These

Universit libre de Bruxelles
Ecole Polytechnique de Bruxelles

Service OPERA
Modulation formats and digital signal processing

for fiber-optic communications with coherent
detection
A thesis submitted for the degree of
Docteur en Sciences de lIngnieur
by
Jessica Fickers
Jury:
Dr. Gabriel Charlet
Prof. Philippe Emplit (Co-promoteur)
Assoc. Prof. Yann Frignac
Prof. Simon-Pierre Gorza (Prsident du jury)
Assoc. Prof. Franois Horlin (Promoteur)
Prof. Jrme Louveaux (Secrtaire du jury)
First of all, I would like to thank my supervisors, the Professors Franois Horlin and
Philippe Emplit as the initiators of my thesis project. Thank you for trusting me to
bridge your respective areas of expertise and for the assistance you provided at all levels
of the research project.
Second, this project would have been impossible without the help of our partners at the
Alcatel-Lucent Bell Labs in France. I would like express my gratitude to Dr. Jean-Pierre
Hamaide who agreed to receive me at Alcatel-Lucent and Dr. Gabriel Charlet who really
integrated me in his research group during my stays. I never felt a stranger at Villarceaux
and I am very grateful for that. It was a privilege for me to be able to work with expert
researchers who were willing to share their knowledge with me: in no particular order
of importance, I express my gratitude to Dr. Oriol Bertran-Pardo, Dr. Annalisa Morea,
Dr. Jeremie Renaudier, Hak Mardoyan, Patrick Brindel and Rafael Rios-Mller. My
special thanks go to Dr. Amirhossein Ghazisaidi, Dr. Massimiliano Salsi and - again Dr. Gabriel Charlet. You never hesitated to take time to help me in my research and in
the laboratory, thank you for your trust, help and interest.
This thesis was funded by the Belgian Fonds National de la Recherche Scientifique
(FNRS) to whom I address my sincere gratitude. Appreciation also goes to my alma
mater the Universit libre de Bruxelles and especially the OPERA department. Thanks
go out to the Professors, colleagues and friends at the wireless communications group:
Professor Philippe De Doncker, Professor Jean-Michel Dricot, Thibault Deleu, Ourouk
Jawad, Theodoros Mavridis, Luca Petrillo, Jonathan Verlant-Chenet, Marc Bauduin, Sullivan Derenne and Jonathan Bodard. I know I am not the most outgoing of collegues,
but rest assured I appreciate the welcoming atmosphere at the group. My special thanks
go to Natascha Vander Heyden and Ibtissame Malouli for all the instances in which their
assistance helped me along the way.
Last but not least kommt die Familie. Da mich ab jetzt keiner von den Nasen da oben
mehr verstehen kann, bin ich mal ehrlich: Ihr wisst ja hoffentlich dass Ihr viel wichtiger
seid als all die anderen zusammen? Mama und Papa, danke dass Ihr die allerbesten
Eltern seid die man sich vorstellen kann. Hallo Jerome und Vicky! Un tout grand merci
va aussi ma deuxime famille Christian, Christine, Yves et Maud. Xavier, que dire?
Nous avons travers cette aventure ensemble. Merci und setzen wir die Segel fr das
nchste Abenteuer!
Contents
Contents
List of acronyms
iv
1 Introduction
2 Fiber-optic communications
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Fiber-optic networks . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 The optical fiber as a transmission medium . . . . . . . . . . . .
2.4 Transmission impairments in optical fibers . . . . . . . . . . . . .
2.4.1 Attenuation . . . . . . . . . . . . . . . . . . . . . . . . . .
2.4.2 Chromatic dispersion . . . . . . . . . . . . . . . . . . . . .
2.4.3 Polarization mode dispersion . . . . . . . . . . . . . . . . .
2.4.4 Nonlinear interactions . . . . . . . . . . . . . . . . . . . .
2.5 The fiber-optic communication system . . . . . . . . . . . . . . .
2.5.1 Laser sources . . . . . . . . . . . . . . . . . . . . . . . . .
2.5.2 Optical modulation . . . . . . . . . . . . . . . . . . . . . .
2.5.3 Digital modulation formats . . . . . . . . . . . . . . . . .
2.5.4 Loss compensation . . . . . . . . . . . . . . . . . . . . . .
2.5.5 Coherent detection . . . . . . . . . . . . . . . . . . . . . .
2.5.6 Trade-off between noise and nonlinear distortions . . . . .
2.6 Mathematical description of the fiber-optic communication system
2.6.1 Nonlinear fiber simulation and linear transfer function . .
2.6.2 Mathematical models of the optical communication chain .
2.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3 Digital signal processing in coherent optical communication systems

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Time/frequency acquisition . . . . . . . . . . . . . . . . . . . . . . . . .
3.3 Chromatic dispersion compensation . . . . . . . . . . . . . . . . . . . . .
3.4 Polarization demultiplexing . . . . . . . . . . . . . . . . . . . . . . . . .
3.5 Carrier frequency estimation . . . . . . . . . . . . . . . . . . . . . . . . .
3.6 Carrier phase estimation . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.6.1 Viterbi-Viterbi carrier recovery algorithm . . . . . . . . . . . . . .
3.6.2 Maximum-Likelihood Phase Estimator . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5
5
6
7
9
10
10
11
11
12
13
13
14
16
16
18
19
19
20
21
.
.
.
.
.
.
.
.
23
23
24
25
26
27
28
28
28
ii
Contents
3.7
3.8
Symbol identification and FEC . . . . . . . . . . . . . . . . . . . . . . . . 29

Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4 Design rules for root-raised-cosine pulse shaping

4.1 Transmitter digital signal processing . . . . . . . . . . .
4.2 Root-raised-cosine pulse shaping . . . . . . . . . . . . . .
4.3 High-spectral efficiency coherent optical communications
4.4 Simulation setup . . . . . . . . . . . . . . . . . . . . . .
4.5 Hardware imperfections . . . . . . . . . . . . . . . . . .
4.6 Rolloff factor optimization . . . . . . . . . . . . . . . . .
4.7 Choice of FEC and modulation format . . . . . . . . . .
4.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
31
31
32
33
35
36
37
39
39
5 Decision feedback equalizers for bandwidth-constrained communications

5.1 Sub-baudrate filtered channels . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Symbol-by-symbol, MAP, and DFE detection . . . . . . . . . . . . . . . .
5.3 Wavelength selective switching in optical mesh networks . . . . . . . . . .
5.4 Equalizer structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.5 Numerical study of MMSE-DFE for bandwidth-constrained communication
5.5.1 Simulation setup . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.5.2 Simulation parametrization . . . . . . . . . . . . . . . . . . . . . .
5.6 Parametrization of MMSE-DFE . . . . . . . . . . . . . . . . . . . . . . . .
5.6.1 Performance of MMSE-DFE . . . . . . . . . . . . . . . . . . . . . .
5.6.2 Comparison with MAP . . . . . . . . . . . . . . . . . . . . . . . . .
5.6.3 Resilience to phase noise . . . . . . . . . . . . . . . . . . . . . . . .
5.7 MMSE-DFE for ROADM-filtered optical networks: experimental study . .
5.7.1 Experimental setup . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.7.2 Experimental results . . . . . . . . . . . . . . . . . . . . . . . . . .
5.7.3 Performance/complexity trade-off . . . . . . . . . . . . . . . . . . .
5.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
41
41
42
44
44
46
46
47
49
49
52
53
54
55
57
59
59
6 Multicarrier Offset-QAM modulations

6.1 Introduction . . . . . . . . . . . . . . . . . . . . .
6.2 MC-OQAM Principle . . . . . . . . . . . . . . . .
6.3 Receiver digital signal processing for MC-OQAM
6.3.1 Time/frequency acquisition . . . . . . . .
6.3.2 Channel estimation and equalization . . .
6.3.3 Phase tracking . . . . . . . . . . . . . . .
6.3.4 Experimental results . . . . . . . . . . . .
6.4 Conclusion . . . . . . . . . . . . . . . . . . . . . .
61
61
62
64
65
66
66
68
73
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
7 Low-complexity equalization for OQAM modulations

75
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
7.2 A low-complexity MC OQAM implementation . . . . . . . . . . . . . . . . 76
7.3 Simulation setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
Contents
7.4
CD compensation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.4.1 SC CD compensation . . . . . . . . . . . . . . . . . . . . . . . . .
7.4.2 OQAM CD compensation . . . . . . . . . . . . . . . . . . . . . .
7.5 Phase noise equalization . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.6 Cross-subcarrier equalizer . . . . . . . . . . . . . . . . . . . . . . . . . .
7.7 Extension to polarization division multiplexing and channel estimation .
7.8 Hybrid SC/MC modulation . . . . . . . . . . . . . . . . . . . . . . . . .
7.9 Computational complexity comparison between single carrier and
multicarrier-OQAM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.10 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
iii
.
.
.
.
.
.
.
79
79
80
82
82
85
86
. 91
. 92
8 Publications
95
9 Conclusion
97
10 Appendix: subcarrier orthogonality in MC-OQAM systems
99
Bibliography
103
List of acronyms
ADC
ASE
BER
BPSK
CD
CFE
CMA
CPE
DAC
DD
DD
DFE
DSP
DU
DU
EVM
FB
FD
FEC
FF
FFT
FIR
ICI
IFFT
ISI
ITU-T
LO
LUT
MAP
MC
ML
Analog to digital converter

Amplified spontaneous emission
Bit error rate
Binary phase shift keying
Chromatic dispersion
Carrier frequency estimation
Constant modulus algorithm
Carrier phase estimation
Digital to analog converter
Decision directed
Direct detection
Decision feedback equalizer
Digital signal processing
DIspersion-unmanaged
Dispersion unmanaged
Error vector magnitude
Feedback
Frequency domain
Forward error code
Feedforward
Fast Fourier transform
Finite impulse response
Inter-carrier interference
Inverse fast Fourier transform
Intersymbol interference
International telecommunications union - telecommunication sector
Local oscillator
Look-up table
Maximum a posteriori
Multicarrier
Maximum likelihood
MMSE
MZM
N-WDM
NLSE
NLT
OFDM
OOK
OQAM
OSNR
PD
PDM
PMD
PME
PRBS
PS
QAM
QPSK
ROADM
RRC
SBS
SC
SE
SMF
SSMF
TD
V+V
WDM
WSS
Minimum mean square error

Mach Zehnder modulator
Nyquist-wavelength division multiplexing
Non linear Schrdinger equation
Nonlinear threshold
Orthogonal frequency division multiplexing
On-off keying
Offset quadrature amplitude modulation
Optical signal to noise ratio
Photodiode
Polarization division multiplexing
Polarization mode dispersion
Polarization divisison multiplexing emulation
Pseudo-random binary sequence
Polarization-switched
Quadrature amplitude modulation
Quaternary phase shift keying
Reconfigurable optical add-drop multiplexer
Root-raised-cosine
Symbol by symbol
Single carrier
Spectral efficiency
Single mode fiber
Standard single mode fiber
Time domain
Viterbi and Viterbi algorithm
Wavelength division multiplexing
Wavelength selective switch
Chapter 1
Introduction
Optical fiber telecommunications form the backbone of the global data network. As such,
they have enabled the exponential growth of the data transfer capacity over the last
decades. Thanks to major technological breakthroughs such as single mode fibers and the
erbium doped fiber amplifier, the transmission distances were progressively increased and
today, long-haul optical links span oceans and continents.
Nowadays, with the speed of electronics progressively catching up with the data rates
of optical signals, a new revolution is taking place: the advent of digital signal processing
for optical communications. Impairments that were to be accounted for by analogical
means - or not at all - can now be corrected after the signal is translated to the digital
domain at the receiver. Moreover, the very form in which the information is transmitted
over the fiber can be re-defined by software. As we will see, these techniques can be used
to increase the data rate on the communication link.
The next-generation optical communication networks will respond to higher requirements in both spectral efficiency and flexibility than the current systems. Moreover, with
an increasing awareness of the need to decrease the power consumption of optical networks, the computational complexity of digital signal processing becomes a research focus.
The context of this PhD thesis is the potential of digital signal processing to achieve these
goals.
In the second chapter of this thesis, we review the fiber-optic transmission channel and
fiber-optic communication architecture for long-haul distances. First, the optical fiber is
introduced as a transmission medium. We study the main effects that distort the optical
signal as it is transmitted over the fiber. We see how digital data can be imprinted on laser
light and injected into the link as a first example of the use of digital signal processing.
Wavelength and polarization division multiplexing are introduced as means to enhance
the communication capacity of a single optical fiber. Next, the global architecture of a
point-to-point optical communication link is introduced. We will see how the same optical
signal can be transmitted over transatlantic distances thanks to optical amplification. We
introduce coherent detection thanks to which both amplitude and phase of the signal can
be detected.
In the third chapter, we present the state of the art in digital signal processing for
coherent optical communications. We present a diagram of the most commonly accepted
digital signal processing architecture for coherent detection. Each block in the diagram is
targeted towards one type of impairment: chromatic dispersion, polarization effects and
Chapter 1. Introduction
phase offsets. In the sections of the chapter, one or several algorithms are presented that
are typically used to mitigate each of the impairments.
The remaining chapters each present a study, experimental or by simulation, of a
digital signal processing solution to mitigate impairments, enhance the spectral efficiency
or the flexibility of a long-haul optical communication link. They progressively evolve from
being very similar to the state of the art to more novel techniques. It should be mentioned
that while the simulations and original digital signal processing algorithms presented
here were designed and implemented by the author of this work, all experimental results
presented in this thesis were carried out at the Alcatel-Lucent Bell Labs in Villarceaux,
France with much help from experienced Bell Labs scientists and technicians for both
laboratory work and state of the art digital signal processing. Luckily, I was also able
to help out with a few digital signal processing solutions (see n. 9 in Publications for
example) that are not included in this thesis because my contributions were minor.
In chapter 4, transmitter digital signal processing is used to optimize the pulse shaping
in order to reduce the optical channel spacing. We use numerical simulations to show that
a trade-off exists between spectral efficiency and the mitigation of hardware impairments.
In chapter 5 we enhance the equalization at the receiver so as to mitigate heavy in-line
filtering. In this study, we first show by numerical simulation that an equalizer that uses
the symbol decisions in a feed-back loop can mitigate the performance penalties arising
from bandwidth reduction at an acceptable computational cost. The proposed equalizer
is then validated in an experimental study in the context of in-line filtering stemming from
reconfigurable optical add-drop multiplexers, components that route the optical channels
at the network nodes. In this study, the experimental waveforms were measured by A.
Ghazisaeidi and M. Salsi at Bell Labs.
In chapter 6 we modify the mode of data transmission in order to increase the flexibility of the transmission scheme. We digitally divide the electrical bandwidth of one
channel into subcarriers. Each subcarrier is modulated independently using Offset-QAM.
In order to successfully detect the multicarrer Offset-QAM, the state of the art digital
signal processing can no longer be used. We propose an equalizer based on the minimum
mean square error criterion and trial phases. We experimentally show the feasibility of
multicarrier Offset-QAM in long-haul link. For this study, measurements were carried out
by the author of this work at the transmissions laboratory of Bell Labs with much help
from A. Ghazisaeidi and P. Brindel.
In chapter 7, we extend the study of chapter 6 to a higher density of subcarriers in
order to significantly reduce the computational complexity of the receiver digital signal
processing. In order to successfully decode the symbols in the presence of the channel
impairments, we propose a hybrid modulation architecture using a single carrier pilot
combined with low-complexity frequency domain equalization and prove its feasibility by
simulation. The different contributions of this thesis are summarized in Fig. 1.1 according
to their main goal(s).
Most of the results presented in this manuscript have already been published and presented in different peer-reviewed journal articles and conference contributions which I have
authored or co-authored. These articles are listed in the last chapter of the manuscript
named Publications.
Spectral
efficiency
Chapter 6
Flexibility
Chapter 4
Chapter 5
Chapter 7
Reduction of
computational
complexity
Figure 1.1: Summary of the contributions presented in this thesis according to their main
goals.
Chapter 2
Fiber-optic communications
Contents
2.1
Introduction
2.2
Fiber-optic networks
2.3
The optical fiber as a transmission medium
. . . . . . . . . . . .
2.4
Transmission impairments in optical fibers . . . . . . . . . . . . .
2.5
2.6
2.7
2.1
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . .
5
6
2.4.1
Attenuation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.4.2
Chromatic dispersion . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.4.3
Polarization mode dispersion . . . . . . . . . . . . . . . . . . . . . . 11
2.4.4
Nonlinear interactions . . . . . . . . . . . . . . . . . . . . . . . . . . 11
The fiber-optic communication system . . . . . . . . . . . . . . . .
12
2.5.1
Laser sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.5.2
Optical modulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.5.3
Digital modulation formats . . . . . . . . . . . . . . . . . . . . . . . 14
2.5.4
Loss compensation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.5.5
Coherent detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.5.6
Trade-off between noise and nonlinear distortions . . . . . . . . . . . 18
Mathematical description of the fiber-optic communication system 19

2.6.1
Nonlinear fiber simulation and linear transfer function . . . . . . . . 19
2.6.2
Mathematical models of the optical communication chain . . . . . . 20
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21
Introduction
In todays telecommunications, silica fibers are the preferred medium to transmit large
amounts of data over long distances. First proposed in 1966 for the transfer of information [1], optical fibers are waveguides based on the principle of total internal reflection.
Chapter 2. Fiber-optic communications
Compared to free-space transmission, they offer very low attenuation of the optical signal
in a broad range of frequencies in the near-infrared range. The data to be transmitted is
imprinted on the laser light which is then injected into the fiber. At the other end of the
link, photodiodes are used to translate the received light into electrical signals so that the
messages can be decoded. Through this basic principle, several Tbit/s can be transmitted on the same fiber over distances ranging up to tens of thousands of kilometers. This
chapter provides a basis for the understanding of the fiber-optical communication link.
First, we define the long-haul optical network in section 2.2. In section 2.3, we describe
the silica fiber as a transmission medium. In section 2.4, we detail the transmission impairments in optical fibers. In particular, we describe the effects of attenuation, chromatic
dispersion (CD) polarization effects and nonlinear distortions originating in the nonlinear
optical Kerr effect. In section 2.5, the general architecture of point-to-point optical transmission links is introduced. We discuss laser sources and optical modulation by which
the numerical data is imprinted on the carrier wave. We introduce digital modulation
formats. Those can be understood as alphabets used to transmit information. Next, we
will see how optical amplifiers compensate the attenuation losses in order to transmit the
same optical signal over thousands of kilometers. Finally, we introduce coherent detection, a receiver architecture that allows us to detect both phase and amplitude of the
optical signal.
2.2
Fiber-optic networks
All optical communication systems propagate data encoded on light waves over a series
of optical fibers. They can be categorized according to the distances covered, the environment in which the fiber is deployed and the network architecture.
In terrestrial long-haul networks, the distances involved are in the order of several
hundreds to several thousands of kilometers. Terrestrial long-haul networks are
conceived to cover the surface of a continent, with large cities at their nodes. In these
systems, each fiber section also called span measures about 80 to 100 kilometers.
Singlemode fibers (see section 2.3) are typically used. Between fiber spans, amplifiers
compensate for the attenuation of the signal power during transmission. At the
nodes of the system, electrical and/or optical routing devices such as wavelength
selective switches (WSS) and reconfigurable add-drop multiplexers (ROADM) insert
or retrieve individual part of the transmitted signal called channels. Each channel
occupies a specific frequency range defined by the laser source providing its carrier
wave. The switching devices are often located at the same places as the amplifiers
and inserted between amplifying stages.
Submarine long-haul networks are conceived to transmit data between continents.
As a consequence, their transmission distances range from 6000 (transatlantic) to
13000 (longest transpacific) kilometers. They also use singlemode fiber, but the
individual spans are shorter than in terrestrial systems (45-60 km). Amplification
is typically done in a single step. Submarine networks have very few ramifications.
2.3. The optical fiber as a transmission medium
Short-distance networks. On the scale of cities, metropolitan networks are designed

to link the long-distance networks to access networks.
The techniques developed in this work and the experimental setups used are representative
of terrestrial long-distance networks.
2.3
The optical fiber as a transmission medium
25
0
12 m
5
m
y
core
cladding
coating
z
x
Refractive index
ng core
di
ad
cl
ng
di
d
cla
Fiber radius
Figure 2.1: Silica fiber

Optical fibers are cylindrical dielectric waveguides made of silica glass. As can be seen
in Fig. 2.1, they are composed of fiber core, cladding and coating. The dielectric properties
of the cladding differ from those of the core. The refractive index is higher in the core
than in the cladding. As a consequence, light propagating approximately in the direction
of the main axis of the fiber z is confined to the core by total internal reflection [2].
The modes of wave propagation inside the optical fiber can be determined by solving
Maxwells equations for the electric field and magnetic fields, taking into account the
boundary conditions at the limits of core and cladding and the cylindrical symmetry. This
yields a discrete number of solutions of guided waves inside the fiber. In singlemode fibers,
the distribution of the refractive index is designed so that only one mode of propagation
is permitted. Singlemode fibers are the most widely used fibers in telecommunications
because the presence of several modes of propagation leads to intermodal interference
of the transmitted signal. The transverse energy distribution of the guided mode in
a singlemode optical fiber is illustrated in Fig. 2.2. It is uniform along the azimuthal
direction and has an approximately Gaussian distribution of energy with respect to the
fiber radius.
After the transverse distribution, we now examine the axial distribution, in other
words the propagation of the guided mode. In order to understand the propagation of
core diameter
Figure 2.2: Energy distribution of the fundamental mode of propagation in singlemode

fibers. Illustration from [3].
optical signals through the fiber, we introduce the concept of wave-packets. A wave-packet
is a superposition of plane waves of different frequencies around a central frequency. In
optical telecommunications, the central frequency is the frequency of the continuous laser
source that is injected into the fiber. The components of the wave-packet that occupy the
frequencies around the central frequency carry the information that is to be transmitter
over the link. The laser source is also called the "carrier" for this information.
The wave-packet A(z, t) at time t and axial position z is expressed as the superposition
of its spectral components:
Z +
A(z, t)
A()
exp (j(()z t)) d
(2.1)
=
In this equation, is the propagation constant, and denotes the pulsation. To calculate
the group velocity of the wave-packet, i.e. the speed at which the information propagates,
we must make the assumption that the wave-packet is quasi-monochromatic around the
carrier pulsation 0 :
() 0 + 1 .( 0 )
(2.2)
.
.
In this equation, 0 = (0 ) and 1 =
. In this case, it is easily shown [2] that
=0
the wave-packet travels at the group-velocity vg = 1/1 without distortion.
In modern optical transmission links, optical signals emitted by different laser sources
at different wavelengths are independently modulated and simultaneously propagated
over the same optical fiber [4]. This technique is called wavelength division multiplexing
(WDM) and exploits the wide frequency range in which optical fibers show very small attenuation. The telecommunication standardization sector of the international telecommunication union (ITU-T) defined six bands for transmission using singlemode fiber. Fig. 2.3
illustrates these bands as well as the attenuation as a function of wavelength/frequency.
The main communication band is the C band (see Fig. 2.3). The bandwidth can however be enlarged by using the L-band. Combining these two bands, approximately 160
wavelength channels can be transmitted on a 50-GHz grid.
The polarization of light refers to the direction of pulsation of the transverse electrical
field as a function of time. In the general case, the electrical field traces an ellipse in the
2.4. Transmission impairments in optical fibers
Figure 2.3: Attenuation of singlemode fiber as a function of wavelength and associated

transmission bands. Illustration from [5].
(x, y)-plane. If the electrical field pulsates along a line, it is said to be linearly polarized
along this direction [6]. Any state of polarization of a wave-packet propagating through
the fiber can be expressed as the combination of two fields that are linearly polarized along
orthogonal axes. Seen this way, the "unique" mode of propagation in singlemode fiber is
in fact the superposition of two degenerate modes polarized along two orthogonal axes in
the transverse plane of the fiber. In todays optical telecommunications, the two linearly
polarized modes of propagation are used to transmit independent streams of information.
First demonstrated in 1992 [7], this technique is called polarization division multiplexing
(PDM).
2.4
Transmission impairments in optical fibers
The undisturbed transmission of the wave-packet in Eq. (2.1) is only a first approximation.
In optical fibers, the main physical effects leading to the distortion of the wave-packet
as it propagates are attenuation, chromatic dispersion, polarization mode dispersion and
the non-linear Kerr effect. The evolution of the optical field along its propagation axis
is described by a partial differential equation called the coupled non-linear Schrdinger
equation [8]:
1 (z)
A
A
0 (z)
2 2 A 3 3 A
+
= Aj
J(z)A
J(z)
+j
z
2 } | 2
} | 2 2 {z 6 3}
| {z
{z 2
(1)
(3)
(2)
1
j |A|2 A (A 3 A)3 A
3
|
{z
}
(2.3)
(4)
In this equation, the bold A designates the propagated wave-packet as Jones vector. In
Jones formalism, the electrical field is expressed using the linearly polarized components
along orthogonal directions. = t z/vg is the time frame moving along the wave-packet
referential. J(z) is a 2 2 unitary Jones matrix. 3 is the third Pauli matrix. is the
. i
attenuation coefficient, is called the nonlinear coefficient, i =
where 0 is
i
=0
10
the central pulsation and () is the propagation constant. In Eq. (2.3), the different
right-hand terms refer to (1) attenuation, (2) chromatic dispersion (limited to the third
order) (3) polarization mode dispersion and (4) the non-linear Kerr effect. These fiber
impairments will be detailed in the subsequent paragraphs.
2.4.1
Attenuation
Ignoring all right-hand terms in Eq. (2.3) but attenuation, we have:
A
= A
z
2
(2.4)
It is easily seen that the power of the electric field A will decrease exponentially with
transmission distance. When an electromagnetic wave propagates through an optical fiber,
its amplitude is attenuated by absorption and scattering. While attenuation does not
result in signal distortion in itself, the need for amplification units along the transmission
line results in signal corruption by noise as we will see in section 2.5.4.
2.4.2
Chromatic dispersion
In an optical fiber, the quasi-monochromatic assumption Eq. (2.2) does only approximately hold. Around 1550 nm (transmission in the C band, see Fig. 2.3) and for signal
bandwidths of several tens of GHz, it is necessary to take into account at least one more
term in the development of the dispersion relation around the central pulsation. This
effect is called group-velocity dispersion and results in distortion of the transmitted signal
because the spectral components of the pulses travel at different speeds. The scalar NLSE
ignoring all effects except CD becomes:
2 2 A 3 3 A
A
=j
+
z
2 2
6 3
We define the Fourier transform of the wave-packet:
Z
)ej d
A(z,
A(z, ) =
(2.5)
(2.6)
In the frequency domain, Eq (2.5) becomes:

A
2 2 3 3
= j 1 + + A
z
2
6
(2.7)
We see that the equation is easily solved in the frequency domain:
) = A(0,
)ej( 22 2 + 63 3 )z
A(z,
(2.8)
As a conclusion, chromatic dispersion manifests as the accumulation of a quadratic phase

for the second order dispersion and a cubic phase for third order dispersion in the frequency
domain and can be inverted by multiplication with the exponential of the opposite phase
in frequency domain. In the fiber-optic community, CD is generally quantified as the
2.4. Transmission impairments in optical fibers
11
dispersion factor D [ps/nm.km] which is linked to 2 and the central wavelength of the
signal in the following way:
2c
(2.9)
D = 2 2
As an example, the dispersion factor is given as a function of wavelength for standard

singlemode fiber (SSMF) [9]:
4z
)
(2.10)
3
In this equation, z = 1320 nm is the zero-dispersion wavelength and Sz = 0.092
ps/(nm2 km) is called the zero dispersion slope.
D() Sz (
2.4.3
Polarization mode dispersion
w
slo
fast axis
Birefringence is the dependency of the refractive index of the fiber on the signal polarization. This phenomenon is caused by fabrication imperfections, stress and temperature
variations. In the presence of birefringence, the propagation constant depends on the
polarization of the light. At each axial position z, the fiber is characterized by an eigenmode corresponding to the field polarization with the slowest propagation constant [10].
At the same time, the orthogonal eigenmode is the mode with the highest propagation
constant. The birefringent fiber model is illustrated in Fig. 2.4. The distribution of the
eigenmodes as a function of distance is stochastic in nature and changes slowly with time.
For singlemode fibers, the correlation length, i.e. the distance after which the eigenmodes
have completely independent orientation, is in the order of 100 m [11].
is
ax
Figure 2.4: Fiber model of randomly birefringent sections

