Sunteți pe pagina 1din 219

Optical performance monitoring for amplified

spontaneous emission related issues in


transparent optical networks




Michal Pawel Dlubek, MSc.








Thesis submitted to the University of Nottingham for the
degree of Doctor of Philosophy

November 2008













































Icv,[o.vnc 1v,.c
Zv o, 2c .,:,nvIv
Abstract
With the increasing complexity oI optical networks, the associated monitoring
requirements have become increasingly demanding. At the same time, traditional
digital signal monitoring in the electronic domain becomes more diIIicult, especially
as new types oI traIIic and data protocols and new network conIigurations emerge. As
a consequence, in transparent optical networks one must rely upon optical
perIormance monitoring (OPM) methods, which essentially measure analogue
parameters oI the signal in the optical domain.
Several aspects oI OPM are discussed within this thesis. However, the main
Iocus has been on the bit error rate (BER) and optical signal to noise ratio (OSNR),
since these are typically the most important signal quality measures in optical
networks.
As a method to generate signal histograms, which are crucial Ior BER
estimation, optical sampling using Iour-wave mixing (FWM) in semiconductor optical
ampliIiers (SOAs) is considered. Experimental results and theoretical noise analysis
oI such an optical sampling system are presented. Preparatory work on SOA
characterization also led to new results on both Iundamental physical understanding
oI FWM in SOAs and on its application to optical regeneration.
Further exploring the issue oI BER estimation, a new signal quality metric is
proposed. This metric is based on statistical moments, hence it is well suited to be
applied in tandem with optical sampling systems. The new metric is shown to possess
certain advantages over the widely used Q-Iactor method.
The subject oI OSNR monitoring has also been investigated. As a result, a
novel method Ior OSNR monitoring, based on modulation spectrum assessment, is
developed. Advantages and limitations oI the new method are discussed.
Finally, nonlinear evolution oI ampliIied spontaneous emission noise in
nonlinear Iibre is investigated. An approximation oI the nonlinear Iibre as a zero-
memory nonlinear device is used. It is shown that in certain situations the presence oI
Iibre nonlinearity results in signiIicantly reduced perIormance and may introduce
errors in BER estimation utilizing the Q-Iactor.
Acknowledgements

I would like to thank my supervisors, Dr Andy Phillips and ProI. Eric Larkins
Ior giving me the ultimate Ireedom in my research. Without the Ireedom I was given,
this thesis would have never taken its present shape. Now I can only vaguely
remember the days when I pondered whether this Ireedom was a privilege or a curse.
DeIinitely a privilege, aIter all.

I wish to thank all the Iellow students and researchers I met during my years at
Nottingham. It is impossible to mention you all, but special thanks go to Steve, who
seemed to have every tool I ever needed, Harvy, who was a very decent housemate
and enjoyable chess partner and also to Fen, Bessie and Simeon Ior the good times oI
procrastination. And to Christian Ior the political debates we had.

I wish to express my gratitude to EPSRC Ior Iunding me throughout my PhD
under the grant GR/T09309/01, which has made this work possible.

And last but not least, I want to thank my mom, dad, brother and my little
sister, who is not so little any more. And, oI course, I owe a lot to Edytka, Ior her
patience.

Acronyms and abbreviations

AB - assist beam
ASE - ampliIied spontaneous emission
ASK - amplitude shiIt keyed
ASON - automatically switched optical network
BDI - birth-death-immigration
BER - bit error rate
BPF- optical bandpass Iilter
CDP - carrier density pulsations
CH - carrier heating
CRC - optical circulator
CW - continuous wave
DFB - distributed Ieedback
DFG - diIIerence Irequency generation
DPSK - diIIerential phase shiIt keyed
DWDM - dense wavelength division multiplexing
EAM - electro-absorption modulator
EDFA - erbium doped Iibre ampliIier
ER - extinction ratio
FEC - Iorward error correction
FWHM - Iull-width at halI maximum
FWM - Iour-wave mixing
HNLF - highly nonlinear Iibre
ISO - optical isolator
mgI - moment generating Iunction
MQW - multiple quantum well
MZ - Mach-Zehnder
NLPN - nonlinear phase noise
NLSE - nonlinear Schroedinger equation
NRZ - non return-to-zero
ODL - optical delay line

OPM - optical perIormance monitoring
OSA - optical spectrum analyzer
OSNR - optical signal to noise ratio
PC - polarization controller
pdI - probability density Iunction
PLL - phase locked loop
PMD - polarization mode dispersion
QAM - quadrature-amplitude modulation
RFSA - RF spectrum analyzer
RIN - relative intensity noise
ROADM - reconIigurable optical add-drop multiplexer
RZ - return-to-zero
SBR - signal to ASE background ratio
SDH - Synchronous Digital Hierarchy
SFG - sum Irequency generation
SGM - selI-gain modulation
SHB - spectral hole burning
SMF - single mode Iibre
SNR - signal to noise ratio
SOA - semiconductor optical ampliIier
SONET - Synchronous Optical Network
SOP - state oI polarization
SPM - selI-phase modulation
TEC - thermoelectric cooler
TIA - transimpedance ampliIier
WDM - wavelength division multiplexing
XGM - cross-gain modulation
XPM - cross-phase modulation
ZMNL - zero-memory nonlinear

.


Contents
1. Introduction the need Ior optical perIormance monitoring.................................1
1.1. The current status and the Iuture oI OPM......................................................4
1.2. Outline oI the thesis .......................................................................................7
1.3. ReIerences......................................................................................................9
2. Four-wave mixing in semiconductor optical ampliIiers ......................................14
2.1. Basic physics oI the SOAs...........................................................................14
2.1.1. Optical gain in semiconductors............................................................15
2.1.2. AmpliIied spontaneous emission noise................................................17
2.1.3. Rate equation and propagation equation..............................................18
2.2. Nonlinear eIIects in SOAs ...........................................................................21
2.2.1. Third order nonlinear processes in SOAs ............................................21
2.2.2. Interband nonlinear processes in SOAs ...............................................23
2.2.3. Intraband nonlinear processes in SOAs ...............................................24
2.3. Four wave mixing in SOAs..........................................................................26
2.3.1. Time domain model oI FWM in SOAs................................................27
2.4. ReIerences....................................................................................................31
3. Optical sampling Ior OPM and signal characterization.......................................34
3.1. Principle oI optical sampling .......................................................................35
3.1.1. Review oI optical sampling gates ........................................................39
3.2. Technical realizations oI the optical sampling system Ior OPM.................42
3.3. Optically sampled histograms and statistical moments ...............................43
3.4. ReIerences....................................................................................................47
4. The Q-Iactor method in OPM..............................................................................52
4.1. Short derivation oI the Q-Iactor...................................................................53
4.2. Q-Iactor measurement methods ...................................................................56
4.3. Limitations oI the Q-Iactor ..........................................................................58
4.4. ReIerences....................................................................................................61
5. Characterization and applications oI SOAs .........................................................64
5.1. Gain and noise measurements......................................................................64
5.2. CW Iour wave mixing in SOAs...................................................................67
5.2.1. Dependence oI CW FWM eIIiciency and SBR on pump power .........69
5.2.2. Dependence oI CW FWM eIIiciency and SBR on detuning...............71
5.2.3. Linearity oI CW FWM.........................................................................72
5.3. Experimental veriIication oI the existence oI optically induced carrier
pulsations in SOAs...................................................................................................74
5.3.1. Experimental setup...............................................................................75
5.3.2. Experimental results.............................................................................76
5.4. Extinction ratio improvement using nonlinear Iour-wave mixing in SOAs
with assist beam.......................................................................................................80
5.4.1. Experimental setup and operating principle ........................................81

5.4.2. Experimental veriIication oI regenerative capabilities oI the nonlinear
FWM , .............................................................................................................83
5.5. Summary......................................................................................................87
5.6. ReIerences....................................................................................................88
6. Optical sampling using FWM in SOAs ...............................................................92
6.1. Analytical model oI picosecond FWM in SOAs .........................................92
6.1.1. Theoretical conversion eIIiciency........................................................93
6.1.2. Dependence oI the conversion eIIiciency on the sampling pulse width
and sampling pulse energy...................................................................................95
6.1.3. Dependence oI the conjugate energy on the SOA gain .......................97
6.1.4. Dependence oI the conjugate energy on the input signal power .........98
6.1.5. Dependence oI the conversion eIIiciency on the Irequency detuning
between the signal and sampling pulses ..............................................................99
6.1.6. Dependence oI the optimum pump energy on the SOA gain ............100
6.2. Noise analysis oI the optical sampling system based on FWM in SOAs ..101
6.2.1. Signal to noise ratio in the considered sampling system...................102
6.2.2. Sensitivity analysis.............................................................................109
6.3. Experimental investigation oI optical sampling using FWM in SOAs......111
6.3.1. Characterization oI the sampling setup..............................................112
6.3.1.1. Detection linearity......................................................................112
6.3.1.2. SNR limits oI the sampling setup ..............................................113
6.3.1.3. Spectral shiIt oI sampling pulses in SOAs.................................116
6.3.2. Experimental results...........................................................................120
6.3.2.1. Dependence oI the conversion eIIiciency on sampling pulse
energy ....................................................................................................121
6.3.2.2. Dependence oI the conjugate energy on the detuning ...............122
6.3.2.3. Dependence oI the conjugate energy and SNR on input signal
power ....................................................................................................124
6.3.2.4. Pulsed bias current injection......................................................127
6.4. Concluding remarks on the applicability oI the investigated method........130
6.5. ReIerences..................................................................................................132
7. Alternative optical signal quality metric Ior BER estimation............................135
7.1. Physical characteristics oI the electrical noise at the receiver output ........136
7.2. Derivation oI the new metric .....................................................................137
7.3. Evaluation oI the proposed signal quality metric ......................................142
7.3.1. Second order version oI the new metric ............................................148
7.4. Summary....................................................................................................152
7.5. ReIerences..................................................................................................153
8. OSNR monitoring method based on modulation spectrum assessment ............156
8.1. State-oI-the-art in OSNR monitoring methods ..........................................156
8.2. Principle oI the proposed OSNR monitoring method................................158
8.3. Robustness analysis against other signal impairments ..............................162
8.3.1. Impact oI SPM in Iibre ......................................................................163
8.3.2. Robustness against degraded extinction ratio....................................165
8.3.3. Robustness against modulator chirp ..................................................165
8.3.4. Robustness against electrical drive characteristic..............................166

8.3.5. Sensitivity to diIIerent pulse shapes ..................................................167
8.3.6. Sensitivity to channel Irequency misalignment .................................168
8.4. Summary and conclusions .........................................................................169
8.5. ReIerences..................................................................................................170
9. Nonlinear evolution oI Gaussian ASE noise in zero-memory nonlinear Iibre ..172
9.1. Propagation oI light in optical Iibre...........................................................173
9.1.1. ZMNL model oI nonlinear Iibre ........................................................175
9.2. Propagation oI purely incoherent Iield in ZMNL Iibre .............................177
9.3. Propagation oI moments oI the quadratures in ZMNL Iibre .....................178
9.3.1. Derivation oI expressions Ior moments .............................................179
9.3.2. Propagation oI kurtosis ......................................................................183
9.3.2.1. VeriIication using Monte-Carlo simulations .............................185
9.4. Propagation oI pdIs oI the quadratures in ZMNL Iibre .............................188
9.4.1. Evolution oI the pdIs oI the ASE quadratures ...................................188
9.4.2. InIluence oI Kerr nonlinearity on BER in ASK systems...................192
9.4.2.1. Direct detection systems ............................................................192
9.4.2.2. Coherent amplitude modulated systems ....................................194
9.5. Summary....................................................................................................196
9.6. ReIerences..................................................................................................197
10. Conclusions and Iuture work .............................................................................200
10.1. Future work................................................................................................202
Appendix A. Derivation oI Iormulae Ior optically sampled moments ......................205
A.1. ReIerences..................................................................................................207
Appendix B. Derivation oI the variance oI the ASE-signal beating in the presence oI
temporal gating in optical domain .............................................................................208
B.1. ReIerences..................................................................................................210


















1
1. Introduction - the need for optical performance
monitoring
Since 1980, when the Iirst commercial optical Iibre link was deployed |1|, the
Iield oI optical communications has seen an enormous development. During only 25
years, the transmission speed has increased Irom a mere 45 Mb/s to the 40 Gb/s being
currently widely commercially deployed, with 100 Gb/s already on the horizon |2, 3|.
This huge growth is driven by people`s need to communicate and exchange
data and inIormation. Initially, predominantly voice (telephone) traIIic has been
carried over the Iirst Iibre links, but the main driver Ior the massive bandwidth
increase turned out to be the introduction oI the Internet, which has gradually become
a commodity Ior a large part oI the world. In particular, it was the increasing
popularity oI various bandwidth-hungry Internet services and applications, especially
those multimedia-related, which has resulted in vast amount oI data generated by the
average Internet user |1|.
On the other hand, this rapid technological progress has been possible thanks
to several important technological breakthroughs |1, 4|, such as low noise
semiconductor lasers operating in 1.3 m and 1.55 m regions, improved
transmission Iibres with low polarization mode dispersion (PMD) and tailored
chromatic dispersion, advanced modulation schemes, optical ampliIiers (mainly
erbium doped Iibre ampliIiers (EDFAs) and Raman ampliIiers) and Iast electronics
and Iorward error correction codes (FEC), to mention just a Iew. At the same time, the
wavelength division multiplexing technology (WDM), Iollowed later by dense WDM
(DWDM), allowed the total system throughput transmitted over a single Iibre strand
to be increased to several terabits |1|. The length oI optical links systematically and
rapidly increased as well. The Iirst transatlantic link was already put into service in
1987 and was soon Iollowed by a transpaciIic connection |5|. However, these
pioneering ultra-long haul links used electronic repeaters, placed every couple oI tens
oI kilometres, to restore the optical signal quality. As a result, these links were more
oI a concatenation oI many short single-span links. It was realized that this approach
is not only expensive, as the upgrade and maintenance costs were prohibitive, but also
limits the practical transmission speed. Since the conception and commercial
availability oI EDFAs, multispan optically ampliIied links and networks have become
2
a paradigm
1
and the Iirst optically ampliIied transoceanic links were laid in 1995 |5,
6|. In practice, the length oI optically ampliIied links is limited by the total
accumulated optical ampliIied spontaneous emission (ASE) noise (assuming optimum
link design). An intense research eIIort is put into pushing the length and speed limits
even Iurther, by introducing optical regenerators |7, 8|.
The very long link length in ultra-long haul networks, combined with the ever
increasing data rates, poses signiIicant problems Ior network design, operation and
maintenance. Traditionally, a very low bit error rate (BER) is required Irom the
optical networks, with the exact requirements varying depending on the application
|1, 4, and 9|. At the same time, the optical signal quality is inevitably degraded during
the transmission, which leads to detection errors. The ASE noise is one source oI
signal degradation, but there are also a host oI other eIIects, including linear eIIects
(e.g. PMD, chromatic dispersion, attenuation, reIlections and multipath interIerence,
chirp, non-ideal Iiltering and linear crosstalk) and nonlinear eIIects (e.g. cross gain
modulation, selI-phase modulation, cross phase modulation, Raman crosstalk,
intrachannel Raman and Brillouin scattering, Iour-wave mixing and intrachannel Iour-
wave mixing) and even more complex phenomena involving interaction between any
oI these eIIects. The severity oI most oI these eIIects scale with increasing the bit rate.
For example, a 40 Gb/s signal is 16 times more sensitive to chromatic dispersion or
requires a 6 dB better optical signal to noise ratio (OSNR) than its 10 Gb/s
counterpart.
In eIIect, as the network complexity and transmission speed increase, the need
Ior accurate signal and link characterization and monitoring intensiIies |10-14|. Not
only do more parameters need to be measured or monitored, but also the requirement
Ior the measurement accuracy becomes more stringent. At the same time, in-service
measurement techniques are preIerred i.e. the measurement should not cause service
interruption. Obviously, end-to-end BER remains the ultimate Iigure oI merit Ior a
digital communication system. II Iorward error correction (FEC) codes are used, FEC
chips have the capability to measure in real time the number oI errors and calculate
the BER. However, FEC chips are normally placed only at the Iar end oI the link
within the receiver. ThereIore, they cannot be used to isolate a Iaulty element nor can

1
Almost all equipment vendors Iollow this paradigm, with the notable exception oI InIinera (Iounded
in 2001), who challenged this concept with their vision oI photonic integration combined with cost-
eIIective and commonplace opto-electronic conversion.
3
they provide any inIormation about the root cause oI the problem |15|. Similar
limitations apply to monitoring using control bytes in the SONET/SDH (Synchronous
Optical Network/Synchronous Digital Hierarchy) Irame header. Moreover, using the
monitoring Iunctions oI higher network levels is eIIective only when the transmission
quality is already aIIected and does not allow any preventive maintenance actions
|15|.
The electronics and FEC chip limitations and the cost reduction rationale both
Iavour the so called optical perIormance monitoring (OPM) approach. In OPM the
physical (analog) properties oI the optical signal are measured without converting it
Iirst into an electrical digital signal in order to measure the electrical signal
characteristics. An important issue is also the optical power budget available Ior
monitoring at each node. Full electrical signal monitoring requires practically the
same optical power as the Iinal receiver, whereas normally only a Iew percent can be
tapped oII the signal. Additionally, digital signal monitoring requires Iull dispersion
compensation, whereas the dispersion map period can be larger than single span.
Moreover, a lot oI important diagnostic inIormation, like optical phase or polarization,
is normally lost aIter the photodetection process.
Even iI, Ior the currently required network Iunctionality, monitoring oI the
channel power, wavelength and OSNR during the normal network operation usually
suIIices |10, 16|, additional pieces oI inIormation are indispensable iI a Iault occurs or
when the BER suddenly worsens aIter a change in the link conIiguration |10|. In this
case, OPM should enable rapid identiIication, location and isolation oI the source oI
signal impairment |16|. Additionally, early detection oI signal degradation beIore it
has an intolerable impact on BER can be used Ior activation oI resilience mechanisms
or as the correction signal Ior diIIerent control loops.
To trace the eIIect responsible Ior the decrease in the transmission quality, it is
oIten necessary to comprehensively characterize the optical signal. This may be a
challenging task in the modern network environment, when, Ior example, tight
Iiltering makes measuring OSNR diIIicult or the state oI polarization driIts due to
unstable inline components. This is even more challenging in optically switched
dynamic networks, as each wavelength can have a diIIerent history |14, 17-19|. The
challenge will be even Iurther exacerbated in data-oriented packet-switched next
generation networks, which seem likely to supersede current SONET/SDH based
networks |20|. In these networks, the perIormance monitoring currently managed at
4
the transport layer by the SONET/SDH protocols will probably have to be perIormed
at the physical layer, relying only upon OPM |10, 16, 20|.
The need Ior and advantages oI OPM and, also, its technological challenges
have been widely recognized. ThereIore, recent years have seen intense research and a
proliIeration oI diIIerent OPM techniques. The ultimate goal is not only to address
this need Ior more Ilexible and powerIul monitoring oI the optical signal, but also to
make sure that the technology is cost eIIective. A short review oI the current state oI
OPM is given in the next section.
1.1. The current status and the future of OPM
A limited implementation oI OPM is currently a reality in reconIigurable
optical networks, and is becoming more common as the commercially available
reconIigurable optical add-drop multiplexers (ROADMs) are becoming more Ilexible
and sophisticated and are being implemented not only in wide-area networks, but also
in metro networks |12|. Some Iorm oI OPM (mainly power and wavelength
monitoring) is also considered necessary in passive access networks |15|, where it
could reduce operational costs. The current, commercially available OPM devices are
actually card-level optical spectrum analyzers (OSAs), which thereIore allow
monitoring oI the channel presence, wavelength and power, and in some case also the
OSNR |21, 22|. Although very useIul, these devices have inherent limitations, as they
can only extract inIormation about the optical spectrum, which has only a limited
connection not only with the BER but also with the many detrimental eIIects in the
link. The OSNR values extracted Irom optical spectrum measurements are likely to be
inaccurate due to tight Iiltering. Some oI the channel monitors make use oI
polarization inIormation to calculate the OSNR, but this also means that they are
prone to polarization related eIIects. Other optical signal properties are usually not
monitored in today`s links. ThereIore, current OPM monitors may not be suIIicient
Ior next generation networks.
As described earlier, the Iield oI OPM is currently receiving a lot oI interest. It
is not certain at the moment what will be the established monitoring standard, i.e.
what signal quantities and properties will have to be measured in practical networks
given budget constraints |10, 19|. This is likely to depend on particular application.
In general, the tasks envisaged Ior OPM mentioned earlier can be categorized into
5
Iour groups (although oIten the boundaries between diIIerent groups are somewhat
unclear) |16, 22|:
- providing correction signal Ior control loops (especially Ior ampliIiers and
compensators, i.e. PMD compensators);
- controlling network elements (e,g, Ior channel levelling and equalizing or Ior
channel pre-emphasis);
- link monitoring, optimization and provisioning;
- Iault detection, location and isolation, Iollowed by connection restoration.
The OPM reIerence model (Iirst proposed in |11|) breaks the possible OPM
Iunctionalities into 3 layers, as shown in Fig. 1.1. The Iirst layer, WDM channel
monitoring, involves the determination oI the most Iundamental, physical properties
oI the optical wave, such as power, wavelength, OSNR and dispersion. The currently
deployed OPM capabilities are limited to the Iirst layer oI the OPM reIerence model.
In some reIerences, the notion oI OPM is actually limited to the Iirst (and sometimes
partly second) layer (see Ior example |10, 18|). The second layer, channel quality
monitoring, involves the measurements oI some average or statistical properties oI the
digital electric signal, without the need Ior a Iull receiver. Examples oI the second
layer quantities include signal histograms, Q-Iactor, extinction ratio and monitoring
the signal level splitting due to various distortion or crosstalk eIIects. Among these,
the Q-Iactor has probably attracted the most attention, due to its link with BER.
However, techniques Ior the measurement oI other quantities are also being
vigorously pursued. In general, this monitoring layer is not yet implemented in
practical networks. The top (third) layer in the OPM reIerence model is the protocol
perIormance monitoring layer. Its realization in the optical domain would require
photonic digital signal processing capabilities, which are not available at the moment.
In the electrical domain, Iull signal processing and decoding is possible only at the
receiver, as explained beIore. Within the all-optical, transparent network paradigm,
there is a shiIt Irom the monitoring done by electronic means (the electrical equivalent
oI the third layer oI the OPM model) to lower layer monitoring (which can be done
only using OPM techniques). As the BER remains the ultimate metric Ior a digital
system, end-to-end BER monitoring will continue to be the Iinal benchmark Ior the
eIIiciency and usability oI the new low level OPM techniques.
6

Fig. 1.1 OPM reIerence model (aIter |11|)
There is a long list oI the optical signal properties measured using OPM
methods. As mentioned above, it is not clear at the moment which oI these quantities
will be measured in next generation networks. The list below presents the various
signal characteristics, which have been shown to be measurable using OPM |11, 23|:
- average power |23|;
- eye diagram |24, 25|;
- waveIorm/pulse shape |26-28|;
- OSNR |19, 23, 29-31|;
- state oI polarization and degree oI polarization |32|;
- carrier wavelength |33, 34|;
- power histogram |25, 35|;
- Q-Iactor |16, 17|;
- PMD |32, 36, 37|;
- chromatic dispersion |23, 33, 36|;
- crosstalk |33, 34, 38|;
- extinction ratio |22|;
- phase and Iield histogram |39|.
Among these many parameters, the likely candidates Ior Iuture advanced
OPM monitors are the in-band OSNR, Q-Iactor, chromatic and polarization dispersion
|10, 11|, accompanied by the currently measured wavelength and channel power.
There are a multitude oI techniques proposed to measure any oI these quantities and it
is diIIicult to predict which method will become the standard. However, there seems
7
to be agreement that the sampling and histogram methods will be very important. This
is because these methods have in principle the potential to measure (depending on
application) the eye diagram and waveIorm shape, the extinction ratio and the
amplitude and power histograms. They are also sensitive to diIIerent signal
distortions. Similarly, the Q-Iactor (or some alternative analogue metric related to
digital BER) and in-band OSNR will be indispensable in Iuture networks. ThereIore,
this thesis explores these important aspects oI optical sampling, BER estimation and
OSNR monitoring.
1.2. Outline of the thesis
The purpose oI this thesis is to investigate techniques Ior advanced optical
perIormance monitoring Ior Iuture transparent and reconIigurable optical networks.
Three main aspects oI OPM are investigated within the scope oI this thesis:
- optical sampling using Iour wave mixing in semiconductor optical ampliIiers;
- BER estimation using analogue parameters describing the optical signal
quality and measured using OPM techniques; and
- in-band OSNR monitoring.
Chapter 2 serves as a theoretical introduction to the matter oI nonlinear eIIects
in semiconductor optical ampliIiers (SOAs) in general and Iour-wave mixing (FWM)
in particular. The applicability oI FWM in SOAs Ior optical sampling Ior OPM will
be investigated later in this thesis. The theoretical model oI FWM brieIly described in
this chapter is also used in Chapter 6 as a tool Ior designing the optical sampling
system.
Chapter 3 discusses the importance oI optical sampling and presents its
diIIerent Ilavours and approaches. In particular, the application oI optical sampling
Ior histogram reconstruction is explored. Based on these histograms, BER can be
estimated without the need Ior Iull digital signal detection.
Chapter 4 brieIly describes the currently most popular method Ior BER
estimation the Q-Iactor. The advantages and shortcomings oI the Q-Iactor are
examined. Various attempts to extend the Q-Iactor validity are discussed. Finally, the
diIIerence between the Q-Iactors calculated using electrically sampled and optically
sampled histograms is explained.
In Chapter 5, the experimental examination oI FWM in SOAs between
continuous waves is reported. This also serves as an introduction to the more diIIicult
8
experiments on optical sampling using FWM. Section 5.3 looks into the importance oI
carrier density pulsations in FWM and discusses the possible physical origin oI this
phenomenon. The work presented in this section will be submitted Ior publication
|40|. Also, a novel application oI FWM Ior signal regeneration is described in Section
5.4. The results presented in this section have led to publication |41|.
In Chapter 6, the Ieasibility oI using optical sampling based on Iour-wave
mixing (FWM) in semiconductor optical ampliIiers (SOAs) is investigated, using both
theoretical models and experiments.
Chapter 7 proposes a novel signal quality metric, which is shown to possess
certain advantages compared to the Q-Iactor. The new metric is based on the Laguerre
expansion and statistical moments, and is well suited Ior use with sampling monitors.
The Iull derivation oI the metric is presented. This chapter has led to publication |42|.
Chapter 8 discusses a novel in-band OSNR monitoring technique based on
modulation spectrum assessment. The main advantages oI the technique are its
simplicity and cost-eIIectiveness. The technique requires narrowband optical Iiltering,
but this is oIten already available in the network nodes as card-level OSAs are widely
deployed, as discussed above. The work presented in this chapter has led to
publications |43, 44|.
Chapter 9 investigates the evolution oI the ASE noise in optical Iibre. As Iibre
is a nonlinear medium, the statistical properties oI the signal and ASE propagating
along the link are inevitably modiIied, which may lead to BER degradation. Using the
zero-memory nonlinear (ZMNL) Iibre approximation Ior highly nonlinear Iibre,
analytical expressions Ior the moments oI the nonlinearly transIormed ASE noise are
derived. The BER degradation Ior the case oI coherent amplitude modulated systems
is examined and it is shown that the Q-Iactor method loses its accuracy in this case.
The work presented in this chapter has led to publications |45, 46|.
Finally, Chapter 10 presents some general conclusions and discusses the need
Ior Iuture work.



9
1.3. References
|1| G. P. Agrawal, Lightwave Technologv. Telecommunication Svstems.
Hoboken: Wiley, 2005.
|2| E. Cornejo, "Choices emerge Ior 40G and 100G applications," Lightwave, vol.
25, 2008.
|3| R. Rubenstein, "Carriers still ponder their high-speed trajectory," fibersvstems
Europe, vol. 12, 2008.
|4| R. Ramaswami and K. Sivarajan, Optical Networks, 2 ed: Morgan KauImann,
2002.
|5| G. Fouchard, "Historical Overview oI Submarine Communication Systems,"
pp. 16-48, in Undersea Fiber Communication Svstems, J. Chesnoy, Ed.,
Academic Press, 2002.
|6| N. Bergano, "Undersea Communication Systems," pp. 154-195, in Optical
Fiber Telecommunications, vol. IVA, I. Kaminov and T. Li, Eds., Academic
Press, 2002.
|7| O. Leclerc, B. Lavigne, D. Chiaroni, and E. Desurvire, "All-Optical
Regeneration: Principles and WDM Implementation," pp. 732-783, in Optical
Fiber Telecommunications, vol. IVA, I. Kaminov and T. Li, Eds., Academic
Press, 2002.
|8| O. Leclerc, B. Lavigne, E. BalmeIrezol, P. Brindel, L. Pierre, D. Rouvillain,
and F. Seguineau, "Optical regeneration at 40 Gb/s and beyond," J. Lightw.
Technol., vol. 21, pp. 2779-2790, 2003.
|9| J. Zyskind, R. Barry, G. Pendock, M. Cahill, and J. Ranka, "High-Capacity,
Ultra-Long-Haul Networks," pp. 198-231, in Optical Fiber
Telecommunications, vol. IVA, I. Kaminov and T. Li, Eds., Academic Press,
2002.
|10| M. N. Petersen, "PerIormance Monitoring in the Next Generation oI Optical
Networks," Int. Conf. Photonics in Switching PS 2006, 2006, pp. 1-3, 2006.
|11| D. C. Kilper, R. Bach, D. J. Blumenthal, D. Einstein, T. Landolsi, L. Ostar, M.
Preiss, and A. E. Willner, "Optical perIormance monitoring," J. Lightw.
Technol., vol. 22, pp. 294-304, 2004.
|12| W. Grupp, "Monitoring requirements Ior optical transparent networks," Opt.
Fiber Commun. Conf. OFC 2006, pp. 3-11, 2006.
10
|13| A. Kirstaedter, M. Wrage, G. Goeger, W. Fischler, and B. Spinnler, "Current
aspects oI optical perIormance monitoring and Iailure root cause analysis in
optical WDM networks," Opt. Transmission, Switching, and Subsvstems Conf.
2005, pp. 362-373, 2005.
|14| L. MeIlah, B. Thomsen, J. Mitchell, P. Bayvel, G. Lehmann, S. Santoni, and
B. Bollezn, "Advanced Optical PerIormance Monitoring Ior Dynamically
ReconIigurable Networks," Europ. Conf. Netw. Opt. Commun. NOC 2005, pp.
, 2005.
|15| J. Hehmann and T. PIeiIIer, "New monitoring concepts Ior optical access
networks," Bell Labs Tech. J., vol. 13, pp. 183-198, 2008.
|16| R. E. Neuhauser, "Optical perIormance monitoring (OPM) in next-generation
optical networks," Network Design and Management Conf. 2002, pp. 34-42,
2002.
|17| A. Richter, W. Fischler, H. Bock, R. Bach, and W. Grupp, "Optical
perIormance monitoring in transparent and conIigurable DWDM networks,"
IEE Proc-Optoelectron., vol. 149, pp. 1-5, 2002.
|18| L. K. Chen, C. C. K. Chan, G. W. Lu, Y. C. Ku, S. T. Ho, and C. Lin, "Optical
perIormance monitoring and network diagnosis in reconIigurable optical
networks," Network Architectures, Management, and Applications J Conf.
2007, pp. 67841I-8, 2007.
|19| D. C. Kilper and W. Weingartner, "Monitoring optical network perIormance
degradation due to ampliIier noise.," J. Lightw. Technol., vol. 21, pp. 1171-
1178, 2003.
|20| J. Yang, M. Y. Jeon, J. Cao, Z. Pan, and S. J. B. Yoo, "PerIormance
monitoring in transparent optical networks using selI-monitoring optical-
labels," Electron. Lett., vol. 40, pp. 1370-1372, 2004.
|21| S. Hardy, "Whatever you call them, monitors measuring up," Lightwave, vol.
24, 2007.
|22| S. Wielandy, M. Fishteyn, and B. Zhu, "Optical perIormance monitoring using
nonlinear detection," J. Lightw. Technol., vol. 22, pp. 784-793, 2004.
|23| G. Rossi, T. E. Dimmick, and D. J. Blumenthal, "Optical perIormance
monitoring in reconIigurable WDM optical networks using subcarrier
multiplexing," J. Lightw. Technol., vol. 18, pp. 1639-1648, 2000.
11
|24| J. Li, J. Hansryd, P. O. Hedekvist, P. A. A.-A. Andrekson, P.A., and S. N. A.-
K. Knudsen, S.N., "300-Gb/s eye-diagram measurement by optical sampling
using Iiber-based parametric ampliIication," IEEE Photon. Technol. Lett., vol.
13, pp. 987-989, 2001.
|25| M. Westlund, H. Sunnerud, M. Karlsson, and P. A. A.-A. Andrekson, P.A.,
"SoItware-synchronized all-optical sampling Ior Iiber communication
systems," J. Lightw. Technol, vol. 23, pp. 1088-1099, 2005.
|26| M. Westlund, P. A. Andrekson, H. Sunnerud, J. A.-H. Hansryd, J., and J. L.
A.-J. Li, "High-perIormance optical-Iiber-nonlinearity-based optical
waveIorm monitoring," J. Lightw. Technol, vol. 23, pp. 2012-2022, 2005.
|27| P. A. Andrekson and M. Westlund, "High-Speed WaveIorm Analysis Using
All-Optical Sampling," pp. 421-504, in Digital Communications Test and
Measurement. High-Speed Phvsical Laver Characteri:ation, D. Derickson;
and M. Mller, Eds., Prentice Hall, 2007.
|28| K.-L. Deng, R. J. Runser, I. Glesk, and P. R. A.-P. Prucnal, P.R., "Single-shot
optical sampling oscilloscope Ior ultraIast optical waveIorms," IEEE Photon.
Technol. Lett., vol. 10, pp. 397-399, 1998.
|29| C. Xie, D. C. Kilper, L. Moller, and R. RyI, "Orthogonal-Polarization
Heterodyne OSNR Monitoring Insensitive to Polarization-Mode Dispersion
and Nonlinear Polarization Scattering," J. Lightw. Technol., vol. 25, pp. 177-
183, 2007.
|30| Z. C. Zhenning Tao, Libin Fu, Deming Wu, Anshi Xu,, "Monitoring oI OSNR
by using a Mach-Zehnder interIerometer," Microwave Opt. Technol. Lett., vol.
30, pp. 63-65, 2001.
|31| J. H. Lee, D. K. Jung, C. H. Kim, and Y. C. Chung, "OSNR monitoring
technique using polarization-nulling method," IEEE Photon. Technol. Lett.,
vol. 13, pp. 88-90, 2001.
|32| S. X. Wang, A. M. Weiner, M. Boroditsky, and M. Brodsky, "Monitoring
PMD-induced penalty and other system perIormance metrics via a high-speed
spectral polarimeter," IEEE Photon. Technol. Lett., vol. 18, pp. 1753-1755,
2006.
|33| H. Ji, K. Park, J. Lee, H. Chung, E. Son, K. Han, S. Jun, and Y. Chung,
"Optical perIormance monitoring techniques based on pilot tones Ior WDM
network applications," J. Opt. Netw., vol. 3, pp. 510-533, 2004.
12
|34| A. Amrani, G. Junyent, J. Prat, J. Comellas, I. Ramdani, V. Sales, J. Roldan,
and A. RaIel, "PerIormance monitor Ior all-optical networks based on
homodyne spectroscopy," IEEE Photon. Technol. Lett., vol. 12, pp. 1564-
1566, 2000.
|35| I. Shake, E. Otani, H. Takara, K. A.-U. Uchiyama, K., Y. A.-Y. Yamabayashi,
Y., and T. A.-M. Morioka, T., "Bit rate Ilexible quality monitoring oI 10 to
160 Gbit/s optical signals based on optical sampling technique," Electron.
Lett., vol. 36, pp. 2087-2088, 2000.
|36| M. Chen, Y. Zhang, L. He, Z. Si, S. Yang, J. Sun, Y. Zhang, H. Chen, and S.
Xie, "Simultaneous monitoring method Ior chromatic dispersion and
polarization mode dispersion based on polarization modulation," J. Optics A.
Pure Appl. Opt., pp. 320, 2007.
|37| G.-W. Lu, M.-H. Cheung, L.-K. Chen, and C.-K. Chan, "Simultaneous PMD
and OSNR monitoring by enhanced RF spectral dip analysis assisted with a
local large-DGD element," IEEE Photon. Technol. Lett., vol. 17, pp. 2790-
2792, 2005.
|38| K.-P. Ho and J. M. Kahn, "Methods Ior crosstalk measurement and reduction
in dense WDM systems," J. Lightw. Technol., vol. 14, pp. 1127-1135, 1996.
|39| C. Dorrer, "Monitoring oI optical signals Irom constellation diagrams
measured with linear optical sampling," J. Lightw. Technol., vol. 24, pp. 313-
321, 2006.
|40| M. P. Dlubek, S. Kaunga-Nyirenda, A. J. Phillips, S. Sujecki and E. C.
Larkins, "Experimental veriIication oI the existence oI optically induced
carrier pulsations in SOAs," to be submitted for publication
|41| M. P. Dlubek, A. J. Phillips, and E. C. Larkins, "Extinction ratio improvement
using nonlinear Iour-wave mixing in SOAs with assist beam," Microwave Opt.
Technol. Lett., vol. 50, pp. 2079-2083, 2008.
|42| M. Dlubek, A. Phillips, and E. Larkins, "Optical signal quality metric based on
statistical moments and Laguerre expansion," Opt. Quantum Electron., vol.
40, pp. 561-575, 2008.
|43| M. P. Dlubek, A. J. Phillips, and E. C. Larkins, "Simple and Cost-EIIective
OSNR Monitoring Method based on Modulation Spectrum Assessment,"
Europ. Conf. Netw. Opt. Commun. NOC 2007, vol. 12, pp. 168 - 175, 2007.
13
|44| M. P. Dlubek, A. J. Phillips, and E. C. Larkins, "Method Ior OSNR
Monitoring Based on Modulation Spectrum Assessment," accepted for
publication in IET Optoelectron.
|45| M. P. Dlubek, A. J. Phillips, and E. C. Larkins, "Nonlinear Evolution oI
Gaussian ASE Noise in ZMNL Fiber," J. Lightw. Technol., vol. 26, pp. 891-
898, 2008.
|46| M. P. Dlubek, A. J. Phillips, and E. C. Larkins, "Evolution oI the probability
density Iunctions oI Gaussian ASE noise in zero-memory nonlinear Iiber,"
accepted for publication in Opt. Fiber Technol.

