Final Project OFDM

ORTHOGONAL FREQUENCY-DIVISION MULTIPLEXING
FOR OPTICAL COMMUNICATIONS

A DISSERTATION
SUBMITTED TO THE DEPARTMENT OF ELECTRICAL
ENGINEERING
AND THE COMMITTEE ON GRADUATE STUDIES
OF STANFORD UNIVERSITY
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR
THE DEGREE OF
DOCTOR OF PHILOSOPHY

Daniel Jose Fernandes Barros
September 2011

http://creativecommons.org/licenses/by-nc/3.0/us/

This dissertation is online at: http://purl.stanford.edu/yz748tf7178

2011 by Daniel Jose Fernandes Barros. All Rights Reserved.
Re-distributed by Stanford University under license with the author.
This work is licensed under a Creative Commons Attribution-
Noncommercial 3.0 United States License.
ii
I certify that I have read this dissertation and that, in my opinion, it is fully adequate
in scope and quality as a dissertation for the degree of Doctor of Philosophy.
Joseph Kahn, Primary Adviser
John Gill, III
Bernard Widrow
Approved for the Stanford University Committee on Graduate Studies.
Patricia J. Gumport, Vice Provost Graduate Education
This signature page was generated electronically upon submission of this dissertation in
electronic format. An original signed hard copy of the signature page is on file in
University Archives.
iii
Abstract
iv
ABSTRACT
The drive towards higher spectral efficiency and maximum power efficiency in
optical systems has generated renewed interest in the optimization of optical
transceivers. In this work, we study the different optical applications: Wide Area
Networks (WANs), Metropolitan Area Networks (MANs), Local Area Networks
(LANs) and Personal Area Networks (PANs).
In WANs or long-haul systems, orthogonal frequency-division multiplexing
(OFDM) can compensate for linear distortions, such as group-velocity dispersion
(GVD) and polarization-mode dispersion (PMD), provided the cyclic prefix is
sufficiently long. Typically, GVD is dominant, as it requires a longer cyclic prefix.
Assuming coherent detection, we show how to analytically compute the minimum
number of subcarriers and cyclic prefix length required to achieve a specified power
penalty, trading off power penalties from the cyclic prefix and from residual inter-
symbol interference (ISI) and inter-carrier interference (ICI). We derive an analytical
expression for the power penalty from residual ISI and ICI. We also show that when
nonlinear effects are present in the fiber, single-carrier with digital equalization
outperforms OFDM for various dispersion maps. We also study the impairments of
electrical to optical conversion when using Mach-Zehnder (MZ) modulators. OFDM
has a high peak-to-average ratio (PAR), which can result in low optical power
efficiency when modulated through a Mach-Zehnder (MZ) modulator. In addition, the
nonlinear characteristic of the MZ can cause significant distortion on the OFDM
signal, leading to in-band intermodulation products between subcarriers. We show that
a quadrature MZ with digital pre-distortion and hard clipping is able to overcome the
previous impairments. We consider quantization noise and compute the minimum
number of bits required in the digital-to-analog converter (D/A). Finally, we discuss a
dual-drive MZ as a simpler alternative for the OFDM modulator, but our results show
that it requires a higher oversampling ratio to achieve the same performance as the
quadrature MZ.
Abstract
v
In MANs, we discuss the use OFDM for combating GVD effects in amplified
direct-detection (DD) systems using single-mode fiber. We review known direct-
detection OFDM techniques, including asymmetrically clipped optical OFDM (ACO-
OFDM), DC-clipped OFDM (DC-OFDM) and single-sideband OFDM (SSB-OFDM),
and derive a linearized channel model for each technique. We present an iterative
procedure to achieve optimum power allocation for each OFDM technique, since there
is no closed-form solution for amplified DD systems. For each technique, we
minimize the optical power required to transmit at a given bit rate and normalized
GVD by iteratively adjusting the bias and optimizing the power allocation among the
subcarriers. We verify that SSB-OFDM has the best optical power efficiency among
the different OFDM techniques. We compare these OFDM techniques to on-off
keying (OOK) with maximum-likelihood sequence detection (MLSD) and show that
SSB-OFDM can achieve the same optical power efficiency as OOK with MLSD, but
at the cost of requiring twice the electrical bandwidth and also a complex quadrature
modulator. We compare the computational complexity of the different techniques and
show that SSB-OFDM requires fewer operations per bit than OOK with MLSD.
In LANs, we compare the performance of several OFDM schemes to that of
OOK in combating modal dispersion in multimode fiber links. We review known
OFDM techniques using intensity modulation with direct detection (IM/DD),
including DC-OFDM, ACO-OFDM and pulse-amplitude modulated discrete multitone
(PAM-DMT). We describe an iterative procedure to achieve optimal power allocation
for DC-OFDM, and compare analytically the performance of ACO-OFDM and PAM-
DMT. We also consider unipolar M-ary pulse-amplitude modulation (M-PAM) with
minimum mean-square error decision-feedback equalization (MMSE-DFE). For each
technique, we quantify the optical power required to transmit at a given bit rate in a
variety of multimode fibers. For a given symbol rate, we find that unipolar M-PAM
with MMSE-DFE has a better power performance than all OFDM formats.
Furthermore, we observe that the difference in performance between M-PAM and
OFDM increases as the spectral efficiency increases. We also find that at a spectral
efficiency of 1 bit/symbol, OOK performs better than ACO-OFDM using a symbol
Abstract
vi
rate twice that of OOK. At higher spectral efficiencies, M-PAM performs only slightly
better than ACO-OFDM using twice the symbol rate, but requires less electrical
bandwidth and can employ analog-to-digital converters at a speed only 81% of that
required for ACO-OFDM.
In PANs, we evaluate the performance of the three IM/DD OFDM schemes in
combating multipath distortion in indoor optical wireless links, comparing them to
unipolar M-PAM with MMSE-DFE. For each modulation method, we quantify the
received electrical SNR required at a given bit rate on a given channel, considering an
ensemble of 170 indoor wireless channels. When using the same symbol rate for all
modulation methods, M-PAM with MMSE-DFE has better performance than any
OFDM format over a range of spectral efficiencies, with the advantage of M-PAM
increasing at high spectral efficiency. ACO-OFDM and PAM-DMT have practically
identical performance at any spectral efficiency. They are the best OFDM formats at
low spectral efficiency, whereas DC-OFDM is best at high spectral efficiency. When
ACO-OFDM or PAM-DMT are allowed to use twice the symbol rate of M-PAM,
these OFDM formats have better performance than M-PAM. When channel state
information is unavailable at the transmitter, however, M-PAM significantly
outperforms all OFDM formats. When using the same symbol rate for all modulation
methods, M-PAM requires approximately three times more computational complexity
per processor than all OFDM formats and 63% faster analog-to-digital converters,
assuming oversampling ratios of 1.23 and 2 for ACO-OFDM and M-PAM,
respectively. When OFDM uses twice the symbol rate of M-PAM, OFDM requires
23% faster analog-to-digital converters than M-PAM but OFDM requires
approximately 40% less computational complexity than M-PAM per processor.

Dedication
vii
DEDICATION
Esta dissertao dedicada
minha irm Brbara Barros.
Acknowledgments
viii
ACKNOWLEDGMENTS
This thesis would not be possible without the help and support of many people.
I would like to express my gratitude and appreciation for my principal adviser,
Prof. Joseph Kahn, for his continuous support and guidance during my doctoral
program at Stanford. Prof. Kahn always paid close attention to the work of his
students and always aimed to perform exceptional work. I am fortunate to have
benefited from his knowledge, and his pursuit for excellence serves as an example for
me.
I would also like to thank my associate adviser, Prof. Bernard Widrow, for his
supervision and availability. I had the opportunity to work with Prof. Widrow in a
medical project where we studied diaphragm electromyography (EMG) signals from
patients with amyotrophic lateral sclerosis (ALS). I learned immensely from Prof.
Widrow and I am very grateful for all his help and advice.
I would like to thank Prof. Robert Twiggs from the Aeronautics and
Astronautics department for all of his support. I had the opportunity to work in his
team where we develop a cube satellite. The purpose of this project was to gather
temperature, acceleration and GPS data at 30000 feet and transmit the data in real-time
through a radio to the base station in the ground. The cube satellite was launched on a
rocket at White Sands missile range, New Mexico.
I would also like to thank the members of my Oral Exam committee, who were
Prof. Joseph Kahn, Prof. Bernard Widrow, Prof. Donald C. Cox and Prof. John Gill
for their willingness to be part of my committee. In particular, I want to thank Prof.
John Gill for his promptness to serve both as my dissertation reader and the chair of
my oral exam.
I also would like to thank Stanford University for the Stanford Graduate
Fellowship (SGF) and the Portuguese Foundation for Science and Technology for the
scholarship SFRH/BD/22547/2005.
Acknowledgments
ix
I would like to express my sincere appreciation to my wonderful girlfriend,
Rita Lopez, for all her understanding and support in the most difficult moments of my
program. I am also grateful for all the fun memories we shared together and without
her help this thesis would have not been possible.
I am also very grateful for all the support and guidance of my family, in
particular my parents, Maria Luisa Fernandes and Jose Manuel de Barros. My parents
were always there for me and supported me in all the decisions I made. I also want to
thank my in-laws Maria Teresa de Oliveira Braga and Rui Fernando Lopez.
Special mention must go to all members of the Optical Communications Group
of Prof. Kahn. I am thankful to Alan Lau, Ezra Ip and Jeff Wilde for all the ideas we
shared during our weekly group meetings. In addition, I also want to thank Rahul
Panicker, Mahdieh Shemirani, Tarek Abouzeid, Gwang-Hyun Gho, Reza Mahalati,
Daulet Askarov and Dany Ly-Gagnon for all the fun and memories.
Last but not least, I also want to thank all my great friends at Stanford: Hugo
Louro, Rita Oliveira, Hugo Caetano, Rita Fragoso, Isaac and Martha Martinez, Josh
and Kelly Alwood, Joo Vicente, Joo Rodrigues, Sabina Alistar, Kristiaan de Greve,
Sandra Beleza, Irina Weissbrot, Rinki Kapoor, Serena Faruque, Henrique Miranda,
Francisco Santos and many others. We shared great and fun moments together which
made life at Stanford much more enjoyable.
Table of Contents
x
TABLE OF CONTENTS
Abstract ........................................................................................................................... iv
Dedication ...................................................................................................................... vii
Acknowledgments .......................................................................................................... viii
1 Introduction ................................................................................................................. 1
2 Wide Area Networks .................................................................................................... 7
2.1 Introduction ........................................................................................................... 7
2.2 WAN and OFDM Review ....................................................................................... 9
2.2.1 WANs ............................................................................................................ 9
2.2.2 OFDM Review ............................................................................................. 11
2.3 Group-Velocity Dispersion .................................................................................... 15
2.3.1 Theory .......................................................................................................... 15
2.3.2 Simulation Results ........................................................................................ 17
2.4 Polarization-Mode Dispersion ............................................................................... 23
2.4.1 Theory .......................................................................................................... 23
2.4.2 Dual-Polarization Receiver ............................................................................ 25
2.4.3 Single-Polarization Receiver .......................................................................... 27
2.5 Fiber Nonlinearity ................................................................................................ 28
2.5.1 Theory .......................................................................................................... 28
2.5.2 Simulation Results ........................................................................................ 30
2.6 Computational Complexity .................................................................................... 33
2.7 Optical Modulator ................................................................................................ 35
2.7.1 PAR and MZ Review .................................................................................... 37
2.7.1.1 Peak-to-Average Power Ratio .................................................................. 37
2.7.1.2 Mach-Zehnder Modulator ........................................................................ 37
2.7.2 Quadrature MZ Optimization ......................................................................... 39
2.7.2.1 Quadrature MZ Optimization ................................................................... 39
2.7.2.2 Clipping Simulation Results ..................................................................... 42
Table of Contents
xi
2.7.2.3 Quantization Effects ................................................................................ 46
2.7.3 Dual-Drive MZ Optimization ......................................................................... 48
2.7.3.1 Dual-Drive MZ Modulator ....................................................................... 48
3 Metropolitan Networks ............................................................................................... 56
3.1 Introduction ......................................................................................................... 56
3.2 Metropolitan Networks and Power Allocation Review ............................................ 58
3.2.1 Metropolitan Networks .................................................................................. 58
3.2.2 Power and Bit Allocation Review ................................................................... 59
3.2.2.1 Gap Approximation ................................................................................. 59
3.2.2.2 Optimum Power Allocation ..................................................................... 61
3.3 OFDM System Model for Metro Links .................................................................. 63
3.4 Analysis of Direct-Detection OFDM Schemes ........................................................ 64
3.4.1 DC-Clipped OFDM ....................................................................................... 64
3.4.2 Asymmetrically Clipped Optical OFDM ......................................................... 66
3.4.3 Single-Sideband OFDM ................................................................................ 67
3.4.4 Effects of Amplifier Noise ............................................................................. 70
3.5 Comparison of Direct-Detection Modulation Formats ............................................. 71
3.5.1 Computational Complexity ............................................................................ 80
4 Local Area Networks .................................................................................................. 82
4.1 Introduction ......................................................................................................... 82
4.2 Local Area Networks and Discrete Bit Allocation Review ....................................... 83
4.2.1 Local Area Networks ..................................................................................... 83
4.2.2 Discrete Bit Allocation Review ...................................................................... 84
4.3 System Model for LANs ....................................................................................... 85
4.3.1 Overall System Model ................................................................................... 85
4.3.2 Multimode Fiber Model ................................................................................. 86
4.3.3 Performance Measures .................................................................................. 88
4.4 Analysis of IM/DD OFDM Schemes ...................................................................... 90
4.4.1 DC-Clipped OFDM ....................................................................................... 90
4.4.2 Asymmetrically Clipped Optical OFDM ......................................................... 91
Table of Contents
xii
4.4.3 PAM-Modulated Discrete Multitone ............................................................... 92
4.5 Comparison of IM/DD Modulation Formats ........................................................... 95
5 Personal Area Networks ............................................................................................ 105
5.1 Introduction ....................................................................................................... 105
5.2 Personal Area Networks ...................................................................................... 106
5.3 System Model and Performance Measures............................................................ 107
5.3.1 Overall System Model ................................................................................. 107
5.3.2 Optical Wireless Channel ............................................................................ 108
5.3.3 Performance Measures ................................................................................ 110
5.3.4 Ceiling-Bounce Model ................................................................................ 111
5.4 Comparison of IM/DD Modulation Formats ......................................................... 112
5.4.1 OOK and OFDM Performance ..................................................................... 114
5.4.2 Unipolar 4-PAM and OFDM Performance .................................................... 119
5.4.3 Outage Probability ...................................................................................... 122
5.5 Computational Complexity .................................................................................. 125
6 Conclusions ............................................................................................................. 127
6.1 Conclusions ....................................................................................................... 127
6.2 Future Work ....................................................................................................... 130
A Appendix...........................................................................................................................131
References .................................................................................................................... 135

List of Tables
xiii
LIST OF TABLES
Table 1.1 Characteristics of the different optical applications. .............................................. 3
Table 1.2 Characteristics of the different optical applications. .............................................. 4
Table 2.1 Number of subcarriers and cyclic prefix required to achieve an overall power
penalty less than 1 dB. We consider 4-QAM subcarriers, R = 26.7 GHz, D = 17 ps/(nmkm)
with 98% inline optical dispersion compensation. .............................................................. 22
Table 2.2 Computational complexity in real operations per bit for the various modulation
formats. We consider 4-QAM subcarriers, R = 26.7 GHz, D = 17 ps/(nmkm) with 98% inline
optical dispersion compensation. ...................................................................................... 35
Table 3.1 Electrical and optical bandwidths required to transmit at a bit rate R. ................... 72
Table 3.2 OFDM parameters for the various dispersion indexes . N is the DFT size, N
u
is the
number of used subcarriers and is the cyclic prefix. ......................................................... 78
Table 3.3 Memory required for various values of the dispersion index for OOK with MLSD
[3]. ................................................................................................................................. 81
Table 3.4 Number of real operations required per bit for SSB-OFDM and OOK with MLSD
for the various dispersion indexes . ................................................................................. 81
Table 5.1 System parameters for the various modulation formats for different bit rates and
symbol rates. N is the DFT size, N
u
is the number of used subcarriers, is the cyclic prefix, N
f

is the number of taps of the feedforward filter, N
b
is the number of taps of the feedback filter
and B
e
is the required electrical bandwidth in MHz. ......................................................... 116
Table 5.2 Computational complexity in real operations per bit for the various modulation
formats. ........................................................................................................................ 125

List of Figures
xiv
LIST OF FIGURES
Fig. 1.1 OFDM and single-carrier waveforms...................................................................... 2
Fig. 2.1 DWDM system with 50 GHz channel spacing. ...................................................... 10
Fig. 2.2 Long-haul link with N
A
spans. .............................................................................. 10
Fig. 2.3 OOK modulation in traditional optical systems. .................................................... 11
Fig. 2.4 Digital implementation of OFDM transmitter and receiver. .................................... 12
Fig. 2.5 OFDM symbol, including the cyclic prefix and windowing. ................................... 13
Fig. 2.6 OFDM spectrum, assuming rectangular windowing (N
win
= 0). ............................... 13
Fig. 2.7 Pulse spreading due to fiber GVD. ....................................................................... 15
Fig. 2.8 OFDM symbol with insufficient cyclic prefix. ...................................................... 16
Fig. 2.9 Probability of symbol error in presence of GVD for L = 80 km, D = 17 ps/(nm.km),
N
c
= 64, N
u
= 52, N
pre
= 5 and R = 26.7 GHz.OFDM symbol with insufficient cyclic prefix. .. 18
Fig. 2.10 ISI + ICI power penalty at P
S
= 10
4
for L = 80 km, D = 17 ps/(nm.km), N
c
= 64, N
u

= 52 and R = 26.7 GHz. ................................................................................................... 19
Fig. 2.11 Overall power penalty at P
S
= 10
4
for L = 80 km, D =17 ps/(nm.km), N
c
= 64, N
u
=
52 and R = 26.7 GHz. ...................................................................................................... 19
Fig. 2.12 Overall power penalties at R = 26.7 GHz for 2% undercompensated GVD lengths of
a) 1000 km b) 2000 km c) 3000 km d) 4000 km and e) 5000 km. The dashed lines are the
analytical values and the solid lines are the simulations results with insufficient cyclic prefix.
...................................................................................................................................... 21
Fig. 2.13 Optimum number of used subcarriers and cyclic prefix length as a function of the
fiber length for R = 26.7 GHz when no optical dispersion compensation is used. ................. 23
Fig. 2.14 Polarization-mode dispersion (PMD. .................................................................. 24
Fig. 2.15 OFDM polarization-multiplexed system. ............................................................ 25
Fig. 2.16 Combined PMD and GVD equalizer for an OFDM polarization-multiplexed system.
...................................................................................................................................... 26
Fig. 2.17 Probability of symbol error in presence of first-order PMD using a single-
polarization receiver, assuming = 21 ps, a| = |b| = 0.5, N
u
= 52, N
c
= 64, N
pre
= 1 and R = 26.7
GHz. .............................................................................................................................. 28
Fig. 2.18 Impact of nonlinearity on the transmission of BPSK: (a) without nonlinearity, (b)
with nonlinearity. ............................................................................................................ 30
Fig. 2.19 Phase noise variance for OFDM and SC transmissions with 100% dispersion
compensation per span: a) for 20 spans, as a function of launched power, and b) as function of
the number of spans. ........................................................................................................ 32
Fig. 2.20 Phase noise variance for OFDM and SC transmissions for 20 spans, as a function of
launched power: a) 90% dispersion compensation per span and b) 0% dispersion
compensation per span. .................................................................................................... 33
List of Figures
xv
Fig. 2.21 Model for single-drive MZ modulator, corresponding to one phase of a quadrature
MZ modulator. ................................................................................................................ 38
Fig. 2.22 Quadrature MZ modulator. ................................................................................ 38
Fig. 2.23 OFDM transmitter using quadrature MZ modulator. ............................................ 39
Fig. 2.24 Pre-distortion transfer characteristic for quadrature MZ modulator. ...................... 41
Fig. 2.25 OFDM transmitter including hard clipping, pre-distortion and electrode frequency
response compensation. ................................................................................................... 41
Fig. 2.26 Electrical and optical MZ waveforms. The solid and dotted lines represent the
waveforms with and without hard clipping, respectively. ................................................... 42
Fig. 2.27 MZ electrode frequency response. ...................................................................... 43
Fig. 2.28 Optical power efficiency for N
c
= 64, N
u
= 52 and R = 29.66 GHz. ....................... 44
Fig. 2.29 Receiver sensitivity penalty at P
S
= 10
4
, N
c
= 64, N
u
= 52 and R = 29.66 GHz. ..... 44
Fig. 2.30 Optical power efficiency for different number of subcarriers. ............................... 45
Fig. 2.31 Receiver sensitivity penalty for different number of subcarriers. ........................... 46
Fig. 2.32 Optical power efficiency for different number of bits in the D/A, N
c
= 64, N
u
= 52
and R = 29.66 GHz. ......................................................................................................... 47
Fig. 2.33 Receiver sensitivity penalty for different number of bits in the D/A, N
c
= 64, N
u
= 52
and R = 29.66 GHz. ......................................................................................................... 47
Fig. 2.34 Model for dual-drive MZ modulator. .................................................................. 49
Fig. 2.35 Dual-drive MZ plane. ........................................................................................ 49
Fig. 2.36 OFDM transmitter using the dual-drive MZ with hard-clipping, pre-distortion and
electrode frequency response compensation. ..................................................................... 50
Fig. 2.37 OFDM span regions for a fixed clipping level. The dotted and solid regions
correspond to the quadrature MZ and the dual-drive MZ, respectively. ............................... 50
Fig. 2.38 Real component of the dual-drive MZ output electric field. .................................. 51
Fig. 2.39 Output electric field of the dual-drive MZ for the cases of no trajectory optimization
and with trajectory optimization before the super-Gaussian filter. The green and blue lines
represent the output electric field and the correct OFDM waveform, respectively. ............... 52
Fig. 2.40 Dual-drive MZ receiver sensitivity penalty with trajectory optimization for M
s
= 2.5,
N
u
= 52 and R = 29.66 GHz. ............................................................................................. 53
Fig. 2.41 Dual-drive MZ optical power efficiency with trajectory optimization for M
s
= 2.5, N
u

= 52 and R = 29.66 GHz. ................................................................................................. 54
Fig. 2.42 Dual-drive and quadrature MZ receiver sensitivity penalties as a function of the
oversampling ratio with no frequency loss compensation and with trajectory optimization for
CR = 2.5, drive voltages range between V
and R = 29.66 GHz. ....................................... 54

Fig. 3.1 Metro network diagram. ...................................................................................... 58
Fig. 3.2 Metro link with an optical filter. ........................................................................... 59
Fig. 3.3 Intensity modulation (IM) by direct modulation of the laser current. ....................... 59
Fig. 3.4 Achievable bit rates as a function of SNR for various values of the gap . .............. 60
Fig. 3.5 Optimal power allocation for OFDM. ................................................................... 62
Fig. 3.6 OFDM system model for metro links. .................................................................. 63
List of Figures
xvi
Fig. 3.7 Block diagram of a DC-OFDM transmitter. .......................................................... 64
Fig. 3.8 Equivalent transfer function for DC-OFDM for R = 20 Gbit/s, L = 100 km and D = 17
ps/nm/km, corresponding to = 0.87. ............................................................................... 66
Fig. 3.9 Block diagram of an ACO-OFDM transmitter. ...................................................... 67
Fig. 3.10 Block diagram of an SSB-OFDM transmitter. ..................................................... 68
Fig. 3.11 PSD of the detected signal in SSB-OFDM. ......................................................... 69
Fig. 3.12 PSD of the different modulation formats. ............................................................ 72
Fig. 3.13 Flow chart of DC-OFDM and SSB-OFDM optimization using the bias ratio (B
R
). . 74
Fig. 3.14 Normalized optical SNRs required for the different OFDM formats for = 0.25 to
achieve P
b
= 10
3
. The number of used subcarriers, N
u
, is indicated for each curve. .............. 75
Fig. 3.15 Subcarrier power distribution for DC-OFDM, ACO-OFDM, SSB-OFDM and to
achieve a P
b
= 10
3
for = 0.25. The number of used subcarriers is 416, 416 and 208 for DC-
OFDM (B
R
= 1.1), ACO-OFDM and SSB-OFDM (B
R
= 1.0), respectively. ......................... 76
Fig. 3.16 Normalized optical SNR required for DC-OFDM to achieve P
b
= 10
3
for = 0.25.
The number of used subcarriers is N
u
= 832. ..................................................................... 77
Fig. 3.17 Minimum normalized optical SNR required for various dispersion indexes for the
different OFDM formats. ................................................................................................. 78
Fig. 3.18 OSNR values (over 0.1 nm) required to obtain P
b
= 10
3
at 10.7 Gbit/s for OFDM
and for OOK with MLSD [3]. .......................................................................................... 79
b
= 10
3
at 10.7 Gbit/s for OFDM
and for OOK with MLSD [3]. In this case, the OFDM signal occupies the full channel
bandwidth (B
0
= 35 GHz). ................................................................................................ 80
Fig. 4.1 LAN block diagram. ........................................................................................... 84
Fig. 4.2 LAN link. ........................................................................................................... 84
Fig. 4.3 OFDM system model for LANs. .......................................................................... 85
Fig. 4.4 (a) Mode power distribution, (b) mode delays and (c) frequency response for fiber
1183. We consider a 1-km length in computing the delays and frequency response. ............. 87
Fig. 4.5 3-dB bandwidth distribution of the multimode fibers simulated. All fibers have 1-km
length. ............................................................................................................................ 88
Fig. 4.6 Rms delay spread (D) distribution of the multimode fibers simulated. All fibers have
1-km length. ................................................................................................................... 90
Fig. 4.7 Block diagram of a PAM-OFDM transmitter. ....................................................... 92
Fig. 4.8 Receiver electrical SNR required to obtain 10 Gbit/s at P
b
= 10
4
for ACO-OFDM and
OOK for fibers with 1-km length. The bit allocation granularity is = 0.25 and ACO-OFDM
has the same symbol rate as OOK (
OFDM
s
R = R
s
= 10 GHz). ................................................. 96
Fig. 4.9 Optical spectra of M-PAM and OFDM. The symbol rate for OFDM is twice that for
M-PAM,
OFDM
s
R = 2R
s
. ...................................................................................................... 97
Fig. 4.10 Receiver electrical SNR required to achieve P
b
= 10
4
at 10 Gbit/s for different
modulations formats in fibers with 1-km length. The bit allocation granularity is = 0.25. The
symbol rate for all OFDM formats is twice that for OOK,
OFDM
s
R = 2R
s
= 20 GHz. ............... 98
List of Figures
xvii
Fig. 4.11 Receiver electrical SNR required for DC-OFDM to achieve P
b
= 10
4
at 10 Gbit/s for
different values of the bias ratio in fiber 295. .................................................................... 99
Fig. 4.12 Receiver electrical SNR required for various OFDM formats with continuous and
discrete bit allocations to achieve P
b
= 10
4
at 10 Gbit/s for fibers of 1-km length. The symbol
rate for all OFDM formats is
OFDM
s
R = 20 GHz. ................................................................. 100
Fig. 4.13 Subcarrier power distribution for ACO-OFDM with continuous bit allocation and
with discrete bit loading (with granularity = 0.25) for 10 Gbit/s at P
b
= 10
4
in fiber 10. The
symbol rate is
OFDM
s
R = 20 GHz. ....................................................................................... 101
Fig. 4.14 Cumulative distribution function (CDF) of the required receiver electrical SNR to
obtain P
b
= 10
4
at 10 Gbit/s for different modulation formats. The symbol rate for OOK is R
s

= 10 GHz and the symbol rate for OFDM is the same or twice that for OOK, as indicated in
the figure. ..................................................................................................................... 102
b
= 10
4
for the different
modulations formats for fibers of 1-km length. The bit allocation granularity is = 0.25 and
the symbol rate for 4-PAM is R
s
= 10 GHz. The symbol rate for ACO-OFDM is the same or
twice that for 4-PAM, as indicated in the figure. .............................................................. 102
Fig. 4.16 Cumulative distribution function (CDF) of the required receiver electrical SNR to
obtain 20 Gbit/s at P
b
= 10
4
for different modulation formats. The symbol rate for 4-PAM is
R
s
= 10 GHz and the symbol rate for ACO-OFDM is the same or twice that for 4-PAM, as
indicated in the figure. ................................................................................................... 103
Fig. 5.1 Indoor optical wireless transmission. .................................................................. 107
Fig. 5.2 OFDM system model for LANs. ........................................................................ 108
Fig. 5.3 Impulse response of an exemplary non-directional, non-LOS (diffuse) channel. This
channel has no LOS component h
(0)
(t). The contributions of the first five reflections, h
(1)
(t),,
h
(5)
(t), are shown. .......................................................................................................... 110
Fig. 5.4 Electrical SNR required to achieve P
b
= 10
4
vs. normalized delay spread D
T
at bit
rates of 50, 100 and 300 Mbit/s (spectral efficiency of 1 bit/symbol) for ACO-OFDM and
OOK. The bit allocation granularity is = 0.25 and the symbol rates for ACO-OFDM are the
same as those for OOK, as indicated. .............................................................................. 114
b
= 10
4
T
at bit
rates of 50, 100 and 300 Mbit/s (spectral efficiency of 1 bit/symbol) for different modulations
formats. The dashed lines correspond to the SNR requirement predicted using the ceiling-
bounce (CB) model. The bit allocation granularity is = 0.25 and the symbol rates for ACO-
OFDM are twice those for OOK, as indicated. ................................................................ 117
Fig. 5.6 Cumulative distribution function (CDF) of the electrical SNR required to achieve P
b
=
10
4
at bit rates of 50, 100 and 300 Mbit/s (spectral efficiency of 1 bit/symbol) for different
modulation formats. The symbol rates for OFDM are the same as or twice those for OOK, as
indicated. ...................................................................................................................... 119
b
= 10
4
T
at bit
rates of 100, 200 and 600 Mbit/s (spectral efficiency of 2 bit/symbol) for different modulations
formats. The bit allocation granularity is = 0.25 and the symbol rates for ACO-OFDM are
the same as or twice those for 4-PAM, as indicated. ........................................................ 120
List of Figures
xviii
b
=
10
4
at bit rates of 100, 200 and 600 Mbit/s (spectral efficiency of 2 bit/symbol) for different
modulation formats. The symbol rates for OFDM are the same as or twice those for 4-PAM,
as indicated. .................................................................................................................. 120
Fig. 5.9 Outage probability for OOK and ACO-OFDM with coding and various bit allocations
averaged over all channels. All modulation formats use the code RS(127,107) over GF(8). The
information bit rate is 300 Mbit/s (spectral efficiency of 1 bit/symbol) and the symbol rates for
OOK and OFDM are 356 MHz and 600 MHz, respectively. The bit allocation granularity is
= 1, i.e., integer bit allocation. ........................................................................................ 124
Fig. A-1 DWDM system with 50 GHz channel spacing .................................................. 134

Introduction
1
1 INTRODUCTION
Research in optical communication systems has received ever-increasing
interest in recent years. The available spectrum in fiber is being rapidly populated with
continued growth of internet traffic driven by bandwidth-hungry applications, such as
video and music sharing, so there is an increasing demand to increase spectral
efficiency transmission while maintaining high SNR efficiency, i.e., minimizing the
required average transmitted energy per bit. A major contributing factor in the
development of new optical systems is the advance in very large scale integration
(VLSI) technology, which enables digital signal processing-based receivers at GHz
clock speeds. When the outputs of an analog optoelectronic downconverter are
sampled at Nyquist rate, the digitized waveform retains the full information of the
received optical electric field, enabling compensation of transmission impairments by
a digital signal processor (DSP). A digital receiver is highly advantageous because
adaptive algorithms can be used to compensate time-varying transmission
impairments. Forward error-correction codes with soft-decision decoding can also be
implemented. Moreover, digitized signals can be delayed, split and amplified without
degradation in signal quality. Digital receivers are already widely used in wireless and
DSL systems that operate at comparatively lower data rates.
Optical networks can be divided into four classes: wide area networks
(WANs), metropolitan area networks (MANs), local area networks (LANs) and
personal area networks (PANs). There are also data center networks (DCNs) which are
very similar to LANs. There have been several experiments showing the increased
performance of using single-carrier modulation with digital equalization for the
different optical applications [1,2,3]. However, with the advances in digital receivers
for optical systems, one can now employ now more advanced modulations formats,
such as orthogonal frequency division multiplexing (OFDM). OFDM is a multi-carrier
modulation format that divides the total bandwidth into N orthogonal slots called
subcarriers, as shown in Fig. 1.1. Among all multicarrier decompositions, OFDM with
Introduction
2
f
Bandwidth
OFDM
f
Bandwidth
Single-Carrier
Time Time
t
T
OFDM
t
1
-1
T
s
cyclic prefix is practical and converges to the optimum performance [4]. Furthermore,
the power of the OFDM subcarriers can be optimized according to the application.