Looking at Eq. (2.3), 0 (z) is the difference in propagation constant between the slow
and fast polarization axes at the central pulsation. 1 (z) is the slope of the propagation
constant difference as a function of pulsation. The Jones matrix J(z), which is defined
on the eigenmodes at position z, acts as projector on the eigenmodes at position z. As
a consequence of PMD, the pulses transmitted on the two polarization modes are both
mixed and broadened. The
strength of PMD in a fiber is generally expressed by the
PMD coefficient, in [ps/ km]. The PMD coefficient, multiplied by the square root of
the fiber length, gives the mean differential group delay between orthogonal polarization
states. Modern optical fibers have a PMD coefficient in the order of 0.05 [ps/ km]. The
distribution of the eigenmodes along the fiber also varies with time on a ms scale [12, 13]
which means that PMD parameters must be tracked at the receiver.
2.4.4
Nonlinear interactions
The optical Kerr effect is the main source of nonlinear distortion in optical communication
systems [14]. It originates in the nonlinear response of the electronic polarization of the
12
medium. As a consequence, the refractive index depends on the power of the electric
field propagating through the fiber. Other nonlinear effects arise from the interaction of
the optical field with the silica molecules: stimulated Raman scattering and stimulated
Brillouin scattering [8]. These nonlinear effects are due to the inelastic scattering of
photons towards lower energy photons. The energy difference is absorbed by molecular
vibrations or phonons in the silica. However, nonlinear Raman or Brillouin scattering is
only efficient above power thresholds that are not obtained in the framework of this thesis
and are therefore ignored.
In Eq. (2.3) the right-hand terms under-braced with (4) define the Kerr effect. The
transmission impairments suffered by a WDM complex propagation through an optical
fiber and stemming from the optical Kerr effect can be categorized:
Self Phase Modulation. The electromagnetic field of one optical channel modifies
the refractive index according to its own instantanenous power. As a consequence,
the field is phase shifted according to Eq. (2.3).
Cross Phase Modulation. Similarly to Self Phase Modulation, Cross Phase Modulation originates from the dependency of the refractive index on the instantaneous
optical power. Here, all co-propagating WDM channels induce a phase modulation.
Four Wave Mixing. Two photons of different energies may be annihilated and two
photons of different energy appear if the net energy and momentum of the transformation are conserved [8]. Phase matching between the different spectral components
is required for the Four Wave Mixing interaction to be effective. On the macroscopic
scale, Four Wave Mixing manifests as a non-linear interaction between up to four
WDM channels.
In addition to this general classification, Eq. (2.3) shows that nonlinear interaction also
occurs between polarization components both in intra-channel and inter-channel frameworks, an effect called Cross Polarization Modulation.
The existence of inter-channel effects such as Cross Phase Modulation and Four Wave
Mixing make the Kerr effect very difficult to compensate in the digital domain because
the information sent over the neighboring WDM channels is generally unknown at the
receiver. To this date, the Kerr effect is not completely compensated by signal processing
in coherent optical communication systems [15].
2.5
The fiber-optic communication system
Fig. 2.5 summarizes the typical architecture of an optical point-to-point long-haul communications link. The information to be transmitted takes the form of a binary sequence.
This sequence undergoes digital signal processing before being translated into an analog
waveform. A continuous laser wave is modulated by this waveform and the resulting signal is fed into the optical fiber. Typical transmission distances range from 1000 to 5000
km for terrestrial transmission links. At the receiver, the optical signal is transformed
back into the electrical domain by photodiodes. The resulting analog waveform is sampled. The digital sequences undergo signal processing again before they are converted
2.5. The fiber-optic communication system

Transmitter
13
Transmission link
Receiver
Photodiode
Fiber
...1010011...
DAC
Ampli
Modulator
Laser
Demodulator
!N
Decision
ADC
...1010011...
Laser
Figure 2.5: Architecture of the fiber-optical communication link

back into a binary sequence. In this section, we introduce the relevant aspects of the
physical long-haul communication system.
2.5.1
Laser sources
Laser sources for telecommunications must operate continuously for long times at room
temperature. They are of small diameter because the light must be fed into the core of
the fiber. The most commonly used sources are light-emitting diodes and semiconductor
lasers. In a semiconductor laser, the performance is mostly limited by spontaneous emission. This results in both amplitude and phase noise and thus broadening of the laser
linewidth. When coherent detection is used (see section 2.5.5), the phase noise beating of
transmitter and receiver lasers imprints itself on the modulated signal. As a consequence,
the received signal shows a phase drift in accordance with the laser linewidths. The laser
frequency noise of telecommunication lasers is commonly assumed a Brownian movement
which results in a lorentzian line shape [16]. The typical linewidth of optical communication lasers is in the MHz range [17]. The frequency noise f (t) is easily computed
as white noise with a spectral density equal to /, where is the full width at half
maximum linewidth. The optical field of the laser is then given in complex notation:
E(t) = E0 ej(2f0 t+(t))
Z t
(t) =
f (t0 )dt0
(2.11)
(2.12)
In this equation f0 is frequency at the maximum of the lorentzian line.

Although the laser frequency can be controlled directly based on a feedback loop,
this type of design is usually complicated and inefficient. Free-running laser sources also
exhibit slow fluctuations of the emitting frequency over the day [18]. The accuracy of
temperature-stabilized lasers is approximately 1 GHz [19].
2.5.2
Optical modulation
The function of the modulator is to convert the electrical data into the optical domain. In
most long-haul high-speed transmission systems, distributed feedback lasers are externally
modulated using a Mach-Zehnder modulator (MZM). The MZM is an interferometer
composed of two 3-dB couplers and two waveguides with equal lengths as shown in Fig. 2.6.
The incoming optical energy is equally divided between the two arms. Each arm comprises
14
an electro-optical cell which induces a phase-shift between the optical signals in each arm
depending on the applied voltage called drive voltage. These phase-shifts translate into
amplitude fluctuations once the optical signals on the two arms are recombined. As a
consequence, the electrical drive voltage can be translated into optical modulation of the
continuous wave. The input-output characteristic of the MZM is given by a cosine function
of the difference of the drive voltages V1 and V2 . The drive voltage difference leading from
maximum output power to zero is called V . When the variations of V1 V2 are centered
around V /2 and small with respect to V , the characteristic can be considered linear and
the distortion between the drive voltage input and the optical output minimized.
Two MZM can be combined in order to imprint complex information on the optical
wave. Again, two 3-dB couplers and two waveguides are used to equally divide the optical
energy. The real("I") and imaginary ("Q") parts of the information stream I + iQ are fed
into their respective MZMs. After phase shifting the imaginary arm by /2, the optical
fields are recombined. This setup is called an IQ-modulator. In polarization-multiplexed
transmission systems (see section 2.2), two IQ-modulators are used. They are followed
by polarizers which impose a linearly polarized state of polarization on two orthogonal
transverse axes. The two signals are then combined in a polarization combiner to compose
the polarization-multiplexed optical signal.
waveguide
V1
electrode
Light out
Light in
3-dB coupler
3-dB coupler
V2
Figure 2.6: Mach-Zehnder modulator
2.5.3
Digital modulation formats
In the previous paragraph, we described how different degrees of freedom - amplitude,

phase, polarization - can be used to transmit information on optical waves. This paragraph deals with the modulation formats, i.e. the way binary sequences can be digitally
translated into these entities before the optical modulation. We review the main systemrelevant modulation formats. They are illustrated in Fig. 2.7.
On-Off Keying with direct detection (OOK-DD). This modulation format is used
in most legacy networks operating at 10 Gbit/s without detecting the phase of the
signal ("direct detection"). While historically important, it is no longer competitive
since neither polarization modes nor optical phase are exploited, which results in a
limited spectral efficiency.
OOK
15
BPSK
R
QPSK
R
16-QAM
Figure 2.7: Modulation formats for optical communications

Polarization Division Multiplexed Binary Phase Shift Keying (PDM-BPSK). With
the advent of coherent detection and digital signal processing, PDM has become
a standard feature permitting to double the data rate at the same symbol rate
as OOK [20]. BPSK encodes one bit per symbol per polarization. Comparing to
higher-order modulation formats, PDM-BPSK is therefore the most suitable choice
for long-distance links with moderate data rate such as 40 Gbit/s [5].
Polarization Division Multiplexed Quaternary Phase Shift Keying (PDM-QPSK).
Encoding two bits per symbol per polarization, PDM-QPSK is the key modulation
format in order to achieve a 100 Gbit/s data rates while keeping the symbol rate
as low as 28 Gbd/s on each optical channel [21]. QPSK and BPSK are part of
the M -phase shift keying modulation format family, characterized by a constant
modulus and equally spaced phase values. This geometry is particularly interesting
for receiver DSP polarization demultiplexing (see section 3.4) and phase equalization
(see section 3.6).
Polarization Division Multiplexed 16-Quadrature Amplitude Modulation (PDM16QAM). To further increase the data rate per optical channel beyond 100 Gbit/s,
high-order modulation formats are currently researched [22, 23]. PDM-16QAM is
illustrated in Fig. 2.7 as an example. The main challenges of these high-order modulation formats are reduced transmission distances due to increased noise and phase
sensitivity with respect to PDM-QPSK [24].
In this listing, we only consider modulation formats in which the two polarization
components are independent (PDM). In fact, the noise sensitivity of PDM-QPSK can be
16
further optimized over the three degrees of freedom phase, amplitude, polarization using
a polarization coding scheme called polarization switched quaternary phase shift keying
(PS-QPSK) [25, 26], which is beyond the scope of this work.
2.5.4
Loss compensation
Nowadays, long optical transmission distances are achieved by periodically incorporating

optical amplification in order to compensate fiber loss and to amplify the whole WDM
multiplex at once. The transmission link consists of consecutive fiber spans separated by
optical amplifiers. Optical amplifiers enhance the weak input signal from the previous span
and launch it into the next span at a high power. Their principle of operation is stimulated
emission. In the optical amplifier medium, electrons are pumped to a high energy level.
This results in population inversion, i.e. a predominance of high-energy electrons between
two energy levels. In these conditions, stimulated emission can take place: the incoming
photons are duplicated during their transmission through the amplifier. At the same time,
spontaneous emission corrupts the signal. Spontaneous photons are decorrelated from the
signal and have arbitrary frequency and polarization. Moreover, the spontaneous photons
are also amplified along the link (ASE for amplified spontaneous emission). The span
length varies depending on system configuration but most of the terrestrial systems use
spans between 80 and 100 kilometers. The most widely used type of amplifier is the
erbium-doped fiber amplifier [27]. It consists of a singlemode fiber of about 10 meters
doped with Erbium ions which is pumped using a laser to achieve the population inversion
necessary to amplify the signal through stimulated emission. The amount of corruption
caused by optical amplification on the transmitted signal is quantified by the optical signal
to noise ratio (OSNR):
OSNR =
Psignal
Pnoise
(2.13)
Where Psignal is the signal power and Pnoise the noise power. The OSNR is usually
normalized to the reference bandwidth of 0.1 nm. Through this work, the signal to noise
ratio is noted as OSNR(0.1nm) .
2.5.5
Coherent detection
Coherent detection relies on detecting a signal through its beating with a reference carrier
supplied by a local oscillator (LO) [28]. As a consequence, both phase and amplitude of
the optical field can be detected. It follows that the information to be transmitted can be
encoded on both amplitude and phase of the optical signal as in the digital modulation
formats introduced in section 2.5.3. While legacy networks still use a simple photodiode
(direct detection) to detect channels at 10 Gbits/s, newly installed transmission links
make use of coherent detection to enhance the throughput of the link and achieve channel
rates of 100 Gbit/s and more. In this work, only coherent detection will be considered.
The architecture and operation of a coherent mixer is explained in Fig. 2.8.
Be ECohM ix (t) the optical field at the entry of the coherent mixer. Be ELO (t) the
unmodulated light of the LO. Both signals first pass polarizers (45 in Fig. 2.8). At the
17
LO
45
/4
Signal
45
4
1
Figure 2.8: The coherent mixer

output of the polarizer, polarized light at 45 between the horizontal ( ) and vertical ()
axes is obtained. Thanks to polarization beam splitters, the PD1 and PD3 photodiodes
receive the polarization whereas the PD2 and PD4 photodiodes receive . As an
example, we calculate the signal received by PD1:
Useful signal The signal is reflected and accumulates a /2 phase shift
Esignal,out (t) = ECohM ix (t)ej/2 ejc t
(2.14)
Local oscillator The signal of the LO goes through the quarter wavelength plate and is
not reflected:
Eol,out (t) = ALO (t)ej/2 ejol t
(2.15)
The output current IP D1 of the photodiode is proportional to the incident power of light.
We define ECohM ix (t) = ACohM ix (t)ej(t)
IP D1 (t) (Es,out + Eol,out )(Es,out + Eol,out )
(2.16)
(2.17)
= |ACohM ix | + |Aol | + 2|ACohM ix ||Aol | cos((c ol )t + (t))

Doing the same for IP D2 , IP D3 and IP D4 yields:
I1 (t) = IP D1 (t) IP D3 (t) = 4|ACohM ix ||Aol | cos((c ol )t + (t))
(2.18)
I2 (t) = IP D2 (t) IP D4 (t) = 4|ACohM ix ||Aol | sin((c ol )t + (t))
(2.19)
The currents I1 (t) and I2 (t) represent the in-phase and quadrature components of the
incoming signal [29]. The frequency offset between the LO and the signal c ol is
corrected via digital signal processing (see section 3.5). In a practical coherent mixer,
the coherent architecture is duplicated as shown in Fig. 2.9. The dual-polarization signal
is separated into two polarization components fed into two coherent mixers. In any
realistic system, the transmitter and receiver polarization components are not aligned. All
polarization effects can however be corrected via digital signal processing (see section 3.4).
18
LO
MIXER
1
x,1 (t)
PBS
Signal
MIXER
2
ADC
x,2 (t)
x,3 (t)
ADC
ADC
x,4 (t)
x,RX [m]
ADC
Sy,RX [m]
Figure 2.9: Coherent detection for polarization-multiplexed systems
2.5.6
Trade-off between noise and nonlinear distortions
is
e
by
ite
d
ts
c
ffe
re
ea
lin
NLT
Li
m
no
AS
E
by
no
ite
Li
Q-factor [dB]
As we will see in the subsequent chapters, many linear transmission impairments in optical fibers can be mitigated using digital signal processing (hereafter abbreviated DSP).
The system performance is ultimately limited only by ASE noise and the nonlinear Kerr
effect. When the input power per optical channel is low, the system is limited by noise.
However, a high input power will result in unacceptable performance penalties stemming
from nonlinear distortions. As a consequence, a trade-off must be found between the
noise-limited and the nonlinear limited regimes as illustrated in Fig. 2.10. The nonlinear
threshold is defined as the power resulting in optimum performance. Its value depends
on the system architecture but also on DSP (modulation format, presence of a mitigation
algorithm for nonlinear effects, etc.).
Link input power per channel [dBm]
Figure 2.10: Performance as a function of the link input power
2.6. Mathematical description of the fiber-optic communication system
2.6
2.6.1
19
Mathematical description of the fiber-optic communication system

Nonlinear fiber simulation and linear transfer function
In order to calculate the optical field at the output of an optical fiber given an input
field, the nonlinear Schrdinger equation Eq. (2.3) must be solved. The most commonly
used technique to propagate the optical field is called the split-step Fourier method or
SSFM [8]. Its principle is to propagate the field over successive fiber segments or "steps".
In each step, one assumes that the linear terms of Eq. (2.3) (contributions (1), (2) and
(3)) can be applied separately from the nonlinear terms (contribution (4)). The SSFM
then consists in applying in each step the two operators separately (splitting). FFT and
IFFT are used to compute the linear part of the step in frequency domain. Split-step
methods are easily accommodated for taking into account polarization effects. The steps
are then used to emulate the fiber model of randomly birefringent sections as described
in section 2.4.3. In practice, a random Jones matrix J(z) and birefringence parameters
0 (z) and 1 (z) (see Eq. (2.3)) must be drawn for each fiber segment. More details
on this computation can be found in [20]. The step size of the SSFM triggers its accuracy
and is often optimized empirically, important parameters include the coherence length of
the PMD, i.e. the distance over which the principal states of polarization of the fiber stay
constant and the nonlinear coefficient . In simulations for fiber-optic communications
with SSMF, the step size is in the order of 100-200 meters.
In fiber-optic communication systems, the nonlinear effects usually have a significant
effect on the transmitted signal because of the ASE noise/Kerr effect trade-off pointed
out in section 2.5.6. However, only linear effects are generally mitigated using digital
signal processing. It is therefore interesting to write out the solution of Eq. 2.3 ignoring
the nonlinear term (4). In this derivation, we will further ignore attenuation (1) and
write out the Jones components as A = [A1 A2 ]T . Over the l-th split-step of length h,
where the birefringence parameters can be considered constant, Eq. 2.3 writes out in the
frequency domain:
A1 (z + h)
A2 (z + h)

1 (z)
H11 (z) H12 (z)
A
=
H21 (z) H22 (z) . A2 (z)
|
{z
}
(2.20)
H[l]
Where
0 (z)
1 (z)
2 2
3 3
.
Hii (z) = ej( 2 + 2 )Jii (z) 2 6 )h
(2.21)
0 (z)
1 (z)
.
Hik (z) = ej( 2 + 2 )Jik (z))h i 6= k
(2.22)
And
The input-output relation can then be obtained by concatenation:

Y
A1 (L)
A1 (0)
=
H[l].
A2 (L)
A2 (0)
l=1:L/h
(2.23)
20
The fiber impulse response is defined as the inverse Fourier transform of Eq. 2.20.
2.6.2
Mathematical models of the optical communication chain
The object of this section is a mathematical description of the optical transmission link
shown in Fig. 2.6.2 where polarization multiplexing is ignored for the sake of simplicity.
If the effects of transmitter, channel and receiver can be assumed linear, their transfer
functions can be described as impulse responses [29]. The fiber impulse response has been
defined in the preceding section. In a real transmission line, noise is added progressively
at each amplification stage and is propagated and amplified along the transmission line.
Here, we use noise loading at the end of the transmission line. This assumption is valid in
a linear model where the noise is assumed to be additive, white and Gaussian. Different
studies [30, 31] proved this assumption to be verified, so that we systematically adopt it
in this work. In Fig. 2.6.2 (b), the noise is represented by n(t), a continuum of white
Gaussian noise. Modulator and demodulator impulse responses are called g(t) and p(t)
respectively. The analog system model between DAC and ADC operations can now be
written out:
z(t) = g(t) (h(t) p(t) s(t) + w(t))
(2.24)
Where is the continuous convolution operator.

(a)
Fibre
DAC
Ampli
Modulator
Laser
!N
ADC
Demodulator
Laser
(b)
DAC
s(t)
g(t)
h(t)
p(t)
z(t)
ADC
w(t)
(c)
s[n]
g[n']
h[n']
p[n']
z[n ]
w[n']
Figure 2.11: (a) Architecture of the communication chain (b) Continuous input-output
model (c) Digital input-output model
We can define the equivalent digital channel model [29] in Fig. 2.6.2 where the operations of pulse shaping, channel and receiver filtering are identified as convolution operations with digital FIR filters g[n0 ], h[n0 ] and p[n0 ] which are sampled version of g(t),
2.7. Summary
21
h(t) and p(t) respectively at twice the symbol rate. In order to avoid aliasing, the symbol
sequences are up ( 2)- and down-sampled ( 2) by two. w[n0 ] are the digitized noise
samples. We can define the input-output model of the communication chain:
0
z[n ] =
s[m]q[n0 m] + w0 [n0 ]
(2.25)
q[n0 ] = (g h p) [n0 ]
(2.26)
m=
w [n] = (w p) [n ]
(2.27)
(2.28)
The digital input-output model is easily generalized to 2 2 multiple-input multipleoutput for taking into account PDM:
0
zi [n ] =
X X
si [m]qik [n0 m] + wi0 [n0 ]
(2.29)
qik [n0 ] = (g hik p) [n0 ]
(2.30)
k=1,2 m=
wi0 [n]
= (wi p) [n ]
(2.31)
(2.32)
In this equation, zi [n] is the received signal on polarization i, hik [n] is the inverse Fourier
transform of Hik (t) (see preceding section), sampled at twice the symbol rate. wi [n0 ] is
the sampled noise on polarization i. In optical communications, the signal to noise ratio
is often identified by the OSNR(0.1nm) (see section 2.5.4) which is particularly suitable for
measurements. In numerical simulations, Eb /N0 , the energy per bit divided by the noise
spectral density, is very useful because it takes into account the modulation format. It
can be calculated from the relative variances of received signal and noise in Eq. 2.29 and
the number of bits per symbol K. By definition, the noise power can be decomposed
into (noise spectral density: N0 ) (spectral width of the receiver filter fs ). Similarly,
we decompose the signal power into (Energy per symbol: Es ) (spectral spectral width
of the signal fs ). Here, we supposed that the receiver filter and signal have the same
bandwidth.
Es fs
Es
Eb
< |z[n]|2 > < |w[n]|2 >
=
=
=
2
K < |w[n]|
KN0 fs
KN0
N0
2.7
(2.33)
Summary
In this chapter, we review the principles of optical telecommunications. We describe the

optical fiber as a transmission medium. We detail the main impairments corrupting the
signal in the fiber channel: CD, PMD, ASE noise and the nonlinear Kerr effect. We outline the architecture of a point-to-point long-haul optical telecommunications link using
coherent detection. We detail the principle of optical modulation and introduce digital
modulation formats. We introduce a mathematical model of the fiber-optic communication link.
Chapter 3
Digital signal processing in coherent
optical communication systems
Contents
3.1
3.1
Introduction
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
23
3.2
Time/frequency acquisition . . . . . . . . . . . . . . . . . . . . . .
24
3.3
Chromatic dispersion compensation . . . . . . . . . . . . . . . . .
25
3.4
Polarization demultiplexing . . . . . . . . . . . . . . . . . . . . . .
26
3.5
Carrier frequency estimation . . . . . . . . . . . . . . . . . . . . . .
27
3.6
Carrier phase estimation . . . . . . . . . . . . . . . . . . . . . . . .
28
3.6.1
Viterbi-Viterbi carrier recovery algorithm . . . . . . . . . . . . . . . 28
3.6.2
Maximum-Likelihood Phase Estimator . . . . . . . . . . . . . . . . . 28
3.7
Symbol identification and FEC . . . . . . . . . . . . . . . . . . . .
29
3.8
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
30
Introduction
In optical telecommunications, the development of DSP is closely related to technological

advances. Its use in association with coherent detection permitted a tremendous increase
in spectral efficiency since the 1990ies [32]. Thanks to coherent detection, both phase
and amplitude of the optical signal can be detected (see chapter 2). As a consequence,
complex and high-order modulation formats can be used [33]. DSP is needed to correct the
transmission impairments for the complex modulation formats. In addition, in order to
benefit from polarization multiplexing, cross-polarization DSP is compulsory [34]. Finally,
before the advent of DSP in optical communications, dispersion compensating fiber were
used to compensate for in-line CD [35]. Thanks to the digital compensation of CD,
dispersion-unmanaged (DU) links have become standard use for newly installed links [36].
Today, the performance of DSP benefits from continuously increasing electrical processing
capacities.
24
Chapter 3. Digital signal processing in coherent optical communication

systems
s1 [n]
g[n']
hik [n ]
s2 [n]
z1 [n ]
p[n']
r1 [n]
eik [n ]
w1 [n ]
g[n']
z2 [n ]
p[n']
r2 [n]
w2 [n ]
Figure 3.1: Digital transmission model and equalization

Fig. 3.1 presents the digital channel model defined in the preceding chapter for a
polarization-multiplexed link. The role of the cross-polarization equalizer eik is to recover
the data symbols si [n]. If the channel impulse responses are known, we can compute the
equalizer guaranteeing the minimum mean square error between si [n] and ri [n]. It is most
conveniently expressed in matrix form. Be Hik (z) the z-transform of hik [n0 ] and Eik (z)
the z-transform of the equalizer. We define MIMO matrices H(z) and E(z):

H(z) =

E(z) =
H11 (z) H12 (z)

H21 (z) H22 (z)
E11 (z) E12 (z)

E21 (z) E22 (z)
(3.1)
(3.2)
The linear minimum mean square error (MMSE) equalizer is:

E(z) = (HH (z)H(z) +
n2
H(z))1 HH (z)
s2
(3.3)
In this equation, n2 is the noise variance and s2 the signal variance.

In principle, Eq. 3.3 solves the question of DSP in optical transmission systems. However, the parameters of several impairments present in these links cannot be known a priori
at the receiver. Phase effects vary very rapidly. CD is known and constant but requires
very long equalizer filters. This variance in impairment results in a DSP architecture
much more complex than shown in Fig. 3.1.
The common configuration of optical coherent receiver associated with DSP algorithms
is shown in Fig. 3.2 [37, 38]. Due to the specificity of the linear impairments in optical
communications, a block architecture has been widely adopted. Each of the blocks addresses only one linear impairment. The architecture is described in the sections of this
chapter.
3.2
Time/frequency acquisition
After passing through the dual-polarization coherent mixer (see section 2.5.5), the electrical signal is first sampled by analog-to-digital converters (ADC). In order to avoid
3.3. Chromatic dispersion compensation
25
LO
MIXER
1
PBS
Signal
MIXER
2
ADC
ADC
ADC
CD
comp
z1
e11
r1
CFE
/
CPE
Decision
CFE
/
CPE
Decision
e12
e21
ADC
CD
comp
z2
e22
r2
Pol Demux
Figure 3.2: Main DSP blocks in the state-of-the-art coherent receiver

aliasing, the sampling frequency must exceed the bandwidth occupancy of the signal [29].
The signal is usually down-sampled to one sample per symbol after equalization of the
main channel perturbations. If the sampling rate is a multiple of the symbol rate, no interpolation is needed to achieve the conversion. As a consequence, the sampling frequency
is often chosen as twice the symbol rate.
3.3
Chromatic dispersion compensation
Chromatic dispersion, as introduced in section 2.4.2, is a static impairment with a wellknown expression in the frequency domain (Eq. (2.8)). As can be observed in the nonlinear
Schrdinger equation (2.3), it is a polarization-independent phenomenon. Moreover, being
directly determined by the length of the link and its type of fiber, the strength of CD
is roughly known at the receiver [39]. Moreover, CD compensation represents the most
complex algorithm in DSP for coherent optical communications [40,41]. These arguments
justify the implementation of an exclusive CD compensation block ("CD comp") as shown
in Fig. 3.2. It is the only static block of the DSP suite.
Using Eq. (2.8) and the expression of the dispersion parameter D from Eq. (2.10), the
frequency-domain filter to compensate for CD is easily designed [42]:
G() = exp(j
2 z 2
)
4c
(3.4)
In this equation, is the centralwavelength of the optical signal, c is the speed of light
z is the fiber length, and j = 1. In practice, the filter Eq. (3.4) is truncated to
the effective bandwidth of the signal and a static finite impulse response (FIR) filter
is designed based on the truncated response. CD compensation can be implemented
in the time domain (TD) or the frequency domain (FD). In a TD implementation, the
signal is convolved with a finite impulse response (FIR) filter implementing a truncated
inverse Fourier transform of Eq. (3.4) [43]. In a FD implementation the incoming data
is transformed with a fast fourier transform (FFT). The result is multiplied with the
filter frequency response Eq. (3.4) and the signal is transformed back into TD by inverse
fast fourier transform (IFFT). The number of taps of the TD equalizer or the size of the
FFT in the FD equalizer is directly proportional to the amount of residual CD [44]. For
26

systems
block k-1
block k
block k+1
input data
overlap of K/2 samples
overlap of K/2 samples