14
2. Four-wave mixing in semiconductor optical amplifiers
Semiconductor optical ampliIiers (SOAs) are considered enabling devices Ior
a wide range oI photonic signal conversion and processing techniques |1, 2| that also
Iind use in the Iield oI optical perIormance monitoring (OPM) (see Chapter 1),
especially Ior optical sampling. Optical sampling, as described in Chapter 3, is
becoming an increasingly important tool Ior ultraIast optical signal characterization
|3, 4|. Among diIIerent possible methods to implement optical sampling, techniques
using SOAs are considered to be some oI the most promising |4|. ThereIore, within
the scope oI this thesis, the Ieasibility oI an optical sampling technique based on Iour
wave mixing (FWM) in SOAs is experimentally investigated and the main results are
reported in Chapter 6. BeIore the experimental results are discussed, it is necessary to
brieIly review the basic physics oI SOAs. Particular emphasis is given to the
nonlinear characteristics oI these devices, as these nonlinearities are the key Ieatures
enabling optical sampling. Among diIIerent nonlinear eIIects observed in SOAs,
FWM will be explained in the most detail, as this process was used in the experiments
presented in Chapter 6 to obtain optical samples. Based on the general SOA model, a
simple time-domain FWM model is described, which is capable oI explaining
nonlinear mixing between optical pulses.
2.1. Basic physics of the SOAs
SOAs are basically semiconductor lasers, pumped (usually electrically) at such
levels that there is not enough gain to achieve the lasing threshold condition |1, 5|.
ThereIore, the basic material physics oI SOAs is the same as that oI a semiconductor
laser. To obtain an appreciable optical gain without IulIilling the lasing condition,
output Iacet reIlectivities are signiIicantly reduced |1, 5| by using antireIlection
coatings, tilted active region or window regions |1|. II the reIlections are not
negligible, the gain spectrum oI the SOA displays residual cavity-imposed
wavelength selectivity due to the residual Fabry-Perot resonances. II the reIlectivity oI
both output Iacets is negligible, the gain spectrum is determined by the material gain
|1, 5, 6|. This type oI SOA is called a travelling wave ampliIier and is generally
preIerred Ior most applications. In the Iollowing, it is always assumed that the Iacet
reIlectivities are negligible; so that the SOA is a travelling wave device.
15
2.1.1. Optical gain in semiconductors
To obtain an ampliIier, a gain medium is required. As mentioned above, the
material physics oI the SOA is the same as that oI the semiconductor laser. Optical
gain in an SOA is achieved only when carrier population inversion exists |5, 6|. This
condition is usually produced by using current injection (electrical pumping). Carrier
population inversion means that more carriers (electrons) are present at the excited
energy level (conduction band) than at the ground energy level (valence band). An
incoming external photon can then be either absorbed by the carrier at the lower
energy level, or can stimulate the excited carrier to release a quantum oI its energy (a
new photon) and to return to the ground level. The Iirst process is called stimulated
absorption, the latter is stimulated emission. For perIect two level systems (no band
broadening) the necessary condition to provide appreciable gain is that the carrier
excitation energy is precisely equal to the photon energy, as the energy conservation
needs to be maintained. In semiconductors, the energy levels are smeared into
relatively broad bands. ThereIore, the allowed range oI Irequencies (energies) is quite
large. This also directly implies quite a wide gain spectrum. Additionally, the
momentum conservation principle requires that the wave vector be conserved in the
absorption/emission process. This explains why direct bandgap semiconductors are
required Ior lasers and ampliIiers |5, 6|. Fig. 2.1 shows schematically how the
transition processes occur. Note that in reality the band structure oI a semiconductor
medium is more complicated |6|.

Fig. 2.1 Schematic representation oI the process oI stimulated emission (a) and absorption (b)
As a result oI the process oI stimulated emission, new photons are Iully
coherent with the coercing photon (i.e. they have the same Irequency, phase,
wavevector and polarization). These two processes oI emission and absorption are
16
clearly competing and their total probabilities oI occurrence are proportional to the
number oI state pairs (electron-hole pairs) available Ior each direction oI transition
(i.e. Irom the conduction band to the valence band or Irom the valence band to the
conduction band). Since the transitions occur only between occupied initial states and
vacated Iinal states, stimulated emission prevails only iI population inversion is
maintained, i.e. iI more electrons occupy the relevant states in the conduction band
than in the valence band |5, 6|. When the occupation oI the contributing states in each
band is equal, the medium is optically transparent, as the gain exactly counterbalances
the absorption. However, since the carrier reservoir in the conduction band is not
inIinite, injected photons deplete the number oI available carriers thereby reducing the
gain. In the limit oI very large input optical power, the gain approaches zero. This
coupling between photon and carrier populations is detailed in the next section.
Besides the energy and momentum conservation principles, quantum selection
rules become increasingly important in determining the allowed transitions as the
dimensionality oI the active region decreases |6|. When the dimensions oI the
semiconductor structure are comparable with the de Broglie wavelength oI the
electrons, quantum eIIects are clearly visible and cannot be neglected |6|. Most oIten
in practical applications, the width oI the active region is reduced, Iorming a quantum
well. In the quantum well the density oI states can no longer be approximated by a
smooth, parabolic Iunction, but instead is given by a staircase-like Iunction, due to
discrete allowed energy levels. When a stack oI such wells is used in an SOA (e.g. to
increase the gain), a multiple quantum well (MQW) ampliIier is Iormed. Such
ampliIiers possess a number oI advantageous Ieatures, such as high gain, wide
bandwidth, lower pumping current and, with proper design, low polarization
sensitivity |1, 6|.
As Iollows Irom the discussion above, the exact calculation oI the gain is quite
involved. The rigorous approach requires taking into account the shape oI the
interacting bands and the density oI states, occupation probability within each band
(usually assumed to be described by the Fermi-Dirac distribution) and the transition
rates, describing the strength oI interaction between each pair oI allowed starting and
Iinal energy levels |6|. Such an approach is oIten cumbersome and in many cases a
simple phenomenological gain model is used, which assumes a linear dependence oI
the gain g on the carrier density N |5, 6|:
17
) ( ) (
0
N N a N g I = (2.1)
where I is the conIinement Iactor (taking into account mode spreading outside the
gain region), a is the (constant) diIIerential gain coeIIicient and N
0
is the carrier
density at transparency. This phenomenological model shows reasonably good
accuracy and is easily understandable in terms oI the stimulated emission and
absorption processes described above. The degree oI population inversion is
proportional to the rate oI the pumping so that the gain increases with pumping above
the transparency point. For these reasons this linear gain model is widely used in the
literature |5, 7-9|. In the SOA and FWM models presented in this thesis, this gain
dependency is always assumed.
2.1.2. Amplified spontaneous emission noise
In addition to the processes oI stimulated emission and absorption described
above, any active ampliIying medium also contributes optical noise through the
process oI ampliIied spontaneous emission. An electron in the excited state can
spontaneously recombine with a hole without an external photon participating in the
transition. As a result, a new photon with random Irequency, phase and wave vector is
created. As these spontaneously generated photons propagate through the SOA, they
are ampliIied. The resulting output optical noise is known as ampliIied spontaneous
emission (ASE). Population inversion is not a necessary condition Ior spontaneous
photons to be generated. In Iact, as the degree oI the inversion decreases, ASE
becomes more signiIicant compared to the coherent photons produced in the
stimulated emission process. This is quantiIied by the noise Iigure F
n
oI the ampliIier.
The noise Iigure is approximately equal to twice the population inversion Iactor n
sp

|5|. The population inversion Iactor in SOAs is given by
0
N N
N
n
sp

= (2.2)
so it improves with increasing the pumping current |5|. An ideal ampliIier has n
sp
1,
which means that the output signal to noise ratio (SNR) is reduced by 3 dB compared
to the input SNR, as F
n
2 in this case. SOAs have relatively large F
n
, typically in the
region oI 4-8 |1, 5|. In Iibre applications this noise Iigure is additionally made worse
by the large Iibre-SOA coupling losses, resulting Irom the severe mode mismatch.
The spectrum oI the ASE noise is determined by the ampliIier bandwidth and
thereIore is usually very broad. Since it can be considered a white noise over
18
reasonably wide bandwidths, the ASE spectral density S
sp
(f) and the ASE power P
ASE

in the bandwidth B
0
are given respectively by
hf G n f S
sp sp
) 1 ( 2 ) ( = (2.3)
0
) ( B f S P
sp ASE
= (2.4)
where G is the ampliIier gain and the Iactor oI 2 in (2.3) accounts Ior the two
orthogonal polarization components oI the ASE.
As the ASE photons are produced in a random process, they need to be treated
using statistical tools. Quantum mechanical theories describe ASE generation as a
birth-death-immigration (BDI) process and, as a result, the total number oI photons
generated in a linear ampliIier is given by the noncentral negative binomial
distribution |10|. This Iormalism is however unwieldy and usually the random Iield at
the ampliIier output is assumed to be described by the Gaussian random process. This
is the so-called additive Gaussian noise model. For any optical ampliIier oI a
reasonable gain, the additive Gaussian and the BDI models result in essentially
identical predictions |10|.
2.1.3. Rate equation and propagation equation
The brieI discussion presented above shows that the Iull description oI the
operation oI SOAs requires taking into account the evolution oI both the photon and
carrier populations, as well as the coupling between them. Such a description is given
by the rate equation Ior carriers and, equivalently, the propagation equation Ior the
optical Iield or photon density. Since the phenomenon oI Iour-wave mixing (FWM),
explained later in this chapter, is most easily explained in terms oI interacting optical
Iields, this convention will be adopted here.
The rigorous derivation oI carrier population evolution is complicated and is
based either on semiclassical density matrix equations or on semiconductor Bloch
equations |6, 11, 12|. Since the derivation oI the carrier rate equation is lengthy and
can be Iound in the literature (see Ior example |6, 12|), only a phenomenological
justiIication Ior this rate equation is given here. There are several processes
inIluencing the carrier density. First oI all, electrical pumping provides new carriers
and thereIore tends to increase the carrier density to the limit given by the current
density. On the other hand, the negative change in the carrier density is due to the
process oI stimulated emission and various undesirable spontaneous processes (Ior
19
example Auger recombination or spontaneous emission). ThereIore, the diIIerential
equation Ior the carrier density N can be written as |5, 6|
) ( ) ( N R N R
qJ
I
dt
dN
st sp
= (2.5)
where I is the pumping current and J is the active volume. In (2.5), current leakage
has been neglected (although it could be implicitly included in the Iirst term on the
right hand side oI the equation), as has the carrier diIIusion, which is usually deemed
negligible Ior typical SOA dimensions |5|. The spontaneous recombination term
R
sp
(N) contains contributions due to radiative spontaneous recombination and non-
radiative eIIects like Auger recombination and recombinations through crystal deIects
and surIace states. As the non-radiative processes do not couple directly with the
propagating optical Iield, it is easiest to absorb them approximately in the threshold
density N
0
in (2.1). Alternatively, R
sp
(N) can be expressed as a third order polynomial
in N (
0
3 2
/ ) ( t N CN BN AN N R
sp
= + + = where the powers oI N are related to the
number oI carriers participating in each recombination process) and this
representation serves as a deIinition oI the spontaneous carrier liIetime t
0
|5|. This
approach simpliIies and linearizes (2.5). The change in N due to stimulated emission
is proportional to the optical gain and to the number oI input photons, which is in turn
proportional to the total optical power P
0
normalized to the photon power |5, 6|
v h
P
N g N R
st
0
) ( ) ( = (2.6)
The Iinal Iorm oI the carrier rate equation, which will be used later on in this thesis, is
v t h
P
N N a
N
qJ
I
dt
dN
0
0
0
) ( = . (2.7)
The propagation oI the electromagnetic Iield is governed by Maxwell`s
equations. Using standard mathematical procedure |6|, these can be simpliIied to
yield the wave equation
t
P
c t
E
c
n
E
2
2
2
0
2
2
2
2
2
1
c
c
=
c
c
V
c
(2.8)
where n is the reIractive index, c is the speed oI light and c
0
is the vacuum
permittivity. P is the polarization oI the medium, which is induced by the input light
and can be expanded into linear and nonlinear parts. P and E are linked through the
20
susceptibility oI the medium _. In general, a rigorous evaluation oI _ is quite diIIicult,
as it requires a quantum mechanical approach, similarly to the evaluation oI gain |7,
8|. Since linear dependence oI the gain g(N) on N is already assumed, and the gain is
the imaginary part oI the susceptibility, a similar Iorm Ior _ is assumed:
) ( ) ( ) (
0
0
N N a f
nc
N + = |
e
_ (2.9)
where is the linewidth enhancement Iactor, which stems Irom the dependence oI the
reIractive index in a semiconductor medium on N |5|. Note that this approach can
only model the interband nonlinear phenomena described by the rate equation (2.7).
There is also a class oI important intraband nonlinear phenomena, which are not
included in (2.9) and a brieI discussion oI all SOA nonlinearities relevant to FWM
(and hence to optical sampling) is given later. Eq. (2.8) can be Iurther simpliIied by
assuming that the optical signal is narrowband with respect to the optical carrier
Irequency, i.e. that it can be represented using the carrier-envelope representation:
) exp( ) , ( ) , (
0
t f t : A t : E e = (2.10)
where it is also assumed that there is no beam diIIraction, due to waveguiding. A(:, t)
is the complex signal amplitude. To obtain the propagation equation, (2.10) is
substituted into (2.8). Now, the terms containing second derivatives, resulting Irom
this substitution, are neglected by invoking the slowly varying envelope
approximation (hence the assumption about the narrowband character oI the signal),
as they are assumed to be negligibly small |7, 8|. This results in a Iirst order partial
diIIerential equation, which is Iurther simpliIied by using the moving reIerence Irame
transIormation t
c
t-:n/c |7|. This gives the Iinal propagation equation Ior the
complex amplitude:
A N g f
:
A
) ( ) 1 (
2
1
| =
c
c
. (2.11)
For convenience, the complex amplitude is normalized such that
) , ( ) , (
0
t : P t : A = . Equations (2.7) and (2.11) with the linking relation (2.1)
constitute the basic SOA model, which is widely used in the literature due to its
simplicity |5, 7, 12|. They also Iorm the basis oI many models describing FWM in
SOAs (see Ior example |8, 13, 14|). In many cases it is convenient to transIorm (2.7)
into a rate equation Ior gain, by using (2.1) |7|. Then, aIter integrating the resulting
21
equation over the ampliIier length L, i.e. by introducing the integrated gain
}
=
L
d: t : g t h
0
) , ( ) ( |7|, (2.11) can be solved. This provides a simple semi-analytical
model oI the propagation in SOAs, given by the system oI equations |7|:
) 1 ) (exp(
) , 0 (
0
0
0

= h
P
P h h
dt
dh
sat
t
t
(2.12)
)) 1 (
2
) (
exp( ) , 0 ( ) , ( | f
t h
t A t L A = . (2.13)
where h
0
is the integrated gain without any external light injection and P
sat
is the
saturation power oI the ampliIier. This system was used in one oI the Iirst
investigations oI the FWM between short pulses in SOAs |15| and thereIore is oI
interest here. With some extensions |16, 17|, a similar model can be used to describe
FWM between subpicosecond pulses. The main limitations oI (2.12) and (2.13) stem
Irom the assumed phenomenological Iorm oI susceptibility _, which turns out to
neglect some important eIIects mediating FWM in SOAs. The discussion oI relevant
nonlinear phenomena in SOAs is given in the next section.
2.2. Nonlinear effects in SOAs
BeIore the invention oI erbium doped Iibre ampliIiers (EDFAs) changed the
world oI optical communications, the SOA was expected to become one oI the most
important building blocks in optical networks |18, 19|. However, EDFAs exhibited
superior practical Ieatures compared to SOAs, among them larger gain, better noise
Iigure and much better linearity. SOAs can never really be considered linear
ampliIiers, due to the Iorm oI rate equation (2.7), which has a signiIicant dependence
on the optical power. Nevertheless, it is the same relatively large nonlinearity that
makes them unsuitable Ior linear ampliIication, which makes it possible to use SOAs
Ior photonic signal processing and, speciIically, Ior optical sampling. This section
brieIly discusses nonlinear eIIects in SOAs with the most attention being given to the
phenomenon oI Iour-wave mixing.
2.2.1. Third order nonlinear processes in SOAs
The nonlinear response oI a medium to electromagnetic stimulation is usually
described in terms oI nonlinear polarization. It is most convenient to expand the
polarization P in a power series in electric Iield E |20|, where the Iirst term is
22
responsible Ior linear polarization and higher terms describe nonlinearities oI
appropriate order
...
) 3 (
0
) 2 (
0
) 1 (
0
+ + + = EEE EE E P _ c _ c _ c (2.14)
Since E and P are vectorial quantities, each oI the susceptibilities _
(n)
is a tensor oI
rank n1. As the coupled equations (2.7) and (2.11) suggest, SOAs exhibit an
(eIIective) third order nonlinearity. This stems Irom the dependence oI the carrier
density (and hence the reIractive index) on the optical power. Since the Iield evolution
equation (2.11) is linearly dependent on carrier density (within the gain approximation
adopted here), it is clear that this dependence may lead to a situation where three
waves interact to produce a Iourth one. Although, in the case oI SOAs, it is usually
assumed that this interaction is due to an intrinsic third order nonlinearity, it will be
discussed in Section 5.3 that an alternative description in terms oI cascaded second
order nonlinear processes is also possible. As discussed in Section 5.3, some
experimental evidence suggests that the cascaded second order nonlinearities may
indeed be present in SOAs |21|. However, as these processes mimic the third order
eIIects in most cases, they contribute to the eIIective third order nonlinearity and can
be described within the same phenomenological Iramework (2.7), (2.11) |21|.
It should be noted that there could be as many as three distinguishable initial
waves participating in this interaction or it could be simply one initial wave,
contributing three photons to the interaction. These photons may be created or
annihilated, as long as energy and momentum conservation are obeyed |20|. Fig. 2.2
schematically shows the quantum mechanical representation oI one possible
interaction. As will be explained later, this transition sequence can describe Ior
example FWM photons i, f are absorbed, whereas photons k, l are emitted. For
cascaded second order processes, this sequence actually consists oI two separate
processes, which may be non-local and not simultaneous |20, 21| and each oI the
contributing interactions may involve a THz photon, as shown in Fig. 2.3. In Fig. 2.3,
SFG and DFG denote sum- and diIIerence Irequency generation, respectively. This
issue is discussed in more detail in Chapter 5. In semiconductors, Ior photons oI
energies within the gain spectral region, the energy states ,g), ,a), ,b), ,u) in Fig. 2.2 are
located within the allowed energy bands, as considerations here are limited to electron
plasma and nonlinearity due to bound electrons is neglected.
23

Fig. 2.2 Quantum transition sequence Ior FWM

Fig. 2.3 Cascaded second order processes contributing to eIIective third order nonlinearity
2.2.2. Interband nonlinear processes in SOAs
Interband processes involve carrier transitions between the valence and
conduction bands and thereIore can demonstrate themselves as detectable electrical
current. The dynamics oI these processes are described by (2.7). The stimulated
recombination term in (2.7) reveals the Iact that the ampliIication oI the input signal
depletes the carrier population. This process subsequently changes the gain and
reIractive index oI the SOA, as evident Irom (2.1) and (2.9). There are several Ilavors
oI this basic mechanism. Firstly, the noncoherent case will be considered, when
optical power due to one beam simply depletes the carrier population. The resulting
change oI the propagation properties is reIlected in the optical signal evolution, as
described by (2.13) and the eIIect oI selI-modiIication oI the SOA gain is known as
selI-gain modulation (SGM). Analogously, the eIIect oI modiIying the phase is
24
known as selI-phase modulation (SPM). Alternatively, a second beam co-propagating
with the strong beam inIluencing the medium experiences essentially the same gain
and reIractive index changes, as both beams couple to the same carrier population.
These eIIects are known as cross-gain modulation (XGM) and cross-phase
modulation (XPM) and their applicability to all-optical signal processing,
regeneration and wavelength conversion has been widely studied |1-3, 22|.
When at least two distinguishable waves are input to the SOA, coherent
eIIects also take place as (2.7) is dependent on the optical Iield. II the two waves,
represented as in (2.10) have diIIerent carrier Irequencies e
1
, e
2
and the same state oI
polarization (SOP), the squaring operation in (2.7) (represented by the dependence oI
the carrier density on the optical power) is responsible Ior mixing between them, as it
introduces nonlinearity. The mixing term at the diIIerence Irequency (e
2
-e
1
) is
usually assumed to cause carrier density pulsations (CDP) |2, 3, 8, 9, 13, 17, 19|,
which Iorm gain and index gratings in the SOA. Based on this reasoning, the solution
oI (2.7) oI the Iorm
) ) ( exp( ) (
1 2
t f N N t N e e A + = (2.15)
is typically assumed |8, 9|, where N is the static carrier density and N A is the small
modulation term.
The input signal is then scattered oII these dynamic gratings, creating new
Irequency components e
1
-(e
2
-e
1
), e
1
(e
2
-e
1
), e
2
-(e
2
-e
1
), e
2
(e
2
-e
1
). The process
is analogous to creating sidebands about a modulated carrier. The eIIiciency oI the
gratings versus the detuning oI the input Irequencies depends upon both the carrier
liIetime t
0
and on the carrier replenishment rate, limited by the pumping term in (2.7).
Among all interband eIIects, this one is oI greatest interest within this thesis, as this is
one oI the processes assumed to mediate FWM, particularly Ior small detunings |8, 9,
13|. In Chapter 5, the assumption oI the existence oI CDP due to wave beating is re-
examined and some experimental data conIirming this model is reported. (This data
also suggests the presence oI cascaded second order nonlinearities in what is
externally seen as a third order process)
2.2.3. Intraband nonlinear processes in SOAs
There exists a large body oI experimental data, which shows FWM in SOAs
Ior detunings at which CDP-induced gratings are no longer eIIicient, due to limited
25
carrier replenishment rate. This suggests the presence oI other nonlinear eIIects, not
included in (2.7), which eIIiciently mediate FWM at large detunings and thereIore
have much shorter characteristic times. These are called intraband eIIects, as they are
not caused by the carrier dynamics between valence and conduction bands. Instead,
they are caused by transitions between diIIerent energy levels within one band. A
more rigorous quantum mechanical approach shows that gain depends on the
occupation oI these levels through diIIerent density oI states and transition rates |6|.
In thermal equilibrium, the probability oI occupation oI an energy level within each
band is given by the Fermi-Dirac distribution |6, 9|. Stimulated recombination
disturbs this equilibrium but strong carrier-carrier and carrier-phonon scattering tends
to restore it on a very short time scale |9|. Beating between the input waves causes
dynamic changes in the occupation probability oI the states in the conduction band,
which can be treated as a Iormation oI the gain and index gratings (a static
contribution to the occupation probability due to each input wave also exists and
maniIests itselI as a nonlinear gain compression |8, 16|, which similarly is not
included in (2.12) and (2.13)). The short intraband relaxation times imply both
relatively low eIIiciency oI the intraband processes and their very large bandwidth.
ThereIore, they become dominant Ior detuning well above the CDP cutoII Irequency.
There are two main kinds oI intraband processes typically considered in the
literature spectral hole burning (SHB) and carrier heating (CH). SHB is typically
pictured as a pulsating 'hole in the carrier distribution, which recovers to the Fermi-
Dirac distribution on a time scale set by the carrier-carrier scattering (t
SHB
50 Is) |8,
9, 13|. CH is caused by the pulsating average carrier temperature relative to the lattice
temperature, due to depletion oI the low energy carriers in the process oI
ampliIication as well as Iree carrier absorption. The response oI CH is set by the
carrier-phonon relaxation time (t
CH
700 Is) |9, 13|.
CH and SHB, together with CDP, are the main processes considered
responsible Ior FWM in SOAs. Rigorous modeling oI SHB and CH requires the
density matrix approach |8, 9, 16|, which is Iairly complicated. ThereIore, most
models use a phenomenological approach, which assumes the same Iorm oI SHB and
CH contributions to FWM as the CDP contribution, with appropriately modiIied
strength and corner Irequencies |13-15, 17|. This approach will also be used in this
thesis.
26
It is worth mentioning that there are other eIIects, which some authors
consider to take a signiIicant part in FWM, like ultraIast nonlinear reIraction |23| or
two photon absorption |23, 24| but there is no general consensus on their practical
importance. ThereIore, they will not be included in the Iorthcoming discussion on the
modeling oI FWM in SOAs.
2.3. Four wave mixing in SOAs
In the previous paragraphs, the physics oI SOAs and FWM in SOAs was
brieIly reviewed. We are now in a position to present the model oI FWM used in this
thesis. This model is later compared in Chapter 6 with experimental results on optical
sampling. There are two main approaches used to model FWM in SOAs. One (more
rigorous) approach is based on density matrix theory |8, 9, 16|. The second approach
is phenomenological, whereby most oI the mathematical diIIiculties are avoided by
assuming the Iorm oI intraband contributions |8, 13-15, 17|. The model adopted here
is phenomenological |15, 17|. FWM models also diIIer in the Iact that some oI them
are Iormulated in the Irequency domain |8, 9, 13, 14| and some in the time domain
|15-17|. Obviously, to analyze optical sampling, a time domain model is much more
useIul. ThereIore, this approach has been chosen in this thesis.
In most schemes, FWM takes place between two input waves. One wave (e
1
)
is usually much stronger and serves as a pump. The other wave (e
2
) is weaker and
carries the inIormation oI interest (i.e. it is the input signal). As a result oI this
asymmetry, only one mixing product (the so-called conjugate wave) is oI appreciable
power, and the output spectrum shows three waves, with the conjugate and the signal
positioned symmetrically on both sides oI the pump Irequency e
1
. Fig. 2.4 presents
schematically the output spectrum Irom the SOA.

Fig. 2.4 Schematic FWM spectrum aIter an SOA
27
The basic mixing scheme presented in Fig. 2.4 suIIers Irom polarization
dependency and non-Ilat mixing eIIiciency with respect to input waves detuning. To
alleviate these problems, more advanced mixing schemes have been proposed based
on two pumps |25, 26| or using the so called assist beam (AB) to improve the
conversion eIIiciency and linearity |27, 28|. In Chapter 5, an application oI the AB
method is presented, which allows obtaining relatively simple all-optical signal
regeneration. Experimental results showing substantial extinction ratio improvement
in this scheme are presented.
2.3.1. Time domain model of FWM in SOAs
Within the scope oI this thesis, the application oI FWM Ior optical sampling is
investigated (see Chapter 6). ThereIore, an appropriate numerical model oI the
underlying physics is useIul to Iully understand the experimental behavior. The
sampling process by its very nature is a time domain process. ThereIore a time
domain model oI FWM is oI most interest. Additionally, it is straightIorward to apply
a time domain model to model quasi-CW conditions by appropriately increasing the
duration oI the propagating optical pulses.
The Iirst models oI FWM in SOAs were usually limited to the Irequency
domain |8, 9, 13| and most oI them were based on or related to the model presented in
Agrawal`s seminal paper |8|. As this model assumes the steady state solution oI the
carrier rate equation (2.7), it is not appropriate Ior modeling time-dependent
interactions. However, another paper by Agrawal and Olsson |7| provides a useIul
and still very popular model Ior pulse evolution in SOAs, given by (2.12) and (2.13),
which then served as the basis Ior several time domain FWM models |15, 17|. It is
also worth noting that Agrawal`s nearly-degenerate FWM model can be considered as
a Irequency-domain, small signal analysis oI the model given by (2.12), (2.13) with
the assumption (2.15). This approach has been Iurther extended as described below.

Shtaif-Eisenstein model
ShtaiI and Eisenstein were among the Iirst researchers who presented a time
domain model capable oI predicting FWM between short optical pulses |15|. This
model is based on the small signal analysis oI the tandem equations (2.12-13). The
total input Iield is assumed to be oI the Iorm
28
t f
e t A t A t A
O
+ = ) , 0 ( ) , 0 ( ) , 0 (
2 1
(2.16)
where O e
1
- e
2
is the detuning Irequency between the pump and the signal. A
similar small signal solution oI the gain equation (2.12) is also assumed:
t f t f
e h he t h t h
O O
A + A + =
*
) ( ) ( . (2.17)
When (2.17) is substituted to (2.12), the equation Ior time dependent saturated
integrated gain ) (t h is Iound, as well as the two terms corresponding to CDP at O and
O, respectively:
sat
sat
h
h
P
A A
f P A e
e
h
2
*
1
0
2
1
, , 1
1
t O + +

= A (2.18 a)
sat
sat
h
h
P
A A
f P A e
e
h
*
2 1
0
2
1
*
, , 1
1
t O +

= A (b)
where P
sat
is the saturation power. To arrive at (2.18), terms oscillating at the same
Irequency, resulting Irom the substitution oI (2.17) to (2.12) are separated, and the
Iollowing small signal expansion is used: + ~ A
O
1 ) exp(
t f
he
t f
he
O
A . Substitution oI
(2.16) and (2.18) into the propagation equation (2.13) leads, upon the use oI the same
expansion as above, to the system oI three coupled output waves. These are given by:
) , 0 (
, ,
, , 1
1
2
1 ) , (
1
2
2
0
2
1
) (
) 1 ) (
) 1 )( ( 2 / 1
1
t A
P
A
f P A e
f e
e t L A
sat
sat
t h
t h
f t h
(
(

O +

=

t
|
|
(2.19 a)
) , 0 (
, ,
, , 1
1
2
1 ) , (
2
2
1
0
2
1
) (
) 1 ) (
) 1 )( ( 2 / 1
2
t A
P
A
f P A e
f e
e t L A
sat
sat
t h
t h
f t h
(
(

O +

=

t
|
|
(b)
) , 0 (
, , 1
1
2
0 ) , (
*
2
2
1
0
2
1
) (
) 1 ) (
) 1 )( ( 2 / 1
3
t A
P
A
f P A e
f e
e t L A
sat
sat
t h
t h
f t h
(
(

O +

=

t
|
|
(c)
where the zero term in (2.19c) signiIies the Iact that the conjugate is generated within
the SOA and no wave at this Irequency is initially input into the ampliIier. The newly
generated wave is a complex conjugate oI the input signal wave, hence the name oI
this mixing product. The spectrum oI this product is a reIlected about the carrier
Irequency version oI the signal spectrum - a Ieature which can be used, Ior example,
Ior dispersion or nonlinearity compensation in optical Iibre links |29|. The model
given by (2.19 a-c) takes into account FWM due to CDP only, which in practical
29
cases is oI low magnitude, since short pulses have wide spectrum, so that the detuning
O is necessarily quite large. Experimental observation shows the importance oI
intraband eIIects |13, 30| Ior correct modeling oI FWM eIIiciency. More rigorous
analysis based on the density matrix approach shows that contributions to FWM due
to intraband eIIects are oI a Iorm similar to (2.18) |9|. ThereIore, the terms
responsible Ior SHB and CH are phenomenologically introduced and the Iinal mixing
Iunction F(O) is given by:
(
(

O +

+
O +

+
O + +

= O
SHB
SHB SHB
CH
CH CH
sat
h
h
f
f
f
f
f P A e
f e
F
t
c |
t
c |
t
|
1
) 1 (
1
) 1 (
, , 1
1
2
1
) (
0
2
1
(2.20)
where each intraband process has its own linewidth enhancement Iactor
CH, SHB
, time
constant t
CH, SHB
andc
CH, SHB
describes the strength oI each process |9, 15|. Although
this model takes into account the contribution oI the intraband processes to FWM, it
does not include the nonlinear gain compression caused by the deviation oI the carrier
distribution Irom a quasi-equilibrium distribution. ThereIore, it is not suIIicient to
describe the propagation oI pulses shorter than about 10 ps, where ultraIast saturation
becomes important |16|. The model oI Mecozzi-Mork extends it into the single-ps
regime.

Mecozzi-Mork model and its simplification
The model presented above overestimates the FWM eIIiciency Ior pulses oI
several ps duration because ultraIast nonlinear gain compression is neglected.
Mecozzi and Mork, starting Irom density matrix equations, obtained a system oI
coupled equations Ior local electron and hole densities, total carrier density and carrier
temperature (energy). They rigorously showed that this system is the minimal set
required to simultaneously treat the eIIects oI CDP, SHB and CH |16|. According to
the authors, this model can be used to describe the propagation oI pulses oI duration
down to about 100 Is. However, such short pulses are not very practical Ior use in
optical sampling, due to their extremely large bandwidth and very low power spectral
density. In the experiments reported in Chapter 6, the shortest pulses used were about
3 ps long. ThereIore, the Iull Mecozzi-Mork model is not required to describe the time
domain FWM in this case.
UseIul simpliIication oI this model was given in |17|, where the equations
describing local carrier density and carrier temperature were adiabatically eliminated
30
by assuming that the pulses are suIIiciently slow so that the local carrier densities and
carrier temperature Iollow input power variations and their only contribution is to
modiIy the carrier rate equation. In this way, the nonlinear gain compression is
included but the speed oI the local density recovery is overestimated. The modiIied
rate equation is
)
`


|
|
.
|

\
|
+

+
=
dt
t dP
h
P
h
P
t P h h
t P e dt
dh
sat sat
h
) , 0 (
) 1 ) (exp(
1
) 1 ) (exp(
) , 0 (
) , 0 ( 1
1
0
0
0
0
0
0
c
t
c
t c
(2.21)
where
SHB CH
c c c + = . The integrated gain given by (2.21) is then combined with
(2.19) and (2.20) to complete the model. As shown in |17|, this model compares
Iavorably with the rigorous and more complicated Mecozzi-Mork model, oIIering
similar simplicity to the basic ShtaiI-Eisenstein model. The validity oI this
approximation is limited to the cases when the duration oI the input pulses is longer
than the SHB and CH characteristic times, i.e. Ior pulses longer than about 1 ps. Since
this was the regime employed in the experiments reported in Chapter 6, this model
was used to compare with the experimental results.

31
2.4. References
|1| L. H. Spiekman, "Semiconductor optical ampliIiers," in Optical Fiber
Telecommunications, vol. IVA, pp. 699-731, I. Kaminov and T. Li, Eds.,
Academic Press, 2002.
|2| C. Schubert, R. Ludwig, and H.-G. Weber, "High-speed optical signal
processing using semiconductor optical ampliIiers," J. Opt. Fiber Commun.
Reports, vol. 2, pp. 171-208, 2005.
|3| L. A. Jiang, E. P. Ippen, U. Feiste, S. Diez, E. Hilliger, C. Schmidt, and H.-G.
Weber, "Sampling pulses with semiconductor optical ampliIiers," IEEE J.
Quantum Electron., vol. 37, pp. 118-126, 2001.
|4| C. Schmidt-Langhorst and H.-G. Weber, "Optical sampling techniques," J.
Opt. Fiber Commun. Reports, vol. 2, pp. 86-114, 2005.
|5| G. P. Agrawal and N. K. Dutta, Semiconductor lasers, 2 ed., Van Nostrand
Reinhold, 1993.
|6| T. Suhara, Semiconductor Laser Fundamentals, 1 ed., Marcel Dekker, Inc.,
2004.
|7| G. P. Agrawal and N. A. Olsson, "SelI-phase modulation and spectral
broadening oI optical pulses in semiconductor laser ampliIiers," IEEE J.
Quantum Electron., vol. 25, pp. 2297-2306, 1989.
|8| G. P. Agrawal, "Population pulsations and nondegenerate Iour-wave mixing in
semiconductor lasers and ampliIiers," J. Opt. Soc. Am. B, vol. 5, pp. 147,
1988.
|9| A. Uskov, J. Mork, and J. Mark, "Wave mixing in semiconductor laser
ampliIiers due to carrier heating and spectral-hole burning," IEEE J. Quantum
Electron., vol. 30, pp. 1769-1781, 1994.
|10| G. Einarsson, Principles of Lightwave Communications, Wiley, 1996.
|11| J. Mork and A. Mecozzi, "Theory oI the ultraIast optical response oI active
semiconductor waveguides," J. Opt. Soc. Am. B, vol. 13, pp. 1803, 1996.
|12| J. Wang and H. C. Schweizer, "A quantitative comparison oI the classical rate-
equation model with the carrier heating model on dynamics oI the quantum-
well laser: the role oI carrier energy relaxation, electron-hole interaction, and
Auger eIIect," IEEE J. Quantum Electron., vol. 33, pp. 1350-1359, 1997.
32
|13| K. Kikuchi, M. Kakui, C.-E. Zah, and T.-P. Lee, "Observation oI highly
nondegenerate Iour-wave mixing in 1.5 μm traveling-wave semiconductor
optical ampliIiers and estimation oI nonlinear gain coeIIicient," IEEE J.
Quantum Electron., vol. 28, pp. 151-156, 1992.
|14| M. A. SummerIield and R. S. Tucker, "Frequency-domain model oI multiwave
mixing in bulk semiconductor optical ampliIiers," IEEE J. Sel. Topics
Quantum Electron., vol. 5, pp. 839-850, 1999.
|15| M. ShtaiI and G. Eisenstein, "Analytical solution oI wave mixing between
short optical pulses in a semiconductor optical ampliIier," Appl. Phvs. Lett.,
vol. 66, pp. 1458-1460, 1995.
|16| A. Mecozzi and J. Mork, "Saturation eIIects in nondegenerate Iour-wave
mixing between short optical pulses in semiconductor laser ampliIiers," IEEE
J. Sel. Topics Quantum Electron., vol. 3, pp. 1190-1207, 1997.
|17| C. Xie, P. Ye, and J. Lin, "Four-wave mixing between short optical pulses in
semiconductor optical ampliIiers with the consideration oI Iast gain
saturation," IEEE Photon. Technol. Lett., vol. 11, pp. 560-562, 1999.
|18| M. J. O'Mahony, "Semiconductor laser optical ampliIiers Ior use in Iuture
Iiber systems," J. Lightw. Technol., vol. 6, pp. 531-544, 1988.
|19| R. M. Jopson and T. E. Darcie, "Semiconductor Laser AmpliIiers in High-Bit-
Rate and Wavelength-Division-Multiplexed Optical Communications
Systems," in Coherence, Amplification and Quantum Effects in Semiconductor
Lasers, pp. 323-366, Y. Yamamoto, Ed., Wiley-Interscience, 1991.
|20| M. Sheik-Bahae and M. P. Hasselbeck, "Third-order optical nonlinearities," in
OSA Handbook of Optics, vol. IV, pp. 17.1-17.34, M. Bass, Ed., McGraw-
Hill, 2001.
|21| R. Paiella, G. Hunziker, U. Koren, and K. J. Vahala, "Polarization-dependent
optical nonlinearities oI multiquantum-well laser ampliIiers," IEEE J. Sel.
Topics Quantum Electron., vol. 3, pp. 529-540, 1997.
|22| M. Usami and K. Nishimura, "Optical nonlinearities in semiconductor optical
ampliIier and electro-absorption modulator: their applications to all-optical
regeneration," J. Opt. Fiber Commun. Reports, vol. 2, pp. 497-529, 2005.
|23| J. M. Tang and K. A. Shore, "InIluence oI probe depletion and cross-gain
modulation on Iour-wave mixing oI picosecond optical pulses in
33
semiconductor optical ampliIiers," IEEE Photon. Technol. Lett., vol. 10, pp.
1563-1565, 1998.
|24| S. Scotti, L. Graziani, A. D'Ottavi, F. A. Martelli, A. Mecozzi, P. Spano, R.
Dall'Ara, F. Girardin, and G. Guekos, "EIIects oI ultraIast processes on
Irequency converters based on Iour-wave mixing in semiconductor optical
ampliIiers," IEEE J. Sel. Topics Quantum Electron., vol. 3, pp. 1156-1161,
1997.
|25| T. J. Morgan, J. P. R. Lacey, and R. S. Tucker, "Widely tunable Iour-wave
mixing in semiconductor optical ampliIiers with constant conversion
eIIiciency," IEEE Photon. Technol. Lett., vol. 10, pp. 1401-1403, 1998.
|26| I. Tomkos, I. Zacharopoulos, E. Roditi, D. Syvridis, and A. Uskov, "On the
polarisation sensitivity and perIormance optimisation oI a dual pump Iour
wave mixing based wavelength converter," Opt. Quantum Electron., vol. 32,
pp. 97-113, 2000.
|27| P.-M. Gong, J.-T. Hsieh, S.-L. Lee, and J. Wu, "Theoretical analysis oI
wavelength conversion based on Iour-wave mixing in light-holding SOAs,"
IEEE J. Quantum Electron., vol. 40, pp. 31-40, 2004.
|28| D.-Z. Hsu, S.-L. Lee, P.-M. Gong, Y.-M. Lin, S. S. W. Lee, and M. C. Yuang,
"High-eIIiciency wide-band SOA-based wavelength converters by using dual-
pumped Iour-wave mixing and an assist beam," IEEE Photon. Technol. Lett.,
vol. 16, pp. 1903-1905, 2004.
|29| A. Mecozzi, G. Contestabile, F. Martelli, L. Graziani, A. D'Ottavi, P. Spano,
R. Dall'Ara, J. Eckner, F. Girardin, and G. Guekos, "Optical spectral inversion
without Irequency shiIt by Iour-wave mixing using two pumps with
orthogonal polarization," IEEE Photon. Technol. Lett., vol. 10, pp. 355-357,
1998.
|30| J. Zhou, N. Park, J. W. Dawson, K. J. Vahala, M. A. Newkirk, and B. I.
Miller, "EIIiciency oI broadband Iour-wave mixing wavelength conversion
using semiconductor traveling-wave ampliIiers," IEEE Photon. Technol. Lett.,
vol. 6, pp. 50-52, 1994.


