AN INVESTIGATION OF CARRIER RECOVERY

TECHNIQUES FOR PSK MODULATED SIGNALS IN

CDMA AND MULTIPATH MOBILE ENVIRONMENTS

by
Steven P. Nicoloso

Thesis submitted to the Faculty of the

Virginia Polytechnic Institute and State University
in partial fulfillment of the requirements for the degree of

MASTER OF SCIENCE
in
The Bradley Department of Electrical Engineering

Approved:

Dr. Jeffrey H. Reed

(Chairman)

Dr. Warren Stutzman Dr. Brian D. Woerner

June 1997
Blacksburg, Virginia
An Investigation of Carrier Recovery Techniques for PSK
Modulated Signals in CDMA and Multipath Mobile
Environments
by
Steven P. Nicoloso
Committee Chairman: Dr. Jeffrey H. Reed
The Bradley Department of Electrical Engineering

Abstract

The challenge of carrier recovery for digitally phase modulated, suppressed-carrier

signals has been the target of much attention in mobile radio system design for many
years. Code division multiple access (CDMA) cellular systems are now being deployed
in mobile environments, and the performance of carrier recovery for CDMA is much
less well understood. CDMA systems typically utilize lower bit energies, relying instead
upon coding gain for reliable performance. As an additional challenge, the multiple ac-
cess interference (MAI) inherent to direct-sequence spread-spectrum (DSSS) signalling
further hampers many carrier recovery techniques.
This thesis first surveys general carrier recovery strategies, conventional as well as
new, and then applies them to the particular task of recovering a suppressed carrier in
mobile CDMA systems. First some of the theory surrounding conventional, closed-loop
carrier recovery techniques and several newer open-loop estimation structures based
on the maximum likelihood (ML) principle is presented. A novel decision-feedback ML
phase estimator is proposed. The statistical behavior of this new structure as well as
the related squaring estimator are examined and are shown to be equivalent to the
stochastic performance of classical closed-loop techniques.
Candidate receivers based upon closed- and open-loop carrier recovery structures
are proposed and then examined via simulation. Both forward and reverse cellular
links are examined under a variety of both of single-path and multipath conditions.
The notion of exploiting cyclostationary MAI in CDMA systems by applying adaptive
receiver techniques is also examined, especially as this scheme relies to some extent
upon coherent carrier recovery for its promising performance enhancements. This work
concludes with recommendations of the best carrier recovery strategies for the variety
of environments examined and suggestions for future research are made.
Acknowledgements

The work presented herein was made possible by the personal contributions of a variety
of individuals with the Mobile and Portable Radio Research Group (MPRG) to whom I
am greatly indebted. Nitin Mangalvedhe has been very involved in every portion of this
research. He has made key contributions, both theoretical and practical, throughout. His
leadership and broad knowledge have been of benefit not only to me, but to many of
my colleagues as well. Francis Dominique, more than any other person, suggested and
motivated me toward this research near the very beginning of my tenure at MPRG. Mike
Buehrer not only contribued excellent edits to the theoretical portions of this work, but
personally contributed to my development as an engineer and thinker. Working side-by-
side with Neiyer Correal and Paul Petrus in coding up algorithms for DSP-implementation,
though well-outside the scope of the work presented here, was nevertheless a pleasure and
an outstanding learning experience for me.
Among the original GloMo-Gang, Nevena Zec̆ević and Milap Majmundar were our orig-
inal investigators of the adaptive filtering receivers examined in this work. I have a memory
of Milap sitting down with me very early in my stint at MPRG and explaining quite clearly
to me the implementation of adaptive equalization—long before I even knew what an FIR
filter was! Dan Bailey, though long matriculated, was very helpful and patient with me
in getting started in coding algorithms for our tarket DSP hardware. Don Breslin, Keith
Blankenship, Pascal Renucci, Tom Biedka, Greg Durgin, Kim Phillips, Rich Ertel, and
Kevin and Donna Krizman have not only made valuable comments to me concerning this
work, but have also been great fun to be around at MPRG. Jeff Laster, a legendary student
of ours in his own right, served me not only as an experienced learner and teacher but also
as a spiritual advisor and friend. His recent moving on to industry creates a void for us that
will undoubtedly be difficult to compensate. A host of new names and faces has arrived at
MPRG, some of which I hardly know. For each of you, I extend a confident hope that your
graduate school experience will be as personally rewarding as has been mine. Many thanks
to all!

iii
 My committee members have been nothing but helpful both inside and outside of class.
Jeff Reed has been everything one could want in an advisor. His recent tenure at Virginia
Tech is well-deserved. He unequivocally sees and brings out the very best in all of his
students. His constant encouragement, liberality, and trust for me have been strong personal
motivators, and I consider Jeff Reed a close personal friend. It has been an honor to
have a professor of the stature of Warren Stutzman on my thesis committee. Though the
distance between Whittemore Hall and Pointe West Commons has separated us, he has
been extremely helpful on a number of occasions, and accomodated my faults. Taking a
class, any class, from Dr. Stutzman is highly recommended, though it may take 6-8 weeks
to realize “he was probably joking!” Brian Woerner’s commitment to his students and
teaching is legendary. His expertise in modulation, coding, and information theory as well
as simulation has often been helpful throughout this work. Many thanks to each of these
men!
The staff of MPRG has been simply superb in supporting each of us. They have not
only helped me, time and again, with specific needs but consistently take care of the ad-
ministrative tasks and day-to-day business of MPRG in such a way as to make these needs
transparent. Their only failure may be in doing the job so well as to cause the rest of us
to take them for granted. Let me avoid such an error here. To our fine staff, both past
and present, Prab Koushik, Wayne Erchak, Annie Wade, Kathy Wolfe, Jenny Frank, Hilda
Reynolds, Aurelia Scharnhorst, Rennie Givens, and Lori Hughes, I salute and thank you
for your hard work on my behalf, and on behalf of all of us.
Much of my financial support for the last two years was provided by the Defense Ad-
vanced Research Projects Agency (DARPA) who expressed their faith in the work of the
MPRG by sponsoring the Global Mobile Communications Initiative (GloMo). As a par-
ticipant in this project, I developed most of the professional relationships I now have and
learned most of what is presented here and even more that is not.
I’d like to point out that my family has endured these last five years of my being
a full-time student with remarkable grace and strength. To Robin, who was there from
the beginning, all of my love and gratitude. And for Daniel and Karen, who ”popped
up” along the way as a serendipitous by-product of love, my committment to getting this
degree, though you probably cannot understand it now, has been both because of you and
for you. To my parents, Luigi and Karen Nicoloso, I give heartfelt thanks not only for
the financial sacrifices which made graduate school possible for me but for the sacrifice of
love and support upon which no price can be placed. The participants of Grace Covenant
Presbyterian Church have been very much like a family for us in Blacksburg. We love each

iv
of you, and have constantly benefited from your love and strength. Finally, to my LORD
and Savior, Jesus Christ, God of very God, I reserve the highest of all praise!—in whose
life I find life, and to whose wisdom I forever humbly submit.

v
Contents

Acknowledgements iii

1 Introduction 1
1.1 Motivations for Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Thesis Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Overview of PSK Reception . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3.1 PSK Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3.2 Coherent Reception . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3.3 Closed Vs. Open Loop Compensation . . . . . . . . . . . . . . . . . 6

2 A Brief Survey of Prior Research 8

2.1 Closed-Loop Synchronization . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2 Open-Loop Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

3 The Phase-Locked Loop 12

3.1 Introductory View . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2 Linearized Analog PLL Model . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2.1 The Second-Order PLL . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2.2 The Third-Order PLL . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3 Derivation of Discrete-time PLL Models . . . . . . . . . . . . . . . . . . . . 16
3.3.1 Discrete-time Second-Order PLL . . . . . . . . . . . . . . . . . . . . 16
3.3.2 Discrete-time Third-Order PLL . . . . . . . . . . . . . . . . . . . . . 18
3.4 Performance Measures for PLLs . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.4.1 Loop Noise Bandwidth . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.4.2 Transient Responses . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.4.3 Cycle Slipping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.4.4 Acquisition Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.5 Performance of Derived Discrete-time Models . . . . . . . . . . . . . . . . . 29

vi
 3.6 The Notion of Optimality for a PLL . . . . . . . . . . . . . . . . . . . . . . 32
3.7 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

4 Coherent Carrier Recovery Structures 35

4.1 Closed-Loop Carrier Recovery Structures . . . . . . . . . . . . . . . . . . . 36
4.1.1 Costas Loop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.1.2 Squaring Loop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.1.3 Noise Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.1.4 Discrete-Time Implementation . . . . . . . . . . . . . . . . . . . . . 39
4.1.5 Four-Phase Costas Loop Implementation . . . . . . . . . . . . . . . 42
4.1.6 Digital Tanlock Loop . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.1.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.2 Open-Loop Carrier Recovery Structures . . . . . . . . . . . . . . . . . . . . 44
4.2.1 First-Order ML Phase Estimator . . . . . . . . . . . . . . . . . . . . 45
4.2.2 Noise Performance of Low SNR ML Phase Estimator . . . . . . . . . 48
4.2.3 Noise Performance of High SNR ML Phase Estimator . . . . . . . . 53
4.2.4 Modified First-Order ML Phase Estimator . . . . . . . . . . . . . . . 55
4.2.5 Second-Order ML Phase Estimator . . . . . . . . . . . . . . . . . . . 58
4.2.6 Simulation Study of Open-Loop Phase Estimators . . . . . . . . . . 59
4.2.7 Application of Open-Loop Estimation to Receivers Employing Adap-
tive Equalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.3 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

5 Automatic Frequency Control 64

5.1 Introduction to AFC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.2 Frequency Difference Detectors . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.2.1 The Balanced Quadricorrelator . . . . . . . . . . . . . . . . . . . . . 67
5.2.2 Angle Doubling AFC . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.3 Acquisition Time of AFC loops . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.4 Costas Loops with AFC-Aided Acquisition . . . . . . . . . . . . . . . . . . . 72
5.4.1 Description of Combined Costas Demodulator with AFC for Acquis-
tion Aiding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.4.2 Performance Issues for Combined Costas Demodulators with AFC . 75

6 Simulation Strategy 82
6.1 BER as a Figure of Merit . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
6.2 CDMA System Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

vii
 6.2.1 System Specifications . . . . . . . . . . . . . . . . . . . . . . . . . . 84
6.2.2 Fading Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
6.3 The Mobile Links . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
6.3.1 The Forward Link . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
6.3.2 The Reverse Link . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
6.4 Candidate Receivers Structures . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.4.1 Original Candidate Structures . . . . . . . . . . . . . . . . . . . . . 93
6.4.2 The “Weeding-Out” Process . . . . . . . . . . . . . . . . . . . . . . 95
6.4.3 Final Candidate Structures . . . . . . . . . . . . . . . . . . . . . . . 99
6.5 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

7 Receiver Performance in AWGN Channels 106

7.1 Notes on Conventional vs. Adaptive Reception . . . . . . . . . . . . . . . . 106
7.2 The Forward Link in AWGN . . . . . . . . . . . . . . . . . . . . . . . . . . 109
7.2.1 Costas Loop Jitter on Synchronous CDMA Links . . . . . . . . . . . 110
7.2.2 Results on the Forward Link in AWGN . . . . . . . . . . . . . . . . 111
7.3 The Reverse Link in AWGN . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
7.3.1 Estimator Bias in Multiple Access Channels . . . . . . . . . . . . . . 116
7.3.2 Results in Strict Power Control . . . . . . . . . . . . . . . . . . . . . 117
7.3.3 Results in Loose Power Control . . . . . . . . . . . . . . . . . . . . . 123
7.4 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126

8 Receiver Performance in Fading Channels 127

8.1 The One-Ray Rayleigh, “Flat Fading” Channel . . . . . . . . . . . . . . . . 128
8.1.1 Performance on the Synchronous Forward Link . . . . . . . . . . . . 128
8.1.2 Performance on the Asynchronous Reverse Link . . . . . . . . . . . 129
8.2 The Two-Ray Rayleigh, “Frequency Selective” Channel . . . . . . . . . . . 132
8.2.1 Performance on the Synchronous Forward Link . . . . . . . . . . . . 132
8.2.2 Performance on the Asynchronous Reverse Link . . . . . . . . . . . 133
8.3 The COST-207 “Typical Urban” Channel . . . . . . . . . . . . . . . . . . . 136
8.3.1 Performance on the Synchronous Forward Link . . . . . . . . . . . . 136
8.3.2 Performance on the Asynchronous Reverse Link . . . . . . . . . . . 140
8.4 The COST-207 “Typical Rural” Channel . . . . . . . . . . . . . . . . . . . 140
8.4.1 Performance on the Synchronous Forward Link . . . . . . . . . . . . 140
8.4.2 Performance on the Asynchronous Reverse Link . . . . . . . . . . . 141
8.5 The Effect of Training Period and Frequency Offset on Receiver Performance 144

viii
 8.5.1 The Impact of Training and Synchronization Period . . . . . . . . . 144
8.5.2 The Impact of Frequency Uncertainty . . . . . . . . . . . . . . . . . 146
8.6 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146

9 Conclusions 149
9.1 Significant Contributions and Findings of this Research . . . . . . . . . . . 149
9.1.1 Theoretical Contributions . . . . . . . . . . . . . . . . . . . . . . . . 149
9.1.2 Experimental Contributions . . . . . . . . . . . . . . . . . . . . . . . 150
9.2 Recommendations for Future Work . . . . . . . . . . . . . . . . . . . . . . . 152
9.2.1 Broadening the Scope of Investigation . . . . . . . . . . . . . . . . . 153
9.2.2 More Robust Adaptive Receiver Designs . . . . . . . . . . . . . . . . 154
9.3 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155

Bibliography 160

Author’s Biographical Sketch 161

ix
List of Tables

3.1 Optimum PLLs for three input signals in AWGN . . . . . . . . . . . . . . . 33

4.1 Intermodulation losses in M th-order phase synchronizers . . . . . . . . . . . 39
6.1 General parameters for simulation study . . . . . . . . . . . . . . . . . . . . 86
6.2 Simulation channels models . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
6.3 Link-specific parameters for simulation study . . . . . . . . . . . . . . . . . 91
6.4 Potential receiver structures with symbol rate processing . . . . . . . . . . . 94
6.5 Final candidate receiver structures simulated and compared in future chap-
ters on the basis of BER performance . . . . . . . . . . . . . . . . . . . . . 99

x
List of Figures

1.1 Basic detection structure for PSK signals . . . . . . . . . . . . . . . . . . . 5

1.2 Example of a feedback control system: A simple PLL model . . . . . . . . . 6
1.3 Example open-loop compensation scheme: ML phase estimator . . . . . . . 7
3.1 Simple illustrative PLL model . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2 Linearized, phase domain model of PLL . . . . . . . . . . . . . . . . . . . . 14
3.3 High gain, active, lowpass filter suitable to realize loop filter in canonical
PLL model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.4 Discrete-time, phase domain model of PLL . . . . . . . . . . . . . . . . . . 17
3.5 Normalized loop bandwidth, BL ωn plotted as a funtion of ζ, for canonical
second-order loop, H(s) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.6 Transient responses of 2nd-order PLL for various values of ζ due to phase
step input of π/2 radians . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.7 Transient responses of 2nd-order PLL for various values of ζ due to frequency
step input: (A) Frequency step, ωof f = ωn ; (B) Frequency step, ωof f = 2ωn 26
3.8 Transient responses of 2nd-order PLL for various values of ζ due to frequency
ramp input (time normalized with ωn ; error normalized with ωn2 /∆ω̇) . . . 26
3.9 Example of cycle slipping (input SNR = -8 dB, Bi /BL ≈ 10) . . . . . . . . 27
3.10 Acquisition time for 2nd-order PLL for various values of initial frequency
offset (ζ = 2− 2 ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1
28
3.11 Frequency responses of second-order, discrete-time PLLs expressed in equa-
tions (3.14) and (3.18) for two choices of ωn (ζ = 2− 2 ) . . . . . . . . . . . .
1
30
3.12 Frequency responses of third-order, discrete-time PLLs expressed in equa-
tions (3.20) and (3.22) for two choices of ω3 . . . . . . . . . . . . . . . . . . 30
3.13 Phase output variance (jitter) of derived discrete-time 2nd-order PLL, versus
theoretical values, for two values of ωn . . . . . . . . . . . . . . . . . . . . . 31
3.14 Phase output variance (jitter) of derived discrete-time 3rd-order PLL, versus
theoretical values, for two values of ω3 . . . . . . . . . . . . . . . . . . . . . 31

xi
4.1 Basic structure of analog Costas demodulator . . . . . . . . . . . . . . . . . 37
4.2 Squaring Loop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.3 Block diagram of Costas demodulator, discrete-time implementation . . . . 40
4.4 Block diagram of four-phase discrete-time Costas demodulator . . . . . . . 43
4.5 Block diagram of digital tanlock loop (DTL) . . . . . . . . . . . . . . . . . 43
4.6 “First-order” open-loop ML phase estimation structure proposed in [41] . . 46
4.7 Decision feedback open-loop phase estimation structure derived from the ML
principle under “high” SNR assumptions . . . . . . . . . . . . . . . . . . . . 48
4.8 Modified first-order ML phase estimation structure, suitable for the condi-
tions of “small” residual frequency offset . . . . . . . . . . . . . . . . . . . . 56
4.9 Second-order ML phase estimation structure with dedicated frequency esti-
mation, suitable for the conditions of “large” residual frequency offset . . . 58
4.10 Variance of output phase estimate of first-order phase estimator, using squar-
ing operation, for a variety of Eb /No . Lines denote theoretical performance
predicted by (4.52), asterisks denote simulation points . . . . . . . . . . . . 60
4.11 Variance of output phase estimate of first-order phase estimator, using deci-
sion feedback, for a variety of Eb /No . Lines denote theoretical performance
predicted by (4.68), asterisks denote simulation points . . . . . . . . . . . . 60
4.12 Open loop carrier estimation applied to an adaptive equalization receiver,
shown here employing decision direction only. This structure was adapted
from [51] for this research . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.13 Differentially coherent receiver employing adaptive interference suppression
for BPSK—a limiting condition of ML phase estimation . . . . . . . . . . . 62
5.1 Linearized, frequency domain model of AFC loop. . . . . . . . . . . . . . . 65
5.2 Block digram of continuous-time balanced quadricorrelator (BQ). . . . . . 68
5.3 Block digram of discrete-time implementation of Angle-doubling AFC (ADAFC)
proposed in [26]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.4 Simulated acquisition time in samples for BQ-AFC and Costas loops versus
normalized loop bandwidth of respective loops; ∗ mark simulation points for
AFC loop; × mark simulation points for Costas loop. . . . . . . . . . . . . . 72
5.5 Receiver block diagram suggesting the combination of AFC with a Costas
Loop. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.6 Block diagram of discrete-time implementation of second-order Costas de-
modulator with a ADAFC loop for acquisition aiding. . . . . . . . . . . . . 74

xii
5.7 Typical acquisition behavior for three styles of carrier trackers: (A) An non-
coherent AFC loop; (B) An unaided Costas loop; C. The same Costas and
AFC loops combined; Input is BPSK signal with Eb /No = 12 dB. . . . . . . 75
5.8 Degradation of phase output estimate for coherent carrier recovery with the
addition of an AFC loop; Eb /No = 20 dB; ∗ denote simulated points; inter-
polation used to help show detail; all values logarithmic (base 10). . . . . . 78
5.9 Degradation of phase output estimate for coherent carrier recovery with the
addition of an AFC loop; Eb /No = 10 dB; ∗ denote simulated points; inter-
polation used to help show detail; all values logarithmic (base 10). . . . . . 79
6.1 Bit error probabilities for three different binary signalling and recovery strate-
gies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.2 Simulation results for bit error probabilities versus Eb /No for three Costas
demodulator structures; frequency uncertainty, 2σf = 1/40Rb ; simulated
results for DBPSK and theoretical DEBPSK performance also shown . . . . 96
6.3 Simulation results for bit error probabilities versus Eb /No for three combined
AFC and Costas loop carrier recovery structures; Costas bandwidth, BL,c is
the same for all three; frequency uncertainty, 2σf = 1/40Rb ; Costas loop act-
ing alone shown for comparison; simulated results for DBPSK and theoretical
DEBPSK performance also shown . . . . . . . . . . . . . . . . . . . . . . . 96
6.4 Simulation results for bit error probabilities versus Eb /No for two combined
AFC/Costas loop carrier recovery structures and two switched loop struc-
tures; Costas bandwidth, BL,c = 10−2 Rb in all cases; frequency uncertainty,
2σf = 1/40Rb ; simulated results for DBPSK and theoretical DEBPSK per-
formance also shown . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
6.5 Simulation results for bit error probabilities versus Eb /No for four ML open-
loop carrier recovery structures; frequency uncertainty, 2σf = 1/40Rb ; simu-
lated results for DBPSK and theoretical DEBPSK performance also shown 98
6.6 Conventional matched filter receiver for DEBPSK with “ideal” coherent car-
rier reference; M1,1,i corresponds to the first arriving fading waveform of user
1 at time i. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
6.7 Conventional matched filter receiver using differentially coherent (DBPSK)
detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
6.8 Conventional DEBPSK matched filter receiver with Costas loop carrier recovery101
6.9 Conventional DEBPSK matched filter receiver with combined Costas and
AFC loops for carrier recovery . . . . . . . . . . . . . . . . . . . . . . . . . 102

xiii
6.10 Conventional DEBPSK matched filter receiver with modified first-order ML
carrier phase estimation circuit . . . . . . . . . . . . . . . . . . . . . . . . . 102
6.11 Three finger differentially coherent RAKE receiver with equal gain combin-
ing; First three arriving rays are assumed known, and first path is assumed
to have delay, τ = 0. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.12 Adaptive filter receiver for DEBPSK with “ideal” coherent carrier reference;
M1,1,i corresponds to the first arriving fading waveform of user 1 at time i. 103
6.13 Adaptive filter receiver employing differentially coherent (DBPSK) detection 104
6.14 Adaptive filter receiver for DEBPSK with conventional Costas loop carrier
recovery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
6.15 Adaptive filter receiver for DEBPSK with either modified first-order or second-
order open-loop phase estimation scheme . . . . . . . . . . . . . . . . . . . 105
7.1 Cross-correlation of all user’s pseudo-random codes with those of user 1 for
matched filter and CWFSLAR filter, the latter exploiting cyclostationarity
in the MAI; adapted from [52]. . . . . . . . . . . . . . . . . . . . . . . . . . 107
7.2 Simulation results for bit error probabilites versus number of users on the
forward link in AWGN channel; Eb /No = 8 dB; performance of conventional
matched and adaptive filters with “ideal” and differentially coherent recep-
tion shown; frequency uncertainty, 2σf = 1/40Rb . . . . . . . . . . . . . . . . 108
7.3 Simulation results for bit error probabilites versus number of users on the
forward link in AWGN channel; Eb /No = 10 dB; performance of conven-
tional matched and adaptive filters with “ideal” and differentially coherent
reception shown; frequency uncertainty, 2σf = 1/40Rb . . . . . . . . . . . . . 108
7.4 Simulation results for bit error probabilites versus number of users on the
forward link in AWGN channel; Eb /No = 8 dB; performance of conventional
matched filter receivers with various Costas loops, “ideal” carrier recovery,
and differentially coherent reception shown; frequency uncertainty, 2σf =
1/40Rb . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
7.5 Simulation results for bit error probabilites versus number of users on the
forward link in AWGN channel; Eb /No = 8 dB; performance of adaptive filter
receivers (CWFSLAR) with various Costas loops, “ideal” carrier recovery,
and differentially coherent reception shown; frequency uncertainty, 2σf =
1/40Rb . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

xiv
7.6 Simulation results for bit error probabilites versus number of users on the
forward link in AWGN channel; Eb /No = 10 dB; performance of conventional
matched filter receivers with various Costas loops, “ideal” carrier recovery,
and differentially coherent reception shown; frequency uncertainty, 2σf =
1/40Rb . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
7.7 Simulation results for bit error probabilites versus number of users on the
forward link in AWGN channel; Eb /No = 10 dB; performance of adaptive
filter receivers (CWFSLAR) with various Costas loops, “ideal” carrier re-
covery, and differentially coherent reception shown; frequency uncertainty,
2σf = 1/40Rb . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
7.8 Simulation results for bit error probabilites versus number of users on the for-
ward link in AWGN channel; Eb /No = 10 dB; performance of conv. matched
filter receivers with various modified 1st-order ML phase estimation struc-
tures, “ideal” carrier recovery, and differentially coherent reception shown;
frequency uncertainty, 2σf = 1/40Rb . . . . . . . . . . . . . . . . . . . . . . . 114
7.9 Simulation results for bit error probabilites versus number of users on the for-
ward link in AWGN channel; Eb /No = 10 dB; performance of adaptive filter
receivers (CWFSLAR) with various modified 1st-order ML phase estima-
tion structures, “ideal” carrier recovery, and differentially coherent reception
shown; freq. uncertainty, 2σf = 1/40Rb . . . . . . . . . . . . . . . . . . . . . 114
7.10 Simulation results for bit error probabilites versus number of users on reverse
link in AWGN channel; “strict” power control, σP2 = 2 dB2 ; Eb /No = 8 dB;
performance of conv. matched filter receivers with various Costas loops,
“ideal” carrier recovery, and differentially coherent reception shown; freq.
uncert., 2σf = 1/40Rb . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
7.11 Simulation results for bit error probabilites versus number of users on reverse
link in AWGN channel; “strict” power control, σP2 = 2 dB2 ; Eb /No = 8 dB;
performance of adaptive filter receivers (CWFSLAR) with various Costas
loops, “ideal” carrier recovery, and differentially coherent reception shown;
freq. uncert., 2σf = 1/40Rb . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
7.12 Simulation results for bit error probabilites versus number of users on reverse
link in AWGN channel; “strict” power control, σP2 = 2 dB2 ; Eb /No = 12 dB;
performance of conv. matched filter receivers with various Costas loops,
“ideal” carrier recovery, and differentially coherent reception shown; freq.
uncert., 2σf = 1/40Rb . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120

xv
7.13 Simulation results for bit error probabilites versus number of users on reverse
link in AWGN channel; “strict” power control, σP2 = 2 dB2 ; Eb /No = 12 dB;
performance of adaptive filter receivers (CWFSLAR) with various Costas
loops, “ideal” carrier recovery, and differentially coherent reception shown;
freq. uncert., 2σf = 1/40Rb . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
7.14 Simulation results for bit error probabilites versus number of users on reverse
link in AWGN channel; “strict” power control, σP2 = 2 dB2 ; Eb /No = 12 dB;
performance of conv. matched filter receivers with various modified 1st-
order ML phase est. structures, “ideal” carrier recovery, and diff. coherent
reception shown; freq. uncert., 2σf = 1/40Rb . . . . . . . . . . . . . . . . . . 121
7.15 Simulation results for bit error probabilites versus number of users on reverse
link in AWGN channel; “strict” power control, σP2 = 2 dB2 ; Eb /No = 12 dB;
performance of adaptive receivers (CWFSLAR) with various modified 1st-
order ML phase est. structures, “ideal” carrier recovery, and diff. coherent
reception shown; freq. uncert., 2σf = 1/40Rb . . . . . . . . . . . . . . . . . . 121
7.16 Same simulation results as shown in Figure 7.15 (check), except BER results
exceeding 0.2 are discarded and replaced with the receiver failure rate. . . . 122
7.17 Simulation results for bit error probabilites versus number of users on reverse
link in AWGN channel; “loose” power control, σP2 = 20 dB2 ; Eb /No = 12
dB; performance of conv. matched filter receivers with various Costas loops,
“ideal” carrier recovery, and differentially coherent reception shown; freq.
uncert., 2σf = 1/40Rb . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
7.18 Simulation results for bit error probabilites versus number of users on reverse
link in AWGN channel; “loose” power control, σP2 = 20 dB2 ; Eb /No = 12 dB;
performance of adaptive receivers (CWFSLAR) with various Costas loops,
“ideal” carrier recovery, and differentially coherent reception shown; freq.
uncert., 2σf = 1/40Rb . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
7.19 Simulation results for bit error probabilites versus number of users on reverse
link in AWGN channel; “loose” power control, σP2 = 20 dB2 ; Eb /No = 12
dB; performance of conv. matched filter receivers with various modified 1st-
order ML phase est. structures, “ideal” carrier recovery, and diff. coherent
reception shown; freq. uncert., 2σf = 1/40Rb . . . . . . . . . . . . . . . . . . 125

xvi
7.20 Simulation results for bit error probabilites versus number of users on reverse
link in AWGN channel; “loose” power control, σP2 = 20 dB2 ; Eb /No = 12 dB;
performance of adaptive receivers (CWFSLAR) with various modified 1st-
order ML phase est. structures, “ideal” carrier recovery, and diff. coherent
reception shown; freq. uncert., 2σf = 1/40Rb . . . . . . . . . . . . . . . . . . 125
8.1 Simulation results for bit error probabilites vs. number of users on the for-
ward link in flat Rayleigh-faded channel; Eb /No = 20 dB; performance of
conventional receivers with “ideal” and differentially coherent reception, wide
BW Costas loop and N = 16 modified 1st-order ML structure; freq. uncer-
tainty, 2σf = 1/40Rb ; vehicle speed is 45 kph. . . . . . . . . . . . . . . . . . 130
8.2 Simulation results for bit error probabilites vs. number of users on the for-
ward link in flat Rayleigh-faded channel; Eb /No = 20 dB; performance of
adaptive receivers with “ideal” and differentially coherent reception, wide
BW Costas loop and N = 16 modified 1st-order ML structure; freq. uncer-
tainty, 2σf = 1/40Rb ; vehicle speed is 45 kph. . . . . . . . . . . . . . . . . . 130
8.3 Simulation results for bit error probabilites vs. number of users on the reverse
link in flat Rayleigh-faded channel; “strict” power control, σP2 = 2 dB2 ;
Eb /No = 15 dB; performance of conventional receivers with “ideal” and
differentially coherent reception, wide BW Costas loop and N = 5 modified
1st-order ML structure; freq. uncertainty, 2σf = 1/40Rb ; vehicle speed is 15
kph. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
8.4 Simulation results for bit error probabilites vs. number of users on the re-
verse link in flat Rayleigh-faded channel; “strict” power control, σP2 = 2 dB2 ;
Eb /No = 15 dB; performance of adaptive receivers with “ideal” and differen-
tially coherent reception, wide BW Costas loop and N = 5 modified 1st-order
ML structure; freq. uncertainty, 2σf = 1/40Rb ; vehicle speed is 15 kph. . . 131
8.5 Simulation results for bit error probabilites vs. number of users on the for-
ward link in 2-ray Rayleigh-faded, frequency-selective channel; Eb /No = 28
dB; performance of conventional receivers with “ideal” and differentially co-
herent reception, wide BW Costas loop and N = 5 modified 1st-order ML
structure; 3-finger RAKE also shown; freq. uncertainty, 2σf = 1/40Rb ; vehi-
cle speed is 15 kph. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134

xvii
8.6 Simulation results for bit error probabilites vs. number of users on the for-
ward link in 2-ray Rayleigh-faded, frequency-selective channel; Eb /No = 28
dB; performance of adaptive receivers with “ideal” and differentially coherent
reception, wide BW Costas loop and N = 5 modified 1st-order ML structure;
3-finger RAKE also shown; freq. uncertainty, 2σf = 1/40Rb ; vehicle speed is
15 kph. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
8.7 Simulation results for bit error probabilites vs. number of users on the re-
verse link in 2-ray Rayleigh-faded, frequency-selective channel; “strict” power
control, σP2 = 2 dB2 ; Eb /No = 28 dB; performance of conventional receivers
with “ideal” and differentially coherent reception, wide BW Costas loop and
N = 5 modified 1st-order ML structure; 3-finger RAKE also shown; freq.
uncertainty, 2σf = 1/40Rb ; vehicle speed is 45 kph. . . . . . . . . . . . . . . 135
8.8 Simulation results for bit error probabilites vs. number of users on the re-
verse link in 2-ray Rayleigh-faded, frequency-selective channel; “strict” power
control, σP2 = 2 dB2 ; Eb /No = 28 dB; performance of adaptive receivers
with “ideal” and differentially coherent reception, wide BW Costas loop and
N = 5 modified 1st-order ML structure; 3-finger RAKE also shown; freq.
uncertainty, 2σf = 1/40Rb ; vehicle speed is 45 kph. . . . . . . . . . . . . . . 135
8.9 Simulation results for bit error probabilites vs. number of users on the for-
ward link in COST-207 “Typical Urban” channel; Eb /No = 24 dB; per-
formance of conventional receivers with “ideal” and differentially coherent
reception, wide BW Costas loop and N = 16 modified 1st-order ML struc-
ture; 3-finger RAKE also shown; freq. uncertainty, 2σf = 1/40Rb ; vehicle
speed is 45 kph. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
8.10 Simulation results for bit error probabilites vs. number of users on the for-
ward link in COST-207 “Typical Urban” channel; Eb /No = 24 dB; perfor-
mance of adaptive receivers with “ideal” and differentially coherent reception,
wide BW Costas loop and N = 16 modified 1st-order ML structure; 3-finger
RAKE also shown; freq. uncertainty, 2σf = 1/40Rb ; vehicle speed is 45 kph. 137
8.11 Results for the same conditions given in Figure 8.10, except at a slower vehicle
speed of 15 kph. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138

xviii
8.12 Simulation results for bit error probabilites vs. number of users on the reverse
link in COST-207 “Typical Urban” channel; Eb /No = 24 dB; “strict” power
control, σP2 = 2 dB2 ; performance of conventional receivers with “ideal” and
differentially coherent reception, wide BW Costas loop and N = 16 modified
1st-order ML structure; 3-finger RAKE also shown; freq. uncertainty, 2σf =
1/40Rb ; vehicle speed is 15 kph. . . . . . . . . . . . . . . . . . . . . . . . . . 139
8.13 Simulation results for bit error probabilites vs. number of users on the revers
link in COST-207 “Typical Urban” channel; Eb /No = 24 dB; “strict” power
control, σP2 = 2 dB2 ; performance of adaptive receivers with “ideal” and
differentially coherent reception, wide BW Costas loop and N = 16 modified
1st-order ML structure; 3-finger RAKE also shown; freq. uncertainty, 2σf =
1/40Rb ; vehicle speed is 15 kph. . . . . . . . . . . . . . . . . . . . . . . . . . 139
8.14 Simulation results for bit error probabilites vs. number of users on the for-
ward link in COST-207 “Typical Rural” channel; Eb /No = 24 dB; perfor-
mance of conventional receivers with “ideal” and differentially coherent re-
ception, wide BW Costas loop and N = 5 modified 1st-order ML structure;
3-finger RAKE also shown; freq. uncertainty, 2σf = 1/40Rb ; vehicle speed is
45 kph. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
8.15 Simulation results for bit error probabilites vs. number of users on the for-
ward link in COST-207 “Typical Urban” channel; Eb /No = 24 dB; perfor-
mance of adaptive receivers with “ideal” and differentially coherent reception,
wide BW Costas loop and N = 5 modified 1st-order ML structure; 3-finger
RAKE also shown; freq. uncertainty, 2σf = 1/40Rb ; vehicle speed is 45 kph. 142
8.16 Simulation results for bit error probabilites vs. number of users on the reverse
link in COST-207 “Typical Rural” channel; Eb /No = 24 dB; “strict” power
control, σP2 = 2 dB2 ; performance of conventional receivers with “ideal” and
differentially coherent reception, wide BW Costas loop and N = 5 modified
1st-order ML structure; 3-finger RAKE also shown; freq. uncertainty, 2σf =
1/40Rb ; vehicle speed is 45 kph. . . . . . . . . . . . . . . . . . . . . . . . . . 143
8.17 Simulation results for bit error probabilites vs. number of users on the for-
ward link in COST-207 “Typical Rural” channel; Eb /No = 24 dB; “strict”
power control, σP2 = 2 dB2 ; performance of adaptive receivers with “ideal”
and differentially coherent reception, wide BW Costas loop and N = 5 mod-
ified 1st-order ML structure; 3-finger RAKE also shown; freq. uncertainty,
2σf = 1/40Rb ; vehicle speed is 45 kph. . . . . . . . . . . . . . . . . . . . . . 143

xix
8.18 Simulation results for bit error probabilites vs. number of users on the for-
ward link in 2-ray Rayleigh channel; 1000 bits allowed for training & synch.;
Eb /No = 28 dB; performance of adaptive receivers with “ideal” and differen-
tially coherent reception, wide BW Costas loop and N = 5 modified 1st-order
ML structure; 3-finger RAKE also shown; freq. uncertainty, 2σf = 1/40Rb ;
vehicle speed is 45 kph. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
8.19 Simulation results for bit error probabilites vs. number of users on the re-
verse link in 2-ray Rayleigh channel; 1000 bits allowed for training & synch.;
“strict” power control; Eb /No = 28 dB; performance of adaptive receivers
with “ideal” and differentially coherent reception, wide BW Costas loop and
N = 5 modified 1st-order ML structure; 3-finger RAKE also shown; freq.
uncertainty, 2σf = 1/40Rb ; vehicle speed is 45 kph. . . . . . . . . . . . . . . 145
8.20 Simulation results for bit error probabilites vs. number of users on the for-
ward link in 2-ray Rayleigh channel; lower frequency offset assumed, 2σf =
1/400Rb ; Eb /No = 28 dB; performance of adaptive receivers with “ideal” and
differentially coherent reception, wide BW Costas loop and N = 5 modified
1st-order ML structure; 3-finger RAKE also shown; vehicle speed is 45 kph. 147
8.21 Simulation results for bit error probabilites vs. number of users on the re-
verse link in 2-ray Rayleigh channel; lower frequency offset assumed, 2σf =
1/400Rb ; “strict” power control; Eb /No = 28 dB; performance of adaptive
receivers with “ideal” and differentially coherent reception, wide BW Costas
loop and N = 5 modified 1st-order ML structure; 3-finger RAKE also shown;
vehicle speed is 45 kph. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147

xx
Chapter 1

Introduction

1.1 Motivations for Research

Nearly three years ago, members of the Mobile and Portable Radio Research Group (MPRG)
began investigating adaptive interference suppression for CDMA systems, sponsored by the
Defense Advanced Research Projects Agency’s (DARPA’s) Global Mobile Information Sys-
tems (GloMo)∗ Intiative. Early results from this research suggested dramatic performance
enhancements and, thus, corresponding capacity increases for certain cellular systems.
These improvements resulted from the ability of a fractionally-spaced adaptive equalizer
to both mitigate code division multiple access (CDMA) interference that shared spectral
properties with the signal of interest (SOI) and coherently combine multiple paths of the
SOI when fractional sampling allows sufficient time resolution for the various paths.
As is common with preliminary investigations, however, certain simplifying and neces-
sary assumptions were made concerning the adaptive receivers under investigation. One
assumption was the availability of an ideal coherent reference carrier for the SOI. This as-
sumption is somewhat realistic when a pilot tone is transmitted on the forward link of a
CDMA system, such as in the IS-95 U.S. digital CDMA standard. On the reverse link of
this system and in many other potential multiple access strategies planned for the Indus-
trial, Scientific, and Medical (ISM) bands in the U.S., however, the existence of pilot tone
is either impossible or not planned.
We therefore began to examine some adaptive receiver structures in the presence of
frequency offsets. The immediate result was catastrophic degradation and failure. This
implied a fundamental need for either carrier recovery or the redesigning of adaptive al-
gorithms to work with differentially coherent reception. The latter solution was quickly
∗
GloMo was known formerly as the Wireless Adaptive Mobile Information System (WAMIS) program.