FFT(N)
G(N)
IFFT(N)
discard K/2 samples

output data
result k-1
discard K/2 samples

result k
result k+1
Figure 3.3: Principle of overlap and save FD CD equalization

a large number of equalizer taps, the complexity of implementation is much lower for
the FD equalizer [40, 45]. This is the case in most realistic link scenarios and a fortiori
in tranmission links without in-line CD compensation. In order to overcome the cyclic
properties of the FFT, the FD CD equalizer is usually implemented using the overlap and
save method [46]. Its principle is illustrated in Fig. 3.3. The incoming data is transformed
with an FFT of length N samples Then the FFT result is multiplied with a frequency
representation of the filter impulse response G and finally the data is transformed back
to TD. K/2 samples are discarded at the beginning and the end of the block respectively
to yield result vector of size N K. In the next step, the FFT is applied to the incoming
data vector delayed by K samples. The overlap and save filtering is applied separately
on both polarization components.
Optical communication systems using electronic CD compensation have a tremendous
resilience to residual chromatic dispersion: given a sufficient size of the equalizer, virtually
any amount of residual CD can be compensated for at the receiver without performance
penalty [47].
3.4
Polarization demultiplexing
The DSP block following the CD compensation targets polarization-dependent effects is

depicted as "Pol Demux" stage in Fig. 3.2. In coherent optical communications using
PDM, polarization-dependent effects (see section 2.4.3) result in mixing and distortion of
the two orthogonal polarization tributaries. In the polarization demultiplexing step of the
receiver DSP the two orthogonal polarization tributaries are recovered using FIR filters
arranged in butterfly structure as shown in Fig. 3.2. The characteristics of the PMD
effects are generally not known at the receiver. Moreover, they vary over time [48, 49].
DSP algorithms to tackle PMD can be data-aided or so-called "blind" algorithms [50]. In
data-aided equalization, part of the communication bandwidth is used to transmit known
training symbols and derive an estimation of the PMD coefficients. For time-varying
channels, the estimation must be regularly repeated. The main drawback of data-aided
equalization is the reduction of transmission capacity due to the training symbols [51].
In opposition, blind adaptive algorithms rely on the statistics of the received waveform
3.5. Carrier frequency estimation
27
to progressively train the equalizer thanks to a error signal that is fed back to adapt
the coefficients of the algorithm [52]. The main drawback of adaptive equalization is
slowness and the risk of divergence of the algorithm [53]. While designed for PMD, the
butterfly stage of the receiver DSP is not restricted to its equalization. Sampling time
offsets or residual CD are other examples of channel perturbations taken care of in this
DSP stage [54]. We denote by zj [n] n = 1... the complex input of the butterfly FIR
filter. zj [n] is a mix of the two polarization tributaries signals sent at the transmitter.
The taps of the butterfly FIR filter between polarizations i and polarization j are noted
eij [p], p = 1...P and rj [n], n = 1... is the complex equalized symbol sequence on receiver
polarization j. We can express the equalization stage in the two following equations:
rj [n] =
P
XX
eij [p] zi [n p] j = 1, 2
(3.5)
i=1,2 p=1
The CMA (Constant Modulus Algorithm) [55] is the most universally used polarization
demultiplexing algorithm in coherent optical communications [56]. It makes use of the
constant amplitude of certain modulation formats such as BPSK and QPSK. The standard
CMA uses the stochastic gradient algorithm as its update device. For each index n, the
update algorithm is as follows:
eij [p] eij [p] + j [n] rj [n] zi [n] p = 1...P, i, j = 1, 2
(3.6)
In this equation, is the update coefficient of the stochastic gradient and j [n] is the error
function on polarization j. In the case of the CMA, continuously updated error signal is
the difference between the modulus of the output of the FIR filters rj [n] and a constant
C:
j [n] = C |rj [n]|2 j = 1, 2
(3.7)
If the sampling rate is twice the symbol rate, the metric in Eq. (3.7) is updated every two
samples, and one sample in two is dropped at the output of the algorithm. At the output
of the butterfly DSP block, the two generated signals are polarization demultiplexed and
the linear distortions occurred within the transmission line are equalized [32].
The error signal in Eq. (3.7) completely ignores the phase of the equalized signal.
However, rather than being a drawback, the simple update mechanism has been proven a
very robust and versatile algorithm for coherent optical communications [33, 5759]. Although it was originally designed for modulations with a constant modulus such as BPSK
or QPSK, the CMA is now used for high-order modulation format such as 64QAM [60]. In
fact, even if the constant-modulus criterion is not exactly met, the high robustness of the
CMA ensures satisfactory polarization demultiplexing and impairment mitigation [61].
Because of the phase independence of the CMA, all linear impairments are equalized
save phase offsets after polarization demultiplexing. All phase offsets are taken care of in
the last DSP block by specialized algorithms.
3.5
Carrier frequency estimation
In coherent transmission systems, the frequency offset between the transmission laser and
the local oscillator is an inevitable problem. In fact, although this frequency mismatch
28

systems
can be compensated by controlling the laser frequency directly based on a feedback loop,
this type of design is usually complicated and inefficient. As a consequence, free-running
lasers are preferred and the frequency offset is corrected by DSP.
The accuracy of temperature-stabilized lasers is approximately 1 GHz [19]. In a popular approach [19] for coarse frequency offset estimation, the incoming signal is processed
in a sliding window approach. For each window, the P th power, where P is the modulation order, is applied to the signal. This way, the modulation of the signal is eliminated.
After that, the signal is translated to FD by FFT. The index of the maximum sample is
indicative of the frequency offset.
3.6
Carrier phase estimation
The carrier phase estimation removes the remaining phase mismatch between the signal and the local oscillator. For P -PSK modulation formats, the Viterbi-Viterbi algorithm [62] is the standard approach for phase control. For high-order modulation formats,
the maximum-likelihood (ML) phase estimator is often used. Both are described in the
following paragraphs.
3.6.1
Viterbi-Viterbi carrier recovery algorithm
In P -PSK modulation formats, the estimated phase offset (n)

on symbol n is calculated
as follows:
U/2
X
1
(n) = arg
(3.8)
rjP (n + u)
P
u=U/2
First, the P th power of the symbol is taken, thereby removing the modulation. Then,
the results is summed over a certain number of precursor and postcursor symbols. The
argument is computed, and the resulting phase is divided by P to correct for the P th
power elevation. Finally, the phase is unwrapped.
3.6.2
Maximum-Likelihood Phase Estimator
For non-P -PSK modulations, the Viterbi-Viterbi algorithm cannot be used since the
modulation can no longer be removed by taking the P -th power. In this case, an algorithm
that has come to be known as maximum likelihood phase estimator or ML [63] in the
optical community is a widely popular choice for phase tracking.
In digital telecommunications, the maximum likelihood criterion refers to the maximization of the probability of manifestation of the received information sequence with
respect to all possible sent sequences [29]. The most straightforward method to achieve
ML detection is to generate all possible sequences and minimize the accumulated euclidian distance to the received sequence. Of course, this method is prohibitively complex.
There are algorithm achieving ML detection at a much lower computational cost [6466].
In optical communications, ML phase estimation refers to the straightforward implementation of ML detection taking into account only phase shifts (since all other channel
3.7. Symbol identification and FEC
29
....
Rotation
(k 1)
rj [n]
Rotation
(k)
Rotation
(k + 1)
DD error
estimation
Average
DD error
estimation
Average
DD error
estimation
Average
Arg
MIN
[n]
....
Figure 3.4: Maximum-likelihood phase estimation

impairments have already been equalized prior to phase tracking) and assuming a memoryless channel. The "trial and compute probability" approach is extended to the current
phase state of the channel. The ML phase estimation algorithm is illustrated in Fig. 3.4.
In practice, test phases are computed from to . The rotational /2-symmetry is taken
into account for QAM modulations. The number of test phases determines the phase resolution of the algorithm. The incoming signal is rotated by each test phase. The euclidean
symbol estimation error is computed for each copy and the results are summed over a certain number of precursor and postcursor symbols to average out the ASE noise. Finally,
the phase the minimizes the mean square error is selected for each symbol. Since the
algorithm is blind, the euclidean estimation error distance must be approximated by the
distance of the symbol with respect to the nearest constellation point (decision-directed
approach, DD). The main drawback of this algorithm is its computational complexity.
3.7
Symbol identification and FEC
A threshold decision is realized on each symbol and the symbol pattern in translated into
bits. In this work, the system performance is evaluated in terms of the bit error rate
(BER, percentage of bit errors). We will also use the equivalent quality factor Q2 :
Q2 = 20 log10

2erfc1 (2 BER)
(3.9)
where erfc1 is the inverse complementary error function and Q2 is in dB.

In all optical communication systems, a forward error correction (FEC) code is applied
to the bit sequence in order to reduce the bit error rate to very small value (< 1014 ) [65].
The operation of the FEC necessitates the inclusion of a overhead in the transmitted bit
sequences. The design and operation of FEC is beyond the scope of this work. However,
its existence dictates our method of performance assessment. In each study presented
in this thesis, we account for a percentage of FEC overhead to yield the net bit rate.
Correspondingly, we use a target pre-FEC Q2 -factor for the considered FEC.
30

systems
The use of coherent detection makes the application of soft-decision FEC possible.
During the last decade, FECs have tremendously evolved towards larger pre-FEC BER at
the cost of a higher overhead [67]. While hard-decision FECs were limited to up to seven
percent overhead and pre-FEC BER in the order of 104 to 103 , the most advanced softdecision low-density parity-check FECs currently in use pick up the BER at approximately
102 at a twenty percent overhead [68].
3.8
Summary
In this chapter, we review the main DSP algorithms used in coherent optical communication. The block structure of the typical DSP suite is motivated by the specificity of the
physical effects encountered in optical communications. The first DSP block is a static
CD compensator in FD using the overlap and save technique. Second, the polarization
components are demultiplexed using adaptive FIR butterfly filters. The most widespread
polarization demultiplexing algorithm, the CMA, implements a stochastic gradient update
optimized according to the constant modulus criterion. Since the CMA stage ignores the
phase information, the varying phase offset between the transmission and receiver laser
can be taken care of in a separate DSP block. The Viterbi and Viterbi algorithm is a
widespread algorithm for P -PSK modulation, and the ML phase estimator for high order modulations. Finally, the symbols are transformed back into binary sequences and
decoded.
Chapter 4
Design rules for root-raised-cosine pulse
shaping
Contents
4.1
4.1
Transmitter digital signal processing . . . . . . . . . . . . . . . . .
31
4.2
Root-raised-cosine pulse shaping . . . . . . . . . . . . . . . . . . .
32
4.3
High-spectral efficiency coherent optical communications . . . .
33
4.4
Simulation setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
35
4.5
Hardware imperfections . . . . . . . . . . . . . . . . . . . . . . . . .
36
4.6
Rolloff factor optimization . . . . . . . . . . . . . . . . . . . . . . .
37
4.7
Choice of FEC and modulation format . . . . . . . . . . . . . . . .
39
4.8
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
39
Transmitter digital signal processing
In digital telecommunications, the conversion from the digital symbol sequence to the
analogical waveform to be transmitted is characterized by the overall transmitter impulse
response g(t) [29]. The impulse response includes digital to analog conversion, driving
circuits for the modulator and the optical modulator itself. Before the advent of DSP in
optical communications, the symbols were directly fed into the driving circuit delivering
the tensions for the MZM. As a consequence, the shape of g(t) was determined by the
electrical driver and could not be optimized. This resulted in signal spectra exhibiting
strong side-lobes. As a consequence, the standard frequency spacing between optical
channel was chosen much larger than the symbol rate. During the late decade, high
speed digital to analog converters (DACs) operating in the range of tens of GHz have
become available so that individualized transmitter DSP becomes implementable in the
laboratory [69]. The symbol sequence can now be processed digitally before going through
modulation. This operation is used to both optimize the spectral shape of the signal
32
Chapter 4. Design rules for root-raised-cosine pulse shaping
[70] and to pre-compensate for transmitter imperfections. The transmitted DSP and
modulation is illustrated in Fig. 4.1.
TX
DSP
DAC
MZM
s[n]
Light out
Light in
MZM
TX
DSP
DAC
Figure 4.1: Transmitter DSP and modulation architecture
4.2
Root-raised-cosine pulse shaping
In conventional WDM systems using the 50 GHz ITU grid and without pulse shaping [71
73], the 10-dB bandwidth of the WDM channel is typically higher than the symbol rate
(see Fig. 4.2).
Power [dB]
Frequency [GHz]
!"
Figure 4.2: Illustration of the difference between optical communications without and
with DSP pulse shaping
In this section, we introduce the digital root-raised-cosine pulse shaping and receiver
filters. We use the digital channel model defined in chapter 2.6 Eq. 2.29.
We first assume the communication chain in Eq. (2.29) to be ideal, i.e. the channel h[n]
is modelled as a Dirac pulse. In this case, it can be shown [29] that the signal-to-noise ratio
is optimized if the transmitter and receiver filters are matched, i.e. complex transposes of
each other: p[n] = g [n]. Intersymbol interference (ISI) is a form of distortion in which
one each symbol interferes with its temporal neighbors. This is an unwanted phenomenon
since the previous symbols have similar a effect as noise, corrupting the communication.
For an ISI free transmission, q[n] has to be a Nyquist pulse with zero crossings at the
points n 6= 0. The raised-cosine filter is a family of filters exhibiting the desired zero
4.3. High-spectral efficiency coherent optical communications
33
crossings. To achieve both matched filtering and zero ISI, the pulse shaping and receiver
filters are consequently root-raised-cosine (RRC) filters (square root of the raised-cosine
in FD). Their expression in FD are given by:
G(f ) =
T
2
1 + cos
if |f |
|f |
1
2T
if
1
2T
1
2T
|f |
otherwise
1
2T
(4.1)
In this equation, T is the symbol period and 0 1 is a free parameter called the
rolloff factor. It can be observed that the RRC filters have a limited support of 1+
and
2T
a 3-dB bandwidth of equal to the symbol rate 1/T .
G(f ) [a.u.]
=0
= 0.2
= 0.4
= 0.6
= 0.8
=1
1
2T
1
2T
g(t) [a.u.]
=0
= 0.2
= 0.4
= 0.6
= 0.8
=1
t [T ]
Figure 4.3: Digital channel model

In this section, both TD and FD representations of the RRC filters are illustrated in
Fig. 4.3 for various rolloff factors. In conclusion, RRC pulse shaping and receiver filters
combine the advantage of zero-ISI filtering and a compact spectrum. As can be seen
in Fig. 4.3, the RRC pulse shape shows significant ripples in TD. Strictly speaking, the
impulse response has an unlimited support for all rolloff factors; however, the TD impulse
response grows progressively broader with an decreasing rolloff factor.
4.3
High-spectral efficiency coherent optical communications
In order to increase the spectral efficiency of optical communication systems, one solution
is to reduce the channel spacing. As the communication channels suffer from inter-carrier
interference (ICI) or crosstalk when the channel spacing is reduced, substantial effort is
currently devoted to optimize pulse shaping for high-SE systems [74]. During the last
years, RRC filtering with = 0 (see Eq. (4.1)) has attracted much research interest.
The technique is known as Nyquist-WDM (N-WDM) in the optical community [7578].
Choosing = 0 ensures minimum spectral support at a given symbol rate. In theory,
the channel spacing can thus be decreased to the symbol rate. In practice, a guard band
is inserted between optical channels to account for laser frequency offset as shown in
Fig. 4.4(b).
34
(a)
f
(b)
1
2T
1
2T
(c)
Figure 4.4: (a) WDM without pulse shaping and large channel spacing (b) Nyquist-WDM
(c) RRC-WDM with = 0.1 and relaxed channel spacing
Because of the infinite TD support and rippled sinc impulse response, N-WDM suffers
from three hardware implementation limitations :
The length of the pulse shaping digital filter is finite in practice. For a given number
of filter taps, the filter will be more heavily truncated for smaller rolloff factors.
The DAC pulse shaping and the receiver ADC sampling are not ideally timed [79].
This phenomenon is called jitter. Since the modulated symbols overlap more for
smaller rolloff factors, the ISI is increased.
The amplitude resolutions (in terms of number of bits) of both DAC and ADC
are finite. The peak to average ratios of RRC-modulated waveforms is larger for
small rolloff factors. As a consequence, the waveform is more heavily distorted as a
consequence of DAC and ADC discretization.
All these constraints translate into ISI and ICI, and therefore affect the performance.
Allowing for nonzero rolloffs relaxes the constraints on the filter length and the tolerable jitter at expense of increasing ICI. As a consequence, there is a trade-off between
spectral efficiency and BER performance. In this chapter, we carry out numerical simulations to optimize the rolloff factor in modulations using RRC pulse shaping. We take
into account transmitter and receiver imperfections. In section 4.4, we explain the simulation setup. Section 4.5 presents numerical results showing the influence of each hardware
imperfection on the performance. In section 4.6, a realistic framework of the overall system limitations is defined and we present numerical results illustrating the performance
trade-off as a function of the rolloff factor. In section 4.7, we use the conclusions of the
preceding sections to study the optimum choice of modulation format and FEC overhead.
4.4. Simulation setup
4.4
s
1 [n]
35
Simulation setup
RRC
()
DAC
Driver
DAC
Driver
IQ
MOD
Laser (-)
s
2 [n]
s1 [n]
RRC
()
RRC
()
DAC
Driver
DAC
Driver
DAC
Driver
DAC
IQ
MOD
LO
IQ
MOD
MIXER
1
Driver
MUX
Laser
ASE
s2 [n]
s+
1 [n]
RRC
()
RRC
( )
R
I
DAC
DAC
Driver
IQ
MOD
Driver
DAC
Driver
DAC
Driver
MIXER
2
ADC
ADC
ADC
ADC
e11
r1 [n]
e12
e21
e22
r2 [n]
Pol Demux:
CMA/DD
IQ
MOD
Laser (+)
s+
2 [n]
RRC
( )
DAC
Driver
DAC
Driver
IQ
MOD
Figure 4.5: The simulation setup. The components with a red background are simulated with imperfections. RRC: root-raised cosine pulse shaper, DAC: digital-to-analog
conversion, IQ-MOD: IQ-modulator, ADC: analog-to-digital conversion, CMA: constant
modulus algorithm, DD: decision-directed, ASE: amplified spontaneous emission, LO:
local oscillator
The simulation setup is explained in Fig. 4.5 and the simulation parameters are listed
in Table 4.1. Pseudo-random binary sequences (PRBS) of length 218 are mapped to
symbol sequences si [n] where i = 1, 2 is the polarization component and n the sequence
index. We carry out simulations with both PDM-QPSK and PDM-16QAM. The symbols
are upsampled by two. They are RRC-shaped by a digital FIR pulse-shaping filter with
Ntaps taps and rolloff . The FIR outputs pass through DACs with Mbits resolution.
The signals are modulated by an ideal IQ-modulator (IQ-MOD) and multiplexed with
WDM channel-spacing f Hz. The amplified spontaneous emission (ASE) is emulated
as additive white Gaussian noise. At the receiver, the signals are mixed with a local
oscillator and sampled at twice the symbol rate by an ADC with resolution Mbits , and
with an electrical front-end frequency response denoted by HADC (f ), where f stands for
36
frequency. We make sure that the overall filter responses of the transmitter and the
receiver are matched, in order to maximize the tolerance to noise. The timing jitter of
ADC and DAC is modeled as a Gaussian process with standard deviation jitter . The
polarizations are demultiplexed using either the constant modulus algorithm (CMA) for
PDM-QPSK as in [80] or a decision directed equalizer (DD) with a stochastic gradient
update for PDM-16QAM as in [24]. We assume perfect CD and zero carrier frequency
offset and carrier phase noise. The signal is downsampled by two to yield the received
symbol sequences rj [n]. After demapping, the bit error rate is estimated by Monte Carlo
simulations, where not only the noise seed is randomly varied, but also the relative delay
and polarization of neighboring channels. We count at least 500 errors, and convert the
bit error ratio into the Q2 -factor. The system performance is quantified by the required
optical signal-to-noise ratio (OSNR) at a given Q2 -factor = Q2target . We use a symbol
rate equal to 32 Gbaud for all simulations. We account for 12 % FEC overhead (OH) for
9
1
Q2target = 8.5 dB. The spectral efficiency is therefore equal to 2P.3210
. 1+O
, where P is
f
the number of bits per symbol (P = 2 for QPSK and P = 4 for 16QAM, multiplied by 2
for PDM), and O is the overhead.
The receiver filter HADC (f ) was RRC, matched to the ideal transmitter pulse shape
and was given the same .
Table 4.1: Simulation parameters
Parameter
Without imperfections
Baud rate [Gbaud]
32
2
Qtarget [dB]
8.5
FEC overhead
12
Filter taps (Ntaps )
DAC/ADC resolution (Mbits )
HADC (f )
RRC rolloff
Jitter std. deviation jitter [ps] 0
4.5
With imperfections
32
8.5
12
16
8
RRC rolloff
1
Hardware imperfections
In optical communication systems, the signal to noise ratio is representative of the transmission distance since the OSNR0.1 nm is degraded with every in-line amplifier: for a
particular choice of link architecture, a given OSNR0.1 nm /Q2 -target couple will correspond to a transmission distance and FEC overhead. As a consequence, we choose to
represent the OSNR0.1 nm for a given Q2 -factor target for the different scenarii.
In Fig. 4.8 we examine each hardware imperfection independently. Fig. 4.8 shows
as an example the OSNR0.1 nm for PDM- 16QAM at a spectral efficiency of 675 % as a
function of a) jitter b) Mbits and c) Ntaps . These effects will affect opt and the penalty at
opt . In Fig. 4.8(a), it can be seen that the minimum of the OSNR8.5 curve varies from 0
to opt = 0.06 when jitter /Ts =0.05, where Ts is the symbol duration. Jitter may have to
be taken into account when optimizing , when jitter /Ts reaches or exceeds 0.05, where
the difference between the penalty at opt and at = 0 starts to be significant, at about
4.6. Rolloff factor optimization
37
0.4 dB. Fig 4.8(b) shows the influence of the DAC amplitude resolution. In this case,
down to a resolution of 4 bits, the influence of on the penalty is limited. Moreover, in
the ISI-limited region (0 < opt < ), opt is not greatly affected by Mbits . To keep the
penalty below 0.2 dB for PDM-16QAM and Q2 -factor target = 8.5 dB, Mbits should be
equal to or larger than 6. Fig. 4.8(c) shows the influence of Ntaps . The optimal rolloff
varies significantly with Ntaps , and opt needs to be increased to 0.1 when Ntaps = 32. For
PDM-16QAM, Q2target = 8.5 dB and a spectral efficiency equal to 675 %, Ntaps should
be equal or larger than 64 to ensure a penalty at opt below 0.2 dB. For PDM-QPSK,
Ntaps = 25 is sufficient as shown by Ref. [81]. As a conclusion, the length of the digital
filter heavily impacts on opt whereas it is nearly constant in the presence of timing jitter
and/or limited DAC/ADC resolution.
24
22
OSNRQ2target [dB]
20
18
16
14
QPSK 20 %
QPSK 12 %
16QAM 20 %
16QAM 12 %
12
10
8
200
300
400
500
SE (%)
600
700
Figure 4.6: Required OSNR as a function of spectral efficiency for PDM-QPSK and
PDM-16QAM.
4.6
Rolloff factor optimization
In this section, we optimize the rolloff factor for the scenarii defined in Table 4.1 . The
optimization is carried out by finding the optimum that minimizes the OSNR0.1 nm for
a given Q2 -factor target. In Fig. 4.7, the OSNR0.1 nm to achieve a Q2 = 8.5 dB is shown
as a function of for PDM-QPSK and PDM- 16QAM and for various values of spectral
efficiency. In the subsequent paragraphs, the penalty is defined as the excess OSNR0.1 nm
with respect to the ASE noise limit. We note OSNRQtarget the OSNR0.1 nm needed to
achieve Q2 = Q2target . We can identify two sources of performance penalty: (1) intrachannel penalties from ISI, dominant for small , (2) WDM penalties stemming from ICI
at large . Based on the values of Table 4.1, Fig. 4.6 shows that an optimal rolloff opt can
be obtained when a trade-off between the ISI and ICI penalties is achieved. The optimal
38
Figure 4.7: Required OSNR8.5 vs. for (a) PDM-QPSK and (b) PDM-16QAM. Solid line:
Single channel. Marked solid lines: WDM. Dashed line: Single channel w/o imperfections.
Dashed marked line: WDM w/o imperfections
Figure 4.8: Required OSNR8.5 vs. (a) jitter (b) Mbits (c) Ntaps for PDM-16QAM at a
spectral efficiency equal to 675 %. The dashed curve shows the result without taking into
account imperfections
4.7. Choice of FEC and modulation format
39
rolloff is approximately equal to 0 (typically opt < 0.05) when terminal imperfections are
not taken into account. In more realistic conditions, ISI precludes the region where opt
is close to zero. Unfortunately, this region is the region sought for high spectral efficiency.
Thus, another lesson to extract from Fig. 4.7 is that the spectral efficiency can only be
optimized at the cost of performance penalties. The penalty is roughly tripled when
increasing the spectral efficiency from 292/583 % to 338/675 % (for PDM-QPSK/PDM16QAM respectively). When comparing modulation formats, it can be noticed that opt
is the same for PDM-QPSK and PDM-16QAM. PDM-16QAM is more vulnerable to both
ISI and ICI. Hence, choosing a very small rolloff is an acceptable solution for PDM-QPSK
while opt must be carefully chosen for PDM-16QAM.
4.7
Choice of FEC and modulation format
We examine the choice of FEC-overhead and modulation format for our rolloff-optimized
RRC modulation scheme in Fig. 4.6. At the same symbol rate, PDM-16QAM offers twice
the spectral efficiency of PDM-QPSK. However, as we have seen in Fig. 4.7, even at the
optimum rolloff factor, PDM-16QAM is more sensitive with respect to the hardware imperfections. The FEC overhead influences the spectral efficiency through the net channel
rate. In the preceding paragraphs, we considered 12 % FEC overhead. This value, associated with a Q2target = 8.5 dB is typical for the last pre-low-density parity check (LDPC)
generation of FECs [67]. Using a higher overhead might be beneficial since OSNR0.1 nm
requirement is relaxed. We considered a FEC overhead of 20 % which is associated with
Q2 -target = 6 dB. Fig. 4.6 shows the required OSNR0.1 nm as a function of the spectral
efficiency taking into account the FEC overhead for Q2target = 8.5 or 6 dB.
PDM-QPSK (20 %) requires the lowest OSNR0.1 nm when the spectral efficiency is
lower than 350 %. PDM-16QAM (20 %) is the most interesting solution for a spectral
efficiency intermediate between 350 and 630 %. For a spectral efficiency higher than 630
%, a 12 % FEC is interesting for PDM-16QAM as it requires a lower OSNR0.1 nm than
20 % overhead at the same spectral efficiency. We highlight the fact that 12 % overhead
offers an advantage at high spectral efficiency for PDM-16QAM only while the required
OSNR0.1 nm is always lower for 20 % FEC in the PDM-QPSK case.
4.8
Conclusion
In this chapter, we investigate RRC pulse shaping as a means to increase the spectral
efficiency of coherent optical long-haul systems thanks to the reduction of channel spacing.
Most recent works address very small rolloff factors to achieve minimum channel spacing
(N-WDM). In this chapter, we show by simulation that a trade-off exists between spectral
efficiency and transmitter/receiver imperfections. As a consequence, the rolloff factor is
optimized. We show that while zero rolloff might be a good choice for PDM-QPSK
systems, the rolloff must be carefully specified in the case of PDM- 16QAM systems. The
specification can be made regardless the jitter and resolution of the ADC/DAC but should
take into account the finite length of RCC-reshaping digital filter in the transmitter. We
examine the effect of the modulation format at FEC overhead on the signal to noise ratio
40
requirement and identify the optimum choices for different regions of spectral efficiency.
Chapter 5
Decision feedback equalizers for
bandwidth-constrained communications
Contents
5.1
Sub-baudrate filtered channels . . . . . . . . . . . . . . . . . . . . .
41
5.2
Symbol-by-symbol, MAP, and DFE detection . . . . . . . . . . .
42
5.3
Wavelength selective switching in optical mesh networks . . . . .
44
5.4
Equalizer structure . . . . . . . . . . . . . . . . . . . . . . . . . . .
44
5.5
Numerical study of MMSE-DFE for bandwidth-constrained

communication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
46
5.6
5.7
5.8
5.1
5.5.1
Simulation setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5.5.2
Simulation parametrization . . . . . . . . . . . . . . . . . . . . . . . 47
Parametrization of MMSE-DFE . . . . . . . . . . . . . . . . . . . .
49
5.6.1
Performance of MMSE-DFE . . . . . . . . . . . . . . . . . . . . . . . 49
5.6.2
Comparison with MAP
5.6.3
Resilience to phase noise . . . . . . . . . . . . . . . . . . . . . . . . . 53
. . . . . . . . . . . . . . . . . . . . . . . . . 52
MMSE-DFE for ROADM-filtered optical networks: experimental study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
54
5.7.1
Experimental setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.7.2
Experimental results . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.7.3
Performance/complexity trade-off . . . . . . . . . . . . . . . . . . . . 59
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
59
Sub-baudrate filtered channels
In chapter 4, we introduced digital pulse shaping as a means to create optical channels with
a 3-dB bandwidth equal to the symbol rate. Using this method, the channel spacing can
42
#$%"
Chapter 5. Decision feedback equalizers for bandwidth-constrained