Fig. 1.1 OFDM and single-carrier waveforms.
OFDM is currently used in wireless and DSL systems for its optimized
performance in frequency-selective channels. Hence, it is important to know if OFDM
can achieve higher performance than single-carrier with equalization for the next
generation optical systems.
This thesis focuses on the analysis and optimization of the OFDM performance
for the different optical applications. The requirements and constrains of the different
optical applications are summarized in Table 1.1.
Applications
WAN
(>1000 km)
MAN
(50 500 km)
LAN
(0.3 5 km)
PAN
(10 m)
Goal 100-1000 Gbit/s 10-100 Gbit/s 10-100 Gbit/s 300 Mbit/s
Medium
Single-Mode
Fiber
Single-Mode
Fiber
Multimode
Fiber
Free Space
Modulation
Electric Field
(Complex)
Electric Field
or Intensity
Intensity Intensity
Detection
Method
Coherent
(Electric field)
Direct Detection
(Intensity)
Direct
Detection
(Intensity)
Direct
Detection
(Intensity)
Dominant
Dispersion
Chromatic/Polar.
(Linear)
Chromatic
(Nonlinear)
Multimode
(Linear)
Multipath
(Linear)
Introduction
3
Dominant
Noise
Amplifier
(AWGN)
Amplifier
(Chi-squared
2
)
Thermal
(AWGN)
Ambient
Light
(Shot noise)
Noise Type Signal-Indep. Signal-Dep. Signal-Indep. Signal-Indep.
OFDM
Formats
OFDM
DC-OFDM
ACO-OFDM
SSB-OFDM
DC-OFDM
ACO-OFDM
DC-OFDM
ACO-OFDM
Other
Effects
Fiber
Nonlinearity
Clipping Noise
Clipping
Noise
Clipping
Noise
Table 1.1 Characteristics of the different optical applications.
Long-haul systems (i.e., WANs) typically use single-mode fiber (SMF) and the
dominant linear impairments are group-velocity dispersion, i.e., different frequencies
travel at different speeds, and polarization-mode dispersion, i.e., different
polarizations arrive at the receiver with different delays (Chapter 2). The different
fiber types are summarized in Table 1.2. Next generation WANs will employ coherent
detection, i.e., detect the phase and amplitude of the optical electric field. In WANs,
the optical amplifiers have sufficient high gain so that its amplified spontaneous
emission (ASE) is dominant over thermal and shot noises. The ASE noise is modeled
as complex additive white Gaussian noise (AWGN). When the input power is high,
fiber nonlinear effects become dominant and degrade the system performance.
Multimode Fiber (MMF) Single-Mode Fiber (SMF)
Large core (62.5 m) Small core (9 m)
Supports many propagation modes Supports only a pair of degenerate modes
o Modal dispersion: different modes
travel with different propagation
constants
o Large delay spreads
o Refractive index changes with
frequency (GVD)
o Fiber asymmetry can cause
polarization-mode dispersion (PMD)
o Small delay spreads
Low bandwidth-distance product
(5 Gbit/skm)
High bandwidth-distance product
(200 Tbit/skm)
Introduction
4
Used in local area networks (LANs) Used in wide area networks (WANs)
Table 1.2 Characteristics of the different optical applications.
Metro systems (MANs) also use SMF but normally use direct-detection at the
receiver, i.e., detect the fiber instantaneous power (or intensity). When direct-detection
is used, the transmitted waveforms have to be non-negative since only the intensity is
detected (Chapter 3). However, the transmitter in metro systems can either modulate
the electric field or the intensity. The amplifier ASE noise is no longer Gaussian due
to the nonlinear photodetection process and becomes Chi-squared
2
distributed.
Furthermore, in the detection process, the amplifier noise interacts with the received
signal and becomes signal-dependent.
LANs typically use multi-mode fiber (MMF) and employ intensity modulation
with direct detection. The main impairment is modal dispersion, i.e., different modes
travel with different propagation constants (Chapter 4). Optical amplifiers are not used
in LANs and the dominant noises are thermal noise or shot noise from the receiver.
The DCNs also use MMF over short distances as LANs and typically operate at 10
Gbit/s.
Indoor optical wireless systems (i.e. PANs) use free-space (i.e., air) and also
employ intensity modulation with direct detection. Usually, optical wireless systems
transmit either in the infrared or visible-light spectra. The main impairment is
multipath distortion, i.e., different reflections arrive at the receiver with different
delays (Chapter 5). The dominant noises are ambient light (shot noise) or thermal
noise from the receiver.
This thesis is organized as: in Chapter 2, we review OFDM modulation and the
major types of impairments in long-haul systems using single-mode fiber and coherent
detection. When only linear effects are present, we derive analytical expressions for
the inter-symbol interference, and inter-carrier interference incurred when an
insufficient cyclic prefix is used. We use these expressions to compute power penalties
for representative examples, comparing the results to simulations. We also present the
required number of subcarrier and cyclic prefix length for different fiber lengths.
Introduction
5
Afterwards, we include fiber nonlinear effects in our analysis and compare the
performance of OFDM to single-carrier systems. We also present two feasible optical
modulators to generate an optical OFDM signal. We show how to minimize the
nonlinear distortion of the modulators for best performance. We also compute the
required number of bits for the digital-to-analog (D/A) converters and sampling
frequencies for the two optical modulators.
In Chapter 3, we review the fundamentals of metro systems. We also review
methods for power and bit allocation for multicarrier systems and describe the optimal
water-filling solution. Furthermore, we introduce the different OFDM formats that can
be used with direct-detection and derive equivalent linear channel models for each
one, assuming only GVD is present in the SMF. In addition, we discuss the effects of
amplifier noise in direct-detection systems and present how to optimize the different
OFDM formats in metro systems. We conclude the chapter by comparing the optical
power required to transmit at a given bit rate for the different optimized OFDM
formats and for OOK with MLSD, making use of previously published results for the
latter.
In Chapter 4, we introduce the multimode fiber model and review different
OFDM formats that can be used with intensity modulation. We then compare the
receiver electrical SNR required to transmit at a given bit rate for the different OFDM
formats and for unipolar PAM with MMSE-DFE equalization at different spectral
efficiencies when the transmitter has channel state information (CSI).
In Chapter 5, we review the indoor optical wireless model for infrared or
visible light communication. We use the same OFDM formats as in chapter 4. We first
assume that the transmitter has CSI and compare the receiver electrical SNR required
to transmit at a given bit rate for the different OFDM formats and for unipolar PAM
with MMSE-DFE equalization at different spectral efficiencies. We then compare the
receiver electrical SNR for the different modulation formats with error correcting
codes when there is no transmitter CSI. We conclude the chapter by comparing the
computation complexity of the different modulation formats.
Introduction
6
In Chapter 6, we give the concluding remarks of this thesis and discuss
directions for future work.

Wide Area Networks
7
2 WIDE AREA NETWORKS
2.1 Introduction
Orthogonal frequency-division multiplexing (OFDM) is a multi-carrier
modulation that has been extensively investigated and deployed in wireless and
wireline communications [5], [6]. It is receiving increased interest in the fiber-optic
research community for its robustness against inter-symbol interference (ISI), since
the symbol period of each subcarrier can be made long compared to the delay spread
caused by group-velocity dispersion (GVD) and polarization-mode dispersion (PMD)
[7], [8]. Although single-carrier and multi-carrier systems using coherent detection
have fundamentally the same power and spectral efficiencies for a given modulation
format in the presence of unitary impairments such as GVD and PMD, there may be
differences due to practical constraints.
Experiments have been performed demonstrating the potential of OFDM with
coherent detection in optical systems [8], [9], but very little closed-form performance
analysis has been performed to date. In principle, the impulse response due to GVD
has infinite duration, so the proper choice of cyclic prefix length is not obvious. If the
cyclic prefix is too short relative to the impulse response duration, ISI and inter-carrier
interference (ICI) will occur. On the other hand, if the cyclic prefix is excessively long
relative to the number of subcarriers, the sampling rate must be increased
significantly, and a large fraction of the transmitted energy is wasted in cyclic prefix
samples, leading to a substantial power penalty. For a given cyclic prefix length, these
penalties can be reduced to an arbitrary degree by increasing the number of
subcarriers.
In practice, however, it may be desirable to minimize the number of
subcarriers employed. For example, it is known that the peak-to-average power (PAR)
ratio is proportional to N
c
, the number of subcarriers [5]. OFDM signals are modulated
using Mach-Zehnder (MZ) modulators having nonlinear, peak-limited transfer
Wide Area Networks
8
characteristics [10], [11], so minimizing N
c
will help maximize optical power
efficiency in the modulator. As another example, laser phase noise destroys the
orthogonality between subcarriers, causing ICI. It has been shown that for a given
laser linewidth, the variance of ICI is proportional to N
c
[12], so minimizing N
c
can
help in combating laser phase noise. Hence, in this chapter, assuming coherent
detection, we find the minimum number of subcarriers and cyclic prefix length that
achieve low power and sampling penalties in the presence of GVD and PMD.
In this chapter, we also study the MZ modulator. Since the MZ has a nonlinear,
peak-limited transfer characteristic, the high peaks of the OFDM signal can degrade
the system performance. A conventional solution to the PAR problem is to reduce the
operating range in the MZ to accommodate the OFDM peak. However, this solution
results in a significant power efficiency penalty, which may require the use of optical
amplification at the transmitter to boost the signal level. An alternative solution would
be using peak-reduction algorithms studied for wireless systems [5] but all these
algorithms present a computational burden to the transmitter [13], which might be
prohibitive at the speeds of optical systems. In this chapter, we present hard clipping
with pre-distortion as a simple and effective approach to combat the nonlinearity in the
quadrature MZ and increase the optical power efficiency. In addition, we study the
combined effects of having a finite number of bits in the D/A and the MZ nonlinearity.
We then extend our study to the dual-drive MZ for generating an optical OFDM signal
since it requires less hardware than the quadrature MZ.
This chapter is organized as follows. In section 2.2, we review the
fundamentals of wide area networks (WANs) and multi-carrier systems. We introduce
the canonical OFDM model and determine the minimum oversampling ratio required
to avoid aliasing. In section 2.3, we focus on GVD, and derive analytical expressions
for the ISI and ICI incurred when an insufficient cyclic prefix is used. We use these
expressions to compute power penalties for representative examples, comparing the
results to simulations. We compare the performance of OFDM to single-carrier
systems. In section 2.4, we consider first-order PMD, and discuss how to extend the
analysis to arbitrary-order PMD. We consider both single- and dual-polarization
Wide Area Networks
9
receivers. In section 2.5, we present the fiber nonlinearity model and compare the
performance of OFDM to single-carrier systems when both linear and nonlinear
effects are present. We calculate the computational complexities of the different
modulation formats in section 2.6. In section 2.7, we review PAR fundamentals in
multi-carrier systems and the quadrature drive MZ as an optical modulator. In
addition, we introduce the MZ canonical model including the electrode frequency
response. We then focus on the quadrature MZ, and study through simulations the
optical power efficiency (OPE) and system performance gains when pre-distortion and
hard clipping are used. Moreover, we consider quantization noise from the D/A and
study through simulations the minimum number of bits required and the optimum
clipping level. Afterwards, we introduce the dual-drive MZ as a simpler option for the
OFDM modulator. We then analyze the oversampling required to achieve low
performance degradation in the dual-drive MZ.
2.2 WAN and OFDM Review
2.2.1 WANs
Modern long-haul optical systems (or WANs) use single-mode fiber (SMF)
and operate around 1.55 m since it is the region with the lowest fiber attenuation. The
total available bandwidth at 1.55 m is around 8 THz (C-band). The total available
bandwidth is partitioned into several channels, typically with 100 or 50 GHz
bandwidth, and is referred as dense wavelength-division multiplexing (DWDM). Fig.
2.1 shows a block diagram of a DWDM system with 50 GHz channel spacing.
Wide Area Networks
10

Fig. 2.1 DWDM system with 50 GHz channel spacing.
Long-haul systems transmit over long distances, often well beyond 1000 km.
The long-haul link is divided into spans, typically 80 to 100 km long. An optical
amplifier is placed at the end of each span to compensate the SMF loss. Many long-
haul systems use inline optical dispersion compensation. In such systems, a
dispersion-compensating fiber (DCF) is also used on each span to cancel some of the
dispersion introduced by the SMF. The DCF usually compensates between 90% and
98% to the SMF dispersion, depending on the dispersion map. A second amplifier is
placed after the DCF in order to compensate the additional attenuation introduced by
the DCF. Fig. 2.2 shows a block diagram of a long-haul link.
Fig. 2.2 Long-haul link with N
A
spans.
Traditional optical systems used binary formats such as on-off keying (OOK)
for each DWDM channel. In OOK, the transmitter turns the laser ON and OFF to
represent bits 1 and 0 as shown in Fig. 2.3.
50 GHz
2
1 N
AWG
TX 1
TX 2
TX N1
TX N
RX 1
RX 2
RX N1
RX N
AWG
2
1 N
SMF
TX RX
Span 1 Span N
A
SMF
DCF
SMF
DCF
Wide Area Networks
11

Fig. 2.3 OOK modulation in traditional optical systems.
Given practical constraints on filters for DWDM, the binary formats using
direct-detection (DD) can only achieve spectral efficiencies around 0.8 b/s/Hz per
polarization; over the usable fiber bandwidth of about 10 THz, this corresponds to an
aggregate capacity of about 8 Tbit/s [14]. Until recently, this was considered much
larger than that required by real world applications. However, with the continued
growth of internet traffic driven by video and music sharing, the available spectrum in
fiber is being rapidly populated, so there is renewed interest in high spectral efficiency
transmission. The most promising detection technique for achieving high spectral and
power (SNR) efficiency is coherent detection with polarization multiplexing. Coherent
detection with polarization multiplexing detects the amplitude and phase of the optical
electrical field in the two fiber polarizations. Hence, higher spectral efficiencies can be
achieved with coherent detection since the information can encoded in all the available
degrees of freedom. Coherent systems use lasers with very low linewidths (e.g. < 300
KHz) to alleviate carrier phase and frequency recovery at the receiver.
2.2.2 OFDM Review
In OFDM, the inverse discrete Fourier transform (IDFT) and DFT are used to
modulate and demodulate the data constellations on the subcarriers, as shown in Fig.
2.4. The IDFT and DFT replace the banks of analog I/Q modulators and demodulators
that would otherwise be required. To show the equivalence between OFDM and
analog multi-carrier systems, we can write the OFDM signal as
Laser
Encoder
Bits
0 1 0 1 0 1 1 0
0 1 0 1 0 1 1 0
RX
Electrical
LPF
Decision
0 1 0 1 0 1 1 0
Fiber
TX
Wide Area Networks
12

( ) ( )
=
=
k
N
q
t qf j
sym k q ofdm
c
d
e kT t b X t x
1
0
2
,

,
(2.1)
where X
q,k
denotes the qth subcarrier constellation symbol transmitted on the kth
OFDM symbol, b(t) is a pulse shape, T
sym
= (N
c
+ N
pre
+ N
win
)T
c
= T
ofdm
+ T
pre
+ T
win
is
the OFDM symbol period and f
d
= 1/(N
c
T
c
) is the frequency separation between
subcarriers. We define T
c
as the sample period, and N
c
, N
pre
and N
win
are integers.
Fig. 2.4 Digital implementation of OFDM transmitter and receiver.
If Eq. (2.1) is sampled every sample period, we have
( ) ( )
=
=
k
N
q
N
qn
j
sym c k q c ofdm
c
c
e kT nT b X nT x
1
0
2
,
.
(2.2)
An OFDM symbol corresponds to N
c
+ N
pre
+ N
win
samples, as shown in Fig.
2.5. The block of N
c
samples in Eq. (2.2) corresponds to the IDFT. The remaining N
pre

and N
win
terms are a periodic extension of the OFDM signal known as the cyclic
prefix, and pulse shaping known as windowing. The cyclic prefix is used so that the
sequence of received samples in one symbol is equivalent to one period of a circular
convolution between the transmitted OFDM symbol x
ofdm
(t) and the sample-rate
samples of the channel impulse response h
channel
(t). In the frequency domain, this
corresponds to the multiplication of subcarrier q by the corresponding sample of the
channel frequency response H
channel
(). Thus, a single-tap equalizer on subcarrier q
can be used to invert any amplitude and phase distortion introduced by the channel,
i.e.,
) / 2 ( ) (
1
c s channel eq
N q f H q H
=
. For our system, the channel frequency response
H
channel
() is given by

) ( ) ( ) ( ) ( V H P H
fiber channel
= ,
(2.3)
Add Prefix
(Npre)
+
Windowing
(Nwin)
.
.
.
X0,k
X1,k
XN-2,k
XN-1,k
.
.
.
.
.
.
P/S Fiber
IDFT
Remove
Prefix
(Npre)
+
Windowing
(Nwin)
.
.
.
.
.
.
DFT
Coherent
O/E Down-
conversion
LO
S/P
.
.
.
Y0,k
Y1,k
YN-1,k
Pulse
Shaping
x
x
x
Heq(N-1)
Heq(1)
Heq(0)
X0,k
X1,k
XN-1,k
p(t)
v(t)
xofdm(t) = I(t)+jQ(t)
hfiber(t)
Anti-
Aliasing
D/A
D/A
Pulse
Shaping
I(t)
Q(t)
p(t)
hchannel(t) = p(t) hfiber(t) v(t)
Anti-
Aliasing
I'(t)
Q'(t)
v(t)
A/D
A/D
MZ laser
Wide Area Networks
13
where P() and V() are the pulse-shaping and anti-aliasing filters and H
fiber
()
represents the fiber frequency response. The cyclic prefix samples (N
pre
= N
pos
+ N
neg
)
are appended to the signal after the IDFT. Ideally, the cyclic prefix length should be
no smaller than the duration of the channel impulse response h
channel
(t).

Fig. 2.5 OFDM symbol, including the cyclic prefix and windowing.
The additional N
win
samples are used to control the OFDM spectrum by
introducing additional pulse shaping. Common choices for the window functions
include rectangular and raised-cosine pulses [5], [6]. The rectangular window
corresponds to N
win
= 0. Assuming a rectangular window and uncorrelated transmitted
symbols, E[x
n,k
x
l,m
*
] = 0 for n l, the power spectrum of x
ofdm
(t) is given by

( ) ( )
=
1
0
2
2
2
2
sinc
c
N
q
d
sym
q
c
sym
xx
qf
T
P
N
T
S
,
(2.4)
where P
q
= E[|X
q
|
2
] is the average power per symbol on sub-carrier q. A plot of
Eq. (2.4) is shown in Fig. 2.6, assuming all subchannels have equal powers.

Fig. 2.6 OFDM spectrum, assuming rectangular windowing (N
win
= 0).
We note in Fig. 2.6 that the OFDM spectrum resembles an ideal rectangular
shape and has a bandwidth approximately given by

c
N
neg
N
pos
N
win
N
win pre ofdm sym
T T T T + + =
Effective Tx time (T
ofdm
)
sym
T
2
( )

+
sym
d c BW
T
f N
2
1 2
d
f 2
( )
xx
S
Wide Area Networks
14

d c
sym
d c BW
f N
T
f N ) 1 ( 2
2
) 1 ( 2 +
+
,
(2.5)
To achieve a desired symbol rate R, the frequency separation between
subcarriers (f
d
=1/N
c
T
c
) must be equal to f
d
= (N
c
+N
pre
)/N
c
R/N
c
. Thus, Eq. (2.5) can
be rewritten as

R
N
N N
N
N
c
pre c
c
c
BW
) (
) 1 (
2
+
+

.
(2.6)
In the case of the rectangular window, the OFDM spectrum has significant
sidelobes, as shown in Fig. 2.6. These can be reduced by using alternate windows,
such as the raised-cosine, which require N
win
> 0. Alternately, the sidelobes can be
reduced by using a pulse-shaping filter p(t) after the digital- to-analog converter
(D/A).
The symbols extensions required for the cyclic prefix (N
pre
) and windowing
(N
win
) represent penalties, since they do not carry useful information and are discarded
at the receiver. They represent a power penalty because a portion of the energy is
wasted on these samples and also a sampling penalty because the sample rate 1/T
c
,
which is equal to the sampling frequency, has to be higher to maintain the desired
symbol rate R, i.e., f
s
= 1/T
c
= R(N
c
+N
pre
+N
win
)/N
c
.
For typical values used in OFDM systems, the OFDM bandwidth given by
Eq. (2.6) is very close to the Nyquist bandwidth, i.e.,
BW

Ny
= 2R. The confined
spectrum of OFDM is a practical advantage over single-carrier systems where, the
90% power bandwidth is typically of order R
sc
2 2 [15]. Another important
advantage of OFDM is in the minimum required oversampling ratio. In OFDM,
oversampling is performed by not modulating the edge subcarriers in Fig. 2.4, i.e., by
inserting zero subcarriers. Assuming that only N
u
of the N
c
subcarriers are modulated,
the oversampling ratio is M
s
= N
c
/N
u
. In OFDM, it is possible to employ arbitrary
rational oversampling ratios, unlike single-carrier transmission, where this may require
complex signal processing. Since the OFDM spectrum falls off more rapidly than a
single-carrier spectrum, lower oversampling ratios can be employed. We have found
Wide Area Networks
15
that when the pulse-shaping and anti-aliasing filters are chosen properly, an
oversampling ratio M
s
= 1.2 is sufficient to avoid aliasing, while single-carrier systems
typically require an oversampling ratio M
s
= 1.5 or 2 [16]. Finally, we note that the
expressions defined previously are also valid when oversampling is used, provided
that N
pre
and N
win
are scaled to reflect the oversampling ratio, i.e., N
pre
(M
s
) =
M
s
N
pre
(M
s
= 1) and N
win
(M
s
) = M
s
N
win
(M
s
= 1) and the sampling frequency is scaled
by the oversampling ratio, i.e., f
s
= 1/T
c
= M
s
R(N
c
+ N
pre
(M
s
)+ N
win
(M
s
))/N
c
.
2.3 Group-Velocity Dispersion
2.3.1 Theory
The refractive index of the fiber changes with frequency so the different
frequencies of a pulse propagate at different speeds. This effect is called group-
velocity dispersion (GVD). Fiber GVD spreads the transmitted symbols, causing ISI
that degrades the error probability, as shown in Fig. 2.7.

Fig. 2.7 Pulse spreading due to fiber GVD.
The fiber frequency response in the presence of GVD is given by

L j
fiber
e H
2
2 2
) (

= ,
(2.7)
where
2
is the fiber GVD parameter and L is the fiber length. Although the receiver
employs a single-tap equalizer on each subcarrier to invert the channel, ISI can occur
when the cyclic prefix is insufficient, so that symbols in a neighboring block overlap
with the symbol of interest. In order to avoid ISI, ideally, the cyclic prefix duration
T
pre
should no smaller than the duration of the impulse response h
channel
(t). However,
GVD leads to an infinite-duration impulse response, and the tails of the impulse
response not covered by the cyclic prefix lead to residual ISI, as shown in Fig. 2.8.
t t
SMF
Input
Output
Wide Area Networks
16

Fig. 2.8 OFDM symbol with insufficient cyclic prefix.
Referring to Fig. 2.8, note that ISI on symbol k comes both from symbols k1
and k+1, since the channel impulse response is two-sided. If N
pre
= N
pos
+ N
neg

samples are used for the cyclic prefix and if the channel impulse response has positive
and negative lengths L
p
1 and L
n
1,

respectively
,
the residual ISI on the nth time-
domain sample in the kth OFDM symbol interval can be written as

) ( ) (
) ( ) ( ) (
1
1
1
1
1
1
u h u N N n x
r h r N N N n x n ISI
channel
N
L u
neg c k
channel
L
N r
pos pre c k k
neg
n
p
pos

+ =
+
+ =

+
+ + + =
.
(2.8)
In Appendix A, we show that if the transmitted symbols are uncorrelated and if
N
c
> max(L
p
N
pos
1, L
n
N
neg
1), after the DFT, the ISI variance on subcarrier q is
given by

+ =

+ =
+ =
1
1
2
1
1
2
2 2
) ( ) ( ) (
neg
n pos
N
L n
n
Lp
N p
p s ISI
q H q H q
,
(2.9)
where
s
2
= E[|x(n)|
2
] is the mean power per sample in the time-domain waveform,
and H
p
(q) and H
n
(q) are the N
c
-point DFTs of the positive and negative tails of the
channel impulse response, respectively. They can be written as

c
c
n c
c
p
N
vq
j
n N
L N v
channel n
N
vq
j
L
p v
channel p
e v h q H
e v h q H
2
1
2
1
) ( ) (
) ( ) (
=
=
.
(2.10)
Moreover, since the linear convolution with the channel can no longer be
considered as one period of a circular convolution, inter-carrier interference (ICI) will
occur within the kth symbol. In Appendix A, we show that the ICI has the same
N
pos

Symbol k+1 Symbol k
h
channel
(t)
t
N
neg
N
pos
N
neg

Wide Area Networks
17
variance as the ISI. Thus, the variance of the total interference on subcarrier q is given
by

+ =

+ =
+ =
+
1
1
2
1
1
2
2 2
) ( ) ( 2 ) (
neg
n
p
pos
N
L n
n
L
N p
p s ICI ISI
q H q H q
.
(2.11)
In order to compute the probability of symbol error, we must also take into
account the power penalty of the cyclic prefix. If we assume that only N
u
of the N
c

subcarriers are used, the probability of symbol error for 4-QAM-modulated subcarriers
can be written as

= +

+
=
1
0
2 2
0
) ( 2
1 1
u
N
q ICI ISI
q
u pre
u
u
s
q
P
N N
N
erfc
N
P

,
(2.12)
where N
u
/(N
u
+N
pre
) is the extra energy wasted on the cyclic prefix samples, P
q
is the
average power per symbol of subcarrier q,
0
2
is the variance of sampled additive
white Gaussian noise (
0
2
= N
0
R), and
2
ISI+ICI
(q) is the variance of the total
interference on subcarrier q.
2.3.2 Simulation Results
We assume that fiber nonlinearity and laser phase noise effects are either
negligible or have been compensated, so the fiber may be modeled as a linear channel.
Therefore, the only impairments in the system are GVD and PMD. In order to
minimize the sampling penalty, at the transmitter, we use a rectangular window, and
perform pulse shaping using a fifth-order Butterworth lowpass filter having a 3-dB
cutoff frequency equal to half the bandwidth of the OFDM signal, given by Eq. (2.6).
At the receiver, the anti-aliasing filter is an identical Butterworth lowpass filter. We
assume transmission of 53.4 Gbit/s in one polarization. The subcarriers are modulated
using 4-QAM and a FEC code with an overhead of 255/239, so the symbol rate is R =
26.7 GHz.
As an initial example, we consider a fiber length of 80 km with a dispersion D
= 17 ps/(nmkm). The FFT size is 64 and the oversampling ratio is M
s
= 1.2, i.e. only
52 subcarriers are used. The minimum required oversampling ratio of M
s
= 1.2 was
Wide Area Networks
18
determined by adding zero subcarriers until noise aliasing became negligible [5]. The
cyclic prefix length, referred to an oversampling ratio M
s
= 1, is 5 samples. A plot of
the symbol-error probability for this example is shown in Fig. 2.9.

Fig. 2.9 Probability of symbol error in presence of GVD for L = 80 km, D = 17 ps/(nm.km), N
c
= 64,
N
u
= 52, N
pre
= 5 and R = 26.7 GHz.OFDM symbol with insufficient cyclic prefix.
In Fig. 2.9, it can be observed that the power penalty increases at small P
s
.
Since the majority of the FEC codes have a threshold around P
s
= 10
3
, we will
measure the power penalty at P
s
= 10
4
in order to have some margin.
Fig. 2.10 and 2.11 illustrate the ISI + ICI penalty and the overall penalty,
respectively, as a function of the cyclic prefix length, holding the data rate constant.
6 7 8 9 10 11 12 13 14 15 16
10
-5
10
-4
10
-3
10
-2
10
-1
E
s
/N
o
(dB)
P
s
Ideal
Cyclic Prefix
Cyclic Prefix + GVD
Simulation
Wide Area Networks
19

Fig. 2.10 ISI + ICI power penalty at P
S
= 10
4
for L = 80 km, D = 17 ps/(nm.km), N
c
= 64, N
u
= 52 and
R = 26.7 GHz.
In Fig. 2.11, we observe that the overall power penalty has two different
regions: ISI + ICI-dominated and cyclic prefix- dominated. In the former, the cyclic
prefix is much shorter than the fiber impulse response and therefore severe ISI + ICI
occur, impairing the system performance

Fig. 2.11 Overall power penalty at P
S
= 10
4
for L = 80 km, D =17 ps/(nm.km), N
c
= 64, N
u
= 52 and R
= 26.7 GHz.
In the latter region, the cyclic prefix is sufficiently long that the ISI + ICI are
negligible, but a large fraction of the transmitted energy is wasted on the cyclic prefix
1 3 5 7 9 11 13 15 17 19 21 23 25 27
1
2
3
4
5
6
7
Prefix Length (N
pre
)
I
S
I

+

I
C
I

P
o
w
e
r

P
e
n
a
l
t
y

(
d
B
)
L = 80 km
Analytical
Simulation
1 3 5 7 9 11 13 15 17 19 21 23 25 27
0
1
2
3
4
5
6
7
8
Prefix Length (N
pre
)
O
v
e
r
a
l
l

P
o
w
e
r

P
e
n
a
l
t
y

(
d
B
)
L = 80 km
Analytical
Simulation
ISI dominated
Prefix dominated
Wide Area Networks
20
samples, leading to a power penalty. For example, for the same scenario as in Fig.
2.11, if one chose a cyclic prefix length such that 98% or 95% of the fibers impulse
response energy was contained in that same duration, one would be required to use a
cyclic prefix length of 14 and 12 samples, respectively, while the optimum cyclic
prefix is 9 samples. The optimum cyclic prefix length results in a penalty from ISI+ICI
equal to the penalty from energy wasted in the cyclic prefix samples.
As explained above, it may be desirable to minimize the number of subcarriers
employed. Hence, we would like to calculate the minimum combination of number of
subcarriers and cyclic prefix that generate a specified power penalty. For concreteness,
we consider a system transmitting at a symbol rate R = 26.7 GHz through fiber spans
having dispersion D = 17 ps/(nmkm), with 98% inline optical dispersion
compensation (i.e., the residual dispersion is 0.34 ps/(nmkm)). Fig. 2.12 shows the
overall power penalties for fiber lengths between 1000 km and 5000 km.

