4.6 for BPSK and metaphorically called an “open
eye.” When the eye is open, symbol decisions are reliable; when
the eye becomes closed, errors occur. In the noiseless case, the eye
may narrow due to ISI, and therefore limit the reliability of decisions
when noise is present.
Symbol timing, the topic of Chapter 5, can likewise be visualized
in the eye diagram. Ideally, the downsampled data, ỹ[n], should
be acquired by taking samples at the time location where the eye
opening is the widest, thereby providing the most noise robustness.
If the symbols are complex-valued, then eye diagrams may be plotted
for both the I and Q channels separately.
The function eyediagram.m takes four inputs: the upsampled
data sequence, ỹ↑ [k], the downsampling factor, L, the number of
symbols, N , and (optionally) the flag ’complex’ if the user wishes
to view both the I and Q channels. The first sample of the input
ỹ↑ [k] is assumed to correspond to the first symbol time instant, with
subsequent symbols occurring every Lth sample.
4.2 Explorations
In this lesson, students implement a digital baseband modem for an
additive white Gaussian noise channel with ideal channel response,
Upsampled RX signal, y [k] Eye Diagram

↑
1.5 1.5
1 1
0.5 0.5
0 0
−0.5 −0.5
−1 −1
−1.5 −1.5
0 200 400 600 800 1000 1200 1400 1600 1800 −0.5 0 0.5
Sample number relative symbol index
Figure 4.6: Left: upsampled received BPSK signal, ỹ↑ [k] with
symbol-rate samples marked as circles. Right: eye diagram.
h̃[k] = δ[k]. The processing chain is illustrated in Figure 4.1 and

Figure 4.3. The baseband modem is implemented by combining the
modulation and demodulation routines from Chapter 3 with the up-
sampling, filtering, and downsampling steps reviewed in Section 4.1.
These new steps can be implemented using the conv command and
the (1:L:end) indexing syntax, both introduced in Chapter 1. The
SRRC filter design is conveniently provided by the function srrc.m
given in Appendix C. Students are reminded to review the engi-
neering skills briefly discussed in Appendix A. Short subsections of
code developed in each chapter will be integrated into a single work-
ing modem; thus, purposeful attention to a few simple practices will
greatly simplify work as the lessons progress.
For your lab report, provide a brief answer to each of the seven
questions enumerated in the page margins. In addition, provide plots
for Questions (b), (e), and (f); with each figure, provide a descriptive
caption. Finally, the report should include your code for Questions
(f) and (g).
Q4.1 (a) Use the functions srrc.m and plottf.m (with Ts=1) to con-
struct and view square-root raise cosine pulse shape filters with
the following values for excess bandwidth, α ∈ [0, 1], trunca-
tion half-width, D, and upsampling, L. (Note L must be an
integer.)
• D = 3, α = 0.10, L = 10.
• D = 3, α = 0.75, L = 10.
• D = 8, α = 0.10, L = 10.
Qualitatively, what do you observe? Does the Nyquist property

approximately hold for gtx with the settings above? Why or
why not?
(b) Using the three settings given in step (a), plot the result of Q4.2
the filter convolved with itself, conv(g tx,g tx). Does the
Nyquist property approximately hold for the settings above?
Why or why not? What do you observe about the effects
of α and D when comparing the three plots to each other?
From the three plots, identify the largest amplitude sample of
conv(g tx,g tx) contributing to ISI; i.e., which symbol-rate
sample deviates most from the ideal Nyquist pulse? Finally,
suppose the symbol ã[20] = 1 is transmitted; for each of the
three settings above; determine the interference (sign and am-
plitude) caused by that symbol at the received symbol for sam-
ple index n = 23.
(c) Short answer: Consider a digital-to-analog converter (DAC) in Q4.3

Figure 4.2 operating at 8000 samples per second.
• Suppose you desire a symbol period of 10 milliseconds.

What should you choose for the integer upsampling factor,
L?
• Can you choose L so that the baud rate is 1500 symbols
per second? Why or why not?
(d) Short answer: Express the quantities listed below in terms of Q4.4
these modem parameters: DAC sampling rate fs ; QAM con-
stellation size, M ; upsampling factor, L; excess bandwidth, α;
SRRC truncation, D.
• Symbol period.
• Baud rate.
• Bit rate.
• Approximate pulse duration, in seconds, for gtx (t).
• Two-sided bandwidth of the baseband analog signal.
• Number of samples in the pulse shaping filter impulse re-

sponse, gtx [k]. (Verify using srrc and length.)
Q4.5 (e) Create a random string of 1000 bits, and encode the bits as
BPSK symbols:
bits=round(rand(1,1000));
a=2*bits-1;
Create an eye diagram for a noiseless channel using the fol-

lowing design parameters for your pulse shape and matched
filter.
• D = 3, α = 0.75, L = 15
• D = 3, α = 0.10, L = 15
• D = 8, α = 0.10, L = 15
Note that the function eyediagram.m expects that the first

sample in ỹ↑ [k] corresponds to the first symbol; therefore, the
filter delay should be considered in passing arguments to the
eyediagram function.
Q4.6 (f) For D = 3, α = 0.75, L = 15, add white Gaussian noise to
the simulated sequence of matched filter outputs, ỹ↑ [k]. Use
noise variance of σ 2 /2 in the real and imaginary parts. For
σ 2 = 0.5 describe what you observe in the eye diagram. Repeat
for σ 2 = 0.02.
Q4.7 (g) In preparation for explorations given in Chapters 5 through
7, create a software template, in four parts, for the acoustic
modem graphically depicted in Figures 1.1, 4.1, and 4.3. The
four parts should be: (i) Modem parameters; (ii) Transmitter;
(iii) Channel Impairments; and, (iv) Receiver. Within each
part, create a section (%%) for each processing step. Provide a
one or two line descriptive caption for the functionality of each
section. Each section should include generation of a graphic for
use in debugging. For a data source at the transmitter, allow
a choice between a text input or random bits.
Fill in your outline with those sections already developed in
Chapters 1 through 4; leave incomplete the other sections.
Preparation of this commented template will serve to accel-
erate your progress through Chapters 5 through 7.
(h) [Optional] From Section 4.1 we can conclude that a desirable

pulse shape has low energy in spectral sidelobes to limit band-
width and has low temporal side-lobes that can contribute to
inter-symbol interference. Are these conflicting or complimen-
tary goals and why?
(i) [Optional] What value of α produces a sinc pulse? Is the sinc
pulse a Nyquist pulse? Is the convolution of a sinc pulse with
itself a Nyquist pulse? Is the convolution of any Nyquist pulse
with itself also a Nyquist pulse?
4.3 Demonstration
Use your complex baseband simulator to consider sample timing er-
rors with and without noise. For example, a sample timing offset of
2 samples can be inserted for a row-vector y up via the command
Offset = 2; y up = [ rand(1,Offset) y up];
First compute the BER and plot an eye diagram for two pulse
designs in the noiseless case with a sample timing offset of 5 samples:
• D = 8, α = 0.1, L = 21
• D = 8, α = 0.75, L = 21
Which pulse is more susceptible to sample timing errors? Why?
Second, repeat with 6 dB SNR in the real part of ỹ↑ [k] and noise
of equal variance in the imaginary part. Note,
1
6 dB = 10 log10 ⇒ σ 2 = 10−6/10 .
σ2
(You are invited to optionally experiment by changing the noise vari-
ance and the timing offset.)
4.4 Summary
In Chapter 4, pulse shaping and matched filtering have been intro-
duced to complete a baseband signal representation for a digital mo-
dem. The eye diagram has been used to visualize the baseband re-
ceived signal and to observe the effects of ISI and symbol timing
error for an AWGN channel. The explorations introduced the new
commands summarized in Table 4.1.

Command Description
srrc function to design square-root raised cosine pulse
eyediagram function to display an eye diagram
clear all removes all variables, globals, functions
close all deletes all figures whose handles are not hidden
CHAPTER 5
Synchronization
In Chapter 5 students build sub-components for a baseband digital

modem with the aim of synchronizing the receiver to the transmitter.
Symbol timing will be estimated for the downsampling of the matched
filter outputs, and frame timing will be recovered to align receiver
processing with a data packet. In conjunction with the symbol timing
and frame timing synchronization, students will also account for a
gain and phase due to the channel.
Non-ideal effects degrade received signals. These effects include:
a) symbol timing and frame synchronization errors; b) offset in the
receiver oscillator’s frequency and phase relative to the transmitter;
c) inter-symbol interference (ISI) due to a frequency-selective fad-
ing channel. These non-ideal effects are viewable in the example
BPSK eye diagrams of Figure 5.1. Ideally, channel impairments due
to timing, frequency offset, phase offset and channel filtering would
be jointly estimated at the receiver, along with the bits encoded in
the transmitted waveform. However, for manageable complexity at
the receiver, the tasks are typically addressed in a staged manner,
with impairments sequentially addressed in several steps. A system-
level view of the receiver processing is depicted in Figure 5.2. This
chapter employs block processing of matched filter outputs to imple-
ment synchronization steps. Processing steps to mitigate additional
impairments are considered in subsequent chapters. Frequency re-
covery is implemented in Chapter 6, and Chapters 8, 10, and 11
consider impairments due to a frequency-selective fading channel.
Adaptive processing is also widely employed to address unknown and
time-varying channel impairments and is considered in Chapter 12.
63
64 CHAPTER 5. SYNCHRONIZATION
2 2
1 1
real
real
0 0
−1 −1
−2 −2
5 10 15 20 25 30 5 10 15 20 25 30
2 2
1 1
imaginary
imaginary
0 0
−1 −1
−2 −2
5 10 15 20 25 30 5 10 15 20 25 30
Figure 5.1: Eye diagrams without (left) and with (right) channel
impairments. Data are shown for a BPSK constellation.
5.1 Background
5.1.1 Symbol timing
The matched filter outputs are fractionally sampled at L times the

symbol rate; that is, there are L samples per symbol period. Our first
step is to select which of the L times best corresponds to the proper
symbol time. We attempt to discover this symbol time in the pres-
ence of channel impairments due to oscillator frequency mismatch,
oscillator phase mismatch and ISI – all of which will be addressed
subsequent to symbol timing. Here we describe a block-processing
version of the constant modulus algorithm (CMA) [17]. Intuition
comes from the eye diagram at the left in Figure 5.1: namely, on
average, the variability in the magnitude, or “modulus,” of the sam-
ples is smallest at the proper sampling instant. Further, notice that
frequency and phase offsets, as seen in Chapter 2, serve only to spin
a PSK constellation in the complex plane, thereby leaving the scatter
plot of matched filter outputs lying in an annulus. With poor symbol
timing, the width of this annulus spreads to fill much more of the
complex plane. This graphical intuition is displayed in Figure 5.3.
Thus, to put this CMA principle into action, we first denote by
γp the mean of the absolute values of matched filter outputs with
5.1. BACKGROUND
Matched Filtering
ỹ[n] ṽ[k]
Channel Symbol Digital Baseband Analog Baseband
Pilot sequence -j2π(fc+fΔ)t+φ Rx Amp

Symbol Timing e
Detection Recovery
Frequency Sync
Channel
Figure 5.2: Receiver processing chain.

65
IQ Samples after Symbol Timing Recovery IQ Samples after Symbol Timing Recovery
1.5 1.5
1 1
0.5 0.5
0 0
−0.5 −0.5
−1 −1
−1.5 −1.5
−1.5 −1 −0.5 0 0.5 1 1.5 −1.5 −1 −0.5 0 0.5 1 1.5
Figure 5.3: Scatter plots of sampled matched filter outputs for QPSK
signalling. Left: poor symbol timing. Right: good symbol timing.
offset p samples and subsampled to the symbol rate:
γp = mean {|ỹ↑ [p + nL]|, n = 1, 2, . . . , ⌊N/L⌋ − 1} . (5.1)
Then, minimization of the variance by selection of p can be written
⌊N/L⌋−1
X
pb = min ỹ↑ [p + nL] − γp 2 . (5.2)
p∈{0,...,L−1}
n=0
That is, which of the L downsampled subsequences has the small-

est variance? See the commands var and min for simple calculation.
Note also the indexing syntax, (p:L:end), for starting at index p,
with 0 ≤ p ≤ (L − 1) and counting by L to the end of a list. In the
laboratory procedures, you will implement this block-processing ver-
sion of the CMA. In some commercial systems, an adaptive version
is implemented [36].
5.1.2 Frame timing for flat channels

Frame synchronization is typically performed through use of a train-
ing sequence of symbols (also called “marker sequence” or “pilots”)
that is known in advance to both the transmitter and receiver. Given
that symbol timing has already been performed, the frame timing
5.1. BACKGROUND 67
step may be computed using the matched filter outputs downsam-

pled to the symbol rate: ỹ[n] = ỹ↑ [nL]. In the explorations, students
will use a training sequence to find the beginning of a received frame.
In a flat fading channel model, the transmitted signal, s(t), sim-

ply experiences a gain, a phase, and a delay. That is, h̃(t) = Aδ(t−τ )
is the model of the channel impulse response for some complex-valued
gain A and a non-negative delay, τ . The channel model is known as
flat fading, because the Fourier spectrum of a delta function is con-
stant versus frequency. (A frequency-selective fading channel model
is considered in Chapters 8, 10, and 11.) Writing the gain and phase
as a complex scalar, H̃, we have
ṽ(t) = H̃ × m̃(t − τ ) + w̃(t), (5.3)
where w̃(t) denotes baseband additive white Gaussian noise (AWGN)

modeling a disturbance due primarily to thermal noise at the receiver.
For the AWGN model, the maximum likelihood (ML) estimate, τb, of
the delay is computed using the absolute value of the matched filter
output:
Z
ỹ(t) = ṽ(λ)m̃∗ (λ − t)dλ

τb = arg max ỹ(t). (5.4)
t
The matched filter output also gives the ML estimate of the unknown
amplitude,
ỹ(τb)
H̃est = R . (5.5)
|m̃(λ)|2 dλ
That is, H̃est is the ratio of the observed peak value to the known
noiseless autocorrelation peak from the training signal m̃(t). Note
that the complex-valued scalar H̃est contains both the gain and the
phase modeling the flat fading channel response. For sampled data
and ±1 marker symbols, Equation 5.5 tells us
H est = peak / length(pilots) (5.6)
where peak is the complex-valued sample of maximum absolute value,

peak = |ỹ(τ̂ )|.
A good marker sequence must have a very sharp autocorrelation
peak to provide noise robustness. For a training sequence, here we
Table 5.1: Code to generate Figure 5.4.

pilots=[1, 1, 1, 1, 1, -1, -1, 1, 1, -1, 1, -1, 1];
N tr=length(pilots);
a=pilots;
xc=conv(a,fliplr(a)); % auto-correlation
figure;stem(-(N tr-1):N tr-1, xc);
xlabel(’lag number’);title(’autocorrelation’)
Ntr −1
consider a Barker code. A length Ntr Barker code {ak }k=0 is a
sequence of binary symbol values, ±1, such that
NtrX
−i−1

ak ai+k ≤ 1 for i 6= 0. (5.7)
k=0
For example, a length-13 Barker code is

{ak }12
k=0 = {1, 1, 1, 1, 1, − 1, − 1, 1, 1, − 1, 1, − 1, 1}. (5.8)
The autocorrelation sequence for the length-13 Barker code is shown

in Figure 5.4, and the code to generate the plot is given in Table 5.1.
Next, we illustrate timing recovery by embedding the Barker code

in noise and computing the cross-correlation between the noisy signal
and the known Barker code. The result is seen in Figure 5.5, where
the visible peak serves to identify the location of the marker sequence
within the noisy signal. The relationship between the sample index
of the correlation peak and the first index of marker sequence is
considered in the Explorations, step (a).
Extension of frame timing for robustness to a frequency-selective
fading channel is considered in Chapters 8, 10, and 11.
5.2 Explorations
In this lesson, students continue development of an acoustic modem
by implementing symbol timing recovery and frame timing recovery
in the baseband receiver processing chain. The baseband channel
model considered is a flat fading channel with additive white Gaus-
sian noise. The processing chain is illustrated in Figures 5.2, 4.1, and
4.3.
autocorrelation
14
12
10
0
−10 −5 0 5 10
lag number
Figure 5.4: Autocorelation for the length-13 Barker code.
For the laboratory exploration, work through the five steps below
to implement a baseband modem in simulation. The suggestions
below explicitly enumerate items for a brief lab report: an equation
for Question 5.1 in step (a), one figure for Question 5.2, a brief
paragraph for Question 5.3, one figure for Question 5.4, and code
appended for Question 5.5.
Important: to aid with debugging and understanding, insert a
graph at each step of the processing chain depicted in Figures 1.1 and
5.2; refer to Table 5.3 for code suggestions to create graphs. Keep
commented and tested code from this chapter for your continuing
use as you build towards a working acoustic modem.
(a) To prepare for frame synchronization in step (d), consider the

following preliminary exercise. Create pilots, a sequence of
pilot symbols, (or “marker symbols”) using a length-13 Barker
code. Insert the pilots into a length 100 signal starting at
location start less than 88. For example,
y = zeros(1,100);
tau = 23; %delay >= 0; start=1+tau
y( (1+tau) : (1+tau)+length(pilots)-1) = pilots;
Use convolution to compute the cross-correlation between y

pulse−shaped pilot sequence

0.6
0.4
0.2
−0.2
−0.4
−0.6
−0.8
0 20 40 60 80 100 120 140
received signal
3
−1
−2
True pilot locations

−3
0 50 100 150 200 250 300 350 400
cross−correlation at symbol rate

12
10
8 Cross−Correlation Peak
6
−2
−4
−6
−8
0 10 20 30 40 50 60 70 80
Figure 5.5: Frame timing example for a flat fading channel. Top:
marker sequence convolved with pulse waveform. Middle: received
waveform before matched filtering. Bottom: correlation of marker
sequence with downsampled matched filter outputs. Peak reveals
location of the marker sequence.
and pilots. Write a general expression to solve for the sample Q5.1
index of y at which the first symbol of the pilot sequence oc-
curs. The solution is found from the sample index of the peak
of the absolute value of the cross-correlation output. Express
your answer in terms of the sample index of the peak of the
correlation output and the length of the pilot sequence. If the
correlation output is named CorrOutput, then the peak value
and the index at which it occurs can be found via
[value,indx] = max(abs(CorrOutput));
peak = CorrOutput(indx);
Test your solution via simulation by changing tau and comput-

ing the cross-correlation. Visualize the technique by plotting
the absolute value of the cross-correlation output.
(b) As a second preparatory step, construct a baseband modem

simulator using the snippets of code you have created in previ-
ous chapters. Refer to Figure 4.3 and Figure 5.2 for processing
roadmaps and re-use your codes from Chapter 4. The purpose
in this step is to implement and test the pulse-shaping and
matched filtering steps, without any synchronization issues.
At the top of your script, assign values for the channel im-
pairments listed in Table 5.2. Verify that the symbol sequence
Table 5.2: Variables to define baseband channel impairments.