communications
#$%"
#$%"
#$%"
#$%"
#$%"
#$%"
#$%"
#$%"
#$%"
#$%"
#$%"
#$%"
#$%"
!"
Figure 5.1: Illustration for single carrier modulation in optical communications using a reduced bandwidth. R = symbol rate. It can be observed that the individual channel bandwidth is smaller than the symbol rate. To mitigate the resulting ISI, decision-feedback
equalization (DFE) is used on each subcarrier.
nearly be reduced to the baudrate. Substantial effort is currently devoted to optimize pulse
shaping for high-spectral efficiency systems with even tighter channel packing [74]. Lately,
a substantial research effort has been put into the development of sub-baudrate-spaced
channels, where the 3-dB bandwidth of the pulse shaping filters is decreased below the
symbol rate to avoid linear crosstalk [82, 83]. In this scenario, the per-channel bandwidth
constraint dramatically increases ISI, so powerful ISI equalization techniques should be
employed [82, 84].
In PDM optical coherent receivers, ISI and ICI are partly equalized by the polarization
demultiplexing DSP block using, for example, the feed-forward (FF) CMA (CMA-FF). In
this chapter, the conventional CMA is denoted CMA-FF to stress the difference between
feed-forward and feedback (FB) equalizers. However, in sub-baudrate filtered channels,
the ISI mitigation offered by the polarization demultiplexing DSP block is not sufficient.
This is expected, since CMA equalizer performance is comparable to linear equalizers [57],
and it is known that linear equalizers perform poorly given strong channel selectivity [29].
In section 5.2, we explain the principle of two nonlinear detection techniques that can be
used to mitigate ISI in bandwidth-constrained communications: maximum a posteriori
detection (MAP) and decision feedback equalization (DFE). In this work, we propose
to use a DFE to mitigate ISI. Section 5.3 explains why high-spectral efficiency optical
networks suffer from heavy in-line filtering which results in additional ISI. In section 5.4
we detail the derivation of the proposed DFE. Section 5.5 presents a numerical study of
the performance of the DFE for bandwidth-constrained communications. In section 5.7,
we assess the performance and complexity of the DFE in an experimental study where
the optical channel bandwidth is reduced because of in-line WSS filtering.
5.2
Symbol-by-symbol, MAP, and DFE detection
In order to mitigate the ISI created by sub-baudrate filtering, some authors have proposed
to use sequence detection [83, 85]. The technique is illustrated in Fig. 5.2(b). In standard
symbol-by-symbol detection, the symbol is identified by minimizing the euclidian distance
with respect to a look-up-table (LUT) of all possible constellation symbols (Fig. 5.2(a)).
In sequence detection, the LUT is extended to small sequences of symbols. The technique
is known as maximum a posteriori (MAP) sequence detection in the optical domain. The
complexity of this detection scheme grows exponentially with the length of the sequence.
5.2. Symbol-by-symbol, MAP, and DFE detection
(a)
(b)
LUT
LUT
43
(c)
LUT
...
MMSE
FF
MMSE
FB
Figure 5.2: (a) Symbol-by-symbol detection, (b): Sequence detection (MAP) (c): MMSEDFE detection
Decision feedback equalizers (DFE) have been proposed for CD compensation [86] as
well as carrier phase estimation [87] in single carrier (SC) coherent optical communications. In [88], the authors propose a DFE to enhance the tolerance of DSP to phase noise.
In this chapter, we propose to use a DFE to mitigate ISI in bandwidth-constrained optical channels. The principle of DFE detection is shown in Fig. 5.2(c). First, the symbol
sequence is filtered using a FF-FIR filter (MMSE-FF). The error signal between the detected symbol and the corresponding constellation symbol is fed back and filtered through
another FIR filter (MMSE-FB). The filters are designed according to the minimum mean
square error (MMSE) criterion. The ISI mitigation offered by MMSE-DFE is superior to
linear equalizers using only a FF-FIR while its complexity remains manageable compared
to MAP.
The amount of comparisons, multiplications and additions required per equalized symbol during the payload transmission are summarized in Table 5.1 for both DFE and MAP
detection for a modulation alphabet size of K symbols.
Table 5.1: Complexity (during payload transmission)
MAP (NM AP )
DFE (L taps)
Real Additions/Subtractions (4NM AP 1)KNM AP 2L + 3K 2
N
Real Multiplications
2N KM
2L + 2K 1
AP
NM AP
Comparisons
K
1
K 1
It can be observed that MAP complexity explodes for high order modulation formats
as well as stringent bandwidth limitations (increase in the requirement for NM AP ) which
44

communications
makes MAP detection very unlikely for future in real-time implementations. On the other
hand, the DFE complexity is linear with its memory depth L.
5.3
Wavelength selective switching in optical mesh networks
In optical long-haul mesh networks, add-drop multiplexers are traffic switching devices at
the wavelength layer that are located as the nodes of the network. The "add" and "drop"
refer to the capability to insert or remove one or several wavelength channels to an existing
WDM signal to pass each signal to its designated network path. In reconfigurable optical
add-drop multiplexers (ROADM), the incoming light is dispersed using diffraction optics.
Mobile switching arrays re-reroute the isolated channels [89].
In the preceding section, we introduced sub-baudrate filtering as a means to reduce
channel spacing. When effectively integrating the reduced channel spacing into the optical network design, we are facing a major challenge at the ROADM traffic switching
nodes [74]. In fact, even the most advanced wavelength selective switches exhibit transfer
functions that can be approached by 3th or 5th order Gaussian filters [90]. As a consequence, when an optical channel passes through cascaded ROADMs on its way through
the network, it will undergo tight filtering or be impaired by ICI [90, 91]. In conclusion, powerful ISI mitigation techniques such as MAP and MMSE-DFE are required in
WDM networks using reduced channel spacing. In the experimental study presented in
this chapter, we apply both MMSE-DFE and MAP detection to experimental waveforms
having cascaded through several wavelength selective switches and compare their performance/complexity trade-off.
5.4
Equalizer structure
In this section, we detail the MMSE-DFE equalizer structure and derive the equalizer FF
and FB filters.
The right side of Fig. 5.3 illustrates the equalizer architecture. The MMSE-DFE
is integrated at the detection step of the receiver DSP in the same manner as MAP
detection. It is standard to use the CMA-FF to demultiplex polarizations and perform
per-channel equalization of residual CD and PMD to some extent. We suppose the signal
is sampled at two samples per symbol at the CMA-FF input and is down-sampled to
one sample per symbol at the CMA-FF output. The MMSE-DFE is implemented on
each polarization separately, which is legitimate since the CMA compensates for the
polarization impairments. The MMSE-DFE makes use of both the received sequence and
the previously detected symbols to remove the interference they cause on the currently
estimated symbol.
The filter taps of the MMSE-FF and MMSE-FB sections are derived based on MMSE
criterion [92, 93]. In our implementation, the calculation of the equalizer filter coefficients
requires the use of a known training sequence.
Referring to Fig. 5.3, the transmitted training sequence on polarization j by sj [],
and their estimation at the output of the MMSE-FF by rj []. The estimation error is
5.4. Equalizer structure
s
1 [n]
RRC
()
DAC
Driver
DAC
Driver
45
IQ
MOD
Laser (-)
s
2 [n]
s1 [n]
RRC
()
RRC
()
DAC
Driver
DAC
Driver
DAC
Driver
DAC
IQ
MOD
LO
IQ
MOD
MIXER
1
Driver
MUX
Laser
ASE
s2 [n]
s+
1 [n]
RRC
()
RRC
()
R
I
DAC
DAC
Driver
IQ
MOD
Driver
DAC
Driver
DAC
Driver
MIXER
2
ADC
ADC
ADC
ADC
e11
MMSE
FF
MMSE
FB
e12
e21
e22
MMSE
FF
Pol Demux:
CMA
CMA-FF
RRC
( )
MMSE
FB
MMSE-DF
IQ
MOD
DAC
Driver
DAC
Driver
r2 [n]
Laser (+)
s+
2 [n]
r1 [n]
IQ
MOD
Figure 5.3: The simulation setup, DAC: digital-to-analog conversion, IQMOD: IQmodulator, ADC: analog-to-digital conversion, CMA: constant modulus algorithm, ASE:
amplified spontaneous emission, LO: local oscillator, MMSE-FF : feedforward
filter of the
MMSE-DFE, MMSE-FB: feedback filter of the MMSE-DFE. j = 1. The symbol rate

B is constant and = 32 Gbaud while the RRC-FIR digital bandwidth is equal to 32
GHz.
46

communications
[i] = sj [i] rj [i], for i = 1, ..., U , where U is the training sequence length. The training
sequence is constructed by l times repeating a training pattern of length N ; therefore
U = lN . We should have N > Neff , where Neff is channel effective memory at twice the
symbol rate. Under this condition the continuous-time equalizer can be derived from the
burst-mode matrix model [94].
Let the error vector calculated over k-th burst (k = 1, ..., l) inside the training sequence
be denoted by k . It is explicitly written as
k = [[(k 1)N + 1], ..., [kN ]]T
(5.1)
where superscript T stands for matrix transpose. The N N error autocorrelation matrix
is denoted by R and calculated as
l
R=
1X H

l j=1 j j
(5.2)
where superscript H stands for Hermitian conjugation [95]. Next we calculate the LDU
decomposition [96] of the positive definite and symmetric R as follows
R = P PH
(5.3)
where is a diagonal matrix and P is an upper triangular matrix. The impulse response
of MMSE-FF, and MMSE-FB parts of the DFE are denoted by hf [] and hb [] respectively,
and are given by the following equations
hf [i] = AN/2,N/2i
(5.4)
hb [i] = hf [i + 1]
(5.5)
.
where A = P 1 , and Am,n is the element of A found on row m and column n. In Eq. (5.5),
i = 0, ..., N/2 1 and in Eq. (5.4), i = 0, ..., N/2 2.
It is important to stress that the impulse responses hf [] and hb [], are derived based
on the training sequence only. During transmission of the payload, the MMSE-DFE is
not updated unlike the CMA-FF. The MMSE-DFE is designed to mitigate ISI stemming
from the bandwidth constraint which is constant in time. In section 5.6.1, we prove that
MMSE-DFE is compatible with the standard DSP for channel tracking in PDM-QPSK
systems. One advantage of a training-based DFE derivation is that the stability of the
equalizer is not impaired at high error rates. As a consequence, very low target BERs can
be considered.
5.5
5.5.1
Numerical study of MMSE-DFE for bandwidthconstrained communication

Simulation setup
In this section, we explain the simulation setup we used to assess the performance of the
MMSE-DFE. It is illustrated in Fig. 5.3. We generated 216 bits using pseudo random
5.5. Numerical study of MMSE-DFE for bandwidth-constrained

communication
47
binary sequence (PRBS). The bits are mapped to QPSK symbols. These sequences are
then upsampled by a factor of two and digitally shaped using a TD RRC FIR filter with
a rolloff factor equal to 0.05. The symbol rate is constant and equal to 32 Gbaud. The
digital bandwidth is defined as the 3-dB bandwidth of the RRC-FIR filter. It is varied
between 27.2 and 32 Gbaud for a constant symbol rate of B = 32 Gbaud. When the
digital bandwidth is equal to 32 Gbaud, no ISI is present provided that the receiver
RRC filter is matched to the transmitter. Three channels are simulated and the BER is
assessed on the central channel. The channel spacing is varied to allow for different values
of spectral efficiency. The signals are digital to analog converted using DACs and passed
to the electrical driver. They are modulated by an ideal IQ-modulator (IQ-MOD) with
a linear transfer function and multiplexed. The bandwidth limitations of the DACs and
the electrical driver are neglected. The amplified spontaneous emission (ASE) is assumed
additive white Gaussian noise. At the receiver, the signals go through the coherent mixer.
The signals are sampled at two samples per symbol by ADCs. The 3-dB bandwidth of
the electrical front-end is set to 32 GHz. The signal are demultiplexed using CMA. To
detect the signal, we performed either MMSE-DFE, MAP, or SBS detection. The BER
is estimated by Monte Carlo simulations, where the noise seed is randomly varied. In
addition, the relative delay and polarization of neighboring channels are randomly varied.
We count at least 500 errors, and convert the bit error rate into the Q2 -factor. In the
following sections, we assume zero carrier frequency offset and carrier phase noise and
polarization alignment on the central channel. In section 5.6.3, we study the resilience of
the MMSE-DFE to laser phase noise. The simulation parameters are further precised in
the subsequent section 5.5.2.
5.5.2
Simulation parametrization
In this section, we specify how we choose the pulse shaping FIR filter length NFIR , CMA
filter length NCMA and digital bandwidth . In Fig. 5.4(a), we investigate the required
pulse shaping length for the bandwidth-constrained systems. The number of taps of the
RRC-FIR pulse shaping filter is varied. We simulate the non-constrained case NFIR =
by applying the shaping filter in the frequency domain on the whole signal vector. In all
other cases, the signal vector is convolved with the truncated impulse response of the pulse
shaper in TD. We conclude that NFIR = 64 is sufficient to achieve a performance similar
to the non-constrained case, whereas an OSNR0.1nm penalty of about 0.5 dB appears when
NFIR is as small as 16. In the rest of the results, NFIR = 64.
In Fig. 5.4(b), we study the impact of the CMA-FF of various lengths NCMA prior to
the MMSE-DFE. System performance at various channel spacings is shown as a function
of NCMA ranging from 0 to 13. The NCMA = 0 represents the case where no CMA-FF is
used. Note that in reality, using CMA-FF is obligatory to retrieve the polarization axes;
however, in the numerical studies presented in this chapter, we are only interested in the
equalization characteristics of CMA-FF and DFE blocks; therefore, in all the simulations
the polarization axes at the transmitter and the receiver are aligned. Comparing system
performance at NCMA = 0 and larger value of NCMA we conclude that the MMSE-DFE
does not benefit from prior CMA-FF. Moreover, increasing NCMA beyond 13 did not influence Q2 so the results are not shown. We conclude that regardless of NCMA , the CMA-FF

communications
48
10
Q2 [dB] @ OSNR(0.1 nm) = 14 dB
10
0.5 dB
9
8
7
9.5
Ntaps =
Ntaps = 64
Ntaps = 32
f = 30 GHz
8.5
f = 28 GHz
Ntaps = 16
8
0 1
28
29
30
31
32
Digital bandwidth [GHz]
(a) RRC-FIR taps. The CMA-FF is not used.
f = 50 GHz.
5
9
CMAFF taps
13
(b) CMA-FF (NCMA ) memory. NFIR = 64.

= f .
10
10
Q2 [dB] @ OSNR(0.1 nm) = 14 dB
f = 50 GHz
8
f = 50 GHz
7
f = 30 GHz
f = 28 GHz
28
29
30
31
32
(c) Different channel spacings. The CMA-FF is

not used. NFIR = 64.
= 32 GHz
= 31.2 GHz
= 30.4 GHz
= 29.6 GHz
7
= 28.8 GHz
= 28 GHz
300 500 1000 2000 4000 8000
Length of the training sequence (M)
8
(d) Length of the training sequence. The CMAFF equalizer is not used. NFIR = 64. The
MMSE-DFE memory depth L = 10.
Figure 5.4: Preliminary results. Q2 @ OSNR0.1nm = 14 dB as a function of different

simulation parameters.
5.6. Parametrization of MMSE-DFE
49
does not have significant influence on the performance of the bandwidth-constrained communication system with small rolloff. As a consequence, the generated results regarding
the MMSE-DFE without CMA-FF can be directly generalized to transmission schemes
using CMA-FF. In particular, links with time-variant polarization impairments can be
studied, because the latter are taken care of by the CMA-FF.
In Fig. 5.4(c), we assess the influence of the digital bandwidth. When the channel
spacing decreases, the system is limited by ICI besides the ISI. A trade-off exists between
the two sources of interference. With decreasing , ICI limitations become less stringent
as the channel bandwidth decreases, but ISI is created. In Fig. 5.4(c) the optimum can
be observed. For each channel spacing where ISI is preset, the optimum performance is
achieved when f . In Fig. 5.4(d), we study the influence of the overall length of the
training sequence U on system performance. Since the MMSE-DFE FF and FB filters
are derived only once at the beginning of the transmission, they do not generate a large
overhead. Nonetheless, it is interesting to evaluate the minimum length of the training
sequence to assure a satisfactory estimation of the autocorrelation matrix. Fig. 5.4(d)
shows that the performance saturates at U = 1000 symbols.
5.6
Parametrization of MMSE-DFE
In the following results, we assess the performance of the MMSE-DFE with respect to its
memory depth L. In this section, only the performance of the MMSE-DFE is assessed. In
section 5.6.1, the MMSE-DFE is combined with linear equalizers. In Fig. 5.5, we study the
performance of the DFE when the channel spacing f = 50 GHz. No ICI is present since
the channel spacing is much larger than the channel bandwidth. Both the Q2 for a fixed
OSNR0.1nm of 14 dB and the OSNR0.1nm at a fixed target of Q2 = 6 dB are presented.
The results are shown as a function of the number of MMSE-FB taps L. When L = 0,
there is no feedback (MMSE-FB) and the feedforward section (MMSE-FF) contains only
one tap equal to 1. When L increases, the FB loop takes into account more and more
past decisions. We therefore assess both the performance gain stemming from the use
of the DFE and the number of taps required for its use. When the digital bandwidth is
decreased, Q2 is degraded. It can be seen that the MMSE-DFE offers a performance gain
when the digital bandwidth is smaller than 30 GHz. When the digital bandwidth is equal
to 28 GHz, the OSNR0.1nm to achieve Q2 = 6 dB raises by 3.5 dB and thanks to the use
of the MMSE-DFE, the penalty is reduced to 1.5 dB. However, there is still residual ISI.
The performance is not further enhanced when L is increased beyond 4 taps.
In Fig. 5.6, we assess the MMSE-DFE gain in the presence of ICI. The OSNR0.1nm at
2
Q = 6 dB is shown as a function of the number of MMSE-DFE taps, for f = 30 GHz
in Fig. 5.6(a) and 28 GHz in Fig. 5.6(b), respectively, for different digital bandwidths.
Comparing Figs. 5.4(a) and 5.6, we can conclude that the optimum bandwidth is equal
to the channel spacing regardless of the length of the MMSE-DFE.
5.6.1
Performance of MMSE-DFE
According to the results of the preceding sections, the simulation parameters are listed
and the notations defined in Table 5.2.

communications
50
14
= 32 GHz
= 31.2 GHz
= 30.4 GHz
= 29.6 GHz
= 28.8 GHz 2 dB
= 28 GHz
OSNR(0.1 nm) @ Q2 = 6 dB
Q2 [dB] @ OSNR(0.1 nm) = 14 dB
10
13
12
32 GHz
31.2 GHz
7
30.4 GHz
1.5 dB
11
29.6 GHz
28.8 GHz
28 GHz
0
4
6
8
Memory of the DFE (L)
10
0
10
2
4
6
8
(a) Q2 @ OSNR = 14 dB
10
(b) OSNR @ Q2 = 6 dB
Figure 5.5: System performance as a function of the MMSE-DFE memory depth. f = 50

GHz. The CMA-FF equalizer is not used. The number of RRC-FIR taps of the pulse
shaping filter NFIR = 64.
15
OSNR(0.1 nm) @ Q = 6 dB
15
14
13
12
32 GHz
31.2 GHz
30.4 GHz
29.6 GHz
28.8 GHz
28 GHz
14
13
12
11
10
11
32 GHz
31.2 GHz
10
30.4 GHz
29.6 GHz
28.8 GHz
9
0
4
6
8
(a) f = 30 GHz
10
9
0
28 GHz
4
6
8
10
(b) f = 28 GHz
Figure 5.6: System performance as a function of the MMSE-DFE memory depth. The
CMA-FF equalizer is not used. The number of RRC-FIR taps of the pulse shaping filter
NFIR = 64.
In Fig. 5.7, we show the performance of the combined CMA-FF (NCMA = 13 taps)
and MMSE-DFE for an optimized digital bandwidth. The DFE performance gain can be
observed, keeping in mind that L = 0 corresponds to CMA-FF-only equalization. When
f = 30 GHz, the OSNR @ Q2 = 6 dB is 0.8 dB lower when using a MMSE-DFE with a
sufficient number of taps, whereas the OSNR gain is as large as 2 dB for f = 28 GHz.
51
Table 5.2: Simulation parameters

Parameter
Symbol
Length of estimation bursts (bits)
N
Length of training sequence (bits)
U
Channel spacing [GHz]
f
DFE memory depth

L
CMA taps
NCMA
Number of taps of the pulse shaping RRC-FIR NFIR
Symbol rate [Gbaud]
B
Value range
50
214
28 - 50
28-32
0-10
13
64
32
In this case, the OSNR0.1nm penalty with respect to the f = 50 GHz is as low as 1.5
dB, as already noted in Fig. 5.6.
14
f = 50 GHz
f = 30 GHz
f = 28 GHz
OSNR(0.1 nm) @ Q2 = 6 dB
Q2 [dB] @ OSNR(0.1 nm) = 14 dB
10
13
12
f = 50 GHz
f = 30 GHz
f = 28 GHz
11
6
0
4
6
8
(a) Q2 @ OSNR = 14 dB
10
10
0
4
6
8
10
Figure 5.7: System performance as a function of the MMSE-DFE memory depth for three
different channel spacings. The digital bandwidth is equal to the channel spacing. The
number of taps of the CMA-FF NCMA = 13. The number of RRC-FIR taps of the pulse
shaping filter NFIR = 64.
In Fig. 5.8, we extend our study to another linear equalizer, a cross-polarization linear
equalizer based on the MMSE-criterion [29] implemented at one sample per symbol. Like
the DFE, it is data-aided and based on channel estimation. Thanks to the training
symbols, a two-input two-output channel matrix is derived, taking into account ISI as
well as polarization effects. Using this matrix, the MMSE burst-mode equalizer is derived
and extended to a continuous-time equalizer using the method presented in [94].
Different combinations of the three equalizers, i.e CMA-FF, linear MMSE and MMSEDFE are studied. We note that the combination of CMA-FF and MMSE fails to mitigate
the ISI stemming from tight filtering. CMA-FF and MMSE have a very similar performance for all channel spacings. Finally, as already noted in Fig. 5.4 for the special case

communications
52
14
CMA
CMA+DFE
MMSE
MMSE+DFE
9.5
OSNR(0.1 nm) @ Q2 = 6 dB
Q2 [dB] @ OSNR
(0.1 nm)
= 14 dB
10
13
8.5
7.5
6.5
5.5
28
12
CMA
CMA+DFE
MMSE
MMSE+DFE
30
32
Channel spacing f [GHz]
11
34
(a) Q2 @ OSNR = 14 dB
10
28
30
32
Channel spacing f [GHz]
34
Figure 5.8: System performance as a function of the channel spacing for different equalizers. The number of FIR taps of the pulse shaping filter NFIR = 64. When used, the
number of taps of the CMA-FF is NCMA = 13. The length of the MMSE equalizer was
equal to 50 taps. When used, the memory depth of the DFE is equal to L = 10 taps. The
digital bandwidth is equal to the channel spacing.
of CMA-FF, the DFE does not benefit from prior linear equalization.
5.6.2
Comparison with MAP
In Figs. 5.9 and 5.10, we compare the performance of the MMSE-DFE with a direct
implementation of the MAP-detector as in Ref. [85]. Similarly to the MMSE-DFE, the
MAP detector operates separately on each polarization tributary. We present results for a
MAP with NMAP = 3 taps and NMAP = 5 taps. The complexity of the equalizer becomes
prohibitive beyond this value. In our simulations, we focused on PDM-QPSK. In a more
general context, due to the number of entries of the MAP look-up-table, the complexity of
the MAP detector is of O(C NMAP ), where C is the alphabet size. This makes it extremely
difficult to implement MAP detection for higher order modulations [97], whereas the
complexity of the MMSE-DFE is largely determined by the cost of LDU decomposition of
the error autocorrelation matrix. This complexity is of O(N 3 ) [96] and does not depend
on the alphabet size. In a MAP equalizer with NMAP = 3, when detecting a given
symbol, the interference from the preceding symbol and the following symbol are taken
into account. On the contrary, for MMSE-DFE with a memory depth L = 1, only the
preceding decision is fed into the feedback loop. It is however interesting to compare these
equalizers at a given memory depth, i.e.: MAPN =3 / MMSE DFEL=1 and MAPN =5 /
MMSE DFEL=2 . For MAPN =5 , the number of entries in the MAP look-up-table is
equal to C NMAP = 1024. For each symbol, every entry is compared with a received
sequence of length NMAP = 5, leading to a number of complex multiplications, additions
and comparisons roughly (256 5) times higher than symbol-by-symbol detection. For
OSNR(0.1 nm) @ Q2 = 6 dB
13
10
Q2 [dB] @ OSNR(0.1 nm) = 14 dB
53
12
MMSEDFE, L=1
MMSEDFE, L=2
MMSEDFE, L=3
MMSEDFE, L=4
MAP, NMAP=3
MAP, N
MAP, NMAP=5
11
=5
MAP
6
28
MMSEDFE, L=1
MMSEDFE, L=2
MMSEDFE, L=3
MMSEDFE, L=4
MAP, NMAP=3
29
30
31
32
(a) Q2 @ OSNR0.1nm = 14 dB
10
28
29
30
31
32
(b) OSNR0.1nm @ Q2 = 6 dB
Figure 5.9: System performance using MMSE-DFE or MAP ISI mitigation for different
memory depths of the equalizers. The channel spacing is equal to 50 GHz. The CMA-FF
equalizer is not used. The number of RRC-FIR taps of the pulse shaping filter NFIR = 64.
MMSE DFEL=4 which shows similar performance, the equalization merely consists of
two convolutions by FIR filters of very short lengths L + 1 and L respectively, subtraction
and symbol-by-symbol decision. We only add 20 complex multiplications and additions
and one subtraction per equalized symbol to the symbol-by-symbol detection complexity.
Fig. 5.9 shows that the MAP equalizer generally performs better than MMSE-DFE at the
same memory depth, the performance difference increasing with the bandwidth constraint.
However with only one increment in memory depth, the MMSE-DFE performs similarly
as the MAP equalizer for both NMAP = 3 and 5. This gives a clear advantage to the
MMSE-DFE, because of its reduced complexity with respect to the MAP equalizer, at
the cost of the presence of a feedback.
The performance of MMSE-DFE at low BER is illustrated in Fig. 5.10, where Q2 is
given as a function of OSNR0.1nm for MMSE DFEL=4 and MAPN =5 . Q2 as low as 4
dB have been considered. It is to be noted that the MMSE-DFE performance does not
degrade with respect to MAP for very low OSNR. We study the feedback error propagation
of MMSE-DFE by adding the MMSE-DFE Q2 when the feedback signal is forced to be
the correct symbols. As already pointed out in section 5.4, thanks to training-based
derivation, the MMSE-DFE does not suffer from catastrophic error propagation at low
BER. The performance loss from feedback error propagation does not exceed 0.6 dB even
at a digital bandwidth of 28 GHz and OSNR0.1nm as low as 8 dB.
5.6.3
Resilience to phase noise
In this section we study the impact of laser phase noise on the performance of the modulation scheme using the MMSE-DFE. We suppose an ideal local oscillator, thus laser phase

communications
54
10
9
Q [dB]
8
7
DFE, L=4
DFE, L=4
DFE, L=4
DFE, L=4, correct FB
DFE, L=4, correct FB
L=4, correct FB
MAP, N
=5
=32 GHz
MAP
MAP, NMAP = 5
6
MAP, NMAP = 5
= 28 GHz
= 30 GHz
5
4
8
10
11
OSNR(0.1 nm)
12
13
14
Figure 5.10: System performance using MMSE-DFE or MAP ISI mitigation for different
OSNR. The channel spacing is equal to 50 GHz. The CMA-FF equalizer is not used. The
number of RRC-FIR taps of the pulse shaping filter NFIR = 64.
noise is simulated for the transmitter laser only. Frequency dithering is ignored. The
phase of the transmission laser is simulated as a random walk whose step size is a zero
mean Gaussian random variable, yielding the approximation of a Lorentzian laser line
shape [98]. The laser phase noise is mitigated using the well-known Viterbi and Viterbi
algorithm [62] taking profit from the symmetry of the QPSK constellation. The slidingwindow averaging length of the Viterbi and Viterbi algorithm is equal to 64 symbols
through all simulations. Because of the /2-ambiguity of the V+V detection, differentially encoded and decoded QPSK is implemented in this simulation. The Viterbi and
Viterbi algorithm is applied after the CMA-FF and prior to the MMSE-DFE. Therefore,
we investigate both the robustness of the V+V to bandwidth constraints and the robustness of the MMSE-DFE to residual phase noise and cycle slips. In Fig. 5.11, the channel
spacing is fixed to 30 GHz. From Fig. 5.11, we conclude that the Viterbi and Viterbi
algorithm is usable in the context of bandwidth-constrained PDM-QPSK transmission
schemes for laser bandwidths up to 1 MHz. When the laser linewidth is equal to 500
kHz, the OSNR penalty @ Q2 = 6 dB is about 0.25 dB. This penalty does not depend on
the number of DFE taps used. As a conclusion, the stability of the MMSE-DFE is not
impaired by residual phase noise.
5.7
MMSE-DFE for ROADM-filtered optical networks:

experimental study
In this section, we address the issue of performance degradation due to optical filtering in
multiple ROADM in coherent optical networks. In Ref. [97] the optical filtering-induced
performance penalty on WDM PDM-8QAM channels was studied in combination with
5.7. MMSE-DFE for ROADM-filtered optical networks: experimental study