1 2 3 4 5 6
0
0.5
1
1.5
2
2.5
3
Prefix Length (N
pre
)
O
v
e
r
a
l
l

P
o
w
e
r

P
e
n
a
l
t
y

(
d
B
)
(a) L = 1000 km, 2% Uncompensated
N
u
= 12
N
u
= 26
N
u
= 52
N
u
=104
N
u
=208
1 2 3 4 5 6 7 8
0
1
2
3
4
5
6
Prefix Length (N
pre
)
O
v
e
r
a
l
l

P
o
w
e
r

P
e
n
a
l
t
y

(
d
B
)
(b) L = 2000 km, 2% Uncompensated
N
u
= 26
N
u
= 52
N
u
=104
N
u
=208
1 2 3 4 5 6 7 8 9 10 11 12
0
1
2
3
4
5
6
7
Prefix Length (N
pre
)
O
v
e
r
a
l
l

P
o
w
e
r

P
e
n
a
l
t
y

(
d
B
)
(c) L = 3000 km, 2% Uncompensated
N
u
= 26
N
u
= 52
N
u
=104
N
u
=208
1 2 3 4 5 6 7 8 9 10 11 12
0
1
2
3
4
5
6
7
8
Prefix Length (N
pre
)
O
v
e
r
a
l
l

P
o
w
e
r

P
e
n
a
l
t
y

(
d
B
)
(d) L = 4000 km, 2% Uncompensated
N
u
= 52
N
u
=104
N
u
=208
Wide Area Networks
21
Fig. 2.12 Overall power penalties at R = 26.7 GHz for 2% undercompensated GVD lengths of a) 1000
km b) 2000 km c) 3000 km d) 4000 km and e) 5000 km. The dashed lines are the analytical values and
the solid lines are the simulations results with insufficient cyclic prefix.
In Figs. 2.10, 2.11 and 2.12, we observe some discrepancies between the
simulation and theoretical results. Eqs. (2.11) and (2.12) are exact given that there is
no correlation between the OFDM samples. This condition is satisfied only if the
transmitted symbols are uncorrelated and if the oversampling ratio M
s
is equal to 1,
i.e., no oversampling. Since an oversampling ratio M
s
= 1.2 is required to avoid
aliasing (with the Butterworth filters we have used), Eqs. (2.11) and (2.12) are only an
approximation of the ICI + ICI variance, as explained in Appendix A. The error is
maximum when the cyclic prefix length is short since it corresponds to the situation of
highest ISI+ICI. However, we verified that for high ISI/ICI, Eq. (2.11) predicts the ISI
+ ICI variance with a maximum error on the order of 10-15% when M
s
= 1.2.
In Fig. 2.12, we can also observe that in some cases increasing the cyclic prefix
length also increases or maintains the power penalty. This is because increasing the
cyclic prefix length while keeping the data rate constant requires increasing the sample
rate which, in turn, increases the OFDM bandwidth, increasing the temporal spread
caused by GVD. On the other hand, a longer cyclic prefix can compensate longer tails
of the impulse response. It is not obvious which effect will dominate. When the
additional pulse spreading is longer than the increased cyclic prefix length, the overall
power penalty increases.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
0
1
2
3
4
5
6
7
8
9
10
Prefix Length (N
pre
)
O
v
e
r
a
l
l

P
o
w
e
r

P
e
n
a
l
t
y

(
d
B
)
(e) L = 5000 km, 2% Uncompensated
N
u
= 52
N
u
=104
N
u
=208
Wide Area Networks
22
From Fig. 2.12, we can extract the minimum number of subcarriers and cyclic
prefix required to achieve a desired power penalty. The values required to achieve a
penalty of 1 dB are given in Table 2.1 .
Total Length
(km)
Residual
Dispersion
DL.(ps/nm)
DFT Size
N
c

Modulated
Subcarriers
N
u

Cyclic Prefix
N
pre

1000 340 32 26 3
2000 680 64 52 4
3000 1020 64 52 6
4000 1360 128 104 6
5000 1700 128 104 9
Table 2.1 Number of subcarriers and cyclic prefix required to achieve an overall power penalty less
than 1 dB. We consider 4-QAM subcarriers, R = 26.7 GHz, D = 17 ps/(nmkm) with 98% inline optical
dispersion compensation.
Note that by making other choices of these parameters, it is possible to make
the total penalty arbitrarily small. As the uncompensated dispersion becomes larger,
this may involve choosing a very large number of subcarriers, which may be
undesirable, for reasons cited above. In choosing these parameters, one should note
that some synchronization schemes used in practice require that the cyclic prefix
length exceeds a certain fraction of the total number of subcarriers [17], [18]. The
results presented in Table 2.1 are compatible with those synchronization schemes.
Finally, in Fig. 2.13 we plot the optimum number subcarriers and cyclic prefix
length as a function of the fiber length when no optical dispersion compensation is
used.
Wide Area Networks
23

Fig. 2.13 Optimum number of used subcarriers and cyclic prefix length as a function of the fiber length
for R = 26.7 GHz when no optical dispersion compensation is used.
2.4 Polarization-Mode Dispersion
2.4.1 Theory
Fiber asymmetry (birefringence) introduces coupling between the two
degenerate modes in the SMF. Furthermore, due to random nature of the fiber
birefringence, the two degenerate modes arrive at the receiver with different delays
[19]. The delay between the degenerate modes is random and is Maxwellian
distributed [19]. This impairment is called polarization-mode dispersion (PMD) and is
illustrated in Fig. 2.14.
200 600 1000 1400 1800 2200 2600 3000
50
100
150
200
250
300
350
400
450
Length (km)
P
r
e
f
i
x

L
e
n
g
t
h

(
N
p
r
e
)
N
u
= 208
N
u
= 416
N
u
= 832
N
u
= 1664
X Pol.
Y Pol.
Electrical
LPF
t
t
t
X
Y
Z
Wide Area Networks
24
Fig. 2.14 Polarization-mode dispersion (PMD.
The average delay between the degenerate modes is given by
,
where D
p
is the PMD parameter and L is the fiber length. A single-mode fiber with
PMD can be described by a frequency-dependent Jones matrix [19,20,21,22]. Input
and output signals can be described by two-component Jones vectors. The input and
output time-domain signals are denoted by x(t)=[x
1
(t) x
2
(t)]
T
and y(t)=[y
1
(t) y
2
(t)]
T
,
respectively. Their Fourier transforms can be related as

) (
) (
) (
) (
) (
2
1
2
1
X
X
U
Y
Y
,
(2.13)
where U() is a frequency-dependent matrix. There is an analytical expression for
U() only for first-order PMD, i.e., the delay is independent of frequency. For first-
order PMD, the matrix U() is given by

1
1 2
) ( ) (

= R R U , (2.14)
where R
1
and R
2
are frequency-independent rotation matrices representing a change of
basis into the input and output principal states of polarization (PSPs), respectively, and
is a frequency-dependent matrix representing the delay between the fast and slow
PSPs [19]. Explicit expressions for R
i
and [21], [22] are

=
2
2
1 2
2 1
0
0
j
j
*
i i
*
i i
i
e
e

r r
r r
R
,
(2.15)
where

i i i i i
i i i i i
j r
j r

sin cos cos sin
sin sin cos cos
2
1
+ =
=
.
(2.16)
The
i
,
i
are independent random variables representing the fast PSP azimuth
and ellipticity angles, respectively. is a random variable representing differential
group delay (DGD). Considering GVD and PMD, the fiber frequency response is then

L
j
fiber
e U
( H
( H
( H
2
1
2
2 2
) (
)
)
)

=
= ,
(2.17)
Wide Area Networks
25
We will now discuss two approaches to reception in the presence of PMD. The
first approach uses a dual-polarization receiver and the second approach uses a single-
polarization receiver with polarization control.
2.4.2 Dual-Polarization Receiver
A dual-polarization receiver enables electronic polarization control and the use
of polarization-multiplexed signals which can double the bit rate. An OFDM
polarization-multiplexed system is shown in Fig. 2.15.

Fig. 2.15 OFDM polarization-multiplexed system.
The transmitter consists of two independent OFDM modulators. The two
modulated optical signals are combined in orthogonal polarizations using a
polarization beam splitter (PBS). After passing through single-mode fiber, the
received signal is split into two copies, which are mixed with the LO in two
orthogonal polarizations. Each polarization is detected using a 90 hybrid and two
balanced photodetectors. The electrical outputs are the in-phase and quadrature
components associated with the two orthogonal polarizations. These signals are then
low-pass filtered and sampled [16].
The dual-polarization receiver can also be used with a single-polarization
transmitter. In this case, one of the transmitters of Fig. 2.15 would not be used.
In order to compensate for PMD and GVD, the receiver must invert the fiber
frequency response. The equalizer corresponds then to a 22 matrix multiplication
between the signals in the two polarizations on each subcarrier q after the DFT
operation. For first-order PMD, it can be written as
Fiber
Dual
Polarization
Coherent
Receiver
OFDM
Tx 1
OFDM
Tx 2
LO
OFDM
Rx 1
OFDM
Rx 2
I
1
(t)
Q
1
(t)
I
2
(t)
Q
2
(t)
x
ofdm,1
(t)
x
ofdm,2
(t)
I'
1
(t)
Q'
1
(t)
I'
2
(t)
Q'
2
(t)
D/A
D/A
D/A
D/A
A/D
A/D PBS
A/D
A/D
MZ
MZ
Wide Area Networks
26

2 2
2
2
2
1
2
1
1
1
2
) (
Lq
N
f
j
c
s
fiber eq
c
s
e R R
N
q f
H q H

+

=
=
.
(2.18)
In writing down Eq. (2.18), for simplicity, we have not included the pulse-
shaping and anti-aliasing filters. Fig. 2.16 shows the combined equalizer for PMD and
GVD.
Fig. 2.16 Combined PMD and GVD equalizer for an OFDM polarization-multiplexed system.
In the absence of polarization-dependent losses, using a dual-polarization
receiver, PMD causes no loss of information, since PMD is a unitary transformation.
In order to avoid ISI, the cyclic prefix length should be no smaller than the duration of
the channel impulse response. For example, to compensate first-order PMD only (in
the absence of GVD), the cyclic prefix duration should be no shorter than the DGD .
A typical PMD parameter is D
p
= 0.1 ps/km
1/2
and therefore, for a fiber length of 5000
km, the mean DGD is E[] 7 ps. Assuming the symbol rate per polarization is R =
26.7 GHz and assuming 4-QAM-modulated subcarriers, the sample period is about T
c

35 ps for 128 subcarriers (Table 2.1). Assuming it is necessary to compensate a
DGD of five times the mean in order to achieve low outage probability, the maximum
DGD is
max
= 5 E[] 35 ps. A cyclic prefix length of only one sample would be
sufficient. In practice, the overall power penalty would be dominated by GVD, as it
requires a cyclic prefix length of several samples.
An exact analytical form for the Jones matrices describing higher-order PMD
has not yet been developed [21], [22]. However, without loss of generality, the
configuration in Fig. 2.16 could be used to mitigate any order of PMD, provided that
the cyclic prefix is chosen to be sufficiently long. If the complete statistics were
2 2
2
2
2
2
2
Lq
N
f
j q
N
f
j
s s
e

+ +
2 2
2
2
2
2
2
Lq
N
f
j q
N
f
j
s s
e

+
Wide Area Networks
27
known for higher-order PMD, Eq. (2.11) could be used to estimate the ISI and ICI
from the tails of the impulse response not covered by the cyclic prefix, and therefore a
design choice could be done for a desired outage probability. However, we believe
that using a cyclic prefix somewhat longer than five times the mean DGD should be
sufficient to combat almost the entire duration of the impulse response of higher-order
PMD [21].
2.4.3 Single-Polarization Receiver
A single-polarization receiver was already shown in Fig. 2.4. We assume that
polarization control is used such that the LO polarization is locked to the polarization
at the carrier frequency. Since only one polarization is detected, the frequency
response corresponds to one of the rows of the Jones matrix given by Eq. (2.14). It can
be written as

2 / 2 /
) (

j j
PMD
be ae H

+ = , (2.19)
where a and b are complex-value constants from the matrices R
1
and R
2
. As one can
observe, Eq. (2.19) is very similar to multi-path propagation in wireless system. The
probability of symbol error for 4-QAM-modulated subcarriers can be written as

+
=
1
0
2 2
0 ) / 2 (
1
2
1 1
u
N
q
c c PMD
q
u pre
u
u
s
N qf H
P
N N
N
erfc
N
P
,
(2.20)
where N
u
/(N
u
+N
pre
) is the extra energy wasted on the cyclic prefix samples, P
q
is the
average power per symbol of each subcarrier,
0
2
is the AWGN variance and H
PMD
()
is the frequency response given by Eq. (2.19). We note again that the cyclic prefix
length should be equal to the DGD to avoid ISI as can be seen from the impulse
response corresponding to Eq. (2.19), i.e., ( ) ( ) 2 2 ) ( + + = t b t a t h
PMD
. However,
we point out that while the cyclic prefix can be made long enough to avoid ISI up to a
desired outage probability, there is SNR degradation in the single-polarization receiver
due to the frequency-dependent attenuation of the channel.
Fig. 2.17 shows the probability of symbol error for transmission at a symbol
rate R = 26.7 GHz using 4-QAM-modulated subcarriers
Wide Area Networks
28

Fig. 2.17 Probability of symbol error in presence of first-order PMD using a single-polarization
receiver, assuming = 21 ps, a| = |b| = 0.5, N
u
= 52, N
c
= 64, N
pre
= 1 and R = 26.7 GHz.
The DGD is = 21 ps, i.e., 3 times the mean of a link with 5000 km with D
p
=
0.1 ps/km
1/2
. As a worst case, equal power splitting between the PSPs (|a| = |b| = 0.5)
is assumed. The error probability computed using Eq. (2.20) is in good agreement with
the simulation results.
2.5 Fiber Nonlinearity
2.5.1 Theory
Nonlinear impairments in fiber arise from the Kerr effect, where the refractive
index of the transmission medium changes with intensity of signal (i.e., the square
magnitude of the applied electric field). Although the Kerr effect is extremely small in
silica, the confinement of light in the core of single mode fiber is such that even at
realistic transmit powers, nonlinear effects can be problematic in long haul systems.
Nonlinearity is ultimately a limiting factor to the theoretical capacity of optical fiber.
In single-polarization transmission where PMD is negligible, signal
propagation is described by a scalar nonlinear Schrdinger equation (NLSE) along the
reference polarization of the fiber
6 7 8 9 10 11 12 13 14 15 16
10
-5
10
-4
10
-3
10
-2
10
-1
E
s
/N
o
(dB)
P
s
Ideal
Cyclic Prefix
Cyclic Prefix + PMD
Simulation
Wide Area Networks
29
E E j E
t
E j
z
E
2
2
2
2
| |
2 2

= +
, (2.21)
where ( ) t z E , is the electric field, is the attenuation coefficient,
2
is the
dispersion parameter, is the nonlinear coefficient, and z and t are the propagation
direction and time, respectively.
Nonlinear effects have deterministic and statistical components. The
nonlinearity experienced by a signal due to its own intensity is known as self-phase
modulation (SPM). In WDM systems, a signal also suffers from nonlinear effects due
to neighboring channels. These are known as cross-phase modulation (XPM) and four-
wave-mixing (FWM) [14], and their impact can be reduced by using non-zero
dispersion fiber which causes pulse walkoff [23]. In the absence of ASE noise, all
nonlinear effects are deterministic, so it is possible in principle to pre-compensate
them at the transmitter [14]. It is also possible to perform multi-channel detection at
the receiver but the system complexity can be prohibitive.
In amplified long-haul systems, interaction between amplifier noise (i.e., ASE
noise) and signal through the Kerr nonlinearity leads to nonlinear phase noise (NLPN)
[24]. The phase noise arising from the interaction of ASE noise with the channel of
interest is called SPM-induced NLPN, while the phase noise arising from the
interaction of ASE noise with the signals of neighboring channels is called XPM-
induced NLPN [14]. Since the nonlinear interactions between signal and noise are
non-deterministic, they cannot be perfectly compensated. Fig. 2.18 shows the
interaction between ASE noise and signal through the fiber nonlinearity when there is
no fiber dispersion.
Re
Im
a)
NL
Re
Im
b)
Wide Area Networks
30
Fig. 2.18 Impact of nonlinearity on the transmission of BPSK: (a) without nonlinearity, (b) with
nonlinearity.
In Fig. 2.18, we observe that the received signal in the presence of nonlinearity
has a spiral-shaped constellation. The spiraling arises because points that are further
away from the origin (i.e., have a higher instantaneous intensity) experience greater
nonlinear phase shift as a result of the Kerr effect.
2.5.2 Simulation Results
Since ASE noise and signal interact through the fiber nonlinearity, the received
nonlinear noise is no longer Gaussian distributed and therefore the conventional SNR
metric cannot be used to evaluate the system performance. A good performance metric
for single-carrier systems when fiber nonlinear effects are present is the phase noise
variance, i.e., ( )
NL
var [14].
When only linear effects are present in the fiber, the phase noise variance is
approximately equal to ( )
SNR 2
1
var
NL
for high SNR. For multi-carrier systems, the
geometric SNR was shown to be good predictor of the performance as a whole [6].
Having these figures in mind, we propose the geometric phase noise variance to
quantify the OFDM performance in the presence of nonlinearity. The geometric phase
noise variance is given by
( ) ( )
u
u
NL NL
N
N
n
n OFDM
/ 1
1
var var

=

=
, (2.22)
where N
u
is the number of used subcarriers.
In this study, we consider a single-wavelength system transmitting over
multiple spans with inline optical amplification and dispersion compensation. We
consider a polarization-multiplexed QPSK system transmitting at 100 Gbit/s; after
inclusion of forward error-correction (FEC) coding with 7% overhead, the symbol rate
is R =26.7 Gsym/s. The FEC limit is at a bit-error ratio BER = 210
-3
, corresponding
to a phase noise variance of 0.08 rad
2
for QPSK systems. For OFDM transmission, we
Wide Area Networks
31
optimize the number of subcarriers and cyclic prefix for each dispersion map
according to section 2.3.
The total amplifier gain is equally distributed between the two amplifiers and
is set to exactly compensate the loss incurred in the SMF and DCF. Each amplifier has
a spontaneous emission factor n
sp
= 1.41. The parameter values for the SMF and DCF
are typical of terrestrial links: L
SMF
= 80 km,
SMF
= 0.25 dB/km, D
SMF
= 17 ps/nmkm,
SMF
= 1.2 W
1
/km,
DCF
= 0.6 dB/km, D
DCF
= 85 ps/nmkm and
DCF
= 5.3 W
1
/km.
The amount of optical dispersion compensation for each dispersion map is determined
by the length of the DCF. At the receiver, DSP-based equalization is implemented for
residual GVD compensation after passing through a 5th order Butterworth lowpass
filter. In single-carrier transmission, the received signal is oversampled by a factor of
2, and is processed by a MMSE linear equalizer. In OFDM transmission, the received
signal is oversampled by a factor close to M
s
1.2., and is processed by a single-tap
equalizer on each subcarrier.
We evaluate system performance by numerical solution of the nonlinear
Schrdinger equation with random noise sources, and we quantify system performance
by phase noise variance. We ignore laser phase noise and PMD. For simplicity, we
perform simulations using single-polarization transmitters and receivers, modeling
propagation in only one polarization. While we neglect the nonlinearity induced by the
orthogonal polarization, it should affect single-carrier and OFDM in a similar fashion,
so including it should not change the conclusions. Considering only one polarization,
in single-carrier systems, nonlinear impairments include intra-channel four-wave
mixing (IFWM) or nonlinear phase noise or a combination of the two, depending on
the dispersion map. In OFDM systems, in the absence of GVD, the impact of
nonlinearity can be understood as four-wave mixing (FWM) between the different
subcarriers, i.e., intermodulation between the subcarriers. In order to compare the
performance of single-carrier and OFDM, we use for OFDM the geometric phase
noise variance given by Eq. (2.22).
Fig. 2.19 a) shows the phase variance comparison for the case of 100%
dispersion compensation per span for 20 spans of propagation.
Wide Area Networks
32
Fig. 2.19 Phase noise variance for OFDM and SC transmissions with 100% dispersion compensation
per span: a) for 20 spans, as a function of launched power, and b) as function of the number of spans.
In the linear regime where signal powers are low, system performance is ASE
noise-limited for both modulation formats and OFDM performs slightly worse than
single-carrier. This is due to the power penalty incurred in the cyclic prefix [25].
However, this penalty can be made arbitrarily small, so the performance becomes
equal to that of single-carrier, by increasing the number of subcarriers. In the high-
signal-power regime, where system performance is nonlinearity-limited, we observe
that single-carrier has significant lower phase noise variance than OFDM. A lower
phase noise variance means that the system can achieve longer distances. Fig. 2.19 b)
shows the minimum phase variance comparison for the case of 100% dispersion
compensation per span as a function of the number of spans. We observe that single-
carrier has always a smaller phase variance than OFDM, and single-carrier can
achieve 52 spans before reaching the FEC threshold, while OFDM is only able to
achieve 33 spans.
The low performance of OFDM for the case of 100% dispersion compensation
is because of the high number of FWM products. For 100% dispersion compensation,
the accumulate phase mismatch between the subcarriers is small and therefore the
FWM efficiency is high [14]. In order to reduce the FWM efficiency, we can
deliberately introduce phase mismatch between the subcarriers by, for example, not
fully compensating the SMF dispersion after each span.
-10 -8 -6 -4 -2 0 2 4
10
-2
10
-1
10
0
Power (dBm)
P
h
a
s
e

n
o
i
s
e

v
a
r
i
a
n
c
e

(
r
a
d
)
2
FEC limit
OFDM SC (4-QAM)
Linear
Nonlinear
(a)
0 10 20 30 40 50 60
10
-4
10
-3
10
-2
10
-1
10
0
Number of Spans
FEC limit
OFDM
SC (4-QAM)
P
h
a
s
e

n
o
i
s
e

v
a
r
i
a
n
c
e

(
r
a
d

)
2(b)
Wide Area Networks
33
Fig. 2.20 a) and b) compare OFDM and single-carrier for 20 spans of
propagation for the cases of 90% and 0% dispersion compensation per span,
respectively.
Fig. 2.20 Phase noise variance for OFDM and SC transmissions for 20 spans, as a function of launched
power: a) 90% dispersion compensation per span and b) 0% dispersion compensation per span.
We note that for the case of 0% compensation, there is no DCF and there is
only one amplifier per span. The small performance difference between OFDM and
SC in the linear regime is again attributed to the cyclic prefix penalty, as described
above. However, in the nonlinearity-limited regime, the performance gap between
OFDM and SC for 90% and 0% is smaller than that of 100% dispersion compensation
per span. This can be understood by the incoherent additions and partial cancelations
of the FWM products from one span to another in the presence of residual dispersion
in each span. Furthermore, we observe that for 0% compensation per span, OFDM and
single-carrier have very similar performance in the nonlinear regime.
2.6 Computational Complexity
For OFDM, the IDFT and DFT operations are performed efficiently using a
fast Fourier transform (FFT) algorithm. An FFT of size N requires 4Nlog
2
(N) real
operations (multiplications plus additions) [26]. Specifically, the number of additions
is 8/3Nlog
2
(N) and the number of multiplications is 4/3Nlog
2
(N). Thus, the number
of real operations required per second for the transmitter is 4Nlog
2
(N)/T
OFDM
, where
T
OFDM
is the OFDM symbol period, which is given in [25]. For the OFDM receiver,
we need to take into account the complex single-tap equalizer on each used subcarrier.
(a)
OFDM
SC (4-QAM)
-10 -8 -6 -4 -2 0 2 4
10
-2
10
-1
10
0
Power (dBm)
FEC limit
P
h
a
s
e

n
o
i
s
e

v
a
r
i
a
n
c
e

(
r
a
d

)
2
-10 -8 -6 -4 -2 0 2 4
10
-2
10
-1
10
0
Power (dBm)
FEC limit
OFDM
SC (4-QAM)
P
h
a
s
e

n
o
i
s
e

v
a
r
i
a
n
c
e

(
r
a
d

)
2(b)
Wide Area Networks
34
Assuming that complex multiplications are implemented with the usual 4 real
multiplications and 2 real additions) [26], the number of real operations required per
second for the receiver is (4Nlog
2
(N) + 6N
u
)/T
OFDM
, where N
u
is the number of used
subcarriers. The overall complexity order in real operations per bit for OFDM is

s
OFDM
s b s
OFDM b
Tx
OFDM
M N
R T M N N
T T N N N O
+
=
=
) ( log 4
/ ) ( log 4 ) (
2
2
.
[ ]
[ ]
s
OFDM
s b s u
OFDM b u
Rx
OFDM
M N
R T M N N N
T T N N N N O
+
+
=
+ =
6 ) ( log 4
/ 6 ) ( log 4 ) (
2
2

(2.23)
For single-carrier (M-QAM), the complexity of directly implementing the
convolution operation for the feedforward MMSE equalizer is 2N
taps
/T
s
complex
operations (multiplications plus additions) per second, where N
taps
is the number of
taps of the feedforward filter and T
s
is the symbol period (T
s
= 1/R
s
= T
b
log
2
M). The
equalizer computes an output once per symbol interval. The overall complexity order
in real operations per bit for single-carrier (M-QAM) in a direct implementation is

[ ]
M N
T T N
/T T N N N O
taps
s b taps
s b taps taps
D
QAM
2
log / 8
/ 8
4 4 ) (
=
=
+ =
.
(2.24)
For a large number of taps, linear equalization is more efficiently implemented
using an FFT-based block processing such as overlap-add or overlap-save [4].
Suppose a block length of L
B
is chosen. An FFT-based implementation has a
complexity of (N
taps
+ L
B
1)(42log
2
(N
taps
+ L
B
1) + 6) real operations per block
L
B
for an equalizer with N
taps
taps. For a given N
taps
, there exists an optimum block
length L
B
that minimizes the number of operations required. We assumed that the
equalizer taps come from a lookup table since the SMF impulse response typically
changes slowly and therefore the same equalizer taps can be used for many symbols.
The overall complexity order in real operations per bit for M-QAM with a MMSE
feedforward equalizer in a FFT implementation is
Wide Area Networks
35

M B
L N L N
N O
f
B taps B taps FFT
QAM
2
2
log
6) ) 1 ( log (8 ) 1 (
) (
+ + +
= .
(2.25)
We compute the number of operations required per bit for single-carrier and
OFDM in Table 2.2 for same system parameters as in section 2.3, i.e., 4-QAM
subcarriers, R = 26.7 GHz, D = 17 ps/(nmkm) with 98% inline optical dispersion
compensation. For single-carrier, we calculate the complexity of direct
implementation and FFT based implementation using an optimized block size L
B
subject to having a FFT size that is an integer power of two. In Table 2.2, we observe
that an OFDM system requires two digital signal processors (DSPs) with
approximately 40% less computational complexity than the single DSP required for
single-carrier. On the other hand, an OFDM system requires an A/D and D/A while
single-carrier requires only one A/D.
Transmission
Distance
(km)
Single-Carrier OFDM
Direct
FFT
Transmitter Receiver Block Size
(L
B
)
Complexity
1,000 48.0 6 53.3 24.6 30.6
2,000 96.0 27 66.4 29.5 35.5
3,000 128.0 25 71.6 29.5 35.5
5,000 208.0 52 78.8 34.5 40.4
Table 2.2 Computational complexity in real operations per bit for the various modulation formats. We
consider 4-QAM subcarriers, R = 26.7 GHz, D = 17 ps/(nmkm) with 98% inline optical dispersion
compensation.
2.7 Optical Modulator
Multi-carrier signals like OFDM have a high peak-to-average ratio (PAR),
which is proportional to the number of used subcarriers [5]. Optical OFDM is
modulated using Mach-Zehnder (MZ) modulators having nonlinear, peak-limited
transfer characteristics. Experiments have been performed demonstrating the potential
of OFDM with coherent detection in optical systems [27,9,10]. Some previous work
has investigated the possibility of pre-distortion and clipping to mitigate the effects of
MZ nonlinearity on OFDM system performance [10], [28]. Our study extends that
previous work by considering the frequency-dependent MZ electrode losses, by
Wide Area Networks
36
quantifying optical power efficiency, and by introducing the dual-drive MZ as an
alternative option for the OFDM modulator.
There are several options to generate an optical OFDM signal. A popular
technique shifts the electrical OFDM signal to an intermediate frequency (IF) and then
uses it to drive a single MZ modulator [27,9,10]. Another option is to use a quadrature
MZ. A quadrature MZ modulates directly the baseband electrical OFDM signal to the
optical domain without the need of an IF. Hence, this technique, sometimes called
direct conversion [10], reduces the required electrical bandwidth by a factor of two, at
the expense of a more complicated modulator design.
A conventional solution to the PAR problem is to reduce the operating range in
the MZ to accommodate the OFDM peak. However, this solution results in a
significant power efficiency penalty, which may require the use of optical
amplification at the transmitter to boost the signal level. An alternative solution would
be using peak-reduction algorithms studied for wireless systems [5]. All of these peak-
reduction algorithms, however, lead to undesired effects, such as increased coding
overhead and/or increased average transmitted power [29,30,31,13]. An increased
coding overhead requires an increased sampling rate in order to maintain the desired
bit rate, exacerbating the impact of GVD and PMD, and requiring faster D/A
converters. An increased average transmitted power can lead to increased nonlinear
impairment; in [11], it was shown that the variance of phase noise caused by four-
wave mixing is proportional to the power in each subcarrier, but is not simply related
to the instantaneous peak power. Furthermore, all of the peak-reduction algorithms
present a computational burden to the transmitter [13], which might be prohibitive at
the speeds of optical systems.
In this section, we present hard clipping with pre-distortion as a simple and
effective approach to combat the nonlinearity in the quadrature MZ and increase the
optical power efficiency. Since the OFDM peaks occur with a very low probability,
clipping can be an effective technique. In addition, we study the combined effects of
having a finite number of bits in the D/A and the MZ nonlinearity. We then extend our
Wide Area Networks
37
study to the dual-drive MZ, which proves to require a higher oversampling ratio to
achieve the same performance as the quadrature MZ.
2.7.1 PAR and MZ Review
2.7.1.1 Peak-to-Average Power Ratio
The PAR is defined as [5]

[ ]
2
2
) (
| ) ( | max
PAR
t x E
t x
t
= ,
(2.26)
where
2
| ) ( | max t x
t
is the peak value squared of the signal x(t) and E[x(t)
2
] is the
average signal power. The peak of the signal determines the dynamic range required
of the D/As and modulators in the circuit. Thus, it is desirable to have signals with low
PAR. In the case of OFDM, i.e., identical constellations on all of the subcarriers, the
PAR can be written as [5], [13]

[ ]
2
2
| | max
PAR
k
k
k
u
X E
X
N = ,
(2.27)
where
2
| | max
k
k
X is the largest symbol magnitude squared on the k
th
subcarrier, E[X
k
2
]
is the average power per symbol on subcarrier k and N
u
subcarriers. In particular, if all the subcarriers are modulated using QPSK, Eq. (2.27)
becomes

u
N = PAR .
(2.28)
2.7.1.2 Mach-Zehnder Modulator
If complementary drive signals are used, the transfer characteristic of a single-
drive MZ modulator is [14]

V
t V
t E
t E
m
in
out
2
) (
sin
) (
) (
,
(2.29)
where E
out
(t) and E
in
(t) are the output and input electric fields, respectively, V
m
(t) is the
electrical modulator signal and V
is the voltage that must be applied to the single

Wide Area Networks
38
electrode to produce a differential phase shift of between the two waveguides. The
modulator electrode has frequency-dependent loss, so that high-frequency components
of the drive signal are attenuated. The electrode frequency response can be modeled
by [14]

int 12 int
int 12 int
2 ) (
) 2 ) ( exp( 1
) (
) (
L fd j L f
L fd j L f
t V
t V
d
m

+
= ,
(2.30)
where V
d
is the input drive voltage, V
m
is the modulating voltage of the MZ transfer
characteristic, (f) is the frequency-dependent loss, d
12
is the velocity mismatch
difference between the optical and electrical waveguides and L
int
is the interaction
length. The model for single-drive MZ modulator is shown in Fig. 2.21.
Fig. 2.21 Model for single-drive MZ modulator, corresponding to one phase of a quadrature MZ
modulator.
Two single-drive MZ modulators can be combined to create a quadrature MZ.
The quadrature MZ comprises two single-drive MZs, whose output optical fields are
added in quadrature. The quadrature MZ is useful to modulate a complex baseband
signal directly to the optical domain without the need of shifting the signal to an
IF. Fig. 2.22 shows the block diagram of a quadrature MZ.
Fig. 2.22 Quadrature MZ modulator.
Laser
I
Q
90
Quadrature MZ
Fiber
MZ
MZ
v
d,i
(t)
v
d,r
(t)
Wide Area Networks
39
For the remainder of the chapter, in block diagrams, we use single lines to
represent electrical signals and double lines to represent optical signals.
2.7.2 Quadrature MZ Optimization
2.7.2.1 Quadrature MZ Optimization
Fig. 2.23 shows an optical OFDM modulator using a quadrature MZ.