H complex-valued gain and phase, flat channel model
noise var variance of complex AWGN
tau channel delay; for simplicity, an integer number
of sampling intervals
is accurately reproduced. Test your baseband modem simulator

for some small N ≤ 10, H=1, and by setting the noise power and
delay time impairment variables listed in Table 5.2 to zero. To
this end, you can directly compute and use the correct timing,
as known from your choice of tau and the filter delays at the
transmitter and receiver, as considered in Section 2.1.5. That
is, in this preparatory step, you do not yet need to implement
symbol timing or frame timing.
Suggestions:
• Recommended pulse shaping parameters: D=5, alpha=0.5,

and L=10.
• For symbols, create a N random BSPK symbols, as in
Chapter 4, and prepend the pilot symbols used in step
(a).
• Implement the sample-rate channel time delay, tau, by
inserting your received message into a longer signal, v,
similar to the construction of signal in step (a) above.
Make the signal, v, 100 samples longer than the sampled
baseband message.
• Add noise to v using variance noise var/2 in each of the
I and Q channels.
• For the baseband transmitter steps, use the conv com-
mand, the indexing syntax (1:L:end), and the srrc.m
function introduced in previous chapters. (Ensure that
srrc.m and any other custom function is present in your
working directory or path.)
(c) [Optional] Using the baseband processing from step (b), view
the changes to the received IQ symbols for channel gain H 6= 1
and nonzero noise variance.
(d) Armed with baseband processing from step (b), next consider
estimation of the proper symbol timing. Augment your base-
band receiver processing chain from step (b) by implement-
ing the CMA symbol timing recovery algorithm described in
Equation 5.1 and Equation 5.2. To begin, review the defini-
tion of variance of a list of numbers and calculation of variance
using var in Matlab. Re-use your code from step (b); set
N = 500. For your tests, a suggested parameter set for pulse
shaping is: D=5, alpha=0.5, and L=10.
Table 5.3 provides suggested plotting commands that you can
insert into your code to use the power of visualization in de-
bugging and interpreting your results. (Note: you may need to
modify variable names to match your code.)
Verify the correctness of your algorithm using scatter plots of
Q5.2 IQ samples before and after symbol timing recovery. Test your
Table 5.3: Code suggestions to generate figures.
% transmitted signal
figure;
plot(m,’bo’);title(’Baseband Tx Message’)
% eye diagram
eyediagram(y_up,L,length(a),’complex’);
% CMA criterion
figure;plot(0:L-1,cma_test);
title(’CMA variance criterion’)
% upsampled MF outputs; superimpose downsampled values
plot(p_est:length(y_up),real(y_up(p_est:end)));hold on
plot(p_est:L:length(y_up),real(y_up(p_est:L:end)),’ro’)
title(’Downsampling of MF Outputs’)
% video illustrating all choices for downsampling
figure
for kk=1:L
plot(kk:length(y_up),real(y_up(p_est:end)));hold on
plot(kk:L:length(y_up),real(y_up(p_est:L:end)),’ro’);
title(’Symbol timing from CMA’,’fontsize’,13);hold off
pause(0.5)% view at two frames per second
end
% correlation output
figure;plot(xc);
title(’Cross-correlation for frame timing’)
% scatter plot
figure;
plot(real(y_up(start:L:end)), ...
imag(y_up(start:L:end)),’ko’);
grid on; axis image;
title{’IQ samples after symbol timing recovery’)
symbol timing recovery using several values for the channel

delay parameters: tau = 0, 2, 4, 6 and keeping other channel
impairments benign: H=1 and noise var=0.
After you have verified noiseless performance, experiment by
re-running your section of code using the following values for
the remaining channel impairments: H= 2*exp(1j*pi/4) and
Q5.3 noise var=0.01. Generate one figure illustrating performance
of the CMA symbol timing recovery; provide a descriptive cap-
tion and list simulation values for both the modem parameters
and the channel impairment parameters. Provide a brief intu-
itive description of why the CMA algorithm works.
(e) Augment your baseband receiver processing chain from step (d)
to insert frame synchronization explored in step (a). Note the
frame synchronization operates on the downsampled matched
filter outputs, ỹ[n]. Verify operation of your frame synchro-
nization by running the simulation with nonzero tau greater
than L.
(f) Augment your baseband receiver processing chain from step (e)
to insert the flat fading channel equalization from Equation 5.5.
Verify operation of your frame synchronization by running the
simulation with several choices of gain and phase for H.
Q5.4 Generate one figure illustrating the combined performance of
the CMA symbol timing recovery, frame timing recovery, and
flat fading channel equalization; provide a descriptive caption
and list simulation values for both the modem parameters and
the channel impairment parameters.
Q5.5 For your report, append code for the software simulator as
augmented at step (f). Provide two or three figures illustrat-
ing execution of your modem simulator as performed in the
Demonstration below; with the figures, provide a descriptive
caption and a table listing both the modem parameters and
the channel impairment parameters.
5.3 Demonstration
Demonstrate your complex-baseband modem simulator using a mix
of parameter settings for the channel impairment variables given in
5.4. SUMMARY 75
Table 5.2.
5.4 Summary
In Chapter 5, techniques have been introduced to recover symbol
timing and frame timing at the receiver. In addition, frame timing
also provided equalization for a flat fading channel response. The
exercises provide a complete working baseband modem; in the next
chapter, the processing will be extended to operate the modem using
bandpass signals at an acoustic carrier frequency. The explorations in
this chapter introduced the new commands summarized in Table 5.4.

Command Description
min returns the value of the smallest element
and the index (or indices)
max returns the value of the largest element
and the index (or indices)
var computes the variance of a list of numbers
fliplr reverses the column ordering of an array
CHAPTER 6
Frequency Recovery
In Chapter 6 students complete a software-defined bandpass digi-

tal modem. The subcomponents developed in previous lessons for
a baseband system are combined with quadrature modulation using
an acoustic carrier frequency. Frequency recovery is implemented at
the receiver to account for the impairment due to a frequency mis-
match between the transmitter and receiver. Additionally, channel
sounding experiments explore flat versus frequency-selective channel
models.
6.1 Background
The complex-baseband equivalent model completed in Chapter 5 is
shown in Figure 6.1. In the illustration, the baseband signal is shown
upsampled to L samples per symbol, and the model is therefore
referred to as “fractionally sampled.” In this chapter, the process-
ing is extended to perform analog quadrature amplitude modulation
(QAM) and demodulation using an acoustic carrier frequency. Due
to the inherent mismatch between transmitter and receiver oscilla-
tors, a frequency recovery step is inserted in the receiver processing
chain. The bandpass model is depicted in Figure 6.2.
6.1.1 Frequency recovery

For coherent demodulation, the local oscillator must have frequency
and phase synchronized with the transmitter. Consider a receiver
oscillator given by ej(2π(fc +f∆ )t+φ) with a frequency mismatch f∆
77
78 CHAPTER 6. FREQUENCY RECOVERY
w̃[k]
ã[n] ã↑ [k] m̃[k] ṽ[k] ỹ↑ [k]

L gtx [k] h̃[k] grx [k] L ỹ[n]
Figure 6.1: Fractionally-sampled complex-baseband model.
DUC s(t)
ã[n] ã↑ [k] m̃[k] s[k]
L gtx [k] Re DAC
ej2πfc Ts k
DDC r(t)
ỹ[n] ỹ↑ [k] ṽ[k] r[k]
L grx [k] ADC
2e−j2πfc Ts k
Figure 6.2: Baseband processing and direct digital quadrature mod-

ulation to an acoustic carrier frequency. Top: Transmitter. Bottom:
Receiver.
Hz and a phase mismatch φ radians. Using the convenient com-

plex baseband representation of Figure 2.4 we have that the received
message is
ṽ(t) = m̃(t)e−j(2πf∆ t+φ) . (6.1)
Suppose ã[n] is a sequence of known message symbols; thus, after
proper receiver filtering, symbol-timing recovery and frame-timing
recovery, the received symbols suffer a time-dependent phase error:
ỹ[n] = ã[n]e−j(2πf∆ nT +φ) , (6.2)
where T is the symbol period and the index n counts the sampled
marker sequence from n = 0. The phase offset term e−jφ may be
6.1. BACKGROUND 79
incorporated into the complex-valued channel gain and therefore is

recovered with frame timing, as seen in Section 5.1.2. Thus, we need
only recover the frequency offset, f∆ .
In Chapter 5, a known sequence of pilot symbols was used to
recover frame timing, channel gain, and channel phase. Here, we use
the same pilot sequence to perform frequency recovery, as well. From
Equation 6.2 and with φ = 0, the phase difference between known
pilots and received symbols is a linear function of sample index,

ã[n]
angle = (2πf∆ T ) n, (6.3)
ỹ[n]
where T is the symbol interval. Notice that the quantity (2πf∆ T )

gives the incremental phase error in radians per symbol and spec-
ifies the slope of the linear function. Thus, a line fit to the un-
wrapped phase angle of the known markers divided by the received
symbols gives an estimate for the frequency recovery, as illustrated
in Figure 6.3. In the example, the true frequency offset of −0.05 ra-
dians per symbol interval is estimated as −0.0570 radians per symbol
using a noisy received marker sequence of 13 pilot symbols.
The Matlab command p = polyfit(x,y,d) finds the coeffi-
cients of a polynomial p(x) of degree d that best fits the data, y(i),
at sampling instants, x(i), in a least squares sense. For additive
white Gaussian noise, the least-squares fit yields the maximum like-
lihood estimate. The result p from the command is a row vector of
length d+1 containing the polynomial coefficients in decreasing order.
In particular, for a linear fit, we set d=1, provide samples x=0:N-1,
and obtain p equal to [2πf∆ T, φ]. Note that the unwrap command
should be applied to the phase difference to avoid jumps in the phase
were it instead reported on the interval [−π, π). Further, recall the
phase difference between a and b can be obtained via angle(a ./
b). The need for the unwrap operation is illustrated in Figure 6.4;
in the example there are 25 symbols, π/10 radians per symbol phase
error, and φ = π/20. To correct the phase of the data symbols, the
linear phase can be added to the symbols for indices beginning with
N ; that is, to properly correct the phase, the indexing must continue
with the same symbol counting used in the line fit.
Frequency Recovery
0
Phase error in received pilot symbols
LS line fit
−0.1
Unwrapped phase (radians)
−0.2
−0.3
−0.4
−0.5
−0.6
−0.7
−0.8
0 2 4 6 8 10 12
Symbol number
Figure 6.3: Frequency recovery example for a flat fading channel.
Wrapped Phase and Linear Fit Unwrapped Phase and Linear Fit
8
3
7
Unwrapped phase (radians)
2
Wrapped phase (radians)
6
1
5
0 4
3
−1
2
−2
1
−3 0
0 5 10 15 20 25 0 5 10 15 20 25
Symbol number Symbol number
Figure 6.4: Phase difference (dots) and linear fit (dashed line) per-
formed without (left) and with (right) phase unwrapping.
6.1. BACKGROUND 81
6.1.2 Frequency-selective fading channel model

Here, as in Chapter 5, the remediation of channel impairments as-
sumes a flat fading channel with impulse response given by
h̃(t) = Aδ(t − τ ). (6.4)
Thus, the impairments in the baseband signal due to the flat fading
channel model are limited to a gain, a phase, and a time delay.
More generally, a frequency-selective fading channel model pro-
vides a gain and phase that vary as a function of frequency. The most
common frequency-selective channel model is a multi-path model, in
which the received signal is a noisy sum of delayed and attenuated
versions of the transmitted signal. For K paths, each with complex-
valued gain Ak and path delay τk , the so-called K-tap impulse re-
sponse model is
K
X
h̃(t) = Ak δ(t − τk ). (6.5)
k=1
The different delayed signal versions result from multiple propagation
paths between the transmitter and receiver. At radio frequencies,
causes of multi-path include atmospheric ducting and reflections from
buildings or mountains. The constructive and destructive interfer-
ence of the multiple arriving signals causes a frequency-dependent at-
tenuation. Timing recovery and equalization for a frequency-selective
fading channel are considered in Chapter 8.
6.1.3 Channel measurement

A frequency-selective fading channel is described by an impulse re-
sponse of length greater than one, in contrast to flat fading channel
characterized by a single complex-valued scalar for gain and phase.
In this subsection, we provide background for an optional exercise in
which the cross-correlation idea from Chapter 5 is used to estimate
a channel impulse response. This measurement is sometimes called
channel sounding.
For a binary marker sequence c[n] of length N with good auto-
correlation properties, the autocorrelation is approximately a scaled
Kronecker delta sequence:
1 X
c[k]c∗ [k − n] ≈ δ[n − N ]. (6.6)
N k
With this observation and the associative property of convolution,

we can compute a cross-correlation between a known marker se-
quence and a noisy received symbol sequence to estimate a symbol-
rate baseband representation of the channel impulse response, h̃[n].
Let ỹ[n] = h̃[n] ⋆ c[n] + w̃[n] be a noisy received version of the marker
sequence, c[n], distorted by convolution with the baseband channel
impulse response, h̃[n]. Here ⋆ denotes linear convolution. The cross-
correlation with a known marker sequence yields
1 1 n o
{ỹ[n] ⋆ c∗ [−n]} = h̃[n] ⋆ c[n] + w̃[n] ⋆ c∗ [−n]
N N
1 1
= h̃[n] ⋆ c[n] ⋆ c∗ [−n] + {w̃[n] ⋆ c∗ [−n]}
N N
1
≈ h̃[n] ⋆ δ[n − N ] + {w̃[n] ⋆ c∗ [−n]}
N
≈ h̃[n − N ]. (6.7)
One choice for a long binary marker sequence is a pseudo-random

noise (PN) sequence of +1 and −1 symbols. Thus, a PN code and
cross-correlation processing can provide an estimate of the channel
impulse response.
6.2 Explorations
In this lesson, students complete development of a bandpass QAM
modem simulator by implementing quadrature modulation to an
acoustic carrier frequency and creating a frequency recovery step
in the receiver processing chain. The baseband channel model con-
sidered is a flat fading channel with additive white Gaussian noise.
The processing chain is illustrated in Figure 1.1 and Figure 5.2.
For the laboratory exploration, work through the three steps be-
low to implement a bandpass modem in simulation. Recommenda-
tions for preparing a brief, descriptive laboratory report are included
in the suggested steps.
Important: to aid with debugging and understanding, insert a
graph at each step of the processing chain depicted in Figure 1.1.
(a) Augment and revise your complex-baseband modem simulator

from Chapter 5 to implement a bandpass model.
• Let the modem parameters, such as L, D, α, fc , and fs ,

be variables clearly assigned using load or set at the be-
ginning of the simulator script, rather than coded as nu-
merical values throughout.
• At the sampled bandpass signal, s[k], insert simulated im-
pairments for a flat fading channel model: a complex-
valued channel gain, a delay, and AWGN.
• Code your quadrature demodulator to include a simulated
frequency offset, f∆ .
• First, verify the operation of your bandpass modem sim-
ulator for zero frequency offset, a channel gain of 1, and
no delay.
• Second, verify the operation of your bandpass modem sim- Q6.1
ulator when channel impairments include a non-zero delay
and a complex-valued channel gain not equal to 1. For re-
porting, show one IQ scatter plot for decoded symbols
and list the simulation parameters of the modem, includ-
ing channel impairments.
(b) Augment your bandpass receiver processing chain from step (a)
by implementing the frequency recovery algorithm described
in Section 6.1.1. And, use the estimated frequency offset to
correct the phase of the received symbols.
To begin, review the online help for the commands unwrap and
polyfit. Test your frequency recovery algorithm using the
following suggested modem parameters:
• acoustic sampling rate of 44100

• carrier frequency of 14.5 kHz
• length 13 Barker sequence for pilot symbols
• N = 200 random BPSK symbols
• Upsampling L to achieve a symbol interval, T , of approx-
imately 2.5 milliseconds
• Pulse shape: half-width D=5; excess bandwidth alpha=0.5
Run tests for four values of frequency offset: 2πf∆ T = 0, Q6.2

−0.00005, −0.0005, and −0.005.
For each test, view the results by plotting a scatter plot before
and after the frequency recovery step. For reporting, simply
provide IQ scatter plots for two of the four cases considered,
along with a list of the simulation parameters for the modem,
including channel impairments.
Q6.3 (c) Throught the explorations in Chapter 6, you have augmented

your acoutic modem software template to include frequency
recovery. Submit you commented code.
(d) [Optional] Use cross-correlation with a pseudo-random BPSK

marker sequence to estimate a channel model, h̃[n]. For your
experiment design, the following parameters are recommended:
• Bandpass operation: sampling frequency fs = 44100 sps;

carrier frequency fc = 5.5 kHz;
• Pulse shape: half-width D = 5; upsampling L = 19; ex-
cess bandwidth α = 0.25;
• Marker sequence: N = 4095 random BPSK symbols.
Place a speaker and microphone approximately 30 cm apart and

approximately 65 cm from a wall (or other large surface). Face
the microphone and speaker directly towards the wall, rather
than towards each other.
(i) Perform acoustic transmission and recording, as explored

in Chapter 1.
(ii) Plot the time and frequency domain views of the transmit-
ted bandpass signal. Include this plot in your report. For
your transmitted signal, what is the fractional bandwidth,
defined as 2W/fc , where W is the one-sided bandwidth of
the baseband message and fc is the carrier frequency?
(iii) Compute the cross-correlation of c[n] and ỹ[n] and plot
the absolute value. Define t = 0 as the first correlation
peak. Using 343 m/s as the approximate speed of sound,
label the horizontal axis as path length in cm. Com-
pare your cross-correlation peaks with the physical path
lengths present in your experimental setup. Include this
plot and a brief statement of comparison in your report.
(iv) Using the complex-valued cross-correlation output, plot

the frequency response of your measured channel. The
command freqz may be useful. Interpret the impulse
response and frequency response plot of your measured
channel. Is the channel accurately described by a single
gain and phase?
(e) [Optional] Modify your code in step (d): select L to imple-

ment a signal with two-sided bandwidth of only approximately
290 Hz. Repeat your channel sounding experiment. In this
case, is the channel accurately described by a single gain and
phase? Why or why not? Do you expect your conclusion to
depend on carrier frequency, fc ? Why or why not?
6.3 Demonstration
Demonstrate your bandpass modem simulator implemented in step
(b) for a combination of channel impairment variables of your choice.
6.4 Summary
In Chapter 6, a technique has been introduced to recover a frequency
offset at the receiver. In addition, flat versus frequency-selective
fading channel models have been explored via an optional channel
sounding experiment. The explorations introduced new commands
summarized in Table 6.1.

Command Description
unwrap undo modulo 2π wrapping of a list of phase values
polyfit compute a polynomial fit to data
freqz plot a discrete-time frequency response
CHAPTER 7
Acoustic Modem
In Chapter 7 students employ the software-defined radio components

completed in previous chapters to implement an acoustic modem. A
series of experiments is performed to quantify the performance of the
modem.
7.1 Background
7.1.1 Mobile app
An acoustic quadrature amplitude modulation transmitter is avail-
able as a mobile app. The app may be obtained for Apple devices at
the iTunes App Store1 , and a version for Android devices is available
at Google Play.
The acoustic transmitter app can be operated in two modes: car-
rier tone and text transmission. In carrier tone mode, the app simply
transmits the waveform cos(2πfc t) for a 400 millisecond duration at a
user-specified carrier frequency fc < 22.05 kHz. In the text transmis-
sion mode, the pilot sequence may be selected as either a 13-symbol
Barker code or a 51-symbol pseudo-random sequence. The routine
makepilots.m is provided in Appendix C to define these pilot se-
quences. The data payload is fixed at 50 ASCII characters; a text
string of fewer than 50 characters entered on the app is augmented
with the underscore symbol, , to achieve a 50 character payload.
The user-defined inputs for text transmission are listed in Table 7.1.
1
App Name: Acoustic Transmitter; App SKU: Acoustic Transmitter.
87
88 CHAPTER 7. ACOUSTIC MODEM
The home page and settings page of the app are illustrated in
Figure 7.1.
Table 7.1: User-defined inputs for acoustic transmitter app.