34
3. Optical sampling for OPM and signal
characterization
Optical sampling has been considered one oI the most promising methods Ior
optical perIormance monitoring and Ior time domain signal characterization. As
reviewed in Chapter 1, OPM is becoming increasingly important in optical networks
and is identiIied as one oI the key elements in Iuture dynamically reconIigurable
automatically switched optical networks (ASONs) |1, 2|. The monitoring techniques
based on optical sampling are potentially the most suitable candidates Ior perIorming
Layer 2 monitoring, as deIined by the OPM reIerence model |3|. Moreover, optical
waveIorm measurement with very high temporal resolution is considered one oI the
enablers Ior optical time-division multiplexing |4|. Although other methods exist (e.g.
optical autocorrelators), optical sampling is regarded as the most powerIul candidate
Ior this task |4|.
As described later in this chapter, optical sampling can be used Ior waveIorm
characterization or histogram reconstruction, which can be later used Ior qualitative or
approximate quantitative signal quality estimation. The qualitative signal monitoring
can include overall signal health inspection, determining iI any waveIorm distortions
characteristic to a particular signal degradation mechanism are visible, determining iI
the eye opening is enough Ior a good transmission and so on. Quantitative monitoring
involves estimating the BER oI the signal under test. This estimation is necessarily
approximate, as it has to be based on an assumed statistical model linking the data
with the BER. Most oIten, this connection exploits the histograms describing the
optical power distribution at the sampling instant. The most popular statistical model
assumes Gaussian statistics Ior the total noise and estimates the BER using the so-
called Q-Iactor |5, 6|. In Chapter 4, an introduction to the Q-Iactor method Ior signal
estimation is given and shortcomings oI this technique are discussed. In Chapter 7, a
novel optical quality metric is proposed, which has similar simplicity to the Q-Iactor
but has certain advantages. As explained in Chapter 1, the main advantages oI
estimating BER based on the histograms are cost-eIIectiveness, the possibility oI bit
rate transparency and the possibility oI early degradation detection. All oI these
Ieatures make optical sampling very attractive Ior use in the Iield oI optical
35
communications and the Iirst investigations oI its Ieasibility Ior this purpose date back
to mid 80`s |7| with an explosion oI research in recent years |8-11|.
However, it should be noted that most oI the aIorementioned capabilities can
be obtained using the more conventional electrical sampling. Clearly, it is possible to
Iirst detect the signal and sample it in the electrical domain to obtain the histograms.
In Iact, OPM methods based on electrical sampling have been simultaneously pursued
|12, 13| but they have several important limitations. First oI all, the bandwidth oI
electrical sampling oscilloscopes is limited to about 75 GHz |9, 14| and the high-end
devices are very expensive. This speed is hardly enough to sample a 40 Gb/s signal,
not to mention any Iuture even higher bitrates. (To measure the signal eye diagram, as
opposed to single shot measurements requiring sampling rate to be at least twice the
highest signal Irequency component, as the Nyquist theorem states.) Secondly, it is
increasingly diIIicult to match the impedance over such a wide bandwidth. ThereIore,
the response Iunction oI very high speed oscilloscopes oIten exhibits severe ringing
|10|. This very large electrical bandwidth also means very large thermal noise, hence
low sensitivity. ThereIore, the applicability oI electrical sampling Ior monitoring
signals above 10 Gb/s becomes questionable. On the other hand, optical sampling can
achieve a much larger bandwidth and is inherently Iree oI ringing. Also, thermal noise
can be signiIicantly reduced by using slow electronics. ThereIore, optical sampling
has greater potential and can also be more cost-eIIective than the ultraIast electrical
sampling, as the most expensive element is usually the sampling source (which can be
shared between many WDM channels in practical conIigurations) and the electronics
can be relatively slow. These advantages are obtained by shiIting the sampling
process Irom the electrical to the optical domain, i.e. the requirement Ior the high
bandwidth is transIerred Irom electronics to inherently Iaster optics. The next section
explains how these advantages are possible.
3.1. Principle of optical sampling
A generic block diagram oI an optical sampling system is shown in Fig. 3.1.
The system consists oI a sampling source (ultraIast sampling laser), optical sampling
gate, slow photodetector and post-processing electronics. Optionally, it may also
contain a clock recovery subsystem, possibly with clock division and Irequency
shiIting. II the clock recovery is used, the system perIorms synchronous sampling,
otherwise asynchronous sampling is obtained. The sampling gate Iunction f(t) is
36
generally proportional to the sampling pulse power P
p
(t), f(t) ~ P
p
(t) but the exact
relationship depends on the particular physical gating process (e.g. linear gating,
second order or third order nonlinearity) and on the medium response time and also
on auxiliary phenomena accompanying the sampling (e.g. data-sampling walk oII,
dispersion, nonlinear losses).

Fig. 3.1 Block diagram oI a generic optical sampling system
As mentioned above, it is possible to achieve very high bandwidth and
temporal resolution using optical sampling with a slow detector and postdetection
electronics. This may seem contradictory but is accomplished by utilizing very short
optical sampling pulses, which then lead to comparably short samples. Even though
the optical samples contain very high Irequencies, which are attenuated by the slow
photodetector, it is the average sample power which is recorded. At this point, it is
helpIul to recall the Iamiliar picture oI a delta excitation applied to a Iinite bandwidth
system: the system will smear the initial delta, but the peak oI the system response is
proportional to the height oI the input excitation. In the case oI optical sampling, Ior a
slow photodetector, the incident optical samples are eIIectively delta pulses.
ThereIore, as long as the bandwidth oI the system is independent oI the incident
energy, the inIormation about the sample power is directly included in the peak oI the
produced electrical pulse and can be easily calculated Irom the calibration curve. It is
straightIorward to express the above argument mathematically. The optical sample
power P
C
(t) is given by:
) ( ) ( ) (
0
2
t t f t P t P
S C
=q (3.1)
where P
S
(t) is the signal power and p is the sampling eIIiciency. The notation f(t-t
0
)
signiIies the Iact that the sampling pulses are strongly localized compared to the data.
The sampling eIIiciency is dimensionless iI the sampling Iunction f(t) is normalized
such that it is dimensionless as well. The sample is photo-detected and passed through
37
a lowpass electrical Iilter with impulse response h
e
(t). The resulting electrical signal
i(t) is given by:
t t t t q d ) ( ) ( ) ( ) ( ) ( ) (
0
2
= - =
}


t h t f P R t h t RP t i
e S e C
(3.2)
By deIinition the data pulses vary much more slowly than the sampling pulses.
ThereIore, (3.2) can be simpliIied to
t t t q d ) ( ) ( ) ( ) ( ) ( ) (
0
2
0 0
= - =
}


t h t f t P R t h t t RP t i
e S e C
(3.3)
Finally, considering that the bandwidth oI the samples is much larger than the Iilter
bandwidth, (3.3) approaches in the limit
) ( ) ( ) (
0 0
t t h t P R t i
e S
q . (3.4)
ThereIore, the peak current is simply a scaled version oI the sampled signal power. It
should be noted that the only theoretical constraint on h
e
(t) is that the electrical
bandwidth needs to be larger than the sampling rate (sampling source repetition
Irequency), as in deriving (3.4) it was tacitly assumed that no intersymbol interIerence
between samples occurs. However, the practical limitation is that the slower the
receiver, the lower the peak oI the impulse response (as the sample energy becomes
smeared over longer time interval) and thereIore the more diIIicult it is to detect.
Since the peak current is proportional to the electrical bandwidth, the electrical power
is proportional to the square oI the electrical bandwidth. At the same time, the thermal
noise variance is typically linear with electrical bandwidth. ThereIore, perhaps
slightly surprisingly, the detection sensitivity is better Ior larger electrical bandwidth.
This issue is explored in detail in Chapter 6, where a particular example oI the optical
sampling system based on FWM in SOAs is analyzed. II the electrical Iilter is
assumed to be an ideal Iinite time integrator, with integration time T, (3.3) simpliIies
at the sampling instant to

T
t P R i
S
t
q
A
= ) (
0
(3.5)
where
}
= A dt t f ) (
2
t is the eIIective optical sampling gate duration. Eq. (3.5) shows
that the optical sampling system eIIectively detects the sample energy but, since the
sampling gate width is much shorter than the receiver integration time, the electrical
signal level is signiIicantly reduced compared to the optical signal level. In Chapter 6,
38
the inIluence oI this energy smearing on the system signal to noise ratio (SNR) is
discussed in detail.
To reduce the requirements on the precision oI electrical analog-digital
converters, the optical sampling rate is usually moderate and less than 1 GS/s (i.e. the
sampling source repetition rate is less than 1 GHz) |9, 14|. This is obviously much too
slow to be able to reconstruct the signal waveIorm, as the bit rates used in
telecommunications are typically larger. However, signals in optical
telecommunications are digital and typically binary. ThereIore, the time shape oI a
particular bit is usually oI less interest than the statistical properties oI the signal. Low
rate synchronous sampling allows the measurement oI the eye diagram oI the signal
under test. Vertical histograms describing power distribution and horizontal
histograms describing signal jitter can be obtained Irom the eye. The noise
distribution obtained in this way can be related to the BER as described in Chapter 4.
Amplitude histograms can also be obtained using asynchronous sampling, but the
quality oI the histograms is lower, as they are contaminated by data points Irom the
signal pulse slopes.
Fig. 3.2 explains how the low rate synchronous sampling can be used to scan
the entire bit slot. In this case, the signal is periodic (which Ior typically random
telecommunication signals may require some Iorm oI waveIorm replication), so that
the envelope oI the samples is a time-magniIied copy oI the individual pulse. In a
typical case encountered in the world oI telecommunications (without the waveIorm
replication), the sampled signal is random so superimposing the samples would
produce the eye diagram.

Fig. 3.2 Principle oI low rate synchronous optical sampling
39
To scan over the bit slot, the sampling period T
g
must diIIer Irom a multiple oI
the bit period T
s
. The phase (position within the bit slot) oI each sample is determined
easily Irom the relationship between the signal repetition rate (f
S
1/T
S
) and the
sampling Irequency (f
g
1/T
g
)
f f k f
g S
A + = (3.6)
or alternatively Irom
S g
kT T t = A (3.7)
where Af is the bit slot scanning rate and At is the bit slot scanning step (one is not a
reciprocal oI the other, as obvious Ior instance Irom inverting (3.6)). k is a natural
number, being eIIectively the bit rate reduction Iactor. The inversion oI the oIIset
Irequency Af represents the sampled image period |15|. Af can be either positive or
negative. There is no principal diIIerence between the two cases, but when the oIIset
Irequency term adds up to the sampling Irequency, the time axis oI the sampled image
is inverted with respect to data reIerence time. Equations (3.6) and (3.7) show that
arbitrarily Iast waveIorms can be sampled to measure the eye diagram as the scanning
step can be made adequately small by adjusting the sampling rate. To display the eye
diagram, samples need to be superimposed aIter the whole bit slot has been scanned.
3.1.1. Review of optical sampling gates
Section 3.1 discussed the principle oI operation oI a generic, black-box type
optical sampling system. In reality, the biggest challenge is to Iind a way to perIorm
the actual optical sampling, i.e. to utilize a suitable physical process, which could be
used as a strobe. Generally speaking, there are two distinct approaches: one is to
utilize linear sampling, whereas the other makes use oI some type oI nonlinearity.
Linear sampling was introduced by Dorrer |16, 17| and basically consists oI a
pair oI balanced detectors, which detect the coherent homodyning term resulting Irom
the beating between the signal and ultraIast sampling pulses on the detectors. Simple
as it sounds, the method is quite challenging in practice, as it places stringent
requirements on the sampling source |18| and on the phase stability oI the whole
setup. This method is capable oI measuring optical amplitude and phase, so it has
been used to measure the constellation diagrams oI phase-modulated signals |16|.
However, due to the carrier-envelope phase slip typically present in the sampling
pulses, retrieval oI the signal phase requires additional processing or assumptions
40
about the nature oI this slip |18|. In principle, this technique is very powerIul, but
concerns about the stability do not make it a likely candidate Ior a common optical
perIormance monitor.
Sampling gates based on nonlinear eIIects have so Iar attracted considerably
more attention |4, 8-10, 19, 20|. They can be classiIied into second order or third
order gates, depending on the order oI the nonlinearity. Chapter 2 brieIly discussed
third order nonlinear eIIects in the SOAs (intrinsic and eIIective, more discussion oI
the possibility oI existence oI eIIective third order eIIects in SOAs is given in Chapter
5), but SOAs are not the only devices used to perIorm sampling as shown in the next
paragraph |4, 8, 9, 14, 20|. Higher order nonlinear processes are generally not used
due to their very small eIIiciencies. An alternative classiIication may be based on the
materials used. In practice three classes oI materials are considered Ior optical
sampling: nonlinear crystals, optical Iibres and semiconductors (among which, SOAs
play a prominent role). The classiIication based on nonlinearity seems to be more
Iundamental, so it is applied here.
Gates utilizing the second order susceptibility rely on nonlinear optical
crystals to perIorm three-wave mixing (usually sum Irequency generation, diIIerence
Irequency generation or second harmonic generation). Several types oI crystals have
been demonstrated to be applicable Ior nonlinear sampling. The most commonly used
are KTiOPO
4
(KTP) |21-23| and periodically poled lithium niobiate (PPLN) |24-26|,
but other crystals like LiIO
3
|7| or magnesium-doped PPLN |27| were also
considered. Also, organic AANP (2-adamantylamino-5-nitropyridine) nonlinear
crystals have been used |28|. The main drawback oI sampling gates based on
nonlinear crystals is that they require high peak powers and phase matching precludes
the use oI long crystals to improve the eIIiciency. However, the use oI periodically
poled (quasi-phase matched) materials signiIicantly improves the conversion
eIIiciency |9|. As mentioned in Chapter 2 and discussed in more detail in Chapter 5,
SOAs also exhibit some second order nonlinearities. However, since the eIIects
mediated by these nonlinearities are cascaded, they are externally seen as an eIIective
third order eIIect. ThereIore, SOA-based gates are not included in this category.
Gates relying on various third order nonlinearities are based either on
nonlinear optical Iibre |10, 14, and 19| or on semiconductors, among them SOAs |9,
11, 29-31|. The ultraIast Kerr nonlinearity in optical Iibre can provide very short
gating windows (limited by the sampling pulse duration and signal-sampling walk-oII
41
in the long Iibre), but the so-called highly nonlinear Iibres (HNLFs) need to be used to
provide an appreciable nonlinearity. Typically, Iibre-based gates utilize FWM |10,
14| or XPM |32|, but more sophisticated interIerometric devices based on XPM (like
the nonlinear optical loop mirror or the nonlinear Mach-Zehnder interIerometer) have
also been considered |9, 33|. The main drawback oI the Iibre-based gates is their
relatively low nonlinearity, even iI HNLFs are used. This requires high sampling peak
power (in |14| sampling pulses with 16 W peak power were used) and long lengths oI
Iibre, which limits the bandwidth and resolution |9, 14|. The high peak power and
long propagation length also leads to large SPM, which can distort the signal and
limits the signal-sampling detuning.
Sampling gates based on semiconductors can utilize the two photon absorption
eIIect |31| or FWM in passive waveguides |30| but the most commonly used
sampling gates are based on SOAs |9|. Among these, the Iunctionally simplest gates
are those employing FWM. FWM in SOAs has been discussed in Chapter 2. FWM in
SOAs has also been used to sample short optical pulses Ior example in |34-37|.
Generally speaking, these devices achieve quite a high conversion eIIiciency
compared to the gates based on passive materials. This is due to the high internal gain
oI the SOA. On the other hand, the price to pay Ior the conversion eIIiciency is a
problematic linearity oI the gate, due to signal induced gain saturation |11, 38|.
Additionally, ASE noise is added, resulting in reduced sample quality. A notable
exception to this is the so-called gain transparent FWM conIiguration |37|, where the
SOA gain is in the 1300 nm region, whereas signal is placed in 1550 nm region.
Beating is realized between two pumps placed in the gain region. This conIiguration
showed very high linearity, but the conversion eIIiciency was only about -50 dB |37|.
In Chapter 6, the perIormance oI the sampling gate based on FWM in SOAs is
thoroughly investigated theoretically and experimentally. The aIorementioned
shortcomings oI the FWM gates have led to development oI more complex
interIerometric gates, which utilize XPM |11, 29, 39, and 40|. In these devices, the
incoming signal pulse is split in the interIerometer and the sampling pulse unbalances
the interIerometer Ior a short moment, which leads to directing the sample to the
normally-destructive port. As usual, achieving stable interIerometric operation is the
main challenge Ior these gates.
42
3.2. Technical realizations of the optical sampling system for OPM
So Iar, only the optical sampling gates were brieIly reviewed. Another
essential element oI an optical sampling system is the sampling source. The
requirements Ior the optical sampling source are stringent. The source needs to
produce very short pulses (whose exact duration depends on the application) with
small timing jitter (much smaller than the pulsewidth) and small amplitude jitter.
Additionally, the preIerred repetition rate is less than 1 GHz |9|. These requirements
mean that only certain types oI lasers are suitable. Most typical sampling sources are
hybrid modelocked semiconductor lasers |9, 36, 37|, gain switched semiconductor
lasers (possible with nonlinear pulse compression) |31, 41|, passively modelocked
Iibre lasers |21, 27|, actively modelocked Iibre lasers |14, 22, 32, 35| or a CW laser
with an external modulator and nonlinear compressor |42|. Generally speaking, Iibre
lasers oIIer the best perIormance in terms oI chirp, peak power and jitter |9|.
Another important part oI the sampling system is the clock recovery circuit
providing sampling synchronization to the signal bit rate. Actually, this subsystem
does not have to be included in the setup, resulting in a much simpler design at the
expense oI lower perIormance. Depending on the type oI synchronization (or its lack),
optical sampling systems can be categorized into one oI the Iollowing groups:
- synchronous sampling, where the recovered clock may be used to
adjust the repetition rate oI the modelocked laser |14, 28|,
drive the external modulator producing pre-compressed pulses |42|,
measure the phase oI the sample relative to the system time base (the
optical clock), with Iree-running sampling laser |27|;
- pseudo-synchronous sampling, without physical clock recovery, but able to
measure the eye diagram by means oI
soItware synchronization, where the eye is recovered Irom
asynchronous samples by a special synchronization algorithm |19|.
The soItware clock recovery algorithm assumes constant signal bit rate
and is thereIore limited by the signal timing driIts.
limited persistence with accumulating time deviation. II the nominal
signal bit rate is known, the ratio oI the bit rate to sampling rate is
approximately known and can be used to determine the phase oI each
sample. Since this ratio is only approximate, the phase deviation would
43
accumulate and lead to the closure oI the eye. ThereIore, only a limited
number oI samples can be superimposed in one series but this
limitation can be circumvented since many series can be superimposed
|15|
waveIorm replication, where a short waveIorm is gated out Irom the
incoming signal and replicated. The resulting replicas are sampled,
where the time spacing between replicas is synchronized with the
sampling laser repetition period |39|.
- asynchronous sampling, where the sampling laser is Iree-running and agnostic
to the signal bit rate, resulting in complete bit rate transparency |41, 43|.
Most oI the clock recovery systems are based on injection locking in Iibre or
semiconductor lasers |4, 9| or on PLL loops, either all-optical or electrooptical |4, 9,
42|. As explained in |9|, only a Iew Iully complete synchronous sampling systems
have been presented and in most papers the sampling clock was derived directly Irom
the local electrical clock driving the signal modulator. This is because oI the technical
diIIiculties in realizing high quality optical clock recovery circuits. For stand-alone
high perIormance optical sampling oscilloscopes the ability to recover the eye
diagram is critical. However, Ior the card-level OPM monitors embedded in optical
nodes, the cost-eIIiciency and size are decisive, even at the expense oI reduced
capability |41, 44|. ThereIore, it appears that the asynchronous sampling method,
possibly combined with soItware synchronization, could well be the method oI choice
Ior OPM in transparent networks.
3.3. Optically sampled histograms and statistical moments
As explained previously, the main application oI optical sampling systems in
optical networks will probably be Ior BER estimation and monitoring. The BER
estimation using any sampling technique is typically based on the histograms.
Although some signal impairments can be identiIied based on the qualitative Ieatures
oI the histograms |45, 46|, this is not easily implemented in automated remote
monitors. Instead, the automated monitor has to operate on some quantitative data.
The most convenient and useIul metrics to describe the histograms are the statistical
moments. These are easily calculated Irom the collected samples by appropriate
averaging. The calculated moments are then Ied into the assumed BER statistical
model Ior BER estimation. The most popular model used Ior BER estimation is the
44
Q-Iactor method |5, 6, 12, 44|, which basically assumes that the total detection noise
has a Gaussian distribution (please note that the Q-Iactor is discussed in detail in
Chapter 4). Thus, this requires knowledge oI the Iirst two moments (the means and
the variances oI the signal logical levels). In Chapter 7, an alternative BER model is
proposed, which requires the knowledge oI the Iirst three moments, but oIIers certain
advantages over the Q-Iactor.
When considering any method utilizing moments to estimate the BER, the
important issue to note is that the measured moments depend on the measurement
system bandwidth. ThereIore, in principle, BER estimated Irom moments using the
monitor receiver can be diIIerent than the BER that would be seen by the real
receiver. This is especially signiIicant in the case oI optical sampling, since, as shown
in Section 3.1, the bandwidth oI the optical sampling is very large - several times
larger than the typical receiver bandwidth. From the physical point oI view this means
that much more noise is captured by the measurement system and this leads to change
in the values oI the calculated moments (e.g. an artiIicially increased variance). As
discussed in |19|, the optically sampled Q-Iactor is thereIore smaller than the
electrical Q-Iactor seen by the end-link receiver, so that the optical BER estimator
needs to be scaled.
To be able to use the optically sampled histograms Ior realistic BER
estimation, the expressions Ior optical moments have to be known to perIorm the
scaling. Likewise, knowledge oI the expressions Ior the electrical moments is
necessary (although an alternative route would be to use an experimental calibration
curve relating the optical moment to its electrical counterpart). Formulae Ior electrical
signal-spontaneous and spontaneous-spontaneous beat noise variances are well known
|19|. Expressions Ior higher order electrical moments are rarely used in the literature.
To analytically calculate these moments, the electrical signal probability density
Iunction (pdI) should in principle be known. It is usually assumed that the optical
ASE noise is Gaussian, which allows the calculation oI the pdI oI the electrical
current. Under the assumption that the ASE noise has a rectangular spectrum and that
the electrical Iilter can be modelled as an integrator (similar to the assumptions used
to calculate the variances oI the electrical noise beating terms), the pdI oI the
electrical current is given by the noncentral chi square distribution. More detailed
discussion oI this noise model is given in Chapter 7, as the novel BER estimation
45
method discussed in that chapter is strongly related to the noncentral chi square
model.
The Laplace transIorm oI a pdI is known as the moment generating Iunction
(mgI) |47| Ior the reason that consecutive diIIerentiation yields progressively higher
moments oI the pdI. To obtain the third moment, the mgI has to be diIIerentiated
thrice. In Chapter 7, the central third moment
3
is required Ior the new BER
estimator. AIter a considerable amount oI algebra, the expression Ior this moment can
be derived as:
( )
2 2 2 3 3
3
3 8
e ASE S o e ASE
B S P B B S R + = (3.8)
where B
e
, B
o
are, the electrical and optical Iilter bandwidth, respectively, and R is the
photodiode responsivity.
In the case oI optically sampled histograms, the expressions Ior variances oI
the signal-spontaneous and spontaneous-spontaneous beating terms are given in |19|:
2 2
2
2
2 2 2
2 1
2
t
t
q o
A +
A
=

o
o
ASE S ASEopt S
B
B
T
S P R (3.9)

2 2
2
2
2
2 2 2 2
2 1 t
t
q o
A +
A
=

o
o
ASE ASEopt ASE
B
B
T
S R (3.10)
ThereIore, knowledge oI the sampling resolution t A and the receiver
integration time T is necessary to be able to scale the optically measured variances
(3.9) and (3.10) to their electrical counterparts. It should be noted that the means oI
the histogram measured by the optical sampling system remain the same as Ior
electrical sampling, even iI the measurement bandwidth is larger. This is because the
mean oI the electrical noise is the DC component (or ASE power). ThereIore, it is
detected by any (DC coupled) receiver. Even iI the measurement bandwidth is larger
in the case oI optical sampling, it does not change the ASE bandwidth, which is the
sole Iactor determining the mean.
Expressions Ior higher order optical moments can be derived in a similar way
to (3.9) and (3.10). In Appendix A, a method Ior calculating these moments is
explained. This method is a generalization and simpliIication oI the derivation
presented in |19|. Here only the Iormula Ior the third central moment is reported. The
optically sampled counterpart oI (3.8) is
46
|
|
.
|

\
|
A +
A
+
A +
A
=
2 2
2
3
3
2
2 2
3
3
3
3 3 3
3
2 1
3
3 1
2
t
t
t
t
q
o
o
ASE S
o
o
ASE opt
B
B
T
S P
B
B
T
S R (3.11)
To summarize, possible physical implementations and system applications oI
optical sampling have been examined in this chapter. Detailed theoretical and
experimental investigations oI the Ieasibility oI optical sampling using FWM in SOAs
are reported in Chapter 6. As explained in the current chapter, optical sampling is
expected to Iind use Ior signal quality estimation and BER monitoring in Iuture
transparent optical networks. The conventional method Ior BER estimation employs
the Q-Iactor, which is discussed in Chapter 4. However, the Q-Iactor method has its
shortcomings and a better technique would be welcomed. In Chapter 7, one such
candidate Ior an optical signal quality metric is proposed and evaluated. This method
will be based (like the Q-Iactor method) on the moments oI the monitored signal.
47
3.4. References
|1| C. Pinart and G. J. Giralt, "On managing optical services in Iuture control-
plane-enabled IP/WDM networks," J. Lightw. Technol., vol. 23, pp. 2868-
2876, 2005.
|2| J. D. Downie and D. J. Tebben, "PerIormance monitoring oI optical networks
with synchronous and asynchronous sampling," Opt. Fiber Commun. Conf.
OFC 2001, vol. 3, pp. WDD50-1-WDD50-3, 2001.
|3| D. C. Kilper, R. Bach, D. J. Blumenthal, D. Einstein, T. Landolsi, L. Ostar, M.
Preiss, and A. E. Willner, "Optical perIormance monitoring," J. Lightw.
Technol., vol. 22, pp. 294-304, 2004.
|4| M. Saruwatari, "High-Speed All-Optical Technologies Ior Photonics," pp.
217-248, in High-speed photonic devices, N. Dagli, Ed., Taylor & Francis,
2007.
|5| F. Matera and M. Settembre, "Role oI Q-Iactor and oI time jitter in the
perIormance evaluation oI optically ampliIied transmission systems," IEEE J.
Sel. Topics Quantum Electron., vol. 6, pp. 308-316, 2000.
|6| N. S. Bergano, F. W. KerIoot, and C. R. Davidsion, "Margin measurements in
optical ampliIier system," IEEE Photon. Technol. Lett., vol. 5, pp. 304-306,
1993.
|7| T. Kanada and D. L. Franzen, "Optical waveIorm measurement by optical
sampling with a mode-locked laser diode," Opt. Lett., vol. 11, pp. 4, 1986.
|8| P. A. Andrekson and M. Westlund, "High-Speed WaveIorm Analysis Using
All-Optical Sampling," pp. 421-504, in Digital Communications Test and
Measurement. High-Speed Phvsical Laver Characteri:ation, D. Derickson;
and M. Mller, Eds., Prentice Hall, 2007.
|9| C. Schmidt-Langhorst and H.-G. Weber, "Optical sampling techniques," J.
Opt. Fiber Commun. Reports, vol. 2, pp. 86-114, 2005.
|10| P. A. Andrekson and M. Westlund, "Nonlinear optical Iiber based high
resolution all-optical waveIorm sampling," Laser Photon. Rev., vol. 1, pp.
231-248, 2007.
|11| L. A. Jiang, E. P. Ippen, U. Feiste, S. Diez, E. Hilliger, C. Schmidt, and H.-G.
Weber, "Sampling pulses with semiconductor optical ampliIiers," IEEE J.
Quantum Electron., vol. 37, pp. 118-126, 2001.
48
|12| P. Andre, J. Pinto, A. Teixeira, J. da Rocha, T. Almeida, and M. Pousa,
"Optical-signal-quality monitor Ior bit-error-ratio assessment in transparent
DWDM networks based on asynchronously sampled amplitude histograms," J.
Opt. Netw., vol. 1, pp. 118-128, 2002.
|13| H. Chen, A. W. Poon, and X.-R. Cao, "Transparent monitoring oI rise time
using asynchronous amplitude histograms in optical transmission systems," J.
Lightw. Technol., vol. 22, pp. 1661-1667, 2004.
|14| M. Westlund, P. A. Andrekson, H. Sunnerud, J. Hansryd, and J. Li, "High-
perIormance optical-Iiber-nonlinearity-based optical waveIorm monitoring," J.
Lightw. Technol.,, vol. 23, pp. 2012-2022, 2005.
|15| I. Shake, H. Takara, and S. Kawanishi, "Simple measurement oI eye diagram
and BER using high-speed asynchronous sampling," J. Lightw. Technol., vol.
22, pp. 1296-1302, 2004.
|16| C. Dorrer, "Monitoring oI optical signals Irom constellation diagrams
measured with linear optical sampling," J. Lightw. Technol., vol. 24, pp. 313-
321, 2006.
|17| C. Dorrer, "High-speed measurements Ior optical telecommunication
systems," IEEE J. Sel. Topics Quantum Electron., vol. 12, pp. 843-858, 2006.
|18| I. Kim, C. Kim, and G. Li, "Requirements Ior the sampling source in coherent
linear sampling," Opt. Express, vol. 12, pp. 2723-2730, 2004.
|19| M. Westlund, H. Sunnerud, M. Karlsson, and P. A. Andrekson, "SoItware-
synchronized all-optical sampling Ior Iiber communication systems," J.
Lightw. Technol., vol. 23, pp. 1088-1099, 2005.
|20| T. Suhara and M. Fujimura, "UltraIast Signal Processing Devices," in
Waveguide Nonlinear-Optic Devices (Springer Series in Photonics, J. 11):
SpringerVerlag, 2003.
|21| H. Ohta, N. Banjo, N. Yamada, S. Nogiwa, and Y. Yanagisawa, "Measuring
eye diagram oI 320 Gbit/s optical signal by optical sampling using passively
modelocked Iibre laser," Electron. Lett., vol. 37, pp. 1541-1542, 2001.
|22| H. Takara, S. Kawanishi, T. Morioka, K. Mori, and M. Saruwatari, "100
Gbit/s optical waveIorm measurement with 0.6 ps resolution optical sampling
using subpicosecond supercontinuum pulses," Electron. Lett., vol. 30, pp.
1152-1153, 1994.
49
|23| H. Takara, S. Kawanishi, and M. Saruwatari, "Optical signal eye diagram
measurement with subpicosecond resolution using optical sampling,"
Electron. Lett., vol. 32, pp. 1399-1400, 1996.
|24| S. Kawanishi, T. Yamamoto, M. Nakazawa, and M. M. Fejer, "High
sensitivity waveIorm measurement with optical sampling using quasi-
phasematched mixing in LiNbO
3
waveguide," Electron. Lett., vol. 37, pp. 842-
844, 2001.
|25| S. Nogiwa, Y. Kawaguchi, H. Ohta, and Y. Endo, "Highly sensitive and time-
resolving optical sampling system using thin PPLN crystal," Electron. Lett.,
vol. 36, pp. 1727-1728, 2000.
|26| S. Nogiwa, H. Ohta, Y. Kawaguchi, and Y. Endo, "Improvement oI sensitivity
in optical sampling system," Electron. Lett., vol. 35, pp. 917-918, 1999.
|27| R. L. Jungerman, G. Lee, O. BuccaIusca, Y. Kaneko, N. Itagaki, R. Shioda, A.
Harada, Y. Nihei, and G. Sucha, "1-THz bandwidth C- and L-band optical
sampling with a bit rate agile timebase," IEEE Photon. Technol. Lett., vol. 14,
pp. 1148-1150, 2002.
|28| H. Takara, S. Kawanishi, A. Yokoo, S. Tomaru, T. Kitoh, and M. Saruwatari,
"100 Gbit/s optical signal eye-diagram measurement with optical sampling
using organic nonlinear optical crystal," Electron. Lett., vol. 32, pp. 2256-
2258, 1996.
|29| C. Schubert, R. Ludwig, and H.-G. Weber, "High-speed optical signal
processing using semiconductor optical ampliIiers," J. Opt. Fiber Commun.
Reports, vol. 2, pp. 171-208, 2005.
|30| A. M. Darwish, E. P. Ippen, H. Q. Le, J. P. Donnelly, S. H. Groves, and E. A.
Swanson, "Short-pulse wavelength shiIting by Iour wave mixing in passive
InGaAsP/InP waveguides," Appl. Phvs. Lett., vol. 68, pp. 2038-2040, 1996.
|31| B. C. Thomsen, L. P. Barry, J. M. Dudley, and J. D. Harvey, "Ultra-sensitive
all-optical sampling at 1.5 μm using waveguide two-photon absorption,"
Electron. Lett., vol. 35, pp. 1483-1484, 1999.
|32| J. Li, M. Westlund, H. Sunnerud, B.-E. Olsson, M. Karlsson, and P. A.
Andrekson, "0.5-Tb/s eye-diagram measurement by optical sampling using
XPM-induced wavelength shiIting in highly nonlinear Iiber," IEEE Photon.
Technol. Lett., vol. 16, pp. 566-568, 2004.
50
|33| V. Marembert, C. Schubert, C. Schmidt-Langhorst, M. Kroh, S. Ferber, and H.
G. Weber, "Investigation oI Iiber based gates Ior time division demultiplexing
up to 640 Gbit/s," Opt. Fiber Commun. Conf. OFC 2006, pp. 3 pp., 2006.
|34| H. Kawaguchi and J. Inoue, "Four-wave mixing among subpicosecond optical
pulses in a semiconductor optical ampliIier and its applications to optical
sampling," SPIE Phvs. Sim. Optoelectron. Dev. JI Conf., vol. 3283, pp. 477-
484, 1998.
|35| M. Jinno, J. B. Schlager, and D. L. Franzen, "Optical sampling using
nondegenerate Iour-wave mixing in a semiconductor laser ampliIier,"
Electron. Lett., vol. 30, pp. 1489-1491, 1994.
|36| S. Diez, C. Schmidt, D. HoIImann, C. Bornholdt, B. Sartorius, H. G. Weber,
L. Jiang, and A. Krotkus, "Simultaneous sampling oI optical pulse intensities
and wavelengths by Iour-wave mixing in a semiconductor optical ampliIier,"
Appl. Phvs. Lett., vol. 73, pp. 3821-3823, 1998.
|37| S. Diez, R. Ludwig, C. Schmidt, U. Feiste, and H. G. Weber, "160-Gb/s
optical sampling by gain-transparent Iour-wave mixing in a semiconductor
optical ampliIier," IEEE Photon. Technol. Lett., vol. 11, pp. 1402-1404, 1999.
|38| A. Mecozzi and J. Mork, "Saturation eIIects in nondegenerate Iour-wave
mixing between short optical pulses in semiconductor laser ampliIiers," IEEE
J. Sel. Topics Quantum Electron., vol. 3, pp. 1190-1207, 1997.
|39| K.-L. Deng, R. J. Runser, I. Glesk, and P. R. Prucnal, "Single-shot optical
sampling oscilloscope Ior ultraIast optical waveIorms," IEEE Photon.
Technol. Lett., vol. 10, pp. 397-399, 1998.
|40| R. J. Runser, D. Zhou, C. Coldwell, B. C. Wang, P. Toliver, K.-L. Deng, I.
Glesk, and P. R. Prucnal, "InterIerometric ultraIast SOA-based optical
switches: From devices to applications," Opt. Quantum Electron., vol. 33, pp.
841-874, 2001.
|41| I. Shake, E. Otani, H. Takara, K. Uchiyama, Y. Yamabayashi, and T. Morioka,
"Bit rate Ilexible quality monitoring oI 10 to 160 Gbit/s optical signals based
on optical sampling technique," Electron. Lett., vol. 36, pp. 2087-2088, 2000.
|42| I. Kang and M. Yan, "Simple setup Ior simultaneous optical clock recovery
and ultra-short sampling pulse generation," Electron. Lett., vol. 38, pp. 1199-
1201, 2002.
51
|43| I. Shake, W. Takara, S. Kawanishi, and Y. Yamabayashi, "Optical signal
quality monitoring method based on optical sampling," Electron. Lett., vol. 34,
pp. 2152-2154, 1998.
|44| I. Shake and H. Takara, "Averaged Q-Iactor method using amplitude
histogram evaluation Ior transparent monitoring oI optical signal-to-noise ratio
degradation in optical transmission system," J. Lightw. Technol., vol. 20, pp.
1367-1373, 2002.
|45| N. Hanik, A. Gladisch, C. Caspar, and B. Strebel, "Application oI amplitude
histograms to monitor perIormance oI optical channels," Electron. Lett., vol.
35, pp. 403-404, 1999.
|46| C. M. Weinert, C. Caspar, M. Konitzer, and M. Rohde, "Histogram method Ior
identiIication and evaluation oI crosstalk," Opt. Fiber Commun. Conf. OFC
2000, vol. 3, pp. 56-58, 2000.
|47| A. Papoulis, Probabilitv, random variables, and stochastic processes, 3 ed:
McGraw-Hill Book Co., 1991.