Fig. 2.23 OFDM transmitter using quadrature MZ modulator.
In Fig. 2.23, the transmitted symbols are modulated using the inverse discrete
Fourier transform (IDFT) and the cyclic prefix is added to the signal. After D/A
conversion, the real and imaginary components are used to drive the MZ modulators.
Then, one of the MZ outputs is delayed by 90
o
in order to generate the optical in-phase
(I) and quadrature (Q) components. These are then summed and injected into the fiber.
We note that in Fig. 2.23 not all the N
c
subcarriers are used to transmit data.
Some of the subcarriers are used for oversampling [25]. The oversampling ratio is
defined as M
s
= N
c
/N
u
, where N
c
is the IDFT size and N
u
subcarriers [25]. Furthermore, the used subcarriers are properly positioned within the
DFT block such that the resulting OFDM spectrum is centered and, typically, the D.C.
subcarrier is not used for data transmission.
The MZ modulators in Fig. 2.23 can introduce two impairments: low optical
power efficiency and intermodulation products between subcarriers. In order to
quantify the first impairment, we define OPE as

Power Laser CW
Power Modulated Average
OPE = , (2.31)
Laser
X
1
IDFT
Parallel
-to-
Serial
+
Cyclic
Prefix
N
X
N
D/A
Im{ }
D/A
Re{ }
I
Q
90
Quadrature MZ
Fiber
MZ
MZ
v
d,i
(t)
v
d,r
(t)
Wide Area Networks
40
A backoff in the MZ operating range is required to accommodate the peak of
any drive signal. Since the OFDM signals have a high PAR, the OPE can be very low.
For example, if N
c
and M
s
= 1.2, i.e., 52 used subcarriers, the OPE is less than 1%
(Fig. 2.28).
The second impairment results from the MZ nonlinear transfer characteristic.
Expanding Eq. (2.29) into a Taylor series, we get
+
=
3
2
) (
2
) (
2
) (
sin
) (
) (

V
t V
V
t V
V
t V
t E
t E
m m m
in
out
(2.32)
The cubic term in Eq. (2.32) will generate in-band intermodulation products
between the subcarriers, which cannot be removed by filtering.
In order to mitigate the previous impairments, we will use pre-distortion to
compensate the MZ nonlinearity and hard clipping to increase the OPE. The clipping
operation is described as

>
< <
<
=
A n x A
A n x A n x
A n x A
n V
in
) ( if ,
) ( if ), (
) ( if ,
) (
(2.33)
where V
in
(n) is the clipped signal at sample time n, and A is the clipping level. The
clipping operation is done separately on each of the real and imaginary components of
the OFDM signal. Furthermore, we will use a super-Gaussian filter (SGF) at the MZ
output in order to eliminate any out of band leakage generated by hard clipping (in a
practical WDM system, this filtering function would be performed by the multiplexer).
The pre-distortion function is given by

=
| ) ( | max
) (
arcsin
2
) (
n V
n V V
n V
in
n
in
out
,
(2.34)
where V
in
(n) and V
out
(n) are the digital input and output voltages of the pre-distortion
device. Note that A n V
in
n
= | ) ( | max because of the clipping operation and such that the
output signal V
out
(n) swing is always between V
and +V
due to the mapping of the

pre-distortion curve. The pre-distortion transfer characteristic is shown in Fig. 2.24.
Wide Area Networks
41

Fig. 2.24 Pre-distortion transfer characteristic for quadrature MZ modulator.
In addition to the pre-distortion curve, we need to compensate the electrode
frequency response of each MZ. We note that if the electrode frequency response is
compensated, then V
m
(n) = V
out
(n). The optimized OFDM transmitter is shown in Fig.
2.25 and the corresponding waveforms are shown in Fig. 2.26.
Fig. 2.25 OFDM transmitter including hard clipping, pre-distortion and electrode frequency response
compensation.
In Fig. 2.26, we note that the peak after the D/A can exceed the value limited
by the digital hard clipping device because of interpolation effects [13], [32]. In order
to avoid an analog clipping device, we will allow the driving signal to exceed V
. Later
on, we will show that this effect has negligible impact on system performance.
-1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1
0
V
in
(t)
V
o
u
t
(
t
)
-V
Hard Clipper
CL
CL CL
CL
sin
1
( )
V
CL CL
V
Hard Clipper
CL
CL CL
CL
sin
1
( )
V
CL CL
V
D/A
X
1
OFDM
TX
X
N
Re{ }
Im{ }
H
elec.
( )
1
f
B
H
elec.
( )
1
f
B
D/A
Quadrature
MZ
Laser
Clipping Pre-Distortion Equalization
V
d,i
(t)
V
d,r
(t)
V
d,r
(n)
V
out,r
(n) V
in,r
(n)
V
out,i
(n) V
in,i
(n)
V
d,i
(n)
Wide Area Networks
42

Fig. 2.26 Electrical and optical MZ waveforms. The solid and dotted lines represent the waveforms with
and without hard clipping, respectively.
2.7.2.2 Clipping Simulation Results
In order to make the simulation results independent of the number of
subcarriers, we will define a normalized clipping level called the clipping ratio (CR)
[33], which is defined as

A
= CR , (2.35)
where A is the clipping level and is the rms power of the OFDM signal, i.e.,
[ ]
2
) ( CR t x E = . For example, if there is no clipping, then PAR CR = . A value CR
= 2 means that the signal is clipped at twice the rms power level. We note that the
PAR of the real and imaginary components is twice the PAR of the complex baseband
OFDM signal, since the real and imaginary parts have half the power but the same
peak value as the complex baseband OFDM signal.
In our simulations, we consider a polarization-multiplexed system with a total
bit rate of 118 Gbit/s (103 Gbit/s with 15% FEC overhead). The subcarriers are
modulated using QPSK, so the symbol rate on each polarization is 29.66 GHz. The
FFT size is 64 and the oversampling ratio is M
s
= 1.2, i.e. only 52 subcarriers are used.
The cyclic prefix is chosen accordingly to [25], and is approximately 1/10 of the total
Wide Area Networks
43
number of subcarriers. We assume for now that the D/A has an infinite number of bits.
We neglect all transmission impairments, such as fiber nonlinearity, GVD and PMD.
We use a homodyne receiver, followed by an anti-aliasing filter that is a 5th-order
Butterworth lowpass filter having a 3-dB bandwidth equal to 17.2 GHz, the first null
in the baseband OFDM spectrum [25]. The 3-dB bandwidth of the SGF is set equal to
the OFDM bandwidth [25]. The MZ 3-dB bandwidth is 30 GHz, which is
representative of currently available devices [34]. The MZ frequency response is
shown in Fig. 2.27.

Fig. 2.27 MZ electrode frequency response.
We note that the compensation of the MZ frequency response shown in Fig.
2.25 requires one extra DFT and IDFT operation to scale the subcarriers by the inverse
of the MZ electrode frequency response. We consider three scenarios for the
modulator: Full Compensation where we compensate the MZ electrode frequency
response as shown in Fig. 2.25, Pre-Compensation where we move the electrode
frequency response compensation to the OFDM transmitter as a pre-distortion
operation, i.e., we are swapping the order of the nonlinear and linear effects and No
Compensation where we do not compensate the MZ electrode frequency response.
The simulation results are shown in Fig. 2.28 and Fig. 2.29.
0 10 20 30 40 50
-9
-8
-7
-6
-5
-4
-3
-2
-1
0
Frequency (GHz)
2
0

l
o
g
1
0
(
|
V
m
(
f
)
/
V
d
(
f
)
|
)

(
d
B
)
Wide Area Networks
44

Fig. 2.28 Optical power efficiency for N
c
= 64, N
u
= 52 and R = 29.66 GHz.
Since typical current FEC codes have a threshold around P
S
= 10
3
, we
measure the receiver sensitivity penalty at P
S
= 10
4
in order to have some margin.
In Fig. 2.28, we observe that the OPE is around 1% if there is no clipping CR =
10.2). As we start clipping, the power efficiency increases significantly. Moreover, we
can also observe that the three compensation options have approximately the same
OPE. For the case of no compensation, the OPE is slightly inferior, due to the
frequency dependence of the MZ electrode frequency response.

Fig. 2.29 Receiver sensitivity penalty at P
S
= 10
4
, N
c
= 64, N
u
= 52 and R = 29.66 GHz.
1 2 3 4 5 6 7 8 9 10 11
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
O
p
t
i
c
a
l

P
o
w
e
r

E
f
f
i
c
i
e
n
c
y

(
W
/
W
)
Clipping Ratio (CR)

Full Comp.
Pre-Comp.
No Comp.
1 2 3 4 5 6 7 8 9 10 11
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
R
e
c
e
i
v
e
r

S
e
n
s
i
t
i
v
i
t
y

P
e
n
a
l
t
y

(
d
B
)
Clipping Ratio (CR)

Full Comp.
Pre-Comp.
No Comp.
Wide Area Networks
45
In Fig. 2.29, we observe that the power penalty is approximately constant
down to about CR = 2.5, below which, the signal becomes severely distorted and the
receiver sensitivity penalty increases rapidly. Furthermore, we observe that the three
compensation scenarios have approximately the same performance. For the case of no
compensation, we note that there is a residual receiver sensitivity penalty even if there
is no clipping. The residual receiver sensitivity penalty is due to the power lost at
some frequencies in the MZ. By going from CR = 10.2 (no clipping) down to CR =
2.5, the OPE increases from less than 1% to 14%, a 12-dB power gain at the cost of a
0.5-dB sensitivity penalty.
Finally, in order to confirm that the CR makes the optimized clipping level
independent of the number of subcarriers, we repeated the simulations with a different
number of subcarriers. Fig. 2.30 and Fig. 2.31 show the OPE and system performance
curves for different numbers of subcarriers for the case of no compensation of the MZ
electrode frequency response.

Fig. 2.30 Optical power efficiency for different number of subcarriers.

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
O
p
t
i
c
a
l

P
o
w
e
r

E
f
f
i
c
i
e
n
c
y

(
W
/
W
)
Clipping Ratio (CR)

N
u
= 26
N
u
= 52
N
u
= 104
Wide Area Networks
46

Fig. 2.31 Receiver sensitivity penalty for different number of subcarriers.
2.7.2.3 Quantization Effects
For the remainder of the chapter, we consider only the situation where the
electrode frequency response is not compensated, since we verified that this
compensation has little effect on OPE and system performance. In addition, not
compensating the electrode frequency response minimizes hardware complexity and
power consumption.
In a practical OFDM transmitter, the D/A necessarily will have a finite number
of bits. Fig. 2.32 and Fig. 2.33 show the OPE and system performance curves for
different number of bits in the D/A.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
0
0.5
1
1.5
2
2.5
3
3.5
4
R
e
c
e
i
v
e
r

S
e
n
s
i
t
i
v
i
t
y

P
e
n
a
l
t
y

(
d
B
)
Clipping Ratio (CR)

N
u
= 26
N
u
= 52
N
u
= 104
Wide Area Networks
47

Fig. 2.32 Optical power efficiency for different number of bits in the D/A, N
c
= 64, N
u
= 52 and R =
29.66 GHz.

Fig. 2.33 Receiver sensitivity penalty for different number of bits in the D/A, N
c
= 64, N
u
= 52 and R =
29.66 GHz.
We note that the D/A reference voltage is always equal to V
regardless of the
CR, since the pre-distortion device maps the clipped peak value to V
.
In Fig. 2.32, we observe that the OPE is independent of the number of bits in
the D/A. On the other hand, in Fig. 2.33 we observe that for a fixed number of bits, the
receiver sensitivity penalty decreases as we further clip the signal until around CR =
2.5. This is because the D/A with a limited number of quantization levels can better
represent the OFDM signal as the peak is further clipped. We also observe that as we
1 2 3 4 5 6 7 8 9 10 11
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
O
p
t
i
c
a
l

P
o
w
e
r

E
f
f
i
c
i
e
n
c
y

(
W
/
W
)
Clipping Ratio (CR)

n = 4 bits
n = 5 bits
n = 6 bits
n = 7 bits
n = 8 bits
1 2 3 4 5 6 7 8 9 10 11
0
0.5
1
1.5
2
2.5
3
3.5
4
R
e
c
e
i
v
e
r

S
e
n
s
i
t
i
v
i
t
y

P
e
n
a
l
t
y

(
d
B
)
Clipping Ratio (CR)

n = 4 bits
n = 5 bits
n = 6 bits
n = 7 bits
n = 8 bits
Wide Area Networks
48
increase the number of bits, the receiver sensitivity penalty decreases and converges to
the case of infinite-resolution value, as expected. If a value CR = 2.5 is used, 6 bits
would be sufficient to obtain good performance. Finally, we note that although we
considered N
u
= 52 in Fig. 2.32 and Fig. 2.33, the receiver sensitivity penalty due to
clipping and quantization is independent of the number of subcarriers for a given value
of CR.
2.7.3 Dual-Drive MZ Optimization
2.7.3.1 Dual-Drive MZ Modulator
We have shown that the quadrature MZ can be used as an efficient OFDM
transmitter. In an attempt to reduce complexity, we consider using a dual-drive MZ as
an OFDM transmitter. The dual-drive MZ has two independent drive electrodes, and
has transfer characteristic given by

=

V
t V
j
V
t V
j
t E
t E
m m
in
out
) (
exp
) (
exp
2
1
) (
) (
2 1
,
(2.36)
where E
out
(t) and E
in
(t) are the output and input electric fields, respectively, V
m1
(t) and
V
m2
(t) are independent electrical modulator signals and V
is the drive voltage that

must be applied differentially between the two electrodes to produce a differential
phase shift of between the two waveguides. The model for the dual-drive MZ is
shown in Fig. 2.34.
+ =
(t) j (t) j
m2
V
m1
V
e e
2
1
(t) E
(t) E
in
out
Wide Area Networks
49
Fig. 2.34 Model for dual-drive MZ modulator.
The frequency response of each electrode in Fig. 2.34 is given by Eq. (2.30). In
the dual-drive MZ, by adding together two independent signals whose phases are
proportional to the two independent drive voltages, we can generate an output electric
field having arbitrary magnitude and phase. In other words, we can generate any point
inside the unit circle of the complex plane E
ou
(t)/E
in
(t), as shown in Fig. 2.35.
Fig. 2.35 Dual-drive MZ plane.
The drive voltages required to generate a point x(n) = A(n)e
j(n)
(0 A(n) 1)
are given by

( ) ( )
( ) ( ) q n A n
V
n V
q n A n
V
n V
out
out

2 ) ( arccos ) ( ) (
2 ) ( arccos ) ( ) (
2
1
+ =
+ + =

(2.37)
We note that due to the symmetry of the problem, the Eqs. for V
out1
(n) and
V
out2
(n) are periodic with period 2V
and that they could be interchanged and the point

x(n) would still be obtained. We note that if the frequency response of each electrode
is compensated, then V
m1
(n) = V
out1
(n) and V
m2
(n) = V
out2
(n).
As for the quadrature MZ modulator, we use hard clipping in the dual-drive
MZ to increase the OPE. The optimized OFDM transmitter using the dual-drive MZ is
shown in Fig. 2.36.
1
x(n)
(t) E
(t) E
Im
in
out
) (
2
2
1
n V
V
j
m
(t) E
(t) E
Re
in
out
) (
1
2
1
n V
V
j
m
e
Wide Area Networks
50
Fig. 2.36 OFDM transmitter using the dual-drive MZ with hard-clipping, pre-distortion and electrode
frequency response compensation.
Besides reduced hardware complexity, another advantage of the dual-drive MZ
over the quadrature MZ is a reduced clipping probability for the same clipping level.
As shown in Fig. 2.37, the dual-drive MZ can span the indicated circle in the complex
plane E
ou
(t)/E
in
(t) for a given clipping level, while the quadrature MZ can only span
the indicated square.
Fig. 2.37 OFDM span regions for a fixed clipping level. The dotted and solid regions correspond to the
quadrature MZ and the dual-drive MZ, respectively.
If the drive signal samples V
m1
(n) and V
m2
(n) are digitally pre-distorted, the
MZ will generate the correct optical electric field at the sampling instants. However,
the drive voltages V
m1
(t) and V
m2
(t) between samples are obtained by interpolation in
the D/A converter, and they are not generally the values required to generate the
correct optical electric field in between the sampling instants. As shown in Fig. 2.38,
the optical waveform will differ from the correct OFDM waveform, which can
degrade system performance. This problem can be controlled by increasing the
oversampling ratio.
D/A
X
1
IDFT
Parallel
-to-
Serial
+
Cyclic
Prefix
N
X
N
Re{ }
Im{ }
H
elec.
( )
1
f
B
H
elec.
( )
1
f
B
D/A
Dual-Drive
MZ
Laser
Hard-Clipping
+
Dual-Drive
Pre-Distortion
V
d1
(t)
V
d2
(t)
V
out1
(n)
V
out2
(n)
V
d1
(n)
V
d2
(n)
1
(t) E
(t) E
Im
in
out
(t) E
(t) E
Re
in
out
Quadrature
MZ
Dual-Drive
MZ
CR
Wide Area Networks
51

Fig. 2.38 Real component of the dual-drive MZ output electric field.
However, for a fixed oversampling ratio, there are some degrees of freedom in
selecting the drive voltage values at the sampling instants since the dual-drive
nonlinear transfer characteristic is periodic with period 2V
and drive voltage values

V
m1
(n) and V
m2
(n) can be interchanged. The correct OFDM values in between the
sampling instants are obtained by smooth interpolation (ideally sinc() interpolation).
So, in order to minimize the error in between the sampling instances, the trajectory
spanned by the drive voltage vectors in the complex plane E
ou
(t)/E
in
(t) should also be
smooth, which is analogous to

1 , ), ( ) ( subject to
) ( ) ( minimize
, 1
2 1
+ = =
+
r r n n V n V
r V r V
i out m
m m

(2.38)
where ) 1 ( ) ( ) ( + = r V r V r V
mi mi mi
and i = 1, 2. In Eq. (2.38), we note that once the
pre-distortion equation for V
m1
(n) is chosen, the equation for V
m2
(n) is constrained
such that the optical electrical field at the sampling instant is equal to the OFDM
sample value.
Fig. 2.39 illustrates the output optical electrical field generated by the dual-
drive MZ for the cases of no trajectory optimization and with trajectory optimization
for an oversampling ratio of M
s
= 2.5 before the SGF.
0 20 40 60 80 100
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
Time (ps)
R
e
a
l
{
O
p
t
i
c
a
l

F
i
e
l
d
}

Dual-Drive
Desired
Wide Area Networks
52

Fig. 2.39 Output electric field of the dual-drive MZ for the cases of no trajectory optimization and with
trajectory optimization before the super-Gaussian filter. The green and blue lines represent the output
electric field and the correct OFDM waveform, respectively.
In Fig. 2.39 a), no trajectory optimization was used and the drive voltages were
set to V
m1
(n) = V
out1
(n) and V
m2
(n) = V
out2
(n) for all discrete time n and they were
limited to the range between V
to +V
. We observe that the optical field in between

the sampling instants is not equal to the correct OFDM waveform for the majority of
the time. However, in Fig. 2.39 b), we observe that the error in between the sampling
instants is significantly minimized when trajectory optimization is used. The receiver
sensitivity penalty is around 1-dB for an oversampling ratio M
s
= 2.5 when trajectory
optimization is used. On the other hand, if no trajectory optimization is used, the
oversampling ratio to achieve 1-dB receiver sensitivity penalty is M
s
5. Finally,
in Fig. 2.39 c), we have expanded the drive voltages range to 2V
and the error in

between samples was further reduced since there are now more degrees of freedom for
the drive voltages values. The receiver sensitivity penalty for this case is around 0.7
dB. We note that increasing the drive voltage range by a factor of two reduces the
effective vertical resolution of the DAC by the same factor which, in turn, means that
0 200 400 600 800 1000 1200
-4
-2
0
2
4
a) No Trajectory Optimization (-v
v
m,i,
v
)
R
e
{
O
p
t
i
c
a
l

F
i
e
l
d
}
0 200 400 600 800 1000 1200
0 200 400 600 800 1000 1200
Time (ps)
-4
-2
0
2
4
R
e
{
O
p
t
i
c
a
l

F
i
e
l
d
}
-4
-2
0
2
4
R
e
{
O
p
t
i
c
a
l

F
i
e
l
d
}
b) Trajectory Optimization (-v
v
m,i,
v
)
c) Trajectory Optimization (-2v
v
m,i,
2v
)
Wide Area Networks
53
a larger number of bits is required to maintain the same performance. However, we
believe that in a practical MZ modulator, the drive voltages values will be constrained
to V
so we will only consider this case for the remainder of the chapter.
Fig. 2.40 and Fig. 2.41 show the dual-drive MZ optical power efficiency and
receiver sensitivity penalty, respectively, for different clipping levels.

Fig. 2.40 Dual-drive MZ receiver sensitivity penalty with trajectory optimization for M
s
= 2.5, N
u
= 52
and R = 29.66 GHz.
We observe in Fig. 2.40 that the OPE increases significantly as we further clip
the signal, like in the quadrature MZ. In Fig. 2.41, we notice that the power penalty is
approximately constant down to about CR = 2.5, below which, the signal becomes
severely distorted and the receiver sensitivity penalty increases rapidly. Finally,
from Fig. 2.40 and Fig. 2.41, we conclude that the optimum clipping level is CR = 2.5,
as for the quadrature MZ.
1 2 3 4 5 6 7 8
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
O
p
t
i
c
a
l

P
o
w
e
r

E
f
f
i
c
i
e
n
c
y

(
W
/
W
)
Clipping Ratio (CR)
Wide Area Networks
54

Fig. 2.41 Dual-drive MZ optical power efficiency with trajectory optimization for M
s
= 2.5, N
u
= 52
and R = 29.66 GHz.
In Fig. 2.42, we compare the receiver sensitivity penalty for the two types of
modulators as a function of the oversampling ratio at the optimum clipping level, CR
= 2.5.

Fig. 2.42 Dual-drive and quadrature MZ receiver sensitivity penalties as a function of the oversampling
ratio with no frequency loss compensation and with trajectory optimization for CR = 2.5, drive
voltages range between V
and R = 29.66 GHz.

We observe in Fig. 2.42 that the quadrature MZ penalty remains constant as we
increase the oversampling ratio. The 0.4 dB residual penalty can be decomposed into
0.15 dB from the MZ frequency-dependent electrode losses and 0.25 dB from the
1 2 3 4 5 6 7 8
0
0.5
1
1.5
2
2.5
3
R
e
c
e
i
v
e
r

S
e
n
s
i
t
i
v
i
t
y

P
e
n
a
l
t
y

(
d
B
)
Clipping Ratio (CR)
1 2 3 4 5 6
0
0.5
1
1.5
2
2.5
3
Oversampling Ratio (M)
R
e
c
e
i
v
e
r

S
e
n
s
i
t
i
v
i
t
y

P
e
n
a
l
t
y

(
d
B
)

Quadrature MZ
Dual-drive MZ
Wide Area Networks
55
clipping noise. On the other hand, the dual-drive MZ penalty decreases as the
oversampling ratio increases. This is expected, as increasing the number of sampling
points reduces the error between the correct and attainable waveform trajectories in
the complex plane. Finally, we note that the dual-drive receiver sensitivity penalty
converges to the same value as that for the quadrature MZ as the oversampling ratio
increases.

Metropolitan Networks
56
3 METROPOLITAN NETWORKS
3.1 Introduction
Receiver-based electronic signal processing in metropolitan networks has also
been the subject of many recent studies. In long-haul systems using coherent
detection, linear equalizers have been shown to fully compensate linear fiber
impairments in single-mode fiber (SMF), such as group-velocity dispersion (GVD)
and polarization-mode dispersion (PMD) [16]. On the other hand, in metropolitan
systems using direct-detection (DD), linear equalizers offer little performance
improvement, because the nonlinear photodetection process destroys information on
the phase of the received electric field. Recently, maximum-likelihood sequence
detection (MLSD) was shown to be effective in mitigating GVD and PMD
impairments in DD links [35], [36]. Although the computational complexity of MLSD
increases exponentially with the channel memory, using low-complexity branch
metrics, near-optimal performance can be achieved with manageable computational
complexity, provided the effective channel memory does not exceed a few symbol
intervals [3].
An alternate approach to combating fiber impairments in amplified DD
systems is to use multicarrier modulation, such as orthogonal frequency-division
multiplexing (OFDM). There are three major approaches for combining OFDM with
DD. The first two techniques are based on intensity modulation (IM), and so require
some means to make the OFDM signal nonnegative. The first technique adds a DC
bias to reduce the negative signal excursions and then clips the remainder of the
negative excursions. This method is called DC-OFDM. The second technique clips the
entire negative excursion of the waveform, avoiding the need for a DC bias [37], [38].
Clipping noise is avoided by appropriate choice of the subcarrier frequencies. This
technique is called asymmetrically clipped optical OFDM (ACO-OFDM) [38]. The
third technique is based on single-sideband modulation of the complex-valued optical
57
electric field by an OFDM signal. A DC bias (carrier component) is added, leaving a
guard band between the carrier and the OFDM signal in order to avoid
intermodulation products caused by photodetection [39], [40]. This method is called
single-sideband OFDM (SSB-OFDM).
There have been several studies of the different OFDM techniques (e.g., [38]),
but there has been no comparison of power efficiencies among the various direct-
detection OFDM methods in SMF, and to conventional baseband methods, such as on-
off keying (OOK). Furthermore, in previous work, the DC bias and the powers of the
subcarriers were not jointly optimized based on the channel response and the nonlinear
beat noises, since there is no closed-form solution for this optimization. We present an
iterative procedure based on known bit-loading algorithms with a new modification,
the bias ratio (BR), in order to obtain the optimum power allocation. We compare the
performance of the three OFDM techniques using optimized power allocations to the
performance of OOK with MLSD.
This chapter is organized as follows. In section 3.2, we present the metro
networks and review methods for power and bit allocation for multicarrier systems
and describe the optimal water-filling solution. We present our system model in
section 3.3. In section 3.4, we review the different OFDM formats and derive
equivalent linear channel models for each one, assuming only GVD is present in the
SMF. Furthermore, we discuss the effects of amplifier noise in direct-detection
systems and compute an expression for the autocorrelation function of the
photodetected noise. In section 3.5, we compare the optical power required to transmit
at a given bit rate for the different optimized OFDM formats and for OOK with
MLSD, making use of previously published results for the latter format. In that
section, we also compare the computational complexities of the various formats.
58
3.2 Metropolitan Networks and Power Allocation Review
3.2.1 Metropolitan Networks
Metropolitan networks (or metro) typically use single-mode fiber (SMF) and
also operate around 1.55 m which is the region with the lowest fiber attenuation.
Metro networks normally use light sources with small linewidths like lasers. Metro
systems demutliplex the incoming high-speed traffic from long-haul systems into
several lower bit rate streams, typically between 1 Gbit/s and 40 Gbit/s. The
demultiplexed traffic is transmitted in the metro network (e.g. core network of New
York city) and the distances vary between 50 km to 500 km. Fig. 3.1 shows a block
diagram of a metro network.

Fig. 3.1 Metro network diagram.
Metro systems typically use a single span between terminal equipment and do
not employ any dispersion-compensating fiber (DCF). At the end of the link, an
optical amplifier is used to compensate the SMF attenuation. An optical filter is also
normally placed after the amplifier in order to reduce out-of-band amplifier noise (i.e.,
ASE noise). Fig. 3.2 shows a block diagram of a metro link.
TX RX
SMF
Amplifier + Filter
59
Fig. 3.2 Metro link with an optical filter.
Metro systems use direct-detection (i.e., detect the instantaneous optical
power) for its simplicity and reduced hardware cost. Furthermore, metro links
normally employ intensity modulation (IM), i.e., the transmitter modulates the
instantaneous optical power (intensity) but electric field modulation can also be used.
In IM, the modulating waveform has to be nonnegative in order to avoid loss of
information. IM can be achieved, for example, by direct modulation of the laser
current, as shown in Fig. 3.3.
Fig. 3.3 Intensity modulation (IM) by direct modulation of the laser current.
We note that in IM, the average transmitted optical power P
t
is proportional to
the average of the modulating waveform E[x(t)], and is given by

= =
T
T
T
T
T
opt
T
t
dt t x
T
K dt t P
T
P ) (
2
1
lim
) (
2
1
lim
, (3.1)
rather than the usual time-average of |x(t)|
2
(i.e., E[|x(t)|
2
]), which is appropriate when
x(t) represents amplitude.
3.2.2 Power and Bit Allocation Review
3.2.2.1 Gap Approximation
On an AWGN channel, the maximum achievable bit rate is given by the
Shannon capacity
Current (mA)
Laser Power
(mW)
I
th
x(t)
P
opt
(t)
60
( ) SNR
B
C
+ = 1 log
2
, (3.2)
where C is the capacity, B is the channel bandwidth and SNR is the signal-to-noise
ratio. Any real system must transmit at a bit rate less than capacity. For QAM
modulation, the achievable bit rate can be expressed approximately as

+ =
SNR
B
R
1 log
2
,
(3.3)
where R is the bit rate and is called the gap constant. The gap constant, introduced
by Cioffi et al [41] and Forney et al [42], represents a loss with respect to the Shannon
capacity.

Fig. 3.4 Achievable bit rates as a function of SNR for various values of the gap .
The gap analysis is widely used in the bit loading of OFDM systems, since it
separates coding gain from power-allocation gain [43]. For uncoded QAM, the gap
constant is given by

3
2
= , (3.4)
where ) (
1
e
P Q
= , P
e
is the symbol-error probability and Q
-1
is the inverse Q
function. As an example, the gap is 8.8 dB at P
e
= 10
6
and is 9.5 dB at P
e
= 10
7
for
uncoded QAM. The use of forward error-correction (FEC) codes reduces the gap. A
2 4 6 8 10 12 14 16
0
1
2
3
4
5
6
R
/
B

(
b
i
t
/
s
/
H
z
)
SNR (dB)
Capacity
( = 0 dB)
3 dB
6 dB
9 dB
61
well-coded system may have a gap as low as 0.5 dB at P
e
< 10
6
. A gap of 0 dB means
the maximum bit rate has been achieved and therefore R = C. Fig. 3.4 shows the
attainable bit rates for various gap values. In the case of real-valued PAM, the gap
approximation is valid, but the bit rate is reduced by a factor of two as compared to
Eq. (3.3). For the remainder of the thesis, we define the normalized bit rate B R b / = ,
which has units of bits/s/Hz.
3.2.2.2 Optimum Power Allocation
The OFDM signal splits the transmission channel into N parallel channels.
When the total average transmitted power is constrained, the maximum obtainable bit
rate can be written as

=
=
+ =
1
0
1
0
2
2
2
P
subject to
1 log b max
n
N
n
n t
N
n
n n
P P
H P
n
, (3.5)
where |H
n
|
2
,
2
n
and P
n
are the channel gain, noise variance and transmitted power at
subcarrier n, respectively, and P
t
is total average transmitted power. The solution of
Eq. (3.5) is the known water-filling solution.

a
H
P
n
n
n
= +
2
2
,
(3.6)
where a is a constant chosen such that
=
n
n t
P P . The optimum power allocation is
then

<

=
2
2
2
2
2
2
, 0
,
n
n n
n
H
a
H
a
H
a
P
n
n n
. (3.7)
The optimum power allocation is illustrated in Fig. 3.5.
62
Fig. 3.5 Optimal power allocation for OFDM.
After the optimum power allocation is determined, the number of bits to be
transmitted on each subcarrier is computed using Eq. (3.3). While the ideal value of b
= R/B is an arbitrary nonnegative real number, in practice, the constellation size and
FEC code rate are adjusted to obtain a close rational approximation. In our analysis,
we will neglect the difference between the ideal value of b and its rational
approximation.
We note that when the total average transmitted power is constrained, the
optimal power allocation yields a variable data rate. For some applications, a fixed
data rate is required. In this case, the optimal design minimizes the average power
required to transmit at a given fixed bit rate. The power minimization can be written as

+ =
=
1
0
2
2
2
1
0
t
P
1 log b subject to
P min
n
N
n
n n
N
n
n
n
H P
P
. (3.8)
The solution of Eq. (3.8) is also the water-filling solution given by Eq. (3.6).
However, in this case the constant a is chosen such that bit rate is equal to the desired
value. We can interpret this solution as the water/power being poured until the
required bit rate is achieved.
[ ]
[ ]
2
2
2
2
H
N

[ ]
[ ]
2
2
3
3
H
N

[ ]
[ ]
2
2
4
4
H
N

[ ]
[ ]
2
2
N H
N
N

n
1 2 3 4 N
water level a
total power P
[ ]
[ ]
2
2
1
1
H
N

[ ] 0 1 = P
[ ] 2 P
[ ] 3 P
[ ] 4 P
[ ] 5 P
n
1 2 3 4 N
63
3.3 OFDM System Model for Metro Links
The OFDM system model is shown in Fig. 3.6.

Fig. 3.6 OFDM system model for metro links.
An optical OFDM modulator encodes transmitted symbols onto an electrical
OFDM waveform and modulates this onto the intensity (instantaneous power) of an
optical carrier. The modulator can generate one of DC-OFDM, SSB-OFDM or ACO-
OFDM. Details of modulators for particular OFDM schemes are described in
section 3.4. After propagating through the SMF, the optical signal is optically
amplified and bandpass-filtered. We assume that the optical amplifier has a high gain
so that its amplified spontaneous emission (ASE) is dominant over thermal and shot
noises. The ASE noise is modeled as complex additive white Gaussian noise (AWGN)
with zero mean. The ASE can be expressed as n
ASE
(t) = n
I
(t) + jn
Q
(t), where n
I
(t) and
n
Q
(t) are uncorrelated Gaussian random processes, each having half the variance of
n
ASE
(t).
In our analysis, PMD and fiber nonlinearity are neglected, and GVD is the only
fiber impairment considered. In the optical electric field domain, the fiber is modeled
as a linear system with transfer function given by

( )
L j
fiber
e H
2
2 2

= ,
(3.9)
where
2
is the fiber GVD parameter and L is the fiber length. The overall optical
transfer function is given by

( ) ( ) ( )
filter fiber
H H H = ,
(3.10)
.
.
.
X
0
X
1
X
N2
X
N1
Optical
OFDM
Tx
Fiber
Optical
Filter
Anti-Aliasing
Filter
OFDM
Rx
Amplifier
.
.
.
Y
0
Y
1
Y
N2
Y
N1
64
where H
filter
() is the transfer function of the optical bandpass filter.
After bandpass filtering, the optical signal intensity is detected, and the
electrical current is lowpass filtered. The OFDM signal is demodulated and equalized
with a single-tap equalizer on each subcarrier to compensate for the channel distortion
[38], [39].
3.4 Analysis of Direct-Detection OFDM Schemes
3.4.1 DC-Clipped OFDM
The main disadvantage of using OFDM with IM is the required DC bias to
make the OFDM signal nonnegative. Since OFDM signals have a high peak-to-
average power ratio, the required DC bias can be excessively high. A simple approach
to reduce the DC offset is to perform hard-clipping on the negative signal peaks [37],
[38]. This technique is usually called DC-OFDM and the transmitter is shown in Fig.
3.7.