Input Description
text a text message of 50 or fewer characters
or
tone carrier tone only
fc acoustic carrier frequency
L upsampling factor (samples per symbol),
resulting in baud rate 44100/L
D pulse shape half-width, in symbols
α pulse shape excess bandwidth
M constellation size (BPSK or QPSK)
pilots Barker-13 or pseudo-random 51
7.1.2 Data packet

A packet is the basic unit of most digital communication systems.
Packets may vary in structure depending on the protocols imple-
menting them, but typically contain a header and a payload. In
Chapter 5 a sequence of pilot symbols was introduced as a header,
and the data symbols comprised the payload. This simple packet
structure is illustrated in Figure 7.2.
7.1.3 Spectral efficiency

Physical layer design for a communication system is the use of con-
strained resources to maximize system performance. The resources
include bandwidth, power, system complexity, and cost. Perfor-
mance objectives may include, for example, bit rate, error rate, out-
age probability, delay, battery life, etc. One performance measure is
spectral efficiency, measured in bits per second per Hertz. For the
QAM modem developed in the previous chapters, spectral efficiency
is given by
log2 M
η= b/s/Hz. (7.1)
1+α
7.1. BACKGROUND 89
Figure 7.1: Communication laboratory mobile app.
header payload
1 … Ntr 1 … Np
Figure 7.2: A simple packet structure with a header of Ntr symbols

followed by a data payload of Np symbols.
7.1.4 Differential PSK modulation

A differentially encoded phase modulation (differential phase shift
keying, or DPSK) allows the phase estimate to be obtained from
the previous symbol interval. This allows for robustness to, for ex-
ample, growing phase errors due to a small inaccuracy in frequency
offset recovery or due to a slowly time-varying mismatch between the
transmitter and receiver oscillators.
Given any M -ary PSK encoding of a bit stream to produce unit-
length complex-valued symbols, ãk , k = 1, 2, . . . , N , the corre-
sponding DPSK symbols, b̃k , k = 0, 1, 2, . . . , N , can be iteratively
obtained:
b̃k = b̃k−1 ãk , k = 1, 2, . . . , N, (7.2)
where b̃0 can be any unit-length initial condition. Let θk denote the
phase of b̃k , then from Equation 7.2 we have that ãk = ej(θk −θk−1 ) .
Following [27], the demodulator outputs, at symbol times k − 1 and
k, are given by
p
ỹk−1 = Es ej(θk−1 −φ) + ñk−1 (7.3)
p
j(θk −φ)
ỹk = Es e + ñk (7.4)
(7.5)
where φ is the (unknown) carrier phase and nk is zero-mean addi-

tive white Gaussian noise at the matched filter outputs. Then, the
decision variable for the phase detector is the difference,
∗
ỹk ỹk−1 = Es ej(θk −θk−1 ) +
p n o
Es ej(θk −φ) ñ∗k−1 + e−j(θk−1 −φ) ñk + ñk ñ∗k−1 . (7.6)
Hence, the mean is independent of the carrier phase, φ. And, if we

neglect the product term ñk ñ∗k−1 at high SNR, then we observe from
the addition of noise sequences in Equation 7.6 that the difference
between DPSK and phase-coherent PSK is roughly that the noise
variance is twice as large. A more careful analysis reveals that 4-
phase DPSK suffers about 2.3 dB SNR loss versus phase-coherent
QPSK, and that the performance difference between 2-phase DPSK
and BPSK is even less. Thus, DPSK versus coherent PSK provides
a design trade-off often used in digital communications: a reduction
of implementation complexity can be obtained at the cost of slightly
inferior noise performance.
7.2 Explorations
In this chapter, students implement and experimentally verify an
acoustic modem. The acoustic modem should be designed to achieve
a spectral efficiency η > 1. For the transmitter, students may use
their code from Chapter 3 and Chapter 6 to transmit from one com-
puter to another. Alternatively, the app described in Section 7.1.1
may be used as an implementation of the transmitter on a mobile
device.
The laboratory report should contain these brief elements:
(i) A table listing choices for all modem parameters: audio sam-
pling frequency, carrier frequency, symbol period, pulse shape
parameters, digital modulation format, pilot sequence;
(ii) Also include in the table the following parameters: packet
length (in symbols); data rate (bits per second); the experimen-
tally estimated frequency offset, in Hertz, between the transmit
and receive devices; bit error rate (BER); the number of bits
measured to report the BER; and, the estimated frequency off-
set, in Hertz, between the transmit and receive devices.
(iii) A short paragraph describing the physical experimental setup
for the transmitter and receiver;
(iv) A time and frequency plot (plottf) of the received audio-
frequency bandpass signal;
(v) Scatter plot of the received symbols before timing recovery and
downsampling;
(vi) Scatter plot of the received symbols after timing recovery and
downsampling;
(vii) A final scatter plot of the received symbols after frequency
recovery.
Students are encouraged to consider four suggestions given below
for achieving the design goals; see Appendix A for a brief discussion
of recommended software design practices.
• Use modular code. Implementation is facilitated by use of sev-

eral modular scripts. For example, digital modulation func-
tions were completed in Chapter 3, and the simulator from
Chapter 6 can be easily partitioned into one routine for trans-

mission and one for reception. The function srrc.m for pulse
design and the function firlpf.m for low-pass filter design have
been used in previous chapters and are provided in Appendix C.
Be sure to have all functions available in your working direc-
tory or have their directories added to the current search path;
see the commands path and addpath to get, set, or add to the
search path.
• Adopt a strategy to set common parameters at both the trans-

mitter and receiver. Consistently use variable names, such as
L or fcarrier, in your codes, rather than directly coding a
numerical value throughout the code. One method is to create
a stand-alone script that simply assigns values to all modem
parameters and stores them to a file. Then, at the beginning of
the transmit and receive functions the common set of param-
eters can be loaded from memory. The commands save and
load can be used for this purpose.
• Use a fixed-length packet structure; a known packet length facil-
itates the receiver implementation. This can be accomplished
in the conversion of text to symbols. For example, a text mes-
sage can be either truncated or padded with spaces (ASCII 32)
to achieve a fixed number of characters. The packet length will
depend on the number of pilot symbols used in the header, the
number of 8-bit ASCII characters in a text message, and the
number, M , of symbols in the symbol constellation.
• The approximate location of a data packet must be detected

within the segment of sound recorded at the receiver. If the de-
tected window were to contain a large percentage of noise-only
symbol-rate samples, then the constant modulus algorithm pre-
sented in Chapter 5 would not effectively compute the variance
of packet symbols. The routine packetdetect.m is provided
in Appendix C for this purpose and implements a simple en-
ergy detector; the routine operates on the fractionally sampled
matched filter outputs.
The implementation of an acoustic demodulator and five optional

exercises for this chapter are enumerated below.
Q7.1 (a) Implement and experimentally verify an acoustic demodulator
with a spectral efficiency greater than 2; decode the first 20

ASCII characters. See items (i)–(vii) above for required ele-
ments in a brief laboratory report.
(b) [Optional] Modify your acoustic demodulator to decode the

first 50 ASCII characters, with the last 30 characters being
some fixed value (e.g., a space or underbar). For a transmitter,
use either your own code or the mobile app. Plot the angles
of the decoded symbols at the receiver. Briefly explain your
observations.
(c) [Optional] Modify your digital modulation and digital demod-

ulation routines from step (b) to implement and experimen-
tally verify differential QPSK modulation. To this end, modify
and replace your routines char2psk.m and psk2char.m to im-
plement digital modulation and demodulation using 4-phase
DPSK. Note that conjugation in Section 7.1.4 can be imple-
mented usign the command conj.
(d) [Optional] Modify your digital modulation and digital demod-

ulation routines from step (a) to implement and experimentally
demonstrate a spectral efficiency greater than 2.
(e) [Optional] Modify your digital modulation and digital demod-

ulation routines from step (a) to implement and experimentally
demonstrate 32-QAM modulation.
(f) [Optional] Derive Equation 7.1.
7.3 Demonstration
Demonstrate your acoustic modem.
7.4 Summary
In Chapter 7, the physical layer communication design steps explored
in Chapters 1 through 6 have been integrated to implement and
verify a spectrally efficient modem operating at an acoustic car-
rier frequency. The implementation is designed for a flat-fading
channel. The baseband processing translates seamlessly to narrow-
band communication using a radio-frequency channel, as explored in
Appendix B. The explorations introduced new commands summa-

rized in Table 7.2.

Command Description
save save workspace variables to file
load load data from memory into a workspace
path get or set search path
addpath add a directory to search path
packetdetect.m function to detect a received packet
makepilots.m function to generate pilot sequence
conj conjugation
CHAPTER 8
Frequency-Selective Fading
The frame synchronization and frequency recovery procedures pre-

sented in Chapters 5 and 6 are done without regard to any distor-
tion due to the inter-symbol interference resulting from a frequency-
selective fading channel. Whereas a flat-fading channel only imparts
a gain and phase, a frequency-selective fading channel performs con-
volution, and thus may impart a gain and phase that varies as a func-
tion of frequency. The insertion of the channel’s filter response, h̃[k],
in the baseband model implies that the net effect of the pulse shap-
ing filter, gtx [k], the channel filter, h̃[k], and the receiver’s matched
∗ [−k] is no longer an ideal Nyquist pulse. For this reason,
filter gtx
the frequency-selective fading channel causes inter-symbol interfer-
ence and is sometimes called an ISI channel. The ISI is illustrated,
for BPSK signaling, in Figure 8.1.
In this chapter, frame timing and frequency recovery procedures

are modified to operate more robustly in the presence of a frequency-
selective fading channel. In addition, a linear equalizer is devel-
oped to combat the symbol distortion caused by an ISI channel.
In Chapters 10 and 11, orthogonal frequency division multiplexing
(OFDM) is presented as an alternative “frequency-domain” approach
to communication over a frequency-selective fading channel; OFDM
essentially reduces a frequency-selective fading channel to multiple,
narrow-band, flat fading channels.
95
96 CHAPTER 8. FREQUENCY-SELECTIVE FADING
Eye Diagram
1.5
0.5
-0.5
-1
-1.5
10 20 30 40 50 60 70 80 90
Eye Diagram
1.5
0.5
-0.5
-1
-1.5
10 20 30 40 50 60 70 80 90
Figure 8.1: BPSK eye diagrams for no ISI (top) and ISI (bottom).
8.1. BACKGROUND 97
8.1 Background
8.1.1 Frame timing for ISI channels
For frame timing in the presence of a frequency-selective fading chan-
nel, we consider the approach proposed by Moose [24]. The method
relies on a periodic pilot sequence and is used, for example, in IEEE
802.11a/g [15].
A length K FIR symbol-rate model of the channel impulse re-
sponse,
{h̃[n]}K−1
n=0 ,
is called a K-tap channel model. The model may physically cor-

respond, for example, to a multi-path channel. Then, the received
signal after matched filtering, symbol timing recovery, and downsam-
pling is given by
K−1
X
ỹ[n] = ej2πf∆ nT h̃[i]ã[n − i] + w̃[n], (8.1)
i=0
where T is the symbol period and f∆ is the frequency offset.

The Moose method is a self-referenced synchronization that avoids
the dependence on the channel response, h̃[n], by using a periodic
pilot sequence; the length 2Np pilot sequence is constructed by re-
peating a given marker code of length Np .
The idea for self-referenced frame timing recovery is to look for
a peak in the cross-correlation between the pair of received distorted
marker sequences:
PNp −1
i=K−1 ỹ[n + i + Np ]ỹ ∗ [n + i]
τbMoose = arg max qP q
PNp −1 .
n Np −1
i=K−1 |ỹ[n + i + Np ]|2
i=K−1 |ỹ[n + i]|2
(8.2)
Because the (K −1) symbols prior to the start of pilots are unknown,
the summation in the cross-correlation in Equation 8.2 runs from
(K − 1) to (Np − 1), rather than 0 to (Np − 1).
8.1.2 Frequency recovery for ISI channels

The periodic pilot sequence is likewise employed in the Moose ap-
proach to provide an estimate of the frequency offset that is robust
to ISI. Assume that, based on the timing recovery from Section 8.1.1
the received pilot sequence begins at index n = 0. Consider again

the convolution in Equation 8.1 and rewrite with an advance of Np
samples:
K−1
X
ỹ[n] = ej2πf∆ nT h̃[i]ã[n − i] + w̃[n] (8.3)
i=0
K−1
X
ỹ[n + Np ] = ej2πf∆ (n+Np )T h̃[i]ã[n + Np − i] + w̃[n + Np ]. (8.4)
i=0
Using the repeated structure of the pilot sequence, we know
ã[n + Np − i] = ã[n − i]. (8.5)
From Equations 8.3, 8.4 and 8.5 we have, for indices K − 1 ≤ n ≤

Np − 1,
K−1
X
j2πf∆ Np T j2πf∆ nT
ỹ[n + Np ] = e e h̃[i]ã[n − i] + w̃[n + Np ]
i=0
= ej2πf∆ Np T ỹ[n] + {w̃[n + Np ] − w̃[n]} . (8.6)
Thus we see that ỹ[n + Np ] equals ej2πf∆ Np T ỹ[n] plus a zero-mean

additive noise. That is, 2πf∆ Np T is the mean angle of the second
half of the received distorted pilots divided by the first half. Again,
we omit use of the first K − 1 samples because of their dependence
on unknown symbols prior to the pilots. Thus, we have a sample
mean to estimate the frequency offset:
 
 Np −1
1/T 1 X ỹ[n + Np ] 
fb∆ = angle . (8.7)
2πNp  (Np − K)
n=K−1
ỹ[n] 
By the 2π periodicity of the function ej2πf∆ Np T , the frequency

offset f∆ is unambiguous for

f∆ ≤ 1
.
2Np T
A longer pilot sequence (i.e., larger Np ) improves the noise averaging

in the estimator, but reduces the unambiguous range for the esti-
mated frequency offset, fb∆ . The Moose approach in Equation 8.7
provides an estimate of frequency offset that is robust to ISI.
8.1. BACKGROUND 99
noise w̃[n]
ỹ[n] q̃[n]
channel h̃[n] equalizer b̃[n]
pilots
+ ẽ[n]
ã[n]
delay ∆
Figure 8.2: Linear equalizer for a frequency-selective fading channel
8.1.3 Linear equalizer for ISI channels

Consider using the known pilot sequence to learn the channel impulse
response; a linear least-squares equalization approach is shown in the
baseband model depicted in Figure 8.2. In the figure, ỹ[n] is the
matched filter output downsampled to the symbol rate and post-
processed for frequency recovery and frame timing. Without loss
of generality, suppose the start of the pilot symbols is indexed by
n = 0. The sequence q̃[n] is the output of the equalization filter with
impulse response {b̃[0], ..., b̃[M ]}. The goal is to choose the filter
so that the output matches a delayed version of the known pilot
symbol sequence: q̃[n] ≈ ã[n − ∆]. The delay is necessary because
the equalizer is a causal filter.
The convolution of the FIR equalizer and the matched filter out-
puts produces q̃[n],
M
X
q̃[n] = b̃[k]ỹ[n − k]. (8.8)
k=0
For a pilot sequence of length 2Np , we seek to minimize the sum of

squared error between the equalizer output and a delayed copy of the
known pilot sequence. Replacing the delay, ∆, in the known pilot
sequence, ã[n], by an advance in q̃[n], we have the sum of squared
errors, J∆ (b), given by
2Np −1 M 2

X X
J∆ (b) = b̃[k]ỹ[n + ∆ − k] − ã[n] . (8.9)

n=0 k=0
The least-squares solution is simplified by writing the expression in

matrix-vector form:
   
ỹ[n+∆] ··· ỹ[n+∆−M ]   ã[0]
  b̃[0]  
 ỹ[n+∆+1] ··· ỹ[n+∆−M +1]    ã[1] 
  b̃[1]   
 ỹ[n+∆+2] ··· ỹ[n+∆−M +2]  .. ≈  ã[2] 
    
 .. .. ..  .   .. 
 . . .   . 
b̃[M ]
ỹ[n+∆+2Np −1] · · · ỹ[n+∆+2Np −M ] | {z } ã[2Np − 1]
| {z } | {z }
b̃ ã
Y∆
(8.10)
The equalizer that minimizes the sum of squared errors is given by
−1
H H
b̃∆,⋆ = Y∆ Y∆ Y∆ ã. (8.11)
The delay, ∆, may be found by computing the squared error for

each choice 0 ≤ ∆ ≤ M and selecting the value that results in the
smallest sum of squared errors. For long pilot sequences, a single
matrix inverse can be used to determine ∆, rather than exhaus-
tive search [18]. An alternative of least-squares is a minimum mean
squared error (MMSE) solution [31]; the MMSE solution balances
noise rejection against the distortion caused by ISI and channel gain.
8.2 Explorations
In this lesson, students modify their acoustic modem simulator from
Chapter 6 to implement frequency and timing recovery that is robust
to a frequency-selective fading channel.
For the laboratory exploration, work through the three steps be-
low to implement in simulation a bandpass modem for ISI channels.
Recommendations for preparing a brief, descriptive laboratory report
are included in the suggested steps below.
(a) Augment and revise your bandpass modem simulation code

from Chapter 6 to implement the Moose algorithm for frame
timing described in Equation 8.2 for an ISI channel.
At the bandpass signal, s[k], insert simulated impairments for
an ISI channel model with additive white complex Gaussian
noise. Separate the non-zero taps of your fractionally-sampled
h̃[k] by at least L samples; and, limit the extent from first to
last nonzero taps to LK or fewer samples.
Verify the operation of your bandpass modem in simulation. Q8.1
(b) Augment and revise your bandpass modem simulation code

from Chapter 6 to implement the Moose algorithm for fre-
quency recovery described in Equation 8.7 for an ISI channel.
Note that the computation in Equation 8.7 can be implemented
using these commands: ./, mean, and angle.
(c) To begin exploration of a least-squares equalizer, work through Q8.3

an illustrative low-order example. Let K = 1 be the order of
the equalizer and 2Np = 4 be the length of the pilot sequence,
ã = {1, −1, 1, −1}. Suppose the true channel is given by the
infinite impulse response system,
ỹ[n] = 0.5ã[n − 1] + 0.25ỹ[n − 1].
(a) With initial condition y[−1] = 1, use recursion to deter-

mine the noise-free sequence ỹ[n] for n = 0, 1, 2, 3.
(b) For ∆ = 0 and ∆ = 1, compute the equalizer, b̃∆,⋆ using
Equation 8.10 and Equation 8.11. Also, compute for each
∆ the corresponding sum of squared errors; select the ∆
yielding the smaller error.
(d) Augment and revise your bandpass modem simulation code

from Chapter 6 to implement a linear least-squares equalizer
to combat the impairment due to an ISI channel. To construct
the matrices in Equation 8.10, the command toeplitz may be
helpful; further, the least-squares solution in Equation 8.11 is
easily computed via b = pinv(Y)*a, where pinv is an abbre-
viation for Moore-Penrose psuedo-inverse.
(e) [Optional] Verify operation of your modem for an acoustic

ISI channel. Construct a frequency-selective channel by posi-
tioning transmitter and receiver to obtain resolved multi-path
interference. This is facilitated by selecting a small symbol
interval and low carrier frequency.
8.3 Demonstration
Demonstrate your modem simulator for a five-tap ISI channel and
other channel impairment variables of your choice.
8.4 Summary
In Chapter 8, a technique has been introduced to recover frame tim-
ing and a frequency offset in the presence of an ISI channel. In
addition, a linear equalizer was developed to combat the effect of the
channel impairment on data symbols. The explorations introduced
new commands summarized in Table 8.1.