52
4. The Q-factor method in OPM
In the preceding chapters the importance and role oI optical perIormance
monitoring in transparent optical networks was discussed. In particular, the need Ior a
means to estimate BER without the necessity oI Iull optical signal demultiplexing and
digital detection has been emphasized. This BER estimation thereIore needs to utilize
some measurable analogue physical properties oI the optical digital signal and has to
be able to inIer the digital signal quality Irom the analogue characteristics. This also
means that the BER evaluation is necessarily approximate, due to the Iact that very
many physical phenomena aIIect the BER and it is virtually impossible to characterize
all oI them at the same time. Furthermore, oIten a statistical model is assumed relating
the BER to the measurable quantities and any such model is inevitably approximate.
Probably the most popular method Ior BER estimation is the so-called Q-Iactor
method, introduced to the Iield oI optical communications by Personick |1| and
popularized by the paper oI Bergano et al. |2|. The Q-Iactor (as deIined below) will
be shown to be eIIectively a measure oI signal to noise ratio in the two level digital
signaling, i.e. it makes use oI the signal average levels and variances about these
levels. To use such a quantity to estimate BER is quite a natural approach, as it is
intuitively obvious that the probability oI incorrect detection is dependent on the
signal spread at the decision circuit and this spread is conveniently quantiIied by
signal to noise ratio. ThereIore, some quantity related to signal to noise ratio (Q-Iactor
Ior the binary digital transmission) can be seen as a natural Iigure oI merit Ior
transmission perIormance |2|. However, the relationship between the signal to noise
ratio and BER depends on the governing statistical model and this relationship is
unambiguous only Ior signal detection in the Gaussian channel (as the Gaussian
distribution is Iully determined by the Iirst two moments) |3|. In general, the
electronic-optical-electronic channel employed in optical communications is not
Gaussian and, moreover, it is usually Iar Irom being Gaussian |4, 5|. Nevertheless, the
assumption about Gaussian noise is widely used in the Q-Iactor method, due to its
simplicity and, in many cases, reasonable accuracy |2| (in Chapter 7 it is discussed in
detail that the oIten sensible accuracy oI the Q-Iactor method is just a Iortunate
coincidence as the electrical noise probability density Iunction (pdI) is very diIIerent
Irom Gaussian and therein an alternative signal quality metric is proposed). Because
53
oI the practical importance oI the Q-Iactor |6|, it is treated here in some detail. The
next section presents a short derivation oI the Q-Iactor. The measurement methods are
discussed later.
4.1. Short derivation of the Q-factor
Due to various random phenomena, generally reIerred to as noise, the
electrical signal at the decision circuit has to be treated as a random variable.
ThereIore, statistical methods are required to appropriately estimate the signal quality.
The random variable is Iully described by its pdI (or, alternatively, as mentioned in
Chapter 3, by its mgI). In binary digital transmission, the noise added to the signal
results in the total pdI being given by a mixture oI the two constituent pdIs
corresponding to the two logical levels. Each oI the constituent pdIs is centred at the
average value
0
(
1
) corresponding approximately to the signal value without noise
(the average optical noise power typically only slightly increases the means
0
,
1
).
The decision circuit needs to decide whether the received bit is a one` or zero` by
comparing its value to the threshold D. This means that iI the value oI a zero` bit at
the sampling instant is larger than the threshold (or less than the threshold Ior the
one` bit), an error is produced. The total probability oI error is thereIore given by the
integral over the tails oI the constituent pdIs Irom the point given by the decision
threshold D. The BER (deIined as the ratio oI the errored bits to the total number oI
bits) measured over suIIiciently long period oI time converges to this probability oI
error, as stated by the law oI large numbers. In what Iollows, it is assumed that the
probability oI error is equivalent to BER. Fig. 4.1 shows an idealized total signal pdI
at the decision point. The notation P(0,1) means the probability oI Ialsely detecting a
zero` when one` was sent. P(1,0) is deIined analogously.
54

Fig. 4.1 Typical probability density Iunction Ior binary signal at the decision circuit
Under the assumption that the Irequency oI occurrence oI zeros and ones is the
same, BER can be calculated as
) 0 , 1 ( 2 1 ) 1 , 0 ( 2 1 ) ( P P D BER + = (4.1)
where the notation BER(D) signiIies the importance oI setting the optimum threshold
level and the probabilities P(1,0), P(0,1) are given by:
}

=
D
dx x p P ) ( ) 0 , 1 (
0
(4.2 a)
}

=
D
dx x p P ) ( ) 1 , 0 (
1
(b)
The pdIs p
0
, p
1
depend on the physics oI the communication channel and on the
details oI the noise sources present. In the Q-Iactor method these pdIs are assumed
Gaussian. Optimal decision threshold D
opt
should be chosen such that (4.1) is
minimized. Calculating the derivative oI (4.1) with respect to D and setting it to 0
leads to the condition
) ( ) (
1 0
D p D p = (4.3)
Utilizing the assumption oI Gaussian noise, P(1,0), P(0,1) are expressed as
}

|
|
.
|

\
|
=
D
dx
x
P
2
0
2
0
0
2
) (
exp
2
1
) 0 , 1 (
o

o t
(4.4 a)
55
}

|
|
.
|

\
|
=
D
dx
x
P
2
1
2
1
1
2
) (
exp
2
1
) 1 , 0 (
o

o t
(b)
where
2
1 , 0
o are the noise variances associated with each signal level. Simple
manipulations lead to the normalized Iorms
}

|
|
.
|

\
|
=
0
2
exp
2
1
) 0 , 1 (
2
Q
dx
x
P
t
(4.5 a)
}

|
|
.
|

\
|
=
1
2
exp
2
1
) 0 , 1 (
2
Q
dx
x
P
t
(b)
where Q
0
(D-
0
)/o
0
and Q
1
(
1
-D)/o
1
. It is possible to calculate the optimum
threshold Irom (4.5 a, b) to IulIil the condition (4.3), but the resulting Iormula is quite
complicated and inconvenient. In practice, a slightly suboptimal threshold is used,
obtained by requiring
) 0 , 1 ( ) 1 , 0 ( P P = (4.6)
which leads to the condition Q
1
Q
0
Q, where Q is given by
0 1
0 1
o o

+

= Q (4.7)
Using (4.7) with (4.5) in (4.1) results in the Iinal Iormula Ior the BER
|
.
|

\
|
=
2
erIc
2
1 Q
BER (4.8)
where erIc(.) is the complementary error Iunction. It is worth observing that this BER
model is remarkably uncomplicated both Irom the analytical and practical points oI
view. It is reasonably straightIorward to measure the quantities necessary to calculate
the Q and then the connection between Q and BER given by (4.8) is very simple. The
model requires only limited inIormation about the signal, which on one hand results in
its simplicity and on the other means that it is very idealized. As the simplicity is the
biggest advantage oI the Q-Iactor technique, in the next section ways used to measure
it are brieIly reviewed.
56
4.2. Q-factor measurement methods
Several methods have been proposed in the literature to measure the Q-Iactor.
The classic approach was presented in the article by Bergano et al. |2|. This method
requires a Iull electrical digital receiver, with means to adjust the decision threshold.
As the threshold is swept, the BER is recorded Ior each threshold value. For very low
values oI the decision threshold, the resultant numerous errors are dominated by
erroneously recorded zeros`. Similarly, Ior very high threshold the BER is very poor
due to many erroneously detected ones`. In this way, the curve BER(D) naturally
separates into two branches. These two branches would converge and intersect Ior
D
opt
. As shown in |2|, it is possible to invert the measured BER curves and estimate

1
, o
1
Irom one measured branch (very high threshold settings) and
0
, o
0
Irom the
other branch (very low threshold settings). The main advantage oI this method is very
short measurement time instead oI long measurements required to directly measure
very low BER, the parameters required Ior the Q are measured very quickly Ior
thresholds Iar Irom optimal. It has to be noted however that a Iull digital receiver is
necessary to perIorm the non-optimal BER measurements, thereIore the method is not
modulation Iormat and bit rate transparent and thus cannot be considered as an OPM
method in the sense employed in this thesis. Nevertheless, it has attracted some
attention in early works devoted to perIormance monitoring in transparent networks
due to its speed and simplicity |7, 8|.
A similar but alternative method to measure the Q was proposed in |9|. Instead
oI estimating the Iirst and second order moments Irom the inverted BER(D) curves,
this method used a variable decision threshold and recorded the number oI zeros` Ior
each threshold setting. For threshold signiIicantly larger than the mean one` level`,
eIIectively all bits are detected as zeros`. By diIIerentiating this Irequency oI
occurrence curve, an estimation oI pdI is obtained and the moments are calculated
Irom this pdI. This method has the same advantages and limitations as the Bergano
method and its main shortcoming is that it lacks the Ilexibility required Irom
monitoring equipment in transparent networks. Although some works reported
utilizing this method to measure the Q-Iactor |10|, in general it has not gained
appreciable popularity and more Ilexible methods were sought.
It appears that the most versatile Q-Iactor monitoring methods utilize sampling
histograms. The histograms serve as an approximation to the true signal pdIs. It is
57
straightIorward to calculate statistical moments Irom the histograms. Even iI the
histograms may poorly approximate the Iar tails oI the true pdIs, the moment
estimators obtained Irom the histograms generally converge to the true moments iI
suIIiciently many samples are included. There are many methods to obtain the
samples and to generate histograms. One classiIication oI the sampling-based
methods depends on whether the sampling process takes place in the optical or
electrical domain. Although many works employed electrical sampling |10-13|, this
approach is limited by the maximum sampling resolution. However, a signiIicant
advantage oI the electrical sampling is relatively simple clock recovery, which (as
discussed in Chapter 3) results in better quality histograms. To show how the presence
oI noise maniIests itselI in the histograms, i.e. how the histograms are sensitive to
signal quality, Fig. 4.2 presents electrically sampled histograms Ior a good quality
signal (upper trace in Fig. 4.2 with a widely open eye diagram) and Ior a severely
degraded signal (lower trace in Fig. 4.2, the eye opening is less due to optical noise
and chromatic dispersion)

Fig. 4.2 Examples oI electrically sampled histograms Ior good quality signal (upper trace, wide open
eye diagram) and degraded signal (lower trace, noisy eye diagram)
58
Optical sampling techniques, as discussed in Chapter 3, are suitable Ior a
much wider range oI bit rates, including bit rates much Iaster than the current state-oI-
the-art 40Gb/s transmission speed |14-17|. However, the ultrawide bandwidth oI
optical sampling systems changes the measured noise variances, as compared to
measured by the standard receiver |18|. This requires an appropriate scaling oI
optically sampled moments, as showed in Chapter 3. One possible implementation oI
the optical sampling system (based on FWM in SOAs) is investigated theoretically
and experimentally in Chapter 6.
Another classiIication oI sampling systems may take into account the presence
or lack oI the signal clock recovery subsystem in the sampling setup. As discussed in
Chapter 3, synchronous sampling produces better quality histograms, as the sampling
takes place at the point oI maximum signal eye opening, whereas asynchronous
sampling is more Ilexible and transparent, simpler and more cost-eIIective. Optical
sampling systems utilizing both synchronous |17-19| and asynchronous |14, 15, 20|
sampling to measure the Q-Iactor have been presented.
4.3. Limitations of the Q-factor
Although the Q-Iactor method proved to be useIul and quite powerIul, it has
some inherent limitations. These are related to the Iact that, on one hand, only limited
inIormation about the signal is used and, on the other, the Gaussian noise model is
generally not valid. In particular, it is known that the standard Q-Iactor method gives
an overly conservative estimation in the presence oI waveIorm distortion, especially
those leading to intersymbol interIerence and splitting the binary logic levels into the
sublevels |21-23|. This can be explained by considering that the eIIect oI level
splitting on the histogram is equivalent to separating the pdIs oI ones` and zeros`
into a mixture oI several Gaussian-like constituent pdIs, each with slightly diIIerent
mean and possibly diIIerent probability oI occurrence. II the eIIective pdI oI each
binary level is approximated by a single Gaussian Iunction, the variance oI this pdI is
larger than sum oI variances oI the underlying components. This in turn leads to Iatter
tails oI the approximating pdI and thereIore to over-pessimistic BER estimation. This
eIIect is schematically shown in Fig. 4.3.
59

Fig. 4.3 Schematic representation oI the eIIect oI level splitting
It was shown that iI, instead oI using a single Gaussian Iunction to
approximate the pdI oI each logic level, a set oI Gaussian Iunctions, each
corresponding to one sublevel, was used, the eIIect oI waveIorm distortions can be
reasonably accurately incorporated in the model in a relatively simple way |21, 23|.
The resulting expression Ior BER is
(
(

|
|
.
|

\
|

+
|
|
.
|

\
|

=

m
m
m
m
n
n
n
n
D
p
D
p BER
, 1
, 1
, 0
, 0
2
erIc
2
erIc
4
1
o

(4.9)
where the weights p
n, m
are the probabilities oI occurrence oI each zero` and one`
sublevel, respectively, and the same subscript notation is used Ior the means and
variances oI each sublevel. Alternatively, instead oI assuming a given number oI
sublevels (based Ior example on the knowledge oI accumulated dispersion), the
experimental histogram can be Iitted with any number oI Gaussian sub-pdIs giving
the best Iit |22|. Since the constituent sub-pdIs are usually oI small variance, they are
decaying Iast and thereIore in most cases BER given by (4.9) is dominated only by
the sublevels closest to the decision threshold |23|.
The other major limitation oI the Q-Iactor technique is not so easily
circumvented, as it is related to the Iact that, Iundamentally, it is not the correct noise
model Ior an optically ampliIied link |4, 5|. This is, Ior instance, especially evident iI
thermal light (ASE) is used to carry the inIormation, which has been considered Ior
low-cost spectrum-sliced systems |24|. It has been shown that in spectrally-sliced
incoherent transmission the Q-Iactor method is inaccurate, unless large optical
bandwidth is used |25|. Even though in other scenarios the BER estimation obtained
using the Q-Iactor is oIten reasonable, a closer look at the shape oI the real pdIs
60
reveals that they are very diIIerent Irom Gaussian. This apparent contradiction is
resolved by the Iact that the optimum threshold predicted by (4.7) is very diIIerent
Irom the true decision threshold |4, 5, 26| and it is only a Iortunate coincidence that
the inaccuracy in the pdI shape and in the threshold setting to a large extent cancel
out. However, it means that Q-Iactor cannot be used to predict the optimum decision
threshold. ThereIore, the Q-Iactor is oI little use in applications requiring the
knowledge oI the real optimum threshold or oI the real shape oI the pdI. Recently,
sequence detection using diIIerent error correction codes or soIt decoding methods
has attracted reasonable attention, as a possible means to improve the transmission
perIormance in optically ampliIied links |27, 28|. It has been shown that the
assumption oI Gaussian noise leads to a very signiIicant perIormance degradation oI
sequence decoders |27, 28|. In Chapter 7, a novel signal quality metric is proposed,
which, whilst retaining similar simplicity as the Q-Iactor method, is capable oI
approximating much better the true shape oI the signal pdI and predicting in the
optimum decision threshold |26|.
Finally, it is worth noting that Ior Gaussian ASE noise, the Gaussian noise
model is the correct model Ior coherent detection amplitude modulated systems (Ior
example quadrature-amplitude modulation |29|) and thereIore the Q-Iactor method is
strictly exact in this case. However, as the optical Iibre is not a linear medium, Ior
large signal powers and low dispersion the output ASE is not Gaussian |30, 31| and
the Q-Iactor method Iails in this case. Interestingly, the Q-Iactor method may be
overly optimistic in this scenario |30|, in contrast with the more common observation
when the Q-Iactor is conservative. This issue is explored in detail in Chapter 9.
61
4.4. References
|1| S. D. Personick, "Receiver design Ior digital Iiber optic communications
systems, I," Bell Svst. Tech. J., vol. 52, pp. 843-874, 1973.
|2| N. S. Bergano, F. W. KerIoot, and C. R. Davidson, "Margin measurements in
optical ampliIier system," IEEE Photon. Technol. Lett., vol. 5, pp. 304-306,
1993.
|3| R. N. McDonough and A. D. Whalen, Detection of signals in noise, 2 ed.,
Academic Press, 1995.
|4| D. Marcuse, "Calculation oI bit-error probability Ior a lightwave system with
optical ampliIiers and post-detection Gaussian noise," J. Lightw. Technol., vol.
9, pp. 505-513, 1991.
|5| P. A. Humblet and M. Azizoglu, "On the bit error rate oI lightwave systems
with optical ampliIiers," J. Lightw. Technol., vol. 9, pp. 1576-1582, 1991.
|6| F. Matera and M. Settembre, "Role oI Q-Iactor and oI time jitter in the
perIormance evaluation oI optically ampliIied transmission systems," IEEE J.
Sel. Topics Quantum Electron., vol. 6, pp. 308-316, 2000.
|7| R. E. Neuhauser, "Optical perIormance monitoring (OPM) in next-generation
optical networks," SPIE Netw. Design Management Conf., vol. 4909, pp. 34-
42, 2002.
|8| A. Richter, W. Fischler, H. Bock, R. Bach, and W. Grupp, "Optical
perIormance monitoring in transparent and conIigurable DWDM networks,"
IEE Proc - Optoelectron., vol. 149, pp. 1-5, 2002.
|9| S. Ohteru and N. Takachio, "Optical signal quality monitor using direct Q-
Iactor measurement," IEEE Photon. Technol. Lett., vol. 11, pp. 1307-1309,
1999.
|10| J. D. Downie and D. J. Tebben, "PerIormance monitoring oI optical networks
with synchronous and asynchronous sampling," Opt. Fiber Commun. Conf.
OFC 2001, vol. 3, pp. WDD50-1-WDD50-3, 2001.
|11| P. Andre, J. Pinto, A. Teixeira, J. da Rocha, T. Almeida, and M. Pousa,
"Optical-signal-quality monitor Ior bit-error-ratio assessment in transparent
DWDM networks based on asynchronously sampled amplitude histograms," J.
Opt. Netw., vol. 1, pp. 118-128, 2002.
62
|12| N. Hanik, A. Gladisch, C. Caspar, and B. Strebel, "Application oI amplitude
histograms to monitor perIormance oI optical channels," Electron. Lett., vol.
35, pp. 403-404, 1999.
|13| C. M. Weinert, C. Caspar, M. Konitzer, and M. Rohde, "Histogram method Ior
identiIication and evaluation oI crosstalk," Opt. Fiber Commun. Conf. OFC
2000, vol. 3, pp. 56-58, 2000.
|14| I. Shake, E. Otani, H. Takara, K. Uchiyama, Y. Yamabayashi, and T. Morioka,
"Bit rate Ilexible quality monitoring oI 10 to 160 Gbit/s optical signals based
on optical sampling technique," Electron. Lett., vol. 36, pp. 2087-2088, 2000.
|15| I. Shake and H. Takara, "Averaged Q-Iactor method using amplitude
histogram evaluation Ior transparent monitoring oI optical signal-to-noise ratio
degradation in optical transmission system," J. Lightw. Technol., vol. 20, pp.
1367-1373, 2002.
|16| I. Shake, H. Takara, and S. Kawanishi, "Simple measurement oI eye diagram
and BER using high-speed asynchronous sampling," J. Lightw. Technol., vol.
22, pp. 1296-1302, 2004.
|17| C. Schmidt, C. Schubert, J. Berger, M. Kroh, H.-J. Ehrke, E. Dietrich, C.
Borner, R. Ludwig, H. G. Weber, F. Futami, S. Watanabe, and T. Yamamoto,
"Optical Q-Iactor monitoring at 160 Gb/s using an optical sampling system in
an 80 km transmission experiment," Conf. Lasers Electro-Opt. CLEO 2002
Tech. Digest., pp. 579-580 vol.1, 2002.
|18| M. Westlund, H. Sunnerud, M. Karlsson, and P. A. Andrekson, "SoItware-
synchronized all-optical sampling Ior Iiber communication systems," J.
Lightw. Technol., vol. 23, pp. 1088-1099, 2005.
|19| M. Westlund, H. Sunnerud, M. Karlsson, J. Hansryd, J. Li, P. O. Hedekvist,
and P. A. Andrekson, "All-optical synchronous Q-measurements Ior ultra-high
speed transmission systems," Opt. Fiber Commun. OFC 2002, pp. 530-531,
2002.
|20| I. Shake, W. Takara, S. Kawanishi, and Y. Yamabayashi, "Optical signal
quality monitoring method based on optical sampling," Electron. Lett., vol. 34,
pp. 2152-2154, 1998.
|21| C. J. Anderson and J. A. Lyle, "Technique Ior evaluating system perIormance
using Q in numerical simulations exhibiting intersymbol interIerence,"
Electron. Lett., vol. 30, pp. 71-72, 1994.
63
|22| L. Ding, W.-D. Zhong, C. Lu, and Y. Wang, "New bit-error-rate monitoring
technique based on histograms and curve Iitting," Opt. Express, vol. 12, pp.
2507-2511, 2004.
|23| S. Norimatsu and M. Maruoka, "Accurate Q-Iactor estimation oI optically
ampliIied systems in the presence oI waveIorm distortions," J. Lightw.
Technol., vol. 20, pp. 19-27, 2002.
|24| V. Arya and I. Jacobs, "Optical preampliIier receiver Ior spectrum-sliced
WDM," J. Lightw. Technol., vol. 15, pp. 576-583, 1997.
|25| M. Leeson, "Determination oI the optimum electrical bandwidth Ior spectral
slicing," Opt. Quantum Electron., vol. 38, pp. 667-673, 2006.
|26| M. Dlubek, A. Phillips, and E. Larkins, "Optical signal quality metric based on
statistical moments and Laguerre expansion," Opt. Quantum Electron., vol.
40, pp. 561-575, 2008.
|27| Y. Cai, J. M. Morris, T. Adali, and C. R. Menyuk, "On turbo code decoder
perIormance in optical-Iiber communication systems with dominating ASE
noise," J. Lightw. Technol., vol. 21, pp. 727-734, 2003.
|28| G. Bosco, G. Montorsi, and S. Benedetto, "SoIt decoding in optical systems,"
IEEE Trans. Commun., vol. 51, pp. 1258-1265, 2003.
|29| J. Hongo, K. Kasai, M. Yoshida, and M. Nakazawa, "1-Gsymbol/s 64-QAM
Coherent Optical Transmission Over 150 km," IEEE Photon. Technol. Lett.,
vol. 19, pp. 638-640, 2007.
|30| M. P. Dlubek, A. J. Phillips, and E. C. Larkins, "Evolution oI the probability
density Iunctions oI Gaussian ASE noise in zero-memory nonlinear Iiber,"
accepted for publication in Opt. Fiber Technol.
|31| M. P. Dlubek, A. J. Phillips, and E. C. Larkins, "Nonlinear Evolution oI
Gaussian ASE Noise in ZMNL Fiber," J. Lightw. Technol., vol. 26, pp. 891-
898, 2008.



64
5. Characterization and applications of SOAs
As reviewed and brieIly explained in Chapters 1 and 2, SOAs are expected to
be the key devices enabling various Iorms oI photonic signal processing |1-3|, all-
optical regeneration |4-6| and also OPM, where they are mainly proposed Ior use in
optical sampling |2, 7, 8|. This wide range oI applications is possible thanks to the
Iast and rich nonlinear dynamics oI the SOAs, which permits the all-photonic signal
manipulation. In Chapter 2, FWM was discussed, which is one oI the most important
nonlinear processes in the SOAs. In Chapter 6, the Ieasibility oI using FWM in SOAs
to perIorm the optical sampling is investigated theoretically and experimentally.
However, beIore SOA-based optical sampling can be discussed, it is necessary to
thoroughly characterize the SOA used in the experiments so as to obtain an
understanding oI its non-trivial gain dynamics and nonlinearity. The SOA used in all
the experiments is the SOA-NL-OEC-1550 device Irom Centre Ior Integrated
Photonics (CIP), which has been optimized Ior nonlinear applications.
The Iirst part oI this chapter presents an experimental characterization oI the
SOA device used later to perIorm the sampling. The gain and spectral properties are
characterized. In particular, a general FWM characteristic was experimentally
investigated. It was shown that beating between the input waves indeed induces the
CDP |9|, which is an assumption on which most oI the theories oI FWM in SOAs are
based |10, 11|. However, it is also discussed that, contrary to the usual explanation,
the CDP may be mediated by second order nonlinearity. ThereIore FWM in SOAs
may be due to an apparent third order nonlinearity (i.e. due to combined intrinsic third
order nonlinearity and cascaded second order eIIects). In the second part oI the
chapter, a novel application oI FWM in SOAs Ior optical signal regeneration is
presented, which allows an extinction ratio improvement |12|.
5.1. Gain and noise measurements
The gain and noise properties are the most basic Ieatures oI any optical
ampliIier. ThereIore, they are investigated Iirst in this chapter. Fig. 5.1 shows the ASE
spectral density S
ASE
Ior diIIerent bias current I
SOA
. It should be noted that the peak
wavelength is dependent on I
SOA
and shiIts towards shorter wavelength Ior higher
pumping. This behaviour is well known in the literature |13|. The ASE spectrum (and
thereIore, approximately, the unsaturated gain) is Ilat over the C-band (1535 1560
65
nm) Ior high pumping current, which is important Irom the practical point oI view.
However, it should be noted that Ior low pumping current, the ASE spectrum varies
by about 8 dB over the C-band. The jagged part oI the 80 mA curve in Fig. 5.1 is due
to the optical spectrum analyser noise and limited sensitivity. This knowledge about
the SOA noise is utilized in Chapter 6 when calculating the signal to noise ratio
(SNR) in the optical sampling system. However, the SOA gain spectrum also depends
on the input optical power, due to the carrier depletion and gain saturation.
Fig. 5.2 shows the gain variation over the C-band (spectral range limited by
the tunability oI the laser used in the experiment) Ior small (P
in
-30 dBm), medium
(P
in
-15 dBm) and large (P
in
-3 dBm) signal conditions, at I
SOA
240 mA. It is
worth noting that the large signal gain varies very little over the whole C-band. This is
due to the selI-clamping eIIect: a larger initial gain results in deeper gain saturation in
the later part oI the SOA, whereas iI the signal experiences lower gain initially, it will
not saturate the SOA section at the output Iacet so much.
Fig.5.3 shows the dependence oI the small signal gain G
0
on the bias current
I
SOA
. The experimental current range was limited by the maximum manuIacturer`s
bias rating oI 300 mA. The probe parameters were: input power P
in
-27 dBm,
wavelength 1556 nm. The maximum gain is over 30 dB, consistent with the
manuIacturer`s speciIication and the gain curve exhibits strong saturation Ior large
pumping current. This shape oI the G
0
(I
SOA
) curve is quite typical and expected |13|.
The device transparency current is about 50 mA. Below this bias, the ampliIier is
absorbing.
ASE spectraI density for different bias current
1.E-19
1.E-18
1.E-17
1.E-16
1.E-15
1500 1520 1540 1560 1580 1600
[nm]
S
ASE
[W/Hz]
80 mA
150 mA
220 mA

SOA

Fig. 5.1 ASE spectral density S
ASE
Ior diIIerent bias conditions
66
Gain spectrum for different input powers
0
5
10
15
20
25
30
35
1530 1535 1540 1545 1550 1555
[nm]
G [dB]
small sig
medium sig
large sig

Fig. 5.2 Gain spectrum Ior diIIerent input powers (P
in
-30, -15 and -3 dBm) Ior I
SOA
240 mA
Gain dependence on the bias current G(I)
-20
-10
0
10
20
30
40
40 80 120 160 200 240 280
I
SOA
[mA]
G [dB]

Fig. 5.3 Small signal gain dependence on the bias current
The last basic experimental curve reported in this section is the gain saturation
curve Ior diIIerent bias currents I
SOA
(Fig.5.4). This curve is very important Irom the
practical point oI view, as it shows how the ampliIier gain is compressed by the input
power. The output power corresponding to the 3 dB gain compression is called the
ampliIier output saturation power P
sat0
|13, 14|. It is desirable to have as high
saturation power as possible, to increase both the dynamic range oI the ampliIier and
the gain available at high input powers. Typical values Ior P
sat0
are in the region oI 10
dBm |3, 13, 14|. The manuIacturer`s speciIication Ior the device used in the reported
experiments was also 10 dBm. Later in this chapter a way to increase P
sat0
will be
discussed and experimentally evaluated. In Chapter 6, it will be shown that low P
sat0

67
may be an important Iactor limiting perIormance oI the optical sampling system based
on FWM in SOAs. It can be observed that the experimentally measured P
sat0
is
consistent with the device speciIication and that it increases with bias current.
Gain saturation curve G(Pin) for differrent bias current
0
5
10
15
20
25
30
35
-15 -10 -5 0 5 10 15
Pout [dBm]
G [dB]
80 mA
150 mA
220 mA
290 mA

Fig. 5.4 Gain saturation curves Ior diIIerent bias conditions
Figs. 5.1-5.4 present the basic noise and gain properties oI the SOA used in
the experiments. The results and plots presented in this section are consistent with
similar graphs reported in the literature. It can be seen that the SOA is a high gain
device, as the maximum gain is larger than 30 dB. The gain sensitivity to input signal
state oI polarization is small and was experimentally determined to be about 1dB,
which is also consistent with values reported by the manuIacturer. Although basic, the
measurements presented in this section are necessary to properly characterize the
device and to understand its perIormance. This understanding is later taken into
account during the design oI diIIerent experimental FWM setups and makes it
possible to explain the obtained results. In particular, in Chapter 6, the measurements
presented here are used to model optical sampling using FWM in SOAs and to
explain the experimental results.
5.2. CW four wave mixing in SOAs
CW FWM in SOAs has long been a subject oI scientiIic research (see
reIerences in Chapter 2). Although this phenomenon seems to be reasonably well
understood, some uncertainty regarding its nature still exists, as discussed later in this
chapter. As explained in Chapter 2, almost all theories oI FWM, including the
68
phenomenological |10| and the more Iundamental density matrix |11| approaches, are
based on the assumption that beating between the input waves causes inter- and
intraband pulsations. The experimental results reported by many researchers agree
well with the theory |15, 16|. However, there seems to be a lack oI experimental data
conIirming the basic assumptions. Later in this chapter we report on experiments
conIirming the existence oI CDP due to wave beating and discuss possible nonlinear
eIIects responsible Ior CDP in view oI our experiments. An analytical time domain
model Ior FWM was described in Chapter 2 and utilized in Chapter 6 to investigate
the perIormance oI the optical sampling scheme based on FWM in SOAs. This model
can be equally well applied to CW FWM iI the appropriate Iorm Ior the input signals
is used. Since numerical investigations oI CW FWM are reported in numerous papers
(Ior example in |10, 11, 15-17|), there is little point in repeating these basic
investigations here. The interested reader is reIerred to any oI the reIerences
mentioned above. However, it is still worth presenting the experimental
measurements oI the CW FWM Ior the particular SOA used in this work. This is to
make sure that its behaviour is in agreement with the theoretical expectations and to
obtain a practical understanding oI the experimentally achievable perIormance. This
is important because the numerical models depend on a number oI parameters and
even a small diIIerence in the value oI each parameter, or oI the relative magnitudes
between a group oI parameters, can sometimes have large inIluence on the results.
Similarly, the experimental results reported in the literature are always device speciIic
and thereIore cannot perIectly describe the case when a diIIerent SOA is used (they
do provide however a useIul sanity check to make sure that the experimental setup is
properly designed and calibrated). ThereIore, in this section some Iundamental plots
showing the essential aspects oI FWM in the CIP SOA-NL-OEC-1550 device are
presented and brieIly discussed. Most oI the basic measurements are taken with the
setup shown in Fig. 5.5. In cases where this setup had to be modiIied, this will be
explained in the text. Both lasers shown in Fig. 5.5 were DFB diode lasers. The signal
laser was tunable Irom 1530 nm to 1560 nm. Optical isolators (ISO) were used to
prevent creation oI any resonant cavities, which could distort the SOA spectrum. An
optical bandpass Iilter (BPF) was used to separate the conjugate wave Irom the rest oI
the spectrum. The power meter (PM) was used to measure the power oI the conjugate.
69

Fig. 5.5 Experimental setup used to investgate CW FWM. ATT attenuator, ISO isolator, PC
polarization controller, BPF bandpass Iilter, PM power meter, OSA optical spectrum analyzer.
Fig. 5.6 shows a typical CW FWM spectrum measured at the SOA output. P
P

and P
S
are the pump and signal power, respectively, and only these waves were input
to the SOA. Their powers at the input were P
P
1.5 dBm, P
S
-5 dBm. Due to FWM,
another two waves P
C
and P
L
are created, which are the conjugate replicas oI P
S
and
P
P
, respectively. In the standard FWM conIiguration it is the wave P
C
which is oI
most interest. It will be shown however that the wave P
L
can also be useIul and may
have applications to signal regeneration |12|. It should also be noted that the
conjugate power is quite high, showing good conversion eIIiciency, which was about
-10 dB in this case. In general, in all practical FWM applications the conversion
eIIiciency p should be as high as possible. Another useIul Iigure oI merit is the signal
to ASE background ratio (SBR), which is a ratio oI the conjugate power to the ASE
power |16|, usually measured using an OSA. ThereIore, in the Iollowing, both the
optimum conversion eIIiciency and SBR are investigated.

Fig. 5.6 Typical FWM spectrum showing presence oI the two mixing products P
C
and P
L

5.2.1. Dependence of CW FWM efficiency and SBR on pump power
From the practical point oI view, it is important to optimize the pump power to
obtain the optimum conversion eIIiciency and SBR. As discussed in Chapter 2, the
70
conjugate power is proportional to the square oI the pump power. At the same time,
however, the gain saturation induced by the pump limits the ampliIication oI the
conjugate. ThereIore, there is an optimum pump power when the two opposing eIIects
are balanced. Fig. 5.7 shows the experimentally measured conversion eIIiciency p as a
Iunction oI pump power. The signal power was used as a parameter to show that it
also contributes to the gain saturation Ior large signal powers. Another consequence
oI this is the nonlinear dependence oI the conjugate power on the signal power, which
is investigated later in this chapter. To obtain Fig. 5.7, the SOA bias current was I
SOA
300 mA and the wavelength detuning between the pump and the signal was 4 nm.
The optimum pump power depends on the signal power. The maximum conversion
eIIiciency was almost -4 dB, but it decreases with the increasing signal power. To
ensure that the SOA is saturated predominantly by the pump, large pump powers are
preIerred. This improves the linearity oI the conversion. A similar characteristic has
been reported in the literature |18|.
Fig. 5.8 presents the SBR corresponding to the conversion eIIiciency in Fig.
5.7. It should be noted that a low signal power results in a low SBR and that the use oI
a high power pump improves the SBR. These Iindings are also consistent with the
literature |18, 19|. The case oI low p and high SBR is more preIerable than the
opposite, as the low conjugate power can always be ampliIied using a booster.
ThereIore high pump powers above 3 dBm are preIerred.
q qq q(P
pump
), I
soa
=300mA
-23
-18
-13
-8
-3
-10 -8 -6 -4 -2 0 2 4
Ppump [dBm]
q

q

q

q

[
d
B
]
Psig=-15dBm
Psig=-10dBm
Psig=-5dBm
Psig=-0.3dBm

Fig. 5.7 Conversion eIIiciency against pump power Ior diIIerent signal values (detuning 4 nm,
I
SOA
240 mA)
71
SBR(P
pump
), I
soa
=300mA
10
12
14
16
18
20
22
24
26
28
-8 -6 -4 -2 0 2 4
Ppump [dBm]
S
B
R

[
d
B
]
Psig=-15dBm
Psig=-10dBm
Psig=-5dBm
Psig=-0.3dBm

Fig. 5.8 SBR against pump power Ior diIIerent signal values (detuning 4 nm, I
SOA
240 mA)
5.2.2. Dependence of CW FWM efficiency and SBR on detuning
It was mentioned in Chapter 2 that the conversion eIIiciency strongly depends
on the Irequency spacing between the input waves. This is because the physical
processes responsible Ior the mixing have Iinite characteristic times so that the
response oI the medium decreases rapidly as the optical beat Irequency exceeds the
reciprocal oI the characteristic time. Even iI the carrier heating (CH) and particularly
the spectral hole burning (SHB) characteristic times are very short, as discussed in
Chapter 5, the carrier density pulsations (CDP) time is much longer. At the same time
the maximum magnitude oI the CDP is much larger than oI CH or SHB (actually, Iast
processes are inherently weak |11|). In practical applications, knowledge oI the
detuning characteristic is very important, so that the maximum useIul detuning is
known. ThereIore, Figs. 5.9 and 5.10 present experimental investigations oI the
conversion eIIiciency and SBR dependency on the detuning. The pump wavelength
was kept constant
P
1542 nm and the pump power was P
P
1 dBm. The detuning
is deIined here as the wavelength diIIerence between the signal and the pump:
P S
= A . It should be noticed that the conversion eIIiciency and SBR is better
Ior positive detuning |11, 16|. This is in agreement with the literature and is due to
coherent adding oI the mixing terms in (1.20) with diIIerent phases and due to the
gain dependence on wavelength. Perhaps surprisingly, Fig. 5.9 shows an improvement
in mixing eIIiciency Ior very large negative detuning. This is a measurement arteIact
caused by the high ASE Iloor in this region (SBR shows that the ASE power is
comparable with the conjugate power).
72
FWM efficiency against detuning
-20
-16
-12
-8
-4
0
-15 -10 -5 0 5 10 15
A A A A [nm]
q qq q [dB]
positive, -16.6dBm
negative, -16.6dBm
positive, -9.6dBm
negative, -9.6dBm

Fig. 5.9 Experimentally measured dependence oI the conversion eIIiciency on pump-probe detuning
SBR against detuning
0
4
8
12
16
20
24
28
32
36
-15 -10 -5 0 5 10 15
A A A A [nm]
SBR [dB]
positive, -16.6dBm
negative, -16.6dBm
positive, -9.6dBm
negative, -9.6dBm

Fig. 5.10 Experimentally measured dependence oI the SBR on pump-probe detuning
5.2.3. Linearity of CW FWM
Another Iundamental characteristic oI a wavelength converter based on FWM
in SOAs is the conversion linearity. In most typical applications it is desirable to have
the transIer Iunction P
C
(P
S
) as linear as possible, so that it does not introduce
additional power-dependent distortions. However, as brieIly mentioned in Section
5.2.1, SOA-based converters are not linear devices due to signal-induced gain
saturation. This eIIect can lead to an extinction ratio reduction and limits the input
power dynamic range. (A novel conIiguration utilizing FWM in SOAs is presented in
Section 5.4, which has the potential to actually improve the extinction ratio.) To some
extent, the linearity can be improved by using large pump powers. Fig. 5.11 shows the
CW FWM transIer curve Ior diIIerent values oI P
P
. For this measurement, the pump
73
wavelength was
P
1548.6 nm and the signal wavelength was
S
1552.4 nm. Fig.
5.12 shows the SBR measured in the same conditions. It can be seen in Fig. 5.11 that
large pump powers give a dynamic range oI reasonable linearity (when the slope oI
the relevant curve is close to 1) oI about 18 dB. Since the actual slope is slightly less
than 1, the dependence is actually sublinear. This is mostly due to residual gain
saturation induced by the signal power. Since, Irom the practical point oI view, the
conjugated signal quality is equally important, the low end oI this input dynamic
range is not preIerable (as seen in Fig. 5.12). ThereIore, SBR considerations reduce
this dynamic range by about 5 dB.
Linearity of FWM for different pump powers
-31
-29
-27
-25
-23
-21
-19
-17
-15
-13
-20 -15 -10 -5 0
P
S
[dBm]
P
C
[dBm]
Pp=-4.6dBm
Pp=-1.6dBm
Pp=1.4dBm
Pp=4.1dBm

Fig 5.11 Linearity oI FWM Ior diIIerent pump powers
Dependence of SBR on the signaI power
10
12
14
16
18
20
22
24
26
28
-20 -15 -10 -5 0
P
S
[dBm]
SBR [dB]
Pp=-4.6dBm
Pp=-1.6dBm
Pp=1.4dBm
Pp=4.1dBm

Fig. 5.12 SBR dependence on input signal power Ior diIIerent pump powers
74
5.3. Experimental verification of the existence of optically induced carrier
pulsations in SOAs
It was explained in Chapter 2 that the current theories oI FWM in SOAs
assume that the various beating-induced pulsations oI the gain medium (and hence
polarization) are responsible Ior the generation oI new Irequency components |10, 11,
16, 17|. In this section, the experimental conIirmation oI the existence oI optically
induced CDP in SOAs is reported |9|, which is an important indication that the FWM
theories are indeed correct. To the best oI the author`s knowledge, this is the Iirst
experimental veriIication oI the existence oI CDP in SOAs.
Almost every theory oI FWM assumes two types oI processes mediating wave
mixing: interband and intraband processes |15-17, 20|. Interband processes (CDP)
involve carrier transitions between the valence and conduction bands and their
response time is limited to tens oI picoseconds by the carrier injection rate. Intraband
eIIects, like spectral hole burning (SHB) and carrier heating (CH) are caused by the
modulation oI the occupation probability within a band and have much lower strength
and much shorter characteristic times (0.1-1 ps).
Even iI FWM in SOAs is well studied in the optical domain, little
experimental work has been done in the electrical domain. It should be noted that
CDP, iI present, should maniIest itselI as a measurable electrical current. We are
aware oI only one work reporting direct electrical measurements oI CDP in
semiconductor lasers |21| and none concerning SOAs. There are also works reporting
the detection oI the terahertz radiation due to FWM in SOAs |22|. At the same time,
in most analytical theories oI FWM in SOAs, the existence and Iorm oI the CDP are
simply assumed and phenomenologically introduced |10, 11, 17| (although,
admittedly, based on strong premises). As the CDP have been predicted to be the
strongest eIIect mediating FWM Ior small detunings, an experimental veriIication oI
this assumption seems important.
In this section, successIul detection oI the CDP in the RF domain is reported.
The measurements conIirm that the centre Irequency oI the CDP is determined by the
optical beat Irequency and the CDP spectrum corresponds to the convolution oI the
two beating optical spectra. Moreover, it is a coherent eIIect, due to adding Iield
amplitudes and not powers, as the CDP vanishes (as one might expect), when the two
input Iields are orthogonally polarized.
75
5.3.1. Experimental setup
The experimental setup is shown in Fig. 5.13. Two diode lasers LD1 and LD2
are temperature Iine tuned, so that the Irequency detuning between them is less than 3
GHz (limited by the bandwidth oI the RF spectrum analyzer (RFSA) used in the
experiment). An OSA was used to monitor the Irequency detuning and to help with
coarse tuning. The lasers used in the experiment were standard telecommunications
DFB signal lasers, both with nominal wavelength oI 1547.02 nm. Optical isolators
were inserted to ensure a stable lasing Irequency and to suppress any back reIlections,
particularly at new Irequencies, as they could produce new beating Irequencies. The
SOA was also temperature stabilized using a thermoelectric cooler (TEC). A DC bias
current was input to the SOA using a BiasT to uncouple it Irom the RF path.
Similarly, the RF spectrum was measured at the RF port oI the BiasT by the RFSA.
RF shielding was used to isolate the SOA and BiasT Irom external electromagnetic
interIerers. Polarization controllers were used to set the polarization states oI the
interacting waves.