1
CHAPTER 1. INTRODUCTION 2

discovered and is outlined in section 4.2.7 of this thesis. Though the performance degrada-
tion associated with moving from coherent to differentially coherent reception is well-known
for conventional receivers, it seemed to be amplified in adaptive receivers, as discussed in
section 7.1. This left us with a nagging question: How can we “buy back” some or all of this
performance degradation caused by differential detection? The question prompted others
and led to the research described in this thesis.

1.2 Thesis Overview

The layout of this thesis serves not only the thematic organization of the investigations
contained herein, but generally represents their chronological order as well. This chapter
begins to describe the challenges that motivated the work. The literature survey presented
in Chapter 2 was really the beginning of investigative process. Therein we searched a wide
range of historically noteworthy as well as more recent literature concerning synchronization,
concentrating on applications to suppressed-carrier recovery. Making sense of this literature
led to the work presented in Chapters 3 through 5. In these chapters, a theoretical substrate
is formed upon which the design of carrier recovery structures relevant to this particular
application, mobile CDMA environments, can be based. As such, these three chapters serve
very much as an introduction to the topic of carrier recovery.
Chapter 6 bridges the gap between the more theoretical observations of the previous
chapters and the results-oriented chapters that conclude this work. Therein, we describe
our simulation strategy, the assumptions that have been kept and those that have been
thrown out, and the details of wireless CDMA system under investigation. We use all
these details to propose candidate receiver structures with and without carrier recovery for
observation. The field is immediately narrowed by some preliminary simulation results.
This yields a field of “finalists,” including both conventional and adaptive receivers, for
further simulation in Chapters 7 and 8.
The additive white Gaussian noise (AWGN) channel is given significant attention in
Chapter 7. The relative performance of carrier recovery strategies is presented, always in
view of the performance of differential detection and ideal coherent reception. Substantive
remarks are made about the susceptibility of carrier recovery strategies to degradation due
to spread-spectrum multiple access interference (MAI). Chapter 8 moves beyond the AWGN
channel to investigate the performance of candidate architectures in more challenging con-
ditions faced in a mobile environment, viz., those of fading due to multipath. Flat fading
and a variety of frequency-selective environments are examined. A three-finger differen-
tially coherent RAKE receiver, similar to one used in IS-95, is also simulated to provide an
CHAPTER 1. INTRODUCTION 3

important basis for comparison. At all times in these chapters, both forward and reverse
cellular links are considered.
Finally, this presentation concludes with a review of the significant observations and
contributions of this research and numerous recommendations for future research. We
ultimately realize that this work has only scratched the surface of problem of carrier recovery
for CDMA, especially as it is linked to adaptive interference suppression. We therefore
intend for this effort to be seen as a solid introduction to this issue, and hope that it will
prove seminal for such future work.

1.3 Overview of PSK Reception

Before moving on to the chronicle the research effort in the chapters that follow, it is useful
to briefly give a high-level view of some important concepts and notation that will be called
upon throughout this work. To this end, we give a brief summary of digital phase shift
keying (PSK) modulation strategies in section 1.3.1. We then define coherent reception in
section 1.3.2, and further explain the notion of differential detection for PSK. Section 1.3.3
concludes this chapter with a brief introduction to the notions of closed- and open-loop
phase compensation. This serves as preview for much of the theoretical and descriptive
work to follow.

1.3.1 PSK Signals

Receiver architectures for digital signals either require or benefit from accurate estimation
of a variety of signal parameters. Among these are amplitude, phase, frequency, and delay.
In land mobile and cellular radio, the use of phase shift keying (PSK) has become a very
popular method for digital signaling because it represents a well-balanced tradeoff between
energy and bandwidth efficiency for the given mobile environment. Receivers for digital
PSK expect to receive a signal of constant magnitude. This simplifies the analysis of the
signal and the design of appropriate receiver structures. The information contained in a
received PSK signal is, by definition, embedded either in the absolute or relative phase of
the received signal. Thus the received signal may be represented by
√
s(t) = 2P · cos (2πfc t + φ(t) + θ(t)) + n(t) (1.1)

where P is the received signal power, fc is the received carrier frequency in Hz, n(t) is, for
now, a random bandpass process, θ(t) is the possibly time-varying phase of the carrier, and
 
2πm
φ(t) ∈ , m = 1, 2, . . . , M − 1
M
CHAPTER 1. INTRODUCTION 4

is the information contained in the signal where M is the number of points in the signal
constellation. The desire to correctly estimate φ(t) motivates the design of the digital
receiver.
Consider next the quadrature downconversion of this signal by a local reference carrier
generated by the receiver as outlined Fig. 1.1, viz.,
   
vd (t) = cos 2πfc t + θ̂(t) and vq (t) = −sin 2πfc t + θ̂(t)

where we have assumed that the instantaneous frequency of the VCO exactly matches the
incoming carrier, and θ̂(t) is an estimate of its phase.† Allowing the dependence of θ̂ on
time allows this simplification without loss of generality. This yields

i(t) = s(t) · vd (t)

s
P h    i
= cos φ(t) + θ(t) − θ̂(t) + cos 4πfc t + φ(t) + θ(t) + θ̂(t)
2
+nd(t) (1.2)

and

q(t) = s(t) · vq (t)

s
P h    i
= sin φ(t) + θ(t) − θ̂(t) − sin 4πfc t + φ(t) + θ(t) + θ̂(t)
2
+nq (t) (1.3)

where nd and nq are, for now, unspecified noise processes. Assuming that the low pass
filters (LPFs) work well, the double frequency terms in (1.2) and (1.3) may be neglected.
Sampling at time, t = kT , then yields
s
P  
ik = · cos φk + θk − θ̂k + nd,k (1.4)
2
and s
P  
qk = · sin φk + θk − θ̂k + nq,k (1.5)
2
as input to the baseband detector where φk is still the transmitted symbol and nd,k and nq,k
are samples of the heretofore unspecified noise processes. It will be convenient to regard ik
and qk together as single complex-valued sample,
√
rk = ik + jqk = P · ej (φk +θk −θ̂k ) + nd,k + jnq,k (1.6)

in the chapters that follow.

†
Throughout the thesis, we consistently view the quadrature downconversion process as multiplication
by cos(·) and −sin(·) due to its natural correspondence with multiplication by exp (−j(·)).
CHAPTER 1. INTRODUCTION 5

ik
x LPF

(
cos 2π f c t + θ$ ) rk Baseband
s(t)
− sin (2π f c t + θ$ )
Detector

qk
x LPF

Figure 1.1: Basic detection structure for PSK signals

1.3.2 Coherent Reception

What are the challenges to obtaining an accurate estimate of θ̂k ≈ θ(kT ), or equivalently,
coherent carrier recovery? There are two broad categories. First of all, θ may be time-
varying for any number of reasons. Residual carrier offset introduces a constant rotation in
the phase parameter that can be a significant fraction of the baud rate in practical systems.
Doppler spread has the same effect as residual offset, but may vary with time, creating
non-zero second and higher order derivatives of the instantaneous phase. Fading due to
multipath further complicates matters by introducing potentially wild variations in carrier
phase over just a few symbols and providing multiple carriers with respectively arbitrary
phases to the synchronization system.
The other major challenge to obtaining θ̂ comes from the noise terms nd and nq . Portions
of these noise terms come from realizations of an additive white Gaussian noise (AWGN)
channel. AWGN introduces random perturbations in the phase and amplitude of rk , that
obviously will hinder the prediction of the carrier phase and thus the transmitted symbol, φ.
Still AWGN is zero mean, uncorrelated, and stationary, the effects of which might somehow
be averaged out. More pathological effects may be observed when co-channel and adjacent-
channel interference are taken into account. These can make the statistics of the noise
non-stationary and quite difficult to predict.
Still there are certain advantages to coherent reception. First, many equalization strate-
gies require coherent carrier recovery. As equalization strategies become more popular in
mobile radio, the motivation to coherently detect PSK signals increases. In addition, coher-
ent reception of PSK provides a 1 to 3 dB power advantage versus noncoherent techniques.
Coherent detection in a mobile channel is extremely challenging. Differentially coherent
phase shift keying (DPSK) simplifies the challenge of phase estimation by using the previous
CHAPTER 1. INTRODUCTION 6

s(t) e(t) Loop

X
Filter
v(t)
VCO

Figure 1.2: Example of a feedback control system: A simple PLL model

phase estimate as a reference for the current one, viz.,

rk · rk−1
?
= P · ej(φk −φk−1 +θk −θk−1 ) + noise . (1.7)

This is based on the generally safe assumption that θk changes very little between successive
symbols, and therefore θk − θk−1 is very close to zero. Note that the transmitted symbol
must first be differentially encoded at the transmitter by ∆φ = φk − φk−1 , the difference
between the current information symbols and the previous one. However, because the
receiver is using a noisy reference, i.e., the previous estimate, M -ary DPSK (MDPSK)
suffers in performance versus coherent M -ary PSK (MPSK). This degradation can be up
to 3 dB for large M but is only about 1 dB for M = 2. Consequently, there is now a great
deal of interest in schemes to perform coherent detection of DPSK. Coherently detected
MDPSK differs only by about a factor of two in BER (for most Eb /No ratios of interest)
from MPSK. Such schemes, of course, lead us back to all the same challenges of estimating
instantaneous carrier phase mentioned above. It is, therefore, one purpose of this work to
present a wide overview of techniques for the estimation and/or compensation of carrier
phase and frequency offset in PSK reception.

1.3.3 Closed Vs. Open Loop Compensation

Phase and frequency synchronization techniques can be broadly categorized as either closed-
loop or open-loop schemes. Closed-loop schemes involve a feedback path in the system to
make corrections for errors. The basic design of such a system is illustrated by the ubiquitous
PLL detailed in Fig. 1.2. The PLL is a very basic feedback control system consisting of
a phase detector (PD), a loop filter, and a voltage controlled oscillator (VCO). This basic
configuration is similar to that of an automatic gain control (ACG) loop or an automatic
frequency control (AFC) loop. The shared feature of all such control schemes is that an
error signal is fed back to the system input to drive the error term to zero at steady-state.
CHAPTER 1. INTRODUCTION 7

s(t) Matched rk rk‘ Baseband

Filter
x Detector
$
e − jθ ML
ML Phase conjugate
Estimator

Figure 1.3: Example open-loop compensation scheme: ML phase estimator

In contrast to closed-loop, open-loop schemes do not contain a feedback path. Open-

loop parameter estimation is classically derived from either maximum likelihood (ML) or
maximum a posteriori (MAP) decision rules. These, in the case of no prior knowledge of the
input parameter vector, are shown to be equivalent in [1]. Figure 1.3 provides an example
of an open-loop phase estimation scheme derived from the ML principle. Note that the
removal of phase occurs at the output of the loop rather than at the input.
Chapters 3 through 5 provide an overview of both closed-loop and open-loop phase and
frequency compensation schemes. First, the PLL is described and analyzed. Performance
measures are described, and the discrete-time implementation of second and third-order
PLLs is derived. Then, in Chapter 4, we apply the concepts of the PLL to the task of
carrier recovery, and review a variety of techniques for this task, and their performance.
Also in Chapter 4, we introduce several open-loop phase estimation structures based on
the ML principle. Both sets of structures are applied to the task of detection of PSK
waveforms. Then in Chapter 5, coverage of AFC and its application to carrier recovery
issues are presented.
Chapter 2

A Brief Survey of Prior Research

Before moving on to detailed descriptions of carrier recovery techniques and structures

presented in Chapters 3 through 5, a brief review of pertinent literature concerning this
problem is in order. In this chapter we review the literature and thus the history concern-
ing this far-reaching topic. Closed-loop synchronization, addressed in section 2.1, has a
very diverse history reflected in its seminal papers over the last seventy years. Open-loop
estimation in section 2.2, while having a rich literature as well, has only more recently been
applied to the task of carrier recovery. Insofar as they are based upon newer concepts,
open-loop techniques therefore get somewhat less attention in this chapter—a deficiency
that will hopefully be made up in later chapters of this work.

2.1 Closed-Loop Synchronization

The subject of classical closed-loop phase and frequency synchronization is both rich and di-
verse. The first description of the phase-locked loop (PLL) was by English physicist Edward
Appleton, appearing in the Proceedings of the Cambridge Philosophical Society in 1923.[2]
The subject remained in relative obscurity until just before World War II when the PLL
began to find its way into the advancing communication technologies of the day, specifically
frequency modulation (FM). The 1950s saw the recommendations of the National Television
Standard Committee (NTSC) for commercial broadcast television as well as an increased
interest in suppressed-carrier techniques for wireless communications. These spawned a rich
period of study in phase and frequency lock techniques. Gruen’s 1953 paper [3] was one of
the first published works to deal specifically with the subject of automatic frequency control
(AFC). This was in an application of retrieving the color subcarrier in the NTSC standard.
Costas [4] proposed a PLL capable of demodulating a suppressed-carrier AM signal. At

8
CHAPTER 2. A BRIEF SURVEY OF PRIOR RESEARCH 9

the time, the double side-band suppressed-carrier (DSB-SC) modulation was presented as
a favorable alternative to single side-band modulation, attracting a great deal of interest
in military applications where low probability of jam and/or intercept was crucial. The
Costas loop and a whole family of related structures are now standard detection schemes
for suppressed-carrier, coherent PSK.
Gupta’s tutorial paper [2] contains 188 references that survey the history of PLLs up to
1975. The paper also provides a concise survey of a wide variety of PLLs and related topics.
In addition to Gupta’s survey, Blanchard [5] and Gardner[6] are standard textbooks on
the subject, that should prove adequate for all but the most serious students of phase-lock
phenomena. Lindsey[7] and Lindsey & Simon [8] are described by Gardner as “massive
assaults” on the difficult problems associated with PLLs. Taken together these works,
without doubt, constitute the most comprehensive study of the subject ever undertaken. In
1966, Viterbi [9] contributed the monumental undertaking of an exact non-linear solution to
the first-order PLL. References [10] and [11] are edited collections that are a rich resource for
seminal papers on the performance and design of PLLs and AFC loops for a large number
of applications.
Over the years, the study of phase and frequency loops has diverged into a number
of separate areas. References [2], [10], and [11] do a good job in pointing the student
or designer interested in these different areas to suitable publications ranging from the
introductory to the advanced. We make a few additional recommendations here, as well.
A good introductory treatment of the subject of digital PLLs may be found in [12]. For
further study, references [13], [14], and [15] should prove helpful.
Application of PLLs to symbol timing recovery and tracking is best summarized in
Franks classic tutorial [16]. J. J. Spilker in [17] and [18] was one of the earliest writers on
the subject, proposing the delay-lock discriminator for binary signals. Early treatments of
carrier tracking and Costas loops may be found in [19] - [25].
Automatic frequency control (AFC) loops, though based on the same error feedback
principle, differ in a strict sense from PLLs. As will be shown, the latter perform phase-
coherent tracking by feeding back an error as a function of the phase difference between
the signal of interest (SOI) to the reference. AFC loops, strictly speaking, are non-coherent
devices where the feedback error depends on the frequency difference. The subject of AFC,
though well-studied, is not nearly so widely published as that of phase-locked loops. Two
reasons for this are:

1. The design and performance AFC loops are a natural extension of well-established
CHAPTER 2. A BRIEF SURVEY OF PRIOR RESEARCH 10

PLL theory, as we show in Chapter 5. If one can build a PLL, then one can prob-
ably build an AFC loop.
2. AFC loops are not nearly as ubiquitous as PLLs. The latter are found in nearly
every conceivable application, from industrial control to radio astronomy. In
contrast, the use of AFC loops is generally limited to RF communications and
ranging applications.

Two references, in particular, are noteworthy in the study of AFC. Natali’s 1984 seminal
paper [26] proposes a number of practical structures for AFC implementation, all but one of
which is digital. Gardner in [27] provides an excellent tutorial on the subject, concentrating
on the structure of frequency difference detectors (FDDs). AFC is often seen as an aid
to frequency acquisition of a PLL. Two early works in this area are from Cahn [28] and
Messershmitt [29]. Later works of interest include [30]- [32].

2.2 Open-Loop Schemes

Open-loop phase and frequency synchronization schemes are based on either the maximum
a priori (MAP) or maximum likelihood (ML) parameter estimation principles. As such,
these are generally feedforward techniques in that the signal parameter is actively estimated
and then treated as a statistic for a detection algorithm. This is a completely different ap-
proach from classical closed-loop schemes, which, as noted above, use error feedback control
to minimize the error in the received signal. Before the days of high-performance digital
computers and signal processors, the idea of open-loop phase and frequency compensation
was primarily an academic one, finding little application in the design of real-time commu-
nication systems. Interest in open-loop approaches has, however, increased tremendously
as it has become more and more feasible in recent years to implement them.
Since these techniques are relatively new in their application to real communication
systems, their literature is scattered and not well-organized. One commonality in many
papers on open-loop estimation is the citation of Van Trees’ [33] - [35] classic texts. Meyr
and Ascheid in [36] devote several chapters to open-loop phase and frequency estimation,
in addition to providing a comprehensive treatment of classical loops. Proakis’ [1] Chapter
6 provides one of the best introductions to the topic. Feedforward estimation techniques
embedded within practical receivers have been suggested by Simon and Divsalar in the
particularly useful papers [37] - [41]. Recent, noteworthy references on the subject also
include [42] - [46]. It is important to note here that likelihood rules for estimation do
not necessarily imply open-loop detection structures. In an unusual and interesting paper,
CHAPTER 2. A BRIEF SURVEY OF PRIOR RESEARCH 11

Simon and others, in [47], rigorously show that closed-loop recovery schemes may be derived
from the likelihood functions of estimation theory.
Chapter 3

The Phase-Locked Loop

Feedback phase and frequency control loops consist of three main components: 1) an error
detector of some sort, which for the PLL is a PD, modeled in Figure 3.1 as a multiplier, 2) a
loop filter, and 3) a VCO. Note that for an AFC loop the error term is given by a frequency
difference detector (FDD). PLLs can either be fully analog circuits, partially analog and
partially digital, or fully digital. Moreover, PLLs find many uses in communications. They
can be used for frequency demodulation, providing superior performance over a conventional
discriminator circuit. They are also used in de-noising oscillators, in frequency multipliers
and dividers, in frequency synthesizers, and in synchronizing digital signals. We are mainly
interested in the PLL for the purpose of carrier recovery, a simple description of which is
as follows.

3.1 Introductory View

Consider the situation portrayed in Figure 3.1 where an input signal,

s(t) = cos(ωc t + θ) (3.1)

where ωc is the radian frequency of the carrier and θ is its arbitrary phase, is applied to the
input of the receiver and multiplied by the local reference,

v(t) = − sin(ωc t + θ̂) . (3.2)

We may assume, without loss of generality, that the input and reference frequencies are
identical if we consider θ̂, an estimate of θ, to be a time-varying parameter. The output of
the PD is an error signal given by

e(t) = s(t) · v(t)

12
CHAPTER 3. THE PHASE-LOCKED LOOP 13

s(t) e(t) Loop

X
Filter
v(t)
VCO

Figure 3.1: Simple illustrative PLL model

= − cos(ωc t + θ) · sin(ωc t + θ̂)

1 1
= − sin(2ωc t + θ + θ̂) + sin(θ − θ̂) . (3.3)
2 2
In practice, the double frequency term will lead to a ripple in the error control signal, which
can be a problem in practical systems and compensated for with a lowpass filter (LPF).
The loop filter is, itself, a LPF with a very narrow bandwidth. Therefore, for the purposes
of introduction, we neglect the double frequency term and simplify (3.3) as
1
e(t) = sin(θ − θ̂). (3.4)
2
The loop filter, F (s), for the moment, is simply some sort of lowpass filter for the error
signal. The design of this filter, the specification of its order and related parameters, will
require the most attention as the discussion of PLL design progresses.

3.2 Linearized Analog PLL Model

Assuming that after some convergence period the error term, e(t), will become small, the
 
approximation sin θ − θ̂ ≈ θ − θ̂ may be applied. This yields the equivalent phase domain
model of Figure 3.2 where the VCO is modeled as a perfect integrator with gain Ko . The gain
factor, Kd , is that of the detector.∗ Linearization allows the use of the Laplace transform,
Z ∞
4
X(s) = L {x(t)} = x(t)e−st dt . (3.5)
0

Thus the transform of the input phase, L{θ(t)} is denoted by Θ(s). Similarly, L{θ̂(t)} =
Θ̂(s), and, for the phase error, L{φ(t)} = Φ(s). We have then, from linear system theory,
that the closed-loop transfer function, i.e., the relationship of the input phase to the output
phase estimate, may be given by
. Θ̂(s) Kd Ko F (s)
H(s) = = . (3.6)
Θ(s) s + Kd Ko F (s)

∗
For a more thorough explanation of detector gain, see section 3.4.1.
CHAPTER 3. THE PHASE-LOCKED LOOP 14

Θ(s) + Φ(s)
Σ Kd F (s)
-
∧
Θ(s) Ko
s
Figure 3.2: Linearized, phase domain model of PLL

3.2.1 The Second-Order PLL

The order of a PLL, based upon servo terminology, is defined by the number of “perfect
integrators,” n, in the loop. An integrator defined by 1/s is merely a pole placed at s = 0, or
DC. Since the VCO represents one such pole, n is one more than the number of integrators
in F (s). Thus, an nth-order PLL may be constructed using an (n − 1)th-order loop filter.
A first-order RC filter built with a high gain DC amplifier (as shown in Figure 3.3) is of the
most interest here since it meets this criterion. This will lead to the very well-known and
widely implemented second-order PLL. Such a filter has a transfer function,
sτ2 + 1
F (s) = (3.7)
sτ1
where τ1 and τ2 are the time constants of the network, defined in Figure 3.3. Letting
Kd Ko 2ζ
τ1 = 2
and τ2 = (3.8)
ωn ωn

we obtain
2ζωn s + ωn2
H(s) = (3.9)
s2 + 2ζωn s + ωn2
where ωn and ζ are the well-known parameters of natural frequency in [rad/s] and damping
factor, respectively. Equation (3.9) shall be called the canonical, continuous-time, second-
order PLL model.
Use of this particular second-order model is advantageous for a couple of reasons. Since
it a true second-order loop† , it is applicable to a wide range of applications including carrier
recovery. Furthermore, specifying ζ = √1 is shown in [6] to produce the optimal second-
2
order PLL as discussed in section 3.6.
†
A second-order loop built with a RC lag filter would technically be second-order, but Gardner [6] shows
that due to a lack of high DC gain, it should be viewed as a modified first-order PLL.
CHAPTER 3. THE PHASE-LOCKED LOOP 15

R2 C

R1
e(t) + -
+
+ ef (t)
- -
τ1 = R1C τ2 = R2C
Figure 3.3: High gain, active, lowpass filter suitable to realize loop filter in canonical PLL
model

The simplest of all possible PLLs, one of first-order, may be constructed with a loop
filter, F (s), equal to a constant. While such a PLL does have many applications, it is
of no value for the task of carrier recovery since it compensates poorly in the presence of
frequency offset and fails utterly to compensate in the presense of a frequency ramp. A
second-order loop filter will lead to a third-order PLL that will prove useful for the task of
carrier recovery.

3.2.2 The Third-Order PLL

Though not as commonly used as the second-order, the third-order PLL does have some
advantages. Most importantly, it is capable of tracking a frequency ramp with zero steady-
state error. The second-order PLL can only do so with a finite steady-state error.
At the heart of the third-order PLL is the second-order loop filter specified by

s2 + a2 s + a3
F (s) = (3.10)
s2
where a2 and a3 are design parameters. Note that the loop filter now has two poles at DC
in addition to that obtained from the VCO. Substituting (3.10) into (3.6), we find


Ko Kd s2 + a2 s + a3
H(s) = 3 . (3.11)
s + Ko Kd (s2 + a2 s + a3 )
For the sake of simplicity, it is convenient to define the various parameters in terms of a
corner frequency, ω3 . Letting Ko Kd = 2ω3 , a2 = ω3 , and a3 = 12 ω32 , we obtain

ω33 + 2ω32 s + 2ω3 s2

H(s) = . (3.12)
ω33 + 2ω32 s + 2ω3 s2 + s3
CHAPTER 3. THE PHASE-LOCKED LOOP 16

Equation (3.12) shall be denoted as the canonical, continuous-time, third-order PLL model.
This loop is shown in [6] to be optimal according to the Wiener criteria discussed in sec-
tion 3.6. Having to specify only the single parameter ω3 simplifies the design process and
streamlines further discussion of the third-order PLL.
The particular loops defined by (3.9) and (3.12) are well-known and well-analyzed in
the classic PLL treatments of [5], [6], and [36]. Due to these and related treatments,
theoretical results for the performance of these PLLs, the steady-state variance of output
phase estimate as well as cycle slipping characteristics, are well known. As such, these will
serve as baselines for the simulations presented later in this work. Finally, we choose to
implement these classical loops because the design parameters, ωn and ζ (which is usually
chosen to be √1 ) for the second-order, and ω3 for the third-order, have simple, intuitively
2
understandable meanings. This simplifies the design process while preserving the well-
known properties of these loops. We will examine these performance characteristics in
section 3.4, but first proceed with a derivation of the discrete-time PLL that will be useful
for computer simulation and, ultimately, for integration in a fully digital receiver.

3.3 Derivation of Discrete-time PLL Models

3.3.1 Discrete-time Second-Order PLL

We begin by considering the canonical continuous-time transfer function given by (3.9).

Note that the poles of H(s) are very near zero, i.e., ωn  2π/T , enabling a straightforward
application of the bilinear z-transform (BZT) without frequency pre-warping. This allows
a simple substitution for s. Let
2 1 − z −1
s= · (3.13)
T 1 + z −1
where T is the sampling interval in seconds. Substituting (3.13) into (3.9) leads to the
following expression for H(z):

Θ̂(z) b0 + b1 z −1 + b2 z −2
H(z) = = (3.14)
Θ(z) a0 + a1 z −1 + a2 z −2

where

b0 = (ωn T )2 + 4ζωn T a0 = 4 (ωn T )2 + ζωn T + 4

b1 = 2 (ωn T )2 a1 = 2 (ωn T )2 − 8
b2 = (ωn T )2 − 4ζωn T a2 = (ωn T )2 − 4ζωn T + 4
CHAPTER 3. THE PHASE-LOCKED LOOP 17

Figure 3.4: Discrete-time, phase domain model of PLL

which we will denote as the canonical discrete-time transfer function. It is expected that
any discrete-time implementation of this transfer function would match the performance of
the canonical continuous-time model.
Discrete-time, digital implementation of the PLL requires adjustment in the modeling
of the loop components. Note that the VCO in Figure 3.2 has been replaced in Figure 3.4
with a numerically controlled oscillator (NCO), N (z). Whereas the VCO was modeled as
a perfect integrator, the NCO is modeled as a perfect accumulator. Thus, the NCO is now
modeled by the transfer function
Ko
N (z) = (3.15)
1 − z −1
where Ko is the output voltage amplitude of the device. This is an accurate model for many
off-the-shelf NCOs currently available. Note also the addition of the delay in the feedback
path. This is needed to make the system causal.
Given these modifications, we now wish to find a model for the loop filter, F (z)‡ , that
will preserve the overall transfer function of (3.14). From linear system theory, we know
that
Kd F (z)N (z) H(z)
H(z) = =⇒ F (z) =
1 + Kd F (z)N (z)z −1 Kd N (z) [1 − H(z)z −1 ]


1 H(z) 1 − z −1
= · . (3.16)
Kd Ko 1 − H(z)z −1

Solving directly for F (z) leads to a third-order loop filter, which is somewhat surprising
since the overall transfer function is only of order two and the continuous-time loop filter is
of order one. It turns out that we can obtain the expected performance of the second-order
discrete-time PLL using only a first-order loop filter.

‡
Note F (z) will not, in general, be the BZT of F (s).
CHAPTER 3. THE PHASE-LOCKED LOOP 18

To create this filter we specify a loop filter, F2 (z)§ , and constrain it to be of first-order,
as is its continuous-time cousin, F (s), with a pole at z = 1, i.e., at DC. For convenience,
we specify the filter by
g0 + g1 z −1
F2 (z) = . (3.17)
1 − z −1
Then H2 (z) is given by
 
Ko z −1
Kd F2 (z) · 1−z −1
H2 (z) =  
Ko z −1
1 + Kd F2 (z) · 1−z −1


Kd Ko g0 z −1 + g1 z −2
= . (3.18)
1+ (Kd Ko g0 − 2) z −1 + (Kd Ko g1 + 1) z −2
Note that H2 (z) cannot, in general, be made to equal the canonical H(z) given in (3.14).
Instead, following the reasoning of [48], we merely equate the denominator of (3.18) with
the normalized denominator of (3.14), i.e., let
a1 a2
Kd Ko g0 − 2 = and Kd Ko g1 + 1 = .
a0 a0
After some algebra, the forward coefficients g0 and g1 in (3.17) are found to be
   
1 a1 1 a2
g0 = +2 and g1 = −1 (3.19)
Kd Ko a0 Kd Ko a0
where the ai s are as defined in 3.14. Thus, the design of the discrete-time second-order PLL
is complete.

3.3.2 Discrete-time Third-Order PLL

For the third-order PLL, we follow exactly the same procedure as for the second-order.
Substituting (3.13) into (3.12), we obtain

Θ̂(z) d0 + d1 z −1 + d2 z −2 + d3 z −3
H(z) = = (3.20)
Θ(z) c0 + c1 z −1 + c2 z −2 + c3 z −3
where

d0 = (ω3 T )3 + 4 (ω3 T )2 + 8ω3 T c0 = (ω3 T )3 + 4 (ω3 T )2 + 8ω3 T + 8

d1 = 3 (ω3 T )3 + 4 (ω3 T )2 − 8ω3 T c1 = 3 (ω3 T )3 + 4 (ω3 T )2 − 8ω3 T − 24
d2 = 3 (ω3 T )3 − 4 (ω3 T )2 − 8ω3 T c2 = 3 (ω3 T )3 − 4 (ω3 T )2 − 8ω3 T + 24
d3 = (ω3 T )3 − 4 (ω3 T )2 + 8ω3 T c3 = (ω3 T )3 − 4 (ω3 T )2 + 8ω3 T − 8 .

§
Use of the subscript 2 , is not intended to cause confusion. It is chosen because it leads to a second-order
discrete-time PLL. Similarly, the subscript 3 is chosen later for the loop filter in a third-order PLL.
CHAPTER 3. THE PHASE-LOCKED LOOP 19

But similar to the loop filter for the second-order PLL, solving directly for F (z) results in
a fourth-order expression, two orders higher than its continuous-time cousin.¶ Rather, we
specify a second-order loop filter
h0 + h1 z −1 + h2 z −2
F3 (z) = . (3.21)
(1 − z −1 )2
Then H3 (z) is given by
 
Ko z −1
Kd F3 (z) · 1−z −1
H3 (z) =  
Ko z −1
1 + Kd F3 (z) · 1−z −1


Kd Ko h0 z −1 + h1 z −2 + h2 z −3
= . (3.22)
1 + (Kd Ko h0 − 3) z −1 + (Kd Ko h1 + 3) z −2 + (Kd Ko h2 − 1) z −3

Note that, as with the second-order model, H3 (z) cannot in general be made to equal the
H(z) of (3.20). Instead, we simply equate the denominators of (3.20) and (3.22) and find
the forward coefficients of the loop filter to be
     
1 c1 1 c2 1 c3
h0 = +3 , h1 = −3 and h2 = +1 (3.23)
Kd Ko c0 Kd Ko c0 Kd Ko c0
where the ci s are defined in 3.20. This completes the design of the discrete-time third-order
PLL.

The second and third-order PLLs may now be accurately modeled and simulated on
a digital computer, as well as implemented in real-time using commercially available inte-
grated circuits (ICs) or using software generated for a digital signal processor (DSP). The
motivations for completely digital and/or software implemented PLLs are manifold. One
advantage is the elimination of device parasitics, which cause unintended poles in the trans-
fer functions of PLLs with analog loop filters. Another advantage is the ability within a
digital filter to actually have a pole, or poles, at DC, providing a loop filter with an open-
loop gain of infinity. This is only approximated with the analog filter of Figure 3.3. Finally,
we maintain that it is actually simpler to build a PLL using fully digital components or in
software, which matches the performance of the canonical analog PLL described in section
3.2, than it is to build one using completely analog components. Digital components give the
designer much more precise control. An important next step to this research is to examine
the performance of these derived discrete-time models. Do they, indeed, give the perfor-
mance we expect? Before this all-important question is addressed, however, performance
measures for these PLLs must be established.
¶
Direct solution for this fourth-order filter has been observed in this research to cause instability due to
machine precision effects. This is yet another motivation to constrain the poles to lie precisely at DC.
CHAPTER 3. THE PHASE-LOCKED LOOP 20

3.4 Performance Measures for PLLs

Any synchronization structure employing the PLL operates in one of two modes: 1) ac-
quisition mode or 2) tracking mode. With each of these modes, we wish to specify a set
of performance measures that will aid in the design of the loop components for a specific
application. In acquisition mode, the most important measure is acquistion time, tacq , how
long it takes for the error signal, e(t), or, in the presence of severe noise, some short term av-
eraged function thereof, to settle below some predefined limit. The parameter tacq turns out
to be a random variable that can be characterized by some probability distribution function,
Pr[tacq < t]. This is an extremely important question, but unfortunately no theory exists to
predict this function [49]. Some experimental results, however, can be obtained and will be
reported in following sections. Intuition relating acquisition time to loop parameters will
also be emphasized.
In tracking mode of operation, the most important questions involve the reliability and
accuracy of the reconstructed phase, θ̂. We are interested in the susceptibility of the loop
to temporarily loose lock. This phenomenon is known as cycle slipping. In addition, we
would like to know how the loop operates in a variety of signal environments. Specifically,
we examine the transient error response seen in e(t) due to a phase step, a phase ramp
(frequency step), and a frequency ramp. Finally, we would like to know if (and how) the
variance of θ̂, denoted output phase jitter, can be related to the input signal to noise ratio
(SNR) and the loop design parameters. We answer this last question first since it gives rise
to the very important notion of loop noise bandwidth.