Command Description
toeplitz construct a Toeplitz matrix
pinv pseudo-inverse
CHAPTER 9
Channel Coding
Fundamental limits for reliable communication over noisy channels

were characterized in the pioneering work of Claude Shannon [33].
Shannon gave answers to several fundamental questions: How does
one measure the amount of information? Can information be trans-
mitted with arbitrarily high accuracy in the presence of noise? If
so, at what rate? What is the minimum rate required to convey an
information source? What is the trade-off between reduced rate and
approximation error?
In this chapter, we consider Shannon’s surprising result that in-
formation can indeed be transmitted with arbitrarily high accuracy
in the presence of noise, and we characterize the rate. A simple block
code is implemented to gain some familiarity with channel coding.
9.1 Background
9.1.1 Channel capacity
Suppose that we encounter a low signal-to-noise ratio (SNR), but
would like to achieve a low probability of error. Into the 1940s,
the conventional wisdom was to repeat the same symbol multiple
times and use a majority vote at the receiver. While this approach
does reduce the probability of error, the rate of information bits per
symbol reduces linearly with the number of repetitions. Indeed, to
drive the error arbitrarily small, the bit rate must go to zero!
Using elementary arguments, Shannon showed that it is indeed
possible to communicate at an arbitrarily low probability of error
yet maintain a non-zero data rate, C, called the channel capacity.
103
104 CHAPTER 9. CHANNEL CODING
For additive white Gaussian noise (AWGN), an ideal band-limited

channel of two-sided bandwidth B has a capacity C given by

P
C = B log2 1+ bits per second (9.1)
BN0
where P watts is the received signal power and N0 /2 watts/Hz is
the power density of the noise. What is the meaning of capacity? If
the information rate R bits/s from the source is less than C, then it
is theoretically possible to achieve reliable transmission through the
channel by appropriate coding [27]. In contrast, if R > C, reliable
transmission is not possible using any processing scheme.
For example, consider telephone line modems. To allow for mul-
tiplexing of many channels, signals are bandlimited to approximately
3300 Hz. A bandwidth of 3300 Hz and signal-to-noise ratio (SNR) of
P/N0 B = 31 dB in Eqn. 9.1 yields a capacity of about 34,000 bits
per second. This limit is approached by commercial modems, which
work at 33.6 kbps, yielding an astounding bandwidth efficiency of
over 10 data bits per second per Hertz. For fax modems, the In-
ternational Telecommunication Union (ITU) V.34 standard is able
to fully adapt to the telephone line quality by adjusting the car-
rier frequency, the data rate, and the transmitted power to fit the
communication channel.
The decades since Shannon’s pioneering work have witnessed a
steady increase in the market demand for data transmission and the
steady progress of engineering techniques pushing ever more closely
to the fundamental theoretical limits [38].
9.1.2 Channel coding

Error correction codes add structured redundancy to an information
sequence in order to detect and correct possible bit errors that occur
during transmission. The basic idea of channel coding in digital
communication is the use of parity bits, and the concept is very
much analogous to spell checking. For example, a reader of English
can identify and correct errors in the following text string:
With malice towark none, weth chariti for all
The three errors are identifiable because not all text strings are al-
lowable words. That is, English has redundancy. A reader’s guess
at the correction may be described as making the fewest changes so
9.1. BACKGROUND 105
that the corrected word is allowable. For example, by changing only

one letter, “k,” the illegal string “towark” becomes the allowable
word “toward.” Further, by considering longer text strings, a reader
may use grammar and context to achieve more sophisticated error
checking.
Table 9.1 illustrates application of the structured redundancy
idea to binary strings. The left columns list the 16 possible infor-
Table 9.1: An example (7, 4) block code.
Information bits Code word

0000 0000000
1000 1000110
0100 0100101
1100 1100011
0010 0010011
1010 1010101
0110 0110110
1110 1110000
0001 0001111
1001 1001001
0101 0101010
1101 1101100
0011 0011100
1011 1011010
0111 0111001
1111 1111111
mation symbols formed using 4 bits; there is no redundancy. In the

right columns, redundancy is introduced by appending three addi-
tional bits; hence, structured redundancy is inserted. This example
is called a (7, 4) block code because blocks of k = 4 bits are mapped
to coded blocks of n = 7 bits. Block codes are perhaps the simplest
of all error correcting codes to both understand and implement.
An (n, k) code has code rate of k/n because the redundancy re-
duces the data transmission rate by the factor k/n. The implemen-
tation of the (7, 4) code may be achieved by binary matrix multipli-
cation using a matrix of four rows and seven columns. For example,
consider the sixth row of Table 9.1.

 
1 0 0 0 1 1 0
h i 0 1 0 0 1 0 1  h i
 
1 0 1 0  = 1 0 1 0 1 0 1
| {z } 0 0 1 0 0 1 1  | {z }
k input bits 0 0 0 1 1 1 1 n encoded bits
| {z }
k×n generator matrix
(9.2)
The arithmetic here, and throughout this chapter, is on the field
of binary numbers, whereby addition is exclusive OR and multi-
plication is logical AND. Equivalently, the arithmetic is modulo-2.
Note that only the last three columns of the generator matrix, G, in
Equation 9.2 need to be stored and used to generate the three bits
appended to the four bit information string. Alternatively, a block
code may be efficiently implemented using a linear feedback shift
register.
The appended bits may also be interpreted as parity bits. Let
H T be a n-by-(n − k) binary matrix satisfying
GH T = 0. (9.3)
Because any allowable code word y is a binary sum of rows from G,
a code word must satisfy
yH T = 0. (9.4)
Thus, H T is called a parity check matrix. From the example above,
let x1 , x2 , x3 , x4 denote the four information bits and let c1 , c2 , c3
denote the three bits appended by the (7, 4) code. We have three
parity check equations:
x1 + x2 + x4 + c1 = 0
x1 + x3 + x4 + c2 = 0 (9.5)
x2 + x3 + x4 + c3 = 0.
For this example, the parity check matrix is given by
 
1 1 0
 
 1 0 1 
 

 0 1 1 

HT =
 1 1 1 .
 (9.6)
 1 0 0 
 
 
 0 1 0 
0 0 1
9.1. BACKGROUND 107
Decoding requires making the best guess of the closest allowable

code word if the received word does not satisfy the parity checks. A
closest allowable word requires a concept of distance. The Hamming
distance between two binary strings of equal length is defined as the
number of positions in which the two strings differ. For example, the
two words
h i
1 0 1 0 0 1 1
h i
1 0 1 0 1 0 1
have a Hamming distance of 2, because the strings differ only in

the fifth and sixth positions. For minimum distance decoding, the
decoder compares the received code word with the M = 2k possible
allowable transmitted code words and decides in favor of the code
word that is closest in Hamming distance. This exhaustive search,
while conceptually simple, is computationally inefficient. A more
efficient method for this minimum distance decoding uses the parity
check matrix, H T , as seen in Section 9.1.3 below.
The minimum distance between any two allowable code words in
the example above is d = 3. Therefore, the code can detect up to
two errors, but three errors might result in an allowable code word.
Additionally, the code permits correction of one error per code word,
because only one unique allowable code word has Hamming distance
1 from a code word received with a single error. In general, an (n, k)
code of minimum distance d can detect d − 1 errors and correct
⌊ 21 (d − 1)⌋ errors, where ⌊x⌋ denotes the largest integer less than or
equal to x.
Block codes are simple to implement. Key characteristics of a
block code are the minimum distance and the rate. A larger block
size permits a higher minimum distance, but requires more delay. A
higher rate code allows for transmission of more data bits per second,
but can correct fewer errors.
9.1.3 Syndrome decoding
Syndrome decoding is an efficient alternative to exhaustive search

for decoding linear block codes. Consider a received string of bits,
y = c + e, that is a codeword, c, plus an error string, e. From
Equation 9.4, we apply the parity check matrix, H T , to learn
yH T = cH T + eH T
= 0 + eH T
= s (9.7)
where s = eH T is the (n − k)-bit syndrome of the error sequence, e.

If the received string has zero or one error, then the syndrome can
be used to return the correct codeword. For example, consider the
(7, 4) block code described above. With a minimum distance of 3,
the code can detect and correct all one-bit errors. For this case, the
syndromes of all one-bit errors are given in Table 9.2. If no errors
Table 9.2: Single error syndromes for a (7, 4) block code.
Error Pattern ×H T Syndrome

[1 0 0 0 0 0 0]H T [1 1 0]
[0 1 0 0 0 0 0]H T [1 0 1]
[0 0 1 0 0 0 0]H T [0 1 1]
[0 0 0 1 0 0 0]H T [1 1 1]
[0 0 0 0 1 0 0]H T [1 0 0]
[0 0 0 0 0 1 0]H T [0 1 0]
[0 0 0 0 0 0 1]H T [0 0 1]
occured, then the syndrome is zero. If one error occurs, then the non-
zero syndrome identifies the most likely error pattern, ê; this error
vector is then added to y to yield the decoded string of bits. For two
or more errors, the syndrome decoding may or may not return the
correct codeword. For block codes with minimum distance greater
than 3, syndrome decoding can be extended to detect and correct
more than one bit error per block of n received bits.
If bit errors occur in bursts, then direct use of a channel code will
not allow detection and correction of the cluster of errors. An effec-
tive method of dealing with burst errors is to interleave (shuffle) the
coded data so that the channel with bursts of errors is approximately
transformed into a channel with independent errors [35]. A simple
approach is the block interleaver, which reads in coded bits row by
row, then reads out bits to the modulator column by column.
The 67 years of coding theory since Shannon’s seminal paper

have witnessed remarkable progress in approximatey achieving chan-
nel capacity through use of more sophisticated codes requiring more
complex processing for decoding. Clever encoding schemes provide
a large Hamming distance between code words, short block lengths
(hence low decoding latency), and simple decoding procedures. For
example, an audio CD uses a cross interleaved Reed-Solomon code;
4G wireless systems use soft-decision decoding, turbo codes, and
trellis-coded modulation.
9.2 Explorations
In this chapter, students implement and experimentally verify a (7, 4)
block code. The brief laboratory report should contain answers to
the questions enumerated in the margins; subroutines for channel
coding and decoding with the (7, 4) block code should be appended.
(a) Block codes are often listed by triples, (n, k, d), where (n−k) is Q9.1
the number of parity bits, k is the number of information bits,
and d is the minimum distance between codewords. Compare
and contrast the rates for three block codes: (7, 4, 3), (8, 4, 4),
(16, 5, 8).
(b) Expand your BPSK simulation from Chapter 3 to include chan- Q9.2
nel coding. Create function calls that perform (7, 4) block cod-
ing and decoding. Repeat the bit error rate simulation from the
demonstration in Section 3.3, but generate the plot for Eb /N0
only on the range 0 to 10 dB. For comparison, overlayp the the-
oretical curve for uncoded BPSK data, BER = Q( 2Eb /N0 ),
found in Chapter 3. For your report, present the BER curve
and append code for your block coding and decoding functions.
Although binary arithmetic is computationally efficient, a Mat-
lab implementation of yG using double precision real variables
is simple to compute using modulo-2 arithmetic, mod(y*G,2).
For example, re-use random bit generation from Chapter 3:
k = 4; N = 250;
bits = round(rand(1,k*N));%random 0’s and 1’s
% use function from part 9(b)
codedbits = blockcode74(bits);
% convert bits to BPSK symbols

symbols = 2*codedbits-1;
Consider these four suggestions for using the syndrome to cor-

rectly decode for 0 or 1 bit error per block of n bits:
(i) Compute the length (n − k) syndrome using the parity

check matrix.
(ii) If the syndrome is zero, then return the first k bits as the
message.
(iii) If the syndrome is not zero, use the syndrome table to
index a single bit error. Flip the detected error and return
the first k bits as the message. (Only the first k syndrome
vectors in Table 9.2 need be used.)
(iv) In general (but not necessary for this (7, 4, 3) block code):
if there is no match of the syndrome to the first k syn-
drome vectors, then return the first k bits unchanged.
Q9.3 (c) Block codes are often listed by triples, (n, k, d), where (n − k)
is the number of parity bits, k is the number of information
bits, and d is the minimum distance between codewords. The
probability of error for a (n, k, d) linear block code is bounded
by [21] s !
k k 2Eb
P ≤ (2 − 1)Q d , (9.8)
n N0
whereas the probability of error for the uncoded case was found
in Chapter 3 to be
s !
2Eb
P =Q (9.9)
N0
The effective coding gain is defined as the additional SNR

required for the uncoded data to experience the same error
rate as the coded data. By using the approximation Q(x) ≤
exp(−x2 /2), we can find the comparison

Eb k k ln 2
P ≤ exp − d− coded
N0 n Eb /N0

Eb
P ≤ exp − uncoded. (9.10)
N0
Thus, we conclude that the effective coding gain for an (n, k, d)

code is approximately
k k ln 2
effective coding gain ≈ d− . (9.11)
n Eb /N0
Referring to Equation 9.11, plot the effective coding gain, in

dB, versus Eb /N0 , for three block codes: (7, 4, 3), (8, 4, 4),
(16, 5, 8). Generate the plot for the range Eb /N0 ∈ [5, 20] dB.
(Note: use 10 log 10 to express these ratios of power quantities
in dB.)
(d) [Optional] Shannon’s capacity implies a lower bound on Eb /N0

below which arbitarily reliable communication cannot occur. In
this exercise, you are asked to derive that bound [21]. To get
started, normalize to B = 1 for simplicity and observe that
ln(1 + x) ≤ x. Then, from Equation 9.1 we have
C = log2 (1 + SNR)
1
= ln (1 + SNR)
ln 2
1
≤ SNR.
ln 2
Continuing, use the definitions SNR = Es /N0 and Es = Eb R,
where Es is energy per symbol, Eb is energy per bit, and Rb
is bit rate. The derivation also requires that bit rate cannot
exceed capacity: Rb ≤ C. Express your answer in dB by com-
puting 10 log10 of the power quantity and rounding to two sig-
nificant digits.
(e) [Optional] Reformat your software from Chapters 3 and 9

into separate, modular functions for the following steps: text
to data bits; data bits to channel-coded bits; bits to symbols.
Similarly, for the receiver processing reformat your software
into separate, modular functions for the following steps: symbol
detection; symbols to bits; channel decoding; bits to text.
9.3 Demonstration
Demonstrate execution of the codes developed in step (b).
9.4 Summary
In Chapter 9, the concept of channel capacity was introduced as a
fundamental bound on the rate at which data can be transmitted
through a noisy channel. Block coding was presented as a simple
way to insert structured redundancy into a transmitted bit stream, in
order to allow detection and correction of errors at the receiver. The
explorations introduced a new command summarized in Table 9.3.

Command Description
mod modulus after division
CHAPTER 10
Orthogonal Frequency Division Multiplexing
Part I
In Chapter 8, we considered approaches to combat channel impair-

ments for a frequency-selective fading channel, also known as an inter-
symbol interference (ISI) channel. The frequency-selective fading
channel is modeled as the convolution of a channel impulse response
with the baseband message, in contrast to the simple complex-valued
gain that serves to model a flat fading channel. In particular, in
Chapter 8 an equalization filter was implemented to approximately
undo the effects of the ISI channel; the equalization filter can be con-
sidered a “time-domain” approach. In contrast, in Chapter 10 and
Chapter 11 we present a “frequency-domain” approach that can of-
fer robust performance for a frequency-selective fading channel with
lower complexity than the time-domain equalization. Conceptually,
the frequency-domain approach is to divide the wide-band ISI chan-
nel into multiple non-interfering narrowband sub-channels, then ex-
ploit the simplicity of channel equalization when the sub-channel
model is merely a complex-valued gain.
The explorations presented in Chapters 10 and 11 provide a tu-
torial for orthogonal frequency division multiplexing (OFDM) com-
munication [3] and present a sequence of exercises guiding students
to the implementation of OFDM in an acoustic modem. OFDM is
used, for example, in 4G mobile communication.
In this chapter, we will derive a frequency-domain equalization
technique; in the next chapter, we then show how this technique
can be further modified to OFDM, where we decompose a multi-tap
channel into a set of parallel single-tap channels.
113
114 CHAPTER 10. OFDM PART 1
10.1 Background
In this and the next chapter, we focus on the discrete-time baseband
equivalent channel model,
ỹ = h̃ ∗ ã + w̃. (10.1)
In Equation 10.1, ã is a vector representing a sequence of complex-

valued symbols and ỹ represents the corresponding received symbols,
after ã goes through an equivalent channel with impulse response h̃
and is corrupted with the white Gaussian noise w̃. In this symbol-
rate model, the convolution of the transmitter’s pulse shaping filter
and receiver’s filter is presumed to yield an ideal Nyquist pulse. To
ease notation, in this chapter and the next, we ignore the noise term
w̃ and drop the tilde signs1 .
Assume that the channel response h is a vector with size L, i.e.,
the channel has L taps. Then, we can rewrite Equation 10.1 as
PL−1
y[n] = l=0 h[l]a[n − l]. (10.2)
Note that, while implementing an acoustic modem in Chapters 1

through 7, the focus was on narrowband communication, where the
bandwidth of the bandpass transmission signal is small enough (or
equivalently, the symbol-rate is low enough) such that the channel
can be effectively modeled using a single-tap response, resulting in
the flat fading input-output relationship:
y[n] = h × a[n]. (10.3)
A nice property of this single-tap channel is that it is memoryless,

meaning that the nth output symbol, y[n], is a function of only the
nth input symbol, a[n]. Therefore, to recover the input symbol a[n],
we can simply estimate the channel gain h and then divide the output
symbol y[n] by h.
Driven by ever-growing traffic demand, most wireless systems op-
erate in the wideband regime, meaning that the symbol rate is suffi-
ciently large so that different frequency components of the transmit-
ted signal experience different channel gains. This wideband phe-
nomenon is often referred to as frequency selectivity or frequency
1
The tilde sign was employed in Chapter 1 through Chapter 8 to differentiate
between real-valued and complex-valued signals.
10.1. BACKGROUND 115
selective fading. As a result, the single-tap model is not able to fully

capture the physical channel effect for a wideband channel in which
an output symbol is affected by more than one input symbol. For a
wideband channel, the multi-tap convolution model of Equations 8.1
and 10.2 is required.
With a multi-tap channel, ISI occurs: each received symbol is
the weighted sum of L input symbols, with the weights determined
by the channel gains on each delayed tap, and each input symbol
affects L output symbols, which is evident from Equation 10.2. As
a result, and different from the single-tap case for a flat fading chan-
nel, the channel effect can no longer be equalized by simply dividing
the received symbols by the estimated channel gain. The challenge
is then to recover the input symbols from the ISI-corrupted output
symbols. In contrast to the time-domain equalization filter explored
in Chapter 8, in this chapter we present a low-complexity, robust al-
ternative that can be interpreted as a frequency-domain equalization
approach. For additional approaches to ISI channels, the reader is
referred to maximum-likelihood sequence detection (e.g., the Viterbi
algorithm) and direct-sequence spread spectrum (CDMA).
10.1.1 Frequency-domain equalization