Fig. 5.13 Experimental setup to measure optical beating induced CDP

Fig. 5.14 Heterodyne spectrum oI the lasers used in the experiment
76
BeIore measuring the CDP spectrum, the reIerence beating spectrum between
the two lasers was measured by heterodyning them on a Iast photodetector. The
detector used was a standard telecommunications 10 Gb/s linear receiver. Fig.5.14
shows the heterodyne spectrum. The 3 dB bandwidth oI the spectrum was equal to
about 40 MHz, which suggests that the laser linewidths were about 20 MHz
(assuming that they were identical). The spectrum has approximately a Lorentzian
shape, as expected.
5.3.2. Experimental results
Firstly, the CDP spectrum was Iound on the RFSA. It was conIirmed that the
CDP is caused by the interaction between the input optical beams as the RF signal
disappeared when one oI the lasers was switched oII. Also, the RF signal was
polarization sensitive and could be completely suppressed by adjusting the two PCs to
the orthogonal states. This is consistent with the theory oI nearly degenerate FWM
|10|. The CDP spectrum corresponding to the heterodyne spectrum Irom Fig. 5.14 is
shown in Fig. 5.15. The CDP spectrum is similar to the heterodyne spectrum, as
expected Irom the Iact that the carrier rate equation (see equation (2.7)) governing the
carrier dynamics is almost linear with the carrier concentration (a small nonlinearity is
introduced by the coupling between the rate equation and the propagation equation).
ThereIore, no new components are produced by electrical nonlinearities. A slight shiIt
in the peak Irequency compared to Fig. 5.14 is due to the slow driIt oI the laser
Irequencies in time. The narrow Irequency lines visible in Fig. 5.15 are due to the
current source and TEC controller switching Irequencies and are independent oI the
optical power. These results give a strong indication that nearly degenerate FWM is
indeed mediated by the CDP.
According to FWM theory, an additional beating is produced between the
newly generated conjugate beam and the input beams. The Irequency oI this beat tone
should be twice the Irequency oI the primary beating. Fig. 5.16 shows that this new
tone can indeed be observed. Markers M1 and M2 indicate the beating spectra. A
weak component at the second harmonic oI the detuning Irequency appeared, which
may be due to beating between one oI the input waves and the second order FWM
product (as predicted by most FWM theories) or due to higher order harmonics in the
CDP due to the asymmetry in the stimulated emission rate caused by gain saturation.
It is worth noting that the bandwidth oI the second order CDP in Fig. 5.16 is larger
77
than that oI the Iirst order CDP but the ratio oI these bandwidths appears to be larger
than 1.5 (as predicted by standard FWM theory). Although the OSA used in these
experiments did not have enough resolution to directly measure the output waves, the
Iact that CDP can produce new Irequencies by scattering the input waves was
conIirmed by observing the heterodyning spectrum. In Fig. 5.16, the peak at about
1860 MHz is contributed by the SOA controller. It was also checked that the peak at
M2 is due to optical beating and not to any electrical nonlinearities in the diode. This
was tested by injecting a third optical wave into the SOA. Only the optical beat
Irequencies were observed in this case (i.e. no additional diIIerence Irequency tones
due to electrical mixing oI the beat Irequency signals). This is consistent with the
commonly used carrier rate equation.

Fig. 5.15 RF spectrum observed with peak due to CDP observed on the RFSA

Fig. 5.16 RF spectrum with CDP peaks due to 1st and 2nd order optical beating tones
78
Further measurements were taken to check the inIluence oI the bias current
and optical power on the CDP. The dependence oI the CDP amplitude on the bias
current is shown in Fig. 5.17. The input optical power was -0.8 dBm (LD1) and -0.7
dBm (LD2). For electrical pumping less than 50 mA, the CDP were below the noise
level. The amplitude oI the CDP increased with bias current and exhibited saturation
Ior bias current oI about 150 mA. This saturation behavior reIlects the gain-current
dependence in multiple quantum well (MQW) SOAs. The largest CDP amplitude was
detected Ior 0 mA bias (cathode and anode grounded). This observation does not
mean, however, that this regime is preIerred Ior a FWM wavelength converter as the
SOA is strongly absorbing in this case. Measurements conIirm that there is no
detectable conjugate power, even though CDP are very large.
-50
-45
-40
-35
-30
-25
-20
0 50 100 150 200 250 300
I
bias
[mA]
U[dBmV]

Fig. 5.17 Dependence oI CDP amplitude U(I
bias
) on bias current
It can also be expected that the amplitude oI the CDP increases with input
power. However, Ior large input powers carrier depletion and gain saturation will
prevent a Iurther increase in the amplitude oI the CDP. Fig. 5.18 shows the
dependence oI the amplitude oI the CDP on the LD2 power Ior two values oI the LD1
power and conIirms this reasoning. The strongest CDP is produced when the two
beams have approximately equal power. When one beam is weak and the other very
strong, the beating is actually reduced due to carrier depletion.
All the results discussed so Iar were Ior co-propagating beams. Some theories
oI FWM predict that CDP also occurs between counterpropagating beams |23|.
Depending on the FWM scheme, a possible decrease in the FWM eIIiciency is
ascribed to the wavenumber mismatch included in the wave equation (not the carrier
rate equation). We have experimentally veriIied that no CDP were detectable when
79
the input beams are counterpropagating, independent oI the SOA current, input
optical power or input states oI polarization.
At this point, it is worth mentioning that even iI most theories describe FWM
in SOAs as a third order nonlinear eIIect, most semiconductor crystals are not centro-
symmetric. ThereIore, they exhibit an eIIective third order nonlinear susceptibility,
due to the combined intrinsic third order susceptibility and cascaded second order
processes. CDP, as discussed in |23| is likely to be due to the second order diIIerence
Irequency mixing (DFG), which is then Iollowed by another second order process oI
photon scattering. Externally, this sequence appears the same as a third order process.
Experimental detection oI THz radiation due to FWM |22| and scattering sidebands
generation by SOA bias current modulation |24| support this physical explanation oI
CDP-mediated FWM. Furthermore, it should be noted that in the counterpropagating
conIiguration, a Iour photon interaction leading to grating Iormation is still phase-
matched whereas DFG is always mismatched Ior counterpropagating waves. The
grating generated by FWM would be both temporally and spatially varying. Even iI it
is argued that the spatial variation in this case would be washed out by the carrier
diIIusion |25| (and it would not be detected anyhow in our experiment), the temporal
aspect should lead to CDP. ThereIore, our experiments appear to support the
possibility that the second order nonlinearity plays a signiIicant role oI in CDP.
-48
-46
-44
-42
-40
-38
-36
-34
-32
-30
-15 -10 -5 0 5 10
P
LD2
[dBm]
U [dBmV]
LD1: -0.8 dBm
LD1: 4.7 dBm

Fig. 5.18 Dependence oI amplitude oI the CDP on the optical input power
The experimental results reported in this section suggest the Iundamental
assumption oI the existence oI the CDP due to beating between the input waves,
mediating FWM in SOAs, is correct. However, they also show that the usual
80
explanation oI CDP as being due to an intrinsic third order nonlinearity may not be
complete, as second order processes seem to play an important role.
5.4. Extinction ratio improvement using nonlinear four-wave mixing in SOAs
with assist beam
It was explained in Section 5.2 that the conjugate wave can have a poorer
extinction ratio than the signal wave, due to signal induced gain saturation. In this
section, a novel scheme is presented |12|, which actually allows Ior an extinction ratio
improvement, and thereIore can be used as a simple 2R (reshape and re-ampliIy)
regenerator.
All-optical signal regeneration is believed to be a key tool Ior Iuture
transparent optical networks and is being researched extensively |4, 6, 26|. The most
advanced optical regenerators are classiIied as 3R devices, as they are able to perIorm
re-ampliIication, reshaping and retiming. However, these capabilities come at a price,
as these devices are usually complicated and bit rate and Iormat speciIic, as the
retiming section is based on some clock extraction procedure. On the other hand,
simpler 2R regenerators may be suIIicient in some network scenarios and can be
made bit rate independent. In this section, it is experimentally shown that a simple
scheme based on nonlinear FWM with an assist beam at transparency can yield
extinction ratio (ER) improvement and that the device can be adjusted to a range oI
input signal powers. (Clearly any FWM is a nonlinear optical process. Here,
'nonlinear is used in the system application sense, when the output signal power is
proportional to the square oI the input signal power.)
It was shown in |27| that nonlinear FWM can indeed improve the ER.
However, in that work a non return-to-zero (NRZ) signal was mixed with a locally
generated return-to-zero (RZ) pulse stream, so that the output ER was Iurther
improved by using sharp RZ pulses and in eIIect NRZ to RZ Iormat conversion was
obtained. The need Ior a locally generated RZ stream (optical clock) is
disadvantageous, as it adds complexity and reduces the Ilexibility oI the scheme. A
similar setup was investigated in |28|, without the use oI the RZ stream, i.e. the NRZ
signal was mixed with a C-band (1550 nm region) CW beam. There it was concluded
that in order to obtain an ER improvement, the signal power needs to be in a narrow
range, such that the zero` power level is less than the C-band CW power, whereas the
one` power level is not much greater than the C-band CW power. In that case, an ER
81
improvement oI 3 dB was reported, but this was at the expense oI a small input
dynamic range and the need Ior tightly controlled input power levels.
Here, the use oI an additional assist beam located close to the transparency
point to Iurther improve ER is proposed and experimentally investigated. Assist
beams have been used beIore to improve the gain saturation |29| and FWM linearity
and equalize the conversion eIIiciency |30, 31|. In this section, Ior the Iirst time to our
knowledge, the Ieasibility oI using the assist beam in a nonlinear FWM conIiguration
is proposed and numerically investigated. The inIluence oI the C-band CW wave
power on the perIormance oI the regenerator is also examined. It is shown here that
the requirement in |28| that the input signal levels to be smaller (Ior the zero` level)
and larger (Ior the one` level) than the CW power can be relaxed by using a high
power C-band CW wave.
5.4.1. Experimental setup and operating principle
The core experimental setup is shown in Fig. 5.19. This setup was
appropriately modiIied to measure diIIerent physical quantities by changing the input
signal conditions and output measurement instruments. The transmitter LD1 at 1551
nm was either a CW wave (when measuring the gain saturation curve and static
transIer Iunction oI the regenerator) or an NRZ 4 Gb/s signal (when measuring ER)
and in what Iollows, its power is denoted as P
S
. Laser LD2 produces a CW wave at
1547.72 nm (to be mixed with the signal P
S
) and its power is denoted P
P
. As
described in Section 5.2, the FWM process produces two Iirst order waves: the
conjugate wave P
C
at 1544.44 nm and the satellite wave P
L
at 1554.32 nm. The
Irequency oI P
C
is related to the input waves Irequencies f
S
, f
P
by f
C
2f
P
- f
S
(compare
Fig. 5.6). Similarly, the Irequency oI P
L
is given by f
L
2f
S
f
P
. In the quantum-
mechanical picture, this means that 2 photons Irom P
P
and 1 photon Irom P
S
are
required to generate a new photon at f
C
and vice versa Ior f
L
. Because oI this
relationship, P
C
is linear with signal power P
S
(within the range where signal
contribution to the gain saturation is small, see Fig. 5.11) and P
L
depends on the
square oI signal power.
82

Fig. 5.19 Experimental setup; LD - laser diode, CRC - circulator
Under the assumption that the pump is not depleted, the output power oI the
conjugate wave is given by |19, 20|
P
C
R(Ae)G
S
(G
P
-1)
2
P
P
2
P
S
(5.1)
where R(Ae) includes all the relevant processes contributing to FWM, as discussed in
Chapter 2. (These processes include most importantly carrier density pulsations,
spectral hole burning and carrier heating |11, 16, 20|). G
S,P
are the gain seen by the
signal and pump input waves, respectively. For small detuning, the gain is
approximately the same at the two Irequencies G
P
G
S
G and the output power
exhibits a cubic dependence on the gain. This is the so-called G
3
theory |16, 19, and
20|. II the gain were constant (no saturation), the conjugate power would be
proportional to the signal power and thereIore no signal regeneration is possible. In
Iact, as experimentally shown in Section 5.2, gain saturation can cause an ER
reduction. However, the satellite power has a quadratic dependence on the signal
power (nonlinear exponent 2, see below), as predicted by the quantum-mechanical
description, iI the gain saturation is negligible. This gives the theoretical limit Ior the
output ER, which is twice the input ER (in dB). In practice, however, the achievable
improvement is smaller. This is Iirst oI all because the gain saturation eIIect is larger
Ior high input powers (i.e. one` level) than Ior low power (i.e. zero` level). The
second reason is the high ASE level Ior the uncompressed gain, which aIIects the
zero` level. ThereIore, in practice the ER gain oIten depends on the input ER value.
In |27| a maximum ER gain oI 5 dB was reported Ior input ER oI 7 dB. However, this
was reduced to 2.5 dB Ior an 11 dB input ER. In |28|, about 3 dB improvement in ER
was obtained. Taking the aIorementioned eIIects into account, the experimental
dependence oI P
L
on P
S
can be phenomenologically expressed as
83
P
L
A(Ae, P
P
)P
S

(5.2)
A method to reduce the gain saturation (and thereIore increase ) is to apply
an assist beam (AB) |29-31|. The high-power AB is usually located close to the
transparency point (when the absorption and gain seen by the beam balance each
other) so that it does not change the net gain oI the device. However, it reduces the
eIIective carrier liIetime and thereIore improves (increases) the saturation power oI
the ampliIier (as the saturation power oI the SOA is proportional to the reciprocal oI
the carrier liIetime) |29|. Additional beneIits are reduced pattern eIIects and lowered
ASE noise level. For these reasons, it is oI interest to combine the AB method with
the nonlinear FWM conversion, as this scheme has the potential to oIIer better
perIormance (i.e. better nonlinear exponent) than that reported in |28|, by reducing
the gain saturation. In our experiments, a 1486 nm pump laser was used to produce
the assist beam. The SOA characterized in Sections 5.1 and 5.2 was also used in these
experiments. The transparency bias Ior this AB wavelength was Iound to be 155 mA.
However, even Ior high bias (300 mA), when the main AB peak was in the gain
region, the gain compression due to the AB was less than 2 dB.
5.4.2. Experimental verification of regenerative capabilities of the
nonlinear FWM ,
To conIirm that the use oI the AB improves the gain characteristics, gain
saturation curves were measured. P
S
was varied by changing the attenuation. Fig. 5.20
shows the gain saturation curve Ior two diIIerent AB powers (21.1 dBm and
16.7 dBm at the SOA input) and Ior diIIerent bias currents I
SOA
(155 mA and 300
mA). Another measurement was perIormed in the presence oI strong P
P
beam oI 1.5
dBm (measured at the SOA input), to investigate the eIIect oI P
P
power and to Iind
the optimum conditions. For this measurement, the polarization controllers were
adjusted so that there was no FWM between the two waves in 1550 nm region. As
expected, the presence oI the AB enhances the saturation power and the eIIect is most
visible Ior low electrical pumping. Depending on the position oI the AB with respect
to the transparency point, the assist light has a diIIerent eIIect on the gain. In
particular, it is seen that it increases the gain when a strong beam P
P
is present, as the
carrier population is depleted in this case. Although a large P
P
signiIicantly reduces
the gain oI the SOA, it also clamps it. For an output power P
S
ranging Irom -15 dBm
84
to 12 dBm, the gain is reduced by only 3 dB. This corresponds to an input signal
dynamic range oI about 30 dB (although the low end oI the dynamic range cannot be
used Ior signal conversion).
10
15
20
25
30
35
-20 -16 -12 -8 -4 0 4 8 12 16 20
1551 signal output power [dBm]
f
i
b
e
r

t
o

f
i
b
e
r

g
a
i
n

[
d
B
]

w/o AB
AB 16.7 dBm
AB 21.1 dBm
soa=155 mA, w/o Pp
soa=300 mA, w/o Pp
soa=300 mA, Pp=1.5 dBm

Fig. 5.20 Gain saturation curves Ior diIIerent assist beam (AB), bias current and C-band pump power
conditions. Combination oI AB and large C-band power gives best gain linearity.
To experimentally veriIy the regeneration capabilities oI this scheme, the
static transIer Iunction oI the regenerator was measured. Fig. 5.21 shows the output
satellite power as a Iunction oI input signal power. To obtain FWM, the polarization
controllers were adjusted so that both input waves P
S
, P
P
have the same polarization.
The tunable Iilter was adjusted to pass the satellite beam. Two cases were
investigated; Ior P
P
-10 dBm and P
P
1.5 dBm. In the Iirst case, the device has
much larger gain, but the gain also saturates Iaster (see Fig.5.20). In the other case,
the gain is reduced but is clamped by the large P
P
power. The transIer Iunctions were
measured in the presence and absence oI the AB. It should be noticed Irom the curves
that the assist light improves the operation oI the regenerator in each case, by
increasing both the gain and the saturation power. Low P
P
is not a Iavourable
operating condition, even though it oIIers much better conversion eIIiciency Ior low
input signal powers, since the gain saturation eIIects are much larger. This is visible
Irom the Ilattening oI the curves Ior large input signal and Irom low value oI the
nonlinear exponent . In the -17 dBm to 8 dBm range (quasi-linear part oI the low P
P

curves in Fig. 5.21), , as Iitted to the data, was about 0.7 dBm/dBm. However, when
P
P
1.5 dBm, was as large as 1.25 and 1.41 dBm/dBm, without and with AB,
85
respectively Ior signal power in the range -12 dBm to -1 dBm. The curves Ilatten out
as the signal power approaches the pump power P
P
, but this point is shiIted towards
large input powers. The Ilat low end oI the transIer curve originates Irom the ASE
Iloor, which also decreases . As discussed beIore, a large value is Iavourable Ior a
regenerator, as this means a steeper transIer Iunction and higher ER gain.
-30
-25
-20
-15
-10
-5
-30 -25 -20 -15 -10 -5 0
1551 signal input power [dBm]
s
a
t
e
l
l
i
t
e

o
u
t
p
u
t

p
o
w
e
r

[
d
B
m
]
w/o AB
AB 21.1 dBm
Pp= -10 dBm
Pp=1.5 dBm

Fig. 5.21 Static transIer Iunction oI the regenerator. Large C-band pump power provides higher
nonlinear exponent . Assist beam additionally improves .
The nonlinear exponent in the log scale is given by 3o2 (as seen Irom
(5.1), assuming that the gain seen by the two mixing beams is equal), where o is the
gain nonlinear exponent and describes the dependence oI the gain on input power. For
a perIectly linear ampliIier o0. The linear Iitting oI the appropriate gain curve in the
region oI interest gives the gain exponent o to be about -0.15 dB/dBm, which is Iairly
consistent with the observed value oI the nonlinear exponent oI 1.41 dBm/dBm.
There is also another eIIect, which is not included in the G
3
theory, but is well-known
and plays a role in decreasing . That is that the ASE level decreases with increasing
input power. Including this eIIect by subtracting the variable ASE background
increases to about 1.6 dBm/dBm and extends the range oI constant down to at
least -20 dBm input power. Obviously, this can only be done numerically and cannot
be used to improve the practical device, as in practice the in-band ASE cannot be
separated Irom the signal.
86
Another important parameter Ior a converter or regenerator is the signal-to-
background ratio (SBR). Fig. 5.22 shows the SBR corresponding to the cases in
Fig.5.21 to discuss their practical applications. The qualitative Ieatures oI the SBR
curves are similar to those in Fig.5.8; with some quantitative diIIerences are due to the
diIIerence in power and detuning. It can be seen that the AB improves the SBR
slightly by increasing the gain. The higher gain is also responsible Ior the improved
SBR at low input powers Ior P
P
-10 dBm. In linear region oI the transIer curves in
Fig. 5.21, the SBR is suIIicient to ensure good signal quality and values oI SBR
higher than 25 dB are obtainable. The SBR can be Iurther improved by reducing the
detuning (see Fig. 5. 10).

Fig. 5.22 Experimentally measured signal to background ratio (SBR) oI the investigated converter
Finally, the ER oI the satellite wave was measured. In this experiment, the
power meter PM shown in Fig.5.19 was replaced by a photodiode and a Iast sampling
oscilloscope. The conditions were arranged to correspond to the nonlinear exponent oI
1.41 dBm/dBm (P
P
1.5 dBm, P
AB
21.1 dBm). The results are presented in Fig. 5.23
Ior two cases, where the input zero` levels are -12 dBm and -9 dBm, respectively.
Clearly, a signiIicant ER improvement is possible and in the best case almost 3 dB
improvement was achieved. The results show also that the improvement in ER
decreases slightly as the input power increases, indicating that the dynamic transIer
Iunction does not Iully overlap with the static transIer Iunction. This may be caused
by the carrier recovery time, even though the carrier liIetime is short, due to strong
electrical pumping and large input optical power. The use oI the AB has an additional
87
beneIicial eIIect in that the FWM eIIiciency varies slower with the detuning as
reported in the literature |30, 31|. ThereIore, the sensitivity oI the regenerator to the
signal wavelength is reduced, Iurther improving the practicality oI the method. This is
another advantage oI our scheme, as compared with the schemes in |27, 28|.

Fig. 5.23 Experimentally measured output ER versus input ER
It was experimentally demonstrated that the use oI an AB can indeed improve
the regenerative capabilities oI the nonlinear FWM scheme. The proposed regenerator
is Iairly simple. Although it requires relatively high signal powers, these can be
achieved by combining it with a preampliIier.
5.5. Summary
In this chapter, CW FWM in SOAs was experimentally investigated. For the
Iirst time to our knowledge, the optically induced CDP were directly measured in the
RF domain, thus providing a strong evidence Ior the validity oI the FWM theories.
However, it was discussed that, contrary to most theories oI FWM in SOAs, cascaded
second order nonlinear processes may also play an important part in generating the
conjugate wave. Based on the understanding oI the properties oI FWM in SOAs, a
novel simple 2R regenerator was proposed and experimentally assessed. The
measurements presented in this chapter also provide a solid grounding Ior the
investigations discussed in Chapter 6, where optical sampling using FWM in SOAs is
considered.
88
5.6. References
|1| H. J. S. Dorren, X. Yang, A. K. Mishra, Z. Li, H. Ju, H. de Waardt, G.-D.
Khoe, T. Simoyama, H. Ishikawa, H. Kawashima, and T. Hasama, "All-optical
logic based on ultraIast gain and index dynamics in a semiconductor optical
ampliIier," IEEE J. Sel. Topics Quantum Electron., vol. 10, pp. 1079-1092,
2004.
|2| C. Schubert, R. Ludwig, and H.-G. Weber, "High-speed optical signal
processing using semiconductor optical ampliIiers," J. Opt. Fiber Commun.
Reports, vol. 2, pp. 171-208, 2005.
|3| L. H. Spiekman, "Semiconductor optical ampliIiers," in Optical Fiber
Telecommunications, vol. IVA, pp. 699-731, I. Kaminov and T. Li, Eds.,
Academic Press, 2002.
|4| O. Leclerc, B. Lavigne, D. Chiaroni, and E. Desurvire, "All-Optical
Regeneration: Principles and WDM Implementation," in Optical Fiber
Telecommunications, vol. IVA, pp. 732-783, I. Kaminov and T. Li, Eds.,
Academic Press, 2002.
|5| M. Usami and K. Nishimura, "Optical nonlinearities in semiconductor optical
ampliIier and electro-absorption modulator: their applications to all-optical
regeneration," J. Opt. Fiber Commun. Reports, vol. 2, pp. 497-529, 2005.
|6| O. Leclerc, B. Lavigne, E. BalmeIrezol, P. Brindel, L. Pierre, D. Rouvillain,
and F. Seguineau, "Optical regeneration at 40 Gb/s and beyond," J. Lightw.
Technol., vol. 21, pp. 2779-2790, 2003.
|7| L. A. Jiang, E. P. Ippen, U. Feiste, S. Diez, E. Hilliger, C. Schmidt, and H.-G.
Weber, "Sampling pulses with semiconductor optical ampliIiers," IEEE J.
Quantum Electron., vol. 37, pp. 118-126, 2001.
|8| C. Schmidt-Langhorst and H.-G. Weber, "Optical sampling techniques," J.
Opt. Fiber Commun. Reports, vol. 2, pp. 86-114, 2005.
|9| M. P. Dlubek, S. Kaunga-Nyirenda, A. J. Phillips, S. Sujecki, and E. C.
Larkins, "Experimental veriIication oI the existence oI optically induced
carrier pulsations in SOAs," to be submitted for publication
|10| G. P. Agrawal, "Population pulsations and nondegenerate Iour-wave mixing in
semiconductor lasers and ampliIiers," J. Opt. Soc. Am. B, vol. 5, pp. 147,
1988.
89
|11| A. Uskov, J. Mork, and J. Mark, "Wave mixing in semiconductor laser
ampliIiers due to carrier heating and spectral-hole burning," IEEE J. Quantum
Electron., vol. 30, pp. 1769-1781, 1994.
|12| M. P. Dlubek, A. J. Phillips, and E. C. Larkins, "Extinction ratio improvement
using nonlinear Iour-wave mixing in SOAs with assist beam," Microwave Opt.
Technol. Lett., vol. 50, pp. 2079-2083, 2008.
|13| G. P. Agrawal and N. K. Dutta, Semiconductor lasers, 2 ed., Van Nostrand
Reinhold, 1993.
|14| T. Suhara, Semiconductor Laser Fundamentals, 1 ed., Marcel Dekker, Inc.,
2004.
|15| M. A. SummerIield and R. S. Tucker, "Frequency-domain model oI multiwave
mixing in bulk semiconductor optical ampliIiers," IEEE J. Sel. Topics
Quantum Electron., vol. 5, pp. 839-850, 1999.
|16| I. Koltchanov, S. Kindt, K. Petermann, S. Diez, R. Ludwig, R. Schnabel, and
H. G. Weber, "Gain dispersion and saturation eIIects in Iour-wave mixing in
semiconductor laser ampliIiers," IEEE J. Quantum Electron., vol. 32, pp. 712-
720, 1996.
|17| A. Mecozzi, "Analytical theory oI Iour-wave mixing in semiconductor
ampliIiers," Opt. Lett., vol. 19, pp. 892, 1994.
|18| M. A. SummerIield and R. S. Tucker, "Optimization oI pump and signal
powers Ior wavelength converters based on FWM in semiconductor optical
ampliIiers," IEEE Photon. Technol. Lett., vol. 8, pp. 1316-1318, 1996.
|19| S. Diez, C. Schmidt, R. Ludwig, H. G. Weber, K. Obermann, S. Kindt, I.
Koltchanov, and K. Petermann, "Four-wave mixing in semiconductor optical
ampliIiers Ior Irequency conversion and Iast optical switching," IEEE J. Sel.
Topics Quantum Electron., vol. 3, pp. 1131-1145, 1997.
|20| J. Zhou, N. Park, J. W. Dawson, K. J. Vahala, M. A. Newkirk, and B. I.
Miller, "EIIiciency oI broadband Iour-wave mixing wavelength conversion
using semiconductor traveling-wave ampliIiers," IEEE Photon. Technol. Lett.,
vol. 6, pp. 50-52, 1994.
|21| R. Nietzke, P. Panknin, W. Elsasser, and E. O. Gobel, "Four-wave mixing in
GaAs/AlGaAs semiconductor lasers," IEEE J. Quantum Electron., vol. 25, pp.
1399-1406, 1989.
90
|22| S. HoIImann, M. HoImann, E. Brundermann, M. Havenith, M. Matus, J. V.
Moloney, A. S. Moskalenko, M. Kira, S. W. Koch, S. Saito, and K. Sakai,
"Four-wave mixing and direct terahertz emission with two-color
semiconductor lasers," App. Phvs. Lett., vol. 84, pp. 3585-3587, 2004.
|23| R. Paiella, G. Hunziker, U. Koren, and K. J. Vahala, "Polarization-dependent
optical nonlinearities oI multiquantum-well laser ampliIiers," IEEE J. Sel.
Topics Quantum Electron., vol. 3, pp. 529-540, 1997.
|24| D. George, H. Fews, and M. McCall, "Enhanced wide-bandwidth wavelength
conversion via IeedIorward photomixing in semiconductor laser ampliIiers,"
IEEE Photon. Technol. Lett., vol. 9, pp. 49-51, 1997.
|25| F. Favre and D. Le Guen, "Four-wave mixing in traveling-wave
semiconductor laser ampliIiers," IEEE J. Quantum Electron.,, vol. 26, pp.
858-864, 1990.
|26| D. Forin, F. Curti, G. Tosi-BeleIIi, F. Matera, A. Reale, S. Betti, S.
Monterosso, A. Fiorelli, M. Guglielmucci, and S. Cascelli, "All Optical 3R
Regeneration and Wavelength Convertion," in Optical Networks and
Technologies, pp. 537-544, K.-I. Kitayama, F. Masetti-Placci, and G. Prati,
Eds., Springer, 2005.
|27| C. Gosset and G.-H. Duan, "Extinction ratio improvement and wavelength
conversion based on Iour-wave mixing in a semiconductor optical ampliIier,"
Photonics Technologv Letters, IEEE, vol. 13, pp. 139-141, 2001.
|28| H. Simos, A. Argyris, D. Kanakidis, E. Roditi, A. Ikiades, and D. Syvridis,
"Regenerative properties oI wavelength converters based on FWM in a
semiconductor optical ampliIier," IEEE Photon. Technol. Lett., vol. 15, pp.
566-568, 2003.
|29| J. L. Pleumeekers, M. Kauer, K. Dreyer, C. Burrus, A. G. Dentai, S. Shunk, J.
Leuthold, and C. H. Joyner, "Acceleration oI gain recovery in semiconductor
optical ampliIiers by optical injection near transparency wavelength," IEEE
Photon. Technol. Lett., vol. 14, pp. 12-14, 2002.
|30| S.-L. Lee, P.-M. Gong, and C.-T. Yang, "PerIormance enhancement on SOA-
based Iour-wave-mixing wavelength conversion using an assisted beam,"
IEEE Photon. Technol. Lett., vol. 14, pp. 1713-1715, 2002.
|31| D.-Z. Hsu, S.-L. Lee, P.-M. Gong, Y.-M. Lin, S. S. W. Lee, and M. C. Yuang,
"High-eIIiciency wide-band SOA-based wavelength converters by using dual-
91
pumped Iour-wave mixing and an assist beam," IEEE Photon. Technol. Lett.,
vol. 16, pp. 1903-1905, 2004.