3.4.1 Loop Noise Bandwidth

What is the effect of noise at the input of a PLL on the steady-state phase estimate available
at feedback? To answer this all-important question, it is convenient to define the notion of
loop bandwidth, BL .∗ Considering once again the simple model of Figure 1.2, we have as
one input to the multiplier, a sinusoidal signal of interest and bandpass Gaussian noise,
√
s(t) = 2P · cos (2πfc t + θ(t)) + n(t) (3.24)

and a pure reference signal,

 
v(t) = −Vo sin 2πfc t + θ̂(t) . (3.25)

∗
This derivation parallels Gardner’s in[6]. His sign and phase conventions, however, differ slightly from
those presented here.
CHAPTER 3. THE PHASE-LOCKED LOOP 21

It is permissible and convenient to express the stationary, bandpass noise as

n(t) = nd (t) cos (2πfc t) − nq (t) sin (2πfc t) . (3.26)

Assume that nd and nq are independent, identically distributed processes. In a manner

completely analogous to (3.3) and (3.4), we may write the multiplier output as
s
P   V n (t) Vo nd (t)
o q
e(t) = Vo sin θ(t) − θ̂(t) + cos θ̂(t) − sin θ̂(t) . (3.27)
2 2 2
p
Define detector gain as Kd = P/2 Vo and now, for notational convenience, assume an
implicit dependence on time for e, θ, θ̂, nd , and nq . Thus, we may express the error as

  
nq nd
e = Kd · sin θ − θ̂ + √ cos θ̂ − √ sin θ̂ . (3.28)
2P 2P
An additive noise equivalent for the linearized phase domain model of Figure 3.2 may now
be given by
nq nd
n0 = √ cos θ̂ − √ sin θ̂ (3.29)
2P 2P
which can be included in the summer. Thus,
h   i
e = Kd · sin θ − θ̂ + n0 (3.30)

where e is now considered a superposition of a signal term, Kd sin(θ − θ̂), and a noise term,
Kd n0 .
It can be assumed that, after some, heretofore unspecified, transient behavior, that a
steady-state will be reached in which θ̂ will have a constant mean. Moreover, we assume
that n0 is zero mean. The variance of n0 may thus be given by
1  2 2 
σn2 0 = nq cos θ̂ + n2d sin2 θ̂ − 2nq nd sin θ̂ cos θ̂ (3.31)
2P

where the overbar is the time averaging operator. Note that n2q = n2d = n2 = σn2 , nq nd = 0,
and cos2 θ̂ + sin2 θ̂ = 1. Thus † ,
σn2
σn2 0 = . (3.32)
2P
The power spectral density of n0 is shown in [6] to be
1
Φn0 (f ) = [Φn (fc − f ) + Φn (fc + f )] (3.33)
2P
†
Here, we are making an assumption that θ̂ is independent of nd and nq , which in general is clearly not
true. This assumption is used, however, in all known treatments and is justified in [6] for the (common)
situation wherein BL  Bi .
CHAPTER 3. THE PHASE-LOCKED LOOP 22

where Φn (f ) is the original single-sided spectral density of the bandpass noise. For the
special case of white noise where Φn (f ) = No , the spectrum becomes
No
Φn0 (f ) = (3.34)
P
where No is the single-sided noise spectral density in [W/Hz]. The variance of the output
phase estimate due to additive noise may thus be calculated using (3.33),
Z ∞
θ̂n2 = Φn0 (f )|H(jω)|2 df [rad2 ] (3.35)
0

which, in general, may or may not be tractable. Given white noise, however, this integral
reduces to
Z ∞
No
θ̂n2 = |H(jω)|2 df [rad2 ]
P 0
Z ∞
1
= |H(jω)|2 df [rad2 ] (3.36)
SNRi Bi 0

where ω = 2πf , SNRi is the input SNR, and Bi is the noise bandwidth of the signal applied
to the phase detector ‡ . This allows a convenient definition of noise bandwidth for the loop
as Z ∞ No BL BL
BL = |H(jω)|2 df [Hz] =⇒ θ̂n2 = = [rad2 ] . (3.37)
0 P SNRi Bi
For the canonical, second-order PLL defined in (3.9), this bandwidth can be shown to equal
 
1 1
BL,2 = ωn ζ + (3.38)
2 4ζ
1 √
which, with ζ = √ , implies BL,2 = ωn 3 8 2 . This value of BL is used to give the theoretical
2
values of phase variance plotted as the lines in Figure 3.13. For the third-order PLL defined
by (3.11), this noise bandwidth is given by

1 a2 Ko Kd + a22 − a3
BL,3 = Ko Kd (3.39)
4 a2 Ko Kd − a3
6
which, when we define the parameters as in (3.12), simplifies to BL,3 = ω3 . This value of
5
BL is that from which the theoretical values of phase variance for third-order PLL plotted
as the lines in Figure 3.14.
Loop bandwidth for the canonical second-order loop is shown in Figure 3.5, normalized
to the choice of ωn as a function of ζ. Note that it is a minimum at ζ = 0.5 but also that
it is nearly minimum in a fairly broad range around 0.5. It turns out that ζ = √1
2
≈ 0.707
will be optimal in terms of acquisition performance and is thus the value most often used in
CHAPTER 3. THE PHASE-LOCKED LOOP 23

3.5

Normalized Loop Bandwidth, BL / ωn

3

2.5

1.5

0.5

0
0 0.5 1 1.5 2 2.5 3 3.5 4

Loop Damping Factor, ζ

Figure 3.5: Normalized loop bandwidth, BL ωn plotted as a funtion of ζ, for canonical

second-order loop, H(s)

practice. This corresponds to the critically damped state of the loop according to classical
control theory.
Given an expression for loop bandwidth, it follows that we may define the notion of loop
signal-to-noise ratio, SNRL .§ The phase jitter of the input signal may be defined by
1
θn2 = [rad2 ] (3.40)
2SNRi
1
because 2 of the noise power is orthogonal to the signal phase. By analogy then, we define
a relationship between output phase jitter and loop SNR by
2 1
θˆn = [rad2 ]. (3.41)
2SNRL
Using (3.37), we define
SNRi Bi
SNRL = (3.42)
2BL
where SNRi and Bi are as defined in (3.36), and BL is defined by (3.37). Thus, a notion
of “gain” may be attributed to the PLL as 10log10 (Bi /BL ) − 3 [dB]. The PLL “cleans
up” the input signal by this amount in providing a locally generated reference. Note that
BL could be made arbitrarily low to allow the PLL to function in an arbitrarily low SNR

‡
For all simulations presented in this work, Bi is taken to be the sample rate, Fs , or, in the case of digital
signalling with matched filters, the baud rate, Rb .
§
There is actually no “signal” in the PLL. This definition is ultimately arbitrary and simply a helpful
measure. Note that the definition in (3.42) is consistent with that used in [6] but is 3 dB lower than the
SNRL expression used in [5], [9], and [36].
CHAPTER 3. THE PHASE-LOCKED LOOP 24

environment. Doing so, however, is at the risk of a potentially long acquisition period that
will be discussed in section 3.4.4.

3.4.2 Transient Responses

We examine the transient behavior of the phase-locked loop for three different inputs: 1)
a phase step, 2) a phase ramp or equivalently, a frequency step, and 3) a frequency ramp.
The attempt is to give a qualitative, versus quantitative, view of loop performance. Note
that these transient behaviors are applicable to the tracking behavior of the PLL whenever
these inputs are encountered after the PLL has reached a steady-state. In other words,
the PLL will essentially re-acquire phase-lock with the measures of performance given. For
brevity, we devote primary coverage to the transient performance of the second-order loop.
All of the transient responses shown are simulation results for H2 (z), defined by (3.18) with
parameters defined by (3.19). Theoretical results in zero noise are available in [6] to which
the following experimental examples match well. For the sake of brevity we choose not to
present these complicated results and refer the interested reader to this excellent reference
for these expressions.
Figure 3.6 shows transient results for H2 (z) due to a phase step of π/2 radians, in the
absense of noise, for a variety of loop damping factors. Time is normalized with ωn to
provide a generalized view of the effect of ζ as a parameter. In part A, the phase error
term, sin φ = sin(θ − θ̂) where the sin(·) functions disappear at steady-state. Note that
the error oscillates for low ζ. For high ζ, the appearance of fast convergence is deceiving
√
since the error decays very slowly. The curve corresponding to ζ = 1/ 2 converges fastest
to zero steady-state error. After convergence, the first, second, and third-order PLLs all
operate with zero steady-state error in the presence of no noise.
Figure 3.6(B) shows the output phase estimate (denoted φo instead of the usual θ̂). The
behavior is very similar to that of the phase error except that here the values converge to
the actual phase value, π/2.
Figures 3.7(A) and 3.7(B) show the error response of the second-order loop due to fre-
√
quency steps of ωn and 2ωn radians respectively. In both cases, once again ζ = 1/ 2 proves
to be optimal in the sense of forcing the steady-state error to zero quickly. For frequency
steps larger than these examined, we would begin to observe a number of sinusoidal oscil-
lations during the transient response before the convergence behavior. The length of this
acquisition period is described in section 3.4.4. Note that the steady-state output phase in
this situation would not converge to a single value but rather to a ramp of slope equal to
the frequency offset.
CHAPTER 3. THE PHASE-LOCKED LOOP 25

1
ζ = 3.0

Phase error, sin (φe) [rad]

ζ = 1.0
0.5

ζ = 0.707
-0.5 ζ = 0.5
ζ = 0.2

0 2 4 6 8 10 12 14 16 18
Normalized time, ωn t
A. Transient Phase error of 2nd order loop

2.5
ζ = 0.2
Output phase estimate, φo [rad]

ζ = 0.5
2
ζ = 0.707

1.5

1
ζ = 1.0
0.5 ζ = 3.0
actual = π/2
0
0 2 4 6 8 10 12 14 16 18
Normalized time, ωn t
B. Transient Output phase estimate of 2nd order loop

Figure 3.6: Transient responses of 2nd-order PLL for various values of ζ due to phase step
input of π/2 radians

The first-order PLL is not capable of responding to a frequency step with non-zero
steady-state error. The error will instead converge to a finite value. The canonical second-
order loop, i.e., one built with a loop filter with infinite DC gain, can track with zero error
as is shown. Second-order loop configurations built without an infinite DC gain loop filter
perform more like a first-order loop with finite steady-state error. This characteristic defends
our assertion in section 3.2 that this type of second-order loop is actually better described
as a modified first-order loop. The third-order PLL, like the second-order canonical loop,
tracks a frequency step with zero error.
In Figure 3.8, we show the response of H2 (z) to a frequency ramp of magnitude ∆ω̇
[rad/s2 ]. This would correspond to a constant acceleration in doppler for mobile or satellite
communications. Note that the error does not now go to zero in steady-state, but rather to
a constant, ∆ω̇/ωn2 . This value is, however, fairly small for reasonable system parameters
and for practical ramps that might be encountered in terrestrial wireless or satellite com-
munications. Therefore, the second-order loop is not disqualified from consideration for the
purpose of carrier recovery. The first-order PLL cannot keep up with a phase acceleration
CHAPTER 3. THE PHASE-LOCKED LOOP 26

0.4 ζ = 0.2

Phase error, sin(φe) [rad]

ζ = 0.5
0.2
0
-0.2
-0.4 ζ = 0.707
ζ = 1.0
-0.6 ζ = 3.0

-0.8
0 2 4 6 8 10 12 14 16 18
Normalized time, ωn t
A. Phase error for 2nd-order loop due to frequency offset, ωoff = ωn
Phase error, sin(φe) [rad]

0.5 ζ = 0.2
ζ = 0.5

-0.5
ζ = 0.707
ζ = 1.0
-1 ζ = 3.0
0 2 4 6 8 10 12 14 16 18
Normalized time, ωn t
B. Phase error for 2nd-order loop due to frequency offset, ωoff = 2ωn

Figure 3.7: Transient responses of 2nd-order PLL for various values of ζ due to frequency
step input: (A) Frequency step, ωof f = ωn ; (B) Frequency step, ωof f = 2ωn

1.5

ζ = 0.2
Normalized Phase error, φ / ( ∆ω / ωn2 )

ζ = 0.5

0.5

ζ = 0.707
ζ = 1.0
ζ = 3.0

0 2 4 6 8 10 12 14 16 18
Normalized time, ωn t

Figure 3.8: Transient responses of 2nd-order PLL for various values of ζ due to frequency
ramp input (time normalized with ωn ; error normalized with ωn2 /∆ω̇)
CHAPTER 3. THE PHASE-LOCKED LOOP 27

such as this and its steady-state error approaches infinity. For zero steady-state error, the
third-order PLL is required.

3.4.3 Cycle Slipping

In some operating conditions, instantaneous noise spikes can cause the PLL to temporarily
loose lock. Figure 3.9 shows an example of this phenomena where the output phase estimate,
θ̂, is plotted as a function of normalized time. The input signal to noise ratio in this
example, SNRi , is set to -8 dB. Loop bandwidth, BL is set to approximately 10 times
the input bandwidth, Bi , yielding a PLL gain of about 10 dB. Loop SNR is thus given by
SNRL ≈ 2 dB. In this region of SNRL , cycle slipping becomes an important attribute of PLL
operation. The actual value of the phase, θ, being tracked is π/2. Note that the PLL quickly
acquires the phase, but after some time, noise dislodges this steady-state condition. The
PLL subsequently re-acquires to various multiples of 2π + π/2, and this behavior continues
randomly over time.

6π+π/2
20
Output phase estimate, θo [rad]

15 4π+π/2

10
2π+π/2

actual value: π/2

0

0 50 100 150 200 250

Normalized time, ωn t

Figure 3.9: Example of cycle slipping (input SNR = -8 dB, Bi /BL ≈ 10)

Viterbi [9], using Fokker-Planck equations, was able to derive a theoretical value for the
average time between cycle slips, TAV , for a first-order PLL, viz.,

π 2 SNRL Io2 (SNRL )

TAV = (3.43)
2BL
where SNRL is expressed in linear form, and Io (x) is the zeroth-order modified Bessel
function of the first kind. For reasonably high SNRL , this expression may be approximated
CHAPTER 3. THE PHASE-LOCKED LOOP 28

by
π
TAV ≈ exp (2SNRL ) . (3.44)
4BL
This expression does not take into account the fact that cycle slips may cluster however.
It is sufficient to note that for values of SN RL > 10 dB, which are of interest for carrier
recovery in PSK modulation schemes, the predicted values of TAV are quite high. Note once
again that BL can be made arbitrarily small to control cycle slipping if necessary.

3.4.4 Acquisition Time

The aquisition process of a phase-locked loop is an inherently non-linear one. No simplifying

linear approximations can hold. Therefore, as noted above, no present exact theory exists
to predict the probability distribution of acquisition time, tacq , in the presence of noise. We
therefore appeal to simulation results to give an intuitive notion of the acquisition behavior
of the PLL.

180

160
Normalized Acquisition time, ωn tacq

140

120

100

80

60

40

20

0
0 1 2 3 4 5

Normalized frequency offset, f off / BL

Figure 3.10: Acquisition time for 2nd-order PLL for various values of initial frequency offset
(ζ = 2− 2 )
1

Figure 3.10 shows the time, normalized to loop natural frequency ωn , for the phase error
out of the detector to reduce to less than 15o , given an initial frequency offset, which is
normalized to loop bandwidth, BL . Note that the choice of 15o as a measure of acquisition
is completely arbitrary and serves only as a basis for comparison. In short, for frequency
offsets less than than BL , the loop converges quite quickly. As the offset increases, however,
the acquisition time rises exponentially. A narrow loop bandwidth is desirable for noise
immunity, but clearly it cannot be arbitrarily narrow if reasonable acquisition time is a
necessity. The exponential behavior of acquisition time in the presence of the frequency
CHAPTER 3. THE PHASE-LOCKED LOOP 29

offset serves to motivate the discussion of AFC-aided acquisition in Chapter 5.

3.5 Performance of Derived Discrete-time Models

We wish to briefly show simulation results that validate the discrete-time models developed
in section 3.3 before moving on to address the topic of PLL optimality in section 3.6. This is
accomplished by 1) comparing the frequency responses of the derived loops to those of the
canonical models defined by (3.9) and (3.12); and 2) comparing the phase jitter performance
of the derived models to that of the canonical models, the theoretical values of which were
developed in section 3.4.
Remember that in section 3.3 we proposed discrete-time second and third-order PLLs,
(3.18) and (3.22), which did not identically match the transfer functions, (3.14) and (3.20),
derived from the BZT of the canonical continuous-time models. The fundamental difference
between the models proposed and those predicted by an exact solution is the lack of a
single zero at z = −1 in the proposed models. Figures 3.11 and 3.12 show the frequency
responses of the proposed second and third-order models versus exact solutions respectively.
In each case, two corner frequency values, ωn or ω3 , are shown. Clearly the lack of a high
frequency zero does change the shape of the frequency response. Nevertheless, the change
is only noticable for the very high frequency values, approaching 12 Fs . Note also that the
differences between ideal H(z)s and the suggested H2 (z) and H3 (z) are less critical as the
1
corner frequencies become smaller, i.e., move farther away from 2 Fs . Since the corner
frequency is usually chosen to be relatively small compared to the sample rate, Fs , we
therefore recommend these PLLs for general use with the caveat that in extremely wide
loop bandwidth conditions, the proposed loops may not perform as well as expected in the
area of noise immunity.
Figures 3.13 and 3.14 show the performance of the second and third-order PLL mod-
els derived in section 3.3 in the important measure of steady-state variance of the phase
estimate, θ̂. In both cases, the theoretical lines are simply plots of (3.37) with BL given
by equations (3.38) and (3.39) respectively. As can be seen in Figure 3.13, the proposed
second-order model almost perfectly matches theory. The match is somewhat less than
perfect for the third-order model in 3.14 where the approximate model of (3.22) actually
performs better than theory. It is unclear exactly why this is so but nevertheless shows that
the performance of the derived third-order loop is approximately consistent with theory and
moreover, is consistent with itself for a wide variety of input SNRs.
CHAPTER 3. THE PHASE-LOCKED LOOP 30

Figure 3.11: Frequency responses of second-order, discrete-time PLLs expressed in equations

(3.14) and (3.18) for two choices of ωn (ζ = 2− 2 )
1

Figure 3.12: Frequency responses of third-order, discrete-time PLLs expressed in equations

(3.20) and (3.22) for two choices of ω3
CHAPTER 3. THE PHASE-LOCKED LOOP 31

Figure 3.13: Phase output variance (jitter) of derived discrete-time 2nd-order PLL, versus
theoretical values, for two values of ωn

Figure 3.14: Phase output variance (jitter) of derived discrete-time 3rd-order PLL, versus
theoretical values, for two values of ω3
CHAPTER 3. THE PHASE-LOCKED LOOP 32

3.6 The Notion of Optimality for a PLL

Loop bandwidth represents the fundamental tradeoff in the design of a PLL. We wish to

1. reduce the output phase jitter due to external noise by making BL as small as
possible, and
2. reduce the acquisition time and transient errors by making BL as large as possi-
ble.

Obviously, we cannot have it both ways. For any particular application, we wish to have
some combination of the above two functions, which would make the performance of the
PLL “optimal” in some sense.¶ Optimality has been defined in the mean square sense for
an input signal in Gaussian noise by minimizing

Σ2 = θ̂n2 + λ2 ET2 (3.45)

where θ̂n is as defined in (3.37) and ET2 is a measure of the total transient error, viz.,
Z ∞
ET2 = e2 (t)dt . (3.46)
0

Note that e(t) is as defined in (3.30), and λ is a Lagrangian multiplier used to establish the
relative weighting of the transient error. Table 3.1 shows optimal loops for three different
input signals in AWGN. The Wiener optimization also specifies the optimal choice of loop
natural frequency as a function of input power to noise spectral density ratio, which is related
to input SNR by P/No = SNRi Bi , where Bi is the baseband equivalent noise bandwidth.
√
The canonical second-order loop described by (3.9) with damping factor ζ = 1/ 2 is
optimal for a frequency step. A frequency offset at loop start up is an important example of
a frequency step. This justifies the attention thus far on the second-order loop. The first-
order loop while optimal for a phase step shows poor tracking capability for a frequency
offset. For this reason, the first-order PLL is virtually ignored in this research. The third-
order loop is optimal for a frequency ramp (or “chirp”) signal. This is very important in
some applications and is of significant interest in mobile communications when a changing
doppler frequency might be expected.
The choice of a particular order of PLL can be based on the particular type of signal
to which we wish to synchronize. The parameters for the loop may be chosen based on the
signal environment in which the loop must operate. In this sense, there are infinitely many
¶
This brief discussion follows mainly from [6].
CHAPTER 3. THE PHASE-LOCKED LOOP 33

Table 3.1: Optimum PLLs for three input signals in AWGN

Input Signal Optimum Loop, H(s) Optimum Bandwidth

s
ω1 P
Phase step: θ(t) = ∆θ ω1 = ∆θλ
s + ω1 No

√ s
ωn2 + 2ωn s P
Frequency step: θ(t) = ∆ωt √ ωn2 = ∆ωλ
ωn2 + 2ωn s + s2 No

s
∆ω̇t2 ω33 + 2ω32 s + 2ω3 s2 P
Frequency ramp: θ(t) = ω33 = ∆ω̇λ
2 ω33 + 2ω32 s + 2ω3 s2 + s3 No

Adapted from [6].

optimal loops for infinitely many different applications. Finally, Gardner points out in [6]
that an extremum for a PLL tends to be quite broad, which means that there are many
values of loop parameters (ζ and ωn in the second-order loop) that give nearly optimum
performance. Given a basic understanding of loop behavior, these nearly optimum solutions
may often be found simply by trial and error.

3.7 Chapter Summary

In this chapter, we have endeavored to provide a cursory introduction to phase-lock loop.
Special attention has been given to second and third-order PLLs, which are of particular in-
terest here in their application to suppressed-carrier recovery. We began with a continuous-
time model, in which linearization and the application of the Laplace transform helped us
to understand the behavior of this feedback control loop and aided in the development of
canonical equations with intuitively satisfying parameters.
In section 3.3, we applied the BZT to these canonical equations to come up with discrete-
time models suitable for implementation with digital components or in software on a DSP
or digital computer. The important point in these derivations was that we preserved the
classical parameters, ζ, ωn , and ω3 , which were developed many years ago for analog circuits.
CHAPTER 3. THE PHASE-LOCKED LOOP 34

Use of these parameters will hopefully serve as an aid to the digital designer, allowing, at
least in first-cut design, rapid development of digital PLLs that closely approximate the
classical continuous-time models.
We devoted significant attention in section 3.4 to defining and describing a variety
of performance measures for PLLs. Where they exist, theoretical expressions for these
measures were stated or referred. The most important measure was steady-state phase
jitter, defined as the steady-state variance of the output phase estimate. This led to the
definition of loop bandwidth, BL , for a PLL.
Note again that in PLL design there is a fundamental tradeoff between acquistion time
and phase jitter. Large BL enables rapid acquisition time and thus small transient error
but allows large phase jitter, which in the more extreme cases can lead more deleterious
effects such as cycle slipping. Narrowing BL , on the other hand, can lead to exponentially
increasing acquisition time and thus large transient errors but can arbitrarily reduce steady-
state phase jitter. The tradeoff between these transient versus steady-state error has been
examined under the Wiener optimization criterion given in (3.45) to which we devoted only
cursory discussion in section 3.6.
We now seek in the following chapters to apply what was learned in this chapter to the
task of recovering a suppressed-carrier in digitally phase modulated signal formats. As will
be shown, the basic PLL structure is at the core of all closed-loop carrier recovery schemes.
Moreover, the performance of both the closed-loop and open-loop schemes presented in
Chapter 4 will be shown to bear striking resemblance to those expressions given in this
chapter.
Chapter 4

Coherent Carrier Recovery

Structures

The fact that the PLL exhibits a transient response given a phase step input makes it an
unlikely candidate for carrier recovery of digitally phase modulated signal formats. In fact,
the PLL is not generally useful for carrier recovery when the signal of interest has phase
discontinuities like those found in PSK. The foregoing coverage of the PLL in Chapter 3
has not been in vain, however. All closed-loop carrier recovery schemes for PSK modulation
are based on the same principle as the PLL but simply add a mechanism to remove the
modulation. As will be shown, this removal is done either by decision feedback (remod-
ulation) or by an M th-order nonlinearity, which provides a discrete spectral line∗ for the
PLL to track. The Costas and squaring loops, outlined in section 4.1, each perform this
task with theoretically equivalent performance. Each is an extension of the PLL to closed-
loop carrier tracking schemes. Open loop schemes operate on the principle of modulation
removal, as well. These, however, use feedforward estimation techniques to remove phase
and frequency errors. We provide a simple overview of open-loop techniques applicable to
the task of carrier recovery in section 4.2. Finally, we conclude the discussion of carrier re-
covery structures with an exposition of the concept of automatic frequency control (AFC),
presented in Chapter 5 as an acquistion aid to phase-coherent techniques.
∗
In the case of the Costas loop, the discrete spectral line might be said to be from a “phantom” carrier
since the discrete spectral line is around 0 Hz.

35
CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 36

4.1 Closed-Loop Carrier Recovery Structures

4.1.1 Costas Loop

The Costas loop is by far the most common implementation of closed-loop carrier recovery
and as will be shown, provides not only a convenient mechanism for suppressed-carrier
tracking but also fully demodulates the received signal. We therefore use the terms, “Costas
loop” and “Costas demodulator,” interchangably throughout the discussion.
√
Consider the classical Costas loop structure shown in Figure 4.1 where s(t) = 2P
d(t) cos(2πfc t + θ) and d(t) ∈ {−1, +1} denotes the transmitted binary signal at time, t.
The Q-arm, taken by itself, provides a PLL-like control signal except for the modulation by
d(t). After low pass filtering
s
P  
q(t) = d(t) sin θ − θ̂ (4.1)
2

where, following our previous conventions, θ is the carrier phase in the SOI and θ̂ is the
output phase estimate from the loop. Note that the fundamental goal of suppressed-carrier
recovery is to provide some mechanism for removing d(t) from a control signal like (4.1).
The I-arm control signal is found to be
s
P  
i(t) = d(t) cos θ − θ̂ . (4.2)
2
Note that the arm LPFs will usually be designed with a cutoff around the baud rate if data
demodulation is desired. Multiplying (4.1) and (4.2), the error control signal is
P 2     P
e(t) = d (t) sin θ − θ̂ cos θ − θ̂ = sin(2φ) (4.3)
2 4

where φ = θ − θ̂ denotes the phase difference between input and reference carriers and d2 (t)
= 1. Thus we have removed the modulation via squaring the received signal and provided
a PLL-like error signal in e(t). For e(t) small, the linearizing approximation, sin(2φ) ≈ 2φ,
can be made and most of our intuition for PLLs can be applied to the Costas loop. Note
also that the I-arm signal at steady-state, 2φ  1 implying cos(2φ) ≈ 1. This means that
i(t) will be proportional to the data signal, d(t), (with additive noise) and will be used for
the receiver decision metric.

4.1.2 Squaring Loop

The Costas loop can be shown to have the same noise performance as the squaring loop
shown in Figure 4.2, and its functionality is simpler to observe. After bandpass filtering for
CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 37

i( t ) to data
X LPF decoder

cos( 2π f c t + θ$ )

Loop Filter
e (t )
s (t ) VCO F(s) X
90o

− sin( 2πf c t + θ$ )

q (t )
X LPF

Figure 4.1: Basic structure of analog Costas demodulator

Loop Filter
BPF BPF
s (t ) ( *) 2 X F(s)
@ ωc @ 2ωc e (t )

− sin( 2ω c t + 2θ$ )

VCO

÷ 2
− sin( ω c t + θ$ )

to data
X LPF decoder

Figure 4.2: Squaring Loop

CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 38

√
noise immunity, we apply a direct squaring operation on s(t) = 2P d(t) cos(ωc t + θ).† We
obtain at the double frequency,

P d2 (t) sin (2ωc t + 2θ) .

Since d2 (t) = 1, a discrete spectral line appears at 2ωc , which may be filtered by a bandpass
filter (BPF) whose bandwidth need only range over the uncertainty of the value for 2ωc . This
is then applied to a PLL tuned to 2ωc which yields a phase estimate, 2θ̂. The local reference
may then be divided to provide − sin(ωc t + θ̂) to downconvert s(t) to baseband. Note that
the multiplication and division operations have led to a 180o phase ambiguity, analogous to
taking the square root of a squared quantity. This is not a shortcoming of the squaring or
Costas loops but a fundamental shortcoming in the detection of phase modulated signals
with locally generated references. If purely coherent detection is required, then some means
of removing this ambiguity, e.g., a training sequence, must be found. Another very common
method is the use of differentially encoded data described in section 1.3.2.
The squaring loop structure may easily be extended to higher-order modulations. For
M -ary PSK signal sets, the squarer can be replaced simply with an M th-law device, and the
divide-by-2 can be replaced by a divide-by-M circuit. Note, however, that an M th-order
phase ambiguity, i.e., 360o /M , is introduced.

4.1.3 Noise Performance

We would expect the Costas and squaring loops to have noise performance similar that of
the PLL. This is true except for an added squaring loss, which comes from a noise analysis
similar to that found in section 3.4.1. It is shown in [6] that the phase estimate of 2θ̂ will
have a steady-state variance due to additive white noise of
 2    
No BL No Bi
2θ̂n = 4 1+ [rad2 ]
P 2P
   
BL 1
= 4 1+ [rad2 ] (4.4)
SNRi Bi 2SNRi
where BL is as defined in (3.37), Bi is the input equivalent noise bandwidth, and SNRi is
the input signal to noise power ratio. Squaring loss arises from the fact that, unlike the
PLL, we perform a squaring operation on noise producing intermodulation products with
non-zero variance.
†
A square-law nonlinearity is used because its analysis proves to be tractable and is in no sense optimal.
Other non-linearities can be used to provide better noise performance. The arctangent operation can in fact
be shown to be optimal and is briefly discussed in section 4.1.6.
CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 39

Table 4.1: Intermodulation losses in M th-order phase synchronizers

M Loss, Ln (SNRi )

1 1

1
2 1+
2SNRi

2 2
3 1+ +
SNRi 3SNR2i

9 6 3
4 1+ + +
SNRi SNR2i 2SNR3i

Adapted from [6].

The notion of squaring loss is actually a special case of intermodulation losses for M th-
order recovery loops. Reference [6] provides the following result for the generalized M -phase
synchronizer:  
 2 No BL
2
M θ̂n = M Ln (SNRi ) [rad2 ] (4.5)
P
where Ln (SNRi ) is a loss defined as a function of input SNR in its linear representation.
Table 4.1 shows the losses in carrier synchronizers for several PSK constellations.‡ Note
that for high input SNR, all such losses are negligible. At low SNR, however, the losses
become very significant, and require a severe restriction on Costas loop bandwidth in order
to provide sufficient loop SNR.

4.1.4 Discrete-Time Implementation

We can, however, improve the SNR seen by the recovery loop by using I&D filters as shown
in the discrete-time structure of Figure 4.3. These filters could easily be taken to be filters
matched to a pulse shape depending upon the application. We note that the sine and cosine
of the phase difference, φ = θ − θ̂, can be averaged over baud periods in the quadrature
arms of the digital receiver. This provides the Costas loop with sufficient information
about sin(2φ) to still acquire the carrier, although it is now operating at the baud rate,
Rb = 1/Tb , rather than at the sample rate, Fs . We note here that in terms of absolute time,
the acquisition process of the Costas loop is slower than a sample rate implementation. In
‡
Although the result for M = 3 seems for the moment to be purely academic!
CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 40

Figure 4.3: Block diagram of Costas demodulator, discrete-time implementation

terms of the number of samples, however, the bit rate implementation aquires more quickly
in general. This baud rate implementation strategy can have tremendous reductions in
required processing budget.
The integrate and dump (I&D) circuit can be replaced with a matched filter in the case
of pulse shaping or baseband equalization. Similarly, in direct-sequence spread spectrum
(DSSS) applications, it can be replaced with a matched filter or binary correlator for de-
spreading. In the case of DSSS, the despreader becomes imperative for two reasons. First,
input SNRs for spread spectrum are typically quite low. Typical values range from 0 to
−50 dB in some applications. These are regimes where the performance of the Costas loop
is marginal at best. Despreading in each of the quadrature arms allows a sufficiently high
input SNR to be seen by the carrier recovery loop. Secondly, in the case of DSSS code
division multiple access (CDMA), there are many interfering signals present in s(t) that are
at roughly equal powers to the SOI and at the same nominal carrier frequency. Without
prior despreading, there would be no way to ensure that the Costas loop would track the
correct carrier.
The structure of Figure 4.3 gives insight into a completely software-based downconver-
sion process. The input signal, s(t), is sampled, creating {si }, where we denote values
obtained at the original sample rate with the subscript i and values sampled at the baud
rate with subscript k. We then apply direct digital downconversion (DDC) by multiplying
with the cosine and negative sine of the NCO output. Arm LPFs can be either finite or
infinite impulse response, (FIR) or (IIR), digital filters with corner frequencies set to the
CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 41

baud rate. These I&D filters can improve the SNR seen by carrier recovery loop by their
averaging of the input noise processes.
Symbol timing is assumed and not explicitly shown in Figure 4.3. In fact, symbol or
code timing is usually linked to the carrier recovery process in a “bootstrapping” manner,
i.e., carrier recovery relies on bit timing and vice-versa. In practice, this means that rough
symbol timing can be gained non-coherently; if the carrier offset is not too large, relative
to the baud rate, then the phase difference, φ, while non-zero, is relatively constant over a
baud period. This allows the quadrature I&D circuits to provide correlation information to
the symbol timing processor, and a proper sampling time, kT , can be selected. Meanwhile,
as ik and qk estimates improve, better phase information will be provided for the error
control samples, ek , allowing carrier recovery. This will in turn provide more robust symbol
timing information and so on.
Hard limiting of the I-arm is commonly applied in practical Costas demodulators. This
practice fundamentally changes how the Costas loop removes the modulation from the SOI.
It essentially becomes a decision feedback, or remodulation, structure. This non-linearity
adds to the complexity of the transient behavior of the loop, but in the steady-state where
the phase error is expected to be small, it is easy to see why this is done. We wish to show
that this practice, common for high SNRs ( 0dB), fundamentally changes the steady-state
stochastic performance of the Costas loop to that of the PLL.
The quadrature arm signal in (4.1) is quite similar to the error signal of a basic PLL
except it is modulated with d(t). In discrete-time, this signal may be expressed by
s
P
qk = dk sin(φk ) (4.6)
2

where φk = θk − θ̂k , the phase error at time, t = kT . In the I-arm the sampled output is
s
P
ik = dk cos(φk ) . (4.7)
2

As φk becomes small, ik ∝ dk . Hard limiting thus produces an estimate, dˆk , for the
q
P
transmitted symbol. Except for bit errors then, ek = Sgn (ik )qk = 2 sin(φk ). Note that
now, in this special case, the Costas loop error term, ek is no longer proportional to 2φ but
rather to φ. Therefore, when examining the variance (jitter) of the phase estimate, (4.4) no
longer applies. Rather, this expression reduces to that of (3.37).
After the multiplication, the discrete-time Costas loop model operates in precisely the
same fashion as the discrete-time PLL in Figure 3.4. The digital loop filter may be a
direct form implementation of either (3.17) or (3.21), depending on the desired order of the
CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 42

system. We now model the NCO as a simple accumulator with “gain” Ko = 1. This, in
turn, produces an estimate of the carrier phase, θ̂k , that will be held constant over the next
baud period. The cosine and sine generators operate at the sample rate.

4.1.5 Four-Phase Costas Loop Implementation

Like the squaring loop, the Costas loop can be extended to higher modulations. A four-
phase Costas loop, suitable for the detection of QPSK signals, is depicted in Figure 4.4.
For convenience, we model the QPSK signal by
√ √
s(t) = 2P x(t) cos (ωc t + θ) − 2P y(t) sin (ωc t + θ) (4.8)

where x(t), y(t) ∈ {±1} are the in-phase and quadrature bits at time t. Sampling at time k
and multiplying digitally by the references
   
Ko cos ωc kTs + θ̂ and − Ko sin ωc kTs + θ̂

produces quadrature samples, after ideal lowpass filtering, of the form

s s
P   P  
ik = Ko xk cos θ − θ̂ − Ko yk sin θ − θ̂ (4.9)
2 2

and s s
P   P  
qk = Ko xk sin θ − θ̂ + Ko yk cos θ − θ̂ . (4.10)
2 2
For small phase error, θ̂ ≈ θ, the outputs of the hardlimiting operations are taken to be
Sgn{ik } = x̂k and Sgn{qk } = ŷk . For reasonably high SNRs, assume that these decisions
on xk and yk are correct, i.e., x̂k = xk and ŷk = yk . Thus the feedback error term is given
by

ek = ik Sgn {qk } − qk Sgn {ik }

s s
P   P  
= Ko xk yk cos θ − θ̂ + Ko x2k sin θ − θ̂
2 2
s s
P   P  
− Ko xk yk cos θ − θ̂ + Ko yk2 sin θ − θ̂
2 2
√  
= 2P Ko sin θ − θ̂ (4.11)

which is the PLL-like control signal useful for tracking the input carrier.
CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 43

Figure 4.4: Block diagram of four-phase discrete-time Costas demodulator

s (t ) ik
X LPF ∫ (∗ ) dt
TB
t = kT B
vd (t )
cos( ∗ )
ek xM
NCO Fd (z) tan-1(*)
mod 2π
Loop Filter
− sin( ∗ ) Phase Error Detector

vq (t )
qk
X LPF ∫ (∗ ) dt
TB
t = kT B
Figure 4.5: Block diagram of digital tanlock loop (DTL)
CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 44

4.1.6 Digital Tanlock Loop

One noteworthy carrier recovery structure, the digital tanlock loop (DTL) shown in Figure
4.5, has only become feasible since the advent of high-speed digital logic circuits. As its name
implies, it relies on an explicit arctangent operation to yield the phase error term. Thus,
no approximations, such as sin θ ≈ θ, need be used. Except for noise and interference, the
phase estimate is accurate during acquisition as well as during steady-state operation. The
tanlock loop is shown in [50] to have several advantages over traditional closed-loop carrier
recovery schemes. Among these is the fact that, since an explicit arctangent operation is
used, the DTL is insensitive to amplitude variations and can thus operate without benefit of
AGC, or as a demodulator for quadrature amplitude modulation (QAM) signalling formats.
Note the multiplication by M , where M is the number of phase dimensions used in the
signal set, introduces an M -ary phase ambiguity just as in more traditional carrier recovery
circuits.

4.1.7 Discussion

Classical closed-loop carrier recovery structures, such as the squaring and Costas loops, were
developed using continuous-time system models and have been implemented for many years
with fully analog circuitry. In the past 30 years, as electronics technology has permitted,
it has become ever more feasible to implement such loops with digital components. Over
the last 15 years, it has become feasible to implement carrier recovery loops entirely with
digitalcomponents. The DTL is an example of a closed-loop carrier recovery that owes
its existence to these new capabilities. In much of the foregoing discussion, therefore,
we have emphasized the discrete-time implementation of carrier recovery structures. This
ever-increasing ability to now demodulate signals using digital, discrete-time components
or software on a DSP leads naturally to the discussion of open-loop parameter estimation
for the purpose of carrier recovery since open-loop structures are really only suited only for
this type of implementation.

4.2 Open-Loop Carrier Recovery Structures

In this section, we survey a variety of open-loop estimation schemes applied to the problem
of carrier recovery. Three open-loop phase estimators proposed in [41] serve well for the
purposes of illustration. These are derived from the maximum likelihood (ML) principle. In
sections 4.2.1, 4.2.4, and 4.2.5, we describe a first-order phase estimator, a modified first-
order phase estimator, and a second-order phase estimator. The use of the term “order” is
CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 45

original to this research as far as is known and is employed because the estimators presented
have precise performance analogies to first-order, modified first-order, and second-order
PLLs. These analogies will be drawn as the discussion progresses.
The derivation in section 4.2.1 suggests different ML estimation structures depending
upon whether low SNR or high SNR conditions are assumed. Sections 4.2.2 and 4.2.3
describe the steady-state noise performance of each of these structures. In section 4.2.2 we
derive the first and second-order statistics for the error of the first-order ML phase estimator.
In a very useful result, we show that the notion of loop noise bandwidth, developed for
PLLs and extended to closed-loop suppressed-carrier recovery structures, can be applied
directly to open-loop estimators as well. Then in section 4.2.3, we do the same for the
ML estimator derived from a high SNR approximation. Here, we show that the structure
amounts to a decision feedback phase estimator, stochastically equivalent to a Costas loop
with hard-limiting in the I-arm. We also show that, under the conditions of high SNR,
the noise performance of the structure derived using the low SNR assumption converges to
that of the optimal estimator. In section 4.2.6, we examine the performance of the phase
estimation structures presented and validate the theoretical results given in sections 4.2.2
and 4.2.3. Finally, section 4.2.7 applies the open-loop phase estimation to the problem of
carrier recovery for systems in which adaptive equalization is employed.