Once again, consider the discrete-time wide-band multi-tap channel
model
y = h ∗ a, (10.4)
where y and h are complex-valued vectors of received symbols and

channel impulse response samples, respectively. The challenge is to
recover the symbol sequence a. In other words, we seek to invert
the convolution operation. The beauty of the discrete-time Fourier
transform is to convert convolution of two sequences into a simple
point-wise multiplication of their transforms. That is, let FDT denote
the discrete-time Fourier transform operator; then, the convolution
y = h ∗ a is equivalent to FDT (y) = FDT (h) × FDT (a), and the
sequence a can be recovered via the inverse discrete-time Fourier
transform of the ratio:

−1 FDT (y)
a = FDT . (10.5)
FDT (h)
For practical implementation, we must work with samples of the

discrete-time Fourier transform operator, rather than directly with
integration and multiplication of the continuous functions, such as
FDT (h), defined on the interval [−π, π). Thus, we resort to the dis-
crete Fourier transform (DFT) and employ fast algorithms (FFT) for
its computation [16]. However, by virtue of working with frequency
samples, the multiplication of frequency-domain samples becomes
equivalent to the circular convolution (periodic convolution, or cyclic
convolution) in the time domain. More precisely, assume that h and
a are complex vectors with dimension L and N , respectively, with
L < N . We have
DFTN (h ⊛ a) = DFTN (h) × DFTN (a), (10.6)
where DFT(·)N is the N -point discrete Fourier transform. The cir-

cular convolution is denoted by ⊛ and defined as
PL−1
(h ⊛ a)[n] = l=0 h[l]a[(n − l) mod N ] with 0 ≤ n ≤ N − 1.
(10.7)
If the the circular convolution of h and a is known at the receiver,

then a can be recovered by

DFTN (h ⊛ a)
a = IDFTN . (10.8)
DFTN (h)
We use uppercase letters to denote the DFT of the vector given by

the corrsponding lowercase symbol; the defining summations for the
DFT and the inverse DFT (IDFT) are given by
1 NX−1
Xk = √ xn e−j2πkn/N (10.9)
N n=0
1 NX−1
xn = √ Xk ej2πkn/N . (10.10)
N k=0
Given the channel model shown in Equation 10.2, the question

is how to modify the sequence of transmitted symbols such that
the linear convolution induced by the channel can yield the circular
convolution as shown in Equation 10.7. To this end, let us first
construct a sequence of symbols a
b with length N + P by inserting a
prefix with length P to a:

a
b= a[N − P ], . . . , a[N − 1] , a[0], a[1], . . . , a[N − 1] . (10.11)
| {z } | {z }
Cyclic Prefix: last P symbols in a a
Since the prefix consists of the last P symbols in a, it is often called

the cyclic prefix (CP). After ab is passed through the multi-tap chan-
nel with channel response h, we can obtain
L−1
X
yb[n] = (a
b ∗ h)[n] = h[l]a
b[n − l]. (10.12)
l=0
If P ≥ L − 1, then for any 0 ≤ n ≤ N − 1, the above equation yields

L−1
X
yb[n + P ] = (a
b ∗ h)[n + P ] = h[l]a
b[n + P − l]
l=0
L−1
!
X a[(n + P − l) − P ] if n + P − l ≥ P
= h[l]
a[(n + P − l) + (N − P )] else
l=0
L−1
!
X a[n + N − l] if n ≤ l − 1
= h[l]
a[n − l] if n ≥ l
l=0
L−1
X
= h[l]a[(n − l) mod N ]
l=0
= (h ⊛ a)[n], (10.13)
where the second equality follows from the definition of a b.

Equation 10.13 reveals that by prepending a cyclic prefix with
length no less than the number of channel taps to the data symbols,
and passing the modified symbols to the channel, the channel output
will contain a copy of the circular convolution of the channel response
and the data symbols (see Figure 10.1 for an illustration). Then,
at the receiver, with the knowledge of h, the data symbols can be
recovered via Equation 10.8.
Thus, the use of a cyclic prefix is a ploy at the transmitter so
that a portion of the received symbol sequence matches the circular
convolution, despite the physical channel behaving as a linear con-
volution. In order to obtain the circular convolution between the
length-N data symbols and the length-L channel filter, the number
of transmitted symbols (which includes both data symbols and the
[1 2 3] [5 4 6 1 3 2]
= 1 x [5 4 6 1 3 2]
+ 2 x [2 5 4 6 1 3]
+ 3 x [3 2 5 4 6 1]
= [18 20 29 25 23 11]
[1 2 3] ∗ [3 2 5 4 6 1 3 2]
= 1 x [3 2 5 4 6 1 3 2]
+ 2x [3 2 5 4 6 1 3 2]
+ 3x [3 2 5 4 6 1 3 2]
= [3 8 18 20 29 25 23 11 13 6]
Figure 10.1: An illustration for Equation 10.13 for a = [5, 4, 6, 1, 3, 2]

and h = [1, 2, 3].
cyclic prefix symbols) must be at least N + L − 1, giving rise to a

transmission overhead that is the fraction (L − 1)/(N + L − 1) of the
total transmitted symbols.
10.1.2 Frequency-domain interpretation

In order to gain more insight into the frequency-domain equalization
technique, let us look at a specific example. Consider the following
complex-valued symbol vector with 10 entries:
0·k 1·k 2·k 3·k 9·k
ak = [ej2π 10 , ej2π 10 , ej2π 10 , ej2π N , . . . , ej2π 10 ].
Observe that
|ak | = [1, 1, 1, 1, . . . , 1],

k 0·k 1·k 2·k 3·k 9·k
angle(a ) = 2π , 2π , 2π , 2π , . . . , 2π ,
10 10 10 10 10
which indicates that the vector ak is a sequence of 10 samples taken

from 10 equally-spaced positions along the unit circle of the com-
plex plane; the sequence ak steps, or “rotates,” exactly k positions
along the circle between consecutive samples. From Equation 10.9
we further know that DFT10 (ak ) is vector with all but the kth entry
being zero. In other words, the symbol sequence ak only has the kth
frequency component.
According to Equation 10.6, which we restate below, if we trans-
mit ak on the channel with response h, then the received signal
also only has the kth frequency component. More precisely, we will
have the received signal being Hk ak , where Hk is the kth element in
DFT10 (h).
DFTN (h ⊛ a) = DFTN (h) × DFTN (a).
In Figure 10.2, we plot both DFT(h) and the absolute value of the
received the symbols after transmitting ak for ten sample times with
k = 8 through a channel with response h. From the figure, we can
verify that the gain in amplitude after transmitting a8 is |H8 |, which
conforms with Equation 10.6.
DFT 10(h)
10 10
9 9 a 8 (10 back-to-back trasnmissions)

received symbols after channel h
8 8
7 7
absolute value
6 6
5 5
4 4
3 3
2 2
1 1
0 0
0 2 4 6 8 10 20 30 40 50 60 70 80
frequency index sample index
Figure 10.2: Absolute values of the received symbols after transmit-

ting the symbol vector ak , k = 8, for ten time steps through a channel
with response h = [4 + 1j, 3 − 1j, 2, 1].
10.2 Explorations
A lab report should provide answers to the following questions.
(a) Given the definition of DFT in Equation 10.9, verify the cor- Q10.1
rectness of Equation 10.6.
Q10.2 (b) Let x1 = [1, 2, 3, 4, 5, 6], x2 = [5, 2, 4, 3, 1, 6], and h = [1, 2, 1, 1].
(i) Find x1 ⊛ h and x2 ⊛ h.

(ii) Find x1 ∗ h and x2 ∗ h.
(iii) Construct a vector x such that x ∗ y contains a copy of
both x1 ⊛ h and x2 ⊛ h.
(iv) What is the length of the shortest x that conforms to step
(iii)?
Q10.3 (c) For N = 6, find DFTN (x) for x = [1, 1, 1, 1, 1, 1]. Repeat for
x = [1, −1, 1, −1, 1, −1].
Q10.4 (d) Reproduce the result in Figure 10.2 with k = 0, 1, · · · , 9. Ex-

plain what you observe.
10.3 Demonstration
Demonstrate the execution of the codes developed in step (d).
10.4 Summary
In Chapter 10, we introduced a frequency-domain channel equal-
ization technique for combating frequency selectivity in wide-band
wireless systems. We showed that by prepending a cyclic prefix with
length larger than the channel memory onto the transmitted packet,
we can equalize the channel gain on a per-frequency-tone basis with
low-complexity FFT and IFFT operations. This chapter serves as
a stepping stone for implementing OFDM in the next chapter. The
explorations introduced new commands summarized in Table 10.1.

Command Description
fft fast algorithm to compute the
discrete Fourier transform
ifft fast algorithm to compute the
inverse discrete Fourier transform
CHAPTER 11
Orthogonal Frequency Division Multiplexing
Part II
In the previous chapter, we introduced a method to equalize the

effect of a multi-tap channel by prepending a cyclic prefix to the
data symbols, converting the received symbols into the frequency
domain, equalizing the channel effects on each frequency component,
and then converting the modified frequency domain signal back to
the time domain. This frequency domain equalization technique is
summarized in Equation 11.1:
add cyclic-prefix L-tap channel discard first P

a −−−−−−−−−→ a
b===========⇒ a
b ∗ h −−−−−−−−−→ (h ⊛ a)
with length P with response h received symbols
N -point DFT Divide H
(h ⊛ a) −−−−−−−−−→ H · A −−−−−−−−−→ a (11.1)
N -point IDFT
Observe two characteristics of this technique. First, most of the

computational complexity resides at the receiver side: the receiver
needs to perform DFT and IDFT, while the transmitter just needs
to prepend a cyclic prefix to the data symbols. Second, the receiver
needs to know the channel impulse response. Therefore, two natural
questions arise: (i) Can we turn the multi-tap channel into a set of
parallel single-tap sub-channels so that the channel effect on each
sub-channel can be estimated and equalized independently? (ii) Can
we modify this technique so that the complexity is shared evenly
between the sender and the receiver? We aim to address these two
questions in this chapter.
121
11.1 Background
11.1.1 Orthogonal frequency division multiplexing
Let us simplify the notation for the procedures shown in Equation 11.1
as the following.
equivalent channel
a ===========⇒ H · A
Note that after a is sent through the equivalent channel, the receiver
obtains the point-wise product of the N -point channel frequency re-
sponse and the N -point DFT of a. This point-wise product is the
key to realizing that it is possible to convert the multi-tap channel
into N single-tap channels, with each channel corresponding to a
term in the point-wise product. The idea is the following: instead of
sending a, let us send the N -point IDFT version of a, denoted as b,
through the equivalent channel, which results in the procedures in
Equation 11.2.
N -point IDFT equivalent channel

a −−−−−−−−−→ b ===========⇒ H · B = H · a (11.2)
The last equation holds because b = IDFTN (a), which leads to

B = DFTN (b) = DFTN (IDFTN (a)) = a. By substituting the
equivalent channel with the procedures in Equation 11.1, we have
the complete procedure:
N -point IDFT add cyclic-prefix L-tap channel

b=
a −−−−−−−−−→ b −−−−−−−−−→ b b∗h
==========⇒ b
with length P with response h
discard first P N -point DFT
b ∗ h −−−−−−−−→ (h ⊛ b) −−−−−−−−−→ H · a
b (11.3)
received symbols
This small modification has two important implications. First,

the multi-tap channel is successfully decomposed into N single-tap
channels. Specifically, for any data symbol a[n] in a, the receiver
will obtain a scaled version of it as H[n]a[n], where H[n] is the nth
element of the N -point DFT of the channel response h. Second,
the computational burden is evenly shared between the transmitter
and the receiver: the transmitter must compute one IDFT, and the
receiver must compute one DFT. The complexity of DFT and IDFT,
using the FFT and IFFT algorithms, is O (N log2 N ).
a b b
...
...
Bits Modulation Symbols
...
Series
Parallel to
Bits Modulation Symbols Add Prepend
Parallel
Series to
Pilot
CP OFDM
...
IFFT Packet
Bits
...
Bits Modulation Symbols Symbols Symbols
... ... ...
...
h Multi-tap
Channel
...
...
Parallel
Series to
CP
Discard
Bits Modulation Symbols Remove
Series
Parallel to
OFDM Pilot Packet

...
Bits FFT
...
... ... ...
...
H·a b h b∗h if P ≥ L − 1
Figure 11.1: Block diagram of an OFDM system.
To gain insight into the first implication, let us consider the

case when a series of N -dimensional symbol vectors, denoted as
a1 , a2 , a3 , . . ., is sent. After the procedures in Equation 11.3, the
receiver will obtain
H[0]a1 [0] H[0]a2 [0] H[0]a3 [0] ...
H[1]a1 [1] H[1]a2 [1] H[1]a3 [1] ...
.. .. ..
. . ... .
H[N − 1]a1 [N − 1] H[N − 1]a2 [N − 1] H[N − 1]a3 [N − 1] . . .
From the above result, we can observe that the nth symbol in each
symbol vector, a1 , a2 , a3 , · · · , experiences the same one-tap channel
gain H[n]. In other words, the procedure in Equation 11.3 effectively
converts the multi-tap channel into N -single tap channels, where for
any symbol vector a, the nth symbol experiences a channel with gain
H[n].
Because the transmitted symbol vector, b, is the IDFT of the
data symbol vector, a, we know that the data symbols in a are the
real component of the 5 eigen-functions of FFT5

1
0.5
-0.5
-1
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
imag component of the 5 eigen-functions of FFT5

1
0.5
-0.5
-1
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Figure 11.2: The time domain waves of the five frequency subcarriers
in DFT5 .
frequency domain representation of the transmitted symbols. In this

sense, OFDM transmission may be interpreted as directly modulat-
ing data as a gain and phase on each of N orthogonal carrier tones
or “subcarrier frequencies,” and the total transmitted waveform is
the sum of these quadrature amplitude modulated subcarrier signals.
The channel gain on the nth frequency subcarrier is equal to H[n]. To
illustrate, we plot in Figure 11.2 the five time-domain signals (eigen-
functions) that correspond to the five frequency subcarriers if IFFT5
and FFT5 are used in the OFDM scheme described in Figure 11.1.
It is easy to verify that these five waves are orthogonal. For any
input vector [a[0], a[1], a[2], a[3], a[4]] to the IFFT5 block, the five
values a[0], a[1], a[2], a[3], and a[4] are modulated directly on these
five orthogonal waves.
The complete OFDM system is illustrated in Figure 11.1. In
Figure 11.3, we highlight the process in the first subcarrier. In an
OFDM system, the symbol vectors are transmitted in blocks of many
symbols concatenated with no delay in between. Since each OFDM
symbol vector corresponds to a group of data symbols modulated
on orthogonal frequencies, the back-to-back transmission results in
a[0]
...
...
...
Series
Parallel to
Bits Modulation Symbols Add Prepend
Parallel
Series to
Pilot
CP OFDM
...
IFFT Packet
Bits
...
... ... ...
...
Single-tap Multi-tap
Channel
H[0] Channel
...
...
Parallel
Series to
CP
Discard
Bits Modulation Symbols Remove
Series
Parallel to
OFDM Pilot Packet

...
Bits FFT
...
... ... ...
...
H[0] · a[0] if P ≥ L − 1
Figure 11.3: Block diagram of a single sub-channel.
a time-frequency resource grid, with each grid entry corresponding

to a single complex-valued data symbol. This time-frequency grid
is depicted in Figure 11.4. Each row in the time-frequency resource
grid represents a frequency subcarrier, while each column represents
the time duration for a single OFDM symbol vector.
Note that the cyclic prefix at the beginning of each OFDM sym-
bol vector not only ensures that the data symbols are modulated
onto orthogonal subcarrier frequencies (in that it provides a way to
implement circular convolution through the channel), but also serves
as a guard interval in time that prevents the transmission of the pre-
vious OFDM symbol vector from interfering with the current OFDM
symbol vector.
Denote the symbol sampling period as TSym second and the time
for a complete OFDM symbol vector as TOFDM second. Because one
symbol is transmitted per subcarrier per TOFDM second, we know
that the bandwidth of a subcarrier is 1/TOFDM Hz. On the other
hand, the total bandwidth is 1/TSym Hz. Therefore, the total num-
ber of subcarriers, including those devoted to the cyclic prefix sym-
bols, is roughly 1/TSym /(1/TOFDM ) = TOFDM /TSym , which conforms
Pilot Symbols CP CP ... CP
OFDM Symbol OFDM Symbol OFDM Symbol
Time-Frequency
Resource Grid
Total frequency bandwidth

am [0] subcarrier 0
Frequency
a1 [0] a2 [0]
a1 [1] a2 [1] am [1] subcarrier 1
a1 [3] a2 [3] ... am [3] subcarrier 3
... ... ...
a1 [N − 1] a2 [N − 1] am [N − 1] subcarrier N-1
Time
Figure 11.4: OFDM time-frequency resource grid. The shaded grid

entries correspond to a possible selection of reference symbols, which
are used by the receiver as known pilot symobls for channel esti-
mation. The channel gain for the remaining grid positions can be
estimated by interpolation.
with our time-domain observation. For a fixed OFDM symbol length

TOFDM second, the number of subcarriers may be varied based on the
total bandwidth available for transmission, leading to a spectrally
flexible system design.
11.1.2 Channel estimation

To estimate the channel gain on different subcarriers, one can send
both data symbols and reference symbols in the time-frequency grid.
Prior to the start of the transmission, the transmitter and the re-
ceiver agree on the values and locations of the reference symbols so
that the channel gain can be measured at the reference symbol lo-
cations. The channel gain at the data symbol locations can then
by estimated by interpolating the complex-valued gains of the ref-
erence symbols locations. Figure 11.4 shows a possible selection of
reference symbol positions, which are marked as shaded rectangles
in the time-frequency grid. The reference symbol locations are typi-
cally designed to stagger across both time and frequency, providing
robustness against channel fading in both time and frequency.
One can also eliminate the need for channel estimation by apply-
ing to each subcarrier a differential modulation/demodulation tech-

nique (such as DBPSK, DQPSK from Section 7.1.4), which relies on
the relative phase to encode/decode information bits. The price to
pay for this simplicity, however, is the overhead caused by the trans-
mission of the initial reference symbol on each subcarrier and a slight
increase in the symbol error rate.
11.2 Explorations
A lab report should provide answers to the following questions.
(a) Verify that the OFDM transceiver architecture described in Q11.1