92
6. Optical sampling using FWM in SOAs
As discussed in Chapter 3, optical sampling has recently attracted a lot oI
interest as a promising method Ior optical perIormance monitoring |1-5|, especially to
monitor very Iast signals (above 10 Gb/s), as an adequately high speed electronic
solution is prohibitively expensive. Moreover, the problem may become even more
diIIicult when 100 Gb/s signals, which is the bit rate currently actively researched and
very likely to become the next standardized transmission speed, either Ior next
generation oI Ethernet or Optical Transport Networks G.709 (although diIIerent
multilevel modulations and subcarrier multiplexing are considered to alleviate the
pressure on electronics) |6, 7|, are introduced. DiIIerent methods to perIorm optical
sampling are brieIly discussed in Chapter 3. Techniques based on SOAs are amongst
the most promising |3, 8|, combining relative simplicity (however, it should be
remembered that the degree oI complexity depends on the particular technique), low
input power requirements (due to the high SOA internal gain) and compactness.
Within the scope oI this thesis, the Ieasibility oI optical sampling Ior optical signal
perIormance monitoring using FWM in SOAs has been investigated. In this chapter,
theoretical analysis and experimental results are reported. These results are compared
to the theoretical conversion and noise models, which provide design guidelines and
explain the limits oI the method. The rest oI this chapter is structured as Iollows: in
the next section analytical predictions oI picosecond FWM in SOAs are discussed
based on the model outlined in Chapter 2, then the noise analysis oI the sampling
system is given. Finally, experimental results are reported and discussed.
6.1. Analytical model of picosecond FWM in SOAs
The FWM model utilized in this thesis has been brieIly outlined in Chapter 2
and is based on ReIs. |9-11|. The analytical equations Ior the evolution oI pump,
probe and conjugate have been derived using small signal analysis introduced by
ShtaiI and Eisenstein |10|. Thanks to the inclusion oI intraband eIIects mediating
FWM, the model is valid Ior large detunings. The modiIied gain equation
adiabatically accounts Ior ultraIast nonlinear gain saturation; hence the model is also
valid Ior pulses much shorter than the carrier liIetime. As discussed in |11|, the range
oI validity oI the model includes pulses longer than 1 ps. As described later in this
Chapter, this regime matches the experimental conditions employed in the present
93
work. The constituent equations oI the model are given in Chapter 2; here they are
repeated Ior convenience:
)
`

|
|
.
|

\
|
+
+

+ +
=
dt
t dP
h
P
h t P
h h
t P e dt
dh
SHB CH
sat
SHB CH
h
SHB CH
) , 0 (
) 1 ) )(exp( (
1 ) (
) 1 ) )(exp( , 0 (
) , 0 ( ) ( 1
1
0
0 0
0
0
0
0
c c
t t
c c
t c c
(6.1)
(
(

O +

+
O +

+
O + +

= O
SHB
sat SHB SHB
CH
sat CH CH
sat
h
h
f
P f
f
P f
f P A e
f e
F
t
c |
t
c |
t
|
1
) 1 (
1
) 1 (
, , 1
1
2
1
) (
0
2
1
(6.2)
| | ) , 0 ( , , ) ( 1 ) , (
1
2
2
) 1 )( ( 2 / 1
1
t A A F e t L A
f t h
O + =
|
(6.3 a)
| | ) , 0 ( , , ) ( 1 ) , (
2
2
1
) 1 )( ( 2 / 1
2
t A A F e t L A
f t h
O + =
|
(b)
| | ) , 0 ( ) ( ) , (
*
2
2
1
) 1 )( ( 2 / 1
3
t A A F e t L A
f t h
O =
|
(c)
where (6.1) is the gain rate equation dependent on the total input power P
0
(0,t), (6.2)
is the mixing Iunction (phenomenologically describing the relevant eIIects) and (6.3a-
c) are the propagation equations Ior the three interacting Iields: pump, signal and
conjugate, respectively. The Iorm oI (6.3a-c) assumes that the dominant eIIects are
XGM and Iirst order FWM, which is an experimentally conIirmed assumption. In the
conIiguration suitable Ior optical sampling, the pump wave is given by the very short
sampling pulses and the conjugate thereIore consists oI the optical samples Irom the
signal beam.
6.1.1. Theoretical conversion efficiency
Based on the model given by the equations (6.1)-(6.3), the sampling process
was theoretically investigated and analyzed by means oI numerical simulations
(required to solve the gain equation (6.1)), in order to determine the optimal sampling
conditions. The propagation equation oI the most interest here is (6.3c), as it describes
the evolution oI the conjugate, i.e. the optical samples. From the practical point oI
view, it is important to achieve as high a conversion eIIiciency p as possible. The
conversion eIIiciency is easily deIined when FWM between CW waves is considered,
as the ratio oI the output conjugate power to the input signal power
) 0 ( ) (
2 3
P L P
CW
= q . Similarly, Ior FWM between pulses, the eIIiciency is deIined as
94
the output conjugate pulse energy to the input signal pulse energy. Since in the case
considered here the signal was a CW laser and the conjugate was a stream oI short
pulses (optical samples), none oI these deIinitions can be directly applied. The
convention adopted in this thesis is that the conversion eIIiciency is given by the ratio
oI the output conjugate energy E
3
(L) to the input signal energy E
2
(0) contained in the
time window under the sampling pulses. ThereIore, the signal energy is given by the
integral oI the signal power P
2
(0,t) times the gating Iunction f(t) (related to the
sampling pulse shape) and the conversion eIIiciency is given by:
}
= dt t f t P L E ) ( ) , 0 ( ) (
2
2 3
q (6.4)
In the experiments reported later in this chapter, the signal was simply a CW
wave Irom a tunable DFB laser Bookham 0224AP. The laser was tunable over the
whole C-band to the Irequencies speciIied by the standard ITU-T 50GHz channel
grid. The sampling pulses were produced by a Pritel FFL laser, which is an all-Iibre,
passively modelocked, soliton laser. The pulse duration depends on the internal
optical Iilter installed, and can range Irom about 2.5 ps to about 8 ps. Since it is a
soliton laser, pulses are close to being transIorm-limited. This issue is important Irom
the practical point oI view, as the FWM eIIiciency strongly depends on the detuning
and Ior best eIIiciency the detuning should be as small as possible. Short sampling
pulses inherently possess a broad bandwidth, but any extra chirp would lead to an
additional broadening oI the spectrum. The shape oI the sampling pulses is close to
sech
2
shape, as expected Ior Iibre solitons |12|. This shape was assumed in the
simulations. The laser repetition rate was 88.43 MHz. The sampling laser was
continuously tunable within a similar range oI Irequencies as the CW signal laser. In
the simulations presented below, conditions matched the actual signal and sampling
(pump) source parameters.
The SOA used in the experiments was the MQW device Irom CIP, which was
extensively characterized in Chapter 5. The SOA parameters used in simulations were
typical Ior InGaAsP-based SOAs (similar to the device used in experiments) and are
taken Irom the literature |9, 11, 13|:
- carrier liIetime t
0
200 ps,
- saturation power P
sat
11 mW,
- linewidth enhancement Iactor 4,
- SHB gain compression Iactor c
SHB
0.35 W
-1
,
95
- CH gain compression Iactor c
CH
0.7 W
-1
,
- SHB linewidth enhancement Iactor
SHB
0.1,
- CH linewidth enhancement Iactor
CH
3,
- SHB characteristic time t
SHB
70 Is,
- CH characteristic time t
CH
700 Is.
Although, as noted in |10, 14|, the parameters describing intraband processes
in SOAs cannot be determined unanimously Irom the experimental data (due to
complex interaction between these processes), the values used in the simulations are
those most oIten encountered in the literature. The unsaturated SOA gain G
0
used in
most simulations was 30 dB, which is similar to the value experimentally measured,
as reported in Chapter 5.
6.1.2. Dependence of the conversion efficiency on the sampling pulse
width and sampling pulse energy
It is well known, that in the context oI FWM between short pulses there is an
optimum pump energy, which maximizes the conversion eIIiciency |9-11|. This
optimum is due to the Iact that, on one hand, mixing strength grows with the pump
power (see (6.3c)) and, on the other hand, strong pump-induced gain saturation limits
the ampliIication available Ior the conjugate. Since the peak power oI the ultrashort
pulses can be very high even Ior very modest average powers, the same mechanism
can be expected to play a similar role in the mixing between pulsed pump and CW
signal. Related to this is the dependence oI the conversion eIIiciency on the sampling
pulse width t
p
(deIined as the Iull-width at halI maximum (FWHM)), as Ior the same
pulse energy, shorter pulses have larger peak power. Figs. 6.1, 6.2 and 6.3 show the
dependence oI the conversion eIIiciency on the pump pulse energy E
p
Ior sampling
pulses with durations oI 3 ps, 6 ps and 12 ps, respectively. Additionally, to show the
eIIect oI gain saturation due to signal power, in each case 3 diIIerent signal powers
are used as the parameter.

96
Conversion efficiency against pump energy
-10
-7
-4
-1
2
5
8
11
14
1.E-04 1.E-03 1.E-02 1.E-01 1.E+00
E
p
[pJ]
q qq q [dB]
0.01mW
0.1mW
1mW
P
signal

Fig. 6.1 Conversion eIIiciency against pump energy, t
p
3 ps, G
0
30 dB, detuning 0.5 THz
As seen in Figs. 6.1-6.3, in the unsaturated regime p improves with increasing
pump energy and the curves Ilatten out as the gain saturation become signiIicant.
Optimal conversion eIIiciency is achieved Ior a broad range oI sampling pulse
energies. Although the peak oI the eIIiciency curves does depend on t
p
, the exact
value oI the optimum E
p
is not critical due to the Ilat nature oI the curves and E
p

0.1 pJ seems to be suIIicient in most practical conditions. Similar shape oI the
eIIiciency curves has been reported Ior FWM between short pulses |9, 11|. It is also
worth noting that p changes with signal power, indicating a nonlinear relationship
between the conjugate energy and input power. This is oI course due to the gain
saturation caused by signiIicant signal CW power. Although in the literature, when
considering FWM between short pulses, conversion eIIiciency is usually considered
as the most important Iigure oI merit, in the case investigated here it might be slightly
misleading to rely only on this quantity to design the experimental setup. This is
because the input signal is a CW beam and the deIinition oI the conversion eIIiciency
adopted here (see equation (6.4)) is dependent on sampling pulse width, thereIore Ior
longer sampling pulses the conjugate energy can be larger, even iI p is smaller. For
example, Ior P
2
(0) 0.2 mW, the maximum conjugate energy is 2.4 IJ, 3 IJ and
3.3 IJ, Ior t
p
3 ps, 6 ps and 12 ps, respectively. As discussed later in this Chapter,
the absolute value oI the conjugate energy is important to achieve a good signal-to-
noise ratio (SNR) aIter detecting the optical samples.

97
Conversion efficiency against pump energy
-10
-7
-4
-1
2
5
8
11
14
1.E-04 1.E-03 1.E-02 1.E-01 1.E+00 1.E+01
E
p
[pJ]
q qq q [dB]
0.01mW
0.1mW
1mW
P
signal

Fig. 6.2 Conversion eIIiciency against pump energy, t
p
6 ps, G
0
30 dB, detuning 0.5 THz
Conversion efficiency against pump energy
-10
-7
-4
-1
2
5
8
11
1.E-03 1.E-02 1.E-01 1.E+00 1.E+01
E
p
[pJ]
q qq q [dB]
0.01mW
0.1mW
1mW
P
signal

Fig. 6.3 Conversion eIIiciency against pump energy, t
p
12 ps, G
0
30 dB, detuning 0.5 THz
6.1.3. Dependence of the conjugate energy on the SOA gain
It can be expected that the conjugate energy E
c
will increase with the
unsaturated ampliIier gain. The exact relationship is perhaps less obvious, but its Iorm
is oI practical interest as the ASE power is also gain-dependent. As discussed in the
noise analysis presented later in this Chapter, the conjugate-ASE beating limits SNR
oI the received samples. To understand the qualitative dependence oI p on the gain to
optimize the experimental setup Ior best SNR, Fig. 6.4 shows the Iunction E
c
(G
0
) Ior
t
p
6 ps and P
2
(0) 0.2 mW. It is clear that to get a usable conjugate energy, high
gain is preIerred.
98
Conjugate energy versus SOA gain
1.E-03
1.E-02
1.E-01
1.E+00
1.E+01
0 5 10 15 20 25 30 35
G
0
[dB]
E
c
[fJ]

Fig. 6.4 Conjugate energy against SOA gain G
0
, t
p
6 ps, P
2
(0)0.2 mW, detuning 0.7 THz
6.1.4. Dependence of the conjugate energy on the input signal power
When discussing the inIluence oI the pump energy on the conversion
eIIiciency, earlier it was mentioned brieIly that the contribution Irom the signal power
to gain saturation causes the conjugate energy to be nonlinearly dependent on the
signal power. Fig. 6.5 shows this dependence Ior 6 ps sampling pulses. An SOA gain
oI 30 dB was used to obtain the curve.
Conjugate energy vs input power
0
1
2
3
4
5
6
0 0.1 0.2 0.3 0.4 0.5 0.6
P
2
[mW]
E
c
[fJ]

Fig. 6.5 Conjugate energy against input signal power, t
p
6 ps, detuning 0.7 THz
The relationship E
c
(P
2
(0)) is clearly nonlinear. This nonlinear dependence is
obviously undesirable, especially iI the optical sampling system is used to
characterize an unknown signal. Moreover, the dependence oI the sample energy on a
multitude oI parameters renders correcting Ior this nonlinear behavior diIIicult in
99
practice. This is the price paid Ior the relatively high conversion eIIiciency due to the
internal SOA gain. Sampling systems based on passive nonlinear devices do not
usually have this limitation |2, 15| unless the signal power is so high that it causes
nonlinear losses |16, 17|, which is not a practical scenario. To improve the linearity oI
the discussed sampling scheme, it has been proposed to use the so-called gain
transparent conIiguration, where conjugate and signal wavelength are outside the
SOA gain region |18|. This approach however compromises the conversion
eIIiciency, as the conjugate does not receive any ampliIication. A slight improvement
in the linearity can be achieved by using the AB approach, investigated in Chapter 5
in the context oI signal regeneration, but simulations perIormed show that the
improvement is marginal.
6.1.5. Dependence of the conversion efficiency on the frequency detuning
between the signal and sampling pulses
As brieIly mentioned in Chapter 2, the eIIiciency oI FWM strongly depends
on the Irequency separation O between the interacting waves producing beating. Eq.
(6.2) shows that the mixing Iunction F(O) is complex and contains three terms
responsible Ior the most important processes mediating FWM. Experimentally
observed FWM eIIiciency depends not only on the magnitude oI each contributing
process but also on their relative phase, as the processes can interIere either
constructively or destructively. The general trend is however that the FWM eIIiciency
quickly decreases with increasing detuning O. Fig. 6.6 shows the dependence oI
conversion eIIiciency p on the detuning O, whereas Fig. 6.7 presents conjugate energy
against detuning. In reality, the exact dependence may be slightly diIIerent, due to the
shape oI the gain curve, not included in the simple model utilized here. It is worth
noting that the FWM eIIiciency is larger Ior positive detunings (i.e. when the pump
Irequency is higher then the signal Irequency). This is due to interIerence eIIects
between the contributing processes |13|. As the detuning approaches zero, CDP
become dominant and the eIIiciency does not depend on the sign oI O. However, the
minimum spacing is limited by the bandwidth oI both pump and signal waves. In the
simulation presented below, the conditions were t
p
6 ps, E
p
2 pJ, G
0
30 dB,
P
2
0.2 mW. As the conjugate energy decreases quite Iast with increasing O, in
practical sampling systems it is desirable to have a tunable sampling laser, which is
able to adjust the detuning to the signal wavelength. This not only simpliIies the
100
calibration, but most oI all improves the SNR, which is, as discussed in the next
section, a crucial Iigure oI merit Ior an optical sampling system.
Conversion efficiency vs detuning
-2
0
2
4
6
8
10
12
14
0.0 0.5 1.0 1.5 2.0 2.5
O OO O [THz]
q qq q [dB]
negative
positive

Fig. 6.6 Conversion eIIiciency against detuning O, t
p
6 ps, G
0
30 dB, P
2
0.2 mW
Conjugate energy dependence on detuning
0
2
4
6
8
10
12
14
-2.5 -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0 2.5
O OO O [THz]
E
c
[fJ]

Fig. 6.7 Conjugate energy against detuning O, t
p
6 ps, G
0
30 dB, P
2
0.2 mW
6.1.6. Dependence of the optimum pump energy on the SOA gain
As investigated in Section 6.1.2, the optimum pump energy depends on the
sampling pulse duration, although the conversion eIIiciency is a slowly decreasing
Iunction Ior pump energies larger than optimum. It is also interesting to know iI the
optimum point shiIts with the SOA gain G
0
. Fig. 6.8 shows the Iunction p(E
p
) Ior
three diIIerent G
0
values: 15 dB, 22 dB and 30 dB. It can be seen that Ior a small gain
101
the optimum pump pulse energy is higher, but it is possible in practice to use an
almost optimal pump energy irrespectively oI the SOA gain.
Conversion efficiency against pump energy for different gain vaIues
-12
-10
-8
-6
-4
-2
0
2
4
6
8
1.E-04 1.E-03 1.E-02 1.E-01 1.E+00 1.E+01
E
p
[pJ]
q qq q [dB]
15 dB
22 dB
30 dB
SOA gain

Fig. 6.8 Conversion eIIiciency against pump energy, with G
0
as a parameter, t
p
6 ps, P
2
0.2 mW
6.2. Noise analysis of the optical sampling system based on FWM in SOAs
In section 6.1, the issue oI conversion eIIiciency was investigated. Although
very signiIicant, conversion eIIiciency is not the only Iigure oI merit Ior the optical
sampling system. Since optical sampling is expected to be an important measurement
and characterization tool, the issue oI the available SNR is crucial Ior the applicability
oI this technique. For example, optical sampling was used to generate histograms |2,
19| describing the power distribution oI an optical signal corrupted by the ASE noise,
as described in Chapter 3. Based on these histograms the signal quality and BER can
be estimated (one popular method oI BER estimation based on Q-Iactor was discussed
in Chapter 4. An alternative novel technique based on statistical moments obtained Ior
example Irom optically sampled histograms will be proposed in Chapter 7). It is clear
that the sampling system itselI should introduce as little noise as possible to be able to
characterize the signal quality prior to optical sampling. Although the noise properties
oI FWM in SOA schemes were investigated previously |9, 20|, these approaches were
usually mainly concerned with diIIerent optical noise contributions to the conjugate
wave and are not directly applicable to the optical sampling system, where the optical
samples are detected using a slow photodetector and are processed in the electrical
domain. The sensitivity oI the optical sampling system was also investigated
102
previously |15, 21|, but these investigations were limited to the passive systems
without an internal saturable optical gain.
6.2.1. Signal to noise ratio in the considered sampling system
To obtain the best possible SNR, all elements oI the sampling system need to
be optimized. This requires a proper choice oI the SOA gain, optical Iilter bandwidth
to separate the conjugate wave and the electrical bandwidth. Additionally, there is a
trade-oII between the resolution and the SNR. In this section, the noise analysis is
presented, giving design guidelines Ior the optimal sampling system. The analysis is
based on the optical sampling system model shown in Fig. 6.9. In Fig. 6.9, o
1
is the
coupler loss (including splitting loss and excess loss) seen by the signal and o
2
is the
insertion loss oI the optical bandpass Iilter (BPF), experienced both by the optical
sample and ASE noise. R is the photodiode responsivity, B
0
is the optical bandwidth,
E
P
is the sampling pulse energy and At is the sampling window related to the
sampling pulse duration as shown in the derivation below.

Fig. 6.9 Optical sampling system model
The noise analysis is based on the assumptions that the optical sampling gates
out signal energy within the sampling window and that the electrical bandwidth is low
compared to the bandwidth oI the conjugate pulses aIter optical Iiltering. These are
valid approximations in most cases oI practical interest (and certainly in the case
investigated here, when a CW signal is assumed), thanks to the very large bandwidth
oI optical sampling, as discussed in Chapter 3. As the sampling pulses are much Iaster
than the signal variations, signal power within the sampling window is assumed
constant and the gated energy depends on the sampling pulse duration. The only
constraint on the receiver is that the electrical bandwidth is larger than the sampling
repetition rate to avoid interIerence between the electrical samples. The optimum
optical and electrical bandwidth is investigated later in this section.
The electrical Iield at the receiver is assumed to be given by:
103
) (
1
) ( )) ( ) ( ( ) (
0
2
0 2
t h A t h t A t f t A
ASE
- + - =
o o
q
(6.5)
where A
2
(t) is the signal Iield at the SOA input and dependence on the spatial
coordinate is dropped to simpliIy notation. o is the total loss
2 1
o o o = , h
0
(t) is the
optical Iilter impulse response and A
ASE
is the white Gaussian ASE at the SOA output.
It is assumed that the optical Iilter`s rejection is so good and the roll-oII is so Iast
compared to the detuning between the sampling and the signal that any crosstalk is
negligible. f(t) is the gating Iunction, assumed to be dependent only on the sampling
pulse shape. This is strictly only an approximation, as XGM in the SOA and SOA-
induced chirping are neglected, but this simpliIication is suIIicient Ior noise analysis.
ThereIore f(t) is proportional to P
1
(t): ) 0 ( ) ( ~ ) (
1 1
P t P t f . The Iorm oI f(t) is a
consequence oI the FWM equation (6.3c). The signal A
2
(t) is assumed to be ideal and
noiseless, i.e. all the degradation oI the electrical samples is contributed by the
sampling system. The optical signal given by (6.5) is detected and electrically Iiltered.
The electrical Iilter is modeled as a Iinite time integrator, with the integration time T.
Detection using a pin photodiode is considered, as the same conIiguration has been
used in experiments. The noisy electrical signal is given by:
}

- + - =
2 /
2 /
2
2
2
) (
1
) ( ) ) ( ( ) (
T
T
o ASE o
dt t h A t h P t f
T
R
t i
o o
q
(6.6)
where CW signal P
2
is assumed. The squaring operation in (6.6) produces a signal
term and signal-ASE and ASE-ASE beating noise terms. The signal term is given by:
T
E R
I
2
q
o
= (6.7)
where E
2
is the gated signal energy E
2
P
2
At and is At the conjugate pulse duration
}


- = A dt t h t f
2
0
) ( ) ( t . Note that the optical Iilter does not change the sampling
resolution, as it may seem, at Iirst glance, to be implied by (6.6), and its inclusion in
the deIinition oI At is only the consequence oI the assumed system model. The
sampling resolution is determined by f(t), which is obvious Irom the physical point oI
view. This is an advantage oI the considered scheme as compared to other sampling
techniques, where the temporal resolution may depend on the chromatic dispersion-
104
induced walk-oII between the sampling pulses and the signal |2, 3, 15|. In SOAs, due
to the very short interaction length, this walk-oII is negligible. The optical Iilter
included in the present scheme broadens the optical samples but at the same time
lowers their peak power so the sample energy is conserved (the Iilter insertion losses
are accounted Ior via o
2
so that h
0
(t) is lossless). ThereIore, the sampling window is
equivalent to
}


= A dt t f
2
) ( t . However, optical Iilter bandwidth does change the
variance oI the noise terms due to ASE, as shown later in the analysis. In the
deIinition oI the sampling window use is made oI the assumption that the receiver
bandwidth is signiIicantly smaller than the optical Iilter bandwidth.
ASE beating terms arise as a result oI the transIer oI the optical noise into the
electrical domain. The Iundamental diIIerence between the ASE-signal term and the
ASE-ASE term is that the latter is created over the whole receiver integration time (as
the ASE noise is quasi-CW, iI the short periods oI reduced ASE spectral density S
ASE

due to additional pump-induced gain saturation are neglected), whereas the Iirst
appears only in the time window when the sample is present. As a result, the standard
expression Ior variance oI the ASE-ASE term can be used |22, 23|, whereas the ASE-
signal variance needs to be rederived. This derivation is relegated to Appendix B.
Here only the Iinal Iormula is presented, derived assuming a Gaussian optical Iilter.
This assumption about the Iilter shape signiIicantly simpliIies the analysis and is
reasonably valid, as the Santec tunable Iilter used in the experiments is indeed
Gaussian-like shaped. The variances oI the ASE-ASE beating
2
A A
o and ASE-signal
beating
2
sig A
o are approximately given by
e
ASE
A A
B B
S
R
0
2
2
2
2 2
4
o
o =

(6.8)
f G
f G
f G
ASE sig A
P P
T
R
t t
t t
t t
oo
q
t o
2 2
2 2
2
2
2
2
2
5 2
2
4
+
+
=

(6.9)
where B
e
1/(2T) is the electrical Iilter bandwidth, t
f
is related to the optical BPF
impulse response as explained in App. B, t
G
is related to the sampling pulse FWHM
width and P
ASE
is the ASE power in one polarization state aIter the optical Iilter.
In addition to the two noise terms given by (6.8) and (6.9), which are speciIic
to the particular type oI the optical sampling system based on FWM in SOAs, the ever
105
present thermal noise and shot noise terms have to be considered as well. The
Iormulae Ior variances oI these noise terms are well known |23| and given by
L e e K B th
R B F T k / 4
2
= o (6.10)
e ASE sh
B P
T
E
eR
|
|
.
|

\
|
+ =
2
2 2
2
2
o o
q
o (6.11)
where k
B
is the Boltzmann constant, T
K
is absolute temperature, F
e
is the electronic
ampliIier noise Iigure, R
L
is the load resistor and e is the electron charge. Note that the
contribution Irom the ASE power to the shot noise included in (6.11) can be large
since ASE noise is not gated.
The total noise variance o
2
is obtained by summing all Iour contributions
(6.8)-(6.11):
2 2 2 2 2
A A sig A th sh
+ + + = o o o o o (6.12)
The total variance can be used to analyze the available SNR oI the electrical samples,
which can be then used Ior Iurther signal processing. The SNR is deIined as:
2
2
o
I
SNR = (6.13)
To Iurther simpliIy the analysis, a Gaussian sampling pulse shape is assumed
(similarly to the approximations used to derive (6.8)), which Ior the soliton sampling
laser is a good approximation, as sech
2
and Gaussian pulses diIIer very little over the
central part (they increasingly diverge in the tails, but these do not contribute
signiIicantly to the conjugate pulses). ThereIore, sampling pulse shape is given
by ) 2 exp
2 2
G p
: t (- P , where the Gaussian width t
G
is related to the FWHM width t
p
by
t
G
t
p
/2.355. The sampling window At used in (6.7) in the case oI Gaussian-shaped
sampling is given by
G
t t t = A or
P
t t 75 . 0 = A . Note the additional improvement in
sampling resolution due to third order nonlinearity, which leads to dependence oI the
conjugate power on sampling power squared (as seen in (6.6)).
Using (6.7)-(6.13), SNR is approximately equal to (assuming no coupler
losses and lossless optical Iilter):
106
ASE e o ASE e L e K B e o ASE
S E R B B S R B R F T k B B eRS eRE
E R
SNR
3
2 2 2
3
2
3
2
2 / 4 ) /( 4 / 4 2 + + + +
=
(6.14)
where
2 3
E E q = is the sample energy as deIined by (6.4). (6.14) is the basic equation
giving direct practical guidelines to design and maximize the system perIormance.
The Iirst obvious observation is that the conversion eIIiciency (included in the
conjugate energy E
3
) should be as large as possible. This limits the practical detuning
range, as evident Irom Figs. 6.6 and 6.7. Moreover, there is a trade-oII between the
SNR and resolution, as longer sampling pulses permit more signal energy to be gated
out. High electrical bandwidth is desirable to reduce the eIIect oI the ASE shot noise,
thermal noise and ASE-ASE beat noise. ThereIore, unlike Ior standard relatively high
duty Iactor signals, large electrical bandwidth actually improves SNR. At the same
time, tight optical Iiltering is beneIicial, as it reduces ASE shot noise and ASE-ASE
noise. The practical constraint is however that usually tighter Iilters introduce more
insertion losses. It is worth noting that the optical Iiltering does not inIluence the
conjugate energy (Ior a lossless Iilter) and at the same time reduces ASE power. This
diIIerent behavior is a simple consequence oI the Iact that the conjugate is an energy
type signal, whereas ASE is a power type signal. Thermal noise (which is typically
signiIicant Ior the discussed system, unless cooling is used) can be also overcome by
using an optical preampliIier. In an optimally designed system, the ultimate sensitivity
limits should be given by the shot noise (quite obviously) and the ASE-signal beating.
It is worth noting that the latter term is independent oI the optical Iilter, and the only
way to combat it is to maximize the conversion eIIiciency and use a low noise SOA.
It may be beneIicial (depending on the input signal power level) to use a second low
noise preampliIier beIore the SOA, as E
3
is proportional to the input power.
Fig. 6.10 shows the SNR against the signal power at the SOA input, obtained
Ior the Iollowing set oI parameters: t
p
6ps, n
sp
3, G
0
30 dB, P
sat
11 mW,
detuning 5.5 nm, E
p
110 IJ, B
e
750 MHz, B
o
40 GHz, o
2
6 dB, R 0.9 A/W,
R
L
0.5 kO, F
e
4. Conversion eIIiciency and saturated gain were numerically
calculated Ior each input power value. The asymptote shown in Fig. 6.10, oI the slope
20 dB/decade, indicates the region dominated by the signal-independent noises.
Additionally, the SNR curve without the thermal noise contribution is plotted, to show
107
the possible scope Ior sensitivity improvement due to using an additional
preampliIier. Sensitivity analysis Ior other system parameters is given later.
Two main signal-independent noise terms are important in practical cases
the thermal noise and ASE-ASE beating. It is thereIore beneIicial to introduce a
preampliIier between the SOA and the receiver, to overcome the thermal noise
contribution. However, it is worth noting that the slope oI the SNR curve changes Ior
relatively low input powers. This is not only because the contribution Irom the signal-
dependent noise terms (mainly signal-ASE beating) becomes signiIicant, but more
importantly because the gain saturation due to the CW beam reduces conversion
eIIiciency. Note that the value oI the saturation power used to calculate the SNR
curves in Fig. 6.10 is consistent with the SOA measurements presented in Chapter 5
and also consistent with the values Iound in the literature Ior MQW SOAs |24|. As
discussed in |25|, Ior a detuning less than 12 nm, the saturation power is independent
oI signal intensity. The saturated gain values Ior diIIerent input powers obtained Irom
simulations using P
sat
11 mW are consistent with experimental values reported in
Chapter 5.
The generally low SNR values show that even a careIul design oI the
experimental setup (which may include Ior example a narrow optical Iilter, a
relatively Iast receiver or an optical preampliIier) can only to some extent improve the
achievable SNR and in practice the limit is set by the conversion eIIiciency and
required resolution. Even iI the thermal noise were eliminated, the highest attainable
SNR is less than 25 dB. This is a disadvantage oI the discussed sampling technique
and is a consequence oI the optically active noisy and saturable element (SOA) used
to perIorm the sampling. Sampling techniques based on passive devices usually have
a much broader SNR range limited by the signal independent noise terms |15, 21|. On
the other hand, they oIten require higher input powers and, due to lower interaction
eIIiciency, require much longer interaction lengths, e.g. in the case oI sampling using
nonlinear Iibre. The long interaction length reduces sampling resolution and
introduces wavevector mismatch.
108
SNR versus input power
-5
0
5
10
15
20
25
0.01 0.1 1 10 Psig [mW]
SNR [dB]
all noise terms
no thermal 20 dB/dec

Fig. 6.10 SNR versus signal power. 20dB/decade asymptote shows region dominated by
signal independent noise terms.
Fig. 6.11 shows magnitude oI each oI the noise terms versus signal power. The
steep Ialling slope oI the ASE-ASE term is a consequence oI the Iact that this noise is
proportional to the SOA gain squared. Thermal noise is obviously constant at given
temperature and its value is typical considering the bandwidth |23|. The Iact that, in
general, none oI the contributing noise terms (except the shot noises) are negligible
makes the inclusion oI a polarizer between the SOA and the receiver impractical. The
polarizer would halve the ASE-ASE beating, but at the same time would introduce
signal losses and increase the relative importance oI the thermal noise. More detailed
design analysis oI the sampling setup is given in the next section.
Dependence of the noise terms on signaI power
1E-17
1E-16
1E-15
1E-14
1E-13
1E-12
0.01 0.1 1 10
Psig [mW]
o oo o
2 22 2
[A
2
]
shot
ASE shot
ASE-ASE
sig-ASE
thermal

Fig. 6.11 Dependence oI diIIerent noise terms on input signal power
109
6.2.2. Sensitivity analysis
The importance oI a proper choice oI each element in the sampling setup
shown in Fig. 6.9 can be investigated by deIining another Iigure oI merit Ior the
optical sampling system - the sampling sensitivity. Here, the sensitivity is deIined as
the input signal power required to obtain SNR 13 dB, when thermal noise is made
negligible. Figs. 6.12-6.14 show how the sensitivity changes with varying parameters
t
p
, B
e
and B
o
, which are the practically adjustable parameters. In each Iigure, only one
parameter Irom the triplet mentioned above was varied. The rest oI the parameters
were kept the same as Ior Fig. 6.10.
Fig. 6.12 presents sensitivity oI the considered sampling scheme to the
sampling pulse width. The trade-oII between the resolution and the sensitivity is
clearly visible. Especially Ior very high sampling resolution (sampling pulse width
less than about 2.5 ps), the sensitivity oI the sampling system is severely reduced.
This may be in many scenarios the limiting Iactor Ior the available resolution. It
seems that the use oI sampling pulse oI about 5 ps duration presents the best trade-oII
between the sensitivity and resolution.
SampIing sensitivity vs sampIing puIse width
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
1 3 5 7 9 11 13
t tt t
p
[ps]
P
sens
[mW]

Fig. 6.12 Sensitivity oI the method to sampling pulse width
Fig. 6.13 shows that, perhaps surprisingly, the sensitivity to the optical Iilter
bandwidth is moderate and almost linear. ThereIore in many cases it is not beneIicial
to use a tighter Iilter at the expense oI larger insertion losses although low-loss tight
optical Iilters are obviously preIerable. On the other hand, a reasonably Iast receiver is
desirable, as Fig. 6.14 shows that the sampling sensitivity improves quickly with
110
increasing the electrical Iilter bandwidth Ior low receiver speeds but the improvement
becomes less pronounced Ior a Iast receiver. It does not seem advantageous to use
electrical bandwidth larger than about 1 GHz and the optimum region is somewhere
between 700 MHz - 1 GHz. Much Iaster receivers have usually lower eIIective load
resistance (including the transimpedance ampliIier (TIA)), which increases thermal
noise contribution and additionally oIIsets the already minor improvement in
sensitivity.
SampIing sensitivity vs opticaI fiIter bandwidth
0.08
0.1
0.12
0.14
0.16
15 20 25 30 35 40 45
B BB B
o
[GHz]
P
sens
[mW]

Fig. 6.13 Sensitivity oI the method to optical Iilter bandwidth
SampIing sensitivity vs eIectricaI fiIter bandwidth
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
100 200 300 400 500 600 700 800 900 1000
B BB B
e
[MHz]
P
sens
[mW]

Fig. 6.14 Sensitivity oI the method to electrical Iilter bandwidth
Analysis presented in Section 6.2, together with the investigations oI the
achievable conversion eIIiciency shown in Section 6.1, provide the practical
111
guidelines Ior design oI the sampling system based on FWM in SOAs. Although there
may be some uncertainty regarding the values oI some oI the parameters used to
investigate both the conversion eIIiciency and the SNR, the qualitative behavior oI the
system should not depend on exact values and thereIore the design rules deduced
Irom the presented analysis should be valid. The practical implementation oI the
sampling system is discussed in Section 6.3. Therein, experimental results are also
presented.
6.3. Experimental investigation of optical sampling using FWM in SOAs
Based on investigations detailed in Sections 6.1 and 6.2, an experimental setup
to perIorm FWM in SOAs has been designed. The experimental setup is very similar
to the model shown in Fig. 6.9, with an additional EDFA preampliIier added aIter the
SOA to boost the conjugate energy. Consequently, two optical bandpass Iilters were
used, beIore and aIter the preampliIier. The Iirst optical Iilter had quite a wide
bandwidth and served to remove most oI the undesirable optical power to prevent the
EDFA saturation without introducing excessive signal losses, whereas the bandwidth
oI the second Iilter was much narrower to perIorm the Iinal tight Iiltering. Compared
with the model in Fig. 6.9, optical isolators were also included in the setup to reduce
any unwanted reIlection.
As a source oI the sampling pulses, a Pritel FFL laser has been used. The laser
was a passively-modelocked, all-Iibre soliton laser, capable oI producing pulses oI
FWHM duration Irom 2.5 ps to 8 ps, depending on the internal optical Iilter installed.
The laser`s repetition rate was about 88.43 MHz. The SOA used to perIorm sampling
was an SOA-NL-OEC-1550 device Irom CIP, extensively characterized in Chapter 5.
The optical detector used in the experiments was a TIA-952 device Irom TTI,
Ieaturing an InGaAs pin photodiode, internal 500 O transimpedance and 800 MHz
bandwidth. The electrical signal was captured using a LeCroy 9362 sampling
oscilloscope with 10 GS/s sampling rate and 750 MHz single-shot bandwidth. AIter
the SOA, the signal oI interest was separated Irom other Irequency components and
broadband ASE noise using a Santec OTF-300 0.6 nm manually tunable Iilter, with
2.6 dB insertion losses. The pre-Iiltered conjugate was subsequently ampliIied using a
low power EDFA preampliIier. The EDFA gain was not Ilattened and the peak small
signal gain was about 19 dB. Final Iiltering was perIormed by a Santec OTF-655 0.34
nm motorized tunable Iilter, with 5 dB insertion losses. Both Iilters had Gaussian-like
112
spectral shapes. Although the analysis given in Section 6.2 predicts a perIormance
improvement when even tighter optical Iiltering is used, the Iilter used was the most
narrowband one available in the laboratory. The two polarization controllers (PC)
seen in Fig. 6.15 were used to adjust the SOP oI the mixing waves as mentioned
beIore, mixing eIIiciency is largest when the two beams have parallel polarization and
no FWM takes place when they are orthogonal, which can be used as a conIirmation
whether the observed signal is in Iact a conjugate. As the signal source, a Bookham
0224AP tunable laser was used, with an isolator added at the output to eliminate
possible relative intensity noise (RIN) enhancement caused by external reIlections.
The sampling laser had an internally installed isolator and adding an external one did
not result in changing the laser RF and optical spectrum.

Fig. 6.15 Experimental setup Ior optical sampling. PM-power monitor, ISO-isolator, PC-polarization
controller
6.3.1. Characterization of the sampling setup
BeIore perIorming any optical sampling using the setup shown in Fig. 6.15, it
is necessary to characterize the setup itselI. The characterization involved measuring
the linearity oI the setup (i.e. the dependence oI the voltage amplitude observed on the
oscilloscope on the sampling pulse energy), investigating spectral evolution oI the
short sampling pulses in the SOA and investigating the noise properties and intrinsic
noise limitations oI the sampling setup.
6.3.1.1. Detection linearity
Although it is usually assumed that the current generated by a photodiode is
linearly proportional to the incident optical power, in practice Ior a high input power
some saturation oI the photodiode response occurs. This may be especially
pronounced when a TIA ampliIier Iollows the photodiode, as any ampliIier may
introduce distortions. In Iact, Ior the TIA-952 receiver, the manuIacturer`s
speciIication is that the maximum linear input power is 2 mW and the maximum
113
linear TIA ampliIier output voltage is 1 V. Although, taking into account relatively
low repetition Irequency oI the sampling laser and its very low duty ratio, it is
unlikely that the average conjugate power would exceed the linearity limit (however,
the Pritel laser is well capable oI producing average power exceeding 2 mW), the
peak power oI the short pulses can easily exceed this limit. Moreover, the
oscilloscope itselI can introduce nonlinearity. ThereIore, it is desirable to characterize
the linearity oI the detection subsystem oI the sampling setup (consisting oI TIA-952
and LeCroy 9362), to make sure its response is understood.
Fig. 6.16 shows the dependence oI the voltage J observed on the oscilloscope
on the incident pulse energy E. The relationship is remarkably linear. The best linear
Iit is also shown Ior comparison. It should be noted that beIore TIA-952 was selected,
another receiver was considered, but it had strongly nonlinear characteristics, possibly
due to nonlinear TIA ampliIier. The oIIset observed in the trend line equation is
probably contributed by the oscilloscope oIIset voltage. The residual nonlinearity may
be partially due to the non-equal vertical gain oI the oscilloscope input buIIers Ior
diIIerent sensitivity scales. The slight amplitude saturation Ior low E is due to the
setup noise Iloor. This calibration curve was independent oI the pulse width.
Linearity of the detection subsystem
y = 0.76x - 1.80
0
20
40
60
80
100
120
0 20 40 60 80 100 120 140 160
E [fJ]
V [mV]

Fig. 6.16 Linearity oI the detection subsystem. Best linear Iit equation displayed in the graph.
6.3.1.2. SNR limits of the sampling setup
Similarly to the detection linearity, Irom the practical point oI view the
intrinsic SNR achievable Ior the experimental setup is an equally important piece oI
inIormation. The best possible SNR
best
was measured in a similar conIiguration as the
114
calibration curve in Fig. 6.16, by using the sampling pulses Irom the Pritel laser
incident directly on TIA-952, without including any optical ampliIiers in the setup
(thereIore, ASE-signal and ASE-ASE beating noise terms are not present).
Best achievabIe SNR of the sampIing setup
0
5
10
15
20
25
30
35
40
1.E+00 1.E+01 1.E+02 1.E+03
E [fJ]
SNR [dB]
10 dB/dec
20 dB/dec

Fig. 6.17 Best SNR achievable in the sampling setup. The two asymptotes denote regions oI signal-
dependent and signal-independent noise domination
Fig. 6.17 shows the dependence oI the experimentally measured SNR
best
on
the input pulse energy. The two asymptotes 10 dB/decade and 20 dB/decade indicate
signal-dependent and signal-independent regimes. It should be noted that Ior low
input pulse energies, the SNR slope is close to 20 dB/decade, which shows dominance
oI the thermal noise. For larger input energies, the slope oI the curve Ilattens and
eventually saturates at about SNR
best
30 dB. This saturation is not expected Irom the
theory presented in Section 6.2 and shows that there is another noise source
proportional to the input optical power or to electrical current. The single break point
(Irom 20 dB/decade to 0 dB/decade) on the SNR curve in Fig. 6.17 shows that there is
no region dominated by signal dependent noise (i.e. shot noise in this case). The SNR
saturation is most likely due to the relative intensity noise (RIN) oI the sampling
pulses. EIIorts to reduce possible RIN enhancement due to reIlections in the setup, by
introducing optical isolators, have not proved successIul. An alternative explanation
Ior the SNR saturation could be an excessive digital conversion noise Irom the
sampling oscilloscope. Although the oscilloscope vertical resolution was 8 bit, which
corresponds to ideal, quantization noise-limited, SNR oI about 43 dB, there may be
some additional sources oI noise, related Ior example to the non-perIect triggering
115
timing, which is critical Ior very short pulses. Note also that the quantization noise
scales in a way with the optical power, as diIIerent sensitivity settings were used
depending on the observed signal voltage.
To compare the noise characteristic oI the setup with optical ampliIiers (which
introduce additional noise components due to ASE) with the best possible SNR, Fig.
6.18 shows similar SNR curve measured in the setup with an SOA and Santec OTF-
300 added between the Pritel FFL and TIA-952. The pulse energies are measured
beIore the SOA, hence the improvement in sensitivity. The most notable Ieature oI the
curve in Fig. 6.18, as compared with the one in Fig. 6.17, is the presence oI the
second break point on the SNR curve, indicating an increasing importance oI signal
dependent noises. This is, oI course, due to the signal-ASE beating term.
SNR achievabIe in sampIing setup with opticaI ampIification
0
5
10
15
20
25
30
35
40
1.E-02 1.E-01 1.E+00 1.E+01 1.E+02
E [fJ]
SNR [dB]
10 dB/dec
20 dB/dec

Fig. 6.18 SNR achievable in the sampling setup with optical ampliIication. The two asymptotes denote
regions oI signal-dependent and signal-independent noise domination
Finally, to show the increased importance oI the signal-dependent noise in the
setup with optical ampliIication, Figs. 6.19 and 6.20 present, respectively, persistence
plots measured Ior the case with no SOA and pulse energy E 9.3 IJ and Ior the case
with SOA and pulse energy beIore the SOA E 0.4 IJ. The broadening oI the upper
rail compared to the Iloor level in Fig. 6.20 is clearly seen.
116

Fig. 6.19 Persistence plot showing dominance oI signal-independent noise in the setup without optical
ampliIiers

Fig. 6.20 Persistence plot showing presence oI signal-dependent noise in the setup with optical
ampliIiers
6.3.1.3.Spectral shift of sampling pulses in SOAs
As brieIly mentioned in Chapter 2 and in Section 6.1 oI the present chapter,
short optical pulses propagating through the SOA may undergo signiIicant spectral
evolution, depending on the pulse energy. This is mostly due to the eIIect oI gain
saturation, which then aIIects the reIractive index. The gain and reIractive index are
117
coupled via the Kramers-Kronig relations. Changes in the reIractive index lead to
SPM. As an eIIect oI SPM, not only the spectrum shape oI the short pulses can
evolve, but it can also acquire a spectral shiIt, i.e. the carrier Irequency oI the output
pulse is diIIerent than the carrier Irequency at the SOA input. This phenomenon was
Iirst studied in detail in the context oI SOAs by Agrawal and Olsson in their seminal
paper |26|. They observed that Ior large pulse energies, the pulse can develop a
multipeak structure and the main peak exhibits a red shiIt.