4.2.1 First-Order ML Phase Estimator

Figure 4.6 outlines the first-order ML phase estimator applied to binary phase modulation.
Intuitively, the structure appears to be a simple averager, summing the squared input over
N symbol observations. The squaring operation doubles the phase of input sample, which
then enters the delay line. This doubling of phase is compensated for later by the square
root operation, noting of course that, as with all closed-loop suppressed-carrier recovery
schemes, the process produces a 180o phase ambiguity. In the proceeding discussion, we
wish to show rigorously that this structure does indeed provide the ML estimate of the
signal phase for binary phase modulation in an AWGN channel.
To derive this structure, we expand the derivation given in [41]. Consider an input to
the receiver given by
√
rk = 2P ej(φk +θ) + nk kT ≤ t ≤ (k + 1)T (4.12)

where φk is the transmitted symbol such that ejφk = dk ∈ {−1, +1} at time kT , θ is the
carrier phase, and nk is a sample of a zero-mean, complex Gaussian process with variance
σn2 = No /T in each signal dimension. Assume, for now, that θ does not vary over an
CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 46

s(t) Matched rk rk‘ Baseband

Filter
x Detector
$
1 2
e − jθ ML
z-1 z-1 ... N z-1 conjugate

(•)2 (•)2 ... (•)2

(•)
Σ
•

Figure 4.6: “First-order” open-loop ML phase estimation structure proposed in [41]

observation interval of length N samples. The likelihood function is thus given by

N  √ 
1 X  j(φk−i +θ)  2
 N −  rk−i − 2P e 
1 2σn2 i=1
p(r | φ, θ) = e
2πσn2
1 X N √ n o
2 −j(φk−i +θ)
 N |rk−i | − 2 2P < rk−i e + 2P
1 2σn2 i=1
= e
2πσn2
X
N n o
α < rk−i e−j(φk−i +θ)
= F e i=1 (4.13)

where <(·) is the real operator,

NP X N
√
 N − 2 + |rk−i |2
1 σn i=1
2P
F = e and α= 2 . (4.14)
2πσn2 σn
Note that F and α are not dependent upon the data φk . We would like the ML estimate
at time k, θ̂M L,k to satisfy
h i
ln p(r, θ̂M L,k ) = Max ln [p(r, θ)] . (4.15)
θ

Now the probability of receiving samples r conditioned upon the carrier phase θ is given as

p(r | θ) = E {p(r | φ, θ)}

 

1 X α<rk−i e−j(φk−i +θ) 
Y
N
= F e
i=1
2φ
k−i

Y
N  n o
= F cosh α< rk−i e−jθ (4.16)
i=1
CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 47

where E(·) denotes the expectation operator. Taking natural logarithms, we have

X
N h  n oi
ln [p(r, θ)] = lnF + ln cosh α< rk−i e−jθ . (4.17)
i=1

Now differentiating with respect to θ and setting equal to zero, we obtain

X
N  n o n o
tanh α< rk−i e−jθ · < jrk−i e−jθ = 0 . (4.18)
i=1

At this point, a solution for θ yields θ̂M L . Unfortunately, no closed form solution yet
exists for (4.18), and more importantly, it gives us no intuitive knowledge about constructing
a receiver based upon it. We now proceed to derive the structure shown in Figure 4.6 by
assuming α < 1, i.e., under low input SNR conditions. Under these assumtions we use
tanh x ≈ x. Then θ̂M L,k ought satisfy

X
N n o n o
< rk−i e−j θ̂ML,k < jrk−i e−j θ̂ML,k = 0 . (4.19)
i=1

Solving directly for ej θ̂ML,k yields

v
u N
u X
u r2
u
u i=1 k−i
ej θ̂ML,k =u
u  N  (4.20)
u X 
t 2 
rk−i 
 
i=1

which is precisely the functionality shown in Figure 4.6.

Now, what about high SNR conditions? After all, if our estimation structure is the best
estimator we can hope for in low SNR, we would surely like it to be optimal or at least
nearly so in the far more common high SNR condition. High SNR allows us to assume
α  1 and <{rk e−jθ } ≈ Aejφk = Adk . Thus we may use tanh(Adk ) ≈ Sgn (Adk ) = dk in
0
(4.18). Under these conditions then, θ̂M L ought satisfy

X
N  
0
dk−i < jrk−i e−j θ̂ML,k =0. (4.21)
i=1

0
Solving for ej θ̂ML,k now yields

X
N
dk−i rk−i
0
j θ̂ML,k i=1
e = N  (4.22)
X 
 
 dk−i rk−i 
 
i=1
CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 48

Figure 4.7: Decision feedback open-loop phase estimation structure derived from the ML
principle under “high” SNR assumptions

which suggests a decision feedback carrier recovery structure shown in Figure 4.7. This
structure seems analogous to the practice of hard limiting the I-arm in a Costas loop in
high SNR conditions.
While this is intuitively satisfying, it nevertheless is not the estimate given by (4.20). In
sections 4.2.2 and 4.2.3, we derive the first and second-order statistics for estimators (4.20)
and (4.22) respectively. Our goal is to show that these converge to the same values under
high SNR assumptions. We further note that the differences between (4.22) and (4.20) are
precisely analogous to the differences between hard-limiting and not limiting the I-arm in
a Costas loop. Moreover, we show that in both cases, a precise analogy exists between the
loop bandwidth of a Costas loop or PLL and the structures suggested by (4.20) and (4.22).

4.2.2 Noise Performance of Low SNR ML Phase Estimator

4
Consider now the error of the estimator given in (4.20), namely θ̂e,k = θ̂M L,k − θ. Squaring
(4.20) and taking natural logarithms yields
!
1 XN
2
θ̂M L,k = arg rk−i (4.23)
2 i=1
 
4 =(·)
where arg(·) = tan−1 and <(·) and =(·) are the real and imaginary operators. We
 <(·)
  
note that −θ = arg e−jθ = 12 arg e−j2θ . Therefore
!
1 XN
θ̂e,k = arg e−j2θ 2
rk−i . (4.24)
2 i=1
CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 49

Now
rk = Adk ejθ + ñk (4.25)
√
where A = 2P , dk = ejφk ∈ {−1, +1} is the transmitted BPSK symbol at time k and
ñk = nd,k + jnq,k , nd , nq ∼ Normal(0, σn2 ). It is convenient to define

4 X
N
2
z2θ,k = rk−i
i=1
X
N
= N A2 ej2θ + 2Adk ejθ (nd,i + jnq,i )
i=1
N 
X  
+ n2d,i − n2q,i + j2nd,i nq,i . (4.26)
i=1

Further, we define

4
zk = e−j2θ · z2θ,k
X
N
= N A2 + 2Adk e−jθ (nd,i + jnq,i )
i=1
N 
X  
+e−j2θ n2d,i − n2q,i + j2nd,i nq,i . (4.27)
i=1

For convenience we also wish to define auxilliary noise variables n0d = nd +nq and n0q = nd −nq
such that n2d − n2q = n0d n0q . We note that n0d , n0q ∼ Normal(0, 2σn2 ) and can be shown to be
mutually independent as well as independent of nd and nq . Thus (4.27) may be rewritten
as
N 
X  
X
N −j n0d,i n0q,i + j2nd,i nq,i
2 −jθ
zk = N A + 2Adk e (nd,i + jnq,i ) + e i=1 . (4.28)
i=1

Thus the error of the phase estimate may now be given by

 
1 1 =(zk )
θ̂e,k = arg(zk ) = tan−1 . (4.29)
2 2 <(zk )

We now define auxilliary random noise variables λ1 and λ2 by

4
λ1,k = =(zk )
!
X
N X
= 2Adk cos θ nq,i − sin θ nd,i
i=1 i
X
N X
+2 cos 2θ nd,i nq,i − sin 2θ n0d,i n0q,i (4.30)
i=1 i
CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 50

and
4
λ2,k = <(zk ) − N A2
!
X
N X
N
= 2Adk cos θ nd,i + sin θ nq,i
i=1 i=1
X
N XN
+2 sin 2θ nd,i nq,i + cos 2θ n0d,i n0q,i (4.31)
i=1 i=1

which allow us to express the estimation error as

!
1 λ1,k
θ̂e,k = tan−1 . (4.32)
2 N A2 + λ2,k

Before proceeding any further, let us consider the first- and second-order statistics of
the noise variables λ1 and λ2 . Considering the first of these, we may write
"N # " # " #
X X
N X
N
E[λ1,k ] = 2AE[dk ] cos θE nq,i − 2AE[dk ] sin θE nd,i + 2 cos 2θE nd,i nq,i
i=1 i=1 i=1
" #
X
N
− sin 2θE n0d,i n0q,i
i=1
= 0. (4.33)

Analogously, it can be shown that E[λ2,k ] = 0. So the expected value of each of the noise
variables is zero. We now consider the variance of λ1 . First of all,
" ! #2
X
N X
N X
N X
N
λ21,k = 2Adk cos θ nq,i − sin θ nd,i + 2 cos 2θ nd,i nq,i − sin 2θ n0d,i n0q,i
i=1 i=1 i=1 i=1
 !2 !2 
X
N X
N X
N X
= 4A2 cos2 θ nq,i + sin2 θ nd,i − 2 cos θ sin θ nq,i nd,i 
i=1 i=1 i=1 i
!2 !2
X
N X
N
+4cos2 2θ nd,i nq,i + sin2 2θ n0d,i n0q,i
i=1 i=1
X
N X
N
+4Adk (four noise terms) − 4 cos 2θ sin θ nd,i nq,i n0d,i n0q,i . (4.34)
i=1 i=1

Examining the expectations of the various noise terms in (4.34), we find

E [dk ] = 0 (4.35)
"N #
X X
E nq,i nd,i = 0 (4.36)
i=1 i
 !2   !2 
X
N X
N
E  nq,i  = E  nd,i  = N σn2 (4.37)
i=1 i=1
CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 51

 !2 
X
N
E  nd,i nq,i  = N σn4 (4.38)
i=1
 !2 
X
N
E  n0d,i n0q,i  = 4N σn4 (4.39)
i=1
"N #
X X
N
and E nd,i nq,i n0d,i n0q,i =0. (4.40)
i=1 i=1
Using (4.35) through (4.40) in (4.34) and taking the expectation, we obtain
h i  
E λ21,k = 4A2 cos2 θN σn2 + sin2 θN σn2 + 4cos2 2θN σn4 + sin2 2θ · 4N σn4
= 4N A2 σn2 + 4N σn4 . (4.41)
h i h i
It can be shown that E λ22,k = E λ21,k .
Now let us reconsider the error or the ML phase estimator given in (4.32). Let us
examine this error for a limited subset of all k, denoted by κ, where the realizations of the
noise processes λ1 and λ2 at times κ produce an argument to the arctangent operation that
is smaller than one. For the assumption of high SNR which is being made, the constant
N A2 in the denominator of (4.32) is much larger than the variance of either λ1 or λ2 . This
leads the useful conclusion that the occurrences κ are highly probable within the set of all
k. In fact, in the limit as SNR → ∞, the set of all κ is approaches the set of all k. For the
realizations at time κ, we therefore allow tan−1 (x) ≈ x. Thus§
1 λ1,κ
θ̂e,κ ≈ · 2
. (4.42)
2 N A + λ2,κ
In order to make use of expectation operators, we now make the assumption that the
output random process is independent of the input noise processes. This is precisely anal-
ogous to the assumption of independence of the steady-state phase error in a PLL to the
additive input noise made in section 3.4.1. As noted in that section, this type of assump-
tion is made in all known PLL treatments, and further yields results consistent with those
observed in practical systems. The independence assuption allows
  1
θ̂e,κ N A2 + λ2,κ ≈ λ1,κ
2
h i  1
=⇒ E θ̂e,κ N A + E [λ2,κ ] ≈ E [λ1,κ ]
2
2
h i 1 E [λ1,κ ]
=⇒ E θ̂e,κ ≈ · (4.43)
2 N A2 + E [λ2,κ ]
§
The Taylor series expansion of arctangent takes three forms. For the region between ±1 the first term of
the approximation produces (4.42), and higher-order terms become neglible based on the assumptions used
here. Use of this expansion, however, still requires the assumption that the argument to the arctangent,
itself a function of random variables, lies between ±1. Our approach does not explicitly rely on such an
assumption, and is therefore presented as an alternative to the Taylor series expansion.
CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 52

as an approximation of the mean of the ML estimator given in Figure 4.6 for times κ.
Similarly, h i
h i 1 E λ21,κ
2
E θ̂e,κ ≈ · h i (4.44)
4 N 2 A4 + E λ2 + 2N A2 E [λ2,κ ]
2,κ

approximates the error variance at times κ. But the noise processes λ1 and λ2 are clearly
stationary upon inspection of (4.30) and (4.31). So we may state with equivalent certainty
that at times κ,
h i 1 E [λ1,k ]
E θ̂e,κ ≈ · (4.45)
2 N A2 + E [λ2,k ]
and h i
h i 1 E λ21,k
2
E θ̂e,κ ≈ · h i . (4.46)
4 N 2 A4 + E λ2 + 2N A2 E [λ2,k ]
2,k

Now the first- and second-order statistics of the estimation error are no longer dependent
upon κ. So we may finally state
h i 1 E [λ1,k ]
E θ̂e,k ≈ · (4.47)
2 N A2 + E [λ2,k ]
and h i
h i 1 E λ21,k
2
E θ̂e,k ≈ · h i . (4.48)
4 N 2 A4 + E λ2 + 2N A2 E [λ2,k ]
2,k
h i h i
Given the values for E [λ1,k ] = E [λ2,k ] and E λ21,k = E λ22,k in (4.33) and (4.41), we
may now solve (4.43) and (4.44). The first- and second-order statistics for the ML estimator
derived from a low SNR assumption are given by
h i
E θ̂e,k ≈ 0 (4.49)

which is not terribly surprising, and

h i A2 σn2 + σn4
2
E θ̂e,k ≈ (4.50)
N A4 + 4A2 σn2 + 4σn4

A2
which proves to be an interesting result indeed. Let ρ = = Eb /No be the symbol SNR.
2σn2
We may then express (4.50) as
h i 2ρ + 1
E θ̂e,k ≈ . (4.51)
4N ρ2 + 4ρ + 4
Maintaining the high SNR assumption and reasonably large N , we may legitimately assume
N ρ2  ρ + 1. Therefore
h i  
2ρ + 1 1 1
E 2
θ̂e,k ≈ 2
= 1+ . (4.52)
4N ρ 2N ρ 2ρ
CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 53

This is a most intuitively satisfying result since it is an exact analogy to the steady-state
noise variance obtained for a Costas loop in (4.4), including squaring loss. Furthermore, the
notion of loop noise bandwidth, developed in Chapter 3 for the PLL has a direct application
to this open-loop estimator. Simply by equating (4.52) with (4.4), we find an equivalent
loop bandwidth for the open-loop estimator, which is given by
Bi
BL,θ̂ = (4.53)
2N
where Bi = Rb = 1/Tb for the estimator of Figure 4.6.¶

4.2.3 Noise Performance of High SNR ML Phase Estimator

Taking natural logarithms and rearranging (4.22), we write

!
X
N
0
θ̂M L,k = arg dk−i rk−i . (4.54)
i=1

It is convenient to define zθ0 as

4 X
N
0
zθ,k = dk−i rk−i
i=1
XN
= Aejθ + ñi (4.55)
i=1

where ñi and A are as defined in 4.25. Note that the modulation is simply removed in the
sum. We further define
4
zk0 = e−jθ · zθ,k
0
!
X
N X
N X
N X
N
= AN + cos θ nd,i + sin θ nq,i + j cos θ nq,i − sin θ nd,i . (4.56)
i=1 i=1 i=1 i=1

Let
4
λ01,k = =(zk0 )
X
N X
N
= cos θ nq,i − sin θ nd,i (4.57)
i=1 i=1
4
and λ02,k = <(zk0 ) − AN
X
N X
N
= cos θ nd,i + sin θ nq,i . (4.58)
i=1 i=1

¶
As far as is known, the expression of the stochastic steady-state performance of open-loop phase esti-
mators in terms of “loop bandwidth” is original to this research.
CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 54

0 , which is the error of the estimator given in (4.22), may be

Then the random variable θ̂e,k
given by !
0
λ01,k
θ̂e,k = arg zk0 = tan −1
. (4.59)
AN + λ02,k
Following precisely the same arguments posited in section 4.2.2, we jump to the approx-
imation
0
λ01,k
θ̂e,k ≈ . (4.60)
AN + λ02,k
0
Again, we now make the assumption of independence between θ̂e,k and the input noise and
write h i
h i E λ01,k
0 h i .
E θ̂e,k ≈ (4.61)
AN + E λ02,k
Clearly, from observation of (4.57) and (4.58), the expected values of λ01,k and λ02,k are zero.
So, as expected, the error is zero mean.
Concerning the variance of the high SNR estimator, we may write

02
λ02
1,k
θ̂e,k ≈ . (4.62)
A2 N 2 + λ02 0
2,k + 2AN λ2,k

Again, assumed independence allows

h i h i
h i E λ02
1,k E λ02
1,k
02 h i h i = h i .
E θ̂e,k ≈ (4.63)
A2 N 2 + E λ02 0
2,k + E 2AN λ2,k A2 N 2 + E λ02
2,k

We note that
!2 !2
X
N X
N X
N X
N
λ02
1,k = cos θ2
nq,i + sin θ 2
nd,i − sin θ cos θ nd,i nq,i
i=1 i=1 i=1 i=1
h i
=⇒ E λ02
1 = N σn2 (4.64)
!2 !2
X
N X
N X
N X
N
and λ02
2 = cos θ2
nd,i + sin θ 2
nq,i + sin θ cos θ nd,i nq,i
i=1 i=1 i=1 i=1
h i
=⇒ E λ02
2 = N σn2 . (4.65)

Thus
h i N σn2
02
E θ̂e,k ≈ . (4.66)
A2 N 2 + N σn2
A2
As in section 4.2.2, letting ρ = = Eb /No yields
2σn2
h i 1
02
E θ̂e,k ≈ . (4.67)
2N ρ + 1
CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 55

Finally, since under high SNR we may assume 2N ρ  1, we write

h i 1
02
E θ̂e,k ≈ . (4.68)
2N ρ

We note that this variance could be obtained by considering an efficient k estimator’s equal-
ity with the Cramer-Rao bound [33], an aspect of estimation theory not explored in this
research. Comparing (4.68) with (4.52), we see that the only difference between them comes
from squaring loss.
So the controversy that arose in section 4.2.1, that a high SNR assumption leads to
a different ML estimation structure than a low SNR assumption, has been solved. First,
the errors of both are shown to be zero mean. Further, by allowing ρ → ∞ in (4.52),
it is clear that the second-order statistics also converge to the same values. Clearly then
the estimation structure suggested by (4.20) and implemented in Figure 4.6, which was
derived from a low SNR approximation, is still nearly optimal under high SNR conditions.
Moreover, we have shown that the only way to improve upon the squaring structure is to
implement decision feedback as shown in Figure 4.7 where, under low noise assumptions
dˆk = dk . In addition, we have shown that the squaring estimation structure suggested
by (4.20) is stochastically equivalent to the generalized Costas loop and that the decision
feedback structure suggested by (4.22) is stochastically equivalent to a Costas loop with
hard-limiting in the I-arm. Finally, we have shown that there is an analogous loop bandwidth
for this class of estimators given by (4.53).

4.2.4 Modified First-Order ML Phase Estimator

Consider now the situation where the carrier phase θ is not constant for the length of time
over which N observations of rk are taken. Assume rather that there is a residual carrier
frequency offset in the received symbols rk , which corresponds to imperfect downconversion.
This condition is almost certain to be the case in any realistic radio receiver. Intuitively, we
would now no longer expect either of the structures proposed in section 4.2.1 to track the
incoming phase with zero error. For this reason alone, we choose to call these first-order
carrier recovery structures since a first-order PLL has precisely this same property. That is,
a first-order PLL cannot track a frequency step, or equivalently a phase ramp, with a zero
mean steady-state error. As with the first-order PLL, then, we do not consider the first-
order estimation structures suggested and analyzed in the preceding three sections to be

k
A ML estimate, when obtainable, is considered to be efficient. Put very simply, the estimator structure
suggested by (4.22) is a simple averager of Gaussian random variables, i.e., receiver decision statistics, with
variance 12 No /Eb . We would intuitively expect this variance to thus decrease by N, the number of statistics
averaged.
CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 56

Figure 4.8: Modified first-order ML phase estimation structure, suitable for the conditions
of “small” residual frequency offset

applicable to the general task of suppressed-carrier recovery. Alas, the foregoing discussion
has not been in vain, however. In sections 4.2.4 and 4.2.5 we undertake a discussion of
modifications to the preceding basic estimation structures that allow them to operate in
the realistic conditions of residual frequency offset.
Assume now that the carrier phase at time kT may be represented by

θk = θo + 2πf kT (4.69)

where f denotes the frequency offset and θo is some initial carrier phase. Similarly, the
phase at time (k − i)T is related to the phase at time kT by

θk−i = θk − 2πf iT . (4.70)

As in section 4.2.1, we are still interested in estimating the phase using N past observations
of received samples, r = {rk−1 , rk−2 , ..., rk−N }T where T denotes the transpose operator.
Now however, θ is not constant, and we must therefore attempt to come up with an estimate
of θk . Specifically, we wish to find
 
p r | θ̂k,M L = Max p (r | θk ) . (4.71)
θk
CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 57

Divsalar and Simon in [41] show that by averaging out equiprobable data and averaging
over a uniform frequency shift distribution from −fmax to fmax , where fmax corresonds
to maximum expected value of doppler shift, we may express the argument to the max
function in 4.71 as
4
p (r | θk ) = E [p (r | φ, θk , f )]
fZmax
F Y
N  n o
= df cosh α< Rk−i e−jθk (4.72)
2fmax i=1
−fmax

4
where F and α are as defined in 4.13 and Rk−i = rk−i ej2πf iT . Thus, to obtain, θ̂k,M L we
must find 
∂p (r | θk ) 
 =0. (4.73)
∂θk θk =θ̂k,ML

Using a low SNR assumption, the solution to 4.73 is given in [41] as

v
u N
u X
u r 2 sinc (4fmax iT )
u
u i=1 k−i
ej θ̂ML,k =u
u  N  (4.74)
u X 
t 2 
rk−i sinc (4fmax iT )
 
i=1

4 sin πx
where sinc(x) = πx , which suggests the receiver structure outlined in Figure 4.8. Note
the addition of the 180o phase jump detector.∗∗ Without this, the detector will make an
error each time the input signal constellation crosses the imaginary axis. This is due the
fact that the combination of the square root and squaring operations limits the range of the
phase estimate to (−π, π], whereas the actual phase is continuous around the unit circle in
the complex plain. This phase jump detector alleviates this problem (with a small amount
of degradation due to noise) by detecting crossings of the imaginary axis and multiplying
the phase estimator by ±1 appropriately.
We denote the structure suggested by (4.74) as the modified first-order ML phase es-
timator since it contains no explicit frequency estimation. A range of possible frequency
offset must be known a priori, and therefore, in an arbitrary frequency offset, this estima-
tion structure would clearly be suboptimal. In addition, those who proposed this scheme in
[41] do so only with the caveat that it will work when the residual frequency offset is fairly
“small.” We therefore liken this structure to the second-order PLL, built with a loop filter
∗∗
The feedback structure seen at the output of the phase jump detector is depicted as a feed forward
structure in [41]. The feed forward structure was clearly the result of an editing error in this excellent
source, and was discovered and corrected as part of this research. Figures 4.8 and 4.9 depict the jump
detector correctly.
CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 58

Figure 4.9: Second-order ML phase estimation structure with dedicated frequency estima-
tion, suitable for the conditions of “large” residual frequency offset

without infinite DC gain. As noted in Chapter 3, such a second-order PLL is better de-
scribed as a modified first-order PLL since it cannot completely compensate for a frequency
offset with a non-zero steady-state error as can the canonical second-order PLL defined by
(3.9). A better match to this canonical second-order PLL is presented in section 4.2.5.

4.2.5 Second-Order ML Phase Estimator

In the final open-loop structure examined in this work, shown in Figure 4.9, we add an open-
loop frequency estimation scheme to the structure outlined originally in Figure 4.6. Note
that this too has the 180o phase jump detector to reconcile the periodic phase ambiguities
arising from a constant rotation in the input constellation. The functionality of Figure 4.9
precisely parallels that of the ML estimator in 4.6 except that a structure that attempts
to estimate the change in the (squared) phase, 2∆θ, is added in parallel. The estimates of
2∆θ are filtered by a lag filter with the transfer function
1−β
F (z) = (4.75)
1 − βz −1
CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 59

where β, though not explicitly defined in [41], is assumed to be a constant relatively close
to but less than one and controls the “memory” of the estimate integration. Finally, note
that for high SNR operating conditions, the squaring operations in both the modified first-
order and second-order phase estimators can be replaced by a decision feedback scheme as
suggested in Figure 4.7.

4.2.6 Simulation Study of Open-Loop Phase Estimators

Before considering a crucial application of open-loop ML estimation, namely adaptive equal-

ization, we present a few results from a simulation study on the performance of the esti-
mation structures derived in the preceding sections. Most importantly, this validates the
theoretical results predicted in sections 4.2.2 and 4.2.3. In addition, it gives a high level view
of the type of conditions under which the proposed estimation schemes can be expected to
work.
Figures 4.10 and 4.11 completely validate, via simulation, the expected noise perfor-
mance of both first-order estimators as given in (4.52) and (4.68). Figure 4.10 shows the
simulation results of the variance in the phase estimator given by (4.20), in which a squaring
operation is used to remove the modulation. Results are plotted versus N , the number of
symbols averaged over, and for Eb /No ranging from 0 to 24 dB. The solid lines show the
variance predicted for this estimator under the high SNR assumption given by (4.52), which
includes squaring loss. Note that the match is almost perfect down to about 0 dB where, as
expected, the high SNR approximation is no longer valid and the estimator’s performance
degrades.
Figure 4.11 shows similar results obtained for the decision feedback phase estimator
suggested by (4.22). Again, results are plotted versus N and for a range of Eb /No from 8 to
24 dB. Simulations were run for 0 and 4 dB, but the high SNR assumption in these cases
was violated and the estimator failed catastrophically. This makes perfect sense since the
estimation structure of Figure 4.7 is based upon making correct decisions of the transmitted
bits, i.e., dˆk is assumed to be equal to dk . Clearly under low SNR conditions, this assumption
fails with ever higher probability.

4.2.7 Application of Open-Loop Estimation to Receivers Employing Adap-

tive Equalization

Many receiver architectures that employ adaptive equalization require coherent demodula-
tion. Many equalization strategies employ the principle of rapid adaptation in the attempt
to create the optimal filter for an arbitrary and rapidly changing channel environment. It
CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 60

Figure 4.10: Variance of output phase estimate of first-order phase estimator, using squaring
operation, for a variety of Eb /No . Lines denote theoretical performance predicted by (4.52),
asterisks denote simulation points

Figure 4.11: Variance of output phase estimate of first-order phase estimator, using decision
feedback, for a variety of Eb /No . Lines denote theoretical performance predicted by (4.68),
asterisks denote simulation points
CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 61

is as yet quite unclear how equalization schemes interact with closed-loop carrier recovery.
Consider, for example, the steepest descent algorithm defined in the normalized least
mean squares (NLMS) update equation
µe∗k
wk+1 = wk + xk (4.76)
xH
k xk + δ

where xk = (xk , xk−1 , · · · , xk−N +1 )T is a vector of current and previous input samples to
the filter w of length N , ek = dk − rk is the filter error at time k, rk = wH
k xk and dk is the
desired symbol or an estimate thereof, µ is an update constant between 0 and 2.0, δ is a
small constant to avoid division by zero, H is the Hermitian transpose operator, and wk is
an iterative solution to the well-known Wiener-Hopf equation

wopt = R−1
xx,k pk (4.77)

where Rxx,k is the autocorrelation matrix of the input samples, xk , and pk represents the
correlation of the input vector to the desired symbol.
The NLMS update is sensitive to the phase of the error ek and therefore will, as we might
expect, attempt to correct for a residual phase offset in the input signal constellation. In fact,
assuming perfect frequency coherence, i.e., zero residual frequency offset, the performance
of the equalizer updated with (4.76) will perform as well (in AWGN) with or without a
coherent phase estimate. The performance quickly degrades, however, in the presence of
even small frequency residuals. We would therefore be inclined to include coherent carrier
recovery when implementing this type of receiver. The problem is that coherent carrier
recovery will also attempt to correct for the instantaneous phase seen by the receiver.
The interaction between these two feedback mechanisms is, at best, unclear and has been
observed in simulation to produce catastrophic failure of the equalization system, and in
bit error performance thereby.
One solution to this problem is to employ adaptive equalization with open-loop carrier
recovery, thereby avoiding this possibly pathological dual feedback effect. Figure 4.12 is
a special case of the structure proposed by [51] where we consider coherent demodulation
of BPSK with updates supplied to the filter weights at the symbol rate. This structure,
however, imposes a vitally important change on the update equation. The input vector,
used as a reference for the update, must also be rotated by the phase estimate. Specifically
we require
µek
wk+1 = wk + · e−j θ̂ML · xk (4.78)
xH
k xk +δ
for the update. Note that for the case of BPSK signalling the error term is no longer
considered to be complex because the expected phase of the rotated decision statistic is
CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 62

Figure 4.12: Open loop carrier estimation applied to an adaptive equalization receiver,
shown here employing decision direction only. This structure was adapted from [51] for this
research

Figure 4.13: Differentially coherent receiver employing adaptive interference suppression for
BPSK—a limiting condition of ML phase estimation
CHAPTER 4. COHERENT CARRIER RECOVERY STRUCTURES 63

zero. Therefore, we no longer expect any useful information in the imaginary part of the
error. In fact, simulation studies have shown a marked performance improvement gained
by using only the real part of the error versus the complex error. This makes sense since
we are effectively eliminating one-half of the noise in the error statistic that is fed back into
the adaptation equation. This naturally improves the performance of the adaptive filter.
Simulations have shown this structure of Figure 4.12 to work well when combined with
fractionally-spaced, adaptive equalization routines for CDMA systems. These results are
discussed later in Chapters 6 through 8.
A very important extension of the application discussed in this section is toward utiliz-
ing equalization within a differentially coherent receiver. In fact, a differentially coherent
receiver may be viewed as a limiting case of the ML phase estimator, namely that of N = 1,
wherein the previous symbol is used as the phase reference for the current one. This sug-
gests the demodulator structure given in Figure 4.13, which was developed independently as
part of this research effort. Note that now, the estimated phase present in (4.78) is replaced
with the complex sign of the previous symbol so that the update equation becomes
µek ∗
wk+1 = wk + · Sgn (rk−1 ) · xk . (4.79)
xH
k xk +δ

Note that the use of the sign function eliminates the one-half of noise power present in the
previous symbol estimate.

4.3 Chapter Summary

Thus concludes the discussion of phase-coherent carrier recovery techniques. Again the
interested reader is urged to consult the many references cited in this chapter as well as
in Chapter 3. Some of the relative merits of the carrier recovery schemes presented in this
chapter are explored later in Chapter 6. One additional portion of background material
is yet to be explored, the concept of automatic frequency control (AFC), which, for the
purposes of this research concerning PSK signalling, will be viewed as an acquisition aid
to phase-coherent structures. As such, this serves as a motivation for the topic of the next
chapter.
Chapter 5

Automatic Frequency Control

Automatic frequency control∗ may be applied to carrier recovery either as an acquisition

aid to phase coherent loops, such as a Costas loop for coherent PSK demodulation, or as a
non-coherent carrier tracker for differentially coherent PSK. In the former class of receivers,
AFC becomes especially important in time division multiple access (TDMA) schemes, where
the communication is of a bursty nature and quick synchronization is required to preserve
bandwidth.
We note that the theory surrounding AFC loops, especially pertaining to their perfor-
mance in noise, is much less mature than that for PLLs. Further, since an AFC loop is
a non-coherent device, the notion of phase jitter, applied universally as a figure of merit
to phase-coherent tracking loops, has no meaning whatsoever in the discussion of AFC. In
fact, an AFC loop will not in general settle to any particular phase relationship with the
input carrier but will randomly “walk” in phase, as it is perturbed by noise, with reference
to the input carrier. This behavior may be seen, for example, in Figure 5.7(A). A notion of
frequency jitter can be applied to AFC as a figure of merit, but the relationship of frequency
jitter and phase jitter is, at best, unclear. We therefore find it difficult to compare AFC
loops with the coherent tracking loops on the steady-state figures of merit established in
Chapter 3. Therefore, only acquisition time will be of primary interest. In section 5.4, where
we directly incorporate AFC as an aid to a Costas loop, we consider only the degradation
of the steady-state phase variance over that which was expected due the presence of the
AFC in the loop. Since no theory presently exists to predict this degradation, we appeal
primarily to simulation results as well as heuristic arguments in the discussion.
∗
Gardner in [27] makes a distinction between frequency-tracking loops (FTLs) and automatic frequency
control (AFC), stating that an AFC loop employs a “conventional” frequency discriminator. While this
reference is one of the very best and most commonly cited on the subject, this distinction is not employed
by other authors, nor aparently is this limited definition of AFC widely subscribed. We therefore make no
such distinction in this work.

64
CHAPTER 5. AUTOMATIC FREQUENCY CONTROL 65

5.1 Introduction to AFC

AFC loops bear striking similarities to PLLs. The only major difference between them and
the variety of control loops discussed thus far in Chapters 3 and 4 is the exchange of a
phase detector (PD) with a frequency difference detector (FDD). In section 5.2, we devote
some attention to the design of an FDD for the purposes of carrier recovery. Consider it
for the moment, however, to be simply a “black box” whose output is proportional to the
frequency difference, ∆ω = ωi − ωo , between the input carrier and the locally generated
reference. We then may view the AFC in the frequency domain as the linearized structure
shown in Figure 5.1. Note that the VCO, modeled as an integrator in the phase domain, is
now modeled simply as a gain in the frequency domain. The closed-loop transfer function
of the AFC loop, relating the input frequency, ωi , to the estimated frequency, ωo , is thus
given by
Ωo (s) Kf Ko Hf (s)
HAF C (s) = = (5.1)
Ωi (s) 1 + Kf Ko Hf (s)
where Kf is the gain of the FDD, analgous to Kd the gain of a phase detector, Ko is, as
defined before, the gain of the VCO, Ωi (s) = L {ωi (t)}, Ωo (s) = L {ωi (t)}, the Laplace
transforms of the instantaneous input and output frequencies respectively, and Hf (s) is the
transfer function of the loop filter.

Figure 5.1: Linearized, frequency domain model of AFC loop.

In order to simplify the on-going discussion, we utilize the second-order loop filter pro-
posed in [27] defined by
1
Hf (s) = (5.2)
sτ1 (sτ2 + 1)
where time constants τ1 and τ2 are carefully chosen to cause the poles of the loop filter,
Hf (s), to coincide with the poles of the closed-loop transfer function, HAF C (s). Note that
this filter is not shown to be optimal in any sense but is simply suggested to be more
effective in supressing noise than one with a single pole. Specifically, define τL to be the
CHAPTER 5. AUTOMATIC FREQUENCY CONTROL 66

time constant of the loop. Then choose

1
τ1 = Kf Ko τL and τ 2 = τL . (5.3)
4
These convenient definitions lead to a closed-loop transfer function of
1
HAF C (s) =  2 . (5.4)
1
2 τL s +1

It is further shown in [27] that the loop noise bandwidth, BL , for a loop so constructed, is
1
BL = , which allows us to conveniently rewrite the transfer function in terms of that
4τL
single design parameter
1
HAF C (s) =  2 . (5.5)
1
s+1
8BL
Relatively complicated theoretical results for the frequency estimate variance due to AWGN,
given the transfer function in (5.5) and the balanced quadricorrelator (BQ) frequency dif-
ference detector (FDD), introduced in section 5.2, are given in [27]. We point the interested
reader to this excellent reference, and do not repeat the results here in the interests of
brevity.
To continue our emphasis on discrete-time implementation, which began in Chapter 3,
we wish to find the digital loop filter, which, when implemented using digital devices or in
software on a digital computer, will obtain equivalent loop performance to the continuous-
time model. Direct application of the BZT to (5.5) yields,
1 + 2z −1 + z −2
HAF C (z) =  2    2   2 
1
4BL T + 1
2BL T + 1 + −2 1
4BL T +2 z −1 + 1
4BL T − 1
2BL T + 1 z −2
(5.6)
where T is the sampling interval in [sec]. The discrete-time loop filter is thus found by
1 HAF C (z) bf,0 + bf,1 z −1 + bf,3 z −2
Hf (z) = · = (5.7)
Kf Ko 1 − HAF C (z) 1 + af,1 z −1 + af,3 z −2
 2
1 1
where, defining αo = + ,
4BL T 2BL T

1 "  2 #
bf,0 = 1 1
αo Kf Ko af,1 = −2
2 αo 4BL T
bf,1 = " 2 #
αo Kf Ko 1 1 1
1 af,2 = − .
bf,2 = αo 4BL T 2BL T
αo Kf Ko
CHAPTER 5. AUTOMATIC FREQUENCY CONTROL 67

So the modeling of the loop filter for AFC is complete. Now what remains is determination
of the structure required to produce ∆Ω(s) shown in Figure 5.1. This motivates a discussion
of the design and implementation of FDDs.

5.2 Frequency Difference Detectors

Several discussions in the relevant literature concerning FDDs begin with the quadricorrela-
tor. Such a device is basically a simple Costas loop with a differentiator in the I-arm. These
produce an error term that is proportional to the frequency difference between the input
carrier and the local reference. The interested reader is urged to review [28], [29], and [27]
for introductory material regarding this type of FDD. Due to the ripple voltage created,
simple quadricorrelators are not suitable for FM demodulation, nor are they advantageous
for PSK demodulation. For this reason, we skip directly to a discussion of the balanced
quadricorrelator (BQ) as a starting place for the discussion of FDDs.