55
15
@ Q2 = 6 dB
OSNR
= 0 MHz, no V+V
= 0 MHz
= 0.1 MHz
= 0.5 MHz
= 1 MHz
14
13
(0.1 nm)
= 0 MHz, no V+V
= 0 MHz
= 0.1 MHz
= 0.5 MHz
= 1 MHz
Q [dB] @ OSNR(0.1 nm) = 15 dB
10
12
6
0
2 3 4 5 6 7 8
(a) Q2 @ OSNR = 15 dB
9 10
11
0
3 4 5 6 7 8
10
Figure 5.11: System performance as a function of the number of MMSE-DFE taps. The
channel spacing is equal to 30 GHz. The digital bandwidth is equal to the channel spacing.
The CMA-FF is not used. The 3-dB transmission laser linewidth is varied.
MAP detection. However, the computational cost of MAP detection makes it prohibitive
for real-time implementation. We use the MMSE-DFE that was developed and validated
in simulation in the preceding sections to mitigate the ISI stemming from optical filtering
of PDM-8QAM channels at acceptable computational cost. We first explain the experimental setup in section 5.7.1. Similarly to the numerical study of the preceding section,
we compare MMSE-DFE with MAP detection performance in section 5.7.2. We conclude
the experimental study by assessing the performance/computational complexity trade-off
of MMSE-DFE in section 5.7.3.
5.7.1
Experimental setup
Figure 5.12 illustrates the experimental setup of the present work. The WDM transmitter uses an array of 41 lasers. Programmable DACs provide 28 Gbaud 8QAM signals at
two samples per symbol modulating the central four external cavity lasers (ECL) spaced
at 37.5 GHz via IQ-modulators. As a result, four test channels with PDM-8QAM at
150Gbit/s at a net spectral efficiency of 4bit/s/Hz are obtained. The remaining 37 distributed feedback (DFB) lasers are modulated with QPSK on the standard ITU frequency
grid. The signals from each modulator are subsequently spliced, one arm delayed by 10
ns and recombined so as to emulate PDM. Three binary De Bruijn sequences of length 213
that were circularly shifted by 0, 1365, and 2730 bits to obtain the 8QAM sequences. We
used linear and non-linear pre-compensation to mitigate the transmitter imperfections, as
well as RRC (rolloff factor 0.1) pulse shaping implemented at two samples per symbols.
The signal spectrum is shown in Fig. 5.14. The transmission line was composed of a circulation loop with four 100 km spans of SSMF. The fiber loss is compensated by hybrid

communications
56
QPSK
even
1
PS
IQ-MOD PME
Tuneable
filter
IQ-MOD PME
DAC
2
Coherent ADC DSP

RX
PDM 8-QAM WSS
IQ-MOD PME
WSS
QPSK
LO
WSS
DAC
odd
ECL (37.5 GHz spacing
37 DFB laser sources, 50 GHz spacing
DAC
LSPS
IQ-MOD PME
EDFA
PS
N
SSMF
100 km
x4
Raman pump
Figure 5.12: Experimental setup. ECL: external cavity laser, DFB: distributed feedback,
IQ-MOD: IQ-modulator, DAC: digital-to-analog converter, PME: polarization division
multiplexing emulator, PS: polarization scrambler, WSS: wavelength selective switch,
SSMF: standard single mode fiber, EDFA: erbium doped fiber amplifier, LO: local oscillator, ADC: analog-to-digital converter DSP: digital signal processing.
5.7. MMSE-DFE for ROADM-filtered optical networks: experimental study

57
Raman-Erbium optical repeaters. The Raman amplifier had two back-propagating laser
diode pumps at wavelengths 1432 nm and 1457 nm, and offered 10 dB on-off gain. Two
wavelength selective switches are incorporated in the loop to emulate the filtering and
crosstalk effects of ROADMs. The 3 dB bandwidth of one wavelength selective switch
stage is 34.5 GHz, and after 10 filtering stages (five loops), it is reduced to 23.3 GHz. One
wavelength selective switch separates odd and even wavelengths while the second combined the wavelengths. At the receiver side, the signal is sampled and recorded by a 40
GSamples/s real-time scope with 20 GHz bandwidth. We use standard DSP for coherent
receivers which included front-end equalization, re-sampling, CD compensation, polarization demultiplexing, carrier and frequency offset estimation as introduced in chapter 3.
We measure waveforms with and without wavelength selective switches. To evaluate the
performance, we measure the bit-error ratio, and translate it to Q2 .The total launched
power is optimized using waveforms after 5 loops.
5.7.2
Experimental results
11
10
Q2 [dB]
9
8
7
6
5
2 loops
4
1
3 loops
4 loops
5 loops
3
4
# DFE taps (L)
6 loops
Figure 5.13: Q2 vs. number of taps of the DFE for different numbers of filtering stages.
The Q2 is measured as a function of distance for with and without ROADM emulation.
The number of recirculation loops varies from two to eight for each configuration, with
two effective filter functions per loop in the ROADM configuration. To combat ISI, we
implement the per polarization MMSE-DFE detailed in section 5.4. First, we study the
memory requirement of the MMSE-DFE. In Fig. 5.13, the performance with ROADM
and different numbers of loops is shown as a function of the MMSE-DFE number of taps
L. It can be seen that the performance saturates around L = 3 or 4. For the following
results, L is fixed to 5. We then compare the performance of MMSE-DFE with MAP and
symbol-by-symbol detection. The results are shown in Fig. 5.15. With symbol-by-symbol
detection, the ROADM configuration suffers from accumulating performance degradation
due to ISI coming from in-loop optical filtering as distance grows. The Q2 -factor penalty
is 1 dB after four filtering stages (two loops), and 2 dB after ten filtering stages.

communications
58
(b)
Power: 5 dB/div
(a)
Frequency: 25 GHz/div
Figure 5.14: (a) Spectra of the four 8-QAM channels (b) Signal bandwidth after 10
filtering stages. Figures are taken from Ref. [97].
12
2 dB
8
0.7 dB
Q [dB]
10
2
2
wROADM 5 tap DFE

wROADM 3 tap MAP
wROADM 4 tap MAP
wROADM SBS
woROADM SBS
4
5
6
# Filtering stages
Figure 5.15: Q2 vs. number of filtering stages for MAP, MMSE-DFE and symbol-bysymbol detection
It has to be noted that the four-tap MAP LUT is only 95% filled due to its large size
and a limited acquisition time for each study. The three-tap MAP LUT is completely
filled. Assuming N is longer than the effective channel length, MAP achieves the optimum maximum likelihood detection performance [85], which is not the case for MMSEDFE. Hence, given a sufficiently large N, MAP is always superior to MMSE-DFE for the
same memory depth. This is confirmed in Fig. 5.15. We observe that the MMSE-DFE
performance is intermediate between 3-tap and 4-tap MAP. Compared to pure symbolby-symbol detection, the Q2 - factor after 5 loops is improved by 0.5 dB with the 3-tap
5.8. Conclusion
59
MAP, 0.7 dB with MMSE-DFE (L = 5) and 1 dB with the 4-tap MAP. About one third
of the considerable ISI penalty can be recovered employing a five-tap MMSE-DFE while
half of the penalty is recovered using 4-tap MAP detection.
5.7.3
Performance/complexity trade-off
We can now compare the complexity (section 5.2) and the experimental results from the
previous section to assess the performance/complexity trade-off of MMSE-DFE and MAP.
It is illustrated in Table 5.3. In can be seen that MMSE-DFE is preferable to three-tap
MAP since it offers a better performance for a computational complexity reduction of
two orders of magnitude. Four-tap MAP offers a noticeable performance advantage over
MMSE-DFE (approximatively 0.5 dB at five loops) but only at the cost of a huge increase
in computational complexity (Table 5.3).
Table 5.3: Complexity (during payload transmission) and performance comparison
Q2 [dB] vs. SBS Add./Sub.(R) Mult.(R) Compar.
@ 5 loops
MMSE-DFE 5 taps 0.7
32
15
3
MAP 3 taps
0.5
5632
3072
511
MAP 4 taps
1
61440
32768
4095
5.8
Conclusion
In this chapter, we use MMSE-DFE to tackle bandwidth constraints in coherent optical

communication systems aiming at high spectral efficiency. In order to reduce the channel
spacing, digital pulse shaping is used at the transmitter to lower the signal bandwidth
below the symbol rate. Moreover, when high-spectral efficiency signals are transmitted
over optical networks, they are heavily filtered by repetitive ROADM passes. As a consequence, powerful DSP solutions are required to reduce ISI at the receiver. MMSE-DFE
is an option that efficiently mitigates ISI at a reasonable computational cost.
We study the impact of the digital channel bandwidth on system performance by simulation and show that the channel spacing should be approximately equal to the digital
bandwidth. The use of a MMSE-DFE greatly enhances the performance when the channel spacing is smaller than the symbol rate. For a channel symbol rate of 32 Gbaud,
the channel spacing could be decreased to 28 GHz thanks to the MMSE-DFE with a
OSNR0.1nm penalty (at Q2 = 6 dB) as low as 1.5 dB with respect to the non-bandwidthconstrained case. In this case, the MMSE-DFE performance gain was as high as 2 dB
in OSNR0.1nm . Thanks to the polarization tracking ability of the blind CMA-FF, our results can be generalized to links with time-variant polarization effects. When the memory
length is carefully chosen, the four-tap MMSE-DFE performs similarly to a five-tap MAP
detector, while offering a reduced computational complexity. MMSE-DFE is also robust
to residual phase noise stemming from imperfect transmission lasers. We experimentally
study the performance of the equalizer in presence of multiple cascaded ROADMs in a 28
60

communications
Gbaud PDM-8QAM system transmitting over a dispersion-unmanaged EDFA+Raman

terrestrial link over 37.5 GHz grid. Thanks to the experimental study, we confirm that
MMSE-DFE achieves better ISI mitigation than three-tap MAP detection while reducing
computational complexity during payload by two orders of magnitude. One third of the
considerable ISI penalty can be recovered employing a five-tap MMSE-DFE. It can be
concluded that the use of MAP detection is the best option only when computational
power is not limited and therefore four-tap (or more) MAP can be implemented, while
MMSE-DFE is always preferable to 3-tap MAP both on complexity and performance
aspects.
Chapter 6
Multicarrier Offset-QAM modulations
Contents
6.1
Introduction
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
61
6.2
MC-OQAM Principle . . . . . . . . . . . . . . . . . . . . . . . . . .
62
6.3
Receiver digital signal processing for MC-OQAM . . . . . . . . .
64
6.4
6.1
6.3.1
Time/frequency acquisition . . . . . . . . . . . . . . . . . . . . . . . 65
6.3.2
Channel estimation and equalization . . . . . . . . . . . . . . . . . . 66
6.3.3
Phase tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
6.3.4
Experimental results . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
73
Introduction
Nowadays, high-speed DACs become more and more available to use in coherent optical
communication systems [69]. It is then possible to digitally divide the available electrical bandwidth into subchannels on each optical carrier [51]. Since modulation of each
subchannel is independent, multicarrier communications attract a growing research interest for applications including multi-user access and code parallelization [51, 99]. Optical
OFDM has attracted much research interest [100103]. In OFDM spectra, each subchannel is sinc-shaped. Bandwidth-limited subcarriers are preferable in order to limit the
subcarrier interference and the performance penalties arising from receiver bandwidth
limitations.
Bandwidth-limited multicarrier transmission can be achieved using conventional multicarrier QAM (MC-QAM) on RRC shaped subcarriers. Subcarrier spacing equal to
the symbol rate results in crosstalk for any rolloff factor greater than zero. Multicarrier offset-quadrature amplitude modulation (MC-OQAM) [104] has recently been proposed as an alternative that achieves orthogonal subcarrier spacing equal to the symbol
rate [99,105,106]. MC-OQAM uses RRC pulse shaping, but the spectra of adjacent channels overlap without resulting in crosstalk between neighboring subcarriers thanks to a
62
Chapter 6. Multicarrier Offset-QAM modulations
!"
Frequency
Frequency
Figure 6.1: Illustration for electrical multicarrier modulation in optical communications.

Conventional single carrier modulation using digital signal processing. R = symbol rate.
Below: Multicarrier modulation using five subcarriers
half-symbol time delay between the in-phase and quadrature components of the signal on
each subcarrier. Hence, non-zero rolloff factors become theoretically accessible without
performance penalty.
Similar to orthogonal frequency-division multiplexing (OFDM), transmission impairments are bound to break the subcarrier orthogonality of MC-OQAM [105]. The benefit
of MC-OQAM therefore strongly depends on the ability to recover subcarrier orthogonality using DSP [107]. Ref. [106] has demonstrated a 224 Gb/s MC-OQAM communication
system with seven 16-Offset-QAM subcarriers in a back-to-back configuration using a
modified blind decision-directed equalization on each subcarrier. Recently, Ref. [108]
has demonstrated the benefits of MC-OQAM for superchannel crosstalk suppression in a
back-to-back experiment.
In this chapter, we demonstrate the potential of MC-OQAM in a coherent long-haul
optical communication experiment. We build a new DSP architecture to tackle the linear
impairments. We compare MC-OQAM with MC-QAM to assess the benefits of crosstalk
mitigation via OQAM. We carry out experiments to test the transmission performance
of MC-OQAM in dispersion-unmanaged long-haul transmission links. The channels are
modulated at an aggregate symbol rate of 32.5 GBaud, to make 100 Gb/s net bit-rate
using multicarrier QPSK (MC-QPSK) or offset-QPSK (MC-OQPSK) on each subcarrier
assuming approximately 32 percent protocol and forward error correction overhead.
In section 6.2, MC-OQAM modulation is mathematically defined. The receiver DSP
is detailed in section 6.3. The experimental results are presented in section 6.3.4.
6.2
MC-OQAM Principle
Figure 6.2 illustrates the mathematical model for the MC-OQAM system. Referring to
Fig. 6.5, the transmitter modulation is implemented prior to DAC loading. The channel
impulse response hij [n] includes all linear impairments of the experimental setup (see
section 6.3.4): the DAC transfer function, the wavelength selective switch induced filtering
of the optical channel, the linear in-line impairments such as CD and PMD, the tunable
receiver filter response and the receiver front end transfer function, polarization separation
and CD compensation since this block is not separated between the subcarriers in our
6.2. MC-OQAM Principle

sR
i,1 [n]
sIi,1 [n]
sR
i,m [n]
sIi,m [n]
sR
i,M [n]
sIi,M [n]
63
ej1 2
2
.
.
.
M/2
z
u[n]
j=
2
.
.
.
M/2 u[n]
z1
1n
jm
2

mn
M
hij [n]
jm
2
ej2
z1
mn
M
ejM 2
ej2
u [n] M/2
.
.
.
ejM 2
M/2 u[n]
.
.
.
R
u [n] M/2
ej2 M
j2 1n
M
ej2
ej1 2
mn
M
Polarization i
ej2
mn
M
u [n] M/2
z1
R
rj,1
[n]
I
rj,1
[n]
R
rj,m
[n]
I
rj,m
[n]
R
rj,M
[n]
I
rj,M
[n]
Polarization j
Figure 6.2: MC-OQAM modulation/demodulation

and channel model. In all exponentials
and complex multiplications, j = 1

DSP setup. The demodulation part of Fig. 6.2 is implemented in the receiver DSP.
The transmitter implements a synthesis filterbank composed of M subchannels, each
modulated with QAM information symbols. For the ease of notation, we separate the
I
complex QAM symbols into real and imaginary parts sR
i,m [n] and si,m [n]. Index i = 1, 2
refers to the two polarizations at the transmitter, index m = 1...M refers to the M
subchannels, while n is the symbol index. The OQAM modulation is implemented by
upsampling the symbols by two and delaying the imaginary part of each subchannel
by half a symbol. The signals are upsampled by M/2 and convolved with the impulse
response of the pulse shaping filter u[n]. Without loss of generality, we assume that the
pulse shaping filter u[n] is an RRC filter. Its expression can be derived from the continous
RRC g(t) given in chapter 4 Eq 4.1 . Since g(t) has an infinite support, it is truncated in
the digital model to L samples.
r

L1 T
T
u[n] =
g (n
)
(6.1)
M
2
M
In this equation, T is the symbol period of one MC symbol.
Its expression can be found in [109]. In order to prevent inter-subchannel interfer
ence, a factor ejm 2 multiplies the signal on the two neighboring subchannels. The M
subchannels are added together after multiplication by complex exponentials to shift each
subcarrier frequency [109]. The signal is transmitted through the frequency selective
channel hij [n]. Index j refers to the receiver polarizations. At the receiver, the analysis
filterbank implements the counter-part of each operation performed at the transmitter.
64
After frequency shifting to process each subchannel around the origin, the received signal
is convolved with the filters matched to the pulse shaping filters and downsampled by
M/2. Real and imaginary parts of the signal are separated. The half symbol delay applied on the imaginary branch is compensated so that both the real and imaginary signals
are received synchronously. The signal is finally downsampled to one sample per symbol.
In MC-OQAM, only neighboring channels interfere because of the spectral shape of
I
R
[n] can
[n] and rj,m
the RRC filters. Thus the real output sequences on subcarrier m, rj,m
R
I
be expressed as a function of the real input sequences si,m0 [n] and si,m0 [n] coming from
both polarizations of the subchannel m and its adjacent subcarriers m 1 and m + 1. The
channel model can be found in [105]. When the channel response hij [n] is approximately
frequency flat, the subcarriers are orthogonal thus do not interfere [110]. The derivation
of the the subchannel orthogonality condition can be found in Ref. [110].
6.3
Receiver digital signal processing for MC-OQAM
In a realistic optical long-haul communication setup, the transmitted signal is altered by

several impairments. For the widespread single carrier QAM modulations, receiver DSP
can be subdivided into three main steps (for more details, see chapter 3).
The FD overlap-and-save-algorithm [111] is used as a first step for CD compensation.
It is not specific to the modulation format.
Polarization demultiplexing is usually carried out afterwards by the adaptive butterfly CMA [55]. The error signal of the CMA is sensitive to the amplitude of the
waveforms only. As a consequence, phase tracking can be carried out independently
in the next step.
For QAM modulations, phase noise and carrier frequency offset result in a rotation
of the constellation diagram in the complex plane. This pattern is exploited in phase
correction algorithms such as the Viterbi and Viterbi phase estimator [62].
Except for the compensation of CD, substantial changes need to be made to the standard algorithms when MC-OQAM is used. First, in OQAM, the symbol peaks on in-phase
and quadrature symbols are delayed by half a symbol time. As a consequence, the modulation does not present constant or even discrete moduli (as in high-order modulations)
at any time, which makes it impossible to use the CMA to demultiplex the polarizations.
Second, phase tracking algorithms that exploit the constellation rotation in presence of
phase noise cannot be used for OQAM. As can be seen in Fig. 6.2, a phase shift will
result in I-Q interference because of the real/imaginary selection in OQAM systems. To
conclude, OQAM polarization demultiplexing and phase tracking are not to be carried
out with the same algorithms as in conventional QAM and independent algorithms for
polarization and phase tracking cannot be used.
As explained in section 6.1, one main motivation for electrical multicarrier modulation
is DSP parallelization which results in a reduction of overall computational complexity.
When designing algorithms to mitigate impairments in MC-OQAM systems, it is therefore
6.3. Receiver digital signal processing for MC-OQAM

sR
1,m [n]
sI1,m [n]
sR
2,m [n]
sI2,m [n]
65
z 1
M/2
u[n]
j2
ejm 2 e
mn
M
h11 [n]
h12 [n]
ej2
mn
M
u [n] M/2
z1
ejm 2
h21 [n]
R
r1,m
[n]
M/2
u[n]
ejm 2 e
j2 mn
M
I
R
h22 [n]
e
j2 mn
M
jm
2
u [n]
M/2
z1
I
r1,m
[n]
MMSE
Equalizer
R
r2,m
[n]
R
r2,m
[n]
Figure 6.3: MC-OQAM subcarrier

m: dual-polarization transmitter, channel, phasecorrected MMSE equalizer. j = 1
desirable to separate the DSP as much as possible among subcarriers of the communication channel. However, per-subcarrier treatment makes the transmission scheme more
vulnerable to variable phase offsets, a fact that is well-known in the wireless OFDM communications [112,113]. Thus MC-OQAM is likely to suffer increasingly from phase related
impairments when the number of subcarriers is increased at a constant aggregate symbol
rate.
In the following paragraphs, we detail the algorithms used for receiver DSP of MCOQAM.
6.3.1
Time/frequency acquisition
At the the receiver side, the signal is first sampled at twice the aggregate symbol rate.
In case the signal is transmitted over a significant length of optical fiber, digital compensation of CD is first carried out using the well-known overlap-and-save method in the
frequency domain [111] where the accumulated CD is supposed known. Since the carrier
frequency offsets encountered in coherent optical communications can be detrimental to
the subcarrier selection, we use a single carrier header to coarsely identify and compensate
for the frequency offset [114]. This header is inserted at the beginning of the transmitted
sequence. Ref. [115] shows that frequency offsets can be coarsely equalized using a repetitive sequence with the pattern (B,B,-B,B), where B is a random sequence. We combine
pseudo-random binary sequences to create QPSK symbols for the header. Increasing the
length of B will improve the frequency resolution but lead to a limitation in the maximum
detectable frequency offset. We have observed experimentally that at a symbol rate of
32.5 GHz, choosing a sequence length of B=32 symbols is sufficient to avoid frequency
ambiguities as well as ensure a remaining frequency offset of less than 10 MHz. In order
to successfully estimate the frequency offset, the digital receiver must first identify the
header in the received signal. The frame synchronization algorithm is also based on [115].
Choosing a sequence length of B=32 symbols ensures a correct frame identification in all
our experiments.
66
6.3.2
Channel estimation and equalization
The subcarriers are separated using the basic scheme shown in Fig. 6.2. From this point,
DSP is carried out in parallel for all subcarriers. On each subcarrier, the complete signal
is shifted to the center frequency. Then the subcarrier is filtered out using the RRC
filter u [n] that is matched to the transmitter filter on the subcarriers. The signal is
down-sampled to two samples per symbol.
The matrix burst model in [50] is used to model the transmission of the training sequence given the channel model in section 6.2. In this model, the channel is supposed
to be of finite length of 30 4 taps at twice the aggregate symbol rate on the four butterfly components. The least-mean square (LMS) method [50] for inverting the channel
using the pseudo-inverse of the burst channel model is used to estimate the channel coefficients. The per-subcarrier channel estimation uses a known pilot pattern of length
384 on each subcarrier as a trade-off between noise sensitivity and resilience to channel
variations for the aforementioned channel model. We derive the MMSE continuous butterfly per-subchannel equalizers [105] based on the subcarrier estimations. The equalizer
architecture is shown in Fig. 6.3 for the subcarrier m. In order to track channel variations mainly due to polarization effects, we implement an adaptive equalizer based on
the stochastic gradient algorithm [57] updated using a decision-directed approach [29].
The MMSE equalizer is initialized using the training sequences. After that, the equalizer
switches to tracking mode.
6.3.3
Phase tracking
In MC-OQAM any phase shift will result in a change of the optimum butterfly filter
values. The phase offset represents both the beating signal between the transmission
and the receiver lasers and nonlinear phase noise stemming from the optical Kerr effect
during transmission. The linewidths of the lasers typically used in coherent optical communications encompass tens of MHz. Even in back-to-back configurations, phase noise
will therefore be the fastest-varying impairment to be corrected [116]. It is therefore crucial to ensure that the speed of convergence of the MC-OQAM butterfly equalizer can
match the characteristic speed of the fast phase distortions in MC-OQAM. The speed of
the stochastic gradient algorithms typically used to update the equalizer in single carrier QAM will not be sufficient. Furthermore, in MC-OQAM, the symbol rate on each
subcarrier is divided by the number of subcarriers. In the referential of one subcarrier,
a given carrier phase variation will therefore vary proportionally faster. It follows that
carrier phase tracking will become increasingly difficult when the number of subcarriers
in MC-OQAM is increased.
To ensure phase tracking when transmitting over substantial length of fiber and/or
using more than one subcarrier, we propose a butterfly equalizer using trial phases. The
architecture presented in the previous section is updated. The algorithm is detailed in
Fig. 6.4. The method is inspired from maximum-likelihood phase tracking [63] but is
less complex to implement. First, trial phases are initialized on the three arms of the
equalizer as [ 0 ] where is a free parameter representing the phase resolution of
the algorithm.
After each block of N symbols, the block-wise average square error is calculated with
67
Phase update
R
Rotation
+
I
R
MMSE
Equalizer
I
R
r1,m
[n]
u [n] M/2
Rotation
0
u [n] M/2
I
R
I
r1,m
[n]
MMSE
Equalizer
R
r2,m
[n]
R
r2,m
[n]
I
R
Rotation
-
MMSE
Equalizer
R
I
Figure 6.4: MC-OQAM phase-corrected MMSE equalizer. (Rotation +) designates a

complex-plane rotation of + degrees, where is the current overall estimated phase
shift and is the test phase of the tracking algorithm. (Rotation 0) is a rotation of
and (Rotation ) is a rotation of degrees.
respect to the nearest constellation points. The phase value minimizing the average square
error is selected. The test phases are then updated so that the phase minimizing the error
migrates to the central arm and the corresponding equalizer is copied on the three arms.
The output signal is always selected on the central arm.
In this updated version, the butterfly equalizer has three convergence parameters: the
conventional of the stochastic gradient, the phase resolution and the block size N
of the fast phase tracking. When the phase variations are sufficiently slow so as to stay
approximately constant over thousands of symbols, the stochastic gradient tracks the
phase variations and the fast phase tracking is useless. However, if the phase variation is
faster, the phase tracking takes over this part of the equalization, leaving the stochastic
gradient with the slowly varying polarization and dispersion impairments. It is important
to note that the fast phase tracking alone cannot equalize all channel impairments as it
is explicitly set up for phase tracking only. The parameters for DSP are summarized in
Table 6.1.
68
Table 6.1: DSP

Parameter
Total length of DAC sequences
Frame header length [symb]
Training sequence length [symb]
Length of channel response [taps]
Stoch. gradient convergence
Phase tracking block size N
Test phase [degrees]
6.3.4
parameters
Value range
218
128
384
304
0.004
40 [1-3 subc.], 20 [5-7 subc.]
5
Experimental results
The experimental setup is illustrated in Fig. 6.5. A programmable DAC provides 32.5
GBd MC-OQPSK or MC-QPSK waveforms at two samples per symbol, to modulate the
continuous-wave optical field of the lasers. For MC-QPSK, we eliminated the half-symbol
delay from the transmitter pulse shaping. We use linear precompensation to mitigate
the transmitter low-pass characteristics. The DAC memory limits the total length of the
sequences to 218 samples. One laser source per MC-(O)QAM waveform is modulated
using a polarization-multiplexed IQ-modulator. At the receiver, the whole MC-(O)QAM
waveform is mixed with a local oscillator and sampled. The demodulation is done in the
digital domain. The WDM transmitter involves an array of 64 distributed feedback lasers
which fall on the standard 50 GHz ITU frequency grid. The central test channel at 1545.72
nm is replaced by an external cavity laser and modulated via a polarization-multiplexed
IQ-modulator. All other sources used are spliced to a polarization division multiplexing
emulator which divides, decorrelates and recombines the data through a polarization beam
combiner with approximately 10 ns delay. The signal is then passed into a low-speed (< 10
Hz) polarization scrambler (PS). The WDM multiplex is sent into a recirculation loop,
composed of four spans of 100 km SSMF. Fiber loss is compensated for by hybrid RamanErbium optical repeaters. The Raman amplifier has two back-propagating laser diode
pumps at wavelengths 1432 nm and 1457 nm and offers approximately 10 dB on-off gain.
At the receiver side, the multicarrier signal is mixed with a local oscillator, sampled, and
recorded, by a 80 GSamples/s real-time scope of 33 GHz bandwidth. Table 6.1 summarizes
the DSP parameters. For MC-OQPSK, the DSP algorithms presented in the preceding
sections are used. For MC-QPSK, the receiver phase tracking is replaced with the Viterbi
and Viterbi algorithm.
We have measured both MC-QPSK and MC-OQPSK waveform sets for different rolloff
factors and subcarrier counts 1, 3, 5 or 7. To evaluate the performance, we have measured
the bit-error ratio (BER) on each subcarrier, calculated the average BER over subcarriers
for the multicarrier signal, and translated the average BER to the Q2 . We have both
performed back to back experiments and transmission experiments using a recirculation
loop of 4100 km of SSMF.
In Fig. 6.7 (a) and (b), we report the back-to-back OSNR0.1 nm necessary to achieve
either Q2 = 8 dB or Q2 = 9.5 dB for MC-QPSK and MC-OQPSK respectively. As noted
in section 3.7, it is interesting to target a certain Q2 -factor or BER because the various
FEC pick up the signal at a fixed BER. The different targets are therefore representative
69
MC(O)QPSK
even
PS
IQ-MOD PME
PDM-MC(O)QPSK
Tuneable
filter
ECL
DAC
Coherent ADC DSP

RX
WSS
IQ-MOD
IQ-MOD
WSS
MC(O)QPSK
LO
WSS
DAC
odd
63 DFB laser sources, 50 GHz spacing
DAC
LSPS
IQ-MOD PME
EDFA
N
SSMF
PS
100 km
x4
Raman pump
Figure 6.5: Experimental setup. ECL: external cavity laser, DFB: distributed feedback,
IQ-MOD: IQ-modulator, DAC: digital-to-analog converter, PME: polarization division
multiplexing emulator, PS: polarization scrambler, WSS: wavelength selective switch,
SSMF: standard single mode fiber, EDFA: erbium doped fiber amplifier, LO: local oscillator, ADC: analog-to-digital converter DSP: digital signal processing.
(a)
(b)
[nm]
[nm]
(a)
Figure 6.6: Measured spectra. (a) Seven-subcarrier MC-QPSK with an increased subchannel spacing (b) Seven-subcarrier MC-OQPSK. The channel spacing equal to the symbol
rate means that the power is constant over the signal bandwidth
70