Fig. 6.21 Spectrum oI the sampling pulses beIore the SOA, E
p
100 IJ

Fig. 6.22 Spectrum oI the sampling pulses aIter the SOA, E
p
100 IJ
118
To provide qualitative understanding oI the possible spectral evolution, Figs.
6.21 and 6.22 show the sampling pulse spectrum beIore and aIter the SOA,
respectively. Due to the low duty Iactor and low repetition rate oI the Pritel laser, the
time-averaged ASE level is quite high. The development oI the second peak is
however visible. The peak wavelength is shown in each Iigure and the spectral shiIt is
apparent. The pulse energy was 0.1 pJ and the pulsewidth was 6 ps.
The magnitude oI the spectral shiIt depends on several parameters, most
importantly on the pulse duration and energy and the SOA gain. This spectral shiIt
may be oI considerable importance Ior optical sampling in the investigated
conIiguration, as it means that the conjugate pulse may be positioned at a diIIerent
Irequency Irom that expected Irom the Irequency detuning between the input signal
and the pump. Considering the Iact that, as shown in Section 6.2, it is beneIicial to use
narrowband Iiltering; the unexpected Irequency shiIt may lead to signiIicant signal
losses, as the Iilter is oIIset Irom the conjugate center Irequency. ThereIore, the
magnitude oI the possible spectral shiIt and its dependence on various parameters
must be given considerable attention.
Spectrum shift against puIse energy
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0 0.2 0.4 0.6 0.8 1 1.2
E [pJ]
A A A A [nm]

Fig. 6.23 Experimental dependence oI the spectral shiIt on pulse energy. t
p
6 ps, I
SOA
300 mA.
Fig.6.23 shows the spectral shiIt against the input pulse energy. The pulse
duration t
p
was 6 ps, I
SOA
300 mA. Additional CW power injected into the SOA was
0.4 mW. It can be seen that Ior very high energies the spectral shiIt can be as large as
0.7 nm. For moderate pulse energies the shiIt is much smaller.
119
Dependence of the spectraI shift on I
soa
0
0.05
0.1
0.15
0.2
0.25
0.3
0 50 100 150 200 250 300
I
soa
[mA]
A A A A [nm]

Fig. 6.24 Dependence oI the spectral shiIt on the SOA current. t
p
6 ps, pulse energy 110 IJ.
Fig. 6.24 shows the experimental dependence oI the spectral shiIt on the SOA
pumping current. The pulse duration and CW power were the same as Ior Fig. 6.23.
The pulse energy was 110 IJ. It can be seen that heavy pumping leads to a larger shiIt.
This is consistent with the investigations reported in the literature that the spectral
shiIt is strongly dependent on the SOA gain |26|. Finally, Fig. 6.25 investigates the
dependence oI the spectral shiIt on the pulse duration, when the pulse energy is kept
constant. Two cases were experimentally investigated, when the pulse energy was 11
IJ and 110 IJ. It is visible that shorter pulses undergo much larger spectral shiIt, as
reported in the literature |9, 27|. This is due to the higher peak power, which also
means more SPM.
Dependence of the spectraI shift on puIsewidth
0
0.2
0.4
0.6
0.8
1
2 3 4 5 6 7 8
t tt t
p
[ps]
A A A A [nm]
110 fJ
11 fJ

Fig. 6.25 Dependence oI the spectral shiIt on the pulse width Ior two diIIerent values oI pulse energy
120
It is important to determine how the sampling pulse Irequency shiIt aIIects the
conjugate output Irequency. First oI at all, it should be noted that, in the scheme
considered here, the probe is given simply by a CW wave. Constant optical intensity
does not induce any SPM. However, the probe experiences the same reIractive index
variations as the pump (short pulses) due to cross phase modulation. ThereIore, during
the short periods where the probe is copropagating with pump pulses, it acquires the
same chirp and spectral shiIt. Hence the Irequency detuning between the probe and
the pump remains constant (subject to minor higher order eIIects due to the
interIerence between diIIerent nonlinear gain components |27|), and the conjugate is
shiIted by the same amount as the pump |9|. Since the pump duty ratio is very low,
the probe Irequency shiIt takes place during almost negligibly short time compared
with the unperturbed probe CW propagation time in the SOA, which does not
experience any chirp. ThereIore, this probe shiIt is not readily visible on the OSA,
which may lead one to the incorrect conclusion that the probe is not shiIted, which
would mean that the conjugate shiIt is twice the pump shiIt. However, it was
experimentally observed that the conjugate shiIt is equal to the pump shiIt, in
agreement with the above reasoning.
The characterization oI the sampling setup presented in the present section,
together with the noise analysis given in Section 6.2 and conversion eIIiciency
analysis discussed in Section 6.1 provide general understanding oI the capabilities and
limitations oI the investigated sampling technique. This understanding is necessary to
be able to correctly interpret the experimental results. In the Iollowing, the
experimental results are reported and discussed.
6.3.2. Experimental results
The setup shown in Fig. 6.15 has been used to perIorm experimental
investigations oI the potential oI the considered optical sampling technique. The
EDFA preampliIier was used because it has been experimentally Iound that otherwise
no FWM signal is detectable. This preampliIier improves the detection sensitivity by
reducing the relative importance oI the thermal noise, as discussed in the previous
section.
121
6.3.2.1.Dependence of the conversion efficiency on sampling pulse
energy .
First oI all, it was experimentally checked iI the optimum sampling pulse
energy dependence on pulse width discussed in Section 6.1.2 is indeed observed in
the measurements. Fig. 6.26 shows the dependence oI the conversion eIIiciency on
sampling pulse energy Ior diIIerent pump pulse widths. Measurement conditions
were: pump wavelength 1544 nm, signal-pump detuning 0.73 THz (5.84 nm), input
signal power 0.4 mW, SOA current 300 mA.
Conversion efficiency against pump energy
-4
-3
-2
-1
0
1
2
10 30 50 70 90
E
p
[fJ]
q qq q [dB]
8 ps
6 ps
4 ps
2.4 ps

Fig. 6.26 Dependence oI the conversion eIIiciency on the pump energy E
p
Ior diIIerent sampling pulse
widths
It can be seen that the conversion eIIiciency curves have similar qualitative
Ieatures as the simulated curves in Section 6.1.2. The curves are quite Ilat Ior E
p
~50
IJ, irrespective oI the sampling pulse width. ThereIore the optimum pump energy is
about 50-100 IJ, which is quite consistent with the numerical results. For energies
larger than 100 IJ, the sampling pulses spectrum becomes heavily distorted due to
SPM in the SOA. Pump energies below 50 IJ should not be used due to low
eIIiciency. Shorter pulses result in higher conversion eIIiciency, as indicated by the
numerical Iindings, but the absolute conjugate energy value is lower (see deIinition
(6.4)), resulting in a worse SNR. There is some measurement uncertainty in the
conversion eIIiciency values, as means to characterize the shape and duration oI the
sampling pulses were not available. As a result, the values provided by the
manuIacturer are used, but the exact pulse width slightly varies with wavelength
122
setting. Similarly, the time-bandwidth product is dependent on the wavelength setting,
as concluded based on the observation oI the optical spectrum. The initial chirp may
play a role in the mixing process. Also, the small possible pulse broadening due to
propagation in the very short length (less than 2 m) oI optical Iibre beIore entering the
SOA is not included as the initial pulse chirp is not known.
6.3.2.2.Dependence of the conjugate energy on the detuning
As discussed in Section 6.1, conversion eIIiciency and thereIore the conjugate
energy is a strong Iunction oI detuning. Generally, the larger the detuning the less
eIIicient the mixing process. Fig. 6.27 shows the experimentally measured conjugate
energy versus pump-signal detuning. The pump wavelength was 1543.94 nm, whereas
the signal wavelength was varied. The signal power P
2
0.4 mW, sampling pulse
width t
P
6 ps and the pump energy was E
P
100 IJ. Conversion eIIiciency is higher
Ior positive detuning, as expected. The maximum and minimum detuning values were
limited by the sampling pulse bandwidth on one hand and by the SNR on the other.
Fig. 6.28 shows the SNR curves Ior the same measurement conditions as Ior Fig.6.27.
As expected, Ior large and Ior negative detuning, the SNR is lower. Perhaps surprising
however is the observation that the SNR also drastically reduces iI the detuning is
small, whereas the conjugate energy is relatively large in this region. This suggests
existence oI another noise term, appearing only in these conditions. The exact nature
oI this noise is not clear. However, this noise enhancement close to the pump
spectrum may be due to some power leaking Irom the sampling pulses into the
conjugate spectral range. As discussed in Section 6.3.1, the pump laser seems to have
some amplitude noise. Additional contribution may be due to interIerence eIIects
(occurring iI there is a spectral overlap between the pump and conjugate waves)
between the conjugate phase noise and the pump phase noise. Due to this interIerence,
the contributing phase noises can be converted into the amplitude noise. It is unlikely
that this degradation in SBR is inherent to FWM process, as corresponding curves Ior
CW FWM shown in Chapter 5 (Fig. 5.10) are monotonic Iunctions oI the magnitude
oI detuning.
123
Dependence of the conjugate energy on detuning
0
1
2
3
4
5
6
7
8
-0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
O OO O [THz]
E
c
[fJ]
positive
negative

Fig. 6.27 Dependence oI the conjugate energy on the pump-signal detuning
ExperimentaI dependence of SNR on detuning
0
2
4
6
8
10
12
14
16
-0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
O OO O [THz]
SNR [dB]
positive
negative

Fig. 6.28 Dependence oI the SNR on the pump-signal detuning
From Figs. 6.27 and 6.28 it is clear that there is only a relatively small
detuning range, which yields reasonable sampling perIormance. For the SOA used in
the experiments, in practice only the positive detuning oI 0.4-0.8 THz is useIul. By
varying the input power slightly diIIerent curves are obtained, but, as Fig. 6.10 shows,
P
2
0.4 mW is close to the optimum SNR conditions. The SNR values obtained
experimentally are quite low and are not really good enough Ior a high perIormance
measurement system. EIIorts made to improve the SNR by reducing all losses and
eliminating possible reIlections were not successIul. Also, a second EDFA
preampliIier was introduced cascaded with the Iirst one, to Iurther reduce the thermal
noise contribution. The detected amplitude was in this case obviously much larger,
124
but the increased ASE level led to increased ASE-ASE and signal-ASE beatings and
the Iinal SNR was almost the same as in the single EDFA case. ThereIore, the
sampling quality is limited by the conversion eIIiciency (as discussed in Section 6.2)
and even careIul experimental design cannot improve it.
6.3.2.3.Dependence of the conjugate energy and SNR on input signal
power .
The conjugate energy versus input power curve was simulated in Section 6.1
and SNR dependence on input power was discussed in Section 6.2. These two curves,
shown in Figs. 6.5 and 6.10, respectively, give probably the most essential Iigures oI
merit Ior the sampling system, thereIore they were experimentally investigated. As
seen in Fig. 6.5, the conjugate energy exhibits saturation when the input CW signal
power is large due to gain compression. This practically limits the system dynamic
range, as a monotonic and preIerably linear transIer characteristic is required Irom a
measurement system. In Fig. 6.10, the theoretical SNR oI the sampling system was
shown. Therein, two curves were plotted, one with the thermal noise contribution and
the other with thermal contribution neglected. As an EDFA was used to ampliIy the
samples in the experiments, the practical SNR curve would lie somewhere between
the two extreme cases shown in Fig.6.10 iI no other noise sources were present. As
discussed previously, depending on experimental conditions other sources oI noise
may appear.
The experimental investigation oI the dependence oI the conjugate energy E
C

and sampling SNR on input signal power are presented in Figs. 6.29 and 6.30,
respectively. The curves were obtained Ior the Iollowing measurement conditions:
pump wavelength 1543.94 nm, detuning 0.54 THz, pump energy 100 IJ and SOA
current 300 mA. The experimentally measured dependence oI E
C
on input power P
2

(Fig.6.29) is qualitatively similar to the numerically calculated plot in Fig.6.5,
exhibiting saturation Ior input powers larger than 300 W. However, the measured
energy values are more than 3 dB lower than the values predicted based on the
numerical model. This might be due to having only approximately correct values oI
many parameters used in the model. The voltage amplitudes corresponding to the
conjugate energies were measured using the oscilloscope`s averaging Iunction
(averaged over 150 waveIorms), to reduce errors due to large noise and to improve
125
sensitivity. ThereIore, it was possible to observe the conjugate signal even Ior SNR
0 dB.
ExperimentaI conjugate energy vs input power
0
0.5
1
1.5
2
2.5
0 100 200 300 400 500 600
P
2
[uW]
E
c
[fJ]

Fig. 6.29 Experimental dependence oI conjugate energy on input power
Fig. 6.30 shows the experimental SNR curve. The highest experimentally
obtained SNR was only about 16 dB and, unlike the theoretical curve in Fig. 6.10, it
exhibits saturation. This may suggest that the sampling source amplitude noise is
transIerred to the conjugate and that possibly a more stable source would result in
slightly better SNR. Due to the nonlinear dependence oI the conjugate amplitude on
the pump amplitude, the pump amplitude noise contribution to the conjugate energy is
Iour times the initial pump noise |20|. However, the expected improvement due to a
more stable pump is not large, as it is limited by the conversion eIIiciency. To make
sure that it is indeed the sampling system intrinsic noise which limits the measured
SNR and not the actual CW signal noise, the RIN oI the signal laser was measured. It
was estimated that the Bookham AP0224 RIN would give an SNR ceiling oI about 40
dB Ior CW operation; thereIore the measured SNR is limited by the sampling system
perIormance and it is not an actual measure oI the signal quality. Another observation
supporting this conclusion is the Iact that Ior large sampling energies (E
P
~ 500 IJ),
the detected conjugate energy is only slightly lower, whereas the SNR is much worse.
126
ExperimentaI SNR vs input power
-5
0
5
10
15
20
0 100 200 300 400 500 600
P
2
[uW]
SNR [dB]

Fig. 6.30 Experimental dependence oI sampling SNR on input power
The experimentally measured SNR is too low Ior a measurement system. To
give a qualitative understanding oI the conjugate signal quality, Fig.6.31 shows
typical train oI samples obtained Irom CW 150 W signal. The amplitude variations
are signiIicant. In the next section some possible ways to improve the SNR are
described and discussed.

Fig. 6.31 Typical measured train oI samples Irom CW 150 W signal.
127
6.3.2.4. Pulsed bias current injection
As shown in Section 6.3.2.3, the SNR obtained in the considered sampling
system is too low Ior signal characterization applications. ThereIore, ways to improve
the system perIormance were sought. Since the system design obeyed the theoretical
guidelines as close as possible, considering resources available in the laboratory, it
was not possible to achieve better SNR by simply replacing some oI the elements in
the setup. Instead, a more complicated setup with pulsed current injection was
investigated. The idea behind this was to reduce the ASE-ASE beating, much in the
same way as the EDFA preampliIier reduces the signiIicance oI the thermal noise.
Since ASE-ASE noise in the sampling conIiguration investigated so Iar is created and
detected over the whole receiver integration time T, whereas the conjugate pulses are
very short, it is beneIicial to reduce the SOA gain outside the short time windows At
when the mixing actually takes place, to reduce the average gain seen by the ASE
noise and thereIore to reduce the ASE-ASE beating. A similar conIiguration was used
in |28, 29| to improve the OSNR in FWM between low-repetition rate subpicosecond
pulses.
The modiIied experimental setup is shown in Fig. 6.32. Compared to the
experimental setup Irom Fig. 6.15, an optical delay line (ODL, General Photonics
VDL-600, 600 ps adjustment range), an RF ampliIier and biasT were added. The
sampling laser has an additional RF output, which is synchronized with the main
optical output. The RF output is internally produced simply by converting a Iraction
oI the laser optical power to electrical signal, by detecting it on a Iast photodiode. As
the amplitude oI the RF output is only 70 mV, as measured on the oscilloscope, the
pulses have to be ampliIied beIore injecting into SOA. A Minicircuits ZFBT-4R was
used as the RF ampliIier. This ampliIier has a bandwidth oI 4200 MHz and the RF
pulses aIter the ampliIier had an amplitude oI 2.7 V and duration oI about 500 ps. The
RF pulses were then combined with the DC bias in a high current biasT (bandwidth 3
GHz). The DC bias I
DC
was adjusted such that the RF gain component was largest.
This was achieved by observing the ASE level on the oscilloscope as the RF bias
current injection produced an ASE pulse train due to gain modulation. For I
DC
90
mA, the amplitude oI ASE pulses was largest. The detected ASE pulse train is shown
in Fig. 6.33. A single captured waveIorm and averaged waveIorm is shown.
128

Fig. 6.32 Optical sampling setup with RF current injection

Fig. 6.33 Train oI ASE pulses due to RF current injection and gain modulation
The ODL was used to align the time position oI the sampling pulses with the
RF gain modulation, so the sampling pulses see the highest SOA gain. The best
overlap was Iound by observing the sampling pulses spectral amplitude on the OSA
(which should, oI course, be maximized) and by observing on the oscilloscope the
relative time position oI the sampling pulses aIter the SOA and the ASE pulses (the
time position oI the peak oI the sampling pulses without the RF modulation was noted
and, similarly, the time position oI the RF-induced ASE pulses was noted without
sampling pulses applied to the SOA). For optimum time overlap between the
sampling pulses and RF modulation, the sampling pulse power has increased by 4 dB,
whereas ASE power increased only by 0.2 dB. The alignment precision is limited by
129
the oscilloscopes resolution to about 100 ps, but the gain modulation pulses are wide
enough to provide a saIety margin. Fig. 6.34 shows the aligned ASE pulses, sampling
pulses aIter SOA and the combined shape detected on the oscilloscope. The
oscilloscope was triggered Irom Pritel RF output and the persistence mode was used
to record diIIerent waveIorms on one screen. The longer slope oI the detected pulse
when both RF current and sampling pulses are present is probably due to gain
saturation.

Fig. 6.34 Time aligned gain modulation and sampling pulses. Outer pulse detected when both RF
injection and sampling pulses present.
It should be noted that whereas the duration oI the sampling pulses is limited
by the speed oI the detection system, the duration oI the ASE pulses is clearly longer.
This may be due to the combined eIIect oI the limited bandwidth oI all the RF
electronics used to drive the SOA, SOA package parasitic capacitance and
spontaneous carrier liIetime. Based on this observation, it may be expected that no
SNR improvement can be obtained in this conIiguration, as the duration oI ASE
pulses is longer than the receiver integration time, so the receiver does not see any
ASE gating. This prediction was indeed conIirmed in the experiments. Moreover, the
conjugate energy seems to be lower than in the system with a DC bias only, due to
lower total peak SOA current and thereIore lower gain. However, exact measurements
130
were diIIicult because oI the very poor extinction ratio oI the samples, i.e. the ASE
pulses always gave a large contribution to the observed amplitude. The decision as to
how much oI the observed amplitude was contributed by the ASE pulses was always
somewhat arbitrary, especially considering the Iact that ASE is inherently random. It
is possible to obtain the ASE gating by using an additional external modulator aIter
the SOA (synchronized with sampling laser), which would allow biasing the SOA at
the maximum current. In this conIiguration, the gating window can be made much
shorter than the receiver integration time, but the reduced extinction ratio oI the
received samples due to ASE pulse would be present as well, thereIore its useIulness
is limited.
Other noise terms (i.e. signal shot noise, ASE shot noise and signal-ASE
beating) cannot be eliminated or even reduced. Their signiIicance can only be made
smaller Ior larger conversion eIIiciency and larger conjugate energy. In the typical
FWM scheme when two slow varying (or CW) waves are mixed, it is not diIIicult to
achieve high OSNR (see experiments described in Chapter 5). Similarly, in the
mixing between short pulses, high eIIiciency is obtained due to very high peak powers
and limited gain saturation. In the mixed regime, considered in this chapter, when a
CW signal and picosecond pump is used, the SNR is limited by strong gain saturation
due to the signal power. II the method was used in the actual network, this eIIect
strongly reduces the method applicability, since, Ior example, the 10 Gb/s signals
(still the most popular bit rate) would eIIectively be quasi-CW compared to the
sampling pulses. The nonlinear transIer Iunction oI the system, shown in Fig. 6.29 is
inherent to a saturable ampliIier and strongly limits the dynamic range oI the system.
The AB method improves somewhat the saturation power, as shown in Chapter 5, and
thereIore improves the dynamic range, but simulations show that in this case the
improvement is minor.
6.4. Concluding remarks on the applicability of the investigated method
In this chapter, the Ieasibility oI using FWM in SOAs to perIorm optical
sampling was studied both theoretically and experimentally. It was shown that when
the sampled signal is a CW wave, the achievable SNR and the nonlinear dependence
oI the conjugate energy on the input power limit the perIormance oI the sampling
system. The possible ways to improve the perIormance were discussed and some oI
them experimentally assessed. It must be concluded that the perIormance oI the
131
system is too low Ior an application oI the method as a measurement tool. Some SNR
improvement can be achieved by using a more stable sampling laser, but the
Iundamental limitations, especially the saturable characteristic oI the SOA, cannot be
circumvented. The only way to extend the dynamic range oI the method would be to
use an SOA with much larger saturation power. In the literature, there is little
discussion on the perIormance oI FWM between CW wave and picosecond pulses.
For example in |30| the applicability oI similar scheme to wavelength conversion was
considered. It was also concluded that the gain saturation strongly limits the SNR in
this case. ThereIore, FWM in SOAs is more useIul Ior sampling short pulses with
pulses, as usually reported in the literature. II relatively slow signals (low bandwidth)
are to be sampled, another technique can be used instead. As discussed in Section
6.3.1.3, the presence oI the sampling pulse causes spectral shiIt oI the signal wave
within the sampling window due to the XPM eIIect. The spectral shiIt can be as large
as 1 nm. II the signal bandwidth is lower than the induced spectral shiIt, this eIIect
can be used to perIorm optical sampling, aIter the shiIted signal is separated by an
optical Iilter. The sample energy in this case is much larger as it is given directly by
the ampliIied input signal energy.













132
6.5. References
|1| J. Li, J. Hansryd, P. O. Hedekvist, P. A. Andrekson, and S. N. Knudsen, "300-
Gb/s eye-diagram measurement by optical sampling using Iiber-based
parametric ampliIication," IEEE Photon. Technol. Lett., vol. 13, pp. 987-989,
2001.
|2| P. A. Andrekson and M. Westlund, "High-Speed WaveIorm Analysis Using
All-Optical Sampling," pp. 421-504, in Digital Communications Test and
Measurement. High-Speed Phvsical Laver Characteri:ation, D. Derickson
and M. Mller, Eds., Prentice Hall, 2007.
|3| C. Schmidt-Langhorst and H.-G. Weber, "Optical sampling techniques," J.
Opt. Fiber Commun. Reports, vol. 2, pp. 86-114, 2005.
|4| K.-L. Deng, R. J. Runser, I. Glesk, and P. R. Prucnal, "Single-shot optical
sampling oscilloscope Ior ultraIast optical waveIorms," IEEE Photon.
Technol. Lett., vol. 10, pp. 397-399, 1998.
|5| I. Shake, W. Takara, S. Kawanishi, and Y. Yamabayashi, "Optical signal
quality monitoring method based on optical sampling," Electron. Lett., vol. 34,
pp. 2152-2154, 1998.
|6| A. Rodriguez-Moral, P. BonenIant, S. Baroni, and R. Wu, "Optical data
networking: protocols, technologies, and architectures Ior next generation
optical transport networks and optical internetworks," J. Lightw. Technol., vol.
18, pp. 1855-1870, 2000.
|7| J.-P. Elbers and N. Peers, "OTN-Compatible 40 GBE and 100GBE
interIaces," presentation for IEEE 802.3 Higher Speed Studv Group Meeting,
http://ieee802.org/3/hssg/public/nov07/index.htm, 2007.
|8| L. A. Jiang, E. P. Ippen, U. Feiste, S. Diez, E. Hilliger, C. Schmidt, and H. G.
Weber, "Sampling pulses with semiconductor optical ampliIiers," IEEE J.
Quantum Electron., vol. 37, pp. 118-126, 2001.
|9| A. Mecozzi and J. Mork, "Saturation eIIects in nondegenerate Iour-wave
mixing between short optical pulses in semiconductor laser ampliIiers," IEEE
J. Sel. Topics Quantum Electron., vol. 3, pp. 1190-1207, 1997.
|10| M. ShtaiI and G. Eisenstein, "Analytical solution oI wave mixing between
short optical pulses in a semiconductor optical ampliIier," Appl. Phvs. Lett.,
vol. 66, pp. 1458-1460, 1995.
133
|11| C. Xie, P. Ye, and J. Lin, "Four-wave mixing between short optical pulses in
semiconductor optical ampliIiers with the consideration oI Iast gain
saturation," IEEE Photon. Technol. Lett., vol. 11, pp. 560-562, 1999.
|12| G. Agrawal, Nonlinear Fiber Optics, 3 ed: Academic Press, 2001.
|13| A. Uskov, J. Mork, and J. Mark, "Wave mixing in semiconductor laser
ampliIiers due to carrier heating and spectral-hole burning," IEEE J. Quantum
Electron., vol. 30, pp. 1769-1781, 1994.
|14| K. Kikuchi, M. Amano, C. E. Zah, and T. P. Lee, "Analysis oI origin oI
nonlinear gain in 1.5 mu m semiconductor active layers by highly
nondegenerate Iour-wave mixing," Appl. Phvs. Lett., vol. 64, pp. 548-550,
1994.
|15| M. Westlund, P. A. Andrekson, H. Sunnerud, J. Hansryd, and J. Li, "High-
perIormance optical-Iiber-nonlinearity-based optical waveIorm monitoring," J.
Lightw. Technol., vol. 23, pp. 2012-2022, 2005.
|16| A. M. Darwish, E. P. Ippen, H. Q. Le, J. P. Donnelly, and S. H. Groves,
"Optimization oI Iour-wave mixing conversion eIIiciency in the presence oI
nonlinear loss," Appl. Phvs. Lett., vol. 69, pp. 737-739, 1996.
|17| J. M. Tang and K. A. Shore, "Four-wave mixing oI strong picosecond optical
pulses in passive semiconductor waveguides," Appl. Phvs. Lett., vol. 74, pp.
2105-2107, 1999.
|18| S. Diez, R. Ludwig, C. Schmidt, U. Feiste, and H. G. Weber, "160-Gb/s
optical sampling by gain-transparent Iour-wave mixing in a semiconductor
optical ampliIier," IEEE Photon. Technol. Lett., vol. 11, pp. 1402-1404, 1999.
|19| I. Shake, E. Otani, H. Takara, K. Uchiyama, Y. Yamabayashi, and T. Morioka,
"Bit rate Ilexible quality monitoring oI 10 to 160 Gbit/s optical signals based
on optical sampling technique," Electron. Lett., vol. 36, pp. 2087-2088, 2000.
|20| K. Obermann, I. Koltchanov, K. Petermann, S. Diez, R. Ludwig, and H. G.
Weber, "Noise analysis oI Irequency converters utilizing semiconductor-laser
ampliIiers," IEEE J. Quantum Electron., vol. 33, pp. 81-88, 1997.
|21| M. Dinu and F. Quochi, "Amplitude sensitivity limits oI optical sampling Ior
optical perIormance monitoring," J. Opt. Netw., vol. 1, pp. 237-248, 2002.
|22| N. A. Olsson, "Lightwave systems with optical ampliIiers," J. Lightw.
Technol., vol. 7, pp. 1071-1082, 1989.
134
|23| G. Agrawal, Lightwave Technologv. Telecommunication Svstems. Hoboken:
Wiley, 2005.
|24| T. W. Berg, J. Mork, and J. M. Hvam, "Gain dynamics and saturation in
semiconductor quantum dot ampliIiers," New J. Phvs., vol. 6, pp. 178, 2004.
|25| S. Diez, C. Schmidt, R. Ludwig, H. G. Weber, K. Obermann, S. Kindt, I.
Koltchanov, and K. Petermann, "Four-wave mixing in semiconductor optical
ampliIiers Ior Irequency conversion and Iast optical switching," IEEE J. Sel.
Topics Quantum Electron., vol. 3, pp. 1131-1145, 1997.
|26| G. P. Agrawal and N. A. Olsson, "SelI-phase modulation and spectral
broadening oI optical pulses in semiconductor laser ampliIiers," IEEE J.
Quantum Electron., vol. 25, pp. 2297-2306, 1989.
|27| J. M. Tang and K. A. Shore, "DiIIerential Irequency shiIt oI picosecond pulse
Iour-wave mixing in semiconductor optical ampliIiers," IEEE Photon.
Technol. Lett., vol. 11, pp. 1129-1131, 1999.
|28| J. Inoue and H. Kawaguchi, "Highly nondegenerate Iour-wave mixing among
subpicosecond optical pulses in a semiconductor optical ampliIier," IEEE
Photon. Technol. Lett., vol. 10, pp. 349-351, 1998.
|29| M. Jinno, J. B. Schlager, and D. L. Franzen, "Optical sampling using
nondegenerate Iour-wave mixing in a semiconductor laser ampliIier,"
Electron. Lett., vol. 30, pp. 1489-1491, 1994.
|30| M. ShtaiI and G. Eisenstein, "Noise characteristics oI nonlinear semiconductor
optical ampliIiers in the Gaussian limit," IEEE J. Quantum Electron., vol. 32,
pp. 1801-1809, 1996.
|31| M. Westlund, H. Sunnerud, M. Karlsson, and P. A. Andrekson, "SoItware-
synchronized all-optical sampling Ior Iiber communication systems," J.
Lightw. Technol., vol. 23, pp. 1088-1099, 2005.

135
7. Alternative optical signal quality metric for BER
estimation
As explained in the Introduction to this thesis, as the complexity oI optical
networks grow and data rates increase, the need to estimate the optical signal quality
in an eIIicient and accurate way also becomes more pressing and apparent. This need
has long been recognized by the industry, and the Q-Iactor |1, 2|, discussed in
Chapter 4, is currently probably the most popular analogue measure oI the optical
signal quality, due to both reasonable accuracy and appealing simplicity |3-8|.
However, the Q-Iactor method has its limitations, as explained in Section 4.3,
originating in the Iact that the ASE-related beat noise terms at the photodetector
(which in most practical cases dominate over other noise sources in optically
ampliIied links) do not possess Gaussian nature |9, 10|. These limitations, in
particular, make this method unsuitable Ior applications requiring accurate
approximation oI not only the BER, but also the overall pdI shape or the decision
threshold. For example, in |11| the eIIect oI diIIerent noise models on the
perIormance oI the turbo code (a kind oI Iorward error correction code) decoders was
investigated and it was shown that the use oI Gaussian approximation may severely
degrade this perIormance. Similarly, it was Iound in |12| that the use oI Gaussian
approximation leads to very signiIicant degradation in the perIormance oI soIt
decoders based on turbo-product error correction codes. Likewise, the popular
approach to optimize the link design by maximizing the Q-Iactor is not necessarily the
truly optimal one (although very intuitive), as relation between the Q-Iactor and the
BER is unambiguous only in the Gaussian noise environment.
ThereIore, there is a need Ior an improved signal quality metric, which should
preIerably retain simplicity comparable with the Q-Iactor and should be easily
translatable into BER, but should also better model the actual noise characteristics at
the receiver. In this chapter, such an optical signal quality metric is proposed. This
metric utilizes knowledge about the Iirst three statistical signal moments and is based
on the Laguerre expansion to approximate the true probability density Iunction oI the
receiver noise. The work presented in this chapter has led to publication |13|.
136
7.1. Physical characteristics of the electrical noise at the receiver output
As argued in Chapters 1 and 4, accurate BER estimation is oI paramount
importance in optical communication links. Due to the very low BER required (oIten
10
-12
or better), direct BER measurements are diIIicult and time consuming.
Moreover, they can only be perIormed on an out-oI-service link, when some a priori
known bit sequence is sent, or at best at the signal termination point, using the error
detection capability oI the FEC chips, iI FEC is used |3, 14|. ThereIore, it is common
to estimate the BER and signal quality based on analytical models oI the output
receiver noise. Most oIten, the electrical noise is assumed to be Gaussian distributed,
i.e. the noise is assumed to have a normal pdI, leading (together with a particular
choice oI the decision threshold) to the Q-Iactor method, as summarized in Chapter 4.
Strictly speaking, in the case oI optically ampliIied links with direct detection, when
ASE noise is dominant, the output electrical noise is not Gaussian. This is because the
square law detection produces signal-ASE and ASE-ASE electrical beat noises and
the resultant pdI is not normal. When other sources oI nonlinearity in the link are
neglected, analytical models predict that the output electrical noise pdIs are given by
the central (Ior pure ASE power) and non-central (Ior signal and ASE both present)
chi-square distributions, under the assumptions that the electrical Iilter is oI the
integrate-and-dump type and the optical Iilter is a perIect rectangular Iilter |9, 10|.
The number oI degrees oI Ireedom in these chi-square distributions depends on the
optical Iilter bandwidth and the pdIs tend to Gaussian only when this number is
unrealistically large. Since the central and non-central chi-square pdIs are deIined
only Ior positive arguments and are not symmetrical, the convergence to the Gaussian
density is slow. Thermal noise, added in the post detection process, is in Iact
Gaussian, but since it is typically much less than electrical noise due to ASE, it does
not change the statistics signiIicantly |15|.
However, these more realistic statistics, even iI well known, are not oIten used
in practice, since they require the knowledge oI the number oI the noise modes and
noise power |9, 10| and also do not lead to convenient expressions. At the same time,
a straightIorward to use BER model better capturing the true characteristic oI the
noise would be oI signiIicant interest, as discussed above. ThereIore, in this Chapter,
a novel method to estimate BER is proposed based on the Laguerre expansion, which
approximates the real pdI using statistical moments and the gamma distribution. It
137
will be shown that this expansion provides a much better approximation to the real
statistics in the regime when ASE beating noises are dominant, thereIore allowing Ior
not only BER estimation, but also the optimum decision threshold estimation, which
cannot be obtained using the Gaussian approximation.
7.2. Derivation of the new metric
The model oI an optical link considered here is shown in Fig.7.1. This model
is essentially the same as considered in |9, 10|. An optical ampliIier with gain G and
ASE noise density S
ASE
is assumed but it can represent an equivalent ampliIier
(modeling a chain oI ampliIiers placed along a link). A similar understanding applies
to the optical Iilter. In this case, the input signal is noiseless and is described by its
average signal power P
av
and extinction ratio r. The electrical Iilter integrates the
signal over bit period T.