5.2.1 The Balanced Quadricorrelator

Consider, for the moment, an unmodulated, noiseless carrier s(t) applied to the input of
the BQ shown in Figure 5.2,
s(t) = A cos (ωi t + θi ) (5.8)
√
where A = 2P is the amplitude, ωi is the carrier frequency, and θi is a random phase
shift induced to the carrier. Note that now we take the phase of the carrier, relative to the
reference, to be constant and allow the frequency difference to be time varying, whereas
in the discussion of phase-coherent tracking, we assumed just the opposite. This signal is
then downconverted in quadrature by the reference signals as shown with frequencies equal
to ωo and amplitudes defined for convenience to be 2Ko . Assume further that the low pass
filters of Figure 5.2 are ideal, completely supressing the double frequency terms. Then

vi (t) = AKo cos (∆ωt + θi ) and vq (t) = AKo sin (∆ωt + θi ) (5.9)

where ∆ω = ωi − ωo is the frequency difference between the input and reference carriers.
Differentiators are used in both arms with gains equal to τd , yielding

v̇i (t) = −AKo ∆ωτd sin (∆ωt + θi ) and v̇q (t) = AKo ∆ωτd cos (∆ωt + θi ) . (5.10)

The difference, the error signal for the AFC loop, is thus given by

vd (t) = −A2 Ko2 ∆ωτd (5.11)

CHAPTER 5. AUTOMATIC FREQUENCY CONTROL 68

vi ( t ) v&i ( t )
LPF τ d d dt

s(t) 2 Ko cos ω o t Σ vd ( t )
-
LPF τ d d dt
vq ( t ) v&q ( t )

− 2 K o sin ω o t

Figure 5.2: Block digram of continuous-time balanced quadricorrelator (BQ).

which is directly proportional to the frequency difference.

Now consider an arbitrary bandpass signal s(t) of the form

s(t) = x(t) cos ωi t − y(t) sin ωi t . (5.12)

The error control signal vd (t) is now given by

h   i
vd (t) = −Ko2τd ∆ω x2 + y 2 + xẏ − y ẋ (5.13)

where now an implicit dependence upon time is assumed for x(t), y(t), ẋ(t), and ẏ(t). Insight
into (5.13) may be gained by considering a noiseless BPSK signal in which x(t) = Ad(t) cos θ
and y(t) = Ad(t) sin θ where d(t) ∈ {±1}. In this case, (5.13) simplifies back down to (5.11),
yielding a time averaged value for vd (t) of

vd (t) = −A2 Ko2 τd σd2 ∆ω (5.14)

4  
where σd2 = E d2 (t) = 1. This suggests that, unlike the PLL fitted with a conventional
multiplier phase detector, the AFC loop fitted with a BQ does in fact form a suitable
tracking device for BPSK signals, granting, of course, that it cannot provide a coherent
phase reference. Further consider the signal in which E [x(t)y(t)] = 0, which is true of all
PSK waveforms modulated with uncorrelated input data bits. The BQ output signal can
be shown to be
 
vd (t) = −Ko2 τd ∆ω σx2 + σy2 . (5.15)

We note that while the BQ is a suitable FDD for MPSK signalling, it cannot, like the
discrete-time PLL proposed in Chapter 3 or the Costas loop proposed in section 4.1.4, be
implemented digitally via baud rate sampling. The 360o /M discontinuities from sample
to sample in such a system cause simply too much jitter for the loop to acquire properly.
CHAPTER 5. AUTOMATIC FREQUENCY CONTROL 69

Due to this, we restrict the application of the BQ to analog FDD implementations only.
Therefore we look for an alternative FDD structure suitable for baud rate implementation.
A robust and simple to understand example of such a structure is angle-doubling AFC.

5.2.2 Angle Doubling AFC

It is worth noting that many FDDs can be proposed that are suitable for baud rate digital
implementation—the only absolute requirement being the necessity for an odd symmetry
of the error response to a frequency difference and the ability of the loop to remove the
modulation in the signal. The richest known coverage of digital FDDs may be found in [26],
which, though in our opinion it lacks sufficient justification for many of the results presented,
is still widely regarded as an important seminal paper on the topic of AFC, especially as
it relates to carrier recovery for digitally modulated signals and digital components. We
mention one structure proposed therein, viz., angle doubling AFC (ADAFC), particularly as
it serves to provide 1) an example of an FDD structure suitable for baud rate implementation
and 2) a motivation for the application of the AFC loop for an acquisition aid to a Costas
loop.
Consider the BPSK signal defined by

s(t) = Adk cos (ωi t + θi ) (5.16)

where A, ωi , and θi are as defined in (5.8) and dk is the data bit at time t = kTb , dk ∈ {±1}.
Once again, assume the down conversion process and lowpass filters are ideal. Thus the
quadrature samples may be represented as

ik = AKo dk cos (∆ωkTb + θi ) and qk = AKo dk sin (∆ωkTb + θi ) (5.17)

where, as before, ∆ω = ωi − ωo , the frequency difference between the input and local
reference carriers. Then

xk = i2k − qk2 = A2 Ko2 cos (2∆ωkTb + 2θi ) (5.18)

and
1
yk = ik · qk = A2 Ko2 sin (2∆ωkTb + 2θi ) (5.19)
2
where, as expected, the modulation had been removed in these control signals. It is impor-
tant to point out here that, with the exception of hard-limiting in the I-arm, yk precisely
represents the phase error signal for a Costas loop. To emphasize this fact, yk is also denoted
eφ,k in Figure 5.3.
CHAPTER 5. AUTOMATIC FREQUENCY CONTROL 70

Figure 5.3: Block digram of discrete-time implementation of Angle-doubling AFC (ADAFC)

proposed in [26].

Now the generalized structure of Figure 5.3 as proposed by [26] includes summers both
for the inputs to the discriminator and to the input of the loop filter Hf (z). Clearly these
summers could be used to perform some sort of short term averaging for the purpose of
noise supression. We choose, for the purposes of this discussion, to move noise suppression
considerations into the design of the quadrature matched filters as well as that of the loop
filter and take, for the sake of simlicity, M = N = 1. Then the error signal out of the
discriminator becomes
1
ef,k = xk−1 · yk − xk · yk−1 = A4 Ko4 sin (2∆ωTb ) . (5.20)
2
Thus ADAFC provides a sinusoidal frequency error characteristic, similar to the phase
error characteristic of a PLL or Costas loop. This is a fundamentally different result than
obtained from the BQ, which was completely linear. Nevertheless, under the fairly easy
1
assumption that ∆ω is less than about 10 of the baud rate, i.e., ∆ω < 0.1/Tb , we may
approximate this error as
ef,k ≈ A4 Ko4 2∆ωTb (5.21)

which, with the exception of a sign difference which must be taken into account, bears
striking similarity to the frequency error produced by the BQ in (5.11) for an unmodulated
input signal.
Finally, to complete the design of the loop shown in Figure 5.3, we include the loop filter,
Hf (z), for which the design proposed in section 5.1 would be appropriate. We include also
the NCO, modeled once again as a simple phase accumulator. Note the addition of a
two sample summation at the input of the NCO. This is used simply to place a zero at
CHAPTER 5. AUTOMATIC FREQUENCY CONTROL 71

z = −1, which is lost upon modeling the NCO as an accumulator versus a strict integrator.
Obviously, the presence of this zero at f = Rb /2[Hz] would provide some additional noise
immunity for the loop.

5.3 Acquisition Time of AFC loops

As mentioned at the beginning of this chapter, no theory presently exists to predict the
aquisition behavior of AFC loops, much less the acquisition time. Before moving on to a
singularly important application of AFC loops in section 5.4, we believe it is appropriate
to show the type of behavior that might be expected from them. To do this, we appeal
strictly to simulation results to suggest typical performance in this all-important matter of
acquisition time.
Figure 5.4 is our attempt at showing the type of acquisition times to be expected relative
to a conventional phase-coherent tracking loop. Here acquistion time, in samples, is plotted
versus the normalized loop bandwidth of the respective loops. Normalization is with respect
1
to the initial frequency offset, fof f = 2π ∆ω(0), of an unmodulated, noiseless carrier, which
served as the input to the systems in this example. It is important to note that two
different initial offsets were used, however. The initial offset for the AFC loop was set
to be 10−1 Rb , whereas for the Costas loop, it was set to be 10−2 Rb because the Costas
loop literally could not capture the carrier at 10−1 Rb offset. Also note that the AFC loop
fails utterly for sufficiently large loop bandwidths, denoted BL,f . We could not obtain
a result for acquisition when BL,f was greater than 10−1 Rb . We believe this behavior
would not necessarily be true for analog implementations of AFC loops. Nevertheless,
we felt that acquisition times approaching 10 samples were sufficient for this illustration.
Acquisition time in each case was defined as the sample beyond which the frequency error
of the respective loop was less than 10−3 Rb .
The most notable aspect of Figure 5.4 is that acquisition time appears to be log-linear for
both types of loops. For the AFC loop, the slope is negative one except as the acquisition
time begins to get quite small. From this we may postulate, in the absense of theory,
that acquisition time is inversely proportional to AFC loop bandwidth for frequency offsets
larger than the said bandwidth. For the Costas loop, on the other hand, the slope is steeper
than -3 for all of the obtainable points. This suggests that for a phase-coherent loop, the
frequency acquisition period is inversely proportional to the cube of loop bandwidth.
These intuitive results allow us to further suggest overarching principles in the design
of the receiver for digitally phase modulated signals. For a fixed Costas demodulator loop
bandwidth, BL,c , (fixed perhaps by the constraint of allowable phase jitter), we expect
CHAPTER 5. AUTOMATIC FREQUENCY CONTROL 72

Figure 5.4: Simulated acquisition time in samples for BQ-AFC and Costas loops versus
normalized loop bandwidth of respective loops; ∗ mark simulation points for AFC loop; ×
mark simulation points for Costas loop.

rapid frequency acquisition when the frequency uncertainty (initial offset) is less than or
approximately equal to BL,c . Under such conditions, there is little or no motivation to
implement AFC. If the uncertainty is larger than this, however, the Costas loop runs the
risk of requiring an exponentially increasing length of time to acquire the input frequency,
and an AFC loop or some other acquisition aid ought to be employed to assist the Costas
demodulator. This is the subject of the next section.

5.4 Costas Loops with AFC-Aided Acquisition

There are a variety of methods to aid a Costas loop with frequency acquisition when the
uncertainty is large. One way is to perform a “slow” frequency sweep of the local oscillator
until the Costas loop acquires. The implementation of such a scheme is highly dependent
upon the design constraints of a particular receiver, and, since a mobile receiver might not
be constructed with such a mechanism, we do not give attention to this scheme in this work.
Another notion might be to utilize two different loop filters, one with a wide bandwidth and
the other one narrow. At initial start up, utilize the wider bandwidth filter and then, upon
detection of course acquistion, switch to the narrower bandwidth filter for tracking mode.
It might be added that an AFC loop could also be utilized in place of the wider bandwidth
Costas loop in this scheme. The notion of switching between two loops is a rather simple
CHAPTER 5. AUTOMATIC FREQUENCY CONTROL 73

one. Gauging the performance of such a scheme amounts quite simply to examining the
performance of two different loops, which has already been accomplished in this chapter
and in Chapter 4. For this reason, we do not give additional attention to this strategy
in this work. A final approach, however, might be to combine the AFC and Costas loop
into one single feedback structure. This is an interesting approach and one that we explore
in some depth in this section. In section 5.4.1, we first describe the proposed combined
carrier recovery structure. Then in section 5.4.2, we examine the very crucial impact of
the combined structure on phase jitter, which is the important figure of merit for coherent
carrier recovery.

5.4.1 Description of Combined Costas Demodulator with AFC for Ac-

quistion Aiding

Consider, for the moment, the heuristically motivated system in which a PD and an FDD
are both incorporated into a carrier recovery loop as shown in Figure 5.5. Such a scheme
was proposed as early as 1977 in [28], wherein we merely sum the filtered error outputs of
the two detectors. Intuitively this ought to work. During acquisition under the conditions
of a large frequency uncertainty, the Costas loop portion will produce a wildly oscillating
phase error, eφ (t). Upon filtering with the quite narrow Costas loop filter, F (s), this error
will not contribute significantly to the integration performed by the VCO. The frequency
error, ef (t), on the other hand, will start out large and have a strong DC component,
assuming the FDD is constructed properly. Thus this error signal will dominate during the
acquisition period and, as the AFC portion was designed to do, bring the frequency error
down close to zero very rapidly. As the frequency error gets small, say less than the closed
loop bandwidth of the Costas portion, ef (t) will become small, eφ (t) will begin to dominate,
thus acquiring the phase, and VCO output will coherently track the input signal s(t).
Extending this intuitively satisfying structure to discrete-time, baud rate implementa-
tion is outlined in Figure 5.6. Here we utilize the basic Costas loop structure with hardlim-
ited I-arm as the PD. The digital loop filter, F2 (z), defined for instance by (3.17), would
be an appropriate choice. For the FDD, we choose to implement ADAFC as discussed
in section 5.2.2. A loop filter, Hf (z) as defined by (5.7) would form another appropriate
choice. Note, once again, that we keep the two-term summation at the output of the AFC
loop filter to preserve that high frequency zero at z = −1, which was lost in modeling the
NCO as an ideal accumulator instead of an ideal integrator.
CHAPTER 5. AUTOMATIC FREQUENCY CONTROL 74

Figure 5.5: Receiver block diagram suggesting the combination of AFC with a Costas Loop.

Figure 5.6: Block diagram of discrete-time implementation of second-order Costas demod-

ulator with a ADAFC loop for acquisition aiding.
CHAPTER 5. AUTOMATIC FREQUENCY CONTROL 75

A. AFC loop only

Qk, quadrature
1

sample value
0

-1
0 100 200 300 400 500 600 700 800 900 1000
k, bit number

B. Costas loop only

Qk, quadrature

1
sample value

-1
0 100 200 300 400 500 600 700 800 900 1000
k, bit number

C. Costas loop aided by AFC

Qk, quadrature

1
sample value

-1
0 100 200 300 400 500 600 700 800 900 1000
k, bit number

Figure 5.7: Typical acquisition behavior for three styles of carrier trackers: (A) An non-
coherent AFC loop; (B) An unaided Costas loop; C. The same Costas and AFC loops
combined; Input is BPSK signal with Eb /No = 12 dB.

5.4.2 Performance Issues for Combined Costas Demodulators with AFC

To give an impression of just how this type of scheme might work under realistic working
conditions, we show an example of the acquistion behavior of three different types of loops
in Figure 5.7. These are A) an AFC loop acting alone; B) a Costas loop acting alone; and C)
the same AFC loop and Costas loop acting together as outlined in Figure 5.6. In each case,
the values for the quadrature samples, qk , are shown versus sample number, k. A BPSK
modulated signal with Eb /No = 12dB is applied to each of the structures for which the
initial frequency offset is set to be 10−1 Rb , which is the sample rate of the carrier recovery
1
portions of the loops. Costas loop bandwidth is 20 Rb , or one-half the initial offset, and
1
AFC loop bandwidth is 200 Rb . Given the careful selection of these loop parameters, the
combined loop was able to exploit the best features of both the Costas and AFC loops.
The quadrature sample values, qk , were chosen to compare the performance of each
loop on equal footing. For the AFC loop there is no phase error estimate, but for a Costas
CHAPTER 5. AUTOMATIC FREQUENCY CONTROL 76

loop with a hard-limited I-arm as is performed here in parts (B) and (C) of Figure 5.7, the
quadrature samples are equal to the phase error samples, eφ,k , assuming no bit errors or
equivalently a relatively high input SNR. For a phase-coherent recovery scheme, we would
expect that qk = eφ,k would go to zero upon phase acquisiton for BPSK. Note that part (A)
of Figure 5.7 shows just what is meant by the AFC loop being an essentially non-coherent
device. Note that after only a couple of oscillations in the phase, the AFC loop quickly
settles down but it never completely acquires phase. For this reason, even at steady-state,
the phase error randomly walks around due to random noise perturbations. The frequency
error is essentially zero, but since the AFC loop is not responsive to phase at all, there is
no coherent phase reference being produced. These random phase perturbations can have
a pathological impact on the combined loop, and this important aspect will be discussed
shortly.
Now in part (B) of Figure 5.7, we see the behavior of the Costas loop acting alone. Note
that since the initial offset is twice the Costas loop bandwidth, acquisition takes a relatively
long time, about 500 symbols. This is possibly an unacceptably long acquisition period,
and as noted in section 5.3, the Costas loop has an acquisition time that is approximately
proportional to the inverse of the cube of the loop bandwidth normalized to the offset.
In other words, if we had chosen a significantly larger initial offset for this example, the
Costas loop might have taken orders of magnitude longer to acquire. Nevertheless, in this
academic example, the Costas loop does indeed acquire and, as expected, the quadrature
samples converge to a noisy zero mean process.
In part (C), we can see the effect combining these loops. The AFC loop, as expected,
quickly acquires the frequency. As soon as the frequency error becomes small, well within
the loop bandwidth of the Costas loop, the latter loop “takes over” and quickly acquires
the phase. Note that, as in part (B), the error goes to zero mean, indicating coherent
phase acquisition. The steady-state jitter in part (C) is similar to that of part (B), but
there is some degradation due to the additional noise allowed in via the AFC portion of the
loop. The nature of this degradation becomes the next important point to make. The loop
bandwidth parameters in the foregoing example were chosen carefully for the given noise
environment with good reason as we shall see.
What is the degradation of a Costas loop’s phase estimate due to the presence of the
parallel AFC loop? This is a most important question, and one that has not, to our knowl-
edge, been explored in the refereed literature previously. Figures 5.8 and 5.9 are the results
of simulations, performed as part of this research, that try to answer this question. These
figures attempt to show a very complex interaction, and thus we spend several paragraphs
CHAPTER 5. AUTOMATIC FREQUENCY CONTROL 77

attempting to explain these plots.

In Figures 5.8 and 5.9, the z-axes gauge the relative difference between the phase jitter
observed in simulation, including the presence of the parallel AFC loop as shown in Figure
5.6, versus that predicted by theory for the Costas loop acting alone. The theoretical values
for Costas loop jitter are predicted by (4.4) or, as is the case here, with I-channel hard-
limiting by (3.37), assuming sufficient input SNR. These z-axes are base-10 logarithms of
the observed values of phase jitter normalized to the theoretical value. The x and y axes
represent the base-10 logarithms of the Costas loop and AFC bandwidths normalized to the
symbol (sample) rate respectively. A z-axis result at or near 0 means that no significant
phase jitter was induced by the AFC loop. Jitter degradation equal to 1 on Figures 5.8
and 5.9 means that the AFC loop added an order of magnitude of jitter to the loop. Such
a degradation may or may not be acceptable depending on a particular application. Note
however, that degradations of 4 and 5 orders of magnitude represent catastrophic failure of
the coherent carrier recovery scheme. In these situations, the noise allowed in by the AFC
loop is dominating the performance of the combined loop. What seems to be happening is
that, for AFC loop bandwidths equal to or greater than the Costas loop bandwidth, the
AFC noise is causing random fluctuations in the estimated phase.† The Costas loop, limited
by its own relatively small loop bandwidth, is attempting but utterly failing to compensate.
The behavior shown in Figures 5.8 and 5.9 is actually rather counter-intuitive. A de-
signer might, as a first pass design of the combined loop, specify a large AFC loop bandwidth
to obtain rapid frequency acquisition and specify a narrow Costas loop bandwidth since the
expected frequency error will then be small. These figures show categorically that “it doesn’t
work that way!” In fact for a given AFC loop bandwidth, say 10−3 Rb in Figure 5.8, the re-
sulting phase jitter, given a Costas loop bandwidth of 10−3 Rb , is actually much worse than
if we chose a Costas loop with a bandwidth of 10−1 . In fact, what has happened in this
case is, as we described above, a complete loss of phase coherence and a complete inability
of the Costas loop, given its narrow bandwidth to acquire. While Figures 4.10 and 4.11 of
Chapter 4 show that the phase jitter of a Costas demodulator is inversely log-linear with
respect to loop bandwidth, an entirely opposite result is occuring here due to the presence
of the AFC loop.
So in a combined Costas and AFC loop, careful attention must not only be given to
absolute loop bandwidths for the purpose of jitter control, but attention must also be paid to
make the AFC loop bandwidth only a small fraction of that of the Costas portion. Moreover,
this consideration, i.e., just how small a fraction this ought to be, is also influenced by the
†
Thus essentially defeating any real notion of steady-state phase behavior and essentially making the
observation of phase jitter a meaningless statistic, except as it shows a complete loss of phaselock.
CHAPTER 5. AUTOMATIC FREQUENCY CONTROL 78

Degradation of steady state phase estimate due to AFC loop BW, Eb/No = 20dB

5
log10 of change in varaince of steady state phase est

0
−1
−1.5 −1
−2 −2
−3
−2.5
−4
log10 of Normalized Costas loop BW −3 −5 log10 of Normalized AFC loop BW

Figure 5.8: Degradation of phase output estimate for coherent carrier recovery with the
addition of an AFC loop; Eb /No = 20 dB; ∗ denote simulated points; interpolation used to
help show detail; all values logarithmic (base 10).
CHAPTER 5. AUTOMATIC FREQUENCY CONTROL 79

Degradation of steady state phase estimate due to AFC loop BW, Eb/No = 10dB

6
log10 of change in varaince of steady state phase est

−1
−1.5 −2
−2 −3
−4
−2.5
−5
log10 of Normalized Costas loop BW −3 −6 log10 of Normalized AFC loop BW

Figure 5.9: Degradation of phase output estimate for coherent carrier recovery with the
addition of an AFC loop; Eb /No = 10 dB; ∗ denote simulated points; interpolation used to
help show detail; all values logarithmic (base 10).
CHAPTER 5. AUTOMATIC FREQUENCY CONTROL 80

expected input SNR. Some intuition concerning this question may be gained by observing
the differences between Figures 5.8 and 5.9. Note that in a relatively high SNR case,
Eb /No = 20dB in Figure 5.8, there is roughly a log-linear line, parallel to the xy plane, to
the left of which the AFC contributes no more than approximately an order of magnitude of
degradation. This line starts around (x = −1, y = −2) and ends around (x = −3, y = −4).
So a fundamental design constraint for the combined loop in this SNR environment would
be one in which BL,c should be greater than or equal to 10 times BL,f . To the right of this
diagonal line we see a strong “knee” behavior, a very rapid rise in observed jitter as we scan
to the right. Beyond this, the AFC contributes so much noise to the loop that it utterly
fails to track phase at all.
In contrast, consider Figure 5.9 where the input SNR is a more moderate Eb /No = 10dB.
We see the same behavior except that it is accentuated, i.e., after a certain point the knee
behavior observed previously is now steeper. Note also that we have used a range of AFC
loop bandwidths an order of magnitude smaller than in Figure 5.8, ranging from 10−6 Rb
to 10−2 Rb . Here the diagonal line of “acceptable” degradation runs from (x = −1, y = −2)
to (x = −3, y = −5). So at the large bandwidth end, i.e., large Costas loop and AFC loop
bandwidths, the constraint of an order of magnitude smaller AFC bandwidth still holds.
But at the low end, i.e., small Costas and AFC loop bandwidths, we are now constrained
to hold the AFC loop bandwidth to be two orders of magnitude smaller.
Basically the FDD operation performed in AFC circuits, which is essentially one of
differentiation is a noisy process. Differentiation of a noisy signal naturally creates an even
more noisy signal. Thus we have observed the overarching principle that the loop bandwidth
of AFC must in general be smaller than that of a Costas loop. This is nowhere more true
than when these loops operate in parallel. We must choose bandwidths for the Costas and
AFC portions such that the AFC noise will not overwhelm the Costas tracking. Further, we
conclude that as input SNR goes down, where we would typically choose a narrower Costas
loop bandwidth to control phase jitter, the AFC bandwidth must become an ever smaller
fraction of the Costas loop bandwidth. It stands to reason that at some low SNR threshold
AFC loop bandwidth must of necessity become exceedingly small for such a combined loop
to work properly. Nevertheless, as noted in section 5.3, the frequency acquisition time
for AFC is only inversely proportional to its loop bandwidth. The Costas loop frequency
acquistion time is approximately inversely proportional to the cube of its bandwidth. Given
this observation, it may still be advantageous to implement AFC in parallel with Costas
tracking even in receivers designed for low SNR environments. We reiterate, however, that
the addition of AFC is motivated only in systems wherein the initial frequency uncertainty
CHAPTER 5. AUTOMATIC FREQUENCY CONTROL 81

is greater than the Costas loop bandwidth. In systems where the frequency uncertainty is
less than this bandwidth, we can expect the Costas loop to acquire frequency and phase
quite quickly on its own.
Chapter 6

Simulation Strategy

This chapter serves as a bridge between the substantive theory presented in Chapters 3
through 5 and the detailed simulation results presented in Chapters 7 and 8. Here we
endeavor to describe and justify the simulation parameters and procedures. First, we briefly
discuss the exchange of the figures of merit, developed in Chapters 3 through 5 for carrier
recovery structures, with BER in section 6.1. Then we give a detailed outline of the PSK
system under consideration in sections 6.2 and 6.3. The former describes the general system
parameters as well as the fading models assumed for Chapter 8. The latter section describes
the differences between the reverse and forward links assumed for this study. Finally, in
section 6.4, we discuss the selection process of candidate algorithms for carrier recovery
applicable to the challenges presented in sections 6.2 and 6.3. Here we present a few
preliminary simulation results to help justify these selections.

6.1 BER as a Figure of Merit

In Chapters 3 through 5, we presented a variety of carrier recovery structures suitable for
implementation in receivers designed to demodulate digital PSK modulated signals. In
these chapters, we examined the carrier recovery performance primarily by scrutinizing
acquisition time and steady-state phase jitter, noting that in general these performance
measures work against each other in the design of the digital receiver.
Beginning with this chapter, we now turn our attention to arguably the most important
figure of merit for the digital receiver, bit error rate (BER) or bit error probability. This
proves to be convenient as we now propose to examine the wide variety of digital receivers
under a wide variety of channel conditions, only the most benign of which, the additive
white Gaussian noise (AWGN), single-user channel, allows a few theoretical expressions for

82
CHAPTER 6. SIMULATION STRATEGY 83

carrier recovery performance. So instead of continuing to examine acquisition time and

steady-state phase jitter, we now take what was learned in chapters 3 through 5, apply it
to the design of receivers that are optimized for lowest possible BER, and measure their
effectiveness thereby.

6.2 CDMA System Overview

Based on the information provided in previous chapters, a virtually infinite number of
PSK receivers might be proposed, with various loop bandwidths, for any particular channel
environment. It is, however, not the intent here to present a nearly infinite number of
results. Some method must be found to eliminate a large number of potential carrier
recovery structures from consideration right from the start. The first step in this process is
clearly defining the type of system in which we will examine the performance of receivers
employing suppressed-carrier recovery.
As noted in Chapter 1, this work was from the very beginning motivated by the develop-
ment of adaptive despreading algorithms for cellular CDMA systems. We noted very early
in the research that the complex-weight fractionally spaced linear adaptive receiver (CW-
FSLAR) examined by our colleagues in [52] and [53] performed quite poorly in the presence
of even small carrier frequency offsets. Ideal coherent (unrealizable) carrier recovery had
been assumed up to this point. Since this research was sponsored in part by the Defense
Advanced Research Projects Agency’s (DARPA’s) Global Mobile Communications (GloMo)
program and working prototypical adaptive radios were part of the promised deliverables, a
method had to be found to make the adaptive receiver work in harsh real-world conditions.
This led quickly to the development of the differentially coherent adaptive receiver outlined
in Figure 4.13 in section 4.2.7. This receiver was found to be robust to frequency offsets
and a variety of channel environments. Nevertheless, a nagging question remained: Is this
the best we can do? Differentially coherent (DPSK) reception is well known to provide
about one dB of degradation in conventional BPSK receivers. However the degradation of
performance for a receiver employing adaptive equalization in a mobile, multiuser, CDMA
environment was unknown. Could we “buy back” some or all of this degradation due to
DPSK by employing robust carrier estimation techniques? Hence, all of the research in
chapters 3 through 5 was motivated.
The purposes of this research, therefore, have been two-fold throughout:

1. discover and test carrier recovery schemes for BPSK signals that may allow co-
herent detection and thus buy back some of the performance penalty of DPSK,
CHAPTER 6. SIMULATION STRATEGY 84

and
2. determine whether such techniques can reliably be applied to adaptive suppres-
sion of CDMA multiple access interference (MAI).

Failure to obtain either of these will still provide a worthwhile effort in showing that DPSK
is the best realizable system to employ in any number of particular channel environments.
As we narrow down the system specifications in section 6.2.1, many of the potential carrier
recovery structures possible are eliminated for one reason or another. This paring down
of structures to a manageable number is discussed in section 6.4 but, as a preview, the
harsh multiple access interference combined with relatively low bit energies seems to have
disqualified most of the contenders. In section 6.2.2, we devote considerable discussion to
the various channels of interest in this simulation work.

6.2.1 System Specifications

The general parameters for the systems that will be simulated are given in Table 6.1. Many
of the parameters have been borrowed from the actual DARPA contract which sponsored
this research. Simply put, this research examines a direct-sequence spread-spectrum (DSSS)
multiple access system. Binary phase shift keying (BPSK) is employed with differential
encoding of the data symbols and code-on-pulse spreading of length 15. Differential encoding
is necessary for the differentially coherent receivers examined and also proves useful in the
coherent receivers since the 180o phase ambiguity is eliminated. Differential encoding does
not automatically imply differential detection. Differentially encoded PSK can be detected
coherently and thus approach the BER, Pe , of “ideal” coherent PSK, i.e., where the absolute
sense of the received data is known. For this reason, we now permanently make a distinction
between BPSK, “ideal” coherent 2-ary PSK, and differentially encoded BPSK (DEBPSK),
a term used in [54] and elsewhere. For all simulations, the latter form of signalling is used.
The BER of BPSK in an AWGN channel is given by
s !
2Eb
Pe = Q (6.1)
No

where Q(x) is the Markum Q-function defined by

Z ∞
4 1
e−u du
2
Q(x) = √ (6.2)
2π x

Eb is the bit energy usually taken to be P Tb , or the signal power integrated over a bit
period, and No is the single-sided noise spectral density in [W/Hz]. The BER of DEBPSK
CHAPTER 6. SIMULATION STRATEGY 85

Figure 6.1: Bit error probabilities for three different binary signalling and recovery strategies

is given by s !
2Eb
Pe = 2Q (6.3)
No
or exactly double that in the former case. This is due quite simply to the fact that in
DEPSK, for any M , the receiver makes exactly two errors for every one error of ideal
PSK. The factor of two increase in BER is for all practical purposes negligible. Differential
detection of DEBPSK employs no coherent reference but rather the conjugate of the previous
received complex-valued symbol as a reference. Since this is presumably a more noisy
reference than that obtained by coherent reception, the system undergoes degradation in
BER. This reception scheme is henceforth denoted by differential BPSK (DBPSK). The
degradation, however, is very small for 2-ary signalling, as can be seen in Figure 6.1. The
BER for DBPSK is given by
1 Eb
Pe = e− No (6.4)
2
which is shown on Figure 6.1 as well. For higher-order modulations, such as 8-PSK or
QAM, this degradation approaches 3 dB and thus provides an ever greater motivation to
implement coherent carrier recovery.
Direct-sequence spreading is with length 15 Gold codes. A set of 15 such codes was
constructed by taking the one maximal sequence (M-sequence) available for this length and
modulo-2 adding it to its time reverse image. See [55] for the details of Gold sequence
generation. When synchronized, Gold codes provide optimally low mutual cross-correlation
between CDMA users, an absolute value of 1/15. When asynchronous, this cross-correlation
CHAPTER 6. SIMULATION STRATEGY 86

Table 6.1: General parameters for simulation study

System Rates:
Carrier Frequency fo = 2050 MHz
Chip Rate fc = 1.92 Mcps
Bit Rate Rb = 128 kbps
Sample Rate fs = 2fc = 3.84 MSps
Other Parameters:
Modulation code-on-pulse DSSS, BPSK—differentially
encoded data bits
Spreading Gain Gp = 15
Spreading Codes set of 15 Gold codes of length 15
Code assignment randomly assigned each Monte Carlo
simulation point
Pulse shaping none
FEC none
Frequency Uncertainty 2σf,1 = 1/40Rb = 3200 [Hz] for SOI
(Gaussian dist.) 2σf,I = 1/20Rb = 6400 [Hz] for SNOIs
Number of users K ∈ {1, 2, 4, 7, 10, 13}
Adaptive Filter µt = 0.5; during training
NLMS step sizes µd = 0.1; during decision direction
Key Simulation Assumptions:
Symbol timing assumed perfect knowledge
v for SOI—first 3 paths
u U
uX
AGC Implementation signal is normalized by t A where A is the
u u
u=1
known amplitude of user u before any fading, and
U is the number of users
Settling time all receivers are given 100 bits to synchronize and,
in the case of adaptive receivers, to train; after
this time all bit errors are counted toward BER
Observation interval each Monte Carlo point is run for 10,000 bits;
minimum of 100 Monte Carlo points per simulation

is bounded by 7/15 and thus these codes perform no better than we might expect from
random codes. Nevertheless, having the cross-correlation bounded by a known value is
still convenient. This short spreading gain is not terribly close to that of IS-95 but is
nevertheless quite comparable to that of spread-spectrum systems currently being deployed
in the Industrial, Scientific, and Medical (ISM) bands as well as systems employing the
IEEE 802.11 standard.
The chip rate, fc , is, of course, an integer multiple, 15, times the bit rate, Rb , due to
the code-on-pulse spreading with the codes described above. This rate is much higher than
the throughput of IS-95 but lower than those rates specified by 802.11. It is representative
for an ISM-band wireless modem. Fractionally-spaced sampling, i.e., two samples per chip,
CHAPTER 6. SIMULATION STRATEGY 87

was chosen as a benefit to the adaptive algorithms under examination. This leads to an
aggregate system sample rate, fs , of 3.84 MSps. Note that three samples per chip was
chosen in [52] and [53], previous works on this subject. The selection of two samples reflects
an attempt to reduce the DSP processing burden in real-time implementation of the systems
proposed by this work.
The carrier frequency, fo , of the system was chosen to be 2050 MHz, a band in which the
MPRG has an experimental license. Note that this is considerably higher than that of US
or European cellular systems, but it quite comparable to PCS systems now being deployed
as well as the 2400 MHz ISM-band. Since the simulations are carried out using complex-
baseband signal representation, the only affect fo has on the system is in the Doppler
frequency, fd , utilized in the fading models discussed in section 6.2.2 and in determining
“reasonable” frequency uncertainties observed at baseband, which we discuss here.
The frequency uncertainty, described by 2σf,1 for the SOI and 2σf,I for the SNOIs
in Table 6.1, is assumed to be a Gaussian random variable distributed with zero mean
and variance σf2 . The quantity 2σf is then equal to two standard deviations and sets an
approximate upper limit on the randomly generated frequency uncertainty, expected at
any particular Monte Carlo simulation point. The values chosen for 2σf are reasonable for
modest-priced oscillators that might be used in a mobile receiver. Consider an oscillator with
a specified frequency stability of 1 part per million (ppm). At 2050 MHz, this corresponds
to a residual baseband offset of 2050 kHz, which is approximately equal to 1/60Rb in the
simulations. We express the offset in terms of Rb since the processing for all of the receivers
considered is implemented at the baud rate. Consider, now a cellular CDMA signal at
900 MHz with the same oscillator, producing a baseband offset of 900 Hz. At an encoded
symbol rate of about 25 kHz, suitable for voice, this corresponds to approximately 1/30Rb .
So the choice of uncertainty equal to 1/40Rb for the SOI reflects a medium point between
the two situations—pessimistic for the former case and slightly optimistic for the latter.
When the users are asynchronous, the interferers are assigned random carrier frequencies
with two standard deviations equal to 1/20Rb . This was chosen since it seemed reasonable
that the frequency range of the interference would be more uncertain than that of the SOI.
Some simple form of automatic gain control (AGC) is assumed in the simulation environ-
ment. AGC was chosen primarily because the adaptive receiver algorithms were sometimes
observed to fail catastrophically if the input signal was too large with respect to a refer-
ence signal, given either by training or decision direction, of amplitude one. Since it is
reasonable that AGC would be implemented in any realistic receiver and we make the very
harsh requirement of only 100 synchronization/training bits, some type of AGC is quite
CHAPTER 6. SIMULATION STRATEGY 88

important. The amplitudes of the various users are normalized to the root sum of squares
of the amplitudes created randomly based upon the log-normal power variance. Specifically
1
A0u = qP Au (6.5)
U 2
u=1 Au

is the AGC-adjusted amplitude of the uth user, where U is the total number of users, and
1
Au = 10 20 σP Pu (6.6)

where A2u is assumed to have a log-normal distribution with standard deviation σP , and
Pu ∼ Normal (0, 1).
To sum up the various simulation parameters, we now describe the unfaded CDMA
signal. In complex baseband representation, it is given by

1 X
U
s(t) = rX Au bu (t − τd,u )cu (t − τd,u ) exp [j (2πfu t + θu )] + n(t) (6.7)
A2u u=1
u

where bu , cu , Au , fu , θu , and τd,u are the differentially encoded bit, the spreading code, the
amplitude, the baseband carrier frequency offset, the carrier phase, and the delay relative
to the SOI of the uth user respectively. U is the number of simultaneous users. Noise,
n(t) (which lies outside of the summation), is a complex valued Gaussian process given by


n(t) = nd (t) + jnq (t); nd , nq ∼ Normal 0, σn2 . The variance is calculated by the rule
s
A02
1 Ns No
σn = (6.8)
2 Eb

where A01 is as defined in (6.5), Ns is the number of samples per bit, 30, and Eb /No is the
bit energy to noise spectral density ratio. The bits and codes are represented by ±1.
The data bits are chosen randomly for each user and differentially encoded. In the
simulation program, the data bits are the output of a 32 stage maximal length sequence
generator, having a repeat period of 232 − 1 ≈ 4 billion bits. The codes are assigned ran-
domly each Monte Carlo simulation with the requirement that no two users get the same
code. Gaussian noise is generated using the modified Box-Muller transform of uniform
random variable created via the Wichman-Hill algorithm, techniques outlined in [56]. The
other parameters are assigned randomly each Monte Carlo simulation as well, but they
are assigned differently depending upon whether we are examining the forward link (syn-
chronous) or the reverse link (asynchronous). Therefore the discussion of these parameters
is put off until section 6.3.
CHAPTER 6. SIMULATION STRATEGY 89

Table 6.2: Simulation channels models

Veh spd # Ray Ray
Channel Eb /No [dB] [kph] rays Delays ampl. env.
AWGN fwd lnk ∈ {8, 10} N/A 1 0.0 µs 1.0 unity
rev lnk ∈ {8, 12}
Flat fading fwd lnk TBD ∈ {15, 45} 1 0.0 µs 1.0 Rayleigh
rev lnk TBD
Rayleigh 2-ray fwd lnk TBD ∈ {15, 45} 2 0.0 µs 1.0 Rayleigh
rev lnk TBD 5.0 µs 0.5 Rayleigh
COST-207: fwd lnk TBD ∈ {15, 45} 4 0.0 µs 1.0 Ricean
“Typical rural” rev lnk TBD 0.2 µs 0.7937 Rayleigh
0.4 µs 0.3162 Rayleigh
0.6 µs 0.1 Rayleigh
COST-207: fwd lnk TBD ∈ {15, 45} 6 0.0 µs 0.7071 Rayleigh
“Typical urban” rev lnk TBD 0.2 µs 1.0 Rayleigh
0.6 µs 0.7937 Gaus 1
1.6 µs 0.5 Gaus 1
2.4 µs 0.4 Gaus 2
5.0 µs 0.3162 Gaus 2
COST-207: fwd lnk TBD 15 6 0.0 µs 0.7071 Rayleigh
“Bad case rev lnk TBD 0.4 µs 1.0 Rayleigh
urban” 1.0 µs 0.7071 Gaus 1
1.6 µs 0.5657 Gaus 1
5.0 µs 0.7937 Gaus 2
6.6 µs 0.6325 Gaus 2
COST-207: fwd lnk TBD 15 6 0.0 µs 1.0 Rayleigh
“Typical Hilly” rev lnk TBD 0.2 µs 0.7937 Rayleigh
0.4 µs 0.6325 Rayleigh
0.6 µs 0.4472 Rayleigh
15.0 µs 0.5 Gaus 2
17.2 µs 0.2449 Gaus 2

6.2.2 Fading Models

The Fading models are described in some detail in Table 6.2. A total of seven distinct
models are assumed. In all but the first, AWGN, complex-valued fading waveforms of one
second duration are created for each path based on one of four probabilistic approaches.
Many of the rays are assumed to have a Rayleigh distribution. These are ones where the
excess delay is not greater than 0.5µs. This well-known Doppler spectrum is given by
A
Sτ (f ) = q τ < 0.5µs (6.9)
1 − (f − fd )2

where A is a complex Gaussian process of baseband bandwidth equal to fd , the maximum

Doppler frequency, and f ∈ {−fd , fd }. When a line-of-sight (LOS) component exists, a
CHAPTER 6. SIMULATION STRATEGY 90

Ricean spectrum is assumed and is modeled by

 
0.41
Sτ (f ) = A  q + 0.91δ (f − 0.7fd ) . (6.10)
2
2πfd 1 − (f /fd )

Two additional Doppler spectra, denoted Gaus 1 and Gaus 2 in Table 6.2, are proposed in
the COST 207 final report. These were developed for the European Groupe Special Mobile
(GSM) standard. [57] Let G(f ) denote a Gaussian spectral density, defined by
(f −f1 )2
−
2f 2
G(f ; A, f1 , f2 ) = Ae 2 (6.11)

where A is again a sample of a complex Gaussian process. For individual rays in which the
excess delay times exceed 0.5µs but are less than 2µs, we apply Gaus 1, given by

Sτ (f ) = G (f ; A, −0.8fd , 0.05fd ) + G (f ; A1 , 0.4fd , 0.1fd ) 0.5µs < τ < 2.0µs . (6.12)

When excess delay times exceed 2µs, the Gaus 2 spectrum, given by

Sτ (f ) = G (f ; B, 0.7fd , 0.1fd ) + G (f ; B1 , −0.4fd , 0.15fd ) τ > 2.0µs (6.13)

is used.
In the simulations, a Doppler spectrum for each ray is created, zero padded out to
1000 samples, and then transformed into the time domain via the inverse discrete Fourier
transform (IDFT). Thus we obtain a fading waveform of duration one second, sampled
at 1 kHz. Then several cubic polynomial interpolations are performed to upsample this
wave to the required 3.84 MSps rate. Note that these channel models allow statistical
variations in the amplitudes and phases of the received rays but do not allow variations
in the relative delays of each arriving component as would be expected in a truly dynamic
channel. Nevertheless, these models give a good idea of what performance to expect from
the realizable carrier recovery structures proposed in a nice variety of channel environments.
As may be noted from Table 6.2, we also select a set of Eb /No ratios at which we examine
each of the channels. Vehicle speeds are also selected. In each case, an attempt is made
to give as wide as possible range of impressions of receiver performance. Note that vehicle
speeds of 15 and 45 kph correspond to Doppler frequencies for the 2050 MHz carrier of 28
and 85 Hz respectively. Note that while these speeds are conservative, at a carrier frequency
of 900 MHz, consistent with cellular systems, these Dopplers reflect much faster speeds of
33.6 and 102 kph.
CHAPTER 6. SIMULATION STRATEGY 91

6.3 The Mobile Links

There are profound differences between the forward and reverse links that are assumed for
this work. A brief outline of these differences is given by Table 6.3. Our motivations for
investigating and presenting both links is to allow for the quite reasonable possibility that
certain receiver or carrier recovery structures might be best applied to only one of these
links. Since both are essential to cellular or mobile radio communications and since perfor-
mance parity between the links is required to maximize system capacity, an examination of
peculiarites of both the forward and reverse link is in order.