21
21
Q2
target
20
= 9.5 dB
OSNR(0.1 nm) [dB]
19
Rolloff 0.01
Rolloff 0.25
Rolloff 0.5
19
18
18
17
17
16
2
target
Q2target = 9.5 dB
Q2target = 8 dB
16
= 8 dB
15
15
14
Rolloff 0.01
Rolloff 0.25
Rolloff 0.5
20
3
5
Number of subcarriers
(a) MC-QPSK back-to-back performance
14
3
5
(b) MC-OQPSK back-to-back performance
Figure 6.7: OSNR0.1 nm required to achieve target Q2 -factor in a back-to-back configuration
of different generations of FEC codes. The rolloff factors vary from to 0 to 0.5 and
the number of subcarriers vary from one to seven. The figure illustrates the crosstalk
limitation of MC-QAM. As can be seen in section 4.2, the support of the RRC filters
increase with the rolloff factor which results in crosstalk. It can moreover be observed that
the performance degrades with an increasing number of subcarriers. For MC-OQPSK,
the required OSNR0.1 nm increases by 1.5 dB/2 dB for target Q2 -factors 8 or 9.5 dB
respectively doing from one to seven subcarriers. This is attributed to increased sensitivity
to phase offsets as the individual subcarrier symbol rate is inversely proportional to the
number of subcarriers. Because of subcarrier crosstalk, the performance also degrades
when the rolloff factor is increased. At a Q2 -factor of 8 dB, we report 2.5 dB OSNR0.1 nm
implementation penalty with respect to the theoretical sensitivity both for MC-QPSK
and MC-OQPSK. It can be seen that the performance is strongly dependent on the rolloff
factor for MC-QPSK because of the subcarrier crosstalk. On the contrary, Fig. 6.7(b)
illustrates the subcarrier orthogonality of OQAM since the performances for all rolloff
factors are similar for a target Q2 = 8 dB. However, the performance is slightly degraded
for large rolloff factors for a target Q2 = 9.5 dB, indicating imperfect channel impairment
mitigation.
For the transmission experiments, we have first identified the optimum loop input
power for the studied multicarrier formats. In Fig. 6.8 (a) and (b), the performance after
4000 km (10 loops) of dispersion-unmanaged SSMF is shown for different input powers for
MC-OQPSK and MC-QPSK respectively. These results illustrate the trade-off between
noise and nonlinear effects in optical communication systems introduced in section 2.5.6.
It can be observed that the nonlinear power threshold does depend neither on the number
of subcarriers nor on offset-modulation. For MC-QPSK, the DSP did not converge for
subcarrier counts five and seven for the input powers larger than the nonlinear threshold,

11
10
10
Q2 [dB]
11
71
8
14
QPSK 1 subc
QPSK 3 subc
QPSK 5 subc
QPSK 7 subc
15
16
17
Loop power input [dBm]
18
(a) MC-QPSK 4000 km transmission performance
8
14
OQPSK 1 subc
OQPSK 3 subc
OQPSK 5 subc
OQPSK 7 subc
15
16
17
Loop power input [dBm]
18
(b) MC-OQPSK 4000 km transmission performance
Figure 6.8: Q2 -factor as a function of the loop input power after 4000 km (10 loops) of
SSMF for a varying number of subcarriers. The rolloff factor = 0.1.
indicating a loss for DSP stability due subcarrier crosstalk in these cases. Even for a
rolloff factor as small as 0.1, MC-OQPSK outperforms MC-QPSK thanks to the subcarrier
orthogonality. The performance degrades roughly by 1 dB/0.8 dB when increasing the
number of subcarriers from one to seven for MC-QPSK/MC-OQPSK respectively.
In Fig. 6.9, we study the performance of each subcarrier separately. It illustrates the
Q2 -factor of seven-subcarrier MC-QPSK/MC-OQPSK. We transmitted over 4000 km of
DU SSMF at the optimum input power and used a rolloff factor of 0.3. In the case of
MC-QPSK, since there is no subcarrier orthogonality, the central subcarriers suffer from
twice as much crosstalk as the edge subcarriers. In the case of MC-OQPSK, the overall
performance averaged over subcarriers is better. We observe that there is a Q2 -factor
deterioration of 0.8 dB on the edge subcarriers, indicating that the transmitter DAC
and/or receiver ADC impairments are more stringent on higher frequencies. The performance distribution is inverted with respect to MC-QPSK. We conclude that in the case
of MC-OQPSK, the receiver DSP has been able to recover the subcarrier orthogonality.
In Fig. 6.10, we compare the Q2 -factor of MC-QPSK and MC-OQPSK in a transmission experiment. The optimal loop input power is used for different rolloff factors on
a 4000 km link of DU SSMF. The results confirm the crosstalk limitation of MC-QPSK
that we already observed in the back-to-back case. Similarly to the back-to-back case, the
performance degrades with an increasing number of subcarriers and rolloff factor. Since
no information is shared between subcarrier processing at the receiver, the phase tracking
is more heavily impaired when the number of subcarriers is increased. In this experiment, one-subcarrier QPSK represents the state of the art single carrier modulation. For
MC-QPSK, the algorithms did not converge for rolloff factors higher than 0.3. Since MCOQPSK offers subcarrier orthogonality, the performance degrades less than MC-QPSK
when the rolloff factor is increased. For seven-subcarrier MC-OQPSK and for an approx-
72
Q2 [dB]
10
OQPSK
QPSK
2
3
4
5
6
Figure 6.9: Assessment of the subcarrier performance after 4000 km of DU SSMF for
7-subcarrier MC-QPSK and MC-OQPSK using a rolloff factor of 0.3.
11
OQPSK
Q2 [dB]
10
QPSK
9
Rolloff 0.1
Rolloff 0.3
Rolloff 0.5
8
1
3
5
Figure 6.10: Comparison of MC-OQPSK and MC-QPSK after 4000 km of DU SSMF for
different rolloff factors.
6.4. Conclusion
73
12
OQPSK 1 subc
OQPSK 3 subc
OQPSK 5 subc
OQPSK 7 subc
Q2 [dB]
10
6
2000
3600
5200
6800
8400
Transmission distance [km]
Figure 6.11: Performance of MC-OQPSK as a function of distance over DU SSMF.

imate Q2 -factor of 10 dB, the performance drops about 0.5 dB going from rolloff = 0.1 to
rolloff = 0.5. This indicates that the receiver DSP has not entirely removed the subcarrier
crosstalk as has already been concluded for the back-to-back case. We further observe
that one and three subcarrier MC-OQPSK perform very similarly to single carrier QPSK,
while five and seven subcarrier MC-OQPSK suffer from a Q2 -factor degradation smaller
than 0.8 dB compared to single carrier QPSK. We conclude that for a small number of
subcarriers, MC-OQAM offers similar performance as single carrier QPSK while enabling
subcarrier parallelization of the DSP algorithms.
In Fig. 6.11, we finally study the performance of MC-OQPSK as a function of transmission distance. One-subcarrier and three subcarrier MC-OQPSK have been transmitted
over 8400 km with a Q2 -factor of 6 dB. Five-subcarrier and seven-subcarrier MC-OQPSK
have been transmitted over 6800 km with a Q2 -factor of 6 dB. The results show that MCOQAM can be transmitted over long-haul distances while enabling DSP parallelization
over the subcarriers.
6.4
Conclusion
In this chapter, we investigate electrical multicarrier offset-QAM for long-haul coherent

optical communications. Offset modulations achieve crosstalk-free communication on each
subcarrier with a channel spacing equal to the symbol rate. We highlight that in coherent
optical communication systems, several physical effects impair the recovering of the payload symbols by breaking the subcarrier orthogonality. We therefore propose a dedicated
DSP using training sequences to estimate the communication channel and a new algorithm
to track the phase in MC-OQAM. We compare MC-OQAM with conventional MC-QAM
modulations so as to assess the benefits of crosstalk mitigation via OQAM. We carry out
experiments to test the transmission performance of MC-OQAM in dispersion-unmanaged
74
long-haul transmission links. MC-OQPSK outperforms MC-QPSK for rolloff factors as

small as 0.1. Rolloff factors superior to 0.1 lead to significant performance degradations
and DSP instabilities for MC-QPSK while MC-OQPSK has been demonstrated at a rolloff
factors up to 0.5 thanks to the subcarrier orthogonality. The performance degrades with
an increasing number of subcarriers due to the increased sensitivity with respect to phase
offsets. For MC-OQPSK, the required OSNR increases by 1.5 dB for a target Q2 -factor
of 8 dB going from one to seven subcarriers. We have shown that at 32.5 Gbd aggregate
symbol rate, one and three subcarrier MC-OQPSK can be transmitted over 8400 km with
a Q2 -factor of 6 dB. Five- and seven-subcarrier MC-OQPSK have been transmitted over
6800 km with a Q2 -factor of 6 dB. We conclude that MC-OQAM can work nearly as well
as single carrier QPSK while enabling DSP parallelization over the subcarriers.
Chapter 7
Low-complexity equalization for
OQAM modulations
Contents
7.1
7.1
Introduction
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
75
7.2
A low-complexity MC OQAM implementation . . . . . . . . . . .
76
7.3
Simulation setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
77
7.4
CD compensation . . . . . . . . . . . . . . . . . . . . . . . . . . . .
79
7.4.1
SC CD compensation
. . . . . . . . . . . . . . . . . . . . . . . . . . 79
7.4.2
OQAM CD compensation . . . . . . . . . . . . . . . . . . . . . . . . 80
7.5
Phase noise equalization . . . . . . . . . . . . . . . . . . . . . . . .
82
7.6
Cross-subcarrier equalizer . . . . . . . . . . . . . . . . . . . . . . .
82
7.7
Extension to polarization division multiplexing and channel estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
85
7.8
Hybrid SC/MC modulation . . . . . . . . . . . . . . . . . . . . . .
86
7.9
Computational complexity comparison between single carrier

and multicarrier-OQAM . . . . . . . . . . . . . . . . . . . . . . . .
91
7.10 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
92
Introduction
In the previous chapter, we investigated the equalization strategy for RRC-shaped OQAM
modulations using a channel spacing equal to the symbol rate for up to seven subcarriers
inside one 32.5 Gbaud optical channel. Our goal was to study the experimental feasibility
for OQAM modulations for coherent optical communications. In the final chapter of this
thesis, we increase the number of subcarriers and use numerical simulations to explore
the potential of OQAM modulation to reduce the computational complexity of the DSP.
76
Chapter 7. Low-complexity equalization for OQAM modulations
Frequency
Frequency
Figure 7.1: Illustration for electrical multicarrier modulation in optical communications.

In the preceding chapter, we used up to seven subcarriers. In this chapter, we increase
the number of subcarriers to up to 8192 for each optical channel.
In wireless communications, MC modulations are popular because highly frequencyselective channels can be subdivided into frequency slots with quasi-flat frequency responses. As a consequence, one-tap equalizers can be implemented on each subchannel,
which results in overall computational complexity savings [117119]. In this chapter, we
apply this principle to OQAM modulations in coherent optical communications. The
number of subcarriers is increased with respect to the previous chapter. Up to 8192
subcarriers will be considered inside the 30 GHz range of one optical channel.
The equalizer architecture is completely reviewed with respect to chapter 6. In fact,
we observed that the sensitivity of OQAM modulation to phase offsets increases with the
number of subcarriers because the dynamics of the phase variations catch up with the
symbol rate of the individual subcarriers. As a consequence, if we wish to further increase
the number of subcarriers at the same aggregate symbol rate, the equalizer architecture
is to be reviewed. In section 7.2, we introduce a low-complexity implementation of the
RRC OQAM modulation format that is particularly suited for high subcarrier densities.
In section 7.3, we detail the parameters of the simulation setup and the architecture of
the equalizer. Section 7.4 studies the potential to equalize CD using the chosen equalizer
structure. In sections 7.5 and 7.6, the influence of phase offsets on the performance of
the equalizer is studied and section 7.7 extends the simulations to polarization effects and
introduces a strategy to estimate the communication channel. In section 7.8, we propose a
hybrid single-/MC modulation in order to better cope with the phase offsets. Finally, we
compare the computational efforts for DSP in standard SC and MC OQAM modulations
in section 7.9.
7.2
A low-complexity MC OQAM implementation
In chapter 6, we introduced a mathematical model for OQAM modulations. In a realistic implementation, the phase shifts are carried out by IFFT and FFT which saves
computational complexity [120].
In this chapter, we design low-complexity digital OQAM equalizers. Therefore, we use
an IFFT/FFT model of OQAM modulation and demodulation. The use of discrete Fourier
transforms for MC communications is quite intuitive: every entry of the FFT or IFFT is
multipled by another complex exponential. As a result, a discrete frequency is assigned
7.3. Simulation setup
77
to every entry. At the receiver, the signal is converted back into baseband by the inverse
operations. The MC OQAM modulation is illustrated in Fig. 7.2. Taken as a whole,
it is mathematically equivalent to the modulation presented in chapter 6 (see Fig. 6.2).
The demodulation is illustrated in Fig. 7.3. Again, it is mathematically equivalent to the
demodulation in Fig. 6.2. For the sake of simplicity, only one polarization component
is shown in both graphs. Both modulation and demodulation require one IFFT/FFT
operation of size M (number of subcarriers) respectively. In this aspect, MC OQAM is
similar to OFDM. However, additional filters Gm [o] are required in order to ensure the
RRC pulse shaping on each subcarrier FD to TD conversion. They represent a polyphase
decomposition of the pulse shaping filters defined in Eq. 6.1. Their derivation is most
convenient in z-transform [120]:
G0l (z) =
u[l + nM ]z n
(7.1)
Gl [z] = G0l (z 2 )
(7.2)
Fig. 7.3 also illustrates a one-tap FD equalizer: on each subcarrier, the receiver FD signal
is multiplied by a complex number before the half-symbol time recovery.
The method used to quantify the computational cost of an algorithm heavily depends
on the platform on which the algorithm is implemented. In application-specific intergrated
circuits (ASICs) or field-programmable gate arrays (FPGAs), the power consumption
associated with multipliers is much larger than for adders. In addition, integration experts
develop cost-efficient solutions for specific DSP algorithms. This discussion being beyond
the scope of this chapter, we choose to assimilate the computational complexity to the
number of real multiplications required per equalized symbol Csymb .
7.3
Simulation setup
We use numerical simulations to assess the performance of OQAM modulation using a FD

equalizer. The simulation is set up according to Fig. 7.2 and Fig. 7.3. The transmitter
implements a synthesis filterbank composed of M subchannels, each modulated with
QPSK information symbols. In our simulations, M is a power of two. This choice ensures
minimum computational complexity for the FFT/IFFT. We generate 219 QPSK symbols
and distribute them among the subcarriers. For the ease of notation, we separate the
I
complex symbols into real and imaginary parts sR
i,m [n] and si,m [n]. Index i = 1, 2 refers to
the two polarizations at the transmitter, index m = 1...M refers to the M subchannels,
while n is the symbol index. The OQAM modulation is implemented by upsampling the
symbols by two and delaying the imaginary part of each subchannel by half a symbol.
The signals are multiplied by m [o] ensuring the equivalence between the IFFT/FFT
implementation and the OQAM principle presented in chapter 6. Their expression can
be found in [120]. The sequence index o indicates the temporal position in the twiceoversampled signals. The signals undergo IFFT of size M , mixing the symbols on all
subchannels. They are then convolved individually with the real FIR filters Gm [o] [121].
The signals are upsampled by M/2. The M subchannels are added together after relative
one-sample delays.
78
sR
i,1 [n]
sIi,1 [n]
sR
i,m [n]
sIi,m [n]
sR
i,M [n]
sIi,M [n]
2
.
.
.
1 [o]
z
1
j=
2
.
.
.
z 1
M/2
m [o]
z
1
IFFT
(M)
Gm [o] M/2
z m
G1 [o]
M [o]
z
1
GM [o] M/2
z M
Figure 7.2: Discrete MC-OQAM transmitter implementation using IFFT.
z1
1 [o]
G1 [o] M/2
z1
.
.
.
Gm [o] M/2
FFT
(M)
m [o]
z1
.
.
.
zM
M [o]
GM [o] M/2

One-tap FD
equalizer
z1
R
ri,1
[n]
I
ri,1
[n]
R
ri,m
[n]
I
ri,m
[n]
R
ri,M
[n]
I
ri,M
[n]
Figure 7.3: Discrete MC-OQAM receiver implementation using FFT. One-tap FD equalizer.
7.4. CD compensation
79
At the receiver, the analysis filterbank implements the counter-part of each operation
performed at the transmitter. The individual delays are recovered and for each subcarrier, the signal is convolved with the matched filterbank Gm [o]. The signals are then
downsampled by M/2 and transformed by FFT of size M . Each subcarrier signal is multiplied by the equalizer. The signals are multiplied by the conjugate complex of m [o] to
recover the original symbols. Real and imaginary parts of the signal are separated. The
half symbol delay applied on the imaginary branch is compensated so that both the real
and imaginary signals are received synchronously. The signal is finally downsampled to
one sample per symbol to yield the real symbols riR [n] and riI [n]. For the simulation of
the in-line impairments such as CD, we suppose an aggregate symbol rate of 30 GHz for
all simulations. We used the same rolloff factor = 1 for all simulations. We assume zero
carrier frequency offset and sample timing jitter.
7.4
7.4.1
CD compensation
SC CD compensation
The standard DSP method to mitigate CD in SC modulations, overlap and save in FD,
was introduced in section 3.3. Supposing that the size of the FFT is a power of two,
we will assume the efficient "split-radix " implementations of FFT and IFFT [122]. The
number of real multiplications required for the calculation of one FFT/IFFT of size NFFT
is:
C = 4NFFT log2 (NFFT ) 6NFFT + 8
(7.3)
Overlap and save equalization requires one FFT, NFFT complex multiplications and one
IFFT operation for the equalization of NFFT K samples. The real multiplier count Csymb
per equalized sample is therefore:
Csymb =
8NFFT log2 (NFFT ) 6NFFT + 16

NFFT K
(7.4)
The overlap K is a central parameter. Its requirement can be linked to the amount of
CD to be compensated for [123]:
Kd
2
c fsymb
fc2
(7.5)
where fsymb is the symbol frequency and fc the optical carrier frequency. c is the speed
of light and d is the residual dispersion in ps/nm. For example, a 30 Gbaud transmission
over 1000 km of a dispersion-unmanaged SSMF link at fc =193 THz, d = 17000 and
K 110. The output of the equalizer is independent of NFFT as long as NFFT is greater
than K. As a consequence NFFT is designed to minimize Csymb in Eq. (7.4). The number
of real multiplications per equalized sample in SC overlap and save FD CD equalization
is shown in Fig. 7.4 as a function of the FFT size NFFT for different lengths of SSMF
links. As an example, for Lf iber = 1000 km, NFFT = 1024 will be chosen and Csymb = 83
multiplications per equalized sample are required.
80

400
Lfibre = 500 km
Lfibre = 1000 km
350
Lfibre = 4000 km
symb
300
250
200
150
100
50
6
9
10
log2(N)
11
12
13
Figure 7.4: Number of real multiplications per equalized sample in SC overlap and save
FD CD equalization as a function of the FFT size NFFT for different lengths of SSMF.
2
1.5
1
0.5
0
0.5
1
1.5
2
2
1
1
(a)
EVM0 = -16 dB
2
2
2
(b) EVM0= -22 dB
2
22
(c) EVM0= -34 dB
Figure 7.5: Recovered QPSK constellations, different EVM. Lf iber = 1000 km
7.4.2
OQAM CD compensation
In OQAM modulations, FD equalization is implemented by multiplying the received samples on each subcarrier by a complex number after the receiver FFT (see Fig. 7.3) similarly
to OFDM [124]. The one-tap CD equalizer is designed by inverting Eq. (2.8) and discretizing over the M subcarriers. Here, no overlap is necessary.
We study the performance of the OQAM system as a function of the number of subcarriers M using the mathematical model defined in the previous section. Here, the channel
includes only CD. As a performance metric, we use the error vector magnitude (EVM),
I
which is the mean square error between the sent symbols sR
i [n]/si [n] and the received symbols riR [n]/riI [n]. The recovered QPSK constellations are illustrated for different EVM in
Fig. 7.5.
In Fig. 7.6, we show the EVM as a function of the number of subcarriers for different
fiber lengths. It can be observed that the EVM increases with Lf iber . When the fiber
7.4. CD compensation
81
EVM [dB]
10
20
Lfiber= 4000 km
Lfiber= 2000 km
30
= 1000 km
fiber
Lfiber= 500 km
40
Lfiber= 200 km
50
3
8
9
log2(M)
10
11
12
13
Figure 7.6: EVM as a function of the number of subcarriers for different fiber lengths.
length increases, so does the accumulated dispersion. The phase shifts on each subcarrier
show a faster variation with respect to the frequency and each subcarrier channel becomes
less frequency-flat. The EVM also increases when M is decreased. At the same aggregate
baudrate, the channel appears less frequency-flat on subcarriers with a higher bandwidth.
For example, in order to achieve EVM < 30 dB, the number of subcarriers must be higher
than 4096 for 2000 km of uncompensated SMF CD and higher than 2048 for 1000 km of
SSMF.
In order to fairly compare SC and OQAM CD mitigation complexity, we assume that
the FFT and IFFT of OQAM modulation and demodulation are part of the equalization
scheme. In fact, the computational complexity of pulse shaping and receiver filtering is
often neglected. This is not a valid assumption for MC OQAM because of the presence
of high-complexity Fourier transform operations. Including FFT/IFFT complexity, the
number of real multiplications required for MC OQAM CD equalization is
Csymb =
8M log2 (M ) 6M + 16
M
(7.6)
For each length of SSMF, we can determine the minimum M to ensure EVM 30
dB and compute the computational complexity. The results are shown in Table 7.1. We
compare the results with the SC complexity. We conclude that taking into account the
Fourier transforms, the complexity is nearly the same. In conclusion, the overlap represents a negligible complexity overhead and MC-OQAM does not reduce the complexity
for CD compensation. It can however be argued that MC OQAM distributes the complexity of equalization between the transmitter and the receiver whereas the SC equalization
complexity is concentrated at the receiver.
82
Table 7.1: Computational complexity per

Lf iber [km] Csymb (SC) Csymb
500
73
1000
83
2000
92
4000
101
7.5
equalized sample
(OQAM)
74
82
90
98
Phase noise equalization
Phase offsets represent a persistent drawback of MC modulations using FD equalization.

In fact, a varying phase mathematically manifests as a TD multiplication by the nonconstant complex number exp(j[n]), where [n] is the phase noise at sample n. In FD,
this represents a convolution. As a consequence the subcarriers interfere. In order to
equalize a time-varying phase, it is therefore necessary to mix neighboring subcarrier signals which is impossible using the one-tap equalizer introduced in the previous section. In
MC OQAM, the TD signal also passes a filterbank before FD transformation. This means
that the TD samples and phase noise samples are mixed. Therefore, even a convolutional
FD equalizer will not be able to completely remove the effects of phase offsets.
The phase noise is simulated as Brownian motion according to the basic assumption
introduced in section 2.5.1. Each TD sample is multiplied by the complex exponential
exp(j[n]). In opposition to CD, phase noise is thus more easily equalized in TD, where it
is sufficient to multiply each sample by the complex conjugate of the phase noise samples.
We will nonetheless attempt to use a pure FD equalizer. The effect of phase noise on
the performance of MC OQAM is shown in Fig. 7.7. Its linewidth is varied in order to
simulate various specifications of transmitter laser and local oscillators. In this simulation,
only the phase noise impairment is considered. Here, we suppose that the mean phase
over one MC symbol duration is known. It can be observed that the EVM increases with
the number of subcarriers M since the duration one MC symbol is proportional to M .
We conclude that the laser linewidth poses a very stringent constraint when using only a
one-tap FD equalizer. The 30 dB EVM requirement can only be satisfied for M 32
for a 500 kHz linewidth and M 512 for a 50 kHz linewidth.
We study the combined effect of CD and phase noise in Fig. 7.8. The one-tap FD
equalizer combines CD compensation and mean phase compensation. It is clear that
the system requirements on M for CD and phase noise compensation conflict: for a
small number of subcarriers, the CD limitation dominates because the FFT are not long
enough to resolve the channel. For a large number of subcarriers, the symbols are too slow
comparing with the phase noise. As a consequence, even for a moderate laser linewidth
of 500 kHz and a propagation distance of 1000 km, the EVM cannot be reduced to an
acceptable level.
7.6
Cross-subcarrier equalizer
In order to enhance the performance of the MC OQAM communication scheme, we implement a cross-subcarrier equalizer as shown in Fig. 7.9. The phase noise is supposed
7.6. Cross-subcarrier equalizer
83
5
= 500 kHz
= 100 kHz
= 50 kHz
= 10 kHz
= 1 kHz
0
5
EVM [dB]
10
15
20
25
30
35
40
45
3
8
log2(M)
10
11
12
13
Figure 7.7: EVM as a function of the number of subcarriers for different phase noise
linewidths.
0
5
EVM [dB]
10
15
Lfiber = 4000 km, = 5 kHz
20

25
30

35
40
8
16
32
64
128
256
M
512
1024
2048
4096
8192
Figure 7.8: EVM as a function of the number of subcarriers for different fiber lengths /
phase noise linewidths.
84
...
FFT
(M)
...
z[1]
Z[1]
z[2]
Z[2]
z[3]
Z[3]
+
+
+
+
FFT
(M)
...
...
+
+
+
(a) One-tap FD equalizer
(b) Three-tap cross-subcarrier FD equalizer
Figure 7.9: Architecture of one-tap and three-tap cross-subcarrier FD equalizer. The

different colors depict the multiplication by a different tap of the equalizer.
ej[n]
...
G1 [n]
M/2
G2 [n]
M/2
G3 [n]
M/2
...
G4 [n]
M/2
G5 [n]
M/2
G6 [n]
M/2
...
G7 [n]
M/2
...
G8 [n]
M/2
G1 [n]
M/2
G2 [n]
M/2
G3 [n]
M/2
G4 [n]
M/2
G5 [n]
M/2
G6 [n]
M/2
G7 [n]
M/2
G8 [n]
M/2
FFT
(M)
FFT
(M)
...
...
+
+
...
Figure 7.10: Principle of the derivation of the cross-subcarrier FD equalizer.
7.7. Extension to polarization division multiplexing and channel estimation

85
5
0
EVM [dB]
5
10
1 tap
3 taps
5 taps
7 taps
15
20
25
3
8
log2(M)
10
11
12
13
Figure 7.11: EVM as a function of the number of subcarriers for different lengths of
cross-subcarrier FD equalizer. Lf iber = 4000 km, = 500 kHz.
known. The principle of equalization is simple: if phase noise is equalized by multiplication in TD, a convolution can be used in FD. More precisely, be z[n], n = 1...M the TD
samples of one MC symbol (see Fig. 7.9) and z 0 [n] the ideally equalized signal. Z[k] and
Z 0 [k], k = 1...M is one FD block after FFT. In FD, we have:
z 0 [n] = exp(j[n]) n = 1...M Z 0 [k] = Z[k] F(exp(j[n])) k = 1...M
(7.7)
In this equation F is the FFT operator and is the circular convolution. The phase noise
samples are passed through the filterbank as shown in Fig. 7.10 to synchronize the phase
noise samples to the z[n]. This approach is only approximate because z[n] is in fact a mix
of several TD samples. The EVM performance of the cross-subcarrier equalizer is shown
in Fig. 7.11 as a function of the number of taps for a fiber length of Lf iber = 4000 km and
a phase noise linewidth = 500 kHz. In conclusion, the EVM of FD equalization in the
presence of CD and phase noise can be reduced by several dB by using cross-subcarrier
equalization. However, the use of cross-subcarrier equalization cannot reduce the EVM
to an acceptable level for realistic system parameters. Moreover, in order to implement
this equalization scheme, the phase noise must be known on each data sample. As a
consequence, we abandon FD equalization of the phase noise in this study.
7.7
Extension to polarization division multiplexing and

channel estimation
In this section, we extend the simulations to PMD. Unlike CD, polarization effects are
time-varying and their parameters are not known a priori. To successfully demultiplex
the polarization components of the PDM signal, we can either implement an adaptive
equalizer that autonomously tracks channel variations or implement a channel estimation
strategy. However, in a MC scenario, adaptive equalizers are challenged by the relative speed of channel variations and subcarrier signals as we experimentally observed in
chapter 6. Since this chapter deals with OQAM configurations using larger numbers of
86
subcarriers, we choose to estimate the communication channel using training symbols.