Fig. 3.7 Block diagram of a DC-OFDM transmitter.
In Fig. 3.7, the transmitted symbols are modulated such that the time-domain
waveform is real. This is achieved by enforcing Hermitian symmetry in the symbols
input to the inverse discrete Fourier transform (IDFT). We note that the 0
th
or the DC
subcarrier is not modulated and it is equal to the DC offset. After D/A conversion, the
Optical Spectrum
X
k
X
k
*
f
X
N/21
X
1
X
0
IDFT
X
1
*
Parallel
to
Serial
+
Cyclic
Prefix
.
.
.
D/A
X
N/21
*
.
.
.
Clip
Real
Signal
Fiber
X
N/2
Laser
DC bias
QAM
QAM
*
65
electrical OFDM signal is hard-clipped such that the waveform is nonnegative and
then the signal is intensity modulated onto the optical carrier.
After propagating through the optical system, the detected signal can be
written as
( ) ( ) ( )
2
2
det
) ( t h t x t E t y = = , (3.11)
where E(t) is the output electric field, x(t) is the DC-OFDM signal and h(t) is the
impulse response of the overall optical system, given by the inverse Fourier transform
of Eq. (3.10). For now, we have neglected amplifier noise in Eq. (3.11).
We can expand the OFDM signal as x(t) = A + x
ac
(t) + n
clip
(t), where A is the
average of the OFDM signal after clipping and x
ac
(t) is the OFDM signal with the DC
subcarrier equal to zero and n
clip
(t) is the clipping noise. We note that if the DC bias is
chosen such that there is no clipping, then A is equal to the DC bias and n
clip
(t) = 0. For
now, we neglect the clipping noise. The detected signal is then
( ) ( )
2
det
) ( t h t x A t y
ac
+ = . (3.12)
We can now use the Taylor series expansion on the square-root term. The
detected signal becomes

( )
( )
*
2 / 3
2
2 / 3
2
det
8
) (
2
) (

8
) (
2
) (
) (
+ +
+ + =
t h
A
t x
A
t x
A
t h
A
t x
A
t x
A t y
ac ac
ac ac
(3.13)
After expanding the terms in Eq. (3.13), the detected signal simplifies to

( ) ( ) ( ) ( )
( ) ( ) ( ) ( )
+
+ =
A
t h t x
A
t h t x
t h t x t h t x
A t y
ac ac
ac ac
8 8

2 2
) (
* * 2 2
* *
det

(3.14)
The first two convolution terms in Eq. (3.14) correspond to an equivalent
linear channel. The remaining terms in Eq. (3.14) correspond to the interaction
between the detector nonlinearity and the fiber GVD. They can be interpreted as an
66
equivalent nonlinear detection noise that will degrade the receiver SNR. We note that
these terms decrease as the bias level increases.
The linear terms in Eq. (3.14) can be further simplified since the OFDM signal
is real valued, i.e., ( ) ( ) t x t x
ac ac
*
= . Thus, we can write the equivalent linear channel for
DC-OFDM as
( )
( ) ( )
2
*
t h t h
t h
eq
+
= . (3.15)
The transfer function is then

( )

=
+
=

L
e e
H
L j L j
eq
2
cos
2
2 2
2 2
2 2 2 2
.
(3.16)
The equivalent transfer function is plotted on Fig. 3.8.

Fig. 3.8 Equivalent transfer function for DC-OFDM for R = 20 Gbit/s, L = 100 km and D = 17
ps/nm/km, corresponding to = 0.87.
In Fig. 3.8, we observe that the equivalent channel transfer function is not
unitary and is frequency-selective. Thus, in order to maximize the bit rate, we need to
use variable bit loading on each subcarrier.
3.4.2 Asymmetrically Clipped Optical OFDM
Armstrong and Lowery [37], [38] proposed adding no DC bias and clipping the
entire negative excursion of an electrical OFDM signal before modulating it onto the
0 2 4 6 8 10 12 14 16 18 20
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
|
H
e
q
(
f
)
|
Frequency (GHz)
67
intensity of an optical carrier. They showed that clipping noise is avoided (at least in
the absence of dispersion) by encoding information symbols on only the odd
subcarriers. This technique is called ACO-OFDM. Fig. 3.9 shows a block diagram of
an ACO-OFDM transmitter.

Fig. 3.9 Block diagram of an ACO-OFDM transmitter.
As in DC-OFDM, the OFDM subcarriers are assumed to have Hermitian
symmetry, so that the time-domain waveform is real. Because only the odd subcarriers
are used to transmit data, for a given choice of signal constellation, ACO-OFDM has
only half the spectral efficiency of DC-OFDM. After D/A conversion, the electrical
OFDM signal is hard-clipped at zero and intensity modulated onto an optical carrier.
The equivalent transfer function of ACO-OFDM is the same as DC-OFDM,
and is given by Eq. (3.16). This can be shown by using the same approach as in
section 3.4.1, setting A equal to the average of the ACO-OFDM waveform.
3.4.3 Single-Sideband OFDM
Another approach to reducing the required DC bias is to use SSB-OFDM. In
SSB-OFDM, only one of the DC-OFDM sidebands is transmitted and there is a
frequency guard band between the sideband and the optical carrier. The bandwidth of
the guard band must be no smaller than that of the OFDM sideband. Fig. 3.10 shows a
block diagram of a digital implementation of a SSB-OFDM transmitter
QAM
QAM
X
k
X
k
*
f
X
1
X
N/21
IDFT
Parallel
to
Serial
+
Cyclic
Prefix
D/A
Fiber
QAM
Real
Signal
0
QAM
.
.
.
0
0
0
0
.
.
.
Clip
X
0
X
1
*
.
.
.
X
N/21
*
.
.
.
X
N/2
Laser
*
*
Optical Spectrum
68

Fig. 3.10 Block diagram of an SSB-OFDM transmitter.
SSB-OFDM can be obtained by applying a Hilbert transform to a DC-OFDM
signal. A Hilbert transform can be generated at the OFDM transmitter by setting the
negative subcarriers to zero, as shown in Fig. 3.10. The frequency guard band is
obtained by modulating the high-frequency subcarriers and setting to zero the low-
frequency subcarriers. The IDFT output is an analytical signal, and its real and
imaginary components are used for the drive voltages of the inphase and quadrature
inputs of a quadrature Mach-Zehnder modulator. Hence, contrary to the previous
techniques, in SSB-OFDM the signal is actually modulated onto the optical electric
field, and not simply onto its intensity, and therefore no hard clipping is required. The
DC bias in SSB-OFDM sets the amplitude of the optical carrier, affecting system
performance, as we show in the following.
SSB-OFDM has the same optical spectral efficiency as DC-OFDM. However,
the transmitter considered here requires twice the bandwidth of DC-OFDM for a
digital implementation of SSB-OFDM. Other transmitter implementations that require
less bandwidth are discussed in [40]. There have been some studies on the
optimization of SSB-OFDM, particularly in the trade-off between spectral efficiency
and performance as a function of the optical carrier power (i.e. DC bias) [44], [45].
However, in the previous work, the DC bias and the powers of the subcarriers were
not jointly optimized based on the channel response and the nonlinear beat noises.
IDFT
Parallel
to
Serial
+
Cyclic
Prefix
.
.
.
D/A
DC bias
.
.
.
Re
0
0
D/A
Quad
MZ
Im
Optical Spectrum
Fiber
X
0
X
1
X
N/2
X
N1
QAM
B B
f
69
After propagating through the optical system, the detected signal can be
written as
( ) ( ) ( ) ( )
2
2
2
det
) ( t h e t s A t E t y
t f j
b
+ = =

, (3.17)
where ( )
t f j
b
e t s
2
is the OFDM sideband centered at frequency f
b
, A is the DC bias and
h(t) is the optical channel impulse response, given by the inverse Fourier transform of
Eq. (3.10). We can expand the detected signal as

( ) ( ) [ ] ( ) ( ) [ ]
( ) ( ) ( ) ( )
( ) ( )
2
* 2 * 2
2
*
2 2
det

) ( ) ( ) (
t h t s
t h e t s A t h e t s A A
t h e t s A t h e t s A t y
t f j t f j
t f j t f j
b b
b b
+
+ + =
+ + =

.
(3.18)
Assuming that the OFDM sideband has a bandwidth B centered at f
b
, the
quadratic term in Eq. (3.18) occupies a bandwidth from B to +B. Thus, if the
frequency guard band is equal or greater than B, the quadratic intermodulation
products are avoided. This concept is illustrated in Fig. 3.11.

Fig. 3.11 PSD of the detected signal in SSB-OFDM.
Finally, since the quadratic intermodulation products are avoided by using the
guard band and the desired signal is the OFDM sideband centered at frequency f
b
, we
conclude from Eq. (3.18) that the transfer function for SSB-OFDM is a scaled version
of the optical transfer function given by Eq. (3.10), and can be written as

( ) ( ) H A H
SSB
= ,
(3.19)
where A is the DC bias.
OFDM
*
OFDM
f
B B
0 f
b
-f
b
Intermodulation
Products
(f) S
det
y
70
3.4.4 Effects of Amplifier Noise
If we include amplifier noise, the detected signal can be written as
( ) ( )
2
det
) ( t n t E t y + = , (3.20)
where E(t) is the output electric field and n(t) is the filtered ASE, i.e.,
( ) ( ) ( ) t h t n t n
filter ASE
= . We can further expand the detected signal as

( ) ( ) ( ) ( )
( ) ( ) ( )
2
* *
2
*
det
) ( ) ( ) (
) ( ) ( ) (
t n t n t E t n t E t E
t n t E t n t E t y
+ + + =
+ + =
. (3.21)
From Eq. (3.21), we observe that the signal is corrupted by signal-spontaneous
and spontaneous-spontaneous beat noises. We define
( ) ( )
2
* *
) ( ) ( ) ( ) ( t n t n t E t n t E t w + + = as the total noise. By using the moment
generating function for n(t) or Gaussian moments theorems [46] and assuming that the
transmitted signal is uncorrelated with the amplifier noise, we can write the
autocorrelation of the total noise w(t) as

( )
( )
) ( 2 ) 0 (
) ( | | ) (
) ( | | ) ( ) (
2 2
2 *
* 2
' '
' '
R R
R m R
R m R R
I I I I
n n n n
nn E
E E
nn E
E E
ww
+ +
+ +
+ =

, (3.22)
where ( )
E
m t E t E = ) (
'
, m
E
is the average value of the received electric field, ) (
' '

E E
R
is the autocorrelation of ( ) t E
'
, R
nn
() is the autocorrelation of the noise n(t) and
( )
I I
n n
R is the autocorrelation of the real component of n(t). In Eq. (3.22), there are
three main noise components:
DC-spontaneous noise: )) ( ) ( ( | |
* 2

nn nn E
R R m +
Signal-spontaneous noise: ) ( ) ( ) ( ) (
* *
' ' ' '

nn
E E
nn
E E
R R R R +
Spontaneous-spontaneous noise: ) ( 2 ) 0 (
2 2
R R
I I I I
n n n n
+
71
From Eq. (3.22), we observe that the DC-spontaneous beat noise power
increases proportionally with the DC bias. In DC-OFDM, if the DC bias is excessively
high, the DC-spontaneous beat noise dominates, and the receiver SNR is low. On the
other hand, if the DC bias is small, the clipping noise and the nonlinear detection noise
from Eq. (3.14) dominate, and the receiver SNR is also low. Thus, the DC bias needs
to be optimized for best performance.
In SSB-OFDM, if the DC bias is high, the receiver SNR is unaffected since
both the OFDM signal and the dominant noise are amplified by the DC offset.
However, the optical power efficiency is very low in this case. On the other hand, if
the DC bias is small, the other noise terms dominate and the receiver SNR is low.
3.5 Comparison of Direct-Detection Modulation Formats
In order to make our results independent of bit rate, we use the dimensionless
dispersion index , defined as

L R
2
2
= ,
(3.23)
where
2
is the fiber GVD parameter, L is the fiber length and R is the bit rate.
Furthermore, in order for the required optical SNR to be independent of the bit rate,
the noise bandwidth should be matched to the signal bandwidth. Thus, we use the
normalized optical SNR defined as

R N
P
SNR
opt
opt
0
= , (3.24)
where P
opt
is the optical power, N
0
is the ASE power spectral density and R is the bit
rate. The normalized optical SNR is related to the conventional optical SNR (OSNR)
measured in a 0.1-nm (12.5-GHz) bandwidth as

R
OSNR SNR
opt
GHz 5 . 12
= , (3.25)
The system model is shown in Fig. 3.6. As mentioned before, we neglect all
transmission impairments except for GVD. We consider a fiber with a dispersion D =
72
17 ps/(nmkm). The optical filter is modeled as a second-order super-Gaussian filter
with a 3-dB bandwidth B
0
= 35 GHz, which is realistic for commercial systems with
50-GHz channel spacing. We also assume that the optical amplifier has a high gain so
that the ASE is dominant over thermal and shot noises from the receiver. At the
receiver, the anti-aliasing filter is a fifth-order Butterworth lowpass filter having a 3-
dB cutoff frequency equal to the first null of the OFDM spectrum [25].
In order to perform a fair comparison with OOK, we maintain OFDM optical
bandwidth constant and equal to the optical bandwidth required to transmit at a bit rate
R using OOK. Table 3.1 summarizes the required electrical and optical bandwidths for
the different modulation techniques.

Fig. 3.12 PSD of the different modulation formats.
Modulation Receiver BW Optical BW
OOK R 2R
DC-OFDM R 2R
ACO-OFDM R 2R
SSB-OFDM 2R 2R
Table 3.1 Electrical and optical bandwidths required to transmit at a bit rate R.
For all the OFDM formats, we use an oversampling ratio of M
s
= 64 /52 1.2
to avoid aliasing. As an initial estimate, we choose the cyclic prefix and the number of
subcarriers using the equivalent linear channel described in section 3.4 for each
OFDM format [25]. Then, we increase the cyclic prefix until the interference penalty
is completely eliminated. Typically, the final result is very close to the initial estimate.
R 2
OFDM
*
( ) f
OFDM DC
S

f
OFDM OFDM
( ) f
OFDM SSB
S

f
R 2
R 2
OFDM
*
( ) f
OFDM ACO
S

f
OFDM
R 2
f
( ) f
OOK
S
73
After selecting the OFDM parameters, we proceed to optimally allocate power
among the subcarriers. We note that we cannot use Eqs. (3.6) and (3.7) directly, since
the nonlinear detection noise in each subcarrier depends on the power of all the
subcarriers. Thus, we perform the power allocation iteratively. For a given power
allocation, the SNR measured at each subcarrier is used to compute an updated water-
filling solution. We repeat this process until the power allocation no longer changes. In
our simulations, we do not actually implement the coding required to transmit the
number of bits given by the water-filling solution. In practice, the constellation used
on each subcarrier will be M-ary QAM (M-QAM) with coding (e.g., trellis coded
modulation) in order to achieve the number of bits given by the water-filling
algorithm. As seen before, the number of bits that can be allocated to a subcarrier
depends only on the received SNR of that subcarrier. For simplicity, we used QPSK
on each subcarrier with the same power given by the water-filling solution in order to
obtain the same SNR that would be required to transmit the specific number of bits
given by the bit allocation.
In addition to optimizing the power/bit allocation at each subcarrier, we need
to optimize the DC bias for DC-OFDM and SSB-OFDM. In order to minimize the DC
bias, we use a bias level proportional to the square-root of the electrical power. We
define the proportionality constant as the bias ratio (B
R
), which is given by

bias
R
DC
B = , (3.26)
where
elect
P = is the standard deviation of the electrical OFDM waveform. Using
the B
R
insures that the water-filling solution minimizes the DC bias, and thus the
optical power required. Fig. 3.13 shows the flow chart for power minimization of DC-
OFDM and SSB-OFDM using the bias ratio.
74
Fig. 3.13 Flow chart of DC-OFDM and SSB-OFDM optimization using the bias ratio (B
R
).
For example, in DC-OFDM, if, at iteration k, the electrical power is high after
subcarrier power allocation, the DC bias will also be high (since it is proportional to
the square-root of electrical power), and the optical power will be high. A high DC
bias reduces both the clipping noise and the intermodulation terms, and therefore the
received SNR on each subcarrier is high. At the next iteration k +1, the water-filling
algorithm removes some of the excess electric power in order to lower the SNR to the
desired value, thereby reducing the DC bias and the optical power. On the other hand,
if, at iteration k, the DC bias is low, the clipping noise and the intermodulation terms
will be high, and therefore the received SNR on each subcarrier is low. At the next
iteration k +1, the water-filling algorithm adds more electrical power to increase the
electrical SNR, and consequently the DC bias and optical power will increase. For a
given value of the bias ratio B
R
, we repeat this process until the power allocation and
DC bias no longer change. The minimum required optical SNR is obtained by
performing an exhaustive search over the value of the bias ratio B
R
, as shown in Fig.
3.14.
Finally, we note that the water-filling solution is optimum for Gaussian noise
but the noises in an optically amplified direct-detection receiver are not Gaussian.
However, if the number of subcarriers is large, we expect the noise to be
approximately Gaussian after the DFT, an assumption confirmed in our simulations.
Power allocation
(water-f illing)
Update DC
subcarrier using
the bias ratio
Measure received
subcarrier
SNRs/gains
Power
allocation
changed?
End
No
Yes
Initialize the
subcarrier
SNRs/gains
75
Typical high-performance FEC codes for optical systems have a threshold of
the order of P
b
= 10
3
, so we compute the minimum optical power required to achieve
P
b
= 10
3
for the different OFDM formats. Fig. 3.14 shows the normalized optical
SNR required for a dispersion index of = 0.25. As an example, = 0.25 corresponds
to 100 km of standard single-mode fiber (SSMF) without optical dispersion
compensation at 10 Gbit/s or 70 km of SSMF with 90% inline dispersion
compensation at 40 Gbit/s.

Fig. 3.14 Normalized optical SNRs required for the different OFDM formats for = 0.25 to achieve P
b
= 10
3
. The number of used subcarriers, N
u
, is indicated for each curve.
In Fig. 3.14, we notice that for all OFDM formats the required optical SNR
decreases slightly with the number of subcarriers, since the prefix penalty decreases as
the number of subcarriers increases. In addition, OFDM uses the available channel
bandwidth more efficiently as the number of subcarriers increases, which contributes
to the reduction of the required optical SNR. In Fig. 3.14, we observe that the
performance achieved with the chosen number of subcarriers is very close to the best
performance achievable for each OFDM technique.
We also verify that SSB-OFDM requires the lowest optical SNR to achieve P
b
= 10
3
. Furthermore, for SSB-OFDM, the required optical SNR increases linearly with
the DC bias for B
R
values greater than 1.4. This is expected since from Eq. (3.19) the
electrical SSB-OFDM signal is scaled by the DC bias, and the dominant noise for a
0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4
11
12
13
14
15
16
17
18
19
20
21
DC-OFDM
N
u
=
208
416
832
ACO-OFDM
N
u
=
104
208
416
SSB-OFDM
N
u
=
52
104
208
R
e
q
u
i
r
e
d

S
N
R
o
p
t
(
d
B
)
Bias Ratio, B
R
76
high bias level is the DC-spontaneous beat noise. As we decrease the B
R
, the signal-
spontaneous beat noise is no longer negligible, and the minimum required optical SNR
is achieved at a B
R
= 1. We note that the optimum B
R
is independent of the number of
used subcarriers. If we further decrease the B
R
, more power has to be allocated to each
subcarrier to compensate for the signal-spontaneous beat and therefore the required
optical SNR increases.
For DC-OFDM, the required optical SNR also increases linearly with the DC
bias for B
R
values greater than 1.6, since the dominant noise is the DC-spontaneous
beat noise. As we decrease the B
R
, the clipping and the nonlinear detection noises are
no longer negligible and the minimum required optical SNR is achieved at a B
R
= 1.1.
We note again that the optimum B
R
is independent of the number of used subcarriers.
If we further decrease the B
R
, more power must be allocated to each subcarrier to
compensate for the clipping and nonlinear noise, and therefore the average value of
the OFDM waveform increases. Thus, the required optical SNR also increases.
Fig. 3.15 shows an example of the subcarrier power allocation for the different
OFDM formats.

Fig. 3.15 Subcarrier power distribution for DC-OFDM, ACO-OFDM, SSB-OFDM and to achieve a P
b
= 10
3
for = 0.25. The number of used subcarriers is 416, 416 and 208 for DC-OFDM (B
R
= 1.1),
ACO-OFDM and SSB-OFDM (B
R
= 1.0), respectively.
We observe that the power allocation is consistent with the equivalent linear
channel described in section 3.4. For DC-OFDM and ACO-OFDM, we observe a
0 0.2 0.4 0.6 0.8 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Normalized Subcarrier Frequency (f
n
/ R)
S
u
b
c
a
r
r
i
e
r

P
o
w
e
r

(
P
n
/

P
n
,
m
a
x
)
DC-OFDM
ACO-OFDM
SSB-OFDM
77
frequency-selective channel with notches, as expected. For SSB-OFDM, there are no
notches, and the power allocation follows the shape of the signal-spontaneous beat
noise.
For DC-OFDM, we can obtain an analytical lower bound on the required
optical SNR by theoretically assuming that the only noises sources are the DC-
spontaneous and the spontaneous-spontaneous beat noises from Eq. (3.22) and the by
using the equivalent linear channel from Eq. (3.16). If we fix the DC bias, the
optimum power allocation is given directly by Eqs. (3.6) and (3.7), since the noises do
not depend on the subcarrier powers. Fig. 3.16 compares the analytical lower bound
with the results obtained by simulation.

Fig. 3.16 Normalized optical SNR required for DC-OFDM to achieve P
b
= 10
3
for = 0.25. The
number of used subcarriers is N
u
= 832.
From Fig. 3.16, we verify that for high B
R
(B
R
> 2), the lower bound is equal to
the simulation results. Thus, we conclude that using water-filling iteratively with the
B
R
normalization for the DC level converges to the optimum power allocation in the
linear regime. Furthermore, from Fig. 3.14 and Fig. 3.16, we observe that DC-OFDM
can never be more power efficient than SSB-OFDM.
Fig. 3.17 shows the minimum required optical SNR for various dispersion
indexes . The OFDM parameters are summarized in Table 3.2.
0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4
12
13
14
15
16
17
18
19
20
21
Bias Ratio, B
R
R
e
q
u
i
r
e
d

S
N
R
o
p
t
(
d
B
)
Lower Bound
Simulation
78

Fig. 3.17 Minimum normalized optical SNR required for various dispersion indexes for the different
OFDM formats.
In Fig. 3.17, we observe that SSB-OFDM requires the lowest optical SNR to
achieve P
b
= 10
3
. The high performance of SSB-OFDM comes at the price of the
extra electrical bandwidth required to avoid the quadratic intermodulation products.
Furthermore, unlike OOK with MLSD, SSB-OFDM requires two D/As and a
quadrature modulator at the transmitter.
DC-OFDM ACO-OFDM SSB-OFDM
0.25
N = 2048
N
u
= 832
= 18
N = 2048
N
u
= 416
= 18
N = 1024
N
u
= 208
= 11
0.5
N = 2048
N
u
= 832
= 31
N = 2048
N
u
= 416
= 31
N = 1024
N
u
= 208
= 21
0.75
N = 2048
N
u
= 832
= 44
N = 2048
N
u
= 416
= 44
N = 2048
N
u
= 416
= 32
1
N = 2048
N
u
= 832
= 59
N = 2048
N
u
= 416
= 59
N = 2048
N
u
= 416
= 42
Table 3.2 OFDM parameters for the various dispersion indexes . N is the DFT size, N
u
is the number
of used subcarriers and is the cyclic prefix.
From Fig. 3.17, we can also verify that the required optical SNR for ACO-
OFDM is higher than DC-OFDM for greater than 0.25. This happens because the
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
10
12
14
16
18
20
22
24
R
e
q
u
i
r
e
d

S
N
R
o
p
t
(
d
B
)
DC-OFDM
ACO-OFDM
SSB-OFDM
79
fiber dispersion destroys the orthogonality between the subcarriers and therefore there
is significant nonlinear inter-carrier interference (ICI). Furthermore, since half the
subcarriers are zero, the used subcarriers need to transmit at twice the data rate, which
exacerbates the ICI. We note that the power allocation for ACO-OFDM does not
converge for greater than 0.5. On the other hand, for low dispersion the subcarriers
remain orthogonal and the ICI is avoided due to the zero subcarriers. In this regime,
ACO-OFDM performs better than DC-OFDM.
In order to compare OFDM with single-carrier, we use results on OOK with
MLSD from [3], since OOK with MLSD achieves the best optical power efficiency
among known IM/DD techniques. For ease of comparison with [3], we present our
results for 10.7 Gbit/s in Fig. 3.18 using the conventional definition of optical SNR
measured over a 12.5-GHz band (OSNR) to achieve P
b
= 10
3
.

b
= 10
3
at 10.7 Gbit/s for OFDM and for
OOK with MLSD [3].
We see that OOK with MLSD performs better than DC-OFDM or ACO-
OFDM. SSB-OFDM performs as well as OOK with MLSD but requires twice the
electrical bandwidth. For 10 Gbit/s systems, we can further optimize the OFDM
signals by using the full available optical bandwidth (B
0
= 35 GHz). In this case, we
are trading optical bandwidth for optical power. Fig. 3.19 shows the OSNR required to
0 50 100 150 200 250 300 350 400
8
10
12
14
16
18
20
22
24
L (km)
R
e
q
u
i
r
e
d

O
S
N
R

(
d
B
)
DC-OFDM
ACO-OFDM
SSB-OFDM OOK with MLSD (Bosco)
80
obtain P
b
= 10
3
when the OFDM signal occupies the full channel bandwidth (B
0
= 35
GHz).

b
= 10
3
at 10.7 Gbit/s for OFDM and for
OOK with MLSD [3]. In this case, the OFDM signal occupies the full channel bandwidth (B
0
= 35
GHz).
As we can observe, SSB-OFDM performs better than MLSD if we use the
extra available optical bandwidth. We also notice that the power allocation for ACO-
OFDM converges because of the extra degrees of freedom as compared to the case
in Fig. 3.17. However, OOK with MLSD still performs better than ACO-OFDM and
DC-OFDM.
3.5.1 Computational Complexity
Another important criterion is the computational complexity of a
modulation/detection technique. For OOK with MLSD, the complexity of the Viterbi
algorithm depends on the number of trellis states, given by N = 2
M
, where M is the
channel memory measured in bit intervals. In the Viterbi algorithm, complexity is
dominated by computation of the branch metrics. In each bit interval, 2NK branch
metrics must be evaluated by the receiver, where K is the number of samples per bit
interval [3]. Thus, the complexity per second for OOK with MLSD is at least
2NK/T
b
, where T
b
is the bit interval. The complexity order per bit for OOK with
MLSD is then at least
0 50 100 150 200 250 300 350 400
8
9
10
11
12
13
14
15
16
17
18
L (km)
R
e
q
u
i
r
e
d

O
S
N
R

(
d
B
)DC-OFDM
ACO-OFDM
SSB-OFDM
OOK with MLSD (Bosco)
81
K M O
M
T MLSD
b
= 2 2 ) (
,
, (3.27)
In our case, K = 2 and the memory M required for different values of
dispersion index are listed in Table 3.3 [3].
0.25 0.5 0.75 1
M 3 5 6 8
Table 3.3 Memory required for various values of the dispersion index for OOK with MLSD [3].
fast Fourier transform (FFT) algorithm. As shown in Chapter 2, an FFT of size N
requires 4Nlog
2
(N) real operations (multiplications plus additions) [26]. The overall
complexity order per bit for OFDM is then

OFDM b
Tx
T OFDM
T T N N N O
b
/ ) ( log ) (
2 ,
=
[ ]
OFDM b u
Rx
T OFDM
T T N N N N O
b
/ ) ( log ) (
2 ,
+ = ,
(3.28)
where T
OFDM
is the OFDM symbol period. Using data from Table 3.2 and Table 3.3
and Eqs. (3.27) and (3.28), in Table 3.4 we compare the computational complexity per
bit for SSB-OFDM and OOK with MLSD.

MLSD SSB-OFDM
Receiver Transmitter Receiver Total
0.25 32 47.9 48.8 96.7
0.5 128 46.8 47.8 94.6
0.75 256 52.1 53.1 105.2
1 1024 51.5 52.4 103.9
Table 3.4 Number of real operations required per bit for SSB-OFDM and OOK with MLSD for the
various dispersion indexes .
We observe that SSB-OFDM requires fewer operations per bit than OOK with
MLSD. The relatively low complexity of OFDM is due to the efficiency of the FFT
algorithm. The high complexity of OOK with MLSD is caused by the exponential
dependence of the number of trellis states on the channel memory length.

Local Area Networks
82
4 LOCAL AREA NETWORKS
4.1 Introduction
Multicarrier modulation has been proposed to combat inter-symbol
interference (ISI) in multimode fibers since the symbol period of each subcarrier can
be made long compared to the delay spread caused by modal dispersion [47,48,49].
The main drawback of multicarrier modulation in systems using intensity
modulation (IM) is the high DC bias required to make the OFDM waveform
nonnegative, as shown in Chapter 3. There are several approaches for reducing the DC
power such as DC-clipped OFDM (DC-OFDM) and asymmetrically clipped optical
OFDM (ACO-OFDM) [37] as discussed in Chapter 3. Recently, a new technique
called pulse-amplitude modulated discrete multitone (PAM-DMT) has been proposed
to reduce the DC power for systems using intensity modulation with direct-detection
(IM/DD). This technique clips the entire negative excursion of the waveform similarly
to ACO-OFDM, but clipping noise is avoided by modulating only the imaginary
components of the subcarriers [50].
There have been several studies of the different OFDM techniques (e.g., [51],
[52]) in multimode fibers, but to our knowledge, there has been no comparison of
power efficiencies among the various OFDM methods, and to conventional baseband
methods, such as on-off keying (OOK). Furthermore, in previous work, the powers of
the subcarriers were not optimized for finite bit allocation based on the channel
frequency response. We present an iterative procedure for DC-OFDM based on known
bit-loading algorithms with a new modification, the bias ratio (B
R
), in order to obtain
the optimum power allocation as presented in Chapter 3.
The optimum detection technique for unipolar pulse-amplitude modulation
(PAM) in the presence of ISI is maximum-likelihood sequence detection (MLSD), but
its computational complexity increases exponentially with the channel memory. ISI in
multimode fiber is well-approximated as linear in the intensity (instantaneous power)
Local Area Networks
83
[53], and for typical fibers, PAM with minimum mean-square error decision-feedback
equalization (MMSE-DFE) achieves nearly the same performance as MLSD and
requires far less computational complexity. Hence, we compare the performance of the
three aforementioned OFDM techniques using optimized power allocations to the
performance of PAM with MMSE-DFE at different spectral efficiencies.
This chapter is organized as follows. In section 4.2, we review local area
networks (LANs) and discrete bit allocation for multicarrier systems, known as the
Levin-Campello algorithm [54]. We present our system and fiber models in
section 4.3. In that section, we also discuss the performance measures used to compare
different modulations formats. In section 4.4, we review the different OFDM formats
and study analytically the performance differences between ACO-OFDM and PAM-
DMT. In section 4.5, we compare the receiver electrical SNR required to transmit at a
given bit rate for the different OFDM formats and for unipolar PAM with MMSE-
DFE equalization at different spectral efficiencies.
4.2 Local Area Networks and Discrete Bit Allocation Review
4.2.1 Local Area Networks
Local area networks (LANs) use multimode fiber (MMF) for its low cost and
ease of connection since MMF is more tolerant to misalignments. LANs typically
operate at 850 nm or 1330 nm and use low cost lasers like vertical-cavity surface-
emitting lasers (VCSEL). LANs can support bit rates between 1 Gbit/s and 10 Gbit/s
and typical distances are between 300 m to 5 km. Fig. 4.1 shows a block diagram of a
LAN.
Local Area Networks
84
Fig. 4.1 LAN block diagram.
LANs use neither optical amplification nor optical dispersion compensation in
the link. Fig. 4.2 shows a block diagram of a LAN link.
Fig. 4.2 LAN link.
LAN systems use intensity modulation with direct-detection (IM/DD) for its
simplicity and reduced hardware cost. IM can be achieved, for example, by direct
modulation of the laser current, as shown in Chapter 3.
4.2.2 Discrete Bit Allocation Review
As shown in Chapter 3, on a bipolar channel with AWGN, the optimum power
and bit allocations for an OFDM system are given by the water-filling solution. While
the optimal value for the number of bits (b = R
b
/B) for each subcarrier is an arbitrary
nonnegative real number, in practice, the constellation size and FEC code rate need to
be adjusted to obtain a rational number of bits. The optimal discrete bit allocation
method is known as the Levin-Campello algorithm [54]. We first choose the desired
bit granularity for each subcarrier, i.e., the bit allocation on each subcarrier will be
an integer multiple of . Next, we choose an initial bit allocation for all the subcarriers
(not necessarily optimal). We make the initial bit allocation optimum by using the
efficientizing (EF) algorithm [54]. Then, depending on the system design, we can
either use the E-tightening algorithm to obtain the maximum achievable bit rate for
a given total power P
t
or the B-tightening algorithm to obtain the minimum total
TX RX
MMF
Local Area Networks
85
power required to transmit at a constant bit rate R
b
. We can summarize the Levin-
Campello algorithm [54] as
Choose an initial bit distribution according to .
Make the initial bit distribution optimal using the EF algorithm.
Either use E-tightening to obtain the maximum achievable bit rate for a
given total power or B-tightening to obtain the minimum total power
required to transmit at a constant bit rate.
4.3 System Model for LANs
4.3.1 Overall System Model