Figure 11.1 is captured by the following equation
N -point IDFT add cyclic-prefix L-tap channel
b=
a −−−−−−−−−→ b −−−−−−−−−→ b b ∗h
==========⇒ b
with length P with response h
discard first P N -point DFT
b ∗ h −−−−−−−−→ (h ⊛ b) −−−−−−−−−→ H · a
b
received symbols
(b) Write and test subroutines for each step in the above equation. Q11.2
(c) Write a function that combines the subroutines from step (b) Q11.3
to create a baseband simulator for OFDM. The example .m file
header below provides defining requirements for the function.
[Output_symbol_matrix]=...
OFDM_Transceiver(Input_symbol_stream,M,N,h,P)
% --------- Input ---------
% Input_symbol_stream is a row vector of length M*N
% N is the IFFT and FFT size.
% P is the cyclic prefix length
% h is the channel impulse response in vector form
% --------- Output ---------
% Output is a M by N matrix, with the n-th row
% being the output of the n-th subcarrier.
% --------- Requirement ---------
% ’conv’ command is invoked only once.
% Calculation of the FFT of h is not allowed.
(d) Test your function with the following input: Q11.4

M=100;
N=1024;
P=10;
h=[1,2,2*1j,-1j];
Input_symbol_stream=ones(1,M*N);
Explain what you observe in the output.
(e) [Optional] Use the baseband simulator from step (d) and the
acoustic modem from Chapter 7 as a foundation from which
to construct an acoustic transmitter and receiver using orthog-
onal frequency division multiplexing. At the transmitter, the
waveform bb will serve as input to the pulse shaping filter. At
the receiver, pilot symbols will be used to estimate a complex-
valued channel gain for each time-frequency grid point; for a
channel that is static across time, a simple strategy is to place
a known pilot symbol at every location in the first column of
the grid shown in Figure 11.4. In an attempt to create an ISI
channel, rather than a narrow-band flat fading channel, select
a low-carrier frequency and short symbol interval. The result-
ing bandpass acoustic signal should haves a large fractional
bandwidth, defined as the bandwidth divided by the carrier fre-
quency.
11.3 Demonstration
Demonstrate the execution of the codes developed in step (c).
11.4 Summary
In Chapter 11, the orthogonal frequency division multiplexing scheme
was derived. ODFM is the fundamental building block in almost
all wide-band wireless systems, and the scheme effectively converts
a wide-band multi-tap channel into multiple orthogonal single-tap
sub-channels. Explorations guided students through implementation
of an OFDM transceiver.
CHAPTER 12
Adaptive Processing
In previous chapters, decoding of a received signal was accomplished

via block processing; the long stream of input samples was stored to
memory and accessible for decoding any particular symbol embedded
in the stream. Channel impairments, such as frequency and phase
offsets, were assumed static throughout the duration of the signal. In
contrast, this chapter introduces adaptive processing techniques that
sequentially process input samples and can provide time-varying es-
timates of channel impairments. In comparison to block processing,
adaptive techniques can lessen memory requirements, reduce process-
ing latency, and, importantly, track time-varying channel conditions.
In order to present representative examples of adaptive process-

ing, this chapter guides students to implement feedback loops for
carrier phase recovery. First, for the special case of a real-valued
baseband message, the Costas loop is considered. Second, decision-
directed phase recovery is presented, whereby the error between a
received value and the nearest symbol is used to track the carrier
phase. In effect, the decoding decisions are used as surrogates for
the true symbols, and the decisions direct the adaptation of the un-
known channel parameters. Remarkably, the strategy is very effective
for non-zero error rates less than 0.01.
129
130 CHAPTER 12. ADAPTIVE PROCESSING
ADC LPF
Costas
T
ejθk Loop
2ej(2π(fc +f∆ ) L k+φ) θk
(a) Costas Loop
ṽ[k] ỹ[n] Symbol

ADC LPF gtx [k] L decision
DD
ejθn Loop
j(2π(fc +f∆ ) T
L k+φ)
2e θn
(b) Decision-directed phase recovery
Figure 12.1: Feedback loops for adaptive carrier phase recovery: (a)
Costas loop; (b) decision-directed phase recovery.
12.1 Background
12.1.1 Carrier phase errors

As explored in Chapter 2 and Chapter 5, imperfections in the re-
ceiver oscillator create a phase mismatch relative to the transmit-
ter. For phase offset, φ, and frequency offset, f∆ , the time-varying
phase error in the fractionally-sampled demodulated signal, ṽ[k], is
φ + 2πf∆ TL k, modulo 2π. Here, T is the symbol interval, L is the up-
sampling factor, and k is the time index for the fractionally-sampled
signal. Thus, a small frequency offset, f∆ , causes a phase error that
grows over time; similarly, frequency drift in the demodulation ref-
erence signal results in a time-varying offset.
The processing structures considered in this chapter for address-
ing phase errors are illustrated in Figure 12.1. In Figure 12.1(a), the
Costas loop operates on the fractionally-sampled demodulated sig-
nal, ṽ[k], to track phase errors due to f∆ and φ. In Figure 12.1(b),
the decision-directed phase tracker operates on the symbol decisions,
a
b[n], to correct the phase of the symbol-rate matched filter outputs,
ỹ[n]. In both cases, feedback is employed, whereby output values
determine future estimates of the time-varying phase correction, θk
or θn .
12.1.2 Costas loop

The Costas loop [4] is a feedback mechanism for generating the cor-
rect carrier phase in a double-sideband suppressed carrier signal. The
basic principle is that for a real-valued message signal, all energy in
a QAM receiver should be in the in-phase channel, with zero signal
energy appearing in the quadrature channel. While the approach
originally employed a voltage-controlled oscillator (VCO) to gener-
ate the reference carrier signal at the receiver, here we present a
software-based, sampled data version. Let the basedband message
be a real-valued signal, m[k]. From Equation 2.12, demodulation at
the receiver with a phase offset φk = 2πf∆ TL k + φ results in the IQ
signal, ṽ[k]
ṽ[k] = m[k] cos(φk ) − jm[k] sin(φk ). (12.1)
We seek a phase correction, θk , so that zero signal energy appears

in the quadrature channel. To this end, we define a cost function,
JC (θk ), and adaptively adjust θk to minimize the cost. Specifically,
let the cost be the mean squared value of the quadrature channel
samples. Define ṽ[k] = Ik + jQk to be the demoduated samples at
the receiver; then, the mean squared value of the quadrature channel,
using phase correction θk , is given by
n h i o
JC (θk ) = E | Im ṽ[k]ejθk |2
n o
= E | Im [(Ik + jQk )(cos θk + j sin θk )] |2
n o
= E |Ik sin θk + Qk cos θk |2 . (12.2)
Above, E{·} denotes expectation, i.e., statistical average. We adopt a

gradient descent method, taking a small step size, µ, in the downhill
direction of the cost, as given by the partial derivative.

∂ ∂
JC (θ) = E |Ik sin θk + Qk cos θk |2 (12.3)
∂θ ∂θ
= E {2(Ik sin θk + Qk cos θk )(Ik cos θk − Qk sin θk )} .
(12.4)
By virtue of the averaging effect of many small steps, we omit the
expectation operator, E, to arrive at the following gradient descent
update equation
θk+1 = θk − 2µ Im{ṽ[k]ejθk } Re{ṽ[k]ejθk }. (12.5)
The same update equation can equivalently be derived by using gra-

dient ascent to maximize the average energy in the in-phase channel.
Rather than using gradient descent, an alternative update strat-
egy may be derived by postulating an error signal and using a tracking
loop, depicted in Figure 12.2. The frequency and phase offsets can
be considered a phase disturbance,
T
dk = 2πf∆ k + φ. (12.6)
L
Thus, we seek a tracking loop to provide zero steady-state error for
a disturbance that contains a ramp, {2πf∆ T /L}k, and a step, φ. To
this end, we use a Type 2 controller with two poles at z = 1 [11],
(K1 + K2 )z − K1
C(z) = . (12.7)
(z − 1)2
To specify an error signal in the tracking loop, observe from Equation

12.5
Im{ṽ[k]ejθk } Re{ṽ[k]ejθk }
n o n o
= Im m[k]e−jdk ejθk Re m[k]e−jdk ejθk
= |m[k]|2 sin(θk − dk ) cos(θk − dk )
= |m[k]|2 sin(2[θk − dk ]). (12.8)
Thus, noting that sin(2[θk − dk ]) ≈ 2(θk − dk ) for small phase errors,

we define the error signal to be
1
e[k] = Im{ṽ[k]ejθk } Re{ṽ[k]ejθk }. (12.9)
|ṽ[k]|2
Disturbance d[k]
Error
e[k]
Reference Controller Plant
C[Z] P [Z]
−
Figure 12.2: Block diagram for tracking interpretation of adaptive

carrier phase recovery.
For discrete-time implementation of the second-order loop filter, we

introduce a state variable, xk , to obtain the update equations:
θk+1 = θk + K1 ek + xk (12.10)
xk+1 = xk + K2 ek . (12.11)
The controller gains should be selected with ratio K1 /K2 approxi-
mately 50 to 100 for a suitable transient response [37, p. 253].
For small phase errors, we adopt the linearization
sin(2[θk − φk ]) ≈ 2(θk − φk ), (12.12)
yielding P (z) = −2 in Figure 12.2. Then, the closed-loop transfer
function from the disturbance, dk , to the tracking error, e[k], is found
using z-transforms:
E(z) −P (z) 2(z − 1)2
= = 2 .
D(z) 1 + P (z)C(z) z − 2(Ka + K2 + 1)z + (2K1 + 1)
(12.13)
12.1.3 Decision-directed phase tracking

Next, we consider carrier phase tracking using a decision-directed
approach, working with symbol rate complex-valued samples of the
matched filter output, ỹ[n]. Define a cost function to be the mean
squared error between the phase corrected received sample, ỹ[n]ejθn ,
and the true transmitted symbol, ã[n]. In the decision-directed ap-
proach, the symbol decisions, ab[n], are used as stand-in values for
the true data symbols. Thus, the cost is
n o
b[n] − ejθn ỹ[n]|2 .
JD (θk ) = E |a
To perform gradient descent, we need the partial derivative of this
cost with respect to the phase correction, θk .
o
∂ ∂ n jθn
∗
jθn
JD (θ) = E a[n] − e ỹ[n]
b a[n] − e ỹ[n]
b
∂θ ∂θ
n o
b[n]ỹ ∗ [n]e−jθ − a
= E j a b∗ [n]ỹ[n]ejθ
n o
b[n]ỹ ∗ [n]e−jθ
= −2 E Im a . (12.14)
Here, the superscript ∗ denotes complex conjugation. Thus, the gra-
dient descent update, with step size µ, is given by
n o
b[n]ỹ ∗ [n]e−jθ .
θk+1 = θk + 2µ Im a (12.15)
As in the previous subsection, we can use a tracking loop in place

of gradient descent. To define an error signal, we observe

Im a b[n]| |ỹ[n]|ej(αn −βn )
= Im |a
= |a
b[n]| |ỹ[n]| sin(αn − βn ), (12.16)
where αn is the phase angle of the symbol a b[n] and βn is the phase
angle of the phase-corrected sample, ỹ[n]ejθn . Thus, we have the
relation n o
Im a b[n]ỹ ∗ [n]e−jθ
sin(αn − βn ) = (12.17)
|a
b[n]| |ỹ[n]|
For small angle errors, we invoke the approximation sin α = α to
define an error signal as
n o
Im a
ek = sin(αn − βn ) = . (12.18)
|a
b[n]| |ỹ[n]|
Armed with this error signal in the decision-directed phase recovery

framework, the tracking loop update equations again take the form
of Equations 12.10–12.11; these update equations can be written as
a second-order difference equation:
θk+1 = 2θk − θk−1 + K1 ek + (K2 − K1 )ek−1 . (12.19)
12.2 Explorations
In this chapter, students implement and experimentally verify adap-
tive carrier phase recovery algorithms. A lab report should provide
answers to the questions enumerated in the page margins and present
commented code, as a subroutine, for the adaptive computations.
Q12.1 (a) Derive Equation 12.5 from Equation 12.4.
Q12.2 (b) Create code to implement decision-directed carrier phase track-

ing, using the gradient descent approach in Equation 12.15.
Apply the adaptive technique as post-processing to receiver
outputs obtained in the simulator constructed at Chapter 6.
Due to mis-estimation of the carrier frequency, the received
symbol phase will drift and eventually cause errors in a long
decoded text string. To view the action of the tracker in your
modem simulator, display the decoded text message both be-

fore and after the decision-directed tracker; also, plot both the
symbol angle error before tracking and the tracker output sig-
nal, θk , from Equation 12.15. An example snippet of code for
plotting is given below:
figure;
plot(angle(symbols./symbols_est)*180/pi,’r’);
hold on;grid;ylabel(’degrees’);
plot(angle(exp(1i*thetaGD))*180/pi);
legend(’OpenLoop Error’,’\theta_{grad}[k]’,...
’location’,’best’)
Choose a text string length greater than 100. Because differ-

ent pseudo-random noise trials will cause different frequency
estimation errors, run your simulator several times to observe
errors in the text string before and after phase tracking. Ex-
periment with several values for step size, µ.
(c) Artificially insert into your modem receiver software a step er- Q12.3
ror at some time point in the demodulation carrier phase. Ap-
ply the gradient descent carrier phase tracking as implemented
in step (b). To view the action of the tracker in your modem
simulator, plot both the symbol angle error before tracking
and the tracker output signal, θk , from Equation 12.15. Try
for several choices of step size, µ. What do you observe?
(d) [Optional] Implement decision-directed carrier phase tracking,

using the second-order tracker given in Equation 12.19. Apply
the adaptive technique as post-processing to receiver outputs
obtained in the simulator developed for Chapter 6.
For numerical experimentation, start with these suggested pa-

rameters:
sampling frequency, fs (sps) 44100

carrier frequency (Hz) 12000
frequency offset, f∆ (Hz) 15
upsampling, L 110
pilot sequence Barker 13
modulation BPSK
number of text characters > 100
passband noise variance 0.001
loop gains K1 = 0.04, K2 = 0.008
Display the decoded text message both before and after the
decision-directed tracker. Due to mis-estimation of the carrier
frequency, the received symbol phase will drift and eventually
cause errors in the decoded text string. To view the action
of the tracker in your modem simulator, plot the angle error
before tracking, and the symbol-rate tracker signals, e[n] and
θ[n]. For example,
figure;
plot(angle(symbols./symbols_est)*180/pi,’r’);
hold on;grid;ylabel(’degrees’);
plot(angle(exp(1i*theta))*180/pi);
plot(PhError*180/pi,’g--’);
legend(’OpenLoop Error’,’\theta[k]’,...
’e[k]’,’location’,’best’)
Because different pseudo-random noise trials will cause dif-

ferent frequency estimation errors, run your simulator several
times to observe errors in the text string before and after phase
tracking. Is the adaptive procedure able to track the correct
phase for long data packets, despite small errors in frequency
offset estimation? Experiment with several values for filter
gains, K1 and K2 .
As alternative to tracking phase drift due to frequency estima-
tion errors, the tracking may be tested by using zero noise at
the receiver and artificially inserting into your modem receiver
software a carrier frequency offset.
(e) [Optional] From step (d), experiment with different values of

K1 and K2 . Choose gains that cause the tracking signal to
diverge. What are the poles of the resulting transfer function

in Equation 12.13?
(f) [Optional] Modify your decision-directed phase tracker in step

(d) to operate for QPSK modulation.
(g) [Optional] Artificially insert into your modem receiver soft-

ware a step error at some time point in the demodulation car-
rier phase. Apply the decision-directed carrier phase tracking
with second-order feedback loop, as implemented in step (d).
Make a plot of tracking error versus symbol index, n. Try for
several choices of controller gains, K1 and K2 . What do you
observe?
(h) [Optional] Derive the Costas loop, Equation 12.5, by using

gradient ascent to maximize the average energy in the in-phase
channel.
(i) [Optional] Use the z-transform to derive the controller trans-

fer function, C(z), from Figure 12.2 and the update equations
given in Equations 12.10 - 12.11.
(j) [Optional] Using z-transforms, derive Equation 12.19 from

Equations 12.10 - 12.11.
(k) [Optional] Derive Equation 12.13, the closed-loop transfer func-

tion from the disturbance, dk , to the tracking error, e[k]. How
would you determine stability for a given pair of controller
gains, K1 and K2 ? What is the DC gain? For your choices
of K1 and K2 , find and plot the step and ramp responses.
12.3 Demonstration
In Chapters 6 and 7, a small error in frequency estimation at the
receiver could lead to an accumulated phase error that destroyed
decoding performance in a long data packet. Demonstrate any one
of your adaptive carrier phase recovery algorithms as a means for
correctly decoding a long data packet in the presence of a small
frequency offset or frequency drift.
12.4 Summary
In Chapter 12, adaptive processing has been introduced as a closed-
loop procedure for tracking carrier phase at the receiver. For real-
valued baseband messages, the Costas loop was derived, using both
gradient descent and a second-order tracking controller for update
equations. For general symbol constellations, a decision-directed
phase tracking procedure was introduced, again using both gradient
descent and a second-order controller to derive update equations.
The adaptive and decision-directed approaches introduced here
are representative of a wealth of procedures developed over the past
four decades (see, e.g., [27, 10, 22, 18]). For example, decision-
directed techniques have been developed for the channel equalization
challenge introduced in Chapter 8.
APPENDIX A
Software Design Suggestions
Students and instructors alike are encouraged to embrace the devel-

opment of engineering skills, and software design skills in particu-
lar, that are part of a laboratory experience. To this end, we high-
light several suggestions as students progress towards implementing
a working acoustic modem.
• Work from a structured outline; follow the graphical roadmaps

provided by Figures 1.1, 4.1, and 4.3. Give a descriptive cap-
tion to each section of code, and segment via the %% section
indicator.
• Use descriptive variable names. For debugging and readability,

generally avoid re-use of variable names during code develop-
ment. Memory efficiency can be sacrificed for clarity during
development.
• Set modem parameter values in a stand-alone section at the

top of your code; or, load parameters from a separate .mat
file, as discussed in Section 7.2. Do not hardcode nor reset
parameters throughout a script. A good rule of thumb is to
ensure that no number appears in your code, except possibly
for the parameter assignment commands at the very beginning
of a script.
• Clear all variables and figures at the beginning of the script