The distribution of the training symbols is shown in Fig 7.12. Because of PMD, the
optical channel is frequency-selective. As a consequence, we evenly distribute estimation subcarriers, where the transmitted symbols are known, in the MC spectrum. The
estimators stemming from the estimation channels are used to equalize the neighboring
subcarriers as shown by the black arrows in Fig. 7.12.
Referring to Fig. 7.3, this graph shows the transmitter OQAM symbols before the
multiplication by m,n as a function of time and subcarrier index. Because of the OQAM
modulation, the neighboring real symbols interfere pair-wise (combined, they compose
one complex QAM symbol). To avoid this, we use the Pairs of Pilots technique [125] in
which the real symbols corresponding to the imaginary part of each complex QAM symbol
is reduced to 0. The Pairs of Pilots cancellation is shown in light grey in Fig. 7.12. In
the presence of channel imperfections, the neighbor subcarriers interfere with the training
channels. To avoid this, we insert guard bands shown as blue bands in Fig. 7.12. Finally,
the training symbols were time-switched between the polarizations. The received training
symbols are extracted at the equalizer stage (see Fig. 7.3). The communication channel
is characterized by a 2 2 matrix on each subcarrier k

H11 [k] H12 [k]

H21 [k] H22 [k]

k = 1...M
(7.8)
The estimated channel response on subcarrier is obtained by simply dividing the received
training symbols by the sent training symbols. The dual-polarization zero-forcing equalizer is obtained by inverting the matrix in Eq. (7.8) on each subcarrier.
In a frequency-selective channel, an insufficient number of test channels will result in
performance penalties. In the following numerical simulations, we study the estimation
subchannel requirement stemming from PMD. In order to emulate this effect, we subdivide
the fiber length into shorter trunks of 400 m. In each trunk, the eigenmodes of polarization
are randomly
selected on the Poincare sphere. We simulated SSMF with a PMD coefficient
of 0.1 ps/ km. The results are shown in Fig. 7.13 for different fiber lengths. Phase noise
is ignored here. It can be observed that PMD results in a moderate frequency selectivity
since less than hundred subcarriers inside the 30 GHz bandwidth are generally sufficient
to ensure an acceptable EVM.
7.8
Hybrid SC/MC modulation
From the previous sections, we conclude that the main limitation of MC OQAM modulations using a large number of subcarriers is phase noise. In fact, CD can be equalized
satisfactorily using a fixed FD equalizer and PMD can be estimated using training subcarriers at an acceptable overhead while even moderate phase noise linewidths result in
EVM exceeding -15 dB. Unfortunately, MC communications are notoriously sensitive to
phase offsets. In summary, all impairments modeled as a convolution in TD are easily
compensated for in FD in MC modulations since it conveniently separates the available
bandwidth into small frequency slots. However, phase noise manifests as TD multiplication by varying complex exponential factors. As we concluded in section 7.6, transforming
7.8. Hybrid SC/MC modulation
87
Polarization 1
Polarization 2
time
subcarriers
0
0
Training symbols (real)

0
Pair-of-pilots cancellation
Guard band
Pilot polarization distribution
Figure 7.12: Distribution of training symbols (red) over the subcarriers and time. The
black arrows indicate the channels that are equalized using the channel estimation from
a particular subcarrier.
20
Lfiber = 1000 km
Lfiber = 2000 km
EVM [dB]
25
Lfiber = 4000 km
30
35
40
45
2
32
128
Number of training subcarriers
512
2048
Figure 7.13: EVM as a function of the number of training subcarriers. M = 8192.
88
P[a.u.]
Figure 7.14: SC/MC hybrid modulation

the one-tap FD equalizer into a convolutional FD equalizer only partially solves this problem since the time samples are scrambled by the filterbank in OQAM modulations. We
conclude that it is much more conventient to equalize the phase offsets in TD. SC modulations are very resilient to phase offsets because every occurrence of the sampled phase
offset only affects one symbol. Moreover, they are very easily usable for phase estimation
because one training symbol is sufficient to estimate the phase offset. Even more interestingly, when using P -PSK modulation formats, the SC phase offsets are eliminated very
reliably using the blind Viterbi and Viterbi algorithm as detailed in chapter 3.
The SC/MC difference of behavior with respect to the physical impairments in optical
fibers motivates the use of a hybrid modulation format as shown in Fig. 7.14. We propose
to replace the central part of the MC OQAM spectrum with a SC P -PSK modulation to
serve as a phase pilot. Phase offset equalization is now carried out in TD. It should be
noted that phase equalization and estimation need not be carried out in the same domain.
In particular, one could derive the phase information from FD pilots then equalize the
phase in TD which would result in a more integrated OQAM modulation scheme. In our
study, we chose to combine both in TD because of the convenient characteristics of the
SC modulation scheme.
We use RRC filters with a very small rolloff factor so that only a very small guard band
must be accommodated between the single- and MC parts of the spectrum. Comparing the
hybrid modulation with previous contributions using a known pilot waveform [126128],
the loss of spectral efficiency with respect to a pure MC signal is limited to the guard
band if the SC modulation format has the same spectral efficiency as the OQAM. We
use PDM-QPSK for the pilot to match the Offset-QPSK on the subcarriers. In Fig. 7.15,
we show the EVM as a function of phase noise linewidth for the case M = 8192. One
eighth of the spectrum is occupied by the pilot. We compare the previously-used one-tap
equalizer to the case where the phase is corrected in TD using the information from the
pilot prior to the subcarrier demodulation. We conclude that the phase can successfully
7.8. Hybrid SC/MC modulation
89
0
5
With TD Viterbi phase tracking

W/o TD Viterbi phase tracking
EVM [dB]
10
15
20
25
30
3
3.5
4.5
log10()
5.5
Figure 7.15: EVM as a function of the phase noise linewidth. M = 8192. Fraction of
the spectrum allocated to SC: 1/8. Viterbi and Viterbi averaging window = 5 symbols.
Number of training subcarriers = 256.
be recovered using a SC pilot occupying part of the spectrum.
In Fig. 7.16, we study the BER performance of the hybrid modulation format in the
presence of ASE noise. The averaged BER of MC subcarriers is shown as a function of
Eb /N0 for different phase noise linewidths. We defined the Eb /N0 in section 2.6. Using
different signal to noise ratios simulates different propagation lengths. The number of
subcarriers M = 8192. One eighth of the spectrum is occupied by the pilot. Thanks
to the TD noise mitigation, the channel estimation can be averaged over subsequent
subcarrier symbols which enhances the accuracy of the channel estimate in the presence
of noise. In our simulations, the channel estimation is averaged over 16 symbols on each
estimation subcarrier. The results show that the proposed equalization architecture is
robust to phase noise.
In Fig. 7.17 we study the impact of the ratio of the spectrum allocated to the SC pilot.
It can be observed that when the SC bandwidth is decreased, it becomes more and more
difficult to mitigate the phase noise. In fact, in the presence of ASE noise, a trade-off
exists for the minimum moving average window for the Viterbi and Viterbi. If the window
is too short, the ASE noise induces unacceptable fluctuations of the phase estimations. If
the window is too long, the algorithm is unable to track the rapid variations of the phase
noise. For a smaller pilot bandwidth, the time interval between SC symbols is increased.
As a consequence, it becomes more and more difficult to track rapid phase variations.
This is illustrated in Fig. 7.17 where it can be observed that for a phase noise linewidth
of = 50 kHz, 1/32th of the spectrum is sufficient for the SC pilot to track the phase. In
opposition, 1/4th of the spectrum is required at = 500 kHz.
So far, we have assumed a standard SC equalization on the pilot, i.e. overlap-andsave CD compensation and butterfly-FIR PMD filters. However, for a pilot with a small
bandwidth, equalization using a small number of taps is sufficient. The final equalization
architecture including the pilot is summarized in Fig. 7.18. In the upper-left corner, the
treatment of the SC pilot is shown: it is filtered-out (LPF or low pass filter), and a
few-tap equalizer is used, principally for polarization demultiplexing and synchronization.
90
10
Theory
= 5 kHz
= 50 kHz
= 500 kHz
10
BER
10
10
10
10
6
7
Eb/N0 [dB]
Figure 7.16: BER as a function of Eb /N0 for different phase noise linewidths. M = 8192.
Fraction of the spectrum allocated to SC: 1/8. Viterbi and Viterbi averaging window =
25 symbols. Number of training subcarriers = 256. The channel estimation is averaged
over 16 symbols on each estimation subcarrier.
10
10
10
10
SC ratio = 1/4
SC ratio = 1/8
SC ratio = 1/16
SC ratio = 1/32
SC ratio = 1/64
BER
BER
SC ratio = 1/4
SC ratio = 1/8
SC ratio = 1/16
10
20
40
60
80
Length of the moving average [symbols]
(a) = 500 kHz
100
10
20
40
60
80
Length of the moving average [symbols]
100
(b) = 50 kHz
Figure 7.17: BER as a function the length of the moving average of the CPE for different
SC to MC ratios. M = 8192. Number of training subcarriers = 256.
7.9. Computational complexity comparison between single carrier and

multicarrier-OQAM
91
V+V
V+V
Few-tap
equalizer
z1
1 [o]
G1 [o] M/2
z1
.
.
.
LPF
zm
Gm [o] M/2
m [o]
FFT
(M)
z1
.
.
.
z1
2 x 2 FD equalizer
zM
GM [o] M/2
G1 [o] M/2
M [o]
z1
1 [o]
z1
.
.
.
zm
Gm [o] M/2
FFT
(M)
m [o]
z1
.
.
.
zM
M [o]
GM [o] M/2
z1
R
r1,1
[n]
I
r1,1
[n]
R
r1,m
[n]
I
r1,m
[n]
R
r1,M
[n]
I
r1,M
[n]
R
r2,1
[n]
I
r2,1
[n]
R
r2,m
[n]
I
r2,m
[n]
R
r2,M
[n]
I
r2,M
[n]
Figure 7.18: Equalizer structure of the hybrid SC/MC modulation. LPF : lowpass filter.
V+V: Viterbi and Viterbi algorithm
The Viterbi and Viterbi algorithm (V+V) is used to estimate the phase. This information
is used to equalize the phase for the MC signal. The MC signal is then demultiplexed as
defined in Fig. 7.3. A one-tap FD equalizer is then used on every subcarrier. Finally, the
offsets between real and imaginary parts of the symbols are eliminated.
7.9
Computational complexity comparison between

single carrier and multicarrier-OQAM
In the previous paragraphs, we showed that the linear impairments of the optical channel
can be equalized in a MC OQAM modulation scheme thanks to the insertion of a SC
pilot. We noted that a low-complexity equalization is sufficient for the SC pilot. We
can now compare the hybrid modulation with a standard SC modulation in terms of
computational complexity. In Table 7.2, we summarize the main steps of the equalization
92
for each modulation format. For each block, we list the number of real multiplications
required for the equalization of one symbol. For the sake of simplicity, we neglect the
multiplications required for the equalization of the pilot in the hybrid scheme which is
legitimate when a few-tap pilot equalizer is used.
Table 7.2: Complexity comparison: DSP detail and real multiplication count per equalized
symbol Csymb . NCMA : number of taps of the CMA in the SC case. rSC is the ratio of the
bandwidth occupied with the SC pilot.
Equ. stage
SC
MC OQAM
FFT/IFFT
1-tap FD equalizer
FFT/IFFT
CD comp/
Convolution w butterfly FIR (NCM A ) 1-tap FD equalizer (22)
Pol Demux
FFT )6NFFT +16
+
Csymb = 8NFFT log2N(NFFT
Csymb = 8M log2 (MM)4M +16
K
2NCMA
Calculate error function
Estimate channel
Pol Demux update Stochastic gradient update
Invert (22) matrix
Csymb = 8NCMA + 6
Csymb = 24
CPE
Multiply by inv. phase noise
Multiply by inv. phase noise
Csymb = 16
Csymb = 16 rSC
We propose to study a concrete example. In the SC case, for Lf iber = 4000 km of
SSMF, and fs = 30 GHz K>425 according to Eq. (7.5) and NFFT = 4096. A typical
value of the length of the butterfly equalizer is NCMA = 15. For the hybrid modulation,
we choose M = 8192. In order to be able to mitigate phase noise with a linewidth up to
500 kHz, the SC to MC ratio rSC = 1/8. Table 7.3 summarizes the real multiplication
count. In this example, the computational complexity has been reduced by half thanks
to the hybrid modulation. It can be observed that a large part of the difference in
computational complexity stems from the CMA application and update. It can be argued
that in case of slow polarization variations, the CMA can also be set up for an update
every R symbols, where R is an integer which would divide the computational complexity
of the CMA update by R. Although the dynamics of PMD have not been studied in
this work, the equalizer dynamics are coherent in our example. In fact, the subcarrier
equalizers are updated once per subcarrier symbol period, which is M times the SC symbol
period. The CMA convergence being in the order of thousands of symbols, the dynamics
of both equalization techniques are comparable in order of magnitude. A finer study of
the update computational complexity should take into account PMD dynamics. However,
even without taking into account the update complexity, hybrid modulation reduces the
computational complexity by roughly 30 percent.
7.10
Conclusion
In this chapter, we investigate the potential of FD equalization in coherent optical communication systems using MC OQAM with a high density of subcarriers. We first ob-
7.10. Conclusion
93
Table 7.3: Csymb comparison between SC

4000 km and fs = 30 GHz. NCMA = 15.
Equ. stage
CD comp/
Pol Demux
Pol Demux update
CPE
Sum
and hybrid modulation. Example for Lf iber =

SC
Hybrid OQAM
131
98
126
16
24
2
273
134
serve that the MC OQAM modulation and demodulation offered a framework for lowcomplexity FD equalization. We apply this idea to the equalization to CD and conclude
that CD can be compensated successfully at the same computational complexity as in SC
systems. We then discuss the equalization of phase offsets and conclude that similarly
to other MC modulations, OQAM is very sensitive to phase offset. Using only a one-tap
FD equalizer does not yield an acceptable equalization even for moderate phase noise
linewidths. Generalizing to inter-subcarrier convolutional equalization results in performance enhancements but the quality of equalization is still limited by the temporal mixing
inherent in the filterbank. We further generalize the simulation taking into account polarization effects. Since the parameters of PMD are generally unkown at the receiver, we
propose a channel estimation strategy based on pilot subcarriers. We conclude that PMD
can be equalized without performance penalties or inacceptable overhead using a Pair
of Pilots channel estimation technique. In conclusion, from the main linear impairments
present in optical fibers, only phase offsets represent a real obstacle to the use of MC
OQAM modulation.
In order to enhance the performance of the phase noise estimation and equalization,
we propose a hybrid SC and MC modulation scheme. The SC part of the spectrum is
used as a pilot to estimate the phase using standard algorithms. Because the SC phase
estimation is blind, the spectral efficiency is not affected by the use of the pilot. We study
the feasibility and computational complexity of the hybrid modulation. We conclude
that in terms of the number of operations per equalized symbol, at least 30 % reduction
can be achieved by using MC OQAM modulations instead of SC. The proposed hybrid
modulation conserves the advantage of SC phase mitigation while offering the flexibility
of MC modulations.
Chapter 8
Publications
1. Jessica Fickers, Yann Lebrun and Franois Horlin. "Resource Allocation for
OFDMA-Based Wireless Mesh Networks with Subcarrier Exchange", IEEE Benelux
Symposium on Information Theory, 2009, Eindhoven.
2. Jessica Fickers, Franois Quitin, Philippe Emplit, and Franois Horlin. "Frequency
Domain Equalization for Dispersive Birefringent Nonlinear Optical Fibers", IEEE
Symposium on Information Theory, 2010, Eindhoven.
3. Franois Horlin, Franois Quitin, Jessica Fickers, and Philippe Emplit. "Polarization Division Multiplexing for SC-FDE Communications over Dispersive Optical
Fibers", IEEE International Conference on Communications, Cape Town, May 2010.
4. Jessica Fickers, Amirhossein Ghaziseidi, Massimiliano Salsi, Gabriel Charlet,
Franois Horlin and Philippe Emplit. "Design Rules for Pulse Shaping in PDMQPSK and PDM-16QAM Nyquist-WDM Coherent Optical Transmission Systems".
European Conference on Optical Communications, 2012, Amsterdam.
Franois Horlin and Philippe Emplit.
"Decision-Feedback Equalization of
Bandwidth-Constrained N-WDM Coherent Optical Communication Systems",
Journal of Lightwave Technology, 31 (10), 2013.
6. Franois Horlin, Jessica Fickers, Philippe Emplit, Andr Bourdoux and Jrme
Louveaux. "Dual-Polarization OFDM-OQAM for Communications over Optical
Fibers with Coherent Detection", Optics Express, 21 (5), 2013.
7. Franois Horlin, Jessica Fickers, Thibault Deleu and Jrme Louveaux.
"Interference-free SDMA for FBMC-OQAM", Eurasip Journal on Advances in Signal Processing, 46 (10), 2013.
Franois Horlin and Philippe Emplit." Decision Feedback Equalization for
Bandwidth-Constrained 28 Gbaud Nyquist-WDM PDM-8QAM over 37.5 GHz
Grid", European Conference on Optical Communications, 2013, London.
96
Chapter 8. Publications
9. Francesco Vacondio, Oriol Bertran-Pardo, Yvan Pointurier, Jessica Fickers,
Amirhossein Ghazisaeidi, Guillem de Valicourt, Jean-Christophe Antona, Pierre
Chanclou and Sebastien Bigo. "Flexible TDMA access optical networks enabled
by data-aided software defined coherent transponders", European Conference on
Optical Communications, 2013, London.
10. Jessica Fickers, Amirhossein Ghaziseidi, Massimiliano Salsi, Franois Horlin,

Philippe Emplit and Gabriel Charlet. "Multicarrier Offset-QAM Modulations for
Coherent Optical Communication Systems", Optical Fiber Conference, 2014, San
Francisco (CA).
Franois Horlin and Philippe Emplit. "Multicarrier Offset-QAM for Long-Haul
Coherent Optical Communications", Journal of Lightwave Technology (review completed and resubmitted).
Chapter 9
Conclusion
Optical fiber telecommunications form the backbone of the global data network. Since
the 1970ies, the transmission distances were progressively increased thanks to several
major technological breakthroughs and today, long-haul optical links span oceans and
continents.
With the speed of electronics progressively catching up with the data rates of optical
signals, the last decade saw the advent of coherent detection combined with digital signal
processing in optical communications. Thanks to coherent detection, all characteristics of
the optical field (amplitude, phase, polarization) can be detected and the noise sensitivity
is optimized. Digital signal processing enables the generation of high-order modulation
formats making full use of coherent detection and the compensation of channel impairments.
The next generation of optical networks will fulfill even higher requirements in terms
of capacity, spectral efficiency and system flexibility. Moreover, as there is an increasing
awareness of the need to decrease the power consumption of optical networks, the computational complexity of digital signal processing becomes a research focus. In this thesis
we explored the potential of digital signal processing to achieve these goals.
In the first part of this thesis, we focused on digital signal processing solutions to either
enhance the performance or spectral efficiency of standard single carrier modulations. In
chapter 4, transmitter digital signal processing was used to optimize the pulse shaping in
order to reduce the optical channel spacing. We showed that in the presence of tranceiver
impairments, there is a trade-off between spectral efficiency and performance. We propose
design rules for the digital transmitter and receiver pulse shaping based on this trade-off.
In chapter 5 we aimed at further reduction of the channel spacing in order to enhance the
spectral efficiency. In this goal, we reduced the individual channel bandwidth below the
symbol rate. The performance penalties arising from this method were mitigated using
a low-complexity decision feedback equalizer. We showed that the performance penalties
from heavy filtering can be significantly reduced using this technique.
In the second part of this thesis, we concentrated on the flexibility of the optical
communication architecture by subdividing the optical channel into electrical subcarriers.
We used Offset-QAM modulation to ensure orthogonal subcarriers at maximum spectral
efficiency. In chapter 6, we proposed a new digital signal processing architecture in order
98
Chapter 9. Conclusion
to cope with the phase offset sensitivity of multicarrier modulations. This enabled us to
experimentally demonstrate the feasibility of multicarrier Offset-QAM with up to seven
subcarriers in a long-haul point-to-point link. In chapter 7 we subdivided the electrical
bandwidth into a large number of Offset-QAM subcarriers and proposed a hybrid single
carrier/multicarrier modulation scheme in order to successfully mitigate phase offsets
while using low-complexity frequency domain equalization technique and benefiting from
the flexibility of multicarrier modulations.
At the end of this project, it is time to discuss future work in the continuity of our
research as well as to point out avenues of research that we did not fully explore.
Looking at the succession of chapters in this thesis, it is clear that several intermediate
steps were necessary in order to successfully transmit the multicarrier signals. In particular, a more focused research on channel estimation techniques based on training symbols
is interesting to pursue. For example, in the context of this thesis, we developed an equalizer based on the minimum mean square error criterion that has been successfully used
for channel estimation in single carrier access networks [129]. In this case, the modulation
format was switched in a time-division multiple access scheme. It was therefore essential
for the equalizer to converge rapidly which is impossible for a blind equalizer. In future
software-defined networks, it is likely that there will be many cases where the capability
to rapidly converge will be important, and equalizers based on channel estimation might
be the way to go.
In the context of multicarrier modulations, the most straightforward next step would
be the experimental validation of the equalization techniques proposed in chapter 7. In
particular, the resilience with respect to timing offsets should be studied. In addition, the
problem of large frequency offsets between the transmitter and receiver lasers has been
left out in chapter 7. Frequency offsets that are large in comparison with the bandwidth
of the single carrier part of the signal will represent a challenge because the receiver filter
will be strongly offset with respect to the pilot. In a more general context, it would be
very interesting to explore the potential of multicarrier modulations in a software-defined
network. Could the subcarriers be realistically assigned to different users? Could the
frequency selectivity of the optical channel justify non-homogeneous modulation distribution formats or coding over the subcarriers? This might be the case if the edge carriers
of an optical channel are heavily filtered during transmission because of the presence of
ROADMs as seen in chapter 5.
Finally, it is clear that the complexity study presented in chapter 7 is a very simplistic
one. Future studies should take into account the chip requirements in terms of power,
chip size and latency.
Chapter 10
Appendix: subcarrier orthogonality in
MC-OQAM systems
The object of this appendix is to derive the subcarrier orthogonality in MC-OQAM communication systems. Fig. 10.1 presents the relevant branches of the (m + 1)-th and
(m 1)-th subcarrier channels that may leak signal to the output of the m-th subcarrier
channel.
First, we inspect the central channel m. In Fig. 10.1 it is easy to see that all complex
exponentials cancel out in this case. By definition of the RRC filter: u[n] u[n] =
[n]. Since the pulse shaping filters are real, the Re and Im cancel out intra-subcarrier
I
I
R
[n] = sR
interference. As a consequence zm
m [n] and zm [n] = sm [n] , in the absence of
neighboring subchannels.
As pointed out in chapter 6, interference is only possible between neighboring subcarriers. According to Fig. 10.1, four interference terms are possible from channel m + 1
towards m:
R
sR
m+1 zm . By inspection of Fig. 10.1, the impulse response between these sequences is given by:
X
n
RR
g+1
[p] = Re[
u[n]ej/2 ej2 M u[pM n]]
(10.1)
n
= Im[
u[n]ej2 M u[pM n]]
(10.2)
= Im[
0
X
u[n0 +
M p j2(n0 + M p )/M M p
2
u[
]e
n0 ]]
2
2
0
X
M p j2(n0 )/M M p
= Im[
u[n0 +
]e
u[
n0 ]]
2
2
n
0
X
u[n0 +
=0 p
Mp
Mp
] sin(j2(n0 )/M )u[
n0 ]
2
2
(10.3)
(10.4)
(10.5)
(10.6)
Because the sum of an odd function over all integrals is equal to zero. The last lines
are obtained by the change of variables n n0 = n pM/2.
100 Chapter 10. Appendix: subcarrier orthogonality in MC-OQAM systems

sR
m1 [n]
sIm1 [n]
sR
m [n]
sIm [n]
sR
m+1 [n]
sIm+1 [n]
ej(m1) 2
2
.
.
.
M/2
z
u[n]
j=
.
.
.
ej2
(m1)n
M
M/2 u[n]
z1
ejm 2