Fig. 4.3 OFDM system model for LANs.
An optical OFDM transmitter encodes transmitted symbols onto an electrical
OFDM waveform and modulates this onto the intensity (instantaneous power) of an
optical carrier. The modulator can generate one of DC-OFDM, ACO-OFDM or PAM-
DMT. Details of modulators for particular OFDM schemes are described in Chapter 3
and in section 4.4.
We assume that there is no optical amplification in the system. After
propagating through the multimode fiber, the optical signal intensity is detected, and
the electrical current is lowpass filtered. Since we are trying to minimize the optical
power required to transmit at given bit rate R
b
, we assume the receiver operates in a
.
.
.
OFDM
Tx
Multimode
Fiber
Anti-Aliasing
Filter
OFDM
Rx
.
.
.
X
1
X
0
X
N1
X
N2
Y
N2
Y
N1
Y
0
Y
1
Laser
Local Area Networks
86
regime where signal shot noise is negligible, and the dominant noise is thermal noise
arising from the preamplifier following the photodetector. We model the thermal noise
as real baseband AWGN with zero mean and double-sided power spectral density
N
0
/2.
After lowpass filtering, the electrical OFDM signal is demodulated and
equalized with a single-tap equalizer on each subcarrier to compensate for channel
distortion [51], [52].
4.3.2 Multimode Fiber Model
There have been several studies on accurately modeling modal dispersion in
multimode fibers [55,56,57,58]. We can express a fibers intensity impulse response as
[55], [57]

( ) ( ) ) ( ) ( ) ( , t h z n t n z t h
pulse
n
MMF

=

, (4.1)
where the set (n) is known as the mode power distribution, (n) is the group delay per
unit length of the n
th
principal mode and h
pulse
(t) is the pulse shape of a principal mode
group. The parameters (n) and (n) depend not only on the fiber index profile but also
on the fiber input coupling, the input excitation, connectors offsets, and fibers bends.
Hence, these parameters are usually modeled as random variables.
The group delay parameter (n) is characterized by the differential-mode delay
(DMD), which is the difference between the fasted and slowest principal mode groups.
In our model, we choose (n) as an exponential random variable, (n) as a Gaussian
power profile and h
pulse
(t) as a second-order super-Gaussian pulse. We choose the full-
width at half-maximum (FWHM) of the pulse shape of a principal mode group
(h
pulse
(t)) to vary inversely with the mean group delay (E[(n)]) such that the fiber 3-
dB bandwidth would scale inversely proportional to the fiber length, as presented in
[57]. We adjust the parameters such that our set of fibers is very similar to the fibers in
[55], [56] at a wavelength of 850 nm. Specifically, we have created 1728 fibers with
DMDs between 0.2 and 0.7 ns/km and with 19 principal mode groups propagating.
Local Area Networks
87
Fig. 4.4 shows the mode power distribution, mode delays and frequency
response for an exemplary fiber from our set. In our study, we choose a fiber length of
1 km.
Fig. 4.4 (a) Mode power distribution, (b) mode delays and (c) frequency response for fiber 1183. We
consider a 1-km length in computing the delays and frequency response.
Fig. 4.5 shows the 3-dB bandwidth distribution of all the fibers used in our
analysis.
0 5 10 15 20
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Principal Mode Number
M
o
d
e

G
r
o
u
p

D
e
l
a
y

(
n
s
/
k
m
)
0 5 10 15 20 25
-60
-50
-40
-30
-20
-10
0
Frequency (GHz)
2
0

l
o
g
1
0
(
|
H
(
f
)
|
)

(
d
B
)
0 5 10 15 20
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
Principal Mode Number
M
o
d
e

P
o
w
e
r

D
i
s
t
r
i
b
u
t
i
o
n
b) a)
c)
Local Area Networks
88

Fig. 4.5 3-dB bandwidth distribution of the multimode fibers simulated. All fibers have 1-km length.
4.3.3 Performance Measures
Our baseband channel model is given by
) ( ) ( ) ( ) ( t n t h t x R t y
MMF
+ = , (4.2)
where y(t) is the electrical detected signal, x(t) is the input intensity waveform, h
MMF
(t)
is the fiber intensity impulse response, n(t) is the thermal noise from the photodetector
preamplifier, and R is the photodetector responsivity (A/W). This model differs from
conventional electrical systems because the channel input represents instantaneous
optical power (i.e. intensity). Hence, the channel input is nonnegative ( 0 ) ( t x ) and
the average transmitted optical power P
opt
is given by
[ ] ) ( ) (
2
1
lim
t x E dt t x
T
P
T
T
T
opt
= =

, (4.3)
2
(i.e., E[|x(t)|
2
x(t) represents amplitude. The average received optical power can be written as

opt MMF
P H P ) 0 ( = ,
(4.4)
where H
MMF
(0) is the DC gain of the channel, i.e.,

= dt t h H
MMF MMF
) ( ) 0 ( .
0 1 2 3 4 5 6 7 8 9 10 11 12 13
0
5
10
15
20
25
3 dB Bandwidth (GHz)
F
i
b
e
r

D
i
s
t
r
i
b
u
t
i
o
n

(
%
)
Local Area Networks
89
In order to facilitate comparison of the average optical power requirements of
different modulations techniques at a fixed bit rate, we define electrical SNR as [59]

b
opt MMF
b
R N
P H R
R N
P R
SNR
0
2 2 2
0
2 2
) 0 (
= = , (4.5)
where R
b
is the bit rate and N
0
is the (single-sided) noise power spectral density. We
see that the SNR given by Eq. (4.5) is proportional to the square of the received
optical signal power P, in contrast to conventional electrical systems, where it is
proportional to the received electrical signal power. Hence, a 2-dB change in the
electrical SNR (Eq. (4.5)) corresponds to a 1-dB change in average optical power.
A useful measure of the ISI introduced by a multimode fiber is the temporal
dispersion of the impulse response expressed by the channel root-mean-square (rms)
delay spread D [60]. The rms delay spread D can be calculated as [60]

2 / 1
2
2 2
) (
) ( ) (
dt t h
dt t h t
D
MMF
MMF
, (4.6)
where the mean delay is given by

=
dt t h
dt t h t
MMF
MMF
) (
) (
2
2
. (4.7)
Fig. 4.6 shows the channel root-mean-square (rms) delay spread D distribution
of all the fibers used in our analysis.
Local Area Networks
90

Fig. 4.6 Rms delay spread (D) distribution of the multimode fibers simulated. All fibers have 1-km
length.
We will also use the normalized delay spread D
T
, which is a dimensionless
parameter defined as the rms delay spread D divided by the bit period T
b
(T
b
= 1/R
b
):

b
T
T
D
D = .
(4.8)
4.4 Analysis of IM/DD OFDM Schemes
4.4.1 DC-Clipped OFDM
We use the same DC-OFDM format with hard-clipping and bias ratio (B
R
)
optimization described in Chapter 3.4.1.
For a high number of subcarriers, we can model the electrical OFDM signal
x(t) as a Gaussian random variable with mean equal to the DC bias and variance equal
to the electrical power ] | ) ( [|
2 2
t x E = . After hard-clipping at zero, we obtain only the
positive side of the Gaussian distribution. The average optical power is equal to the
average of the clipped waveform and can be written as

[ ] dx N x t x E P
clip opt

+
) = =
0
2
, ( ) ( , (4.9)
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
0
2
4
6
8
10
12
Rms Delay Spread, D (ns)
F
i
b
e
r

D
i
s
t
r
i
b
u
t
i
o
n

(
%
)
Local Area Networks
91
where N(,
2
) is the Gaussian pdf with mean and variance
2
. Doing a variable
change, z = x , we obtain

[ ]
2
2
2
2 2
2
2
1
, 0 ( , 0 (
, 0 ( ) ( ) (
=
) + ) =
) + =

e Q
dz N z dz N
dz N z t x E
clip
,
(4.10)
where Q is the Gaussian Q function [4]. The average optical power for DC-OFDM is
then
[ ]

+ = =

Q e t x E P
clip OFDM DC
1
2
) (
2
2
2
. (4.11)
We note that if the DC bias is chosen such that there is no clipping ( ),
then the optical power is equal to the DC bias , as expected. If we choose the DC bias
to be proportional to the square root of the electrical power, i.e., = B
R
where B
R
is
the bias ratio discussed in Chapter 3.4.1, then minimizing the electrical power also
minimizes the optical power required.
4.4.2 Asymmetrically Clipped Optical OFDM
In our study, we use the same ACO-OFDM format described in Chapter 3.4.2.
For a high number of subcarriers, we can model again the electrical OFDM signal x(t)
as a Gaussian random variable but for ACO-OFDM the electrical signal has a zero
mean. After hard-clipping at zero, we obtain again only the positive side of the
Gaussian distribution. Armstrong and Lowery showed that for ACO-OFDM, the
average and variance of the electrical waveform are 2 / and 2 /
2
, respectively
[61]. Hence, the average optical power for ACO-OFDM is [61]

2
=
OFDM ACO
P .
(4.12)
Local Area Networks
92
As we can observe in Eq. (4.12), the average optical power is proportional to
the square root of the electrical power. We can obtain the same result as Armstrong for
ACO-OFDM [61] by setting the DC bias to = 0 and Eq. (4.11) simplifies to
Eq. (4.12).
4.4.3 PAM-Modulated Discrete Multitone
Similar to ACO-OFDM, PAM-DMT [50] clips the entire negative excursion of
the electrical waveform to minimize average optical power. Lee and Koonen [50]
showed that if data is modulated using PAM only on the imaginary components of the
subcarriers, clipping noise does not affect system performance since that noise is real
valued, and is thus orthogonal to the modulation. Fig. 4.7 shows a block diagram of a
PAM-DMT transmitter.

Fig. 4.7 Block diagram of a PAM-OFDM transmitter.
As in DC-OFDM and ACO-OFDM, the OFDM subcarriers for PAM-DMT are
assumed to have Hermitian symmetry, so that the time-domain waveform is real.
Contrary to ACO-OFDM, in PAM-DMT all of the subcarriers are used, but the
modulation is restricted to just one dimension. Hence, PAM-DMT has the same
spectral efficiency as ACO-OFDM. After D/A conversion, the electrical OFDM signal
is hard-clipped at zero and intensity modulated onto an optical carrier.
Optical Spectrum
X
k
X
k
*
f
X
N/21
X
1
X
0
IDFT
X
1
*
Parallel
to
Serial
+
Cyclic
Prefix
.
.
.
D/A
X
N/21
*
.
.
.
Clip
Real
Signal
Fiber
X
N/2
0
jPAM
jPAM
Local Area Networks
93
Using an analysis similar to that for ACO-OFDM, Lee and Koonen showed
that the average transmitted optical power for PAM-DMT is given by [50]

2
=
DMT PAM
P ,
(4.13)
where is the square root of the electrical power of the unclipped OFDM waveform.
Up to our best knowledge, no comparison has been done between the
performance of ACO-OFDM and PAM-DMT in frequency selective channels. We
first compare the performance in the limit of a small number of subcarriers. The
minimum number of subcarriers is five, such that two are used for data, the other two
are Hermitian conjugates of the data subcarriers and the fifth is the DC subcarrier,
which is set to zero. For ACO-OFDM, only one of the two possible subcarriers is used
for data transmission. Thus, for a given total power P
t
, the normalized bit rate is given
by

+ =
2
2
1
2
| |
1 log
H P
b
t
ACO
,
(4.14)
where H
1
is the channel gain at the first subcarrier.
In PAM-DMT, all the subcarriers are used for data transmission but the
modulation is restricted to one dimension, i.e., PAM. The optimum power allocation
for PAM-DMT is given by the waterfiling algorithm in Chapter 3.2.2.2 and can be
written as

( )
a
H

P
n
n
n
= +
2
2
2
.
(4.15)
We note that the noise variance in Eq. (4.15) is half than in ACO-OFDM
because in PAM-DMT only one dimension is being modulated. The constant a can be
found by knowing that P
t
= P
1
+ P
2
. The optimum power allocation for the two
subcarriers is then
Local Area Networks
94

=
+ =
2
2
2
1
t opt
t opt
P
P
P
P
,
(4.16)
where is given by

=
2
1
2
2
2
1 1
4
1
H H
.
(4.17)
The maximum normalized bit rate for PAM-DMT is

( ) ( )
+
+ =
+ +
+ =
2
2
1
2
2
1 log
2
1 log
2
1
2
1 log
2
1
2
2
2
2
2
1
2
2
2
2 2
2 2
2
1 1
2

H P H P
H P H P
b
t t
opt opt
PAM
.
(4.18)
We now assume that the channel has a lowpass frequency response, such that
|H
2
| < |H
1
|. This assumption is valid for the majority of multimode fibers. We obtain an
upper bound on Eq. (4.18) as

( ) ( )
( )
ACO
t
t
t t
PAM
b
H P
H H P
H P H P
b
<

+ <
+
+
2
2
1
2
2
2
2
1
2
2
2
1
2
2
2
1
2
2
1
2
1 log
2 2
2
1 log
2
2
1
2
2
1 log
.
(4.19)
From Eq. (4.19), we observe that ACO-OFDM is more power efficient than
PAM-DMT for a small number of subcarriers on lowpass channels. Furthermore, from
Eq. (4.19), we see that if the channel is flat, i.e., |H
1
| = |H
2
|, then = 0 and both
techniques have exactly the same power efficiency.
For a high number of subcarriers, we can apply a similar analysis for each pair
of used subcarriers. Since |H
n+1
| |H
n
| for a high number of subcarriers, the power
Local Area Networks
95
efficiency of PAM-DMT converges asymptotically to that of ACO-OFDM for
lowpass channels in the limit of a high number of subcarriers.
4.5 Comparison of IM/DD Modulation Formats
The system model is shown in Fig. 4.3. As mentioned previously, we neglect
all transmission impairments except for modal dispersion of the multimode fiber. We
employ the fiber model discussed in section 4.3.2, assuming a fiber length of 1 km and
transmission at 850 nm.
We assume the dominant noise is thermal noise, modeled as real baseband
AWGN with zero mean and double-sided power spectral density N
0
/2. We choose N
0

= 10
22
A
2
/Hz, which is a typical value for commercial optical receivers. We assume a
photodetector quantum efficiency of 90%, corresponding to responsivity R = 0.6 A/W
at 850 nm. At the receiver, the anti-aliasing filter is a fifth-order Butterworth lowpass
filter. For OFDM, we set the 3-dB cutoff frequency of the anti-aliasing filter equal to
the first null of the OFDM spectrum [25]. For M-PAM, we set the 3-dB cutoff
frequency to 0.8R
s
,

where R
s
is the symbol rate.
Typical high-performance FEC codes for optical systems have a threshold bit-
error ratio (BER) of the order of P
b
= 10
3
. In order to provide a small margin, we
compute the minimum required SNR to achieve P
b
= 10
4
for the different modulation
formats. A BER P
b
= 10
4
corresponds to a gap of = 6.6 dB. We choose a granularity
= 0.25 for the discrete bit loading algorithms, since it is straightforward to design
practical codes whose rates are multiples of 0.25.
We let R
s
denote the symbol rate for unipolar M-PAM (in the special case of 2-
PAM or OOK, R
s
= R
b
). We let
OFDM
s
R denote the equivalent symbol rate for OFDM
[25]. In an attempt to provide a fair comparison between unipolar M-PAM and
OFDM, unless stated otherwise, we will let the two symbol rates be equal,
s
OFDM
s
R R =
. For all OFDM formats, we use an oversampling ratio of M
s
= 64 /52 1.23 to avoid
noise aliasing. In this case, OFDM requires an analog-to-digital (A/D) converter
Local Area Networks
96
sampling frequency of
OFDM
s s
R M . While an equalizer for M-PAM can employ an
arbitrary rational sampling frequency such as 3/2 to achieve good performance, an
oversampling ratio of 2 is often chosen because it yields slightly better performance
than 3/2, while greatly simplifying the equalizer structure [16]. Assuming a bit rate R
b

= 10 Gbit/s and
s
OFDM
s
R R = , the required A/D sampling frequency is 12.3 GHz for
OFDM and 20 GHz for OOK.
Fig. 4.8 shows the receiver electrical SNR required to achieve a bit rate of 10
Gbit/s for several fibers of 1-km length using ACO-OFDM and OOK.

b
= 10
4
for ACO-OFDM and OOK
for fibers with 1-km length. The bit allocation granularity is = 0.25 and ACO-OFDM has the same
symbol rate as OOK (
OFDM
s
R = R
s
= 10 GHz).
In order to make our results independent of the bit rate, we present our results
in Fig. 4.8 as a function of the normalized delay spread D
T
. The symbol rate is the
same for both modulation schemes, i.e.,
OFDM
s
R = R
s
= 10 GHz. For ACO-OFDM, the
FFT size is N = 1024 and the number of used subcarriers is N
u
= 208. We set the cyclic
prefix equal to the duration (in samples) of the worst fiber impulse response,
corresponding to N
pre
= 14 samples measured at the OFDM symbol rate. For OOK, we
use a fractionally spaced MMSE-DFE at an oversampling ratio of two. We use the
0.1 0.2 0.3 0.4 0.5 1 2
8
10
12
14
16
18
20
22
24
26
Normalized Delay Spread (D
T
)
R
e
c
e
i
v
e
r

E
l
e
c
t
r
i
c
a
l

S
N
R

(
d
B
)
ACO-OFDM (R
s
)
MMSE-DFE OOK (R
s
)
Local Area Networks
97
same number of taps for all fibers, which is chosen based on the worst fiber: 41 taps
for the feedforward filter and 15 taps for the feedback filter.
In Fig. 4.8, we observe that ACO-OFDM requires a higher SNR than OOK for
all fibers in our set when both modulations use the same symbol rate. In this case,
which corresponds to a spectral efficiency of 1 bit/s/Hz, the performance difference is
about 1 dB. ACO-OFDM requires a higher SNR than OOK because it requires an
average of 4 bits (16-QAM) on the used subcarriers to compensate the information
rate loss when half of the subcarriers are set to zero. However, we can improve the
performance of OFDM by using more optical bandwidth (which also requires more
electrical bandwidth). At a symbol rate (or baud rate) R
s
, M-PAM requires an optical
bandwidth of approximately 2R
s
, which corresponds to the interval between spectral
nulls on either side of the carrier. On the other hand, OFDM requires a bandwidth of
approximately
OFDM
s
R [25]. Hence, for the same symbol rate as M-PAM (
OFDM
s
R = R
s
),
OFDM requires half the bandwidth of M-PAM. If we double the symbol rate for
OFDM, both modulations schemes use approximately the same optical bandwidth
1

2R
s
, as shown in Fig. 4.9.
Fig. 4.9 Optical spectra of M-PAM and OFDM. The symbol rate for OFDM is twice that for M-PAM,
OFDM
s
R = 2R
s
.
Fig. 4.10 shows the receiver electrical SNR required to achieve a bit rate of 10
Gbit/s for several fibers of length 1 km when all the OFDM formats use twice the
1
The optical bandwidths stated here assume that the source linewidth is small compared to the
modulation bandwidth. If the source linewidth is large, changing the symbol rate may have little effect
on the optical bandwidth, although it does affect the electrical bandwidth.
s
R 2
f
( ) f
s
R PAM
S
) (
( ) f
s
R OFDM
S
) 2 (
OFDM
*
f
OFDM
s
R 2
Local Area Networks
98
symbol rate of OOK, i.e.,
OFDM
s
R

= 2R
s
= 20 GHz. The sampling frequency required
for OFDM in this case is 1.2320 = 24.6 GHz. We set again the cyclic prefix equal to
the duration of the worst fiber impulse response, obtaining N
pre
= 24 samples
measured at the OFDM symbol rate. The prefix penalty is computed as in [25]. We
choose the number of used subcarriers N
u
equal to 416 for DC-OFDM, such that the
prefix penalty is negligible. The FFT size is N = 1024 for DC-OFDM. For ACO-
OFDM, the FFT size is N = 2048, such that the number of used subcarriers is the same
as DC-OFDM. For PAM-DMT, we set the FFT size to N = 2048 (N
u
= 832), in order
to make a fair comparison with ACO-OFDM.

Fig. 4.10 Receiver electrical SNR required to achieve P
b
= 10
4
at 10 Gbit/s for different modulations
formats in fibers with 1-km length. The bit allocation granularity is = 0.25. The symbol rate for all
OFDM formats is twice that for OOK,
OFDM
s
R = 2R
s
= 20 GHz.
For DC-OFDM, we perform the power allocation iteratively using the bias
ratio (B
R
) as described in Chapter 3.5. The minimum required optical SNR is obtained
by performing an exhaustive search on the B
R
value, as shown in Fig. 4.11.
0.1 0.2 0.3 0.4 0.5 1 2
5
10
15
20
25
30
T
)
R
e
c
e
i
v
e
r

E
l
e
c
t
r
i
c
a
l

S
N
R

(
d
B
)
ACO-OFDM (2R
s
)
MMSE-DFE OOK (R
s
)
DC-OFDM (2R
s
)
Local Area Networks
99

Fig. 4.11 Receiver electrical SNR required for DC-OFDM to achieve P
b
= 10
4
at 10 Gbit/s for
different values of the bias ratio in fiber 295.
In Fig. 4.10, we observe again that OOK with MMSE-DFE requires the lowest
SNR to achieve 10 Gbit/s at P
b
= 10
4
for most fibers, even when all OFDM formats
use twice the symbol rate of OOK. We note that when OFDM uses twice the symbol
rate of OOK, OFDM is allowed to use twice the bandwidth it would normally require
to transmit the same bit rate as OOK. Even with this advantage, all OFDM formats
still perform worse than OOK for the majority of the fibers, as shown in Fig. 4.10.
However, for fibers with a large bandwidth (> 7 GHz), ACO-OFDM outperforms
OOK. We note that in computing Fig. 4.10 for OOK with MMSE-DFE, we have used
correct decisions at the input of the feedback filter. If we use detected symbols for the
feedback filter input, the required SNR increases by 0.4 dB for fiber bandwidths less
than 4 GHz, due to error propagation. For fiber bandwidths beyond 4 GHz, the
increase in required SNR is less than 0.2 dB. The difference between ACO-OFDM
and OOK is within 0.8 and 1 dB, so OOK with error propagation is still more power
efficient than ACO-OFDM.
DC-OFDM requires the highest SNR because of the DC bias required to make
the OFDM waveform non-negative. Fig. 4.11 shows the receiver electrical SNR
required for DC-OFDM for different B
R
values. We can see that if the B
R
is too high,
then the required SNR is also high, because of the power wasted in the DC bias. On
1 1.5 2 2.5 3 3.5
24.5
25
25.5
26
26.5
27
27.5
Bias Ratio
R
e
c
e
i
v
e
r

E
l
e
c
t
r
i
c
a
l

S
N
R

(
d
B
)
Local Area Networks
100
the other hand, if the B
R
is too low, more power has to be allocated to each subcarrier
to compensate for the high clipping noise, and therefore the required SNR increases.
Fig. 4.12 compares ACO-OFDM and PAM-DMT using = 0.25 with ACO-
OFDM using continuous bit allocation. In Fig. 4.12, we observe that there is no
significant performance difference between PAM-DMT and ACO-OFDM. The
difference in SNR between ACO-OFDM and PAM-DMT is less than 0.1 dB. This is
to be expected, since for a high number of subcarriers, PAM-DMT converges
asymptotically to the same performance as ACO-OFDM. Since PAM-DMT and ACO-
OFDM have the same performance, we choose ACO-OFDM as the OFDM format for
comparison for the remainder of the chapter.

Fig. 4.12 Receiver electrical SNR required for various OFDM formats with continuous and discrete bit
allocations to achieve P
b
= 10
4
at 10 Gbit/s for fibers of 1-km length. The symbol rate for all OFDM
formats is
OFDM
s
R = 20 GHz.
We also see that there is no significant difference between ACO-OFDM with
discrete loading ( = 0.25) and ACO-OFDM with continuous bit allocation. This
means that using OFDM with coding rates that are multiples of 0.25 is sufficient to
achieve optimal power performance. The subcarrier power distribution for discrete and
continuous ACO-OFDM is shown in Fig. 4.13.
0.1 0.2 0.3 0.4 0.5 1 2
6
8
10
12
14
16
18
20
22
24
26
R
e
c
e
i
v
e
r

E
l
e
c
t
r
i
c
a
l

S
N
R

(
d
B
)
T
)
ACO-OFDM (Discrete, = 0.25)
PAM-DMT (Discrete, = 0.25)
ACO-OFDM (Continuous)
Local Area Networks
101

Fig. 4.13 Subcarrier power distribution for ACO-OFDM with continuous bit allocation and with
discrete bit loading (with granularity = 0.25) for 10 Gbit/s at P
b
= 10
4
in fiber 10. The symbol rate is
OFDM
s
R = 20 GHz.
Fig. 4.14 shows the cumulative distribution function (CDF) of the required
electrical SNR, which is defined as
( ) SNR x SNR = Prob ) ( CDF , (4.20)
where SNR takes all possible values for the required SNR for a given modulation
format. We assume that all 1728 fiber realizations occur with equal probability. The
CDF corresponds to the fraction of channels on which the target P
b
is reached at a
given SNR. For example, Fig. 4.14 shows that if the electrical SNR is at least 25 dB,
the target P
b
10
4
is met for all the channels when using MMSE-DFE OOK at 10
Gbit/s.
In Fig. 4.14, we observe that OOK with MMSE-DFE requires less SNR for the
majority of the fibers, as expected. Furthermore, we also notice that doubling the
symbol rate for ACO-OFDM only improves the performance for fibers with low RMS
delay spreads. For high RMS delay spreads, doubling the symbol rate gives practically
no performance increase.

0 100 200 300 400 500
0
0.2
0.4
0.6
0.8
1
1.2
Data Subcarrier Number
S
u
b
c
a
r
r
i
e
r

P
o
w
e
r

(
P
n
/

P
n
,
m
a
x
)
Continuous
Discrete ( = 0.25)
Local Area Networks
102

Fig. 4.14 Cumulative distribution function (CDF) of the required receiver electrical SNR to obtain P
b
=
10
4
at 10 Gbit/s for different modulation formats. The symbol rate for OOK is R
s
= 10 GHz and the
symbol rate for OFDM is the same or twice that for OOK, as indicated in the figure.
It is also interesting to check if the same performance difference is obtained at
higher spectral efficiencies, i.e., for unipolar M-ary PAM (M-PAM). Fig. 4.15 and Fig.
4.16 show the receiver electrical SNR required to achieve a bit rate of 20 Gbit/s for
ACO-OFDM and unipolar 4-PAM. The symbol rate for 4-PAM is kept constant at 10
GHz.

b
= 10
4
for the different
5 10 15 20 25 30
0
0.2
0.4
0.6
0.8
1
1.2
Receiver Electrical SNR (dB)
C
D
F
(
S
N
R
)
ACO-OFDM and
PAM-DMT (2R
s
)
DC-OFDM (2R
s
)
MMSE-DFE OOK (R
s
)
ACO-OFDM and
PAM-DMR (R
s
)
0.2 0.3 0.4 0.5 1 2 3 4
5
10
15
20
25
30
35
R
e
c
e
i
v
e
r

E
l
e
c
t
r
i
c
a
l

S
N
R

(
d
B
)
ACO-OFDM (2R
s
)
ACO-OFDM (R
s
)
MMSE-DFE 4-PAM (R
s
)
T
)
Local Area Networks
103
modulations formats for fibers of 1-km length. The bit allocation granularity is = 0.25 and the
symbol rate for 4-PAM is R
s
= 10 GHz. The symbol rate for ACO-OFDM is the same or twice that for
4-PAM, as indicated in the figure.
In Fig. 4.15, we observe that 4-PAM requires the lowest SNR to achieve 20
Gbit/s at P
b
= 10
4
for most fibers. This can also be easily seen in Fig. 4.16. We
observe that the difference in SNR between ACO-OFDM and unipolar 4-PAM is
about 4 dB when both modulations use the same symbol rate. We conclude that
increasing the spectral efficiency from 1 bit/s/Hz (Fig. 4.8) to 2 bit/s/Hz (Fig. 4.15),
increases difference in SNR requirement between ACO-OFDM and M-PAM from 1
dB to 4 dB. This is to be expected, since increasing the spectral efficiency from 1
bit/s/Hz to 2 bit/s/Hz for ACO-OFDM requires doubling the average number of bits on
each subcarrier from 4 bits (16-QAM) to 8 bits (256-QAM) when the symbol rate for
ACO-OFDM is the same as that for M-PAM.

Fig. 4.16 Cumulative distribution function (CDF) of the required receiver electrical SNR to obtain 20
Gbit/s at P
b
= 10
4
for different modulation formats. The symbol rate for 4-PAM is R
s
= 10 GHz and
the symbol rate for ACO-OFDM is the same or twice that for 4-PAM, as indicated in the figure.
We also observe that ACO-OFDM with twice the symbol rate of 4-PAM has
practically the same performance as 4-PAM. Furthermore, ACO-OFDM now
outperforms M-PAM for a larger fraction of fibers. However, this performance
increase for ACO-OFDM requires doubling the electrical bandwidth and increasing
the sampling frequency from 12.3 GHz to 24.6 GHz, which is 23% higher than the
5 10 15 20 25 30 35 40
0
0.2
0.4
0.6
0.8
1
1.2
Receiver Electrical SNR (dB)
C
D
F
(
S
N
R
)
MMSE-DFE 4-PAM (R
s
)
ACO-OFDM and
PAM-DMT (2R
s
)
ACO-OFDM and
PAM-DMR (R
s
)
Local Area Networks
104
sampling frequency required for 4-PAM using an oversampling ratio of two, i.e., 20
GHz.