Table 6.3: Link-specific parameters for simulation study

Forward Link Reverse Link
Power variance equal powers
  for all users; Log normal distribution;
σp2 = 0 dB2 amplitudes Au = 101/20σp Pu
where σp2 ∈ {2, 20} dB2 and
Pu ∼ Normal (0, 1)
Carrier phases uniform dist. on [0, 2π)—all uniform dist. on [0, 2π)
users identical
  —randomly  assigned

Carrier frequencies ∼ Normal 0, σf,12
—all ∼ Normal 0, σf,12
SOI
 
users identical ∼ Normal 0, σf,I2
SNOIs
Symbol timing all users synchronous; all users have random delays;
mutual Xcorr is 1/15 Xcorr is bounded by 7/15
Fading single carrier; each path multiple carriers; each path
gets multiplied by one from each user gets multiplied
fading envelope by one fading envelope

6.3.1 The Forward Link

The forward link is the much more forgiving link in the simulation strategy. First of all,
we assume complete synchronization of the forward link. This means that there is only one
carrier for all CDMA users. There is therefore only one carrier phase to track regardless
of the level of MAI. The length 15 Gold codes, when synchronous, have an optimally low
mutual cross-correlation of 1/15. In addition, the multiple access signal is constrained to
have all users at equal powers. Thus the MAI is quite benign up to a fairly large number
of users.
Finally, since there is only one carrier, only one set of fading waveforms need be gener-
ated. The received forward link signal gets multiplied by only one set of fading waveforms.
CHAPTER 6. SIMULATION STRATEGY 92

The complex baseband forward link signal is thus given by

1 X X
U #Paths
s(t) = √ bu (t − τp)cu (t − τp)Ap δ (t − τp ) Mp (t) exp [j (2π (f1 t − fo τp ) + θ1 )] + n(t)
U u=1 p=1
(6.14)
where bu , cu , U , and n(t) are as defined in (6.7), Ap and τp are the deterministic amplitude
and time delay of the pth ray component as given in Table 6.2, f1 and θ1 are the frequency
offset and phase of the SOI (which are shared by all users), fo = 2050MHz, and Mp (t) is
the complex-valued fading waveform discussed in section 6.2.2.
The use of a forward link pilot channel is common in CDMA systems and is a key feature
of IS-95. We make no such assumption here as we are interested in the receiver’s ability to
perform coherent carrier recovery, relying solely upon the suppressed-carrier received signal.
Moreover, as pilot pollution, the corruption of reference tones by those of surrounding cells,
has become a noted weakness of IS-95, we believe it is helpful to further consider systems
that place no reliance on dedicated carrier references.

6.3.2 The Reverse Link

The reverse link assumed in this research combines all that is challenging in CDMA sys-
h i
tems. Power variances of 2 and 20 dB2 are examined with 2 dB being the “tight” power
control case, and 20 dB corresponding to a virtual lack of power control. In addition, the
codes are completely asynchronous, which increases the mutual cross-correlation. Carrier
frequencies and phases are also different for each user. This, as will be noted, can have
some deleterious effects on carrier recovery structures that employ M th-order nonlinearites
to remove modulation.
Finally, in contrast to the forward link, each user’s signal gets multiplied by mutliple
fading waveforms, denoted Mp,u . This ultimately makes the received complex baseband,
reverse link signal look like

1 X X
U #Paths
s(t) = rX Au bu (t − τu − τp )cu (t − τu − τp )Ap δ (t − τu − τp )
A2u u=1 p=1
u
×Mp,u (t) exp [j (2π (fu t − fo τp ) + θu )] + n(t) (6.15)

where all of the variables have been previously defined in either (6.7) or (6.14). Once again
the the noise process n(t) lies outside of the summations. Note that the consideration of
fading simultaneous with power variances may give pessimistic results, especially in the case
of loose power control.
CHAPTER 6. SIMULATION STRATEGY 93

6.4 Candidate Receivers Structures

At the start of the simulation portion of this work, we nominated a large number of potential
receiver structures that might be applied to receiving the SOI in the conditions described
above. In this section, we discuss the process of how the field was narrowed down from the
very large number of possible carrier recovery candidates, discussed in section 6.4.1, to the
more manageable number presented in section 6.4.3. Section 6.4.2 discusses this “weeding
out” process and presents some early simulation results that helped make the determination
of the final candidate algorithms and receiver structures.

6.4.1 Original Candidate Structures

Table 6.4 outlines a representative sampling of the original candidate structures of interest
in the simulation work. Parameters, especially loop bandwidth for Costas and AFC loops,
or, equivalently, N , the number of symbols over which the estimator is averaged, for open-
loop structures, were iteratively tweaked to test candidate algorithms. Table 6.4 shows the
best designs we were able to come up with for some of the various structures discussed in
Chapters 3 through 5. Most of these preliminary simulations were run in single-user AWGN
channel environments over a fairly low range of Eb /No ratios with a conventional matched
filter receiver. Since the performance of such a receiver is known from theory for both ideal
coherent and differentially coherent cases, we can see very clearly the BER impact of the
carrier recovery structures.
We provide diagrams of most of these candidate receiver structures in figures at the
end of this chapter where we denote all real scalars by single thin lines, complex scalars
by double thin lines, and vector values by bold lines. Note that decision feedback, or
equivalently, hard-limiting the I-channel of a Costas loop, while beneficial in forgiving high
Eb /No and low MAI conditions, was not employed in any of the receivers examined since
most of the conditions to be examined in Chapters 7 and 8 will not be so forgiving. Finally
note that “ideal” carrier recovery means “perfect knowledge of the phase of the first arriving
component of the SOI” is assumed as shown in Figures 6.6 and 6.12. While this is clearly
unrealizable, it ought provide a nearly best case performance in the experiments in fading
detailed in Chapter 8. It is conceivable, however, that the RAKE receiver, or a realizable
adaptive receiver could perform better than this ideal receiver in any particular fading
environment. This will, in fact, be observed in certain situations examined in Chapter 8.
As shown in Table 6.4, in many cases we implement a bank of similar carrier recovery
circuits with different parameters. For example, five different Costas demodulators differing
only in loop bandwidth were implemented for the conventional matched filter receiver; six
CHAPTER 6. SIMULATION STRATEGY 94

Table 6.4: Potential receiver structures with symbol rate processing

Description # Rxs Parameters Comments
Matched Filter Receivers:
Ideal coh. 1 Conforms to theory, 1 usr
Diff. coh. 1 Conforms to theory, 1 usr
Costas loop 5 BL,c = 10−3 Rb Acqu. time way too long
alone BL,c = 10−2.5 Rb ”
BL,c = 10−2 Rb ”
BL,c = 10−1.5 Rb pretty good
BL,c = 10−1 Rb very near ideal
Costas loop 6 BL,c = 10−3 Rb , BL,f = 10−5 Rb Very bad performance
w/ combined BL,c = 10−3 Rb , BL,f = 10−4 Rb ”
AFC loop BL,c = 10−3 Rb , BL,f = 10−3 Rb ”
BL,c = 10−2 Rb , BL,f = 10−4 Rb ”
BL,c = 10−2 Rb , BL,f = 10−3 Rb ”
BL,c = 10−2 Rb , BL,f = 10−2 Rb best but still not good
AFC loop 3 BL,c = 10−2 Rb , BL,f = 10−2 Rb Very bad performance
switched to BL,c = 10−2 Rb , BL,f = 10−1.5 Rb ”
Costas loop BL,c = 10−2 Rb , BL,f = 10−1 Rb ”
Modified 1st 5 N = 500 Very high complexity
order, open N = 168 high complexity
loop est. N = 50 reasonable complxty
N = 16 good performance
N =5 ”
RAKE Receiver:
3-Finger 1 Identical perf. to
diff. coh. diff. coh. MF in AWGN
Adaptive Filter Receivers:
Ideal coh. 1
Diff. coh. 1
Costas loop 3 BL,c = 10−2 Rb Acqu. time way too long
alone BL,c = 10−1.5 Rb okay performance
BL,c = 10−1 Rb Pretty good
Modified 1st 5 N = 500 Very high complexity
order, open N = 168 high complexity
loop est. N = 50 reasonable complexity
N = 16 good performance
N =5 ”
Second-order 6 N = 16 β = 0.9 fair performance
open-loop N = 16 β = 0.8 bad performance
estimator N = 16 β = 0.7 ”
N = 5 β = 0.9 good but same as Mod 1st
N = 5 β = 0.8 bad performance
N = 5, β = 0.7 ”
CHAPTER 6. SIMULATION STRATEGY 95

different second-order ML open-loop estimation structures were examined for the adaptive
receiver; etc. In both the conventional matched filter and adaptive receiver cases, we im-
plement ideal coherent carrier recovery to provide a baseline for the simulation results. We
also implement differentially coherent receivers in which no carrier recovery is performed.
The performance of these inherently realizable receivers will tell us whether carrier recov-
ery was worth the trouble since it is the performance of these receivers that we are trying
to better with coherent carrier recovery. In addition, a three-finger differentially coherent
RAKE receiver is compared side-by-side with the others.

6.4.2 The “Weeding-Out” Process

It could be seen from very early in the simulation study that many receiver structures were
simply not going to make the final cut. Costas loop receivers with bandwidths wider than the
frequency uncertainty were penalized severely by the 100-bit synchronization requirement.
This can be seen in Figure 6.2 where BER results for the widest three Costas loops are
plotted. The loop bandwidths are 10−2 Rb , 10−1.5 Rb , and 10−1 Rb respectively and improve
in performance in that order. Note how close the widest bandwidth Costas loop conforms to
theory. This shows just how relatively immune binary PSK is to phase jitter. No errors were
generated for 12 dB, but we have no reason to expect a sample point here not to conform
to theory as well. Narrower bandwidths, 10−3 Rb and 10−2.5 Rb , were also examined but
fared even worse than the top curve in Figure 6.2. The acquisition time for the narrowest
three Costas loops was judged to be far too long. The widest two, BL,c = 10−1.5 Rb and
BL,c = 10−1 Rb , were chosen for the final competition.
We chose to apply AFC in combination with the middle Costas loop, i.e., the one with
the Costas bandwidth, BL,c = 10−2 Rb, to see if AFC could buy back some of the BER
penalty incurred by the requirement of synchronization in 100 symbols. Figure 6.3 shows
the results of this portion of the preliminary investigation. As can be seen, widening the
AFC loop bandwidth, BL,f , incrementally improved the BER performance by improving
the acquisition time relative to the Costas loop acting alone. Still, at no Eb /No did the
performance of the combined AFC/Costas loop outperform the simulated DBPSK receiver.
All receiver structures employing AFC performed dismally for apparently one of two
reasons. When the AFC loop bandwidth was narrow relative to the Costas loop bandwidth,
as was suggested in Chapter 5, the acquisition time became too long. Thus a severe BER
penalty was incurred, just as with the narrowbandwidth Costas loops. When the AFC loop
bandwidth was widened, as was proscribed in Chapter 5, the acquisition time improved
dramatically and incrementally improved BER, but in these situations steady-state phase
CHAPTER 6. SIMULATION STRATEGY 96

Figure 6.2: Simulation results for bit error probabilities versus Eb /No for three Costas
demodulator structures; frequency uncertainty, 2σf = 1/40Rb ; simulated results for DBPSK
and theoretical DEBPSK performance also shown

Figure 6.3: Simulation results for bit error probabilities versus Eb /No for three combined
AFC and Costas loop carrier recovery structures; Costas bandwidth, BL,c is the same for all
three; frequency uncertainty, 2σf = 1/40Rb ; Costas loop acting alone shown for comparison;
simulated results for DBPSK and theoretical DEBPSK performance also shown
CHAPTER 6. SIMULATION STRATEGY 97

jitter became such a problem that BER was impacted anyway. If the frequency uncertainty
had been much higher or if a higher-order signal constellation was used that proscribed the
use of a wide bandwidth Costas loop, the addition of AFC might have been shown to be
beneficial. In no observed cases in the preliminary study, however, did AFC prove helpful.
A wide bandwidth Costas loop acting alone is a better choice given the constraints of the
system described in sections 6.2 and 6.3. Thus all AFC implementations were removed from
the final competition.
As a “sanity check” for the combined AFC/Costas loop, we examined a scheme where
the AFC loop was allowed to run alone during the synchronization period and then was
switched off in favor of the Costas loop after that. To our surprise, this switched scheme
fared no better than a combined AFC/Costas loop where the outputs of both loop filters
were summed. In fact, in some cases, it fared worse as shown in Figure 6.4. Reasonably
good performance was only obtained from the widest bandwidth, BL,f = 10−2 Rb , combined
structure. Note, however, that it still performs worse than DBPSK. We believe that the
inherently noisy differentiation process required by AFC limits its usefulness to high signal
energy situations, and the low Eb /No constraints in our simulations proved catastrophic for
AFC in terms of BER performance.
Early simulation results for the ML open-loop estimation schemes described in Chapter
4 proved quite illuminating. The modified first-order ML structure of section 4.2.4 proved to
be quite adequate for the range of frequency offsets encountered in the preliminary investi-
gation. In no case did the second-order structure of section 4.2.5, which utilizes a dedicated
frequency offset estimator, perform better than the former structure. We observed, in fact,
quite deleterious effects when increasing the value of N for the second-order structure, as
shown for N = 16, β = 0.9, in the highest curve in Figure 6.5. This can only be due to
the error of the frequency estimator propagating through the delay-line sum as shown in
Figure 4.9. This error can be minimized by making the “forgetting factor,” β, approach
unity. This dramatically increases the acquisition time of the loop, however, and the 100-
bit period synchronization requirement proved a handicap when this was tried. Since the
second-order ML structure provided no appreciable gains in the simulation conditions, we
removed it from consideration.
The values of N ∈ {500, 168, 50, 16, 5} set for the modified first-order structures directly

corresponded the loop bandwidths BL,c ∈ 10−3 , 10−2.5 , 10−2 , 10−1.5 , 10−1 × Rb of the
Costas loops examined above. As noted in Chapter 4, an equivalent loop bandwidth for
open-loop estimators may be defined as
Rb
BL,θ̂ = . (6.16)
2N
CHAPTER 6. SIMULATION STRATEGY 98

Figure 6.4: Simulation results for bit error probabilities versus Eb /No for two combined
AFC/Costas loop carrier recovery structures and two switched loop structures; Costas
bandwidth, BL,c = 10−2 Rb in all cases; frequency uncertainty, 2σf = 1/40Rb ; simulated
results for DBPSK and theoretical DEBPSK performance also shown

Figure 6.5: Simulation results for bit error probabilities versus Eb /No for four ML open-
loop carrier recovery structures; frequency uncertainty, 2σf = 1/40Rb ; simulated results for
DBPSK and theoretical DEBPSK performance also shown
CHAPTER 6. SIMULATION STRATEGY 99

Table 6.5: Final candidate receiver structures simulated and compared in future chapters
on the basis of BER performance
Description # Rxs Parameters
Matched Filter Receivers:
Ideal coh. DEBPSK 1

Diff. coh. DBPSK 1

Costas loop 2 BL,c = 10−1.5 Rb

alone BL,c = 10−1 Rb
Modified 1st-order 3 N = 50
open-loop est. N = 16
N =5
RAKE Receiver:
3-Finger diff. 1
coherent
Adaptive Filter Receivers:
Ideal coherent 1
Diff. coherent 1
Costas loop 2 BL,c = 10−1.5 Rb
alone BL,c = 10−1 Rb
Modified 1st-order 3 N = 50
open-loop est. N = 16
N =5

The complexity of the ML schemes for large N became a significant drawback when consider-
ing real-time implementation. Since these seemed to perform no better than those of smaller
N , the N = 500 and N = 168 structures were disqualified for their complexity. The widest
“bandwidth” structures, i.e., N ∈ {50, 16, 5}, were selected as final candidates. Note that
the value of fmax for the modified first-order structure is set to 2σf,1 = 1/40Rb = 3200Hz
in all simulations.

6.4.3 Final Candidate Structures

The weeding out process described in section 6.4.2 led to the final nomination of receiver
structures outlined in Table 6.5. Five receivers, of the original total of 15, the ideal and
differentially coherent matched filter receivers, the three-finger differentially coherent RAKE
receiver, and the ideal and differentially coherent adaptive receivers, employ no carrier
recovery scheme at all. (The “ideal” coherent receivers, as noted above, simply make use of
perfect knowledge of the phase of the first arriving SOI component.) These are presented in
the simulations to provide baseline performance comparisons. The RAKE receiver, shown
CHAPTER 6. SIMULATION STRATEGY 100

in Figure 6.11, is particular interesting as it provides a more fair comparison to the adaptive
equalizing receiver in the presence of multiple signal rays. Moreover, a three finger RAKE
is implemented in the IS-95 CDMA standard. So five candidate carrier recovery algorithms
made the final cut, two Costas loops and three modified first-order open-loop estimators.
These are each applied to both the conventional matched filter and adaptive receivers to
make up the final 10 receivers under consideration in Chapters 7 and 8.

6.5 Chapter Summary

The selection of candidate receiver structures for examination in the next two chapters is
complete. In this chapter, we first presented detailed simulation procedures and assumptions
in sections 6.1 through 6.3. We then used the system so defined to perform preliminary
investigations into a large number of potential carrier recovery structures and weeded out all
those whose performance did not seem worth additional attention. This exercise is detailed
in section 6.4. This process led to the selection of the final candidates presented in Table
6.5. All that is left now is to move on to Chapters 7 and 8 where simulation results for
these structures is presented.
CHAPTER 6. SIMULATION STRATEGY 101

Figure 6.6: Conventional matched filter receiver for DEBPSK with “ideal” coherent carrier
reference; M1,1,i corresponds to the first arriving fading waveform of user 1 at time i.

Figure 6.7: Conventional matched filter receiver using differentially coherent (DBPSK)
detection

Figure 6.8: Conventional DEBPSK matched filter receiver with Costas loop carrier recovery
CHAPTER 6. SIMULATION STRATEGY 102

Figure 6.9: Conventional DEBPSK matched filter receiver with combined Costas and AFC
loops for carrier recovery

Figure 6.10: Conventional DEBPSK matched filter receiver with modified first-order ML
carrier phase estimation circuit
CHAPTER 6. SIMULATION STRATEGY 103

Figure 6.11: Three finger differentially coherent RAKE receiver with equal gain combining;
First three arriving rays are assumed known, and first path is assumed to have delay, τ = 0.

Figure 6.12: Adaptive filter receiver for DEBPSK with “ideal” coherent carrier reference;
M1,1,i corresponds to the first arriving fading waveform of user 1 at time i.
CHAPTER 6. SIMULATION STRATEGY 104

Figure 6.13: Adaptive filter receiver employing differentially coherent (DBPSK) detection

Figure 6.14: Adaptive filter receiver for DEBPSK with conventional Costas loop carrier
recovery
CHAPTER 6. SIMULATION STRATEGY 105

Figure 6.15: Adaptive filter receiver for DEBPSK with either modified first-order or second-
order open-loop phase estimation scheme
Chapter 7

Receiver Performance in AWGN

Channels

This chapter denotes the beginning of the presentation of results obtained via the Monte
Carlo simulation methods discussed in Chapter 6. Results for the forward link and the
reverse link, under two different power control assumptions, are presented in sections 7.2
and 7.3 respectively. In these sections, we also briefly discuss peculiar problems, caused
by spread-spectrum MAI, faced by coherent carrier recovery structures in CDMA systems.
Before discussing these results, however, this juncture serves as an opportune location to
point out, in section 7.1, some of the relevant differences between the conventional and
adaptive CDMA receivers examined in this research.

7.1 Notes on Conventional vs. Adaptive Reception

Before discussing performance results of several carrier recovery techniques in AWGN chan-
nels, it is important to give some high level discussion to increased performance degradation
of adaptive reception, with respect to conventional reception, when differential detection is
employed. The adaptive filter solution reduces the correlation, ρ1,u between the spreading
code of the SOI (user 1) and those of the SNOIs (users u). This correlation is the MAI that
incrementally degrades DSSS signals as more users are overlapped in time and frequency.
Figure 7.1 shows this ability. It displays the correlation of the SOI’s spreading code with
10 spreading codes in the system for a conventional matched filter receiver and of (bor-
rowing the nomenclature [52]) a complex-weight fractionally-spaced linear adaptive receiver
(CWFSLAR) after adaptation. Figure 7.1 clearly shows a dramatic reduction of MAI and a
corresponding dramatic improvement in such a receiver’s decision statistic. Time-dependent

106
CHAPTER 7. RECEIVER PERFORMANCE IN AWGN CHANNELS 107

Figure 7.1: Cross-correlation of all user’s pseudo-random codes with those of user 1 for
matched filter and CWFSLAR filter, the latter exploiting cyclostationarity in the MAI;
adapted from [52].

adaptive filtering for CDMA has been shown in [52] and [53] to be most effective when the
CDMA users in a system share code repeat and chip rates, which allows the exploitation of
cyclostationary signal properties.
The impact of reduced correlation between users can also be seen in Figures 7.2 and
7.3, where the BER performance on a synchronous CDMA forward link for matched and
adaptive filter receivers are shown. The Eb /No ratios are 8 and 10 dB respectively. The
ideal coherent and differentially coherent matched filter receivers are those shown in Figures
6.6 and 6.7, and the ideal and differentially coherent adaptive receivers are those shown in
Figures 6.12 and 6.13. The reduced slope of the adaptive receiver performance curves shows
adaptive reception has a lower sensitivity of BER to the number of simultaneous users. This
is due entirely to the ability of the adaptive receiver to reduce cross-correlation with the
MAI. We have noted in previous chapters that differential detection of encoded BPSK
suffers about 1 dB of power degradation relative to ideal coherent reception of the same, a
loss that hardly seems worth buying back given the simplicity of the DBPSK demodulator.
This result however is for single user, AWGN environments, for receivers not employing
adaptive equalization. What Figures 7.2 and 7.3 show dramatically is that the performance
CHAPTER 7. RECEIVER PERFORMANCE IN AWGN CHANNELS 108

Figure 7.2: Simulation results for bit error probabilites versus number of users on the
forward link in AWGN channel; Eb /No = 8 dB; performance of conventional matched
and adaptive filters with “ideal” and differentially coherent reception shown; frequency
uncertainty, 2σf = 1/40Rb .

Figure 7.3: Simulation results for bit error probabilites versus number of users on the
forward link in AWGN channel; Eb /No = 10 dB; performance of conventional matched
and adaptive filters with “ideal” and differentially coherent reception shown; frequency
uncertainty, 2σf = 1/40Rb .
CHAPTER 7. RECEIVER PERFORMANCE IN AWGN CHANNELS 109

degradation of differential with respect to ideal coherent detection is far worse for adaptive
receivers than for conventional matched filter receivers. Note the distances denoted by
arrows (A), are much shorter than the distances marked by (B). So the news that adaptive
reception can increase CDMA system capacity is now accompanied by the news that much
of its promised capacity is dependent upon coherent carrier recovery. This observation is the
most critical motivation for examining carrier recovery for adaptive receivers as we attempt
to “buy back” as much of this degradation as possible.
Observe, too, that in all cases in Figures 7.2 and 7.3, the performances of conventional
and adaptive receivers cross at some small number of users. Theoretically the performance
of the matched and adaptive filter receivers should be precisely equivalent for one user,
since the adaptive filter ought converge to the matched filter which is the optimal filter,
according to the Weiner criterion, for a single user receiver in AWGN. However, due to the
non-zero NLMS stepsize, µ, which is equal to relatively large values of 0.5 during training
and 0.1 during decision direction, we can expect additional mean squared error in the
adaptive algorithm. It is therefore not surprising that the matched filter receiver performs
better than the adaptive in single-user AWGN environments. That the differential receiver
performance lines cross at 5 users in both figures is somewhat surprising. We point out,
however, that the mutual cross-correlation between the users in this forward link situation
is 1/15. Therefore, the adaptive receiver really has very little interference to mitigate in
this benign environment. In short, at Eb /No of 8 and 10 dB, the conventional receivers
perform quite well and the adaptive receivers offer relatively little enhancement.

7.2 The Forward Link in AWGN

Figures 7.4 through 7.7 show the performance of the two candidate Costas loops relative
to ideal and differentially coherent receivers. In all cases, the Costas loop of bandwidth,
BL,c = 10−1.5 Rb , performed much worse than the expectations given in Chapter 6. This is
very likely due to relatively high value of frequency uncertainty, 2σf = 1/40Rb ≈ 10−1.6 Rb ,
used for all simulations. What was even more unexpected was the terrible performance
of the widest bandwidth Costas loop with BL,c = 10−1 Rb for multiple user cases. This
bandwidth is sufficiently wide for the typical values of frequency offsets simulated. Note that
in each of Figures 7.4 through 7.7, showing conventional and adaptive receiver performance
at both 8 and 10 dB respectively, this latter Costas loop does exactly what we expect for
one user—it performs better than the respective differentially coherent receivers and closely
approaches the performance of ideal coherent reception. But adding as few as one user, for
a total of two, causes the performance of this Costas loop to dramatically degrade. It is
CHAPTER 7. RECEIVER PERFORMANCE IN AWGN CHANNELS 110

quite clear that even in this seemingly benign forward link environment, MAI wreaks havoc
for conventional Costas demodulators. We attribute this problem to additional amplitude
jitter, which is described in section 7.2.1. Then in section 7.2.2, we give further discussion
to the simulation results obtained for Costas loops (Figures 6.8 and 6.14 for conventional
and adaptive receivers respectively) and open-loop estimators (Figures 6.10 and 6.15) on
the forward link in AWGN.

7.2.1 Costas Loop Jitter on Synchronous CDMA Links

Recall that the forward link CDMA signal may be expressed in complex baseband repre-
sentation by !
X
U
s(t) = du (t)cu (t) ej (2πf1 + θ1 ) + n(t) (7.1)
u=1

where, as in previous chapters, du and cu are the differentially encoded data symbol and
the spreading code of the uth user at time, t; f1 and θ1 are the residual frequency offset
and phase of user 1, the SOI, which are shared by all users on the single carrier; U is
the total number of users on the system; and n(t) is a complex-valued Gaussian noise
h  i
process. Downconversion by a complex-valued reference,by exp −j 2πfref + θ̂ , as shown
in Figure 6.8 and sampling at time, t = kTb , results in a receiver statistic
!
X
U
rk = d1,k + du,k ρ1,u ejφk + nk (7.2)
u=2

where we have subsumed frequency and phase errors between the reference and transmitted
carriers into the possibly time-varying phase error, φ. ρ1,u represents the cross-correlation
between the spreading code of user 1 and user u. Note that in this particular case, all
cross-correlation values between all users’ time-synchronous bits are 1/15. So let ρi,j = ρ =
1/15, ∀i 6= j, and note that ρi,i = 1. In-phase and quadrature arm samples may thus be
represented by !
X
U
ik = d1,k + du,k ρ1,u cos φk + nd,k (7.3)
u=2
and !
X
U
qk = d1,k + du,k ρ1,u sin φk + nq,k . (7.4)
u=2

The error term, ek , is no longer a simple function of one symbol and noise as it was in
Chapter 4. Rather it is given by
4
ek = ik qk
CHAPTER 7. RECEIVER PERFORMANCE IN AWGN CHANNELS 111

  
1   X
U X
U X
U

= sin 2φ d21,k + ρ2  d2u,k + du,k dv,k  + f (nk )
2 u=2 u=2 v=2
v6=u
  
U (U −1)
1 X
= sin 2φ 1 + ρ2 (U − 1) + d0v  + f (nk ) (7.5)
2 v=1

where f (nk ) is some complicated function of the noise that is ignored for now and d0 is
simply a binary random variable taking values ±1. Note that at steady-state, when 2φ is
supposed to be small, the effect of the cross-correlation between the SNOIs and user 1 on
the error term should be small, especially since ρ2 = (1/15)2 = 1/225. There are two things
to note, however. First of all, the summation in (7.5) of U (U − 1) binary random variables
is well-known to be itself a random variable with variance U (U − 1). This wide variance
would seem to contribute significantly to the jitter of the Costas loop even at steady-state
for even a relatively small number (say about 4) of users. Since we have chosen a very wide
Costas bandwidth to meet the specification of acquisition within 100 symbols, this jitter may
not be getting mitigated sufficiently in the loop filter. Secondly, the acquisition behavior of
the Costas loop due to these added random terms in the error signal could become severely
compromised. We believe that some combination of these issues contributes to the poor
performance of even the best Costas loop shown in Figures 7.4 through 7.7.

7.2.2 Results on the Forward Link in AWGN

The performance degradation of the Costas loop in the adaptive filter receiver, as seen in
Figures 7.5 and 7.7, is more striking than that seen in the conventional receivers in Figures
7.4 and 7.6. As noted briefly in section 4.2.7, the interaction between dual feedback systems
as depicted in Figure 6.14 is complicated and seems likely to lead to catastrophic effects. In
combining closed-loop carrier recovery with complex-weight adaptive filtering, we operated
on the assumption that the time constant of the carrier recovery loop ought to be shorter
than that of the adaptive algorithm. That is, we desired the carrier recovery structure
to respond more quickly to changes in phase and correct for them before the adaptive
algorithm could react. We therefore only examined relatively wide Costas bandwidths for
the adaptive receiver, the widest being BL,c = 10−1 Rb as shown in Figures 7.5 and 7.7.
That this structure performed well for one user is encouraging. But the nearly catastrophic
performance degradation seen in these plots for even a few additional users suggests that
close-loop carrier recovery circuits interact poorly in adaptive receivers designed to mitigate
spread-spectrum MAI.
One final note on Costas loop performance here is that remodulation or hard-limiting
CHAPTER 7. RECEIVER PERFORMANCE IN AWGN CHANNELS 112

Figure 7.4: Simulation results for bit error probabilites versus number of users on the
forward link in AWGN channel; Eb /No = 8 dB; performance of conventional matched filter
receivers with various Costas loops, “ideal” carrier recovery, and differentially coherent
reception shown; frequency uncertainty, 2σf = 1/40Rb .

Figure 7.5: Simulation results for bit error probabilites versus number of users on the
forward link in AWGN channel; Eb /No = 8 dB; performance of adaptive filter receivers
(CWFSLAR) with various Costas loops, “ideal” carrier recovery, and differentially coherent
reception shown; frequency uncertainty, 2σf = 1/40Rb .
CHAPTER 7. RECEIVER PERFORMANCE IN AWGN CHANNELS 113

Figure 7.6: Simulation results for bit error probabilites versus number of users on the
forward link in AWGN channel; Eb /No = 10 dB; performance of conventional matched
filter receivers with various Costas loops, “ideal” carrier recovery, and differentially coherent
reception shown; frequency uncertainty, 2σf = 1/40Rb .

Figure 7.7: Simulation results for bit error probabilites versus number of users on the
forward link in AWGN channel; Eb /No = 10 dB; performance of adaptive filter receivers
(CWFSLAR) with various Costas loops, “ideal” carrier recovery, and differentially coherent
reception shown; frequency uncertainty, 2σf = 1/40Rb .
CHAPTER 7. RECEIVER PERFORMANCE IN AWGN CHANNELS 114

Figure 7.8: Simulation results for bit error probabilites versus number of users on the
forward link in AWGN channel; Eb /No = 10 dB; performance of conv. matched filter
receivers with various modified 1st-order ML phase estimation structures, “ideal” carrier
recovery, and differentially coherent reception shown; frequency uncertainty, 2σf = 1/40Rb .

Figure 7.9: Simulation results for bit error probabilites versus number of users on the for-
ward link in AWGN channel; Eb /No = 10 dB; performance of adaptive filter receivers
(CWFSLAR) with various modified 1st-order ML phase estimation structures, “ideal” car-
rier recovery, and differentially coherent reception shown; freq. uncertainty, 2σf = 1/40Rb .
CHAPTER 7. RECEIVER PERFORMANCE IN AWGN CHANNELS 115

the I-channel was not applied in any of the results presented here. This practice would un-
doubtedly mitigate the problems shown in (7.5) due to squaring but would create additional
problems in low Eb /No or high MAI where as noted in Chapter 4, the receiver statistics are
highly corrupted and dˆk might often be 180o out of phase with dk . A possible work-around
for this problem would be to create a receiver that could arbitrarily switch between squaring
and remodulation (hardlimiting) based on some knowledge of received Eb /No and/or the
number of CDMA users on the system. An algorithm would have to be developed for such
a receiver and would incrementally increase mobile unit complexity for the forward link. As
such, this possibility was not pursued for this research but is posited as a possibly fruitful
area of future work in carrier recovery.
Figures 7.8 and 7.9 show the performance of the modified first-order ML phase estimator,
at Eb /No = 10 dB, relative to the ideal and differentially coherent conventional and adaptive
receivers respectively. Here quite opposite behavior to that of the Costas loop seems to be
taking place. Especially for the N = 50 and N = 16, performance is degraded for a low
number of users, but performance is improved with respect to the Costas loop for a large
number of users. In a most encouraging result, the N = 5 modified first-order estimator
performed quite well in both the conventional and adaptive receivers. In Figure 7.8, the
conventional receiver case, this estimator did slightly better than the differentially coherent
receiver at all numbers of users examined. In the adaptive receiver case shown in 7.9, the
N = 5 estimator did significantly better than the differentially coherent scheme at 1 through
7 users, equally well at 10 users, and quite a bit worse at 13.
The reasons for the degraded performance of the two ML estimators for one or few users
are not exactly known. According to the theory presented Chapter 4, the performance be-
tween an ML estimator and a Costas loop is identical in a single-user, AWGN environment,
which is the case at one user in Figures 7.8 and 7.9. This prediction only holds, however, for
channels in which the received phase, θk , is not time-varying. Remember that the modified
first-order ML structure, described in section 4.2.4, relies on a sinc(∗) weighting function to
compensate for a frequency offset, which is reflected in a constant rotation in the samples
θk . This structure was shown to be the ML estimator for a frequency offset equal to fmax .
In these simulations, however, fmax = 2σf,1 , and not the actual value of the frequency
offset. Some degradation is therefore expected. This fact alone probably explains how the
N = 5 estimator usually outperformed the N = 16 and N = 50 estimators. The longer
the averaging period, the greater the impact of the suboptimality of choosing a fixed fmax
parameter. In addition, the performance of the 180o phase jump detector is not perfect and
can be expected to contribute to BER degradation. For these reasons, we are not terribly
CHAPTER 7. RECEIVER PERFORMANCE IN AWGN CHANNELS 116

surprised with the performance curves of the ML estimators shown in Figures 7.8 and 7.9
and are in fact quite pleased with the performance of the N = 5 estimator.