(clear all;close all) to avoid debugging difficulties that
may arise from re-use of variable values inherited from previous
executions of the script.
139
140 APPENDIX A. SOFTWARE DESIGN SUGGESTIONS
• At the conclusion of each step along the roadmap, generate a

new graph to visualize the result of the step. The visualization
can be an indispensible aid in debugging and understanding.
The command figure initiates a new figure window.
• Develop, test, and debug steps along the roadmap one at a

time. Steps may be re-used via cut-and-paste of in-line code,
or may be re-used as function calls. Efficient progress through
the lessons presented in Chapters 1 through 7 will depend on
re-use of debugged code as students proceed through staged
development of an acoustic modem. Unused sections can be
bypassed by use of comment symbol, %. A code section may
be created using the double percent marker, %%, and a section
may be individually executed using the menu bar selection,
“Run section,” at the top right (or via Ctrl + Enter). Addi-
tionally, breakpoints can provide a useful debugging tool.
• Use complex-valued arithmetic to conveniently implement quad-

rature modulation and to maintain the I and Q samples linked
in a simple format.
APPENDIX B
RF Experiments
The explorations in Chapters 1 through 7 culminate in students im-

plementing a half-duplex modem operating at an acoustic carrier fre-
quency. The same baseband processing can be used to implement a
digital communication system operating at a radio frequency carrier.
This appendix provides suggestions and links to additional informa-
tion for students and instructors interested in implementing a radio
frequency modem. Although radio frequency propagation has been
understood since the 1880s, wireless broadcast can be an ethereal
experience, especially when acomplished with the most rudimentary
of materials.
B.1 Low-cost DIY Modulation

For do-it-yourself (DIY) construction of a transmitter and receiver
from simple materials, both AM and FM approaches provide solu-
tions accessible to students in middle school [30] and beyond.
B.1.1 AM
Mathematically, an envelope detector for large-carrier AM reception
is described by
π
v(t) = LPF{ |r(t)| } − A ≈ m(t). (B.1)
2
The envelope detector is a noncoherent receiver in that it does not use
knowledge of the phase of the carrier signal. The gain π2 compensates
141
142 APPENDIX B. RF EXPERIMENTS
red black C3
L C1 R1 C2 R2
Figure B.1: Circuit model for an envelope detector.
for the loss incurred when lowpass filtering the rectified signal,
1
R 4fc
0 cos(2πfc t)dt 2
1 = . (B.2)
4fc
π
A circuit model of Equation B.1 using passive devices is shown

in Figure B.1. The absolute value, | · |, is easily approximated using
a diode (or, for example, piece of carbon with a blued razor blade).
An RC filter provides the lowpass filtering; and, an additional capac-
itor provides the DC block to achieve the level shift of −A. Further,
a resonant LC circuit provides tuning of the antenna to the car-
rier frequency, fc . The demodulated signal is the voltage m(t); a
piezo-electric crystal in parallel with R3 can provide audible out-
put. By converting approximately one trillionth Watt of power from
the transmitted waveform into acoustic energy, the circuit is able to
produce sound pressures perceptible by the human ear. Many DIY
guides are available, e.g. [9, 40, 6, 26].
QA.1 As an exercise, the reader is invited to derive the resonant fre-
quency of a parallel LC circuit. Specifically, find the transfer func-
tion, H(jω), from x(t) to y(t) for the circuit shown in Figure B.2.
(In the figure, the resistance R approximates an antenna given by a
long insulated wire.)
An AM transmitter can be constructed from an audio trans-
former and crystal oscillator [9]. The baseband BPSK signal used in
the acoustic modem may serve as single-channel audio input; more
generally, a complex-valued QPSK or QAM baseband signal mod-
ulated to an acoustic carrier frequency, as in Chapter 7, may serve
as a real-valued intermediate frequency signal prior to RF amplitude
modulation.
B.1. LOW-COST DIY MODULATION 143
x L C y
Figure B.2: RLC circuit model.
B1
9V
C1 R2 C5
1-33μF 27k L1 10pF
Audio IN Rf OUT
(mono) to antenna
R1 C2 C3 C4 (optional)
10k 0.01μF 0.01μF 10pF
Q1
2N3904
R3
470
Figure B.3: Circuit schematic for a single-transistor FM transmitter.
B.1.2 FM
The basic regenerative circuit for FM modulation and demodula-
tion was introduced by E. H. Armstrong in 1915 [1, 19]. A super-
regenerative FM transmitter can be implemented using a single tran-
sistor design [28, 20]. The circuit schematic is given in Figure B.3
[28]; note that capacitor C4 is a variable capacitor for tuning. Many
other designs can be found from DIY sources [34].
Similarly, a single-transistor design may be used to implement a
super-regenerative FM receiver. A circuit schematic [2, 7, 5] is given
in Figure B.4; note that IC1 is an LM386 audio amplifier. Many
alternatives, e.g. [23], can be found online.
Rather than build an FM modulator and demodulator, students
L2 10k
10k 0.1μF
MPF102 G 0.01μF
G
0.047μF
6 inches
of 18 D
guage + 1
wire 10μF IC1
15pF
1k
+9V
+
+
+ 0.047μF 220μF
15-120pF
L1 100μF 0.047μF
0.001μF
Figure B.4: Circuit schematic for single-transistor FM receiver.
can construct a modem using commercially available integrated cir-

cuits. The example shown in Figure B.5 employs ES series modules
from Linx Technologies operating at 916 MHz.
In addition, monaural audio input and output can be easily taken
from a cordless phone and base to provide a low-cost FM wireless
solution. Retailers sell phones for under US$15, and used phones
can be found at thrift stores and consignment shops. In 1994, digital
cordless phones were introduced operating in the 900 MHz frequency
range. Telephones sold in the US use the 900 MHz, 1.9 GHz, 2.4 GHz,
or 5.8 GHz bands; phones operating at 900 MHz are typically inex-
pensive analog models.
Nonlinearities in low-cost FM transmitters and receivers are poorly
Figure B.5: FM transmitter (left) and receiver (right) constructed

from low-cost RF modules.
B.2. COMMERICAL SDR SOLUTIONS 145
suited for use with general QAM constellations; but, PSK digital
modulation can perform well.
B.2 Commerical SDR Solutions

Several software-define radios are commercially available. Costs range
from several hundred to several thousand US dollars per unit.
The PicoZed SDR Z7035/AD9361 System-on-Module (SOM)
is a credit-card size form-factor PicoZed SDR solution that provides
frequency-agile wideband 2 × 2 receive and transmit paths from RF
to baseband. The unit was introduced in 2015 by Avnet in conjunc-
tion with Xilinx, Analog Devices, and Mathworks. A bandwidth to
56 MHz can be obtained with a tunable carrier from 70 to 6000 MHz.
The Wireless Open-Access Research Platform (WARP) provides

a scalable and programmable hardware for software-agile physical
layer prototyping. The WARP v3 shown in Figure B.6 integrates
a high performance FPGA, two flexible RF interfaces, and multiple
peripherals to facilitate rapid prototyping of custom wireless designs
[25]. Reference designs and support materials are available online
[39]. In addition, teaching materials for use in a ten week labora-
tory course are available, including slides and lecture videos [8]. The
course introduces basics of implementing digital communication sys-
tems on real-time hardware and uses the Xilinx System Generator
environment.
For radio operation at 2 to 50 MHz, GenesisRadio sells trans-
ceiver kits and associated SDR software for under US$400.
Ambitious students may wish to contruct their own software-agile
transceiver, for example using the Maxim 2830 evaluation kit. The
MAX2830 is a direct conversion, zero-IF, RF transceiver designed
specifically for the 2.4GHz to 2.5GHz ISM band. The chip integrates
all circuitry required to implement the RF transceiver function, pro-
viding an RF power amplifier (PA), an Rx/Tx and antenna diversity
switch, RF-to-baseband receive path, baseband-to-RF transmit path,
voltage-controlled oscillator (VCO), frequency synthesizer, crystal
oscillator, and baseband/control interface.
Digital video broadcast USB dongles based on the Realtek RTL
2832U can be used as an inexpensive SDR receiver, since the chip
allows transferring the raw I/Q samples to a host computer. The
Figure B.6: WARP provides a scalable and programmable hard-

ware platform for prototyping software-agile communication systems.
(Photograph used with permission.)
RTL2832U outputs 8-bit I/Q-samples at approximately 2.5 MS/s.

The frequency range depends on the tuner employed [29]. The GNU
Radio collection of tools can be used to build custom radio devices
[14].
APPENDIX C
Functions
This appendix provides code for eight custom functions referenced in

the text and listed in Table C.1. Scripts and functions are available
for download at OpenStax CNX1 and Mathworks Matlab Course-
ware2 .
Table C.1: Custom functions.
Function Description
plottf.m Make time and frequency plots
firlpf.m Design low-pass filter design
char2psk.m Template for PSK digital modulation
psk2char.m Template for PSK digital demodulation
srrc.m Design square-root raised cosine filter
eyediagram.m Make eye diagram
makepilots.m Make pilot sequence
packetdetect.m Detect data packet
1
http://cnx.org
2
http://www.mathworks.com/academia/courseware
147
148 APPENDIX C. FUNCTIONS
plottf
%PLOTTF Plot sampled signal in time and frequency domains

% PLOTTF(x,Ts) plots the time-domain samples in vector x, assuming
% that Ts is the sampling interval in seconds, and also plots the
% Riemann-sum approximation of the Fourier transform between the
% frequencies of -1/(2Ts) and 1/(2Ts) Hertz.
%
% PLOTTF(x,Ts,’t’) plots only the time-domain signal.
%
% PLOTTF(x,Ts,’f’) plots only the frequency-domain signal.
%
% In all cases, PLOTTF returns handles to the graphical objects.
% P. Schniter. Used with permission.
function hh = plottf(x,Ts,str)
plot_type = 0;
if nargin==3,
if str==’f’,
plot_type = 1;
elseif str==’t’,
plot_type = 2;
end;
end;
N=length(x); % discrete signal length

t=Ts*(0:N-1); % time vector
if 2*floor(N/2)==N,
f=(-N/2:N/2-1)/(Ts*N); % frequency vector
else
f=(-(N-1)/2:(N-1)/2)/(Ts*N); % frequency vector
end;
X=Ts*fft(x); % do DFT
Xs=fftshift(X); % % shift it for plotting
if plot_type==1,
hh = plot(f,abs(Xs)); % plot magnitude spectrum
xlabel(’frequency [Hz]’);
ylabel(’magnitude’);% label the axes
elseif plot_type==2,
if isreal(x),
hh = plot(t,x);% plot the real waveform
xlabel(’time [sec]’);
ylabel(’amplitude’); % label the axes
else
hh = plot3(t,real(x),imag(x)); % plot the complex waveform
149
ylabel(’real’); zlabel(’imag’); % label the axes
end;
else
subplot(2,1,1);
if isreal(x),
hh = plot(t,x); % plot the real waveform
ylabel(’amplitude’); % label the axes
else
hh = plot3(t,real(x),imag(x)); % plot the complex waveform
ylabel(’real’); zlabel(’imag’); % label the axes
end;
subplot(2,1,2);
hh = [hh,plot(f,abs(Xs))]; % plot magnitude spectrum
xlabel(’frequency [Hz]’);
ylabel(’magnitude’) % label the axes
subplot(2,1,1);
end;
firlpf
function [h]=firlpf(Lh,fpass,fstop,fsample)
%FIRLPF Least-squares design of linear-phase low-pass filter.
%
% [h] = firlpf(Lh,fpass,fstop,fsample) designs Type I, length Lh,
% linear-phase low-pass filter via least-squares design.
% Lh must be odd.
%
% Lh filter length (odd)
% fpass edge of passband (Hz)
% fstop edge of stopband (Hz)
% fsample sampling rate (Hz)
% h designed impulse response
% Digital Communication Laboratory

% Autumn 2014
%% argument check
if(nargin < 4)
error(’Too few input arguments for FIRLPF.’)
end
if(fpass >= fstop)
error(’Passband edge must be less than stopband edge’)
end
if(fsample <= 2*fstop)
error(’Sampling rate is sub-Nyquist’)
end
if( mod(Lh,2) ~= 1);
Lh=Lh+1;
warning(’Lh incremented for Type I FIR filter.’);
end
%% normalized frequency (radians per sample)
fpass = fpass / (fsample/2) * pi;
fstop = fstop / (fsample/2) * pi;
%% apply orthogonality principle for LS solution
indx=(1:(Lh-1)/2);
RHS = [fpass sin(fpass*indx)./indx].’;
% diagonal of matrix
tmp=zeros(1,length(indx)+1);
tmp(1) = pi - (fstop - fpass);
tmp(2:end)=tmp(1)/2+(sin(2*indx*fpass)-sin(2*indx*fstop))./(4*indx);
LHS = diag(tmp);
% off-diagonals
for row=0:(Lh-1)/2;
for col=row+1:(Lh-1)/2;
LHS(row+1,col+1) = ...
(sin((col+row)*fpass)-sin((col+row)*fstop))/2/(col+row)+...
151
(sin((col-row)*fpass)-sin((col-row)*fstop))/2/(col-row);
LHS(col+1,row+1) = LHS(row+1,col+1);%symmetry
end
end
g = LHS\RHS;g=g.’;%solve linear system
h = [g(end:-1:2)/2 g(1) g(2:1:end)/2];%linear phase impulse response
%end of function
char2psk template
%% template for char2psk.m
function [a,bits] = char2psk(str,M);
% char2psk.m
% Function to map a charater string to PSK symbols
% (See also psk2char.m)
%
% Inputs:
% str string variable with text
% M size of PSK constellation (2 or 4)
% or, user can select ’bpsk’ or ’qpsk’
% Outputs:
% a output list of PSK symbols (complex-valued)
% bits output list of bits (logical)
% Digital Communications Laboratory

% Autumn 2014
%% error checks
if(nargin ~= 2)
error(’Error: char2psk.m requires two input arguments.’)
end
%% text to symbols
% first, string to decimal to binary
str_binary = dec2bin(double(str),8);
% convert array row-by-row to one long string
bits = reshape(str_binary.’,1,8*length(str));
% binary to symbols
switch M
case {2,’bpsk’,’BPSK’}
M = 2;
a = (2*bin2dec(bits.’)-1).’;
case {4,’qpsk’,’QPSK’}
M = 4;
%
%create your QPSK code here
%
end
bits=(bits==’1’);% convert char to logical
% end of function
153
psk2char template
%% template for psk2char.m
function [str,bits] = psk2char(y,M)
% psk2char.m
% Function to decode M-PSK symbols to a string of ASCII text.
% (See also char2psk.m)
%
% Inputs:
% M size of PSK constellation (2 or 4; or ’bpsk or ’qpsk’)
% y list of PSK IQ points (complex numbers)
% Outputs:
% strbin string variable with ASCII text
% bits list of 0/1 bits (as logical variables)
%
% Autumn 2014
%% error checks
if(nargin ~= 2)
error(’Error: psk2char.m requires two input arguments’)
end
if (isnumeric(y) ~= 1 || isempty(y))
error(’Error: the input y must be numeric.’)
end
%% symbol slicing
% first, symbol string to string of integer labels: 0,1,...,log2(M)
% labels are integers; use double data type
switch M
case {2,’bpsk’,’BPSK’}
M = 2;
labels = (real(y) >= 0);
case {4,’qpsk’,’QPSK’}
M = 4;
%
%create code here; see Figure 3.1(b) for labeling quadrants
%
end
% second, labels into string of bits
N = length(labels)*log2(M)/8;%number of characters
tmp = dec2bin(labels,log2(M));%label to binary
strbin = reshape(tmp.’,8,N).’;%array, labels to 8 bits/character
% third, convert from char to logical:
bits = logical(bin2dec(reshape(tmp.’,1,N*8).’).’);
% fourth, convert binary to decimal ASCII to character
str=char(bin2dec(strbin)).’;%
% end of function
srrc
function g = srrc(D,alpha,L)
% SRRC Fractionally-sampled square-root raised cosine pulse.
%
% [g] = srrc(D,alpha,L) creates a fractionally-sampled square-root
% raised cosine pulse
%
% D one-half the length of srrc pulse in symbol durations
% alpha excess bandwith (valude between 0 and 1);
% alpha=0 gives a sinc pulse
% L samples per symbol (oversampling factor);
% L must be a positive integer
% g samples of the srrc pulse

% Autumn 2014
%% argument check
if(nargin < 3)
error(’Too few input arguments for SRRC’)
end
if( min( [D,alpha,L] ) < 0 )
error(’Inputs must be non-negative for SRRC.’)
end
if( round(L) ~= L )
error(’Input L must be a non-negative integer for SRRC.’)
end
%% compute samples
k = -D:(1/L):D; % k is t/T
g = (sin(pi*(1-alpha)*k)+(4*alpha*k).*cos(pi*(1+alpha)*k)) ./ ...
((pi*k) .* (1-(4*alpha*k).^2))/sqrt(L);
%% fill in for denominator zeros
g(k==0) = (1 + (4/pi-1)*alpha)/sqrt(L);
g(abs(abs(4*alpha*k)-1) < sqrt(eps)) = ...
alpha/sqrt(2*L)*( (1+2/pi)*sin(pi/4/alpha) + ...
(1-2/pi)*cos(pi/4/alpha));
%end of function
155
eyediagram
function [] = eyediagram(y_up,L,N,IQ)
% eyediagram.m
% Function to plot an eye diagram
%
% y_up upsampled matched filter outputs
% L downsampling factor (integer)
% N number of symbols
% IQ string, ’complex’, if both I and Q are desired
%
% Note that first sample of y_up is assumed to occur at a symbol time.

% Autumn 2014
%% error checks
if(nargin < 2)
error(’Error: eyediagram.m requires two input arguments’)
end
if(nargin == 3), IQ=’real’; end
if(strcmp(IQ,’complex’))
IQ=1;%plot both I and Q
else
IQ=0;
end
%N-2 segments for N symbols
%% extract symbol period length segments

% start of first full interval
start = 1 +ceil(L/2);
Y_up = reshape(y_up(start:start+(N-2)*L-1),L,N-2); %extract segments
%% plots
if(IQ == 0)
% plot eye diagram, I channel only
figure;
plot(-floor(L/2)/L+(0:L-1)/L,real(Y_up));% superimpose them
title(’Eye Diagram’,’fontsize’,13);
xlabel(’relative symbol index’,’fontsize’,13);
else
% plot eye diagram, I & Q channels
figure;
subplot(211);plot(-floor(L/2)/L+(0:L-1)/L,real(Y_up));
title(’Eye Diagram’,’fontsize’,13);
subplot(212);plot(-floor(L/2)/L+(0:L-1)/L,imag(Y_up));
xlabel(’relative symbol index’,’fontsize’,13);
end
makepilots
function [pilots] = makepilots(opt)
%MAKEPILOTS generate pilot sequence
%
% [pilots] = makepilots(opt) generates a sequence of pilot symbols
%
% opt indicator to select pilot sequence:
% Barker: ’barker13’, ’B’, ’b’, ’Barker’, or ’barker’ (default)
% PN: ’pn51’, ’PN’, ’pn’, or ’pseudorandom’
% split Barker: ’splitbarker’ or ’splitBarker’
%
% The length-13 Barker code and length-51 pseudo-random code
% generated here are employed in the acoustic transmitter mobile
% app available at iTunes Apps or Google Play.