Personal Area Networks
105
5 PERSONAL AREA NETWORKS
5.1 Introduction
Indoor optical wireless transmission has been studied extensively in recent
decades [62,59,63,64]. The visible and infrared spectral regions offer virtually
unlimited bandwidth that is unregulated worldwide. Light in the infrared or visible
range penetrates through glass, but not through walls or other opaque barriers, so that
optical wireless transmissions are confined to the room in which they originate.
Furthermore, in a visible or infrared wireless link employing intensity modulation with
direct detection (IM/DD), the short carrier wavelength and large-area photodetector
lead to efficient spatial diversity that prevents fading. Nevertheless, the existence of
multiple paths between the transmitter and receiver causes multipath distortion,
particularly in links using non-directional transmitters and receivers, or in links relying
upon non-line-of-sight propagation [62], [59]. This multipath distortion can lead to
inter-symbol interference (ISI) at high bit rates.
Multicarrier modulation has been proposed to combat ISI in optical wireless
links, since the symbol period of each subcarrier can be made long compared to the
delay spread caused by multipath distortion [65]. Multicarrier modulation is usually
implemented by orthogonal frequency-division multiplexing (OFDM) [63,66,38]. The
main drawback of multicarrier modulation in systems using intensity modulation (IM)
is the high DC bias required to make the multicarrier waveform nonnegative. There
have been several approaches for reducing the DC bias in IM OFDM systems, such as
DC-clipped OFDM (DC-OFDM) [63], [66], asymmetrically clipped optical OFDM
(ACO-OFDM) [38] and PAM-modulated discrete multitone (PAM-DMT) [50].
There have been several studies comparing the performance of different
OFDM techniques (e.g., [38]) but these comparisons have been made for ideal
additive white Gaussian noise (AWGN) channels. To our knowledge, previous studies
have not compared the OFDM methods to conventional baseband methods, such as
106
on-off keying (OOK) or unipolar pulse-amplitude modulation (PAM), nor have they
considered the dispersive nature of optical wireless channels. Furthermore, in previous
work, the powers of the subcarriers and the DC bias for DC-OFDM were not jointly
optimized according to the channel frequency response in order to obtain the lowest
required optical power. We present an iterative procedure for DC-OFDM based on
known bit-loading algorithms with a new modification, the bias ratio, in order to
obtain the optimum power allocation.
The optimum detection technique for unipolar PAM in the presence of ISI is
maximum-likelihood sequence detection (MLSD), but its computational complexity
increases exponentially with the channel memory. ISI in optical wireless links is well-
approximated as linear in the instantaneous power [59], and for typical wireless links,
PAM with minimum mean-square error decision-feedback equalization (MMSE-DFE)
achieves nearly the same performance as MLSD and requires far less computational
complexity. Hence, we compare the performance of the three aforementioned OFDM
techniques using optimized power allocations to the performance of PAM with
MMSE-DFE at different spectral efficiencies.
This chapter is organized as follows. We introduce the optical wireless
networks in section 5.2. We present our system and indoor optical wireless models in
section 5.3. In section 5.4, we compare the receiver electrical SNR required to transmit
at several bit rates for the different OFDM formats and for unipolar M-PAM with
MMSE-DFE equalization at different spectral efficiencies. Furthermore, we compare
the receiver electrical SNR required for the different modulation formats when there is
no channel state information (CSI) available at the transmitter. Finally, we also
compare the computational complexity required for OFDM and M-PAM at different
bit rates in section 5.5.
5.2 Personal Area Networks
Personal area networks (PANs) typically use optical wireless for its low cost.
Indoor optical wireless networks operate at infrared or visible light since these spectral
107
regions offer virtually unlimited bandwidth that is unregulated worldwide. Optical
wireless networks typically use noncoherent light sources such as light-emitting
diodes (LED). LEDs emit an average power of several tens of mW that is concentrated
within a semiangle of 15
30
. LED emission wavelength typically lies between 850

and 950 nm and are cheaper than lasers. PANs can support bit rates between 10 Mbit/s
and 300 Mbit/s and the transmission is normally confined within a room. Fig. 5.1
shows a block diagram of an indoor optical wireless system.
Fig. 5.1 Indoor optical wireless transmission.
Indoor optical wireless use neither optical amplification nor optical dispersion
compensation in the link. Typically, the transmitter is placed in the ceiling and the
receiver is located at desk floor (~ 1 m), as shown in Fig. 5.1. Optical wireless systems
use intensity modulation with direct-detection (IM/DD) for its simplicity and reduced
hardware cost.
5.3 System Model and Performance Measures
5.3.1 Overall System Model
108

Fig. 5.2 OFDM system model for LANs.
An OFDM modulator encodes transmitted symbols onto an electrical OFDM
waveform and modulates this onto the instantaneous power of an optical carrier at
infrared or visible frequencies. The modulator can generate one of DC-OFDM, ACO-
OFDM or PAM-DMT. Details of modulators for particular OFDM schemes are
described in Chapter 4.4. After propagating through the indoor wireless link, the
optical signal intensity is detected and the electrical photocurrent is low-pass filtered.
Since we are trying to minimize the optical power required to transmit at given bit rate
R
b
, we assume the receiver operates in a regime where signal shot noise is negligible,
and the dominant noise is the shot noise from detected background light or thermal
noise from the preamplifier following the photodetector. After low-pass filtering, the
electrical OFDM signal is demodulated and equalized with a single-tap equalizer on
each subcarrier to compensate for channel distortion [25].
5.3.2 Optical Wireless Channel
Multipath propagation in an indoor optical wireless channel [67], [68] can be
described by an impulse response h(t) or by the corresponding baseband frequency
response

= dt e t h f H
ft j 2
) ( ) ( . Including noise, the baseband channel model is
[59]:

) ( ) ( ) ( ) ( t n t h t x R t y + = , (5.1)
where y(t) is the detected photocurrent, x(t) is the transmitted intensity waveform, R is
the photodetector responsivity, and n(t) represents ambient light shot noise and
.
.
.
Optical
OFDM
Tx
Indoor Wireless
Channel
Anti-Aliasing
Filter
OFDM
Rx
.
.
.
X
1
X
0 Y
0
Y
1
h(t)
Noise
+
X
N2
X
N1
Y
N2
Y
N1
109
thermal noise. Optical wireless channels differ from electrical or radio frequency
channels because the channel input x(t) represents instantaneous optical power. Hence,
the channel input is nonnegative ( 0 ) ( t x ) and the average transmitted optical power
P
t
is given by
[ ] ) ( ) (
2
1
lim
t x E dt t x
T
P
T
T
T
opt
= =

, (5.2)
2
(i.e., E[|x(t)|
2
x(t) represents amplitude. The average received optical power can be written as

t
P H P ) 0 ( = , (5.3)
where H(0) is the DC gain of the channel, i.e.,

= dt t h H ) ( ) 0 ( .
We use the methodology developed by Barry et al [67] to simulate the impulse
responses of indoor optical wireless channels, taking account of multiple bounces. A
similar model can be found in [69]. The algorithm in [67] partitions a room into many
elementary reflectors and sums up the impulse response contributions from k
th
-order
bounces,
( )
, 2 , 1 , 0 ), ( = k t h
k
. More recently, Carruthers developed an iterative
version of the multi-bounce impulse response algorithm which greatly reduces the
computational time required to accurately model optical wireless channels [70]. In this
study, we use a toolbox developed by Carruthers to implement the algorithm in [70].
We place the transmitter and receiver in a room with dimensions 8 m 6 m
3 m (length, width, height). Room surfaces are discretized with a spatial resolution of
0.2 m, and are assumed to have diffuse reflectivities as in configuration A in [67]. All
the source and receiver parameters are the same as in configuration A in [67]. We
assume the transmitter has a Lambertian radiation pattern, with intensity per unit solid
angle proportional to the cosine of the angle with respect to the transmitter normal
[67]. We also assume that the receiver area is equal to 1 cm
2
and it only detects light
whose angle of incidence is less than 90 with respect to the receiver normal [67].
According to [70], considering three to five bounces in calculating the impulse
110
response should be sufficient to accurately model most indoor environments with
typical reflectivities and geometries. Hence, we consider five bounces.
We generate 170 different channels by placing the transmitter in different
locations within the room. In all channel realizations, the receiver is placed in the
middle of the room 1 m above the floor and pointed upward. Our channel ensemble
includes line-of-sight (LOS) configurations (transmitter placed at the ceiling and
pointed down) and diffuse configurations (transmitter placed 1 m above the floor and
pointed up). In order to simulate shadowing by a person or object next to the receiver,
we block the LOS path (i.e., h
(0)
(t)) of some of the impulse responses. Fig. 5.3 shows
the impulse response for an exemplary diffuse channel from our set.

Fig. 5.3 Impulse response of an exemplary non-directional, non-LOS (diffuse) channel. This channel
has no LOS component h
(0)
(t). The contributions of the first five reflections, h
(1)
(t),, h
(5)
(t), are
shown.
5.3.3 Performance Measures
In order to compare the average optical power requirements of different
modulation techniques at a fixed bit rate, we define electrical SNR as [59]

b
t
b
R N
P H R
R N
P R
SNR
0
2 2 2
0
2 2
) 0 (
= = ,
(5.4)
10 20 30 40 50 60 70 80 90 100
0
2
4
6
8
10
12
14
16
18
20
Time (ns)
I
m
p
u
l
s
e

r
e
s
p
o
n
s
e

(
s
-
1
)
h
(1)
(t)
h
(2)
(t)
h
(3)
(t)
h
(4)
(t)
h
(5)
(t)
h(t)
111
where R
b
is the bit rate and N
0
is the (single-sided) noise power spectral density. We
see that the SNR given by Eq. (5.4) is proportional to the square of the received
optical signal power P, in contrast to conventional electrical systems, where it is
proportional to the received electrical signal power. Hence, a 2-dB change in the
required electrical SNR Eq. (5.4) corresponds to a 1-dB change in the required average
optical power.
Given a channels impulse response h(t), we compute the channels root-mean-
square (rms) delay spread D using [68], [60]
( ) ( )
2 / 1
2
2 2
) (
) ( ) (

dt t h
dt t h t
t h D

,
(5.5)
where the mean delay is given by

=
dt t h
dt t th
) (
) (
2
2
.
(5.6)
We also compute the normalized delay spread D
T
, which is the rms delay
spread D divided by the bit period T
b
(T
b
= 1/R
b
).

b
T
T
D
D = .
(5.7)
The normalized delay spread D
T
is known to be a useful measure of multipath-
induced ISI, at least when using OOK or pulse-position modulation (PPM). In [68],
[60], it was found that over a wide ensemble of experimentally measured channels, for
a given modulation and equalization method, the ISI power penalty on a given channel
can be predicted with reasonable accuracy based solely on the normalized delay
spread D
T
, independent of the particular time dependence of the impulse response h(t).
5.3.4 Ceiling-Bounce Model
Carruthers and Kahn [60] derived a channel model based on a diffuse link
comprising a Lambertian transmitter co-located with a non-directional receiver, both
112
directed toward an infinite diffuse reflector. Considering a single bounce, the impulse
response can be obtained in closed form as

( )
) (
6
) (
7
6
t u
a t
a
t h
c
+
= ,
(5.8)
where a is the round-trip propagation time and u(t) is the unit step function. The
impulse response given by Eq. (5.8) is normalized such that 1 ) 0 ( = H . The rms delay
spread is related to a by
( )
11
13
12
) (
a
t h D
c
= . (5.9)
Carruthers and Kahn referred to the impulse response in Eq. (5.8) as the
ceiling-bounce (CB) model. They found that the simple functional form of the impulse
response given by Eq. (5.8) exhibits approximately the same relationship between ISI
penalty and normalized delay spread D
T
as a wide ensemble of experimentally
measured channels, at least for OOK and PPM [60].
5.4 Comparison of IM/DD Modulation Formats
In the following analysis, we use the system model shown in Fig. 5.2. We
employ the wireless channel model discussed in section 5.3.2. In our analysis, we
consider the IM/DD OFDM formats described in Chapter 4, such as DC-OFDM,
ACO-OFDM and PAM-DMT.
We model the dominant noise as real baseband AWGN with zero mean and
double-sided power spectral density N
0
/2. We assume a photodetector quantum
efficiency of 90%, which corresponds, for example, to a responsivity of R = 0.6 A/Hz
at 850 nm. At the receiver, the anti-aliasing filter is a fifth-order Butterworth low-pass
filter. For OFDM, we set the 3-dB cutoff frequency of the anti-aliasing filter equal to
the first null of the OFDM spectrum [25]. For M-PAM, we set the 3-dB cutoff
frequency to 0.8R
s
,

where R
s
is the symbol rate. We note that the anti-aliasing filter
cannot cause noise enhancement for any of the modulation formats, since the noise is
added before the anti-aliasing filter. In other words, the performance comparison
113
between the modulation formats is independent of the anti-aliasing filter type, given an
adequate receiver oversampling ratio and an adequate cyclic prefix length or number
of equalizer taps.
We would like to minimize the oversampling ratios for OFDM and M-PAM in
order to minimize the A/D sampling frequency while still obtaining optimal
performance. In OFDM, oversampling is performed by inserting zero subcarriers, and
hence, it is possible to employ arbitrary rational oversampling ratios. For OFDM, we
find that an oversampling ratio M
s
= 64 /52 1.23 is sufficient to obtain optimal
performance. For M-PAM, arbitrary rational oversampling ratios are possible, but
require complex equalizer structures [16]. While good performance can be achieved at
an oversampling ratio of 3/2, we choose an oversampling ratio of 2, since it achieves
slightly better performance than 3/2 while minimizing equalizer complexity. Note that
in this case, OOK requires an A/D sampling frequency 63% higher than OFDM.
Typical high-performance forward error-correction (FEC) codes for optical
systems have a bit-error ratio (BER) thresholds of the order of P
b
= 10
3
[71]. In order
for our system to be compatible with such FEC codes while providing a small margin,
we compute the minimum required SNR to achieve P
b
= 10
4
for the different
modulation formats.
On a bipolar channel with AWGN, the optimum power and bit allocations for
an OFDM system are given by the water-filling solution [4]. While the optimal value
for the number of bits for each subcarrier is an arbitrary nonnegative real number, in
practice, the constellation size and FEC code rate need to be adjusted to obtain a
rational number of bits. The optimal discrete bit allocation method on bipolar channels
is known as the Levin-Campello algorithm [54]. In this algorithm, we first set the
desired bit granularity for each subcarrier, i.e., the bit allocation on each subcarrier is
an integer multiple of (Chapter 4). In our study, we choose a granularity = 0.25,
since it is straightforward to design codes whose rates are multiples of 0.25 [72].
Furthermore, we observe that there is no significant performance difference between
OFDM with discrete loading ( = 0.25) and with continuous loading (Chapter 4).
114
5.4.1 OOK and OFDM Performance
In this section, we analyze the effect of delay spread on the required SNR for
different modulation formats at a spectral efficiency of 1 bit/symbol. We present our
results using the normalized delay spread D
T
, given by Eq. (5.7).
We let R
s
denote the symbol rate for unipolar M-PAM (for 2-PAM or OOK, R
s

= R
b
). We let
OFDM
s
R denote the equivalent symbol rate for OFDM [25]. In order to
perform a fair comparison at a given fixed bit rate R
b
between unipolar M-PAM and
OFDM, we initially set the two symbol rates to be equal:
s
OFDM
s
R R = [4].
Fig. 5.4 shows the electrical SNR required to achieve bit rates of 50, 100 and
300 Mbit/s for all channels using ACO-OFDM and OOK. In Fig. 5.4, the symbol rates
are the same for both modulation schemes, i.e.,
OFDM
s
R = R
s
, and they are chosen to be
50, 100 and 300 MHz, respectively. The receiver and transmitter electrical bandwidths
are scaled accordingly to the symbol rate in use.

b
= 10
4
T
at bit rates of 50,
100 and 300 Mbit/s (spectral efficiency of 1 bit/symbol) for ACO-OFDM and OOK. The bit allocation
granularity is = 0.25 and the symbol rates for ACO-OFDM are the same as those for OOK, as
indicated.

0.001 0.01 0.1 1 10
5
10
15
20
25
30
ACO-OFDM (R
s
)
OOK MMSE-DFE (R
s
)
Duration Bit
Spread Delay RMS
=
T
D
R
e
q
u
i
r
e
d

E
l
e
c
t
r
i
c
a
l

S
N
R

(
d
B
)
115
For ACO-OFDM with a symbol rate of 300 MHz, the FFT size is N = 1024
and the number of used subcarriers is N
u
= 208. We set the cyclic prefix equal to three
times the rms delay spread (D) of the worst impulse response, corresponding to

= 18
samples measured at the OFDM symbol rate. For OOK, we use a fractionally spaced
MMSE-DFE at an oversampling ratio of 2. We use the same number of taps for all
wireless channels, which is chosen based on the worst channel: 98 taps for the
feedforward filter and 28 taps for the feedback filter when the symbol rate is 300
MHz. The parameters for the various modulation formats and bit rates are listed
in Table 5.1.
In Fig. 5.4, we notice that as the normalized delay spread increases, the
available electrical bandwidth decreases and hence more power is required to maintain
the bit rate. We also observe that ACO-OFDM requires a higher SNR than OOK for
all channels in our set when both modulations use the same symbol rates to transmit
the same bit rates. In this case, which corresponds to a spectral efficiency of 1
bit/symbol, the performance difference is about 1.3 dB. ACO-OFDM requires a higher
SNR than OOK because it requires an average of 4 bits (16-QAM constellation) on the
used subcarriers to compensate the information rate loss of setting half of the
subcarriers to zero.
One option to improve the performance of OFDM is to use more electrical
bandwidth in order to reduce the constellation size on each subcarrier. We define
electrical bandwidth (B
e
)
2
as the span from DC to the location of the first spectral null
of the power spectral density (PSD) of the transmitted waveform x(t) [59]. At a
symbol rate R
s
, M-PAM with rectangular non-return-to-zero (NRZ) pulses requires an
electrical bandwidth of approximately
s
PAM
e
R B . On the other hand, OFDM at
symbol rate
OFDM
s
R only requires an electrical bandwidth of only 2 /
OFDM
s
OFDM
e
R B
[25], since it has a more confined spectrum. Hence, for the same symbol rate as M-
2 We note that the electrical bandwidth of a modulation technique has little influence on the optical
bandwidth B
o
occupied by an IM signal when using typical broadband light sources (e.g., a light
emitting diodes or multimode laser diodes) [59], since the optical bandwidth B
o
is dominated by the
source linewidth. For example, a 1-nm linewidth corresponds to 469 GHz, assuming a wavelength of
800 nm.
116
PAM (
OFDM
s
R = R
s
), OFDM requires about half the bandwidth of M-PAM with
rectangular NRZ pulses. If we double the symbol rate for OFDM, both modulations
schemes use approximately the same electrical bandwidth:
s
PAM
e
OFDM
e
R B B = = . The
electrical bandwidth required for all modulation formats is summarized in Table 5.1.
Bit Rate
(Mbit/s)
DC-OFDM ACO-OFDM PAM-DMT OOK
R
S
2R
S
R
S
2R
S
R
S
2R
S
R
S

50
N = 512
N
u
= 208
= 10
B
e
= 25
N = 512
N
u
= 208
= 12
B
e
= 50
N = 1024
N
u
= 208
= 10
B
o
= 25
N = 1024
N
u
= 208
= 12
B
e
= 50
N = 1024
N
u
= 416
= 10
B
e
= 25
N = 1024
N
u
= 416
= 12
B
e
= 50
N
f
= 28
N
b
= 6
B
e
= 50
100
N = 512
N
u
= 208
= 12
B
e
= 50
N = 512
N
u
= 208
= 15
B
e
= 100
N = 1024
N
u
= 208
= 12
B
o
= 50
N = 1024
N
u
= 208
= 15
B
e
= 100
N = 1024
N
u
= 416
= 12
B
e
= 50
N = 1024
N
u
= 416
= 15
B
e
= 100
N
f
= 44
N
b
= 10
B
e
= 100
300
N = 512
N
u
= 208
= 18
B
e
= 150
N = 1024
N
u
= 416
= 26
B
e
= 300
N = 1024
N
u
= 208
= 18
B
o
= 150
N = 2048
N
u
= 416
= 26
B
e
= 300
N = 1024
N
u
= 416
= 18
B
e
= 150
N = 2048
N
u
= 832
= 26
B
e
= 300
N
f
= 98
N
b
= 28
B
e
= 300
Table 5.1 System parameters for the various modulation formats for different bit rates and symbol rates.
N is the DFT size, N
u
is the number of used subcarriers, is the cyclic prefix, N
f
is the number of taps of
the feedforward filter, N
b
is the number of taps of the feedback filter and B
e
is the required electrical
bandwidth in MHz.
However, we note that when OFDM uses twice the symbol rate of M-PAM,
OFDM uses twice the electrical bandwidth it would normally require to transmit the
same bit rate as M-PAM. We also note that when the OFDM symbol rate is doubled,
OFDM requires an analog-to-digital (A/D) converter sampling frequency of
s s
OFDM
s s
R M R M 2 = , which is 23% higher than the sampling frequency required for
M-PAM using an oversampling ratio of 2. We also want to point out that the usable
electrical bandwidth becomes eventually limited by the channel multipath distortion.
Hence, for channels with high delay spreads, there is very little benefit in increasing
the transmitter and receiver electrical bandwidths.
Having the previous considerations in mind, we study the trade-off between
electrical bandwidth and SNR performance for OFDM in Fig. 5.5 and Fig. 5.6. Fig.
5.5 shows the receiver electrical SNR required to achieve the bit rates of 50, 100 and
300 Mbit/s when ACO-OFDM uses twice the symbol rate of OOK, i.e.,
OFDM
s
R

= 2R
s
.
117
The symbol rates for OOK are 50, 100 and 300 MHz, respectively and the symbol
rates for ACO-OFDM are 100, 200 and 600 MHz, respectively. The receiver and
transmitter electrical bandwidths are scaled according to the symbol rate in use.

b
= 10
4
T
at bit rates of 50,
100 and 300 Mbit/s (spectral efficiency of 1 bit/symbol) for different modulations formats. The dashed
lines correspond to the SNR requirement predicted using the ceiling-bounce (CB) model. The bit
allocation granularity is = 0.25 and the symbol rates for ACO-OFDM are twice those for OOK, as
indicated.
In Fig. 5.5, we observe that ACO-OFDM using twice the symbol rate of OOK
requires the lowest SNR to achieve a P
b
= 10
4
for all channels at the bit rates of 50,
100 and 300 Mbit/s. For low delay spreads (D
T
< 0.1), corresponding to LOS
channels, ACO-OFDM significantly outperforms OOK. For high delay spreads,
however, OOK with MMSE-DFE performs very close to ACO-OFDM. We again note
that when OFDM uses twice the symbol rate of OOK, OFDM uses twice the
bandwidth it would normally require to transmit the same bit rate as OOK. Even with
this advantage, for a given bit rate, ACO-OFDM achieves approximately the same
performance as OOK for high delay spreads, as shown in Fig. 5.5. ACO-OFDM
performs better for low delay spreads because it can use more bandwidth and transmit
lower constellation sizes on each subcarrier. For high delay spreads, the wireless
channel becomes more bandwidth-constrained, and there is very little gain by using
more bandwidth for ACO-OFDM.
ACO-OFDM (2R
s
)
OOK MMSE-DFE (R
s
)
CB for ACO-OFDM (2R
s
)
CB for
OOK (R
s
)
0.001 0.01 0.1 1 10
5
10
15
20
25
30
Duration Bit
Spread Delay RMS
=
T
D
R
e
q
u
i
r
e
d

E
l
e
c
t
r
i
c
a
l

S
N
R

(
d
B
)
118
In Fig. 5.5, we also show the SNR requirements for OOK and ACO-OFDM
estimated using the ceiling-bounce (CB) model [60]. Carruthers and Kahn [60] found a
simple functional approximation of the impulse response that exhibits approximately
the same relationship between ISI penalty and normalized delay spread D
T
as a wide
ensemble of experimentally measured channels, at least for OOK and PPM [60]. We
observe that the CB model provides a reasonable is estimate of the SNR requirement.
We note that in computing Fig. 5.5 for OOK with MMSE-DFE, we have assumed
correct decisions at the input of the feedback filter.
In order to make a fair comparison with ACO-OFDM, the number of used
subcarriers in DC-OFDM should be the same as for ACO-OFDM. For PAM-DMT,
since only one dimension is used to transmit data, the number of used subcarriers
should be twice that for ACO-OFDM. In computing the optimized subcarrier power
allocation for DC-OFDM, we use an iteratively water-filling solution with the bias
ratio B
R
, as described in Chapter 3.5, performing an exhaustive search over the value
of B
R
as in [73]. We have plotted the SNR requirements for DC-OFDM and PAM-
DMT separately in Fig. 5.6, in order to make Fig. 5.5 more legible.
Fig. 5.6 shows the cumulative distribution function (CDF) of the required
electrical SNR, which is defined as:
( ) SNR x SNR = Prob ) ( CDF , (5.10)
where SNR takes all possible values for the required SNR for a given modulation
format. We assume that all 170 channels realizations occur with equal probability. The
CDF corresponds to the fraction of channels on which the target P
b
is reached at a
given SNR. For example, Fig. 5.6 shows that if the transmitter has CSI and the
electrical SNR is at least 30 dB, the target P
b
10
4
is met for all the channels when
using ACO-OFDM (or PAM-DMT) at any of the three bit rates, 50, 100 or 300
Mbit/s.
119

b
= 10
4
at
bit rates of 50, 100 and 300 Mbit/s (spectral efficiency of 1 bit/symbol) for different modulation
formats. The symbol rates for OFDM are the same as or twice those for OOK, as indicated.
We observe in Fig. 5.6, as in Fig. 5.5, that ACO-OFDM using twice the
symbol rate of OOK requires the lowest SNR to achieve a P
b
= 10
4
. On the other
hand, if ACO-OFDM uses the same symbol rate as OOK, OOK is more power
efficient. Furthermore, we verify that there is no significant performance difference
between ACO-OFDM and PAM-DMT. This is expected, since it was proven in [74]
that ACO-OFDM and PAM-DMT achieve very similar performance on low-pass
channels. Finally, we observe that DC-OFDM requires the highest SNR because of the
optical power used in the DC bias to reduce clipping noise.
5.4.2 Unipolar 4-PAM and OFDM Performance
In this section, we compare the performance of the different modulation
formats at higher spectral efficiencies to check if there are any significant differences
from the previous section. Fig. 5.7 and Fig. 5.8 show the electrical SNR required to
achieve bit rates of 100, 200 and 600 Mbit/s for different modulation formats at a
spectral efficiency of 2 bit/symbol. The symbol rates for unipolar 4-PAM are 50, 100
and 300 MHz, respectively. The receiver and transmitter electrical bandwidths are
scaled according to the symbol rate in use.
5 10 15 20 25 30 35
0
0.2
0.4
0.6
0.8
1
1.2
Required Electrical SNR (dB)
C
D
F
(
S
N
R
)DC-OFDM (2R
s
)
ACO-OFDM and
PAM-DMT (R
s
)
OOK MMSE-DFE (R
s
)
ACO-OFDM and
PAM-DMT (2R
s
)
120

b
= 10
4
T
at bit rates of
100, 200 and 600 Mbit/s (spectral efficiency of 2 bit/symbol) for different modulations formats. The
bit allocation granularity is = 0.25 and the symbol rates for ACO-OFDM are the same as or twice
those for 4-PAM, as indicated.
In Fig. 5.7 and Fig. 5.8, we observe that ACO-OFDM and PAM-DMT using
twice the symbol rate of 4-PAM again requires the lowest SNR for all channels.

b
= 10
4
at
bit rates of 100, 200 and 600 Mbit/s (spectral efficiency of 2 bit/symbol) for different modulation
formats. The symbol rates for OFDM are the same as or twice those for 4-PAM, as indicated.
ACO-OFDM (2R
s
)
ACO-OFDM (R
s
)
4-PAM MMSE-DFE (R
s
)
0.01 0.1 1 10
10
15
20
25
30
35
40
45
R
e
q
u
i
r
e
d

E
l
e
c
t
r
i
c
a
l

S
N
R

(
d
B
)
Duration Bit
Spread Delay RMS
=
T
D
DC-OFDM (R
s
)
10 15 20 25 30 35 40 45
0
0.2
0.4
0.6
0.8
1
1.2
C
D
F
(
S
N
R
)
DC-OFDM (2R
s
)
ACO-OFDM and
PAM-DMT (R
s
)
4-PAM MMSE-DFE (R
s
)
ACO-OFDM and
PAM-DMT (2R
s
)
121
In Fig. 5.8, we observe again that PAM-DMT has the same performance as
ACO-OFDM for any symbol rate. Since both modulations have the same performance,
we choose ACO-OFDM as our basis of comparison for the remainder of the chapter.
The performance improvement for ACO-OFDM (and PAM-DMT) in Fig. 5.7 and Fig.
5.8 requires doubling the bandwidth and increasing the sampling frequency, which
becomes 23% higher than the sampling frequency required for 4-PAM with an
oversampling ratio of 2. Comparing Fig. 5.6 and Fig. 5.8, we see that the performance
difference between ACO-OFDM with twice the symbol rate and 4-PAM increases
with the spectral efficiency. We also observe in Fig. 5.8 that DC-OFDM with twice
the symbol rate performs close to 4-PAM.
When all the modulations use equal symbol rates, 4-PAM requires the lowest
SNR of all the modulation formats, and the difference in SNR between ACO-OFDM
(or PAM-DMT) and unipolar 4-PAM is about 4 dB. We conclude that increasing the
spectral efficiency from 1 bit/symbol (Fig. 5.6) to 2 bit/symbol (Fig. 5.8) increases
differences in SNR requirements between ACO-OFDM and M-PAM from 1 dB to 4
dB. This is to be expected, since increasing the spectral efficiency from 1 bit/symbol
to 2 bit/symbol for ACO-OFDM requires doubling the average transmitted number of
bits on each subcarrier from 4 bits (16-QAM) to 8 bits (256-QAM). For higher
spectral efficiencies, ACO-OFDM require constellations that are increasingly large
and therefore M-PAM has increasingly better performance than ACO-OFDM with
equal symbol rates.
Finally, we also note in Fig. 5.8 that DC-OFDM performs very close to ACO-
OFDM (or PAM-DMT) when the OFDM symbol rate is the same as that for 4-PAM
and the spectral efficiency is 2 bit/symbol. This is because the clipping noise penalty
in DC-OFDM becomes less significant when compared to the additional SNR required
to transmit higher constellation sizes in each ACO-OFDM subcarrier.
Although DC-OFDM uses twice as many subcarriers as ACO-OFDM (or
PAM-DMT) to transmit data, the clipping noise and DC bias limit the power
efficiency of DC-OFDM for a spectral efficiency of 1 bit/symbol, as shown in Fig.
5.6. However, as spectral efficiency increases, the clipping noise penalty in DC-
122
OFDM becomes less significant when compared to the additional SNR required to
transmit higher constellation sizes in each ACO-OFDM subcarrier, and eventually
DC-OFDM performs better than ACO-OFDM.
5.4.3 Outage Probability
It is also useful to analyze the receiver performance when the transmitter does
not have CSI. This happens, for example, in a broadcast configuration where the
transmitter sends the same information to several receivers or when there is no return
path from the receiver (i.e., uplink). When CSI is unavailable at the transmitter, some
data-bearing subcarriers might be lost due to multipath distortion. Hence, we use
coding across the subcarriers such that the information can be recovered from the good
subcarriers. We choose a shortened Reed-Solomon (RS) code with (n, k) = (127, 107)
over GF(8) for simplicity
3
. In order for the comparison between OFDM and OOK to
be fair, OOK also employs this code.
We consider a channel realization to be in outage if P
b
> 10
4
for a desired bit
rate. We define outage probability P
outage
as the fraction of channels with P
b
> 10
4
over the entire ensemble of channels realizations. In practice, a system might not be
expected to work on the most severe channel realizations (e.g., shadowed channels),
but in our study we include all channels as a worst-case scenario.
We assume a desired bit rate of 300 Mbit/s. For OOK, we increase the symbol
rate from 300 MHz to 356 MHz to compensate for coding overhead. For OFDM, we
use ACO-OFDM with twice the data symbol rate of OOK, i.e., 600 MHz. We
compensate for the information loss due to coding by increasing the constellation size
on each subcarrier. The same performance could be achieved with PAM-DMT.

3 In an optimized implementation, this code (used to recover subcarriers lost due to multipath
distortion) might be integrated with the outer FEC code (used to combat additive noise) into a single
concatenated coding scheme. Likewise, more powerful coding schemes (such as turbo or low-density
parity-check codes) could be used to recover the lost subcarriers. Given the simple nature of indoor
optical wireless channels, we would expect our conclusions to remain valid even if more optimized
coding schemes were employed.
123
We assume that the channel changes slowly enough such that many OFDM
blocks are affected by a channel realization. In this scenario, there is no benefit to
using time diversity (e.g., interleaving) in the bit allocation on the subcarriers. We also
assume that the transmitter has knowledge of the mean channel, which has a
frequency response defined by

( ) ( )
=
=
realiz
N
k
k
realiz
mean
f H
N
f H
1
1
,
(5.11)
which is an average over all 170 channel realizations. For ACO-OFDM, we
consider three different bit allocation schemes: mean channel loading, where the
transmitter does integer bit loading using the frequency response of the mean channel;
ceiling-bounce loading, where the transmitter performs integer bit allocation on a
ceiling-bounce channel having the same rms delay spread D as the mean channel,
4
and
equal loading, where all the subcarriers carry the same number of bits. After having
performed the bit allocation, we sweep the transmitter power and observe which
channel realizations are in outage, i.e., have P
b
> 10
4
.
Since each RS block in OFDM has a different uncoded P
b,
we simulate RS
decoding with error counting in order to obtain the correct P
b
after RS decoding. For
OOK, we use a binomial expansion to obtain P
b
after RS decoding, since each bit has
the same uncoded P
b
.
Fig. 5.9 shows the outage probability for coded ACO-OFDM and coded OOK
with MMSE-DFE for a desired bit rate of 300 Mbit/s.
4 For the purpose of computing the rms delay spread D of the mean channel, the impulse response of
the mean channel is computed as the inverse Fourier transform of H
mean
(f).
124

Fig. 5.9 Outage probability for OOK and ACO-OFDM with coding and various bit allocations
averaged over all channels. All modulation formats use the code RS(127,107) over GF(8). The
information bit rate is 300 Mbit/s (spectral efficiency of 1 bit/symbol) and the symbol rates for OOK
and OFDM are 356 MHz and 600 MHz, respectively. The bit allocation granularity is = 1, i.e.,
integer bit allocation.
We observe that OOK has a significantly lower outage probability than OFDM
for the same SNR. In other words, when the transmitter does not have CSI, OOK with
MMSE-DFE performs better than ACO-OFDM with twice the symbol rate, in contrast
to Fig. 5.6 and Fig. 5.8. Results for 50 and 100 Mbit/s are qualitatively similar to those
for 300 Mbit/s, particularly at low outage probabilities. These results are not shown
in Fig. 5.9 to maximize legibility. At low outage probabilities (P
b
< 10%), CB loading
performs slightly better than the other loading schemes. For outage probabilities
between 25% and 70%, CB loading is significantly better than the other schemes. For
high outage probabilities, equal loading performs the best. Finally, we observe that
over a wide range of outage probabilities, mean channel loading generally outperforms
equal loading. We note that when the outage probability is high, the system can only
operate on channels having low delay spreads, which are generally those without
shadowing. In practice, a user could improve performance by moving his receiver to
avoid shadowing.
5 10 15 20 25 30 35 40
0
20
40
60
80
100
120
P
o
u
t
a
g
e
(
%
)
ACO-OFDM
(Mean Channel Loading)
ACO-OFDM
(Equal Loading)
OOK
MMSE-DFE
ACO-OFDM (CB Loading)
125
5.5 Computational Complexity
fast Fourier transform (FFT) algorithm. For M-PAM, the DFE equalization can be
done by direct implementation or by FFT-based block processing, as discussed in
Chapter 2. In our comparison, we use the same analysis and equations from Chapter
2.6.
Using the parameters in Table 5.1, we compute the number of operations
required per bit for OOK and OFDM in Table 5.2.
Bit Rate
(Mbit/s)
50 100 300
ACO-OFDM
(R
S
)
Tx 48.5 48.5 48.1
Rx 50.0 50.0 49.6
ACO-OFDM
(2R
S
)
Tx 97.0 96.7 106.5
Rx 100.0 99.6 109.4
PAM-DMT
(R
S
)
Tx 48.5 48.5 48.1
Rx 51.5 51.5 51.1
PAM-DMT
(2R
S
)
Tx 97.0 96.7 106.5
Rx 102.9 102.6 112.4
DC-OFDM
(R
S
)
Tx 43.0 43.0 42.4
Rx 45.9 45.9 45.3
DC-OFDM
(2R
S
)
Tx 86.1 85.4 95.3
Rx 91.9 91.2 101.1
OOK
(R
S
)
L
f
101 213 415
L
b
59 119 229
FFT 137.2 150.8 174.5
Direct 272 432 1008
Table 5.2 Computational complexity in real operations per bit for the various modulation formats.
For OOK, we calculate the complexity of direct implementation and FFT
based implementation using optimized block sizes L
f
and L
b
subject to having a FFT
size that is an integer power of two. In Table 5.2, we observe that an OFDM system
with twice the symbol rate of OOK requires two digital signal processors (DSPs) with
approximately 40% less computational complexity than the single DSP required for
OOK. On the other hand, an OFDM system requires an A/D and D/A while OOK
requires only one A/D. We also note that DC-OFDM has the lowest computational
126
complexity among OFDM formats, since it uses smaller FFT sizes than ACO-OFDM
or PAM-DMT.