7.3 The Reverse Link in AWGN

7.3.1 Estimator Bias in Multiple Access Channels

The jitter problem for the Costas loop on the synchronous forward link in CDMA channels
described in section 7.2.1 has something of an analogous problem on asynchronous reverse
links. This is seen most easily in a bias that can creep in to the ML estimators studied.
Consider the reverse link desicion statistic seen by the base station at the output of the
matched filter,
X U
rk = d1,k ejθ1,k + Au du,k ρ1,u ejθu,k + nk (7.6)
u=2
where du,k , ρ1,u , U , and nk are as defined in section 7.2.1; the respective, generally time-
varying, carrier phases, θu,k , are now arbitrary for each user, u; and Au is the amplitude of
the uth user defined via
1
Au = 10 20 σP Pu , Pu ∼ Normal (0, 1) (7.7)

or the square root of the uth user’s power, which is distributed log-normally with variance
h i
σP2 dB2 .
Assuming that the use of decision feedback for the phase estimation scheme is unwar-
ranted or unavailable, we remove the modulation on the statistics rk by squaring them.
Consider rk2 on this asynchronous reverse link. Squaring (7.6) we get

X
U X
U X
U
rk2 = d21,k ej2θ1,k + d2u,k A2u ρ21,u ej2θu,k + du,k Au ρ1,u ej2θu,k dv,k Av ρ1,v ej2θv,k + f (nk )
u=2 u=2 v=2
v6=u

X
U X
U X
U
= ej2θ1,k + A2u ρ21,u ej2θu,k + du,k dv,k Au Av ρ1,u ρ1,v ej2(θu,k +θv,k ) + f (nk ) (7.8)
u=2 u=2 v=2
v6=u

where, as before, d2i = 1 and f (nk ) is some function of the noise. As discussed in Chapter
4, we then sum these statistics over N observations, possibly weighting each statistic. The
complex square root and signum function then yields the phase estimate of ejθ1,k .
The double summation term in (7.8) clearly might cause jitter in this estimate, but
note that summing over N observations reduces the impact of this term since the term is
randomly modulated by du dv . Observe, however, that the second, single-sum term could
CHAPTER 7. RECEIVER PERFORMANCE IN AWGN CHANNELS 117

become quite pernicious to efforts of extracting an estimate. This term is no longer depen-
dent upon the randomly modulated data, and thus produces a potentially large bias in the
estimator. The effects of the second term can only be averaged out over N observations if
the frequency offsets of user 1 and users u, which produce the time-varying quality of θk ,
are quite different. But it is very easy to imagine a situation where the frequency offsets of
the various users, while different, are nevertheless close, and thus in an observation interval
of N symbols, θk of the various users will change very little. This will hinder the attempt
at constructing an estimate for the phase of the SOI, ejθ1,k , and, at any rate, cause the
estimator to no longer be ML.
The bias produced by the second term of (7.8) could be assumed to be small if the various
Au are relatively equal and the ρ1,u are small. However, on the reverse link, we cannot make
this assumption in general. The amplitudes Au , relative to A1 , are now random parameters
based on the level of power control available and/or the instantaneous fading levels and ρ1,u
is bounded by the fairly high value of 7/15. This weakness therefore might be expected to
show itself in sections 7.3.2 and 7.3.3 where results on the reverse link are discussed. Note
finally that this perceived weakness of the ML estimators developed in Chapter 4 could be
overcome by developing, from first principles, a ML phase estimator for the peculiarities of
the multiple access link under consideration. Though such an estimator is not attempted
in this work, it is believed to be a possibly fruitful area of future research. Moreover, we
believe that if such an estimator were constructed, it would necessarily rely upon knowledge
of the MAI, specifically the time-varying phase of all the users in the system. Since such
information could reasonably be expected to be available in base station technology, the
development of such a structure might possibly have profound implications and benefits for
multiuser detection.

7.3.2 Results in Strict Power Control

Figures 7.10 through 7.13 show the performance of Costas loop carrier recovery for con-
ventional matched filter and adaptive filter receivers at Eb /No of 8 and 12 dB respectively.
These results are obtained on an asycnhronous reverse link with “strict” power control, a
situation defined by σP2 = 2 dB2 . As in section 7.2, the narrower bandwidth Costas loop,
BL,c = 10−1.5 Rb , performs much worse than hoped. Again this is most likely due simply
to the large frequency uncertainty, 2σf = 3200 Hz, present in all of these simulations. Also
in very much the same way as it did for the synchronous forward link, the Costas loop
helps the adaptive receivers in Figures 7.11 and 7.13 only for one user. For two or more
CDMA users, the Costas loop quickly degrades in these plots and performs far worse than
CHAPTER 7. RECEIVER PERFORMANCE IN AWGN CHANNELS 118

the differentially coherent adaptive receiver.

The one bright spot in Figures 7.10 through 7.13 is how well the wider bandwidth, BL,c =
10−1 Rb , Costas carrier recovery circuit works for the conventional matched filter receivers
in Figures 7.10 and 7.12. At Eb /No = 8 dB (Figure 7.10) this Costas loop outperforms
differentially coherent reception for 1 through 7 users and performs equivalently for a higher
number. At Eb /No = 12 dB (Figure 7.12), this Costas loop performs roughly the same as
differentially coherent reception. The reasonably good performance of this Costas loop in
these two plots is in stark contrast to its performance in Figures 7.4 and 7.6 where quick and
almost catastrophic degradation is seen for even a few users added. Again, we believe that
this is due to the jitter problem discussed in section 7.2.1. Interestingly, the Costas loop
examined performs better relative to the differentially coherent receiver on the asynchronous
reverse link where the MAI is undoubtedly worse, than on the synchronous forward link
where there is only a single carrier to track.
The performance of the modified first-order ML∗ phase estimation structures is shown in
Figures 7.14 and 7.15 for the conventional and adaptive receivers respectively at Eb /No = 12
dB. For the conventional receiver (Figure 7.14), the open-loop schemes perform resonably
well although in no case do they do better than differentially coherent reception. For
the adaptive filter receiver (Figure 7.15), the “narrow bandwidth,” N ∈ {50, 16} recovery
schemes fare quite a bit worse than differentially coherent in all cases. The N = 5 estimator
performs reasonably well for U = {1, 2, 4} users, but this receiver’s performance degrades
catastrophically for higher numbers of users.
We point out here that the adaptive receiver is subject to complete failure, i.e., it gets
into situations where decision directed adaptation with a high density of incorrect deci-
sions produces a highly suboptimal filter, and the BER approaches 0.5. This problem is
exacerbated by the stringent 100-symbol synchronization requirement and shows itself more
and more as the channel conditions get worse and worse, i.e., as the fading rate or level
of MAI increases or as received Eb /No drops. Such catastrophic failure is not observed
in the matched filter receiver, but a steady, more predictable degradation due to these
impairments. This can strongly bias the BER results seen in the plots presented for the
adaptive receiver. In a realistic receiver, this failure would result in a “dropped call” in
cellular telephony or in a request for retransmit in a packet-based wireless data system, and
therefore, though such failures ought to be considered a penalty against the receiver, they
would nevertheless not likely result in a steady-state BER as bad as is being counted in
these simulations.
∗
We now use the “maximum likelihood” designation loosely, only to bring consistency to the discussion.
As pointed out in section 7.3.1, this estimator is not ML for the asynchronous CDMA environment.
CHAPTER 7. RECEIVER PERFORMANCE IN AWGN CHANNELS 119

Figure 7.10: Simulation results for bit error probabilites versus number of users on reverse
link in AWGN channel; “strict” power control, σP2 = 2 dB2 ; Eb /No = 8 dB; performance
of conv. matched filter receivers with various Costas loops, “ideal” carrier recovery, and
differentially coherent reception shown; freq. uncert., 2σf = 1/40Rb .

Figure 7.11: Simulation results for bit error probabilites versus number of users on reverse
link in AWGN channel; “strict” power control, σP2 = 2 dB2 ; Eb /No = 8 dB; performance
of adaptive filter receivers (CWFSLAR) with various Costas loops, “ideal” carrier recovery,
and differentially coherent reception shown; freq. uncert., 2σf = 1/40Rb .
CHAPTER 7. RECEIVER PERFORMANCE IN AWGN CHANNELS 120

Figure 7.12: Simulation results for bit error probabilites versus number of users on reverse
link in AWGN channel; “strict” power control, σP2 = 2 dB2 ; Eb /No = 12 dB; performance
of conv. matched filter receivers with various Costas loops, “ideal” carrier recovery, and
differentially coherent reception shown; freq. uncert., 2σf = 1/40Rb .

Figure 7.13: Simulation results for bit error probabilites versus number of users on reverse
link in AWGN channel; “strict” power control, σP2 = 2 dB2 ; Eb /No = 12 dB; performance
of adaptive filter receivers (CWFSLAR) with various Costas loops, “ideal” carrier recovery,
and differentially coherent reception shown; freq. uncert., 2σf = 1/40Rb .
CHAPTER 7. RECEIVER PERFORMANCE IN AWGN CHANNELS 121

Figure 7.14: Simulation results for bit error probabilites versus number of users on reverse
link in AWGN channel; “strict” power control, σP2 = 2 dB2 ; Eb /No = 12 dB; performance
of conv. matched filter receivers with various modified 1st-order ML phase est. structures,
“ideal” carrier recovery, and diff. coherent reception shown; freq. uncert., 2σf = 1/40Rb .

Figure 7.15: Simulation results for bit error probabilites versus number of users on reverse
link in AWGN channel; “strict” power control, σP2 = 2 dB2 ; Eb /No = 12 dB; performance of
adaptive receivers (CWFSLAR) with various modified 1st-order ML phase est. structures,
“ideal” carrier recovery, and diff. coherent reception shown; freq. uncert., 2σf = 1/40Rb .
CHAPTER 7. RECEIVER PERFORMANCE IN AWGN CHANNELS 122

Figure 7.16: Same simulation results as shown in Figure 7.15 (check), except BER results
exceeding 0.2 are discarded and replaced with the receiver failure rate.

We therefore reserve the right to discount simulation points resulting in a sufficiently

high BER for the adaptive receivers and substitute instead a failure rate for the various
receivers. Sufficiently high BER is henceforth arbitrarily defined as 0.2 because once a
receiver begins to make errors at this rate, incorrect decision direction will force the filter
to a abitrarily incorrect solution and it is highly atypical for such a filter solution to be
rectified solely by continued decision-direction. When the adaptive filter fails, retraining
is necessary. Using this threshold, we point out that for Figure 7.15, the failure rates are
zero for the ideal coherent adaptive receiver at all numbers of users. For the differentially
coherent and the N = 5 open-loop estimation receivers the failure rates are {0%, 3%, and
10%} and {5%, 13%, and 17%} at 7, 10, and 13 users respectively. This different view of
the results in Figure 7.15 is given in Figure 7.16 where the performance of the N = 5 ML
estimation receiver is seen to improve dramatically with respect to differentially coherent
detection. In fact, ignoring for the moment the unpleasant fact of the high failure rate, we
get expected performance from the realizable coherent receiver, i.e., the performance curve
lies pretty much between the differential and ideal receiver and “buys back” some of the
performance degradation.
CHAPTER 7. RECEIVER PERFORMANCE IN AWGN CHANNELS 123

7.3.3 Results in Loose Power Control

The asynchronous reverse link with users’ powers distributed log-normally with a variance
of 20 dB2 is the most challenging environment yet discussed in this work. Large power
variances are sometimes referred to as the Near-Far (NF) problem and are a widely rec-
ognized “Achilles heel” of conventional CDMA systems. Figures 7.17 and 7.18 show the
performance of Costas loops at Eb /No = 12 dB for the conventional and adaptive receivers
respectively. For the conventional matched filter receiver in Figure 7.17, all strategies, in-
cluding ideal coherent carrier recovery, perform almost equally poorly, and therefore very
little can be said concerning the efficacy of carrier recovery strategies for such receivers
in this environment. The NF problem does its expected work in severely degrading the
performance of all the matched filter receivers.
The adaptive receiver, on the other hand, with its ability to mitigate MAI, shows great
promise in Figure 7.18, at least for the ideal coherent receiver. The adaptive receiver is at
its best when reducing the correlation between CDMA users and as such, advertises itself as
a Near-Far resistant receiver strategy. The very disappointing feature of this figure is that
even differentially coherent reception is degraded severely relative to the ideal coherent case.
The two Costas loops examined, BL,c equal to 10−1.5 and 10−1 Rb , give no improvement.
The great distance seen between the ideal and differentially coherent curves in Figure 7.18
continues to motivate an on-going search for carrier recovery structures that can better
approach the performance provided by ideal.
The performance of open-loop estimation structures is shown for conventional and adap-
tive receivers in Figures 7.19 and 7.20 respectively. In the same manner as with the Costas
loop, all conventional receivers degrade badly in these severe NF conditions. For the adap-
tive receivers in Figure 7.20, the N = 50 and N = 5 modified first-order estimation receivers
show some promise in at least approaching the performance of the differential receiver. But
the distance between these curves and that of ideal coherent reception is disappointing.
As was true in section 7.3.2, the BER simulated for adaptive filter receiver suffers dra-
matically from fairly high failure rates in Figures 7.18 and 7.20. These rates, not surpris-
ingly, are much higher than those seen in the “tight” power control conditions of the previous
section. These failures, while no small problem for receiver design, nevertheless bias the
BER curves in these figures. Without plotting the adjusted results, we point out that even
the ideal coherent adaptive receiver experienced failures at rates of {1%, 6%, and 7%} at 7,
10, and 13 users respectively. The diffentially coherent and modified first-order receivers,
however, failed at much higher rates. The differentially coherent receiver failed {1%, 9%,
19%, 35%, and 43%} of the time, for 2, 4, 7, 10, and 13 users respectively. Similarly, the
CHAPTER 7. RECEIVER PERFORMANCE IN AWGN CHANNELS 124

Figure 7.17: Simulation results for bit error probabilites versus number of users on reverse
link in AWGN channel; “loose” power control, σP2 = 20 dB2 ; Eb /No = 12 dB; performance
of conv. matched filter receivers with various Costas loops, “ideal” carrier recovery, and
differentially coherent reception shown; freq. uncert., 2σf = 1/40Rb .

Figure 7.18: Simulation results for bit error probabilites versus number of users on reverse
link in AWGN channel; “loose” power control, σP2 = 20 dB2 ; Eb /No = 12 dB; performance
of adaptive receivers (CWFSLAR) with various Costas loops, “ideal” carrier recovery, and
differentially coherent reception shown; freq. uncert., 2σf = 1/40Rb .
CHAPTER 7. RECEIVER PERFORMANCE IN AWGN CHANNELS 125

Figure 7.19: Simulation results for bit error probabilites versus number of users on reverse
link in AWGN channel; “loose” power control, σP2 = 20 dB2 ; Eb /No = 12 dB; performance
of conv. matched filter receivers with various modified 1st-order ML phase est. structures,
“ideal” carrier recovery, and diff. coherent reception shown; freq. uncert., 2σf = 1/40Rb .

Figure 7.20: Simulation results for bit error probabilites versus number of users on reverse
link in AWGN channel; “loose” power control, σP2 = 20 dB2 ; Eb /No = 12 dB; performance of
adaptive receivers (CWFSLAR) with various modified 1st-order ML phase est. structures,
“ideal” carrier recovery, and diff. coherent reception shown; freq. uncert., 2σf = 1/40Rb .
CHAPTER 7. RECEIVER PERFORMANCE IN AWGN CHANNELS 126

N = 5 open-loop estimator receiver failed {1%, 12%, 25%, 40%, and 50%} of the time, for
2, 4, 7, 10, and 13 users. This dismal performance in bad, though still reasonable, channel
conditions continues to motivate investigation not only into carrier recovery circuits that
can approach the performance of ideal coherent reception but also into ways of preventing
the failure of the adaptive algorithms to begin with.

7.4 Chapter Summary

In this chapter, we have examined the performance of conventional matched filter and
adaptive (CWFSLAR) receivers with ideal coherent carrier recovery, with differentially
coherent detection, and with several realizable, closed-loop and open-loop carrier recovery
schemes. Where expedient, we have endeavored to give some mathematical justification for
the types of results seen in the simulations. Specific attention was given to the problem
of added Costas loop jitter being caused by the synchronous MAI in section 7.2.1. The
problem of estimator bias in an asynchronous reverse link was also discussed in section
7.3.1.
The results are somewhat disappointing for the carrier recovery circuits examined as
we had hoped that they could consistently narrow the gap between the performance of
ideal coherent carrier recovery and that of differentially coherent detection. That this has
not been observed is nevertheless illuminating and lends credence to early suspicions that
differential reception might be a very good alternative in mobile channels. We now move
on to the final area of investigation in this research in Chapter 8 where we will expand
the investigation to include a variety of fading channels commonly encountered in mobile
systems.
Chapter 8

Receiver Performance in Fading

Channels

In this chapter, we submit the various receivers proposed in Chapter 6 to even more chal-
lenging conditions, those of fading that are so often encountered in mobile environments.
In section 8.1, we briefly examine receiver performance in flat fading—a situation where
a single path is received, in which all frequency components of the signal are attenuated
identically, at any given time. Then in sections 8.2 through 8.4, we examine frequency
selective channels where 2, 6, and 4 rays, respectively, reach the receiver spaced by periods
greater than the channel coherence time. In the on-going consideration of these channels,
we continue to examine both synchronous forward link results as well those of asynchronous
reverse links.
We caution the reader to be prepared to see some very high BER values in the proceeding
discussion. This motivates two points. First of all, the high BERs and high failure rates for
adaptive receivers, commonly seen in this chapter, remind us that BER is not the only figure
of merit for wireless digital systems. Rate of packet loss or requests for retransmit for data
communications and blocking probability or rates of call-dropping for isochronous voice
or multimedia applications are all valid bases, other than steady-state BER, upon which
to judge the performance of wireless systems. Further, these auxilliary figures of merit
can provide a different and useful perspective for such systems. Because of the additional
complexity required by the simulations to measure these characteristics, we were not able
to investigate them in this present research. We acknowledge this weakness in this work
only to suggest and motivate future investigation strategies. Interleaving, packetization
or frame-based transmission strategies, and forward error correction, should be considered
ripe and relatively untouched topics for research in the area of suppressed-carrier recovery.

127
CHAPTER 8. RECEIVER PERFORMANCE IN FADING CHANNELS 128

Note, however, that relative merits of various reception strategies are presented herein, and
they nevertheless provide a fairly wide view of the subject.
A second point to be made is that due to the high BERs observed, especially in sections
8.2 through 8.4 and additionally in the interest of time and of brevity, we do not present
results for two of the channel models mentioned in Chapter 6. Specifically, the “Hilly”
and “Bad Urban” channels are left relatively untouched in this work. These are the two
most challenging of all the channels described and as such can be expected and have been
observed to give even worse results than those given attention here.
Finally, in section 8.5, we briefly re-focus attention on the simulation procedures used.
Specifically, the impact of the stringent 100-symbol synchronization and training require-
ment and the impact of the choice of wide frequency uncertainty upon the BER performance
is examined. This examination is part of an attempt to explain the relatively poor BER
results which follow. We ultimately find, however, that relaxation of the simulation param-
eters mentioned does not provide any additional information concerning the receivers under
examination.

8.1 The One-Ray Rayleigh, “Flat Fading” Channel

The flat-fading model is not one in which we expect to see the adaptive receivers significantly
outpace the conventional receivers. As we will see in proceeding sections, adaptive reception
has the capability of coherently combining multipath in addition to suppressing MAI, but
in this particular model there is no multipath to combine, only small-scale Rayleigh fading
applied to a single ray. That being so, there is still some benefit of adaptive reception in the
suppression of MAI, but the dramatic improvements, relative to conventional matched filter
reception seen when the adaptive filter additionally exploits multipath, will be delayed until
sections 8.2 through 8.4 where frequency-selective environments are considered. Simulations
for this the one-ray Rayleigh channel model were carried out at Eb /No of 15 and 20 dB at
vehicle speeds of 15 and 45 kph for both forward and reverse links. On the reverse link,
power variances of 2 and 20 dB2 were examined. In section 8.1.1, we briefly examine the
results for forward link and then move on to those of the reverse link in section 8.1.2.

8.1.1 Performance on the Synchronous Forward Link

Figures 8.1 and 8.2 show the performance of receiver structures designed for conventional
matched filter and adaptive receivers respectively. Eb /No is set to be 20 dB in these par-
ticular results, and the vehicle speed is 45 kph. The relative flatness of most of the curves
CHAPTER 8. RECEIVER PERFORMANCE IN FADING CHANNELS 129

clearly shows the low level incremental MAI added for each user in a synchronous CDMA
system. The most noteworthy feature of both plots is the bad performance of the Costas
loop, BL,c = 10−1 Rb , for all situations except one user. As was discussed in section 7.2.1, we
believe this is due to the additional amplitude jitter induced into the Costas error signal by
the MAI. The N = 16 modified first-order estimation structure, in contrast, performs quite
well for both the conventional receivers in Figure 8.1 and the adaptive receivers in Figure
8.2. Nevertheless, at no time does this open-loop estimation scheme outperform differential
detection, which tends in this case to discourage the use of carrier recovery at all.

8.1.2 Performance on the Asynchronous Reverse Link

Figures 8.3 and 8.4 show the BER performance of conventional and adaptive receivers,
respectively, in the one-ray Rayleigh fading model. The conditions are an Eb /No of 15 dB,
a vehicle speed of 15 kph, and “tight” power control. Results for the conventional receivers
in Figure 8.3 are not terribly noteworthy except perhaps in that they show quite poor BER
performance. Realizable carrier recovery structures show little or no improvement over
differentially coherent detection. As expected, the ideal, unrealizable, coherent reception of
differentially encoded PSK fares slightly better.
Results for the adaptive receivers in Figure 8.4, on the other hand, do show at least
one interesting result. Ideal coherent reception, as we have come to expect, performs best
for all numbers of users. For the first time in this work, however, the Costas loop, with
BL,c = 10−1 Rb , outperforms differential detection for the adaptive receiver. This is in stark
contrast to the rather dismal results obtained in Chapter 7 where use of closed-loop carrier
recovery was all but ruled out in combination with the adaptive receiver. The modified
first-order open-loop estimator performs consistently worse than differential in this case.
To explain these unexpected results, it is important to point out the fundamental differ-
ence between the Costas loop scheme depicted in Figure 6.14 and the open-loop estimation
receiver depicted in Figure 6.15 in their application to adaptive reception. Note that the
Costas loop relies upon dedicated matched filter circuitry to compensate for phase and fre-
quency errors. The open-loop estimator, however, relies upon the output of the adaptive
filter itself to do the same.
Consider a situation where each receiver goes into a deep fade. Both receivers can
be expected to do quite badly for a certain amount of time. The adaptive filter, in each
case, will very likely get perturbed from its nearly optimal solution. Upon coming out
of the fade, the Costas loop adaptive receiver has a decisive advantage. It will be able
to quickly begin making very good estimates of the phase of the received signal again,
CHAPTER 8. RECEIVER PERFORMANCE IN FADING CHANNELS 130

Figure 8.1: Simulation results for bit error probabilites vs. number of users on the forward
link in flat Rayleigh-faded channel; Eb /No = 20 dB; performance of conventional receivers
with “ideal” and differentially coherent reception, wide BW Costas loop and N = 16 mod-
ified 1st-order ML structure; freq. uncertainty, 2σf = 1/40Rb ; vehicle speed is 45 kph.

Figure 8.2: Simulation results for bit error probabilites vs. number of users on the forward
link in flat Rayleigh-faded channel; Eb /No = 20 dB; performance of adaptive receivers with
“ideal” and differentially coherent reception, wide BW Costas loop and N = 16 modified
1st-order ML structure; freq. uncertainty, 2σf = 1/40Rb ; vehicle speed is 45 kph.
CHAPTER 8. RECEIVER PERFORMANCE IN FADING CHANNELS 131

Figure 8.3: Simulation results for bit error probabilites vs. number of users on the reverse
link in flat Rayleigh-faded channel; “strict” power control, σP2 = 2 dB2 ; Eb /No = 15 dB;
performance of conventional receivers with “ideal” and differentially coherent reception,
wide BW Costas loop and N = 5 modified 1st-order ML structure; freq. uncertainty,
2σf = 1/40Rb ; vehicle speed is 15 kph.

Figure 8.4: Simulation results for bit error probabilites vs. number of users on the reverse
link in flat Rayleigh-faded channel; “strict” power control, σP2 = 2 dB2 ; Eb /No = 15
dB; performance of adaptive receivers with “ideal” and differentially coherent reception,
wide BW Costas loop and N = 5 modified 1st-order ML structure; freq. uncertainty,
2σf = 1/40Rb ; vehicle speed is 15 kph.
CHAPTER 8. RECEIVER PERFORMANCE IN FADING CHANNELS 132

regardless of the state of the adaptive filter. The open-loop estimation scheme as depicted
in Figure 6.15, on the other hand, cannot quickly do so. The error in the adaptive filter
will compound itself by failing to produce good phase estimates. This “double jeopardy”
inherent to the latter receiver will increase the likelihood that the receiver makes a high
density of symbol estimation errors, which, due to decision-directed adaptation, will in turn
cause an increasingly worse adaptive solution and ultimately cause the receiver to fail. We
believe this is precisely what is happening in Figure 8.4.
This suggests an additional receiver structure for future consideration: one in which an
open-loop phase estimation structure is used but which utilizes the outputs of a dedicated
matched filter instead of the adaptive filter. Investigation of such a structure was not
possible for this work, however. Note that the rationale for using the adaptive filter as we
have is to reduce the impact of MAI on the received symbols, which in turn yield the phase
estimate. This impact, which is especially pernicious on the reverse link, as discussed in
Chapter 7, can be significant to the open-loop phase estimation scheme.

8.2 The Two-Ray Rayleigh, “Frequency Selective” Channel

Although it has only two signal paths, the two-ray Rayleigh channel model chosen for this
research is deceptively pernicious. The second signal component, as shown in Table 6.2, is
only 6 dB down from the first-arriving faded component. Its delay of 5 µs, however, is what
really causes problems for matched filter detection. In the system, this delay corresponds to
approximately 10 chips, or about 2/3 a symbol period. In lieu of coherent combining, this
dramatically increases the MAI seen at the output of a matched filter. It is in this section
that we can introduce the recombining capabilities of adaptive reception. In addition,
we introduce the performance of the differentially coherent, three-finger RAKE receiver,
depicted in Figure 6.11, to provide the adaptive filter with a more fair comparison than
that given by a simple matched filter. Simulations were run at Eb /No of 20, 24, and 28
dB, for vehicle speeds of 15 and 45 kph for both the forward link and the reverse link with
power variances of 2 and 20 dB2 . Sections 8.2.1 and 8.2.2 discuss results obtained for the
forward and reverse links respectively.

8.2.1 Performance on the Synchronous Forward Link

Figures 8.5 and 8.6 show the forward link performance of conventional and adaptive receivers
respectively. In each case, the performance of the RAKE receiver is shown for the purposes
of comparison. The bit energy to noise spectral density ratio is 28 dB,and vehicle speed
CHAPTER 8. RECEIVER PERFORMANCE IN FADING CHANNELS 133

is 15 kph. For conventional matched filter receivers in Figure 8.5, no realizable carrier
recovery strategies outperform the differentially coherent receiver. The Costas loop with
BL,c = 10−1 Rb performs slightly worse than the modified first-order N = 5 phase estimator.
The adaptive receivers in Figure 8.6 once again produce a much more interesting result.
Here, two realizable detection structures perform better than the “ideal” coherent receiver
at a number of points. Specifically the differentially coherent and the N = 5 open-loop
estimator perform better than “ideal” at 4 or fewer users. Remember that the definition
of “ideal” carrier recovery in a multipath environment merely means that we assume the
receiver has perfect knowledge at all times of the phase of the first arriving component,
which is arbitrarily considered to have delay τ1 = 0. No knowledge is assumed about other
paths. Clearly in this frequency-selective environment, realizable receivers, at least those
equipped with an adaptive equalizer capable of coherently combining multipath, can do
better than this “ideal” situation.
The wide bandwidth Costas loop applied to the adaptive receiver in Figure 8.6 performs
quite poorly on the forward link. We are convinced that this is due to the dramatically
increased jitter in the Costas error term induced by the MAI as described in Chapter 7.
Note, finally, that at all numbers of users, the RAKE receiver outperforms all others. This
is due to the fact that the RAKE receiver is a better recombiner of multipath than is
the adaptive filter in this particular environment. Since second path is delayed by 2/3 of a
baud period, only 1/3 of the energy in the second component is available in the time window
of the adaptive filter.∗ The MAI, the mitigation of which is the primary purpose of the
adaptive filter, is quite benign on the forward link. Borrowing the analogy of “interference-
limited” versus “noise-limited,” we denote this situation as “mulipath-limited” rather than
“interference-limited.” The dramatically better performance of the RAKE receiver versus
all adaptive receivers is therefore not suprising.

8.2.2 Performance on the Asynchronous Reverse Link

Simulations results at Eb /No = 28 dB on the reverse link with “tight” power control are
shown for conventional and adaptive receivers in Figures 8.7 and 8.8. Vehicle speed is
45 kph corresponding to a doppler frequency, fd = 85 Hz. Once again, the curves for
conventional receivers in Figure 8.7 tell us very little. In fact the RAKE receiver only
deviates significantly from the performance of all the other receivers examined at 1 and 2
users.
∗
Extension of the adaptive filter to an arbitrary length, based on expected multipath delays, is quite
feasible, but was not examined in this work.
CHAPTER 8. RECEIVER PERFORMANCE IN FADING CHANNELS 134

Figure 8.5: Simulation results for bit error probabilites vs. number of users on the forward
link in 2-ray Rayleigh-faded, frequency-selective channel; Eb /No = 28 dB; performance of
conventional receivers with “ideal” and differentially coherent reception, wide BW Costas
loop and N = 5 modified 1st-order ML structure; 3-finger RAKE also shown; freq. uncer-
tainty, 2σf = 1/40Rb ; vehicle speed is 15 kph.

Figure 8.6: Simulation results for bit error probabilites vs. number of users on the forward
link in 2-ray Rayleigh-faded, frequency-selective channel; Eb /No = 28 dB; performance of
adaptive receivers with “ideal” and differentially coherent reception, wide BW Costas loop
and N = 5 modified 1st-order ML structure; 3-finger RAKE also shown; freq. uncertainty,
2σf = 1/40Rb ; vehicle speed is 15 kph.
CHAPTER 8. RECEIVER PERFORMANCE IN FADING CHANNELS 135

Figure 8.7: Simulation results for bit error probabilites vs. number of users on the reverse
link in 2-ray Rayleigh-faded, frequency-selective channel; “strict” power control, σP2 = 2
dB2 ; Eb /No = 28 dB; performance of conventional receivers with “ideal” and differentially
coherent reception, wide BW Costas loop and N = 5 modified 1st-order ML structure;
3-finger RAKE also shown; freq. uncertainty, 2σf = 1/40Rb ; vehicle speed is 45 kph.

Figure 8.8: Simulation results for bit error probabilites vs. number of users on the reverse
link in 2-ray Rayleigh-faded, frequency-selective channel; “strict” power control, σP2 = 2
dB2 ; Eb /No = 28 dB; performance of adaptive receivers with “ideal” and differentially
coherent reception, wide BW Costas loop and N = 5 modified 1st-order ML structure;
3-finger RAKE also shown; freq. uncertainty, 2σf = 1/40Rb ; vehicle speed is 45 kph.
CHAPTER 8. RECEIVER PERFORMANCE IN FADING CHANNELS 136

Results for the adaptive receivers, however, show us more illuminating results in Figure
8.8. Here the N = 5 open-loop estimator and the differentially coherent adaptive receivers
perform almost identically. Note that at 1 user, these outperform the “ideal” coherent
receiver by a wide margin. The performance of the adaptive receiver equipped with the
parallel Costas loop structure is perhaps the most noteworthy. At all but one user, the
performance of this receiver lies between that of differential and ideal coherent, which is
precisely why we wish to do carrier recovery. Note also that in contrast to the forward link
case in Figure 8.6, the adaptive receiver, at least the one with “ideal” coherent reception,
now performs better than the RAKE receiver in all cases except one user.

8.3 The COST-207 “Typical Urban” Channel

Six distinct faded paths comprise the COST-207 “Typical Urban” channel. This is arguably
the most challenging environment in which we examined the performance of the candidate
receivers. As given from Table 6.2, the second arriving path is actually the strongest
component, 3 dB higher than the first arriving ray. But note that since the delay of the
second component, 0.2 µs, is less than one-half of the chip period of approximately 0.52µs,
this really should not cause too big a problem for spread spectrum receivers. By the time
we get out to the last path, arriving 5 µs later as in the two-ray Rayleigh model in section
8.2, the power is 10 dB down from the strongest component. So while “typical urban”
model is certainly more complex than the two-ray case, we don’t expect significantly worse
results than obtained in the previous section. Simulations were run at Eb /No of 24 and 28
dB for both forward and reverse links, the results of which are given in sections 8.3.1 and
8.3.2 respectively. Vehicle speeds of 15 and 45 kph were considered, and for the reverse link,
power variances of 2 and 20 dB2 were examined.

8.3.1 Performance on the Synchronous Forward Link

Figures 8.9 and 8.10 show results on the forward link for the urban channel. Specific param-
eters for these plots include an Eb /No of 24 dB and a vehicle speed of 45 kph. The perfor-
mance of all the single-path conventional receivers shown in Figure 8.9 is almost identical.
Interestingly, the “ideal” coherent matched filter receiver performs worst. This is likely due
to the fact that the “ideal” coherent tracker is tracking the first arriving component. Since
this ray is weaker than both the second and third, it is not surprising that performance so
degrades. As would be expected the three-finger RAKE receiver dramatically improves the
performance on the channel.
CHAPTER 8. RECEIVER PERFORMANCE IN FADING CHANNELS 137

Figure 8.9: Simulation results for bit error probabilites vs. number of users on the forward
link in COST-207 “Typical Urban” channel; Eb /No = 24 dB; performance of conventional
receivers with “ideal” and differentially coherent reception, wide BW Costas loop and N =
16 modified 1st-order ML structure; 3-finger RAKE also shown; freq. uncertainty, 2σf =
1/40Rb ; vehicle speed is 45 kph.

Figure 8.10: Simulation results for bit error probabilites vs. number of users on the forward
link in COST-207 “Typical Urban” channel; Eb /No = 24 dB; performance of adaptive
receivers with “ideal” and differentially coherent reception, wide BW Costas loop and N =
16 modified 1st-order ML structure; 3-finger RAKE also shown; freq. uncertainty, 2σf =
1/40Rb ; vehicle speed is 45 kph.
CHAPTER 8. RECEIVER PERFORMANCE IN FADING CHANNELS 138

Results for the adaptive receivers shown in Figure 8.10 allow more discrimination be-
tween the different receiver structures. The RAKE still performs best over a wide range of
users. The performances of differentially coherent and N = 16 open-loop phase estimator
receivers are almost equivalent, and no errors were generated for either at one user, sug-
gesting BERs in the range of 10−7 . Again the ideal coherent performs pretty badly, based
on the now less than ideal assumption regarding “ideal” coherent carrier recovery. Only the
Costas loop with BL,c = 10−1 Rb performs worse.

Figure 8.11: Results for the same conditions given in Figure 8.10, except at a slower vehicle
speed of 15 kph.

In this particular channel model, it is quite interesting to see the results for adaptive
filter receivers under conditions of the lower vehicle speed of 15 kph. Conventional receivers
as well as the RAKE show little change and are not therefore plotted. But for the adaptive
receivers, dramatic changes in the relative performance between receivers can be seen in
Figure 8.11. In stark contrast to Figure 8.10, now the “ideal,” differential, and open-loop
phase estimator receivers perform significantly better than the RAKE over a range of users
from 1 to about 7. This reinforces our earlier assertion of the ability of adaptive receivers
to combine multipath in addition to suppressing spread-spectrum MAI. Note again that
the differentially coherent and N = 16 open-loop receivers achieved no errors for one user,
suggesting a BER less than 10−7 .
CHAPTER 8. RECEIVER PERFORMANCE IN FADING CHANNELS 139

Figure 8.12: Simulation results for bit error probabilites vs. number of users on the reverse
link in COST-207 “Typical Urban” channel; Eb /No = 24 dB; “strict” power control, σP2 = 2
dB2 ; performance of conventional receivers with “ideal” and differentially coherent recep-
tion, wide BW Costas loop and N = 16 modified 1st-order ML structure; 3-finger RAKE
also shown; freq. uncertainty, 2σf = 1/40Rb ; vehicle speed is 15 kph.

Figure 8.13: Simulation results for bit error probabilites vs. number of users on the revers
link in COST-207 “Typical Urban” channel; Eb /No = 24 dB; “strict” power control, σP2 = 2
dB2 ; performance of adaptive receivers with “ideal” and differentially coherent reception,
wide BW Costas loop and N = 16 modified 1st-order ML structure; 3-finger RAKE also
shown; freq. uncertainty, 2σf = 1/40Rb ; vehicle speed is 15 kph.
CHAPTER 8. RECEIVER PERFORMANCE IN FADING CHANNELS 140

8.3.2 Performance on the Asynchronous Reverse Link

For the reverse link, we show results in a system with a “tight” power variance of 2 dB2
at a received Eb /No of 28 dB for the SOI, prior to fading. These are given in Figures 8.12
and 8.13 for conventional and adaptive receivers respectively. Vehicle speed is 15 kph. The
conventional receivers again show little discrimination between various receiver structures
with only the RAKE receiver performing slightly better. Again, the “ideal” coherent receiver
performs worst of all.
Results for the adaptive receivers in Figure 8.13 again show the “ideal,” differential,
and N = 16 modified first-order open-loop estimator performing much better than the
RAKE. While it is encouraging to see at least two realizable receivers perform better than
the three-finger RAKE, we note with some disappointment how much better the “ideal”
coherent receiver performs for 4 to 13 users than all others. It has been our hope that
realizable carrier recovery structures could come close to this peformance or even exceed it
under certain conditions. Figure 8.13 shows once again how very difficult and elusive this
goal has been.