% Autumn 2014
if(nargin == 0), opt=’barker13’; end % default

switch opt
case {’barker13’, ’B’, ’b’, ’Barker’, ’barker’}
pilots=[+1 +1 +1 +1 +1 -1 -1 +1 +1 -1 +1 -1 +1];
case {’pn51’, ’PN’, ’pn’, ’pseudorandom’}
pilots = [+1,+1,-1,+1,+1,-1,-1,+1,+1,+1, ...
-1,+1,+1,-1,+1,-1,-1,+1,+1,+1, ...
+1,-1,+1,+1,+1,+1,+1,-1,+1,-1,...
+1,-1,-1,-1,-1,+1,+1,-1,+1,-1, ...
-1,-1,+1,+1,-1,-1,-1,+1,+1,+1,-1];
case {’splitbarker’, ’splitBarker’}
pilots=[+1 +1 +1 +1 +1 -1 -1 +1 +1 -1 +1 -1 +1 ...
+1 +1 +1 +1 +1 -1 -1 +1 +1 -1 +1 -1 +1];
otherwise
warning(’Unexpected pilot type. Default barker13.’)
pilots=[+1 +1 +1 +1 +1 -1 -1 +1 +1 -1 +1 -1 +1];
end
157
packetdetect
function [ rx_Trim ] = packetdetect( yup , L, pkt_len)
%PACKETDETECT energy detector to approximately locate a data packet;
% trims a received upsampled signal
%
% [rx_Trim] = packetdetect(yup,L,pkt_len)
%
% yup fractionally sampled matched filter output
% L upsampling factor
% pkt_len packet length (including both header and payload)

% Autumn 2014
Mask=ones(1,pkt_len*L);
Energy_Corr=conv(abs(yup).^2,Mask);
Energy_Corr=Energy_Corr(length(Mask):end);
[~,Max_Energy_Index]=max(Energy_Corr);
End_Packet_Index=min(length(yup),Max_Energy_Index+(pkt_len+3)*L);
rx_Trim=yup(Max_Energy_Index-3*L-round(max(10,L)*rand(1,1)) : ...
End_Packet_Index);
%for educational use:
% randomized offset to test frame timing;
% presumes at least 4L noise samples precede packet
end
Glossary
Notation
The indexing conventions f (t), f [k], and f [n] are adopted to suggest
a continuous-time signal representation, a fractionally-sampled base-
band signal representation, and a symbol-rate sampled signal repre-
sentation, respectively. A tilde, as in m̃, denotes a complex-valued
quantity carrying both in-phase and quadrature components.
Symbol Quantity
fc carrier frequency
fs sampling frequency
Ts sampling period
T symbol period
M constellation size
gtx (t), gtx [k] pulse shaping filter
Gtx (f ) Fourier transform of gtx (t)
D pulse half-width, in symbol intervals
L upsampling factor
α excess bandwidth; a pulse shaping parameter
W one-sided baseband bandwidth (Hertz)
p(t) convolution of pulse shape and matched filter
P (f ) Fourier transform of p(t)
Ntr length of pilot, or training, symbol sequence
159
160 GLOSSARY
Symbol Quantity
ã[n] symbol sequence, including pilots

a↑ [k] upsampled symbol sequence
a
b[n] decoded symbol sequence
m̃[k], m̃(t) baseband message at transmitter
mI (t) in-phase (real) component of message m(t)
mQ (t) quadrature (imaginary) component of m(t)
s(t), s[k] bandpass transmission
f∆ frequency offset
φ phase offset
N0 noise power spectral density, Watts per Hertz
Eb energy per bit
Es energy per symbol
K order of multi-path channel model
δ[k] Kronecker delta sequence
δ(t) Dirac delta function √
j imaginary unit, j = −1
Re{·} real part of
Im{·} imaginary part of
r(t), r[k] bandpass signal at receiver
ṽ(t), ṽ[k] baseband message at receiver
vI (t) in-phase (real) component of ṽ(t)
vQ (t) quadrature (imaginary) component of ṽ(t)
Bp pass-band edge frequency (Hertz)
Bs stop-band edge frequency (Hertz)
ỹ↑ [k] fractionally sampled matched filter output
ỹ[n] symbol-rate matched filter outputs at receiver
161
Acronyms
ADC analog-to-digital converter
AM amplitude modulation
ASCII American standard code for information interchange
AWGN additive white Gaussian noise
BER bit error rate
BPSK binary phase-shift keying
CDMA code-division multiple access
CMA constant modulus algorithm
CP cyclic prefix
DAC digital-to-analog converter
DDC digital downconverter
DFT, IDFT discrete Fourier transform, inverse DFT
DDS direct digital synthesis
DPSK differential phase-shift keying
DSP digital signal processor
DUC digital upconverter
FFT, IFFT fast Fourier transform, inverse FFT
FIR finite impulse response
I in-phase
I/O input/output
IF intermediate frequency
ISI inter-symbol interference
PA power amplifier
PSK phase-shift keying
LNA low-noise amplifier
LPF low-pass filter
NCO numerically controlled oscillator
OFDM orthogonal frequency division multiplexing
PHY physical layer
Q quadrature
QAM quadrature amplitude modulation
QPSK quadrature phase-shift keying
RF radio frequency
SDR software-defined radio
SNR signal to noise ratio
sps samples per second
SRRC square-root raised cosine pulse
UHF ultra-high frequency (300 – 1,000 MHz)
VCO voltage-controlled oscillator
162 GLOSSARY
Bibliography
[1] Edwin H. Armstrong. Some recent developments in the audion

receiver. Proc. Inst. Radio Engineers (IRE), 3:215–247, 1915.
143
[2] Patrick Cambre. A one transistor super-regenerative FM

receiver. http://web.archive.org/web/20090121123846/
http://braincambre500.freeservers.com/rss(1)(1)(1).
htm. Archived 2009-01-21; accessed 2015-06-09. 143
[3] R. W. Chang. Synthesis of band-limited orthogonal signals for

multi-channel data transmission. Bell System Technical Journal,
45(10):1775–1796, 1966. 113
[4] J.P. Costas. Synchronous communications. Proceedings of the

IRE, 44(12):1713–1718, Dec 1956. 131
[5] Martijn de Milliano. Simple FM radio. www.millibyte.nl/

index.php?page=fm-radio, 2002. Accessed: 2015-06-09. 143
[6] Steven Dufresne. Crystal radios. www.rimstar.org/

equip/crystal_radios.htm and www.youtube.com/watch?
v=VqdcU9ULAlA. Accessed: 2015-06-09. 142
[7] electronicsNmore. One transistor FM receiver. www.youtube.

com/watch?v=Wdi0X0YQSxM, 2013. Accessed: 2015-06-09. 143
[8] Evan Everett and Michael Wu. Eec 433 architeture for
wireless communications. www.warpproject.org/trac/wiki/
Rice_ELEC_433. Accessed: 2015-06-12). 145
[9] Simon Quellen Field. A simple AM transmitter. http://

sci-toys.com/scitoys/scitoys/radio/am_transmitter.
html. Accessed: 2015-06-09. 142
163
164 BIBLIOGRAPHY
[10] Michael Fitz. Fundamentals of Communications Systems.

McGraw-Hill, New York, NY, 2007. viii, 138
[11] Gene F. Franklin, David J. Powell, and Abbas Emami-Naeini.
Feedback Control of Dynamic Systems. Prentice Hall PTR, Up-
per Saddle River, NJ, USA, 4th edition, 2001. 132
[12] L. E. Frenzel. The elusive software-defined ra-
dio. http://electronicdesign.com/communications/
elusive-software-defined-radio, 2006. 2
[13] Robert G. Gallager. Principles of Digital Communication. Cam-
bridge University Press, New York, NY, 2008. viii
[14] GNUradio. GNURadio: The Free & Open Software
Radio Ecosystem. www.gnuradio.org/redmine/projects/
gnuradio/wiki. Accessed: 2015-06-09. 146
[15] Robert W. Heath, Jr. Digital Communications: Physical Layer
Exploration Lab using the NI USRP. NTS Press, Allendale, NJ,
2012. ix, 97
[16] M. Heideman, D.H. Johnson, and C.S. Burrus. Gauss and the
history of the fast Fourier transform. IEEE ASSP Magazine, 1
(4):14–21, October 1984. 116
[17] R. Johnson, Jr., P. Schniter, T.J. Endres, J.D. Behm, D.R.
Brown, and R.A. Casas. Blind equalization using the constant
modulus criterion: a review. Proceedings of the IEEE, 86(10):
1927–1950, Oct 1998. 64
[18] C. Richard Johnson Jr. and William A. Sethares. Telecommu-
nications Breakdown: Concepts of Communications Transmited
via Software-Defined Radio. Prentice Hall, Upper Saddle River,
NJ, 2004. viii, ix, 55, 100, 138
[19] Charles Kitchin. High peformance regenerative receiver
design. QEX, pages 24–36, 1998. Available on-
line www.arrl.org/files/file/Technology/tis/info/pdf/
9811qex026.pdf. 143
[20] Tetsuo Kogawa. Making the simplest FM transmitter. www.
translocal.jp/radio/micro/simplestTX01.pdf, 2007. Ac-
cessed: 2015-06-09. 143
BIBLIOGRAPHY 165
[21] C. Emre Koksal. ECE 5000, Introduction to Analog and Digi-

tal Communication. unpublished lecture notes, Columbus, OH,
2015. 110, 111
[22] Upamanyu Madhow. Fundamentals of Communications Sys-
tems. Cambridge University Press, Cambridge, UK, 2008. viii,
138
[23] Andrew R. Mitz. Build a one transistor FM radio. www.
somerset.net/arm/fm_only_one_transistor_radio.html,
2013. Accessed: 2015-06-09. 143
[24] P. Moose. A technique for orthogonal frequency division multi-
plexing frequency offset correction. IEEE Trans. Commun., 42
(10):2908–2914, October 1994. 97
[25] Patrick Murphy. WARP v3 Kit. Mango Communications
http://mangocomm.com/products/kits/warp-v3-kit. Ac-
cessed: 2015-07-13). 145
[26] Bre Pettis. How to make a foxhole radio. www.youtube.com/
watch?v=skKmwT0EccE, 2007. Accessed: 2015-06-09. 142
[27] John Proakis. Digital Communications. McGraw-Hill, New
York, NY, 2001. viii, 38, 90, 104, 138
[28] Sean Michael Ragan. Make-zine: Super-simple iPod
FM transmitter. http://makezine.com/projects/
super-simple-fm-transmitter, 2013. Accessed: 2015-06-09.
143
[29] rtl-sdr. rtl-sdr. sdr.osmocom.org/trac/wiki/rtl-sdr. Ac-
cessed: 2015-06-09. 146
[30] H. C. Sandifer. Telegraph plant by Amateur Work readers. Am-
ateur Work, Illustrated, page 223, June 1904. Available online,
www.earlyradiohistory.us/1904ama.htm. 141
[31] Philip Schniter. Introduction to Analog and Digital Communi-
cations [Kindle edition]. Amazon Digital Services, Inc., Seattle,
WA, 2011. ASIN B005JU4GTM. ii, viii, ix, 100
[32] SDR Forum. SDRF cognitive radio definitions,. http://
www.sdrforum.org/pages/documentLibrary/documents/
SDRF-06-R-0011-V1_0_0.pdf, 2007. Accessed: 2007-11-15. 2
166 BIBLIOGRAPHY
[33] Claude Shannon. A mathematical theory of communication.

Bell System Technical Journal, 27(3):379–423, 1948. 103
[34] Art Swan. Art Swan’s electronic circuits. www.angelfire.com/

art2/artswan. Accessed: 2015-06-09. 143
[35] O.Y. Takeshita. On maximum contention-free interleavers and

permutation polynomials over integer rings. IEEE Transactions
on Information Theory, 52(3):1249–1253, March 2006. 108
[36] J. Treichler and M. Larimore. New processing techniques based

on the constant modulus adaptive algorithm. IEEE Transac-
tions on Acoustics, Speech and Signal Processing, 33(2):420–431,
Apr 1985. 66
[37] Steven A. Tretter. Communication System Design Using

DSP with Laboratory Experiments for the TMS320C6713 DSK.
Springer, New York, NY, 2008. ix, 133
[38] S. Verdu. Guest editorial. IEEE Transactions on Information

Theory, 44(6):2042–2044, Oct 1998. 104
[39] WARP. Warp project. www.warpproject.org. Accessed: 2015-

06-09. 145
[40] Charles Wenzel. Crystal radio circuits. www.techlib.com/

electronics/crystal.html. Accessed: 2015-06-09. 142
Index
adaptive processing, 63, 129 clear all, 139

addpath, 92 close, 139
aliasing, 6 CMA, 64
AM transmitter, 142 code rate, 106
amplitude demodulation, 9 coding gain, 111
amplitude modulation, 7 colon, :, 33
angle, 34 complex-baseband signal, 25, 37
app, 87 conj, 93
Armstrong, E. H., 143 constant modulus algorithm, 64
audioplayer, 14 constellation, 38
audiorecorder, 13 conv, 16
cordless phone, 144
bandpass sampling, 7 Costas loop, 131
Barker code, 68 cyclic prefix, 117
baud rate, 38, 59
bit, 37
bit error rate, 43 dec2bin, 45
bit rate, 38 decimation, 11
block code, 105 decision directed, 133
block processing, 63, 129 decision region, 42
BPSK, 38 decision-directed tracking, 129
DFT, 116
carrier frequency, 7 differential modulation, 127
CDMA, 115 differential PSK, 90
channel capacity, 104 digital downconverter, 32
channel coding, 105 digital modulator, 37
channel impairments, 63 digital upconverter, 31
channel model, 28 digital-to-analog conversion, 10
channel sounding, 81 direct digital synthesizer, 32
char2psk.m, 45, 152 double, 45
circular convolution, 116 double frequency term, 25
167
168 INDEX
downconversion, see amplitude de- IDFT, 116

modulation image, 6
DPSK, see differential PSK in-phase signal, 24
inter-symbol interference, 53
envelope detector, 141 interleave, 108
Euler identity, 7 interpolation, 10
excess bandwidth parameter, 56 ISI, 53
expectation, 131 ISI channel, 95, 113
eyediagram.m, 57, 60, 155
Kronecker delta sequence, 54
feedback, 130
FFT, 116, 122 linear phase filter, 29
figure, 15 load, 92
filter, 16 low-pass filter, 9
filter delay, 57 lower sideband, 8
FIR, 56 LPF, see low-pass filter
firlpf.m, 16, 150
flat fading channel, 67, 81 makepilots.m, 87, 156
FM, 143 marker sequence, see pilot
Fourier spectrum, 6 max, 71
fractional bandwidth, 84, 128 Maxim 2830, 145
fractional sampling, 77 memoryless modulation, 37
frame timing, 66, 67 min, 66
frequency drift, 130 mixing, see amplitude modulation
frequency offset, 10, 29, 78 mobile app, 87
frequency recovery, 77, 98 mod, 110
frequency-selective channel, 67 Moose method, 97
frequency-selective fading, 81, 95, Morse code, 38
115 multi-path channel, 81
generator matrix, 106 Nyquist criterion, 6

GenesisRadio, 145 Nyquist pulse, 54
get, 13
getaudiodata, 13 OFDM, 113
gradient descent, 131
packet, 88
Gray code, 40
packetdetect.m, 92, 157
Hamming distance, 107 parity bits, 106
header, 88 parity check matrix, 106
path, 15, 92
I, see in-phase signal payload, 88
INDEX 169
phase offset, 10, 29, 78 tap, 81, 97

phase tracker, 133 time-frequency grid, 125
phase-shift keying, 38 timing recovery, 68
physical layer, 1 training code, see pilot
PicoZed, 145
pilot, 66, 126 unit-distance code, 40
pinv, 101 unwrap, 79
play, 14 upconversion, see amplitude mod-
plottf.m, 14, 148 ulation
PN code, 82 upper sideband, 8
polyfit, 79 upsampling, 10, 53
pseudo-inverse, 101 var, 66
PSK, see phase-shift keying variance, 66
psk2char.m, 45, 153 Viterbi algorithm, 115
pulse shape, 38, 51
Warp, 145
Q, see quadrature signal wideband channel, 115
Q function, 43
QAM, 23, 37 Xilinx, 145
qfunc, 43
QPSK, 38
quadrature signal, 24
randn, 16
recordblocking, 13
rtl, 146
sampling, 6
sampling rate, 6
save, 92
scatter plot, 40
signal space, 38
sinc interpolation, 10
software-defined radio, 2
spectral efficiency, 44, 88
square-root raised cosine, 55
srrc.m, 58, 154
subcarrier frequency, 124
subplot, 34
symbol timing, 64

CommunicationsLaboratory 20161206

Uploaded by

Copyright:

Available Formats

CommunicationsLaboratory 20161206

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

CommunicationsLaboratory 20161206

Uploaded by

Copyright:

Available Formats

A Digital Communication

Lee C. Potter and Yang Yang

Terms of Use: This work is licensed under a Creative Commons

This volume was typeset by the authors in LATEX.

First edition, December 2015

Matlab and Simulink are trademarks of MathWorks. LabVIEW, USRP, Uni-

2 Quadrature Amplitude Modulation 23

4 Pulse Shaping & ISI 51

7.1.4 Differential PSK modulation . . . . . . . . . . 90

9 Channel Coding 103

10 OFDM Part 1 113

11 OFDM Part 2 121

12 Adaptive Processing 129

A Software Design Suggestions 139

This text is intended as a student guide for hands-on exploration of

inevitably experience rapid successions of updates or replacements.

on communication system performance measured as bits per second

The goals of this introductory exercise are to review basic concepts

1.1 A Physical Layer Model

1.2 Software Radio

Pilot sequence -j2π(fc+fΔ)t+φ Rx Amp

to provide new features to existing systems. The steady advancement

Im Pulse Shaping Filter

Figure 1.2: Illustrations of (a) an example symbol constellation; (b)

plifier and filtered before downconversion to in-phase (I) and quad-

A similar set of steps occurs on transmission. Digital upconver-

Local oscillator Baseband

Figure 1.3: A typical SDR architecture.

1.3.1 Sampling and aliasing

where M (f ) is the spectrum of the original analog signal, m(t). Each

In this manner, an alias of M (f ) appears with frequency support

1.3.2 Frequency upconversion

Upconversion (amplitude modulation or mixing) of a real-valued

s(t) = m(t) cos(2πfc t) (1.3)

and serves to translate the message to a desired frequency band cen-

The spectrum, S(f ), is found via the Fourier transform

Because m(t) is real-valued, M (f ) is conjugate symmetric about f =

Figure 1.5: Amplitude modulation performs a translation in fre-

Figure 1.6: Magnitude spectrum resulting from amplitude modula-

1.3.3 Frequency downconversion

v(t) = LPF{s(t) × 2 cos(2πfc t)}

This process in depicted in Figure 1.7 and Figure 1.8.

r(t) × LPF v(t)

Figure 1.8: Low-pass filter (dashed line) used in amplitude demodu-

When the receiver oscillator has frequency offset, f∆ , or phase

The time-varying attenuation can be viewed as an upconversion of

The processing to produce the interpolated samples can be viewed

Figure 1.9: Block diagram for ideal interpolation.

DAC: sinc interpolation

where T is the sampling period for x[k]. The summation is the

function) with the impulse train of samples, x[k]δ(t − kT ). This

Figure 1.10: Illustration of sinc interpolation.

Decimation is a digital filtering procedure that decreases the sam-

Figure 1.11: Block diagram for decimation by L; the anti-aliasing

1.4.1 Getting started with audio I/O

>> recObj = audiorecorder(Fs,Nbits,nChannels) sets the

>> playObj = audioplayer(Y,Fs,Nbits) creates an audio-

The following code snippet provides two examples.

1.4.2 Spectral content of signals