Conclusions
127
6 CONCLUSIONS
6.1 Conclusions
We have shown that in long-haul systems using coherent detection, OFDM
with a one-tap equalizer is able to compensate for linear distortions such as GVD and
PMD, provided that an adequate cyclic prefix length is chosen. Compensation of PMD
requires a dual-polarization receiver. We have derived analytical penalties for the ISI
and ICI occurring when an insufficient cyclic prefix is used. Using these penalties, we
have computed the minimum number of subcarriers and cyclic prefix length required
to achieve a specified power penalty for GVD and first-order PMD. We observed that
GVD is the dominant impairment, since it requires several cyclic prefix samples,
whereas first-order PMD typically requires only one cyclic prefix sample. We verified
that an oversampling ratio of 1.2 is sufficient to minimize noise aliasing. By contrast,
single-carrier systems typically require an oversampling ratio of 1.5 or 2 to avoid
noise aliasing. We have also shown that when nonlinear effects are present in the
fiber, single-carrier with digital equalization outperforms OFDM for various
dispersion maps. However, when no inline dispersion compensation is used, OFDM
has a similar performance as single-carrier because of incoherent additions and partial
cancelations of the FWM products. We have also shown that OFDM requires less
computational complexity per DSP than single-carrier but requires two DSPs. In
terms of the optical modulator, we have shown that the quadrature MZ with pre-
distortion and hard clipping is able to achieve good performance without additional
oversampling, while achieving high optical power efficiency. We observed that the
MZ electrode frequency dependent losses can be neglected, since they have very little
impact on system performance. We have also shown that the optimum clipping level is
approximately CR = 2.5, since this value yields a 12-dB gain in optical power
efficiency with minimal receiver sensitivity degradation. Furthermore, we verified that
the D/A requires at least 6 bits, and that the optimum clipping level is also
approximately CR = 2.5 when quantization noise is present. We have also shown that
Conclusions
128
the dual-drive MZ can also be used as an efficient OFDM modulator. It requires less
hardware than the quadrature MZ and minimizes the clipping probability. On the other
hand, the dual-drive MZ requires a higher oversampling ratio which, in turn, means
that faster D/As and wider electrode frequency responses are necessary.
Considering metropolitan area networks, we have evaluated the performance
of three different OFDM formats in amplified DD optical systems: SSB-OFDM, DC-
OFDM and ACO-OFDM. We have derived equivalent linear channel models for each
format in the presence of GVD. We showed how to minimize the average optical
power required to achieve a specified error probability by iteratively adjusting the
subcarrier power allocation and the bias ratio (B
R
). We found that for a given
dispersion, SSB-OFDM requires the smallest optical power. We presented an
analytical lower bound for the minimum required optical SNR for DC-OFDM and
concluded that DC-OFDM can never achieve the same optical power efficiency as
SSB-OFDM. We showed that at low dispersion, ACO-OFDM performs close to SSB-
OFDM but at high dispersion ACO-OFDM performs worse than DC-OFDM, because
of nonlinear ICI. Using published results on OOK with MLSD, we showed that SSB-
OFDM can achieve the same optical power efficiency as OOK with MLSD, at the
expense of requiring twice the electrical bandwidth and therefore requiring also a
higher A/D sampling rate. Furthermore, SSB-OFDM requires a quadrature modulator,
which also increases the hardware complexity. On the other hand, we showed that
SSB-OFDM requires significantly lower computational complexity than OOK with
MLSD.
For application to local area networks, we have evaluated the performance of
three different IM/DD OFDM formats in multimode fibers: DC-OFDM, ACO-OFDM
and PAM-DMT. We have derived the optimal power allocation for PAM-DMT and
have shown that the performance of PAM-DMT converges asymptotically to that of
ACO-OFDM as the number of subcarriers increases. We have also shown how to
minimize the average optical power required for DC-OFDM to achieve a specified
error probability by iteratively adjusting the subcarrier power allocation and the bias
ratio (B
R
). For a given symbol rate, we have found that unipolar M-PAM with MMSE-
Conclusions
129
DFE has a better optical power performance than all OFDM formats. Furthermore, we
have found that the performance difference between M-PAM and OFDM increases as
the spectral efficiency increases. We have shown that at a spectral efficiency of 1
bit/symbol, OOK performs better than ACO-OFDM with twice the symbol rate of
OOK. For higher spectral efficiencies, ACO-OFDM at twice the symbol rate of M-
PAM performs close to M-PAM, but requires more electrical bandwidth and 23%
faster A/D converters than those required for M-PAM at 2 samples per symbol.
Considering indoor optical wireless systems, we have evaluated the
performance of the three IM/DD OFDM formats, DC-OFDM, ACO-OFDM and
PAM-DMT, and have compared them to unipolar M-PAM with MMSE-DFE. When
using the same symbol rate for all modulation methods, we have found that unipolar
M-PAM with MMSE-DFE has better optical power efficiency than all OFDM formats
over a range of spectral efficiencies. Furthermore, we have found that as spectral
efficiency increases, the performance advantage of M-PAM increases, since the
OFDM formats require increasingly large signal constellations. We have also found
that ACO-OFDM and PAM-DMT have virtually identical performance at any spectral
efficiency. They are the best OFDM formats at low spectral efficiency, but as spectral
efficiency increases, DC-OFDM performs closer to ACO-OFDM, since the clipping
noise penalty for DC-OFDM becomes less significant than the penalty for the larger
constellations required for ACO-OFDM. When ACO-OFDM or PAM-DMT are
allowed to use twice the symbol rate of M-PAM, these OFDM formats have better
performance than M-PAM. However, at a spectral efficiency of 1 bit/symbol, OOK
with MMSE-DFE has performance similar to ACO-OFDM or PAM-DMT for high
delay spreads. When CSI is unavailable at the transmitter, M-PAM significantly
outperforms all OFDM formats even when they use twice the symbol rate of M-PAM.
When using the same symbol rate for all modulation methods, M-PAM requires
approximately three times more computational complexity per DSP than all OFDM
formats and 63% faster A/D converters. When OFDM uses twice the symbol rate of
M-PAM, OFDM requires 23% faster A/D converters than M-PAM but OFDM
requires 40% less computational complexity than M-PAM per DSP.
Conclusions
130
6.2 Future Work
In long-haul systems, future work could focus on nonlinearity compensation
algorithms in order to improve the OFDM performance, and possibility, determine the
maximum performance achievable.
Future work in metro systems is needed to determine new techniques or
algorithms that can achieve good performance while requiring much less
computational complexity.
Finally, future work in IM/DD systems, such as LANs and indoor optical
wireless, is still needed to compute the information theoretic channel capacity of non-
negative channels with bipolar Gaussian noise. The main difficulties are that the
transmitted waveform x(t) is constrained to be non-negative and the parameter to be
minimized is the optical power, which corresponds to the first moment of x(t).

Appendix A
131
A APPENDIX
Derivation of the ISI and ICI variance on the different subcarriers is crucial for
determining the probability of symbol error. A similar derivation for unilateral
channels can be found in [75]. Here, we generalize to bilateral channels and include
the effect of correlation when oversampling is used.
If N
pre
= N
pos
+ N
neg
samples are used for the cyclic prefix, and if the channel
impulse response has positive and negative lengths L
p
1 and L
n
1,

respectively
,
the
residual ISI on the nth time-domain sample in the kth OFDM symbol can be written as

) ( ) (
) ( ) ( ) (
1
1
1
1
1
1
u h u N N n x
r h r N N N n x n ISI
channel
N
L u
neg c k
channel
L
N r
pos pre c k k
neg
n
p
pos

+ =
+
+ =

+
+ + + =

A.1
where n varies from 0 to N
c
1. Eq. A.1 represents the ISI as a linear function of the
OFDM samples. As the number of subcarriers increases, by the Central Limit
Theorem [5], the pdf of the samples approaches a Gaussian. Since the ISI is a linear
function of the samples, its pdf is also Gaussian. The signal after the DFT becomes

c
neg
n
p
pos
c
N
nq
j
channel
N
L u
neg c k
channel
L
N r
pos pre c k
N
n
k
e u h u N N n x
r h r N N N n x q ISI
2
1
1
1
1
1
1
1
0
) ( ) (
) ( ) ( ) (

+ =
+
+ =

+ + + =

A.2
Performing a variable change on the different sums, we get

c
neg c
n c
c
c
p
pos
c
N
nq
j
c channel
n N N
n L N z
neg k
N
n
N
nq
j
channel
n L
n N v
pos pre c k
N
n
k
e N z n h z N x
e v n h v N N N x q ISI
2
2
1
1
0
2
1
1
1
1
0
) 1 ( ) 1 (
) ( ) ( ) (

=
+

+ =

=
+ + +
+ + + =

A.3
Next we interchange the inner and outer sums. In both of the double
summations, we require that N
c
> max(L
p
N
pos
1, L
n
N
neg
1). This means that the
OFDM block must be longer than the impulse response duration or interference from
symbols k+2 or k2 will take place. In the first double sum, as n increases, the upper
Appendix A
132
limit of the inner sum v decreases accordingly. The lower limit N
pos
+1 is reached
when n = L
p
N
pos
1. If we now take the sum over v for the outer sum of the first
double sum, we obtain the upper limit for the inner sum n as L
p
1v. The same idea
can be applied to the second double sum, and the lower limit of the inner sum z is
reached when n = L
n
N
neg
1. If we now take the sum over z for the outer sum of the
second double sum, we obtain the lower limit for the inner sum n as N
c
L
n
z. Thus,
Eq. A.3 becomes

+

+ =

=
+ =
+ + +
+ + + =
1
2
1
1
1
1
0
2
1
1
1
) 1 ( ) 1 (
) ( ) ( ) (
c
n c
c
neg
n
p
c
p
pos
N
z L N n
N
nq
j
neg k c channel
N
L z
v L
n
N
nq
j
pos pre c k channel
L
N v
k
e z N x N z n h
e v N N N x v n h q ISI

A.4
Substituting n + v= p and n+ z = m in the first and second double sums,
respectively, we get

) ( ) 1 (
) ( ) (
) 1 ( ) 1 (
) ( ) (
) 1 ( ) 1 (
) ( ) ( ) (
1
1
2
1
2
1
1
1
1
2
1
1
2
1
1
2 2
1
1
1
1
) ( 2
1
1
1
1
) ( 2
1
1
1
q H e z N x
q H e v N N N x
e N m h e z N x
e p h e v N N N x
e N m h z N x
e p h v N N N x q ISI
z
N
L z
N
zq
j
neg k
v
N
vq
j
L
N v
pos pre c k
z N
L N m
N
mq
j
c channel
N
L z
N
zq
j
neg k
L
v p
N
pq
j
channel
N
vq
j
L
N v
pos pre c k
z N
L N m
N
q z m
j
c channel
N
L z
neg k
L
v p
N
q v p
j
channel
L
N v
pos pre c k k
neg
n
c
c
p
pos
c
n c
c
neg
n
c
p
c c
p
pos
c
n c
c
neg
n
p
c
p
pos

+ =
+
+ =

=

+ =
+
+ =

=

+ =
+
+ =

+
+ + =
+ +
+ + =
+ +
+ + =

A.5
Note that the expressions for H
v
(q) and H
z
(q) correspond to the DFTs of the
positive and negative tails of the channel impulse response h
channel
(n). After the DFT,
the variance
2
ISI
(q) on each subcarrier q is
Appendix A
133

[ ]
[ ]
[ ]
[ ]
[ ]
c
neg
n
p
pos
c
p
pos
neg
n
c
neg
n
neg
n
c
p
pos
p
pos
N
q v z
j
pos pre c k neg k
N
L z
L
N v
v z
N
q z v
j
neg k pos pre c k
L
N v
N
L z
z v
N
q z z
j
neg k neg k
N
L z
N
L z
z z
N
q v v
j
pos pre c k pos pre c k
L
N v
L
N v
v v
ISI
e v N N N x z N x E
q H q H
e z N x v N N N x E
q H q H
e z N x z N x E
q H q H
e v N N N x v N N N x E
q H q H
q ISI q ISI E q
) ( 2
*
2 1 2 1
1
1 2
1
1 2
*
2 2
) ( 2
*
1 1 1 1
1
1 1
1
1 1
*
1 1
) ( 2
*
2 1 1 1
1
1 1
1
1 2
*
2 1
) ( 2
*
2 1 1 1
1
1 1
1
1 2
*
2 1
* 2
2 2
1 1
2 1
2 1
) ( ) 1 (
) ( ) (
) 1 ( ) (
) ( ) (
) 1 ( ) 1 (
) ( ) (
) ( ) (
) ( ) (
) ( ) ( ) (
+

+ =
+ =
+ =

+ =
+ +

+ =

+ =
+ =
+ =
+ +
+
+ +
+

+
+ + + +
=
=

A.6
If we assume that the symbols are uncorrelated, i.e., E[x(n)
k
x(l)
m
*
] = 0 for n
l, Eq. A.6 simplifies to

+ =

+ =
+ =
1
1
2
1
1
2
2 2
) ( ) ( ) (
neg
n
p
pos
N
L z
z
L
N v
v s ISI
q H q H q

A.7
where
s
2
= E[|x(n)|
2
] is the mean power per sample of the time-domain waveform.
Eq. A.7 gives the variance of the interference induced on subcarrier q in the OFDM
symbol k by the previous and subsequent symbols, k1 and k+1. It remains now to
calculate the interference caused by the current symbol k on itself, i.e., the ICI. ICI
occurs because when the cyclic prefix is not sufficiently long, linear convolution
between the channel and symbol k does not correspond to one period of a circular
convolution. In order to compute the ICI, we assume an extension within the received
symbol k so that it is equivalent to one period of a circular convolution between the
channel and the transmitted symbol k. The removal of this extension would then be the
ICI. The extension concept is illustrated in Fig. A-1.
Appendix A
134

Fig. A-1 DWDM system with 50 GHz channel spacing.
The ICI is then similar to Eq. A.1 and can be written as

) ( ) (
) ( ) ( ) (
1
1
1
1
u h u N N n x
r h r N N N n x n ICI
channel
N
L u
neg c k
channel
L
N r
pos pre c k k
neg
n
p
pos

+ =
+ =

+ + + =

A.8
If we follow the same steps as in derivation of the ISI, we see that the ICI and
ISI differ only by a sign, which disappears when squaring to compute the variance.
Hence, the ISI and ICI have the same variance,
2
ISI
(q) =
2
ICI
(q). Thus, the variance
of the total interference on subcarrier q can be written as

+ =

+ =
+ =
+
1
1
2
1
1
2
2 2
) ( ) ( 2 ) (
neg
n
p
pos
N
L z
z
L
N v
v s ICI ISI
q H q H q

A.9
Eqs. A.7 and A.9 are exact when the OFDM samples are uncorrelated.
However, if oversampling is used, the samples are no longer uncorrelated and
therefore Eqs. A.7 and A.9 represent an approximation.

N
pos
N
neg
N
pos
N
neg
Ext
p
Ext
n
Ext
p
Ext
n
-
Symbol k with ICI
Symbol k without ICI
ICI
0 0 0
References
135
REFERENCES
[1] S. J. Savory, G. Gavioli, R. I. Killey, and P. Bayvel, "Transmission of 42.8 Gbit/s
polarization multiplexed NRZ-QPSK over 6400 km of standard fiber with no
optical dispersion compensation," in Proc. Optical Fiber Commun. Conf. (OFC
2007), OTuA1, 2007.
[2] J. B. Carruthers and J. M. Kahn, "Modeling of Nondirected Wireless Infrared
Channels," IEEE Trans. on Commun., vol. 45, no. 10, pp. 12601268, 1997.
[3] G. Bosco, P. Poggiolini, and M. Visintin, "Performance Analysis of MLSE
Receivers Based on the Square-Root Metric," J. Lightw. Technol., vol. 26, no. 14,
pp. 20982109, 2008.
[4] J. G. Proakis, Digital Communications, 4th ed.: McGraw-Hill, 2002.
[5] G. L. Stuber, Principles of Mobile Communication, 2nd ed.: Kluwer Academic
Publishers, 2001.
[6] K. Sistanizadeh, P. Chow, and J. M. Cioffi, "MultiTone Transmission for
Asymmetric Digital Subscriber Lines (ADSL)," in ICC'91, Switzerland, 1993, pp.
756760.
[7] W. Shieh, "PMD-Supported Coherent Optical OFDM Systems," IEEE Photon.
Technol. Lett., vol. 19, no. 3, pp. 134136, 2007.
[8] S.L. Jansen, I. Morita, N. Takeda, and H. Tanaka, "20-Gb/sOFDM Transmission
over 4160 km SSMF Enabled by RF-Pilot Tone Phase Noise Compensation," in
Proc. Optical Fiber Commun. Conf. (OFC 2007), PDP15, 2007.
[9] B. Schmidt, A. Lowery, and J. Armstrong, "Experimental Demonstration of 20
Gbit/s Direct-Detection Optical OFDM and 12 Gbit/s with a colorless
transmitter," in Proc. Optical Fiber Commun. Conf. (OFC 2007), PDP18, 2007.
References
136
[10] Y. Tang, W. Shieh, X. Yi, and R. Evans, "Optimum Design for RF-to-Optical
Up-Converter in Coherent Optical OFDM Systems," IEEE Photon. Technol.
Lett., vol. 19, no. 7, pp. 483485, 2007.
[11] A. Lowery, S. Wang, and M. Premaratne, "Calculation of power limit due to fiber
nonlinearity in optical OFDM systems," Optics Express, vol. 15, no. 20, pp.
1328213287, 2007.
[12] S. Wu and Y. Bar-Ness, "OFDM Systems in the Presence of Phase Noise:
Consequences and Solutions," IEEE Trans. on Commun., vol. 52, no. 11, pp.
19881996, 2004.
[13] J. Tellado, Multicarrier Modulation with Low PAR: Applications to DSL and
Wireless.: Kluwer Academic Publishers, 2000.
[14] K.-P. Ho, Phase-modulated optical communication systems, Springer, Ed., 2005.
[15] K. Ho and J. M. Kahn, "Spectrum of Externally Modulated Optical Signals," J.
Lightw. Technol., vol. 22, no. 2, pp. 658663, 2004.
[16] E. Ip and J. M. Kahn, "Digital Equalization of Chromatic Dispersion and
Polarization Mode Dispersion," J. Lightw. Technol., vol. 25, no. 8, pp. 2033
2043, 2007.
[17] T. M. Schmidl and D. C. Cox, "Robust frequency and timing synchronization for
OFDM," IEEE Trans. on Commun., vol. 45, no. 12, pp. 16131621, 1997.
[18] J. J. van de Beek, M. Sandell, and P. O. Borjesson, "ML estimation of time and
frequency offset in OFDM systems," IEEE Trans. Signal Process., vol. 45, no.
12, pp. 18001805, 1997.
[19] G. J. Foschini and C. D. Poole, "Statistical theory of polarization dispersion in
single mode fibers," J. Lightw. Technol., vol. 9, no. 11, pp. 14391456, 1991.
[20] E. Forestieri and G. Prati, "Exact analytical Evaluation of Second-order PMD
Impact on the Outage Probability for a Compensated System," J. Lightw.
References
137
Technol., vol. 22, no. 4, pp. 988996, 2004.
[21] E. Forestieri and L. Vincetti, "Exact evaluation of the Jones matrix of a fiber in
the presence of polarization mode dispersion of any order," J. Lightw. Technol.,
vol. 19, no. 12, pp. 18981909, 2001.
[22] H. Kogelnik, L. E. Nelson, J. P. Gordon, and R. M. Jopson, "Jones Matrix for
second-order polarization mode dispersion," Optics Letters, vol. 25, no. 1, pp.
1921, 2000.
[23] A. P. T. Lau and J. M. Kahn, "Signal design and detection in presence of
nonlinear phase noise," J. Lightw. Technol., vol. 25, no. 10, p. 30083016, 2007.
[24] J. P. Gordon and L. F. Mollenauer, "Phase noise in photonic communications
systems using linear amplifiers," Optics Letters, vol. 15, no. 23, p. 13511353,
1990.
[25] D. J. F. Barros and J. M. Kahn, "Optimized Dispersion Compensation Using
Orthogonal Frequency-Division Multiplexing," J. Lightw. Technol., vol. 26, no.
16, pp. 28892898, 2008.
[26] S.G. Johnson and M. Frigo, "A modified split-radix FFT with fewer arithmetic
operations," IEEE Trans. Signal Process., vol. 55, no. 1, pp. 111119, 2007.
[27] S. Jansen, I. Morita, T. Schenk, N. Takeda, and H. Tanaka, "Coherent optical
25.8-Gb/s OFDM transmission over 4160-km SSMF," J. Lightw. Technol., vol.
26, no. 1, pp. 615, 2008.
[28] Y. Tang, K. Ho, and W. Shieh, "Coherent Optical OFDM Transmitter Design
Employing Predistortion," IEEE Photon. Technol. Lett., vol. 20, no. 11, pp. 954
956, 2008.
[29] B. Krongold and D. Jones, "PAR reduction in OFDM via active constellation
extension," IEEE Trans. Broadcast., vol. 49, no. 3, pp. 258268, 2003.
References
138
[30] S. Sezginer and H. Sari, "OFDM peak power reduction with simple amplitude
predistortion," IEEE Commun. Lett., vol. 10, no. 2, pp. 6567, 2006.
[31] C. Ciochina, F. Buda, and H. Sari, "An analysis of OFDM peak power reduction
techniques for WiMAX systems," in Proc. IEEE ICC06, 2006, pp. 46764681.
[32] J. Armstrong, "Peak-to-average power reduction for OFDM by repeated clipping
and frequency domain filtering," Electronics Letters, vol. 38, no. 5, pp. 246247,
2002.
[33] X. Li and J. L. Cimini, "Effects of clipping and filtering on the performance of
OFDM," IEEE Commun. Lett., vol. 2, no. 5, pp. 131133, 1998.
[34] 40 Gbps LiNbO3 External Modu. (2004) Fujitsu.
[Online]. http://www.fujitsu.com/global/services/telecom/optcompo/lineup/40gln/
[35] O. E. Agazzi, M. R. Hueda, H. S. Carrer, and D. E. Crivelli, "Maximum-
likelihood sequence estimation in dispersive optical channels," J. Lightw.
Technol., vol. 23, no. 2, pp. 749763, 2005.
[36] T. Foggi, E. Forestieri, G. Colavolpe, and G. Prati, "Maximum-likelihood
sequence detection with closed-form metrics in OOK optical systems impaired by
GVD and PMD," J. Lightw. Technol. , vol. 24, no. 8, pp. 30733087, 2006.
[37] J. Armstrong, B. J. C. Schmidt, D. Kalra, H. A. Suraweera, and A. J. Lowery,
"Performance of asymmetrically clipped optical OFDM in AWGN for an
intensity modulated direct detection system," in IEEE GLOBECOM, 2006.
[38] J. Armstrong and B. J. C. Schmidt, "Comparison of Asymmetrically Clipped
Optical OFDM and DC-Biased Optical OFDM in AWGN," IEEE Comm. Letters,
vol. 12, no. 8, pp. 343345.
[39] A. J. Lowery, "Improving Sensitivity and Spectral Efficiency in Direct-Detection
Optical OFDM Systems," in Proc. Optical Fiber Commun. Conf. (OFC 2008),
OMM4, 2008.
References
139
[40] B. J. C. Schmidt, A. J. Lowery, and J. Armstrong, "Experimental demonstrations
of electronic dispersion compensation for long-haul transmission using direct-
detection optical OFDM," J. Lightw. Technol., vol. 26, no. 1, pp. 196203, 2008.
[41] J. M. Cioffi, G. P. Dudevoir, M. V. Eyuboglu, and G.D. Forney Jr, "MMSE
decision-feedback equalizers and coding Part II: Coding results," IEEE Trans.
on Commun., vol. 43, no. 10, pp. 25952604, 1995.
[42] G. D. Forney and G. Ungerboeck, "Modulation and coding for linear Gaussian
channels," IEEE Trans. on Info. Theory, vol. 44, no. 6, pp. 23842415, 1998.
[43] W. Yu, G. Ginis, and J. M. Cioffi, "An adaptive multiuser power control
algorithm for VDSL," in IEEE GLOBECOM, 2001.
[44] A. Ali, J. Leibrich, and W. Rosenkranz, "Spectral Efficiency and Receiver
Sensitivity in Direct Detection Optical-OFDM," in Proc. Optical Fiber Commun.
Conf. (OFC 2009), OMT7, 2009.
[45] A. Ali, H. Paul, J. Leibrich, W. Rosenkranz, and K. D. Kammeyer, "Optical
Biasing in Direct Detection Optical-OFDM for Improving Receiver Sensitivity,"
in Proc. Optical Fiber Commun. Conf. (OFC 2010), JThA12, 2010.
[46] J. W. Goodman, Statistical Optics.: Wiley-Interscience, 1985.
[47] J. M. Tang, P. M. Lane, and K. A. Shore, "Transmission performance of
adaptively modulated optical OFDM signals in multimode fiber links," IEEE
Photon. Technol. Lett., vol. 18, no. 1, pp. 205207, 2006.
[48] S. Randel, F. Breyer, and S. C. J. Lee, "High-Speed Transmission over
Multimode Optical Fibers," in Proc. Optical Fiber Commun. Conf. (OFC 2008),
OWR2, 2008.
[49] J. M. Tang, P. M. Lane, and K. A. Shore, "High-speed transmission of adaptively
modulated optical OFDM signals over multimode fibers using directly modulated
DFBs," J. Lightw. Technol., vol. 24, no. 1, pp. 429441, 2006.
References
140
[50] S. C. J. Lee, S. Randel, F. Breyer, and A. M. J. Koonen, "PAM-DMT for
Intensity-Modulated and Direct-Detection Optical Communication Systems,"
IEEE Photon. Technol. Lett., vol. 21, no. 23, pp. 17491751, 2009.
[51] X. Q. Jin, J. M. Tang, P. S. Spencer, and K. A. Shore, "Optimization of
adaptively modulated optical OFDM modems for multimode fiber-based local
area networks," J. Optical Networking, vol. 7, no. 3, pp. 198214, 2008.
[52] S. C. J. Lee, F. Breyer, S. Randel, H. P. A. van den Boom, and A. M. J. Koonen,
"High-speed transmission over multimode fiber using discrete multitone
modulation," J. Optical Networking, vol. 7, no. 2, pp. 183196, 2008.
[53] M. B. Shemirani and J. M. Kahn, "Higher-Order Modal Dispersion in Graded-
Index Multimode Fiber," J. Lightw. Technol., vol. 27, no. 23, pp. 54615468,
2009.
[54] J. Campello, "Optimal discrete bit loading for multicarrier modulation systems,"
in IEEE Symp. Info. Theory, 1998.
[55] P. Pepeljugoski, S. E. Golowich, A. J. Ritger, P. Kolesar, and A. Risteski,
"Modeling and simulation of next-generation multimode fiber links," J. Lightw.
Technol., vol. 21, no. 5, pp. 12421255, 2003.
[56] P. Pepeljugoski et al., "Development of system specification for laser-optimized
50-m multimode fiber for multigigabit short-wavelength LANs," J. Lightw.
Technol., vol. 21, no. 5, pp. 12561275, 2003.
[57] X. Q. Jin, J. M. Tang, K. Qiu, and P. S. Spencer, "Statistical Investigations of the
Transmission Performance of Adaptively Modulated Optical OFDM Signals in
Multimode Fiber Links," J. Lightw. Technol., vol. 26, no. 18, pp. 32163224,
2008.
[58] X. Zheng, J. M. Tang, and P. S. Spencer, "Transmission Performance of
Adaptively Modulated Optical OFDM Modems Using Subcarrier Modulation
References
141
over Worst-Case Multimode Fibre Links," IEEE Commun. Lett., vol. 12, no. 10,
pp. 788790, 2008.
[59] J. M. Kahn and J. R. Barry, "Wireless Infrared Communications," Proc. of the
IEEE, vol. 85, no. 2, pp. 265298, 1997.
[60] J. B. Carruthers and J. M. Kahn, "Modeling of Nondirected Wireless Infrared
Channels," IEEE Trans. on Commun., vol. 45, no. 10, pp. 12601268, 1997.
[61] J. Armstrong and A. J. Lowery B, "Power efficient optical OFDM," Electronics
Letters, vol. 42, no. 6, pp. 370372, 2006.
[62] F. R. Gfelle and U. Bapst, "Wireless In-House Data Communication via Diffuse
Infrared Radiation," Proc. of the IEEE, vol. 67, no. 11, pp. 14741486, 1979.
[63] O. Gonzalez, R. Perez-Jimenez, S. Rodriguez, J. Rabadan, and A. Ayala,
"Adaptive OFDM system for communications over the indoor wireless optical
channel," IEE Proc. Optoelectronics, vol. 153, no. 4, pp. 139144, 2006.
[64] H. Joshi, R. J. Green, and M. S. Leeson, "Multiple Sub-carrier Optical Wireless
Systems," in ICTON 2008, vol. 4, 2008, pp. 184188.
[65] J. B. Carruthers and J. M. Kahn, "Multiple-Subcarrier Modulation for Non-
Directed Wireless Infrared Communication," IEEE J. Sel. Areas in Commun., vol.
14, no. 3, pp. 538546, 1996.
[66] S. K. Wilson and J. Armstrong, "Transmitter and Receiver methods for improving
Asymmetrically-Clipped Optical OFDM," IEEE Trans. on Wireless Commun.,
vol. 8, no. 9, pp. 45614567, 2009.
[67] J. R. Barry, J. M. Kahn, W. J. Krause, E. A. Lee, and D. G. Messerschmitt,
"Simulation of Multipath Impulse Response for Wireless Optical Channels,"
IEEE J. Sel. Areas in Commun., vol. 11, no. 3, pp. 367379, 1993.
[68] J. M. Kahn, W. J. Krause, and J. B. Carruthers, "Experimental Characterization of
Non-Directed Indoor Infrared Channels," IEEE Trans. on Commun., vol. 43, no.
References
142
234, pp. 16131623, 1995.
[69] J. Grubor, O. C. Gaete Jamett, J. W. Walewski, S. Randel, and K.-D. Langer,
"High-Speed Wireless Indoor Communication via Visible Light," in ITG
FACHBERICHT, 2007.
[70] J. B. Carruthers and P. Kannan, "Iterative site-based modeling for wireless
infrared channels," IEEE Trans. on Antennas and Propag., vol. 50, no. 5, pp.
759765, 2002.
[71] E. Forestieri, Optical Communication Theory and Techniques.: Springer, 2005.
[72] T. J. Richardson and R. L. Urbanke, Modern Coding Theory.: Cambridge Univ.
Pr., 2008.
[73] D. J. F. Barros and J. M. Kahn, "Comparison of Orthogonal Frequency-Division
Multiplexing and On-Off Keying in Amplified Direct-Detection Single-Mode
Fiber Systems," J. Lightw. Technol., vol. 28, no. 12, pp. 18111820, 2010.
[74] D. J. F. Barros and J. M. Kahn, "Comparison of Orthogonal Frequency-Division
Multiplexing and On-Off Keying in Direct-Detection Multimode Fiber Links," in
press for J. Lightw. Technol., 2011.
[75] W. Henkel, G. Taubock, P. Odling, P.O. Borjesson, and N. Petersson, "The
Cyclic Prefix of OFDM/DMT An Analysis," in Proc. International Seminar on
Broadband Communications, 2002.

Final Project OFDM

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Final Project OFDM

Uploaded by

Copyright:

Available Formats

ORTHOGONAL FREQUENCY-DIVISION MULTIPLEXING

FOR OPTICAL COMMUNICATIONS

and R = 29.66 GHz. ....................................... 54

is the voltage that must be applied to the single

due to the mapping of the

is the drive voltage that

and that they could be interchanged and the point

and drive voltage values

. We observe that the optical field in between

and the error in

and R = 29.66 GHz.

. LED emission wavelength typically lies between 850

You might also like