8.4 The COST-207 “Typical Rural” Channel

The rural channel provides a slightly different view of the receiver structures examined in
this work. The first of four, arriving rays is assumed to have a Ricean fading envelope,
wherein a strong line-of-sight (LOS) component exists. Since the rays are very tightly
spaced, this is not a channel that we expect the adaptive receivers, with fractional sampling
of only two samples per chip, to be able to combine multipath. This limits the ability of
the adaptive, fractionally-spaced receivers to resolve the first from the second component,
and the third from the fourth. Nevertheless MAI is still mitigated to some extent, and
we therefore see the adaptive receivers giving a small improvement over the conventional
matched filter. Simulations were run for forward and reverse links at Eb /No of 24 and 28
dB for the SOI prior to fading. On the reverse link only the “tight” power control situation
with power variance of 2 dB2 was examined. Simulations were run on both links for vehicle
speeds of 15 and 45 kph.

8.4.1 Performance on the Synchronous Forward Link

Figures 8.14 and 8.15 show the forward link results for conventional and adaptive receivers,
respectively, at an Eb /No of 24 dB and a vehicle speed of 45 kph. For the conventional
CHAPTER 8. RECEIVER PERFORMANCE IN FADING CHANNELS 141

receivers in Figure 8.14, there is not much new information. The RAKE receiver dramati-
cally outperforms all single-path matched filter receivers, and the Costas loop and N = 5
open-loop estimation receiver are unable to do better than the differentially coherent. The
“ideal” coherent receiver, as it did for the urban channel forward link in section 8.3.1, per-
forms quite badly for the conventional receivers as well as the adaptive receivers in Figure
8.15.
The results for the adaptive filters are, as they have usually been, slightly more illumi-
nating. As we can see in Figure 8.15, the N = 5 receiver does quite well and outperforms
differentially coherent detection for 1 and 2 users. Still the RAKE receiver does better than
all adaptive receivers and is therefore recommended for the forward link in a rural channel.

8.4.2 Performance on the Asynchronous Reverse Link

Very little discriminatory information concerning carrier recovery was gained in examination
of the reverse link in the rural channel. Figures 8.16 and 8.17 show results at Eb /No = 24
dB with a vehicle speed of 45 kph for conventional and adaptive receivers respectively. The
conventional receivers in Figure 8.16 are very closely bunched together at all but one user.
The “ideal” coherent again performs slightly worse than the others.
However, in Figure 8.17, for the adaptive receivers, “ideal” coherent performs better
than all other receivers for a wide range of users. This is in stark contrast to the other
results obtained in this section and in the section prior wherein the first arriving ray is not
the strongest. We believe this is because the first arriving component in this particular
channel is assumed to have a Ricean spectral density. The LOS component present in the
first arriving ray may, at least in this instance, be sufficient to tip the balance and enable
“ideal” coherent recovery to perform well once again. Differentially coherent and N = 5
open-loop estimation receivers once again perform almost equivalently.
The Costas loop with BL,c = 10−1 Rb , on the other hand, outperforms both of these
and, in fact, along with the “ideal” coherent receiver, outperforms even the RAKE for a
wide range of users. Given that one of our major research goals has been to find realizable
carrier recovery structures for adaptive receivers that are capable of “buying back” the
performance generally lost in going from “ideal” to differentially coherent, we consider the
performance of this Costas loop in this situation to be one of the bright spots.
CHAPTER 8. RECEIVER PERFORMANCE IN FADING CHANNELS 142

Figure 8.14: Simulation results for bit error probabilites vs. number of users on the forward
link in COST-207 “Typical Rural” channel; Eb /No = 24 dB; performance of conventional
receivers with “ideal” and differentially coherent reception, wide BW Costas loop and N = 5
modified 1st-order ML structure; 3-finger RAKE also shown; freq. uncertainty, 2σf =
1/40Rb ; vehicle speed is 45 kph.

Figure 8.15: Simulation results for bit error probabilites vs. number of users on the forward
link in COST-207 “Typical Urban” channel; Eb /No = 24 dB; performance of adaptive
receivers with “ideal” and differentially coherent reception, wide BW Costas loop and N = 5
modified 1st-order ML structure; 3-finger RAKE also shown; freq. uncertainty, 2σf =
1/40Rb ; vehicle speed is 45 kph.
CHAPTER 8. RECEIVER PERFORMANCE IN FADING CHANNELS 143

Figure 8.16: Simulation results for bit error probabilites vs. number of users on the reverse
link in COST-207 “Typical Rural” channel; Eb /No = 24 dB; “strict” power control, σP2 = 2
dB2 ; performance of conventional receivers with “ideal” and differentially coherent recep-
tion, wide BW Costas loop and N = 5 modified 1st-order ML structure; 3-finger RAKE
also shown; freq. uncertainty, 2σf = 1/40Rb ; vehicle speed is 45 kph.

Figure 8.17: Simulation results for bit error probabilites vs. number of users on the forward
link in COST-207 “Typical Rural” channel; Eb /No = 24 dB; “strict” power control, σP2 = 2
dB2 ; performance of adaptive receivers with “ideal” and differentially coherent reception,
wide BW Costas loop and N = 5 modified 1st-order ML structure; 3-finger RAKE also
shown; freq. uncertainty, 2σf = 1/40Rb ; vehicle speed is 45 kph.
CHAPTER 8. RECEIVER PERFORMANCE IN FADING CHANNELS 144

8.5 The Effect of Training Period and Frequency Offset on

Receiver Performance
As noted in Chapter 6, we have made stringent sycnhronization and training requirements
on the receivers examined for this research. Moreover, we have assumed a rather large
uncertainty in carrier frequency. Throughout the discussion of results in both this chapter
and the last, we have often remarked that these two factors might play a role in producing
the less than stellar results obtained and most importantly, the unacceptably high failure
rates of many adaptive receivers. In this section, we relax these parameters and examine the
system in slightly more forgiving conditions. It is hoped that modifying these simulation
parameters would provide additional information regarding receiver performance. Alas, for
reasons to which we can now only speculate, this will turn out not to be true in the sections
below. First in section 8.5.1, we change one simulation parameter, viz., the length of the
period given for receiver training and synchronization, and examine the difference achieved
in BER results. Then in section 8.5.2, we dramatically reduce the frequency uncertainty in
the simulations and consider the impact this has on receiver performance.

8.5.1 The Impact of Training and Synchronization Period

The first modification was to allow 1000 training symbols for the adaptive filters, as opposed
to 100 given in previous trials. This is a generous allowance of time for an adaptive filter
to converge, but is not unthinkably long when considering a telephony application where
millions of bits could be transmitted over a link after call setup and synchronization is
acheived. We examined both reverse and forward links with this parameter change in the
two-ray Rayleigh fading model examined in section 8.2.
To our disappointment, this relaxation provided virtually no improvement in perfor-
mance for any of the receiver structures under consideration. We expected the adaptive
receivers to perform better given this additional training period, but as can be seen in Fig-
ures 8.18 and 8.19, which show results for the forward and reverse links respectively, this
simply was not the case. For the forward link in Figure 8.18, the results are almost exactly
the same as those obtained in the very similar conditions of Figure 8.6. No additional infor-
mation concerning carrier recovery or receiver performance can be drawn. Similarly, results
for the reverse link under tight power control with the 1000 bit training period, shown in
Figure 8.19, are almost exactly the same as those seen under identical conditions in Figure
8.8 where only 100 such bits were allowed.
From these disappointing results, it is clear that the length of training period given
CHAPTER 8. RECEIVER PERFORMANCE IN FADING CHANNELS 145

Figure 8.18: Simulation results for bit error probabilites vs. number of users on the forward
link in 2-ray Rayleigh channel; 1000 bits allowed for training & synch.; Eb /No = 28 dB;
performance of adaptive receivers with “ideal” and differentially coherent reception, wide
BW Costas loop and N = 5 modified 1st-order ML structure; 3-finger RAKE also shown;
freq. uncertainty, 2σf = 1/40Rb ; vehicle speed is 45 kph.

Figure 8.19: Simulation results for bit error probabilites vs. number of users on the reverse
link in 2-ray Rayleigh channel; 1000 bits allowed for training & synch.; “strict” power
control; Eb /No = 28 dB; performance of adaptive receivers with “ideal” and differentially
coherent reception, wide BW Costas loop and N = 5 modified 1st-order ML structure;
3-finger RAKE also shown; freq. uncertainty, 2σf = 1/40Rb ; vehicle speed is 45 kph.
CHAPTER 8. RECEIVER PERFORMANCE IN FADING CHANNELS 146

for the adaptive receiver does not significantly impact the very bad BER and high failure
rates seen in previous sections of this chapter. We next turn attention to the frequency
uncertainties assumed in this work to see if new results here can shed any light on this
problem.

8.5.2 The Impact of Frequency Uncertainty

Simulations throughout chapters 7 and 8 up to this point have assumed a frequency un-
certainty, 2σf = 3200 Hz, equal to 1/40Rb , for the SOI. We relaxed this parameter by one
order of magnitude to 320 Hz. This is a reasonable figure given a more accurate, and thus
slightly more expensive, set of oscillators are available for any particular receiver design.
We then re-examined the receivers in the two-ray Rayleigh channel as in section 8.5.1. The
results appear in Figures 8.20 and 8.21 for the forward and reverse links respectively. We
had hoped that the relaxation of the relatively wide frequency offsets would result in better
BER performance of at least some of the receivers and lower failure rates for the adaptive
receivers. Once again this is not the case.
The forward link results shown in Figure 8.20 show no significant departure from those
obtained in similar conditions for Figure 8.6. BERs are similar, and the relative performance
between users remains essentially unchanged. This observation can be extended to the
reverse link results shown in Figure 8.21 where there is similarly little departure from
those shown in Figure 8.8. The failure in this and the preceding subsection to obtain an
explanation for high BER and adaptive receiver failure rates suggests that there are more
fundamental challenges to the design of CDMA receivers employing adaptive interference
suppression, and further study is needed on this problem. We will suggest briefly in Chapter
9 that more advanced algorthims are needed, and the linking of equalization to channel
coding should provide fertile ground in which to examine this issue.

8.6 Chapter Summary

In this chapter, we have presented selected results from the many hundreds of simulation
points that were run in a variety channels for this research. We examined receiver per-
formance for the flat fading case in section 8.1 and then moved on to examine frequency
selective environments in sections 8.2 through 8.4. The most noteworthy finding in this
chapter is arguably the difficulty the adaptive receivers had in fading channels. We di-
gressed briefly in section 8.5 from the original simulation parameters in an attempt to
explain this lackluster behavior. It was, however, to no avail and leaves us no choice but
CHAPTER 8. RECEIVER PERFORMANCE IN FADING CHANNELS 147

Figure 8.20: Simulation results for bit error probabilites vs. number of users on the forward
link in 2-ray Rayleigh channel; lower frequency offset assumed, 2σf = 1/400Rb ; Eb /No = 28
dB; performance of adaptive receivers with “ideal” and differentially coherent reception,
wide BW Costas loop and N = 5 modified 1st-order ML structure; 3-finger RAKE also
shown; vehicle speed is 45 kph.

Figure 8.21: Simulation results for bit error probabilites vs. number of users on the reverse
link in 2-ray Rayleigh channel; lower frequency offset assumed, 2σf = 1/400Rb ; “strict”
power control; Eb /No = 28 dB; performance of adaptive receivers with “ideal” and differen-
tially coherent reception, wide BW Costas loop and N = 5 modified 1st-order ML structure;
3-finger RAKE also shown; vehicle speed is 45 kph.
CHAPTER 8. RECEIVER PERFORMANCE IN FADING CHANNELS 148

to recommend further study of realizable carrier recovery techniques suitable for receivers
employing adaptive interference suppression.
Based on the observations in this chapter, the best performing realizable carrier recovery
structures are the N = 5 and N = 16 modified first-order phase estimator for the forward
CDMA link and a Costas loop with BL,c = 10−1 Rb for the reverse link. Both of these
proved capable of beating the performance of the differentially coherent receiver, at least
some of the time, on the respective links. These can thus be said, at least under certain
conditions, to “buy back” part of the performance degradation between ideal coherent and
differential reception.
Chapter 9

Conclusions

We set out very early in this research to see if carrier recovery structures could be found
to work in CDMA environments. Specifically, we sought such structures that could “buy
back” the well-known performance degradation caused by resorting to a simpler differen-
tially coherent reception scheme. That this degradation seemed to be wider in systems
employing adaptive interference suppression further motivated and broadened the work.
The research began in Chapter 2 by surveying the available literature concerning synchro-
nization, especially where this historical knowledge had been applied to suppressed-carrier
recovery. Digestion of these concepts led to the introductory material provided in Chapters
3 through 5. With this knowledge as a foundation, we sought in Chapter 6 to apply it to
the nomination of candidate receiver structures and candidate carrier recovery algorithms
for the specific conditions of interest—a typical CDMA cellular system. Investigation via
simulation was then carried out in Chapters 7 and 8 for both forward and reverse links in
a variety of realistic mobile environments.
In this chapter, we review the salient points of the research process described above.
In section 9.1, these points are broken up into a set of more theoretical contributions, and
a set of more practical observations. Then in section 9.2, we conclude this thesis with
recommendations for future research strategies—promising areas of investigation that time
does not permit us to explore further in this work.

9.1 Significant Contributions and Findings of this Research

9.1.1 Theoretical Contributions

The major contributions of this work may be delineated by those of a theoretical nature and
those of a more practical nature—the bulk of them belonging to the latter category. With

149
CHAPTER 9. CONCLUSIONS 150

regard to substantive theory, the concise development of discrete-time transfer functions in

Chapter 3, for PLL loop filters in terms of classical and intuitively understandable analog
PLL paramaters, was the first major contribution of this work. These expressions allow
designers of digital systems to very quickly specify parameters for a discrete-time, digitally
implemented PLL that will have performance predicted by theory.
The exposition of several open-loop carrier phase estimation schemes based upon the
maximum likelihood (ML) principle in Chapter 4 was our second major theoretical contribution—
and perhaps the most important. The development of the decision feedback open-loop es-
timator based upon a high-SNR assumption in the ML equation given by (4.18) in that
chapter is, to our knowledge, unique. Further, the evaluation of first- and second-order
error statistics for both squaring and decision feedback structures shown in Figures 4.6 and
4.7 is unique to this work. That these statistics converged to those expected for the Costas
loops with “squaring” and with I-channel hard-limiting provided a new and important in-
sight into open-loop ML phase estimation structures. Finally, the development of a notion
of “loop bandwidth” for such structures was original to this work and will hopefully prove
helpful to future researchers.

9.1.2 Experimental Contributions

Experimental contributions of this research have been numerous. The first such contribution
was the marriage of open-loop ML phase estimation schemes to the fractionally-spaced
adaptive CDMA receiver first mentioned in section 4.2.7. This came very early in our
research as a result of applying the estimation schemes presented in [41] to the general
equalization strategies proposed in [51] for differentially coherent detection. It turned out
to be a very robust receiver structure in the examinations performed in Chapters 7 and 8.
From the very beginning, its performance served as a benchmark to be beaten by any and
all realizable carrier recovery techniques.
In Chapter 5, after a brief introduction to the topic of automatic frequency control
(AFC), we looked in some depth at the notion of combining AFC with a Costas loop by
summing the outputs of respective loop filters. Although this idea is at least 20 years
old, no known references have addressed the issue of the degradation of the Costas phase
estimate due to the presence of the AFC as we did in section 5.4. This was observed in a
few representative simulations. Framing the discussion in terms of Costas and AFC loop
bandwidths, we were able to posit some general guidelines for the design of such a system
that were not apparently well-known.
In Chapter 6, we spent a great deal of effort in helping to define the simulation procedures
CHAPTER 9. CONCLUSIONS 151

and assumptions for this research. We specified a CDMA system where bit energy to noise
or interference spectral density ratios were expected to be quite low. In narrowing down
the field of candidate carrier recovery structures and algorithms to be applied to such an
environment, we were able to observe major drawbacks to some of them. All receivers
employing AFC were disqualified due to their poor performance in high noise or when AFC
loop bandwidth was sufficiently narrow for the noise conditions, their unacceptably long
acquisition time. We observed that AFC loops ought be reserved for situations in which
the input SNR is sufficiently high, well above 10 dB, or in situations where the frequency
uncertainty is very wide and the acquisition time of an AFC loop with even with a narrow
bandwidth constraint will still be necessary. Neither of these conditions were examined in
our simulation work.
We further noted in Chapter 6 that the second-order ML phase estimation structure
proposed in Chapter 4, i.e., one implementing a dedicated frequency estimation loop, seemed
to perform no better in the conditions in question than the modified first-order estimation
structure. We observed, in fact, that for large N , the second-order structure performed
significantly worse than the other. We suggested that this was probably due to the fact
that jitter in the frequency estimate is multiplicative over the N -tap delay line in the second-
order estimator. Such jitter can be reduced arbitrarily by making the forgetting factor β
closer to unity but is paid for by an increasingly long acquisition period for the second-order
loop. This is not a penalty incurred by the modified first-order estimator. For this reason,
we applied no further investigation to the second-order structure and submitted that it be
reserved for situations of far more extreme carrier offsets.
Chapters 7 and 8 finally brought us to the long-awaited simulation experiments for
the final candidate receiver structures and carrier recovery algorithms. In Chapter 7, we
addressed some of the problems expected when applying the candidate carrier recovery
algorithms to CDMA environments. We observed in section 7.2.1 that Costas loops could
particularly be expected to have difficulty on a synchronous spread-spectrum multiple-access
forward link due to the rapid amplitude fluctuations due to MAI of the received samples.
The open-loop structures examined avoid this weakness on the forward link since a distinct
complex signum operation essentially eliminated this problem. We consquently observed
that the Costas loop almost universally performed poorer than all other algorithms on the
forward CDMA link at all points except one user. This is so in spite of the relatively small
level of forward link MAI contributed by each user in our simulations.
The open-loop schemes on the other hand showed some promise on the forward link
CHAPTER 9. CONCLUSIONS 152

in AWGN. The N = 5 modified first-order phase estimation structure significantly outper-

formed the differentially coherent adaptive receiver over a wide range of users at Eb /No
of 10 dB (Figure 7.9) and performed slightly better than differentially coherent reception
when applied to conventional matched filter receivers (Figure 7.8).
On the asynchronous CDMA reverse link, a new set of challenges is encountered. It was
shown in section 7.3.1 that a potentially strong estimator bias can occur due to the squaring
operation of our open-loop estimation schemes. Thus, we noted that the schemes derived
from the ML principle in AWGN, single-user environments, are no longer ML when applied
to CDMA. In reverse link situations of both “tight” power control, σP2 = 2dB2 , and “loose”
power control, σP2 = 20dB2 , we found very little consistent improvement to be gained
by implementing carrier recovery versus simply using differentially coherent detection. At
Eb /No = 8 dB in an AWGN channel under tight power control, we saw a wide bandwidth
Costas loop, BL,c = 10−1 Rb , give slight improvement over DPSK for conventional receivers
but none at all for adaptive receivers. The N = 5 modified first-order open-loop estimation
structure provided some performance enhancement for the adaptive filters in this situation.
In “loose” power control, near-far conditions, no realizable carrier recovery structures were
able to outperform DPSK in AWGN.
In fading channels, based on the observations in Chapter 8, we continued to see the
poorest performance in traditional Costas loops for either conventional matched filter or
adaptive filter receivers in forward link situations. Only open-loop estimation schemes
were able to beat differential detection on the forward link and this only some of the
time. Typically this improvement was seen for a relatively small number of users. On the
reverse link in fading channels, the Costas loop redeemed itself, outperforming the open-
loop estimation receivers most of the time and outperforming differential detection some of
the time. See, for example, the results for flat fading in Figure 8.4 and those of the two-ray
Rayleigh faded channel in Figure 8.8 where the Costas loop performance curve lies between
that of “ideal” and differentially coherent reception for nearly all numbers of users. That
this improvement was seen in adaptive receivers even more than in conventional receivers
was somewhat unexpected.

9.2 Recommendations for Future Work

While a representative variety of structures in a wide variety of conditions have been exam-
ined in this work, its observations have in many ways only scratched the surface of the issues
concerning carrier recovery in CDMA systems. In this section, we therefore endeavor to ex-
pose the shortcomings of our work with the goal of motivating future directions for research.
CHAPTER 9. CONCLUSIONS 153

Many of these recommendations have been hinted at throughout the previous chapters. We
summarize them here, delineating each as belonging to one of two broad categories: 1)
broadening the scope of the carrier recovery investigation to produce more interesting re-
sults, and 2) finding new methods to increase the robustness of adaptive equalization for
spread-spectrum MAI suppression.

9.2.1 Broadening the Scope of Investigation

As has been alluded to abundantly in previous chapters, the performance degradation be-
tween ideal coherent and differentially coherent reception of PSK is minimal for sparse
signal constellations such as M = 2. Binary phase signalling is incredibly robust to phase
jitter and thus, in the conditions assumed for this work, the widest bandwith systems,
the BL,c = 10−1 Rb Costas loop and its corresponding N = 5 phase estimation structure,
have consistently performed best of all realizable carrier recovery systems. For higher-order
signal constellations, we could expect just the opposite result to be true. Extending the
investigation to QPSK, 8-PSK, and QAM systems is likely to result in observations quite
different from those obtained in this present research.
Similarly, the assumptions in this work of low bit energy to noise and interference
spectral density ratios as well as relatively large frequency offsets proved prohibitive for a
large class of receiver strucutres that would otherwise have been worthy of investigation.
AFC was left relatively untouched due especially to the high noise densities. Narrower
bandwidth Costas loops were not seriously examined due to their inability to acquire quickly
in the presence of a large frequency offset. Future investigations ought to investigate more
forgiving Eb /No ratios and smaller frequency offsets that would go hand-in-hand with the
higher-order signal constellations suggested in the paragraph above.
The amplitude jitter problem for Costas loops on a single-carrier forward link signal, as
discussed in section 7.2.1, brings up a significant oversight in this work. In highsight, the
digital tanlock loop (DTL) introduced in section 4.1.6 might well have proved an interesting
work-around to this pernicious problem. Since the phase detector (PD) in the DTL is an
explicit arctangent operation, its performance is expected to be invariant to amplitude
fluctuations. Since it is a relatively obscure carrier recovery technique with only a small
handful of papers addressing it, a relative lack of attention was given it. In defense of
our lack of regard, we point out that an explicit four-quadrant arctangent operation is a
rather exotic scheme for real-time systems. Even on a DSP, the processing required for such
an operation would likely be considered prohibitive. Nevertheless, we consider the relative
lack of consideration given the DTL to have been an oversight and suggest, especially as the
CHAPTER 9. CONCLUSIONS 154

speed and flexibility of digital circuitry is increasing, that future investigations consider this
relatively obscure technique. It seems particularly promising for QAM signals or a CDMA
single-carrier (forward link) signal.
A final recommendation to broaden the scope of carrier recovery research comes out of
the problem of estimator bias due to asynchronous CDMA reverse link interference. Our
brief analysis in section 7.3.1 suggests an investigation of a ML approach to phase estimation
in CDMA systems. We believe that such an approach will necessarily utilize estimates of
the relative phases of all CDMA users. As such, we believe that such an approach will be
a perfect match for a variety of multiuser detection schemes, some of which are currently
under investigation by my colleagues at the MPRG.

9.2.2 More Robust Adaptive Receiver Designs

High failure rates observed for adaptive receivers in both Chapters 7 and 8 continue to
motivate further investigation in adaptive interference suppression for CDMA. One aspect
of future investigation ought to dispense with BER as the primary figure of merit for such
systems. Instead, a new and arguably better view could be provided by frame or packet
error rates for wireless data systems and call-dropping or blocking rates for isochronous
telephony. Such investigations would still pertain to carrier recovery in that the degradation
of adaptive receiver systems due to differentially coherent reception remains demonstrably
greater than that of conventional matched filter systems.
The fact that increasing the training interval from 100 to 1000 symbols for adaptive
receivers did not significantly affect the high failure rates serves to point out that there may
be a more fundamental weakness of the adaptive receivers in bad wireless channels. In most
wireless digital systems and, even more so, in those employing CDMA, interleaving and
forward error correction (FEC) are used to combat the periodic and dramatic reductions
in received signal strength. Time did not permit the inclusion of these in this research and
hence left us with raw, uncoded BER as the main figure of merit. We therefore believe
that an important, perhaps the most important, future direction of research in adaptive
interference suppresion for CDMA is in linking adaptive equalization at the front of the
baseband receiver to the FEC at the back end as proposed by [58]. We, in fact, find the
paucity of literature following up on this 1993 paper to be surprising. The authors thereof
suggest that a highly reliable decision of whether or not to update the adaptive filter can
be provided by the output of a decoder. If the decision to update is made, the Euclidean
distance between the received vector of statistics and the highly reliable estimate of the
transmitted symbol vector provide a highly reliable error for the update equation. The
CHAPTER 9. CONCLUSIONS 155

marriage of adaptive equalization and simpler FEC methods may suggest lower-complexity
or more bandwidth efficient alternatives to such powerful schemes as the constraint-length
9, rate 1/3 Viterbi decoder used on the IS-95 reverse link.

9.3 Chapter Summary

In this chapter, we have given a brief synopsis of the work described in this thesis. The
emphasis in section 9.1 was primarily on pointing out the most important results among
the dizzying myriad of lesser results contained in preceding chapters. This review served
additionally as both a thematic and a chronological overview of the research. We concluded
this report with suggestions and motivations for future endeavors in section 9.2. These are
areas left relatively untouched not only by this work but by most known literature. Finally
we point out, as we have abundantly before, that the surface has merely been scratched
concerning the topic of carrier recovery for CDMA. What we have presented is intended to
be a useful survey of the subject and to create fertile ground for future discovery.
Bibliography

[1] J. G. Proakis, Digital Communications, McGraw-Hill, NY, 3rd edition, 1995.

[2] S. C. Gupta, “Phase-locked loops”, Proc. of IEEE, vol. 63, no. 2, pp. 291–306, Feb.
1975, Classic Tutorial: 188 refs.

[3] W. J. Gruen, “Theory of AFC synchronization”, Proc. of IRE, vol. 41, pp. 1043–1048,
Aug. 1953.

[4] J. P. Costas, “Synchronous communications”, Proc. of IRE, vol. 44, pp. 1713–1718,
Dec. 1956.

[5] A. Blanchard, Phase-locked Loops: Application to Coherent Receiver Design, Wiley &
Sons,, NY, 1976.

[6] F. M. Gardner, Phaselock Techniques, Wiley & Sons, NY, 2nd edition, 1979.

[7] W. C. Lindsey, Synchronization Systems in Communications and Control, Prentice-

Hall, Englewood Cliffs, NJ, 1972.

[8] W. C. Lindsey and M. K. Simon, Telecommunication Systems Engineering, Prentice-

Hall, Englewood Cliffs, NJ, 1973.

[9] A. J. Viterbi, Principles of Coherent Communication, McGraw-Hill, NY, 1966.

[10] W. C. Lindsey and M. K. Simon, eds., Phase-Locked Loops & Their Application, IEEE
Press, NY, 1978.

[11] W. C. Lindsey and C. M. Chie, eds., Phase-Locked Loops, IEEE Press, Pitscataway,
NJ, 1986.

[12] W. C. Lindsey and C. M. Chie, “A survey of digital phase-locked loops”, Proc. of

IEEE, vol. 69, no. 4, pp. 410–431, Apr. 1981.

156
BIBLIOGRAPHY 157

[13] S. C. Gupta, “On optimum digital phase-locked loops”, IEEE Trans. Comm. Tech.,
vol. 16, pp. 340–344, Apr. 1968.

[14] G. S. Gill and S. C. Gupta, “First order discrete phase-lock loop with applications to
demodulation of angle-modulated carrier”, IEEE Trans. Comm., vol. 20, pp. 454–462,
June 1972.

[15] I. S. Xezonakis and M. S. Sangriotis, “An all digital Costas loop-like PLL circuit with
a wide locking range”, Int’nat’l Journal of Electronics, vol. 71, no. 6, pp. 951–965,
1991.

[16] L. E. Franks, “Carrier and bit synchronization in data communication – a tutorial

review”, IEEE Trans. Comm., vol. 28, no. 8, pp. 1107–1120, Aug. 1980, Classic
Tutorial.

[17] J. J. Spilker, “The delay-lock discriminator – an optimum tracking device”, Proc. of

IRE, vol. 49, pp. 1403–1416, Sep 1961.

[18] J. J. Spilker, “Delay-lock tracking of binary signals”, IEEE Trans. on Space Electronics
and Telemetry, vol. SET-9, pp. 1–8, Mar. 1963.

[19] R. L. Didday and W. C. Lindsey, “Subcarrier tracking methods and communications

system design”, IEEE Trans. Comm. Tech., vol. 16, no. 4, pp. 541–550, Aug. 1968.

[20] W. C. Lindsey and M. K. Simon, “The performance of suppressed carrier tracking loops
in the presence of frequency detuning”, Proc. of IEEE, vol. 58, no. 9, pp. 1315–1321,
Sep 1970.

[21] W. C. Lindsey and M. K. Simon, “Carrier synchronization and detection of polyphase

signals”, IEEE Trans. Comm., Concise Papers, vol. 20, no. 6, pp. 441–454, June 1972.

[22] G. L. Hedin, J. K. Holmes, W. C. Lindsey, and K. T. Woo, “Theory of false lock in

costas loops”, IEEE Trans. Comm., vol. 26, no. 1, pp. 1–12, Jan. 1978.

[23] M. K. Simon, “The false lock performance of costas loops with hard-limited in-phase
channel”, IEEE Trans. Comm., vol. 26, no. 1, pp. 23–34, Jan. 1978.

[24] M. K. Simon, “Tracking performance of costas loops with hard-limited in-phase chan-
nel”, IEEE Trans. Comm., vol. 26, no. 4, pp. 420–432, Apr. 1978.

[25] W. C. Lindsey and M. K. Simon, “Optimum performance of suppressed carrier receivers

with Costas loop tracking”, IEEE Trans. Comm., vol. 25, no. 2, pp. 215–227, Feb. 1977.
BIBLIOGRAPHY 158

[26] F. D. Natali, “AFC tracking algorithms”, IEEE Trans. Comm., vol. 32, no. 8, pp.
935–947, Aug. 1984, Seminal Paper on AFC.

[27] F. M. Gardner, “Characteristics of frequency-tracking loops”, in W. C. Lindsey and

C. M. Chie, eds., Phase-Locked Loops, IEEE Press, Pitscataway, NJ, 1986, Good Intro
to analog AFC.

[28] C. R. Cahn, “Improving frequency acquisition of a Costas loop”, IEEE Trans. Comm.,
vol. 25, no. 12, pp. 1453–1459, Dec. 1977.

[29] D. G. Messerschmitt, “Frequency detectors for PLL acquisition in timing and carrier
recovery”, IEEE Trans. Comm., vol. 27, no. 9, pp. 1288–1295, Sep 1979.

[30] H. Sari and S. Moridi, “New phase and frequency detectors for carrier recovery in PSK
and QAM systems”, IEEE Trans. Comm., vol. 36, no. 9, pp. 1035–1043, Sep 1988.

[31] A. N. D’Andrea and U. Mengali, “Performance of a quadricorrelator driven by modu-

lated signals”, IEEE Trans. Comm., vol. 38, no. 11, pp. 1952–1957, Nov. 1990.

[32] A. N. D’Andrea and U. Mengali, “Design of quadricorrelators for automatic frequency

control systems”, IEEE Trans. Comm., vol. 41, no. 6, pp. 988–997, June 1993.

[33] H. L. Van Trees, Detection, Estimation, and Modulation Theory, part I: Detection,
Estimation, and Linear Modulation Theory, Wiley & Sons, NY, 1968.

[34] H. L. Van Trees, Detection, Estimation, and Modulation Theory, part II: Nonlinear
Modulation Theory, Wiley & Sons, NY, 1971.

[35] H. L. Van Trees, Detection, Estimation, and Modulation Theory, part III: Radar-Sonar
Signal Processing and Gaussian Signals in Noise, Wiley & Sons, NY, 1971.

[36] H. Meyr and G. Ascheid, Synchronization in Digital Communications: Phase-,

Frequency- Locked Loops, and Amplitude Control, vol. 1, Wiley & Sons, NY, 1990.

[37] M. K. Simon and D. Divsalar, “The performance of trellis coded multilevel DPSK on
a fading mobile satellite channel”, IEEE Trans. on Vehicular Tech., vol. 37, no. 2, pp.
78–91, May 1988.

[38] M. K. Simon and D. Divsalar, “Doppler-corrected differential detection of MPSK”,

IEEE Trans. Comm., vol. 37, no. 2, pp. 99–109, Feb. 1989.

[39] D. Divsalar and M. K. Simon, “Multiple-symbol differential detection of MPSK”, IEEE

Trans. Comm., vol. 38, no. 3, pp. 300–308, Mar. 1990.
BIBLIOGRAPHY 159

[40] M. K. Simon and D. Divsalar, “Multiple-symbol, partially coherent detection of

MPSK”, JPL New Technology Report NPO-18932, July 1994.

[41] D. Divsalar and M. K. Simon, “Pseudocoherent demodulation of DPSK radio signals”,

JPL New Technology Report NPO-19205, June 1995.

[42] P. Y. Kam, “Maximum likelihood carrier phase recovery for linear suppressed-carrier
digital data modulations”, IEEE Trans. Comm., vol. 34, no. 6, pp. 522–527, June
1986.

[43] O. C. Mauss, F. Classen, and H. Meyr, “Carrier frequency recovery for a fully digital
direct-sequence spread spectrum receiver: A comparison”, in Proc. of IEEE Vehicular
Technology Conference, Secaucus, NJ, May 1993, pp. 392–395.

[44] F. Classen, H. Meyr, and P. Seheir, “Maximum likelihood open loop carrier synchro-
nizer for digital radio”, in Proc. of IEEE Int’nat’l Conf. on Communications (ICC
’93), Geneva, Switz., 1993, pp. 493–497.

[45] K. E. Scott and E. B. Olasz, “Simultaneous clock phase and frequency offset estima-
tion”, IEEE Trans. Comm., vol. 43, no. 7, pp. 2263–2270, July 1995.

[46] G. Karam, et al., “A reduced-complexity frequency detector derived from the

maximum-likelihood principle”, IEEE Trans. Comm., vol. 43, no. 10, pp. 2641–2650,
Oct. 1995.

[47] M. K. Simon, H. Tsou, and S. M. Hinedi, “Closed loop carrier phase synchronization
techniques motivated by likelihood functions”, JPL New Technology Report NPO-
19377, Apr. 1996.

[48] B. Y. Chung, et al., “Performance analysis of an all-digital BPSK direct sequence

spread spectrum IF receiver architecture”, IEEE JSAC, vol. 11, no. 7, pp. 1096–1106,
Sep 1993.

[49] C. M. Chie and W. C. Lindsey, “Phase-locked loops: Applications, performance mea-

sures, and summary of analytical results”, in W. C. Lindsey and C. M. Chie, eds.,
Phase-Locked Loops, IEEE Press, Pitscataway, NJ, 1986.

[50] C. H. Wei and W. J. Chen, “Digital tanlock loop for tracking pi/4-DQPSK signals in
digital cellular radio”, IEEE Trans. on Vehicular Tech., vol. 43, no. 3, pp. 474–479,
Aug. 1994.
BIBLIOGRAPHY 160

[51] K. M. Aleong, H. Leib, and P. Kabal, “A technique for combining equalization with
generalized differential detection”, in Proc. of IEEE 12th Annual Phoenix Conf. on
Computers and Communications, Phoenix, AZ, Mar. 1993, pp. 416–422.

[52] M. V. Majmundar, “Adaptive single-user receivers for direct sequence CDMA systems”,
Master’s thesis, Virginia Polytechnic Institute and State University, Blacksburg, VA,
April 1996.

[53] N. Zecevic, “Techniques and adaptation algorithms for direct-sequence spread-

spectrum CDMA single-user detection”, Master’s thesis, Virginia Polytechnic Institute
and State University, Blacksburg, VA, July 1996.

[54] H. Taub and D. L. Schilling, Principles of Communication Systems, McGraw-Hill,

NY, 2nd edition, 1986.

[55] R. C. Dixon, Spread Spectrum Systems with Commercial Applications, Wiley & Sons,
NY, 3rd edition, 1994.

[56] M. C. Jeruchim, P. Balaban, and K. S. Shanmugan, Simulation of Communication

Systems, Plenum Press, NY, 1992.

[57] COST 207: Digital Land Mobile Radio Communications, “Final report”, Office for
Official Publications of the European Communities, Brussels, 1989, ISBN 92-825-9946-
9.

[58] R. Sharma, W. D. Grover, and W. A. Krzymien, “Forward-error-control (FEC)-assisted

adaptive equalization for digital cellular mobile radio”, IEEE Trans. on Vehicular
Tech., vol. 42, no. 1, pp. 94–102, Feb. 1993.

[59] W. H. Beyer, Ed., CRC Standard Mathematical Tables and Formulae, CRC Press,
Boca Raton, FL, 29th edition, 1981.
Author’s Biographical Sketch

Steven P. Nicoloso was born on September 20, 1966 in Portland, OR. He received his
B.S. degree, summa cum laude, in social science education, from Liberty University in 1988.
In 1992, after the birth of his first child, Mr. Nicoloso changed career paths and returned
to college to pursue a degree in electrical engineering. In 1993, he entered Virginia Tech to
take undergraduate background coursework. From 1995 to the present, he has worked as a
graduate research assistant for the Mobile and Portable Radio Research Group. His research
interests center in applying DSP techniques and hardware to the communication systems
and include interference rejection, adaptive antenna arrays, and software radio concepts.
He is happily married and the proud father of currently two wonderful children.

161