ej2
mn
M
ej2
z1
mn
M
M/2 u[n]
ej2
.
.
.
ej(m+1) 2
u [n] M/2
u [n] M/2
(m1)n
M
ejm 2
j2
.
.
.
ej(m1) 2
(m+1)n
M
j(m+1)
2
ej2
u [n] M/2
(m+1)n
M
z1
R
zm1
[n]
I
zm1
[n]
R
zm
[n]
I
zm
[n]
R
zm+1
[n]
I
zm+1
[n]
Figure 10.1: MC-OQAM system, channels m 1, m and m + 1 are represented.
I
sR
m+1 zm . By inspection of Fig. 10.1, the impulse response between these sequences is given by:
RI
g+1
[p] = Im[
u[n]ej/2 ej2 M u[M/2 + pM n]]
(10.7)
= Re[
u[n]ej2n/M u[
= Re[
0
X
= Im[
n
0
X
n
M
+ pM n]]
2
(10.8)
u[n0 +
M
M p j2(n0 + M + M p )/M M
Mp
4
2
+
]e
u[ +
n0 ]]
4
2
4
2
u[n0 +
M
M p j2 n0 M
Mp
+
]e M u[ +
n0 ]]
4
2
4
2
(10.10)
=0 p
The last lines are obtained by the change of variables n n0 = n
(10.9)
(10.11)
M
4
Mp
.
2
R
sIm+1 zm
. By inspection of Fig. 10.1, the impulse response between these se-
101
quences is given by:
X
n
M
IR
g+1
[p] = Re[
u[n]ej/2 ej/2 ej2 M u[ + pM n]]
2
n
X
n
M
= Rre[
u[n]ej2 M u[ + pM n]]
2
n
(10.12)
(10.13)
0
X
M
M p j2(n0 + M p )/M M p M
2
= Re[
u[n0 +
+
]e
u[
+
n0 ]]
4
2
2
4
n
(10.14)
0
X
M p M j2 n0 M p M
= Im[
u[n0 +
+ ]e M u[
+
n0 ]]
2
4
2
4
n
(10.15)
=0 p
(10.16)
M
4
Mp
.
2
I
. By inspection of Fig. 10.1, the impulse response between these se sIm+1 zm
quences is given by:
X
n
II
g+1
[p] = Im[
u[n]ej/2 ej/2 ej2 M u[pM n]]
(10.17)
n
= Im[
= Im[
X
n
0
X
u[n]ej2 M u[
= Im[
(10.18)
u[n0 +
M p j2(n0 + M p )/M M p
2
]e
n0 ]]
u[
2
2
(10.19)
u[n0 +
M p j2 n0 M p
]e M u[
n0 ]]
2
2
(10.20)
n
0
X
M
+ pM n]]
2
=0 p
(10.21)
Mp
.
2
Using the same technique, the four interference terms leaking from the (m 1)-th channel can be proven to be equal to zero as well. In conclusion, the OQAM modulation/demodulation defined in chapter 6 ensures interference-free communication between
overlapping subcarriers.
Bibliography
[1] K.C. Kao and G.A. Hockham. Dielectric-fibre surface waveguides for optical frequencies. Proceedings of the Institution of Electrical Engineers, 113(7):11511158,
1966.
[2] G. P. Agrawal. Fiber-Optic Communication Systems. John Wiley and Sons, New
York, 2002.
[3] J.A. Buck. Fundamentals of optical fibers, volume 1. Wiley, 2004.
[4] I.P. Kaminow and T.L. Koch. Optical fiber telecommunications III, volume 1. Academic Press, 1997.
[5] O. Bertran-Pardo. On coherent detection for optical transmissions at 40 Gbit/s and
100 Gbit/s. PhD thesis, Tlcom ParisTech, 2010.
[6] M. Born and E. Wolf. Principles of optics. Pergamon Press, 1964.
[7] S.G. Evangelides, L.F. Mollenauer, J.P. Gordon, and N.S. Bergano. Polarization
multiplexing with solitons. Journal of Lightwave Technology, 10(1):2835, 1992.
[8] G. P. Agrawal. Nonlinear Fiber Optics. Academic Press, New York, 2001.
[9] Corning. Corning single mode fiber datasheet. In PI1036, 2002.
[10] M. Brodsky, N.J. Frigo, M. Boroditsky, and M. Tur. Polarization mode dispersion
of installed fibers. Journal of Lightwave Technology, 269(12):45844599, 2006.
[11] M. Wuilpart, C. Crunelle, and P. Mgret. High dynamic polarization-OTDR for
the PMD mapping in optical fiber links. Optics Communications, 24(12):315321,
2007.
[12] DL Peterson Jr, P.J. Leo, and K.B. Rochford. Field measurements of state of polarization and PMD from a tier-1 carrier. In Optical Fiber Communication Conference,
2004.
[13] N. Sambo, M. Secondini, N. Andriolli, M.N. Petersen, and P. Castoldi. Accounting for PMD temporal correlation during lightpath set up in transparent optical
networks. Journal of Lightwave Technology, 28(22):33073313, 2010.
[14] R. W. Boyd. Nonlinear Optics. Academic Press, San Diego, California, 2006.
104
Bibliography
[15] G. Gao, X. Chen, and W. Shieh. Influence of PMD on fiber nonlinearity compensation using digital back propagation. Optics Express, 20(13):1440614418, 2012.
[16] G. Di Domenico, S. Schilt, and P. Thomann. Simple approach to the relation
between laser frequency noise and laser line shape. Applied optics, 49(25):4801
4807, 2010.
[17] S. Bartalini, S. Borri, I. Galli, G. Giusfredi, D. Mazzotti, T. Edamura, N. Akikusa,
M. Yamanishi, and P. De Natale. Measuring frequency noise and intrinsic linewidth
of a room-temperature DFB quantum cascade laser. Optics Express, 19(19):17996
18003, 2011.
[18] T. Heil, I. Fischer, W. Elsasser, J. Mulet, and C.R. Mirasso. Statistical properties
of low-frequency fluctuations during single-mode operation in distributed-feedback
lasers: experiments and modeling. Optics Letters, 24(18):12751277, 1999.
[19] L. Li, Z. Tao, S. Oda, T. Hoshida, and J.C. Rasmussen. Wide-range, accurate and
simple digital frequency offset compensator for optical coherent receivers. In Optical
Fiber Communication Conference, 2008.
[20] F. Horlin, F. Quitin, J. Fickers, and P. Emplit. Polarization division multiplexing
for SC-FDE communications over dispersive optical fibers. In EEE International
Conference on Communications, 2010.
[21] O. Bertran-Pardo, J. Renaudier, G. Charlet, P. Tran, H. Mardoyan, M. Salsi, and
S. Bigo. Impact of nonlinear impairments on the tolerance to PMD of 100Gb/s
PDM-QPSK data processed in a coherent receiver. In European Conference on
Optical Communications, volume O, pages 12. IEEE, 2008.
[22] C. Behrens, R. Killey, C. Ming, and P. Bayvel. Nonlinear transmission performance
of higher-order modulation formats. Photonics Technology Letters, 23(6):377379,
2011.
[23] P. Boffi, P. Martelli, M. Cirigliano, M. Magarini, M. Martinelli, M. Bertolini,
M. Pepe, D. Mola, and G. Gavioli. PDM-16QAM transmission performance over
uncompensated fiber links. Optics Express, 19(22):2189821903, 2011.
[24] I. Fatadin, D. Ives, and S.J. Savory. Blind equalization and carrier phase recovery in
a 16-QAM optical coherent system. Journal of Lightwave Technology, 27(15):3042
3049, 2009.
[25] P. Poggiolini, G. Bosco, A. Carena, V. Curri, and F. Forghieri. Performance evaluation of coherent WDM PS-QPSK (HEXA) accounting for non-linear fiber propagation effects. Optics Express, 18(11):1136011371, 2010.
[26] M. Karlsson and E. Agrell. Which is the most power-efficient modulation format in
optical links? Optics Express, 17(13):1081410819, 2009.
[27] E. Desurvire, D. Bayart, B. Desthieux, and S. Bigo. Erbium-Doped Fiber Amplifiers.
Device and System Developements. Wiley Interscience, 2002.
Bibliography
105
[28] E. Ip, A.P.T. Lau, D.J.F. Barros, and J.M. Kahn. Coherent detection in optical
fiber systems. Optics Express, 16(2):753791, 2008.
[29] J. Proakis. Digital Communications. McGraw-Hill, New York, 2001.
[30] Edouard Grellier and Alberto Bononi. Quality parameter for coherent transmissions
with gaussian-distributed nonlinear noise. Optics express, 19(13):1278112788, 2011.
[31] Pierluigi Poggiolini, Andrea Carena, Vittorio Curri, Gabriella Bosco, and Fabrizio
Forghieri. Analytical modeling of nonlinear propagation in uncompensated optical
transmission links. Photonics Technology Letters, IEEE, 23(11):742744, 2011.
[32] G. Charlet. Coherent detection associated with digital signal processing for fiber
optics communication. Elsevier, Comptes Rendus de Physique, 9:10121030, 2008.
[33] J. Renaudier, G. Charlet, M. Salsi, O.B. Pardo, H. Mardoyan, P. Tran, and S. Bigo.
Linear fiber impairments mitigation of 40-gbit/s polarization-multiplexed QPSK by
digital processing in a coherent receiver. Journal of Lightwave Technology, 26(1):36
42, 2008.
[34] E. M. Ip and J.M. Kahn. Fiber impairment compensation using coherent detection
and digital signal processing. Journal of Lightwave Technology, 28(4):502 5519,
2010.
[35] L. Zhu and G. Li. Nonlinearity compensation using dispersion-folded digital backward propagation. Optics Express, 20(13):1436214370, 2012.
[36] A. Ghazisaeidi, J. Renaudier, and M. Salsi. System benefits of digital dispersion precompensation for non-dispersion-managed PDM-WDM transmission. In European
Conference on Optical Communications, 2013.
[37] O. Bertran-Pardo, J. Renaudier, G. Charlet, M. Salsi, M. Bertolini, H. Mardoyan,
P. Tran, C. Koebele, and S. Bigo. PDM-QPSK: on the system benefits arising
from temporally interleaving polarization tributaries at 100Gb/s. Optics Express,
17(22):1990219907, 2009.
[38] O Bertran-Pardo and J Renaudier. Digital signal processing in coherent receivers
for ultra long-haul applications. Signal Processing in Photonic Communications,
2010.
[39] Xu Zhang. Digital signal processing for optical coherent communication systems.
PhD thesis, Technical University of Denmark, 2012.
[40] J.C. Geyer, C.R.S. Fludger, T. Duthel, C. Schulien, and B. Schmauss. Efficient
frequency domain chromatic dispersion compensation in a coherent polmux QPSKreceiver. In Optical Fiber Communication Conference, 2010.
[41] C. Dorize, Y. Pointurier, F. Vacondio, J.C. Antona, and S. Bigo. Adaptive power
efficiency for chromatic dispersion compensation. In European Conference on Optical
Communications, 2013.
106
Bibliography
[42] K. Ishihara, T. Kobayashi, R. Kudo, Y. Takatori, A. Sano, E. Yamada, H. Masuda,

and Y. Miyamoto. Frequency-domain equalization for optical transmission systems.
Electronics Letters, 44(14):870 871, 2008.
[43] E. Ip and J.M. Kahn. Digital equalization of chromatic dispersion and polarization
mode dispersion. Journal of Lightwave Technology, 25(8):20332043, 2007.
[44] S.J. Savory. Digital filters for coherent optical receivers. Optics Express, 16(2):804
8817, 2008.
[45] B. Spinnler. Complexity of algorithms for digital coherent receivers. In European
[46] P. Poggiolini, A. Carena, V. Curri, and F. Forghieri. Evaluation of the computational effort for chromatic dispersion compensation in coherent optical PM-OFDM
and PM-QAM systems. Optics Express, 17(3):13851403, 2009.
[47] E. Ip and J.M. Kahn. Compensation of dispersion and nonlinear impairments using
digital backpropagation. Journal of Lightwave Technology, 26(20):34163425, 2008.
[48] C. Antonelli, A. Mecozzi, M. Brodsky, and M. Boroditsky. A simple analytical
model for PMD temporal evolution. In Optical Fiber Communication Conference,
2006.
[49] C. Antonelli and A. Mecozzi. Theoretical characterization and system impact of the
hinge model of PMD. Journal of Lightwave technology, 24(11):40644074, 2006.
[50] B. Hassibi and B.M. Hochwald. How much training is needed in multiple-antenna
wireless links? IEEE Transactions on Information Theory, 49(4):951963, 2003.
[51] A. Barbieri, G. Colavolpe, T. Foggi, E. Forestieri, and G. Prati. OFDM versus singlecarrier transmission for 100 gbps optical communication. Journal of Lightwave
Technology, 28(17):25372551, 2010.
[52] L. Mazet, P. Ciblat, and P. Loubaton. Fractionally spaced blind equalization: CMA
versus second order based methods. In IEEE Workshop on Signal Processing Advances in Wireless Communications, 1999.
[53] C.B. Papadias and A. Paulraj. Decision-feedback equalization and identification of
linear channels using blind algorithms of the bussgang type. In Asilomar Conference
on Signals, Systems and Computers, volume 1, 1995.
[54] A. Leven, N. Kaneda, and Y.K. Chen. A real-tme CMA-based 10 Gb/s polarization demultiplexing coherent receiver implemented in an FPGA. In Optical Fiber
Communication Conference, 2008.
[55] D. Godard. Self-recovering equalization and carrier tracking in two dimensional
data communications system. IEEE Transactions on Communications, 28(11):1867
1875, 1980.
Bibliography
107
[56] S.J. Savory. Digital coherent optical receivers: Algorithms and subsystems. IEEE
Journal Selected Topics in Quantum Electronics, 16(5):11641179, 2010.
[57] A. Neves and C. Panazio. A class of channels resulting in ill-convergence for CMA in
decision feedback equalizers. IEEE Transactions on Signal Processing, 58(11):5736
5743, 2010.
[58] G. Charlet, M. Salsi, P. Tran, M. Bertolini, H. Mardoyan, J. Renaudier, O. BertranPardo, and S. Bigo. 72x100Gb/s transmission over transoceanic distance, using large
effective area fiber, hybrid raman-erbium amplification and coherent detection. In
Optical Fiber Communication Conference, 2009.
[59] J. Renaudier, G. Charlet, O Bertran-Pardo, H. Mardoyan, P. Tran, M. Salsi, and
S. Bigo. Transmission of 100Gb/s coherent PDM-QPSK over 16 times 100 km of
standard fiber with all-erbium amplifiers. Optics Express, 17(7):5112, 2009.
[60] S. Randel, S. Corteselli, and P.J. Winzer. Generation of a digitally shaped 55GBd 64-QAM single-carrier signal using novel high-speed DACs. In Optical Fiber
Communication Conference, 2014.
[61] A. Nespola, S. Straullu, G. Bosco, A. Carena, J. Yanchao, P. Poggiolini, F. Forghieri,
Y. Yamamoto, M. Hirano, and T. Sasaki. 1306-km 20x124. 8-Gb/s PM-64QAM
transmission over PSCF with net SEDP 11,300 (b.km)/s/hz using 1.15 samp/symb
DAC. Optics Express, 22(2):17961805, 2014.
[62] A. Viterbi and A. Viterbi. Nonlinear estimation of psk-modulated carrier phase
with application to burst digital transmission. IEEE Transactions on Information
Theory, 19(4):543551, 1983.
[63] P. Y. Kam. Maximum likelihood carrier phase recovery for linear suppressed-carrier
digital data modulations. IEEE Transactions on Communications, 34(6):522527,
1986.
[64] W. Liu, J. C. Prncipe, and S. Haykin. Kernel Adaptive Filtering. Wiley, Hoboken,
New Jersey, USA, 2010.
[65] I. Djordevic. Coding for Optical Channels. Springer, New York, 2010.
[66] X Shao, PY Kam, and C. Yu. Maximum likelihood sequence detection in laser phase
noise-impaired coherent optical systems. Optics Express, 19(23):2260022606, 2011.
[67] I.B. Djordjevic, M. Arabaci, and L.L. Minkov. Next generation FEC for highcapacity communication in optical transport networks. Journal of Lightwave Technology, 27(16):35183530, 2009.
[68] X. Zhou, L.E. Nelson, P. Magill, R. Isaac, B. Zhu, D.W. Peckham, P.I. Borel, and
K. Carlson. High spectral efficiency 400 gb/s transmission using PDM time-domain
hybrid 32-64 QAM and training-assisted carrier recovery. Journal of Lightwave
Technology, 31(7):9991005, 2013.
108
Bibliography
[69] T. Sugihara, T. Kobayashi, T. Fujimori, and T. Mizuochi. Electronic preequalization technologies using high-speed DAC. In European Conference on Optical
Communications, 2011.
[70] N.C. Beaulieu and A.A. Abu-Dayya. Nyquist shaped QPSK in cochannel interference and fading. In Global Telecommunications Conference, 1993.
[71] M. Salsi, C. Koebele, P. Tran, H. Mardoyan, E. Dutisseuil, J. Renaudier, M. BigotAstruc, L. Provost, S. Richard, and P. Sillard. Transmission of 96* 100Gb/s with
23% super-FEC overhead over 11,680 km, using optical spectral engineering. In
[72] D. Foursa, Y. Cai, J.X. Cai, C. Davidson, O.V. Sinkin, W.T. Anderson, A. Lucero,
A. Pilipetskii, G. Mohs, and N.S. Bergano. Coherent 40 gb/s transmission with
high spectral efficiency over transpacific distance. In Optical Fiber Communication
Conference, 2011.
[73] M.F. Huang, Y.K. Huang, E. Ip, Y. Shao, and T. Wang. WDM transmission of 152Gb/s polarization multiplexed RZ-16QAM signals with 25-GHz channel spacing over
15x80-km of SSMF. In Optical Fiber Communication Conference, 2011.
[74] X. Zhou, L.E. Nelson, P. Magill, B. Zhu, and D.W. Peckham. 8x450-Gb/s, 50-GHzspaced, PDM-32QAM transmission over 400km and one 50GHz-grid ROADM. In
[75] G. Bosco, A. Carena, V. Curri, P Poggiolini, E. Torrengo, and F. Forghieri. Investigation on the robustness of a Nyquist-WDM terabit superchannel to transmitter
and receiver non-idealities. In European Conference on Optical Communications,
2010.
[76] G. Bosco, A. Carena, V. Curri, P. Poggiolini, and F. Forghieri. Performance limits of
nyquist-WDM and CO-OFDM in high-speed PM-QPSK systems. IEEE Photonics
Technology Letters, 22(15):11291131, 2010.
[77] J. Fickers, A. Ghazisaeidi, M. Salsi, G. Charlet, F. Horlin, P. Emplit, and S. Bigo.
Design rules for pulse shaping in PDM-QPSK and PDM-16QAM Nyquist-WDM
coherent optical transmission systems. In European Conference on Optical Communications. Optical Society of America, 2012.
[78] R. Schmogrow, M. Winter, M. Meyer, D. Hillerkuss, S. Wolf, B. Baeuerle, A. Ludwig, B. Nebendahl, S. Ben Ezra, J. Meyer, M. Dreschmann, M. Huebner, J. Becker,
C. Koos, W. Freude, and J. Leuthold. Real-time Nyquist pulse generation beyond
100 gbit/s and its relation to OFDM. Optics Express, 20(1):317337, 2011.
[79] A. Khilo, S.J. Spector, M.E. Grein, A.H. Nejadmalayeri, C.W. Holzwarth, M.Y.
Sander, M.S. Dahlem, M.Y. Peng, M.W. Geis, and N.A. DiLello. Photonic ADC:
overcoming the bottleneck of electronic jitter. Optics Express, 20(4):44544469,
2012.
Bibliography
109
[80] S.J. Savory, A.D. Steward, S. Wood, G. Gavioli, M.G. Taylor, R.I. Kiley, and
P. Bayvel. Digital equalisation of 40gbit/s per wavelength transmission over 2480km
of standard fibre without optical dispersion compensation ECOC 2006. In European
[81] J. Wang, C. Xie, and Z. Pan. Generation of spectrally efficient nyquist-WDM
QPSK signals using DSP techniques at transmitter. In Optical Fiber Communication
Conference, 2012.
[82] J.-X. Cai, Y. Cai, C.R. Davidson, D.G. Foursa, A. Lucero, O. Sinkin, W. Patterson,
A. Pilipetskii, G. Mohns, and N.S. Bergano. Transmission of 96 times 100G prefiltered PDM-RZ-QPSK channels with 300% spectral efficiency over 10,608 km and
400% spectral efficiency over 4,368 km. In Optical Fiber Communication Conference,
2010.
[83] J. Wang, X. He, K. Gao, and L. Myers. 8 times 400-Gbit/s PDM-QPSK with 100
GHz channel spacing over 2000km transmission using MAP detection. In Conference
on Lasers and Electro-Optics, 2011.
[84] J.-X. Cai, Y. Cai, C.R. Davidson, A. Lucero, H. Zhang, D.G. Foursa, O.V. Sinkin,
W.W. Patterson, A. Pilipetskii, G. Mohs, and N.S. Bergano. 20 tbit/s capacity
transmission over 6,890 km. In Optical Fiber Communication Conference, 2011.
[85] J.X. Cai. 100G transoceanic length transmission with high spectral efficiency using
bandwidth constrained PDM-QPSK. In Optical Fiber Communication Conference,
2011.
[86] M.S. Alfiad, D. Van den Borne, M. Kuschnerov, B. Spinnler, T. Wuth, A. Napoli,
S.L. Jansen, and H. de Waardt. FFE, DFE and MLSE equalizers in phase modulated
transmission systems. In LEOS Annual Meeting, 2009.
[87] Y. Mori, K. Igarashi, K. Katoh, and K. Kikuchi. Decision-feedback carrier-phase
estimation for digital coherent optical receivers. In Australian Conference on Optical
Fiber Communications, 2008.
[88] X. Zhang, X. Pang, L. Deng, D. Zibar, I.T. Monroy, and R. Younce. High phase
noise tolerant pilot-tone-aided DP-QPSK optical communication systems. Optics
Express, 20(18):1999019995, 2012.
[89] T. Zami, B. Lavigne, and E. Balmefrezol. Crosstalk analysis applied to wavelength
selective switches. In Optical Fiber Communication Conference, 2011.
[90] T. Zami, P. Henri, L. Lorcy, and C. Simonneau. Impact of the optical routing on
the transmission in transparent networks. European Conference on Optical Communications, 2009.
[91] T. Zami, B. Lavigne, E. Balmefrezol, M. Lefrancois, and H. Mardoyan. Comparative
study of crosstalk in 50 Ghz-spaced wavelength selective switch for various modulation formats at 43 Gbit/s. In European Conference on Optical Communications,
2008.
110
Bibliography
[92] A.P. Liavas. On the robustness of the finite-length MMSE-DFE with respect to
channel and second-order statistics estimation errors. IEEE Transactions on Signal
Processing, 50(11):28662874, 2002.
[93] N. Al-Dhahir and A.H. Sayed. The finite-length multi-input multi-output MMSEDFE. IEEE Transactions on Signal Processing, 48(10):29212936, 2000.
[94] M. Miyoshi and Y. Kaneda. Inverse filtering of room acoustics. IEEE Transactions
on Acoustics, Speech and Signal Processing, 36(2):145152, 1988.
[95] R. Arablouei, K. Doanay, and S. Perreau. Partial-update adaptive decisionfeedback equalization. In 19th European Signal Processing Conference (EUSIPICO),
2011.
[96] G. Strang. Introduction to linear algebra. Wellesley Cambridge Pr, 2003.
[97] A. Ghazisaeidi. Impact of tight optical filtering on the performance of 28 Gbaud
Nyquist-WDM PDM-8QAM over 37.5 GHz grid. In Optical Fiber Communication
Conference, 2013.
[98] R. Paschotta, A. Schlatter, S.C. Zeller, H.R. Telle, and U. Keller. Optical phase
noise and carrier-envelope offset noise of mode-locked lasers. Applied Physics B,
82(2):265273, 2006.
[99] S. Randel, A. Sierra, X. Liu, and P. J. Winzer. Study of multicarrier offset-QAM
for spectrally efficient coherent optical communications. In European Conference on
Optical Communications, 2011.
[100] S. Chandrasekhar, X. Liu, B. Zhu, and D. W. Peckham. Transmission of a 1.2-tb/s
24-carrier no-guard-interval coherent OFDM superchannel over 7200-km of ultralarge-area fiber. In European Conference on Optical Communications, 2009.
[101] N. Dupuis, C.R. Doerr, L. Zhang, L. Chen, N.J. Sauer, L. Buhl, and D. Ahn.
Inp-based comb generator for optical OFDM. In Optical Fiber Communication
Conference, 2011.
[102] K. Hueske and J. Gotze. Ov-OFDM: A reduced PAPR and cyclic prefix free multicarrier transmission system. In International Symposium on Wireless Communication Systems, 2009.
[103] S. Kilmurray, T. Fehenberger, P. Bayvel, and R. I. Killey. Comparison of the nonlinear transmission performance of quasi-nyquist WDM and reduced guard interval
OFDM. Optics Express, 20(4):41984205, 2012.
[104] B.R. Saltzberg. Performance of an efficient parallel data transmission system. IEEE
Transations on Communications, 15:805811, 1967.
[105] F. Horlin, J. Fickers, P. Emplit, A. Bourdoux, and J. Louveaux. Dual-polarization
OFDM-OQAM for communications over optical fibers with coherent detection. Optics Express, 21(5):64096421, 2013.
Bibliography
111
[106] S. Randel, S. Corteselli, A. Chandrasekhar, S. and.Sierra, X. Liu, P.J. Winzer,

T. Ellermeyer, J. Lutz, and R. Schmid. Generation of 224-gb/s multicarrier offsetQAM using a real-time transmitter. In Optical Fiber Communication Conference,
2012.
[107] T. Ihalainen, T.H. Stitz, M. Rinne, and M. Renfors. Channel equalization in filter bank based multicarrier modulation for wireless communications. EURASIP
Journal on Applied Signal Processing, 2007(1):140140, 2007.
[108] Z. Li, T. Jiang, H. Li, X. Zhang, C. Li, Ch. Li, R. Hu, M. Luo, X. Zhang,
and X. Xiao. Experimental demonstration of 110-Gb/s unsynchronized bandmultiplexed superchannel coherent optical OFDM/OQAM system. Optics Express,
21(19):2192421931, 2013.
[109] F. Horlin, J. Fickers, T. Deleu, and J. Louveaux. Interference-free SDMA for FBMCOQAM. EURASIP Journal on Advances in Signal Processing, 1, 2013.
[110] B. Farhang-Boroujeny and C. Yuen. Cosine modulated and offset QAM filter bank
multicarrier techniques: A continuous-time prospect. EURASIP Journal on Advances in Signal Processing, 2010:1, 2009.
[111] W. Bocquet, K. Hayashi, and H. Sakai. On the block size design for single
carrier overlap frequency domain equalization. Journal of Lightwave Technology,
28(17):3721 33728, 2010.
[112] I. Djordevic and W. Shieh. OFDM for Optical Communications. Elsevier, New
York, 2010.
[113] N. Benvenuto and S. Tomasin. On the comparison between OFDM and single
carrier modulation with a DFE using a frequency-domain feedforward filter. IEEE
Transactions on Communications, 50(6):947955, 2002.
[114] R. Elschner, F. Frey, C. Meuer, and J.K. Fischer. Experimental demonstration
of a format-flexible single carrier coherent receiver using data-aided digital signal
processing. In European Conference on Optical Communications, 2012.
[115] K. Shi and E. Serpedin. Coarse frame and carrier synchronization of OFDM systems:
a new metric and comparison. IEEE Transactions on Wireless Communications,
3(4):12711284, 2004.
[116] J Fickers, A Ghazisaeidi, M Salsi, and F Horlin. Multicarrier offset-QAM modulations for coherent optical communication systems. In Optical Fiber Communication
Conference, 2014.
[117] T. Ihalainen, A. Ikhlef, J. Louveaux, and M. Renfors. Channel equalization for
multi-antenna FBMC-OQAM receivers. IEEE Transactions on Vehicular Technology, 60(5):20702085, 2011.
112
Bibliography
[118] A. Ikhlef and J. Louveaux. Per subchannel equalization for MIMO FBMC-OQAM
systems. In IEEE Pacific Rim Conference on Communications, Computers and
Signal Processing, pages 559564. IEEE, 2009.
[119] A. Ikhlef and J. Louveaux. An enhanced MMSE per subchannel equalizer for highly
frequency selective channels for FBMC-OQAM systems. In IEEE 10th Workshop
on Signal Processing Advances in Wireless Communications, 2009.
[120] P. Siohan, C. Siclet, and N. Lacaille. Analysis and design of OFDM/OQAM systems
based on filterbank theory. IEEE Transactions on Signal Processing, 50(5):1170
1183, 2002.
[121] M.G. Bellanger. Specification and design of a prototype filter for filter bank based
multicarrier transmission. In Conference on Acoustics, Speech, and Signal Processings, 2001.
[122] S.G. Johnson and M. Frigo. A modified split-radix FFT with fewer arithmetic
operations. IEEE Transactions on Signal Processing, 55:111139, 2007.
[123] J. Leibrich and W. Rosenkranz. Frequency domain equalization with minimum complexity in coherent optical transmission systems. In Optical Fiber Communication
Conference, 2010.
[124] D.J.F. Barros and J.M. Kahn. Optimized dispersion compensation using orthogonal
frequency-division multiplexing. Journal on Lightwave Technology, 26(16):2889
2898, 2008.
[125] C. Ll, P. Siohan, R. Legouable, and J-.P. Javaudin. Preamble-based channel
estimation techniques for OFDM/OQAM over the powerline. In IEEE International
Symposium on Power Line Communications and Its Applications, 2007.
[126] T. Kobayashi, A. Sano, A. Matsuura, Y. Miyamoto, and K. Ishihara. Nonlinear tolerant long-haul WDM transmission over 1200km using 538 Gb/s/ch PDM-64QAM
SC-FDM signals with pilot tone. Optical Fiber Communication Conference, 2012.
[127] B.Y. Liang and A.J. Lowery. Pilot-based XPM nonlinearity compensator for COOFDM systems. Optics Express, 19(26):B862B867, 2011.
[128] M.H. Morsy-Osman, L.R. Chen, and D.V. Plant. Joint mitigation of laser phase
noise and fiber nonlinearity using pilot-aided transmission for single-carrier systems.
In European Conference on Optical Communications, 2011.
[129] F. Vacondio, O. Bertran-Pardo, Y. Poinurier, J. Fickers, A. Ghazisaeidi, and
G. De Valincourt. Flexible TDMA access optical networks enabled by burst-mode
software defined coherent transponders. In European Conference on Optical Communications, 2013.

These

Enviado por

Dados do documento

Descrição original:

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

These

Enviado por

Direitos autorais:

Formatos disponíveis

Universit libre de Bruxelles

Ecole Polytechnique de Bruxelles

Modulation formats and digital signal processing

3 Digital signal processing in coherent optical communication systems

Symbol identification and FEC . . . . . . . . . . . . . . . . . . . . . . . . 29

4 Design rules for root-raised-cosine pulse shaping

5 Decision feedback equalizers for bandwidth-constrained communications

6 Multicarrier Offset-QAM modulations

7 Low-complexity equalization for OQAM modulations

10 Appendix: subcarrier orthogonality in MC-OQAM systems

Analog to digital converter

Minimum mean square error

The optical fiber as a transmission medium

Transmission impairments in optical fibers . . . . . . . . . . . . .

Polarization mode dispersion . . . . . . . . . . . . . . . . . . . . . . 11

The fiber-optic communication system . . . . . . . . . . . . . . . .

Digital modulation formats . . . . . . . . . . . . . . . . . . . . . . . 14

Trade-off between noise and nonlinear distortions . . . . . . . . . . . 18

Mathematical description of the fiber-optic communication system 19

Nonlinear fiber simulation and linear transfer function . . . . . . . . 19

Mathematical models of the optical communication chain . . . . . . 20

Chapter 2. Fiber-optic communications

2.3. The optical fiber as a transmission medium

Short-distance networks. On the scale of cities, metropolitan networks are designed

The optical fiber as a transmission medium

Figure 2.1: Silica fiber

Chapter 2. Fiber-optic communications

Figure 2.2: Energy distribution of the fundamental mode of propagation in singlemode

2.4. Transmission impairments in optical fibers

Figure 2.3: Attenuation of singlemode fiber as a function of wavelength and associated

Transmission impairments in optical fibers

Chapter 2. Fiber-optic communications

Ignoring all right-hand terms in Eq. (2.3) but attenuation, we have:

In the frequency domain, Eq (2.5) becomes:

We see that the equation is easily solved in the frequency domain:

As a conclusion, chromatic dispersion manifests as the accumulation of a quadratic phase

2.4. Transmission impairments in optical fibers

As an example, the dispersion factor is given as a function of wavelength for standard

Polarization mode dispersion

Figure 2.4: Fiber model of randomly birefringent sections

Chapter 2. Fiber-optic communications

The fiber-optic communication system

2.5. The fiber-optic communication system

Figure 2.5: Architecture of the fiber-optical communication link

In this equation f0 is frequency at the maximum of the lorentzian line.

Chapter 2. Fiber-optic communications

Digital modulation formats

In the previous paragraph, we described how different degrees of freedom - amplitude,

2.5. The fiber-optic communication system

Figure 2.7: Modulation formats for optical communications

Chapter 2. Fiber-optic communications

Nowadays, long optical transmission distances are achieved by periodically incorporating

2.5. The fiber-optic communication system

Figure 2.8: The coherent mixer

= |ACohM ix | + |Aol | + 2|ACohM ix ||Aol | cos((c ol )t + (t))

I2 (t) = IP D2 (t) IP D4 (t) = 4|ACohM ix ||Aol | sin((c ol )t + (t))

Chapter 2. Fiber-optic communications

Figure 2.9: Coherent detection for polarization-multiplexed systems

Trade-off between noise and nonlinear distortions

Link input power per channel [dBm]

Figure 2.10: Performance as a function of the link input power

2.6. Mathematical description of the fiber-optic communication system

Mathematical description of the fiber-optic communication system