Fig. 7.1 Schematic oI the considered system
It has been shown |9, 10| that electrical noise pdIs due to ASE in direct
detection systems like the one in Fig.7.1 are analytically given by the central and non-
central chi-square distributions, Ior the spaces and marks, respectively. The non-
central chi-square distribution Ior the observed electrical current is given by:
|
|
.
|

\
|
|
|
.
|

\
| +
|
.
|

\
|
=

ASE
M
ASE
M
ASE
S
xE
I
S
E x
E
x
S
x p 2 exp
1
) (
1
2 / ) 1 (
1
(7.1)
whereas the central chi-square is given by
|
|
.
|

\
|

|
|
.
|

\
|

=

ASE
M
ASE ASE
S
x
S
x
M S
x p exp
)! 1 (
1
) (
) 1 (
0
(7.2)
where S
ASE
is the ASE power spectral density, E is the signal energy contained in one
bit and M is the number oI noise modes |9, 10|. These results have been derived using
a number oI assumptions and approximations, regarding Ior example the shape oI the
138
optical (perIect rectangular) and electrical (integrate-and-dump) Iilters and also the
neglect oI shot and thermal noise. Chan and Conradi |16| have experimentally
veriIied that this noise model is remarkably accurate and predicts the correct optimum
threshold. Other reIerences, however, have reported that in many practical lightwave
links, the marks pdI is indeed well approximated by the non-central chi-square
distribution but the distribution oI the spaces is diIIerent than predicted by the chi-
square |17, 18|. This is mainly due to non perIect extinction ratio and non ideal
Iiltering. In |17|, it was numerically Iound based on Monte Carlo simulations that the
pdI oI spaces can be well approximated by the generalized gamma distribution. All oI
these analytical pdIs (central chi-square, non-central chi-square and generalized
gamma) can be expressed in terms oI the standard gamma distribution, or reduce to
this distribution in limiting cases. They also share a number oI qualitative properties
with the gamma distribution: they are deIined Ior the positive axis only; they are non-
symmetrical and have exponentially decaying tails. ThereIore, we propose to
approximate the real noise statistics using statistical moments and the gamma pdI,
because oI its qualitative properties mentioned above. This approximation is based on
the Laguerre expansion |19|. The idea is similar to the wider-used Edgeworth
expansion, when an unknown density Iunction is expanded in terms oI the sum oI
Gaussian pdIs weighted by Hermite polynomials and coeIIicients dependent on the
moments oI the original Iunction |19|. However, the Edgeworth expansion is known
to perIorm poorly when used to approximate pdIs that are deIined only over the
positive axis or are asymmetrical |19, 20|. In these cases a much better expansion can
be developed in terms oI the Laguerre polynomials and gamma density.
Consider Iormally an expansion oI the unknown pdI p(x) into inIinite series:

=
=
0
0
) (
) ( ) ( ) (
n
n n
x p x L c x p
o
(7.3)
where p
0
(x) is a weight Iunction and L
n
(o)
(x) are some polynomials, orthogonal with
respect to the weight Iunction p
0
(x). II p
0
(x) is chosen as the (not normalized) gamma
distribution:
x
e x x p

=
o
) (
0
(7.4)
then L
n
(o)
(x) are the generalized Laguerre polynomials, given by the expression:
) (
!
) (
) ( o
o
o +

=
n x
n
x
n
x e
dx n
d x
e x L (7.5)
139
This system is known to Iorm a complete basis so it can be used to represent
any Iunctions deIined over the same domain |19|. Note that slightly diIIerent
deIinitions oI Laguerre polynomials are also sometimes used in the literature, the
Iorm used here is consistent with |19|. The Iirst Iew polynomials are
1 ) (
) (
0
= x L
o
(7.6a)
x x L + = ) 1 ( ) (
) (
1
o
o
(b)
2 ) (
2
2
1
) 2 ( ) 2 )( 1 (
2
1
) ( x x x L + + + + = o o o
o
(c)

3 2 ) (
3
6
1
) 3 (
2
1
) 3 )( 2 (
2
1
) 3 )( 2 )( 1 (
6
1
) ( x x x x L + + + + + + + = o o o o o o
o
(d)
and the pattern should be clear. The Laguerre polynomials are orthogonal over the
positive axis with the weight Iunction p
0
(x)
mn m n
n
n
dx x p x L x L o
o
o o
!
) 1 (
) ( ) ( ) (
0
0
) ( ) (
+ + I
=
}

(7.7)
where o
mn
is the Kronecker delta and I (.) is the (complete) gamma Iunction (not to be
conIused with the gamma distribution, given by (7.4)). Since the polynomials are
orthogonal, the coeIIicients c
n
can be calculated Irom (7.3) and (7.7) to be
}

+ + I
=
0
) (
) ( ) (
) 1 (
!
dx x p x L
n
n
c
n n
o
o
(7.8)
and, considering the Iorm oI the Laguerre polynomials (7.5, 7.6a-d), are related to
various statistical moments oI the unknown pdI p(x) (as (7.8) contains an integral oI
this pdI multiplied by a polynomial). ThereIore, in principle, the knowledge oI
moments can be used to recover the pdI. In practice, it is necessary to truncate (7.3)
aIter the Iirst Iew terms and, to Iurther reduce the number oI retained terms, it is
desirable to zero some oI the low-order terms in the expansion (7.3). To be able to
reduce (zero) the terms c
1,2,3
, a normalized random variable can be considered, given
by | / ) ( a x v = , where a, are, at this point, arbitrary numbers. The original pdI
p(x) can be recovered Irom the new pdI p
v
(v) by using the transIormation |21|
|
|
.
|

\
|
=
| |
a x
p x p
v
1
) ( (7.9)
and the ordinary moments m
vn
oI the new variable v can be expressed in terms oI
moments oI x using the relation
140
}

|
|
.
|

\
|
=
0
) ( dx x p
a x
m
n
vn
|
(7.10)
The expansion coeIIicients b
n
Ior p
v
(v) are calculated analogously to (7.8). The
Iirst Iew coeIIicients are given by
) 1 (
1
0
+ I
=
o
b (7.11a)
) 1 (
) 2 (
1
1
1
|
o
o
a m
b

+
+ I
= (b)
(

+
+

+ + +
+ I
=
2
2
1 2 1
2
2
) 2 ( 2 ) 2 )( 1 (
) 3 (
1
| |
o o o
o
a am m a m
b (c)
(

( +

+
+ +


+ + + + +
+ I
=
3
3
1
2
2 3
2
2
1 2
1
3
3 3 2
) 3 ( 3
) 3 )( 2 ( 3 ) 3 )( 2 )( 1 (
) 4 (
1
| |
o
|
o o o o o
o
a m a am m a am m
a m
b
(d)
where m
n
are the ordinary moments oI p(x). Higher order coeIIicients become more
and more complicated. Since at this point the coeIIicients o, , a are arbitrary, they
can be determined by requiring b
1,2,3
0 to reduce the number oI terms in the
expansion Ior required accuracy. This system oI equations has only one solution, and
this triplet is given by:
1
) (
, ,
2
2
2
1
2
1
2
3

= = =
o
o
|
o
o

|
m a
m a (7.12)
where
n
are the central moments oI p(x). Higher order terms can be derived iI
required. Finally, the approximated pdI p(x) can be expressed as

=
|
|
.
|

\
|
|
|
.
|

\
|
=
0
0
) (
) (
n
n
n
a x
p
a x
L
b
x p
| | |
o
(7.13)
The number oI terms taken into account depends on the required accuracy. In
this chapter, mainly the third order approximation will be considered (i.e. retaining
only b
0,1,2,3
terms, albeit with b
1,2,3
zeroed due to the choice oI parameters) to show
that it can give reasonable results. The Iourth order term (which is non-zero Ior the
parameterization used) could be easily added, at the expense oI increased complexity
(mostly length) oI the equations and similarly with any other higher order terms.
Later, we will also discuss the second order expansion, with simpliIied
141
parameterization. In the third order case considered here, the approximated pdIs oI
marks and spaces ) (
~
), (
~
1 0
x p x p are given by
|
|
.
|

\
|
|
|
.
|

\
|
+ I
=
0
0
0
0
0 0
0
exp
) 1 (
1
) (
~
0
| | | o
o
a x a x
x p (7.14)
|
|
.
|

\
|
|
|
.
|

\
|
+ I
=
1
1
1
1
1 1
1
exp
) 1 (
1
) (
~
1
| | | o
o
a x a x
x p (7.15)
where o
0,1
,
0,1
, a
0,1
are calculated separately Ior marks and spaces Irom (7.12).
ThereIore, the estimated BER, depending upon the decision threshold D is calculated
Irom
} }

+ =
D
D
dx x p dx x p D BER
0
0 1
) (
~
2
1
) (
~
2
1
) ( (7.16)
The optimum threshold D
opt
is Iound when the BER is minimum, which is
equivalent to requiring that ) (
~
) (
~
1 0
D p D p = (i.e. the pdIs intersect). Hence, the
optimum threshold is the solution oI the Iollowing nonlinear equation
C AD B
a D
a D
+ =

0
1
) (
) (
ln
0
1
o
o
(7.17)
where
1 0
1 0
/
o o
| | = B ,
1 0 1 0
/ ) ( | | | | = A ,
0 0
1 1
1 0
1 0 0 1
) 1 (
) 1 (
ln
| o
| o
| |
| |
+ I
+ I
+

=
a a
C .
This equation has to be solved numerically in general case but this task is not
too diIIicult. When the threshold is Iound, BER is Iinally given in terms oI the (upper)
incomplete gamma Iunctions as
) 1 ( 2
) ) ( , 1 (
) 1 ( 2
) ) ( , 1 ( ) , 1 (
0
0 0 0
1
1 1 1 1 1 1
+ I
+ I
+
+ I
+ I + I
=
o
| o
o
| o | o a D a D a
BER
opt opt

(7.18)
where the deIinition oI the incomplete gamma Iunction I(s,v) adopted here is
}


= I
v
t s
dt e t v s
1
) , ( .
The ratio oI the incomplete gamma Iunction to the (complete) gamma Iunction
appearing in (7.18) is sometimes known as the regularized or normalized (Ior obvious
reason) incomplete gamma Iunction Q(s,v), so
142
. , 1 , 1 , 1
2
1
0
0
0
1
1
1
1
1
1
(
(

|
|
.
|

\
|
+ +
|
|
.
|

\
|
+
|
|
.
|

\
|
+ =
|
o
|
o
|
o
a D
Q
a D
Q
a
Q BER
opt opt
(7.19)
The Iunction Q(s,v) is implemented directly in many mathematical packages,
thereIore the Iormula (7.19) is straightIorward to use. Eq. (7.17) and (7.19) with the
appropriate parameters given by (7.12) constitute the new signal quality metric. As
can be observed, the new metric is based on the statistical moments oI the noise
distribution, and thereIore is well suited to be used with histograms, generated Ior
example using optical sampling, as discussed in Chapter 3. Unlike the Gaussian
approximation, the new metric makes use oI the higher order moments as well, and
can be, in principle, extended to include moments oI an arbitrary order. Eq. (7.19) is
based on the third order Laguerre expansion; thereIore it requires the knowledge oI
the Iirst three moments.
7.3. Evaluation of the proposed signal quality metric
In this section, the new metric derived above is compared with the Gaussian
approximation and the exact (Ior the simpliIied system model considered here) BER
given by the chi-square model (7.1) and (7.2). It is relevant to note that the use oI the
real statistics is not Ieasible in practice, as they require the knowledge oI the number
oI the noise modes in the channel optical bandwidth and oI the ASE noise power,
render the calculations diIIicult and also are derived under a number oI assumptions.
ThereIore, it is more practical to use a good and Ilexible approximation, which can be
Iitted to the pdI obtained in non-ideal conditions. The evaluation oI the new metric
presented in this section is given Ior 10 Gb/s NRZ signal. The system shown in Fig.
7.1 is considered again, with a Iairly typical choice oI its parameters. The system
parameters were: the optical preampliIier gain G 30 dB, the ampliIier spontaneous
emission coeIIicient n
sp
2 and the carrier Irequency f 193 THz. The optical noise
density is given by ) 1 ( = G hf n S
sp ASE
, where h is Planck`s constant, and the optical
bandwidth B
o
was 10bit rate. Initially, a Iinite extinction ratio r oI 10 dB is
considered. ThereIore, both mark and space pdIs are given by the non-central chi-
square densities (7.1), with diIIerent non-centrality parameters. Input average signal
power is denoted by P
av
. Because thermal noise is neglected, this amounts to the case
when ASE (i.e. beating noises due to ASE in the electrical domain) is the dominant
signal impairment, which is in most cases the practical assumption.
143
Fig. 7.2 compares exact pdIs, given by the non-central chi-square
distributions, with their respective third order Laguerre approximations in the region
oI practical interest. The best Gaussian Iit is also plotted in Fig. 7.2 Ior comparison
purposes. The signal and noise conditions corresponded to BER just better that 10
-9
.
Normalized current is plotted Ior numerical convenience and the normalization
constant was M/(P
N
R) (i.e. the multiplication oI the normalized current by P
N
R/M
would recover the current in amperes). Here M is the number oI noise modes in given
optical bandwidth, P
N
is the ASE noise power and R is the receiver sensitivity.
Although single polarization is used, including the other polarization state simply
doubles M |9, 15|. A logarithmic scale is used Ior clarity. The same normalization is
consistently used throughout this chapter. It can be seen that the proposed
approximation provides good Iit and the intersection point is very close to the one
given by the real statistics, whereas the Gaussian Iit is dramatically diIIerent Irom the
real pdIs. This good quality oI the Laguerre approximation is consistent with that
reported previously in the relevant statistical literature |22|. Fig. 7.2 suggests that not
only should accurate BER estimation be possible, but also that good optimum
threshold estimation can be given by the proposed method. This is in contrast with the
Gaussian approximation (Q-Iactor), which is known to seriously underestimate the
decision threshold |9, 10|, as indeed is clear in Fig. 7.2.

Fig. 7.2 Comparison oI the Laguerre expansion (LE) with the exact pdIs (CHI). Gaussian Iit (GA) is
also shown. The normalization constant is M/(P
N
R), conditions correspond to BER10
-9
: P
av
-36.4
dBm, r10 dB, G30 dB, n
sp
2, B
o
10bit rate
144
To Iurther investigate this issue, Fig. 7.3 plots BER vs. decision threshold Ior
the same conditions as in Fig. 7.2. Additionally, the Gaussian approximation case is
plotted in the same Iigure, Ior comparative purposes. It can be seen that optimum
threshold prediction given by the new metric is indeed very close to the real optimum
threshold. Gaussian approximation provides reasonably good BER estimation, but
does it accidentally`, in the sense that the decision threshold is Iar Irom the optimum
value.

Fig. 7.3 Comparison oI BER vs. decision threshold Ior the Laguerre expansion (LE), exact pdIs (CHI)
and Gaussian approximation (GA). Conditions: P
av
-36.4 dBm, r10 dB, G30 dB, n
sp
2, B
o
10bit
rate
Fig. 7.4 compares the BER calculated Irom all the 3 statistical models (real
pdIs, 3rd order Laguerre expansion and Gaussian approximation) Ior diIIerent average
signal power P
av
beIore the ampliIier. The BER Irom the new metric is calculated
straightIorwardly using (7.17), and (7.18). To calculate the expansion parameters,
(7.12) was used. It can be seen that in this case both approximations provide very
good estimation over a wide range oI BER. Both estimations are slightly conservative.
A slight loosening oI tightness oI the new metric can be seen Ior very low BER. It
might be surprising that Gaussian approximation provides good BER estimation,
whereas Figs. 7.2 and 7.3 clearly show that the Gaussian model is inadequate. The
reason is that Iortuitously the large errors due to the use oI wrong approximating pdIs
Ior marks and spaces cancel each other. Fig. 7.2 shows that Gaussian approximation
(working Irom its intersection point) seriously underestimates the weight oI the tail oI
145
the spaces pdI, but overestimates the tail oI the marks pdI. The two errors have
diIIerent sign and it is very Iortunate that they largely compensate each other. On the
other hand, Fig. 7.2 shows that the error in the third order Laguerre expansion is very
small, but the two error contributions have the same sign, thereIore they add up.
Hence Ior very low BER the Gaussian estimation is actually slightly better. For this
set oI system parameters however, the main advantage oI the new metric can be seen
when the optimum threshold obtained using this metric is compared with the one
given by Gaussian approximation.

Fig. 7.4 Comparison oI BER vs. average signal power Ior the 3rd order Laguerre expansion (LE), exact
pdIs (CHI) and Gaussian approximation (GA), G30 dB, n
sp
2, B
o
10bit rate, r10 dB
Fig. 7.5 shows the optimum decision threshold predicted by the two
approximations and compares it with the true one, Ior diIIerent average input signal
powers. The threshold given by the Laguerre expansion is very close to the real
optimum threshold, whereas Gaussian approximation gives a threshold very diIIerent
Irom the real optimum. ThereIore the Laguerre expansion should be used in
applications which require a good estimation oI the general shape oI the pdIs and oI
the optimum threshold. For example, as discussed in |11, 12|, poor estimation oI the
decision threshold may signiIicantly degrade the perIormance oI turbo-code decoders
and in this case the new model is expected to perIorm much better. II the model
presented here is applied to a channel with dispersion, the presence oI dispersion
eIIectively splits the two logical levels into a number oI sublevels, as discussed in
Chapter 4. Each sublevel is described by its own conditional pdI, where the
146
conditioning is on the value (mark or space) oI the signiIicant preceding and
Iollowing bits. The standard procedure in this case is to calculate the unconditional
pdI by averaging over all possible cases |12|. The Laguerre expansion model should
lend itselI to such extension, in a similar way the Gaussian approximation does. In
many practical scenarios, usually only one preceding and one Iollowing bit has to be
considered |12|.

Fig. 7.5 Comparison oI the normalized decision threshold Ior the 3rd order Laguerre expansion (LE),
exact pdIs (CHI) and Gaussian approximation (GA), G30 dB, n
sp
2, B
o
10bit rate, r10 dB.
The good BER estimation given by the Gaussian approximation in the
discussed case is consistent with the results reported previously |15|. Low extinction
ratio (r10 dB) means that the output noise Ior both marks and spaces have quite
large Gaussian contributions (ASE-signal beat noise). Although the decision threshold
is very Iar Irom the real optimum, the noise variances Ior both marks and spaces
change in approximately the same way as input signal power is changed, as signal
dependent noise dominates Ior the two signal levels. ThereIore the errors resulting
Irom the Gaussian approximation oI each individual pdI can cancel out, even iI it
happens accidentally`. The choice oI a relatively large value oI M (10) also slightly
improves the quality oI Gaussian approximation, although Fig. 7.3 shows that the pdIs
are still very diIIerent Irom Gaussian.
It is however known that in other network scenarios BER estimation given by
the Gaussian approximation can be much less accurate. For example, this can be
expected iI ASE-ASE noise is dominant Ior the space level, i.e. iI a good extinction
147
ratio is used. In these cases Gaussian approximation is expected to be inaccurate and
usually conservative |15|. To test how the new metric based on the 3rd order Laguerre
expansion perIorms in these cases, the BER Ior a new optical link conIiguration was
calculated. In this case, the system is again as shown in Fig. 7.1, and the only
parameter changed is the extinction ratio r20 dB. Fig. 7.6 shows BER predicted
using the new metric and Gaussian approximation and compares it with the BER Irom
the real statistics. It can be clearly seen that Gaussian approximation overestimates
BER in this case by about an order oI magnitude and this inaccuracy depends on the
signal power, whereas the new metric provides a much better estimate. Obviously, in
this case too the threshold predicted by Gaussian approximation is very inaccurate,
whereas the one given by Laguerre expansion is a very good estimate, as seen in Fig.
7.7. Figs. 7.6 and 7.7 show the superiority oI the new metric in this network scenario.

Fig. 7.6 Comparison oI BER vs. average signal power Ior the 3rd order Laguerre expansion (LE), exact
pdIs (CHI) and Gaussian approximation (GA), G30 dB, n
sp
2, B
o
10bit rate, r20 dB
It is also worth mentioning that the Gaussian approximation may be grossly
inaccurate in incoherent spectrum sliced transmission, which has received some
attention |23, 24|. This inaccuracy is due to the noise-like character oI the inIormation
signal. Since the statistics oI the received signal in this case are described by the
(central) chi-square distribution |23| (assuming Ilat spectrum), the Laguerre
expansion method is expected to be well-suited here, as the basis oI the expansion is
the gamma distribution, which is a generalization oI the chi-square.
148

Fig. 7.7 Comparison oI the normalized decision threshold Ior the 3rd order Laguerre expansion (LE),
exact pdIs (CHI) and Gaussian approximation (GA), G30 dB, n
sp
2, B
o
10bit rate, r20 dB.
7.3.1. Second order version of the new metric
The use oI the metric based on 3rd order Laguerre expansion requires the
knowledge oI the Iirst three statistical moments. Additional inIormation about the
signal, contained in the third moment, allows accurate estimation oI the optimum
threshold or improved BER estimation, as seen in Fig. 7.6. InIormation about the
moments can be conveniently obtained Irom the sampled histograms, constructed Ior
example using the methods oI optical sampling, discussed in Chapter 3. However, one
may not be willing to use third order moments, as it involves some additional signal
processing, or perhaps the Iirst two moments only are estimated in some other way,
not Irom the histograms. In this case, it is still possible to derive an approximation
based on the Laguerre expansion making use oI only the Iirst two moments. In this
section, the quality oI such a second order metric is brieIly evaluated, although it
might be expected to be less accurate than the third order expansion, as it uses less
inIormation about the signal. To derive the second order expansion, one proceeds
exactly in the same way as it was done to derive the third order one. The only
diIIerence is that now the auxiliary normalized random variable v is given by vx/,
i.e. only a 2-parametetrized expansion is used (with parameters o and and with a
restricted to a0). ThereIore, only the coeIIicients b
1,2
can be reduced (zeroed). II all
higher order non-zero terms are neglected, this gives a second order expansion. To
zero b
1,2
,
1
~
o ,
1
~
| are now given by:
149
1
~
,
~
2
2
1
1
1
2
1
= =
o
o
o
|
m
m
(7.19)
BER is again estimated using (7.17) and (7.18) with this new set oI
parameters. To assess the quality oI this simpliIied metric, the same network scenarios
are considered as beIore. Figs. 7.8 and 7.9 show the BER and optimum decision
threshold estimation obtained using the 2nd order expansion Ior extinction ratio r10
dB.

Fig. 7.8 Comparison oI BER vs. average signal power Ior the 2nd order Laguerre expansion (LE(2nd)),
exact pdIs (CHI) and Gaussian approximation (GA), G30 dB, n
sp
2, B
o
10bit rate, r10 dB
In this case, the simpliIied metric gives good BER estimation; however the
decision threshold estimation is visibly worse than given by the third order Laguerre
expansion (Fig. 7.9). Nevertheless, it is still reasonably accurate and certainly better
than that given by the Gaussian approximation. The expected poorer BER accuracy oI
the simpliIied metric has not shown up in this case but in general the third order
expansion is preIerred. This is evidenced in Figs. 7.10, 7.11 when the high extinction
ratio case is shown. In this case the inaccuracy oI the simpliIied metric is similar to
the inaccuracy oI the Gaussian approximation, but the sign oI the estimation error is
opposite. In practice, the conservative error in the Gaussian approximation estimation
is usually more acceptable. The threshold estimation also in this case is worse than
that given by the third order expansion, but much better than that Irom Gaussian
approximation. ThereIore, having only the Iirst two moments available, one may opt
150
Ior calculating BER Irom the Q-Iactor (Gaussian approximation) and the
approximated optimum decision threshold Irom the second order Laguerre expansion.

Fig. 7.9 Comparison oI the normalized decision threshold Ior the 2nd order Laguerre expansion (LE),
exact pdIs (CHI) and Gaussian approximation (GA), G30 dB, n
sp
2, B
o
10bit rate, r10 dB.

Fig. 7.10 Comparison oI BER vs. average signal power Ior the 2nd order Laguerre expansion
(LE(2nd)), exact pdIs (CHI) and Gaussian approximation (GA), G30 dB, n
sp
2, B
o
10bit rate, r20
dB
151

Fig. 7.11Comparison oI the normalized decision threshold Ior the 2nd order Laguerre expansion (LE),
exact pdIs (CHI) and Gaussian approximation (GA), G30 dB, n
sp
2, B
o
10bit rate, r20 dB.
The region oI validity oI the new metric is restricted to the cases when ASE
noise is the dominant signal impairment, which is usually the case in quasi-linear
optically ampliIied links, and when direct detection is used. It is currently the most
common practice to extend the achievable link reach by using optical ampliIication,
as outlined in Chapter 1. II however the thermal noise is dominant (no optical
ampliIication), the total noise is approximately Gaussian and the Q-Iactor method is
preIerred. Similarly, iI amplitude modulation with coherent detection is used, the
Gaussian nature oI the ASE noise is retained aIter the detection process, and the Q-
Iactor method should be strictly valid, iI no other sources oI nonlinearity are present
in the link. However, as shown in Chapter 9, in particular cases nonlinearity
introduced by the optical Iibre itselI may be non-negligible. In this case the Q-Iactor
method ceases to be valid even in coherent detection systems and its application may
lead (unexpectedly) to very optimistic BER estimation.
Within its region oI validity, the tightness oI the Laguerre expansion is slightly
reduced Ior very low BER, but it is still accurate in most cases oI practical interest. As
the method is based on moments, it is well suited to be used with sampling
histograms, although it should be remembered that iI optical sampling is used, the Iar
greater sampling bandwidth needs to be taken into account, as discussed in Chapter 3,
Ior example by using the Iormulae presented in Chapter 4. Similarly, the metric
presented here may be used as a perIormance measure Ior optimizing link design
during the simulation phase.
152
7.4. Summary
In this chapter, we have presented a novel optical signal quality metric, based
on the Laguerre expansion and statistical moments. The metric can be used to
estimate the BER in direct detection systems in cases when ASE noise is the dominant
phenomenon aIIecting the signal quality. The conditions oI validity oI the metric have
been identiIied and discussed. Numerical calculations show that the metric gives
accurate estimation oI the BER and also oI the optimum decision threshold, which is a
signiIicant advantage over the Gaussian approximation. Additionally, the new model
can accurately approximate the output noise pdIs due to the beating between the
signal and ASE and ASE-ASE. At the same time, the new method retains simplicity
comparable with the Gaussian approximation. The method can be used with signal
histograms, as it is based on statistical moments. A simpliIied version oI the new
metric, based on the second order Laguerre expansion, have been also investigated
and shown that usually the more complex version is preIerred. However, the second
order metric can be also oI some value, when only the knowledge oI the Iirst two
moments is available and a reasonably good estimation oI the decision threshold is
needed, in addition to the BER estimation.



153
7.5. References
|1| S. D. Personick, ""Receiver design Ior digital Iiber optic communication
systems"," Bell Svst. Tech. J., vol. 52, pp. 843-886, 1973.
|2| F. Matera and M. Settembre, "Role oI Q-Iactor and oI time jitter in the
perIormance evaluation oI optically ampliIied transmission systems," IEEE J.
Sel. Topics Quantum Electron., vol. 6, pp. 308-316, 2000.
|3| N. S. Bergano, F. W. KerIoot, and C. R. Davidsion, "Margin measurements in
optical ampliIier system," IEEE Photon. Technol. Lett., vol. 5, pp. 304-306,
1993.
|4| C. J. Anderson and J. A. Lyle, "Technique Ior evaluating system perIormance
using Q in numerical simulations exhibiting intersymbol interIerence,"
Electron. Lett., vol. 30, pp. 71-72, 1994.
|5| S. Norimatsu and M. Maruoka, "Accurate Q-Iactor estimation oI optically
ampliIied systems in the presence oI waveIorm distortions," J. Lightw.
Technol., vol. 20, pp. 19-27, 2002.
|6| C. Schmidt, C. Schubert, J. Berger, M. Kroh, H.-J. Ehrke, E. Dietrich, C.
Borner, R. Ludwig, H. G. Weber, F. Futami, S. Watanabe, and T. Yamamoto,
"Optical Q-Iactor monitoring at 160 Gb/s using an optical sampling system in
an 80 km transmission experiment," Conf. Lasers Electro-Opt. CLEO 2002
Technical Digest., vol. 1, pp. 579-580, 2002.
|7| I. Shake and H. Takara, "Averaged Q-Iactor method using amplitude
histogram evaluation Ior transparent monitoring oI optical signal-to-noise ratio
degradation in optical transmission system," J. Lightw. Technol., vol. 20, pp.
1367-1373, 2002.
|8| M. Westlund, H. Sunnerud, M. Karlsson, J. Hansryd, J. Li, P. O. Hedekvist,
and P. A. Andrekson, "All-optical synchronous Q-measurements Ior ultra-high
speed transmission systems," Opt. Fiber Commun. Conf. OFC 2002, pp. 530-
531, 2002.
|9| D. Marcuse, "Derivation oI analytical expressions Ior the bit-error probability
in lightwave systems with optical ampliIiers," J. Lightw. Technol., vol. 8, pp.
1816-1823, 1990.
|10| P. A. Humblet and M. Azizoglu, "On the bit error rate oI lightwave systems
with optical ampliIiers," J. Lightw. Technol., vol. 9, pp. 1576-1582, 1991.
154
|11| Y. Cai, J. M. Morris, T. Adali, and C. R. Menyuk, "On turbo code decoder
perIormance in optical-Iiber communication systems with dominating ASE
noise," J. Lightw. Technol., vol. 21, pp. 727-734, 2003.
|12| G. Bosco, G. Montorsi, and S. Benedetto, "SoIt decoding in optical systems,"
IEEE Trans. Commun., vol. 51, pp. 1258-1265, 2003.
|13| M. Dlubek, A. Phillips, and E. Larkins, "Optical signal quality metric based on
statistical moments and Laguerre expansion," Opt. Quantum Electron., vol.
40, pp. 561-575, 2008.
|14| A. F. Richter, W. Bock, H. Bach, R. Grupp, W., "Optical perIormance
monitoring in transparent and conIigurable DWDM networks," IEE Proc.-
Optoelectron., vol. 149, pp. 1-6, 2002.
|15| D. Marcuse, "Calculation oI bit-error probability Ior a lightwave system with
optical ampliIiers and post-detection Gaussian noise," J. Lightw. Technol.,
vol. 9, pp. 505-513, 1991.
|16| B. Chan and J. Conradi, "On the non-Gaussian noise in erbium-doped Iiber
ampliIiers," J. Lightw. Technol., vol. 15, pp. 680-687, 1997.
|17| Y. Kopsinis, J. S. Thompson, and B. Mulgrew, "System-independent threshold
and BER estimation in optical communications using the extended generalized
gamma distribution," Opt. Fiber Technol., vol. 13, pp. 39-45, 2007.
|18| N. Alic, G. Papen, R. Saperstein, L. Milstein, and Y. Fainman, "Signal
Statistics and Maximum Likelihood Sequence Estimation in Intensity
Modulated Fiber Optic Links Containing a Single Optical Pre-ampliIier," Opt.
Express, vol. 13, pp. 4568-4579, 2005.
|19| S. Primak, V. Kontorovitch, and V. Lyandres, Stochastic Methods and their
Applications to Communications. Stochastic Differential Equations Approach:
John Wiley & Sons, 2005.
|20| E. Gaztaaga , P. Fosalba , and E. Elizalde, "Gravitational Evolution oI the
Large-Scale Probability Density Distribution: The Edgeworth and Gamma
Expansions," Astrophvs. J., vol. 539, pp. 522-531, 2000.
|21| A. Papoulis, Probabilitv, random variables, and stochastic processes, 3 ed:
McGraw-Hill, 1991.
|22| M. L. Tiku, "Laguerre series Iorms oI non-central X2 and F distributions,"
Biometrika, vol. 52, pp. 415-427, 1965.
155
|23| V. Arya and I. Jacobs, "Optical preampliIier receiver Ior spectrum-sliced
WDM," J. Lightwave Technol., vol. 15, pp. 576-583, 1997.
|24| M. S. Leeson, "PerIormance Analysis oI Direct Detection Spectrally Sliced
Receivers Using Fabry-Perot Filters," J. Lightwave Technol., vol. 18, pp. 13,
2000.


156
8. OSNR monitoring method based on modulation spectrum
assessment
It was outlined in the Introduction to this thesis, that even iI there are
numerous optical signal parameters measurable using OPM techniques, some oI these
parameters are oI much greater practical signiIicance than other. Furthermore, it was
discussed that optical signal to noise ratio (OSNR) is among the most important
optical quantities characterizing the signal quality due to both widespread practice oI
using optical ampliIiers to extend the link length and due to its relationship with the
Q-Iactor (and thereIore with BER) in linear optically ampliIied links |1, 2|. When
other signal impairments are present, the simple Iormulae linking OSNR and BER are
no longer valid, but even then, due to its simplicity, OSNR is still used to estimate the
link quality, with some penalties added to account Ior other eIIects.
ThereIore, OSNR monitoring is a very important aspect oI OPM. At the same
time, modern optical networks pose certain diIIiculties Ior accurate measurement oI
this quantity, mainly due to the Iact that each channel can have diIIerent history and
due to tight Iiltering. In this chapter, it is discussed how these diIIiculties impact the
traditional out-oI-band OSNR measurement techniques. An alternative novel cost-
eIIective in-band OSNR monitoring technique is proposed to overcome these
problems. The robustness oI the method against other signal impairments is assessed
through simulations. The work presented in this chapter has led to publications |3, 4|.
8.1. State-of-the-art in OSNR monitoring methods
Because oI the importance oI OSNR, there has been considerable research
eIIort directed towards Iinding a reliable and practical technique to measure this
quantity |2, 5-9|. The task oI accurately measuring the OSNR becomes more and
more diIIicult as the complexity oI optical networks increases, especially as the ever
higher spectral eIIiciency is required and practically implemented and as the
ROADMs become truly Ilexible, allowing ASONs to slowly become a reality |10,
11|. This evolution means that the channel spacing bandwidth is reduced and the
neighbouring channels may have traversed through diIIerent number oI ampliIiers.
The still most commonly used OSNR measurement technique is based on the
linear interpolation oI the ASE noise level measured between the optical channels and
157
on the assumption that this level is representative Ior ASE noise within each channel
|12|. However, this out-oI-band measurement technique Iails when either WDM
channels are spaced too closely, tight optical Iiltering is used in the network or when
the network is reconIigurable, i.e. channels have propagated through diIIerent routes.
For these reasons, an in-band monitoring technique is required to accurately measure
the OSNR in each channel. In the case oI in-band techniques, the signal and noise
spectra are overlapping, so it is necessary to Iind some other Ieatures which make it
possible to distinguish between the signal power and the noise power.
One way is to use the polarisation properties oI the signal and ASE noise, e.g.
to perIorm measurements based on orthogonal polarisation heterodyning |8| or
polarisation-nulling |7|. Another way is to use the diIIerent correlation properties oI
the signal and uncorrelated ASE noise. For example, the uncorrelated beat noise
measurement perIormed in |5, 9| utilized this eIIect. Finally, a method was proposed,
which utilized some knowledge oI the modulation Iormat (i.e. made use oI the radio-
Irequency (RF) power spectral density at low Irequency) combined with the
incoherent property oI the ASE noise in RF spectrum analysis |6|.
The above mentioned methods, although quite successIul, still suIIer Irom
some shortcomings. This could be polarisation dependency |7| or relatively
complicated setup, as ReIs. |5, 8| need both narrowband optical and electrical
Iiltering. The setup complexity is particularly disadvantageous in metro networks,
which have recently adopted the dynamic ROADM technology and which are very
cost-sensitive |13|. On the other hand, application oI the method presented in |6| to
intensity modulated signals (still prevalent in metro networks) is more problematic.
Commercially available approaches to tackle the problem oI measuring the in-band
OSNR in tightly Iiltered networks include using an optical spectrum analyser (OSA)
with very narrow Iilters, possibly combined with polarisation diversity |14, 15|, which
however inherits some oI the problems oI polarisation dependency. ThereIore, it
appears that even iI an inexpensive but robust and reasonably accurate OSNR
measurement technique is oI great practical importance, such a technique has not yet
been Iound.
In this chapter, a novel OSNR monitoring method is proposed, which requires
only that the modulation Iormat and bit rate are a priori known. Those requirements
are also implicitly present in the case oI most other OSNR monitoring. Compared to
other monitoring methods, the method proposed here requires narrowband optical
158
Iiltering, similarly to methods proposed in |5, 8| but does not require any other
complex instrumentation. It is also worth noting that a precision OSA is oIten already
available (as discussed in |14, 15|), thereIore upgrade to the method proposed here
may be very cost-eIIective. The simplicity oI the method and oI its implementation
makes it well-suited Ior cost-sensitive applications, such as metro networks.
8.2. Principle of the proposed OSNR monitoring method
The method proposed in this chapter is based on the simple observation that
Ior a given modulation Iormat and bitrate, the power spectrum S(f) oI the optical
signal is known. Although this spectrum depends on the transmitted bit pattern, it
converges to some average value Ior all practical bit sequences, iI the measurement
time is suIIiciently long compared with the bit period. However, the required
measurement time is still short compared with the typical requirements Ior the
monitoring equipment. As an example, it may be seen that at 10 Gb/s, a 1 ms
measurement time covers 10 million bits. It was Iound Irom simulations that this
should be suIIicient to give representative spectrum. In practice, even measurement
times longer by an order oI magnitude should be acceptable.
The method proposed uses two narrowband optical Iilters with the transIer
Iunctions F
1
(f), F
2
(f) to extract two spectral slices Irom the channel spectrum. The two
optical powers are then measured using standard slow optical power meters. The ratio
oI the optical powers in the two spectral intervals in the no-noise case is given only by
the modulation Iormat and bit rate. However, iI a uniIorm ASE spectral density is
present, it will shiIt up the entire spectrum by the same amount, thereby changing the
power ratio. We propose to use this power ratio as an OSNR indicator. Simple
analysis shows that Ior suitably chosen Iilters, this new indicator can be made quite
linear. Fig. 8.1 shows a simpliIied optical spectrum with the two optical Iilters to
extract the spectral slices. The mathematical derivation oI the proposed OSNR metric
presented below gives practical guidelines as to how the Iilters should be chosen.
During the derivation the ASE noise Iloor is assumed Ilat throughout the channel. The
EDFA ASE spectrum usually does not vary signiIicantly over channel bandwidth.
Even iI optical Iilters introduce ASE reshaping, this eIIect is much smaller in the
passband region to avoid signal distortion. It should be appreciated that the bandwidth
oI the inline Iilters is likely to impose practical limitations on the spacing between the
OSNR monitoring Iilters and the channel centre Irequency. Also, it is worth noting
159
that the assumption that the ASE slices are representative Ior the majority oI the in-
band noise is common to most methods utilizing narrowband Iiltering (e.g. |5, 8|).

Fig. 8.1 Principle oI OSNR monitoring using modulation spectrum assessment with the two Iilters
F
1
(f), F
2
(f). The inset shows the same idealized spectrum without the ASE contribution
The optical powers aIter the Iilters oI the transIer Iunctions F
1
(f), F
2
(f) are
denoted by P
1
and P
2
and are given by (neglecting the thermal noise contribution,
negligible Ior averaging power meters)
}


+ =
ASE
S df f S f F P
1
2
1 1
NEB ) ( ) ( (8.1)
} }


+ = + =
ASE ASE
S X df f S f F S df f S f F P
1
2
2 2
2
2 2
NEB ) ( ) ( NEB ) ( ) ( (8.2)
where NEB
1,2
are the noise equivalent bandwidths oI the two Iilters, with X
NEB
2
/NEB
1
, and S
ASE
is the (Ilat) ASE spectral density. In the noiseless case, the ratio
K
P
oI the two powers is determined by the modulation Iormat and given by
} }


= df f S f F df f S f F K
P
) ( ) ( ) ( ) (
2
1
2
2
(8.3)
The noiseless ratio K
P
can be conveniently obtained Irom simulations or
measurements during precalibration. In the noisy case, the measured ratio K
N
would
be:
|
|
.
|

\
|
+
|
|
.
|

\
|
+ =
} }


ASE ASE N
P df f S f F XP df f S f F K ) ( ) ( ) ( ) (
2
1
2
2
(8.4)
where P
ASE
is the noise power in NEB
1
. The OSNR indicator AK is given by
160
P N
K K K = A (8.5)
The precise relationship between AK and the OSNR is not easy to express in
terms oI analytical expressions as it necessitates a mathematical Iormulation oI S(f),
which requires the knowledge oI the analytical expression Ior the single pulse
spectrum multiplied by the spectrum oI a random data sequence, which usually yields
very unwieldy Iormulae |16|. ThereIore, in practice the relationship between AK and
the OSNR would be obtained experimentally, in the Iorm oI a calibration curve. It can
be seen that AK(OSNR) is a monotonic Iunction and also it can be easily shown that
AK(P
ASE
) is approximately linear iI the condition (8.6) is IulIilled
ASE
P df f S f F >>
}


) ( ) (
2
1
(8.6)
This condition is satisIied in most practical situations, Ior suitably chosen Iilters. In
this case, (8.5) becomes:
}


= A df f S f F P K X K
ASE P
) ( ) ( ) (
2
1
(8.7)
To assess the potential oI the method, numerical simulations were perIormed using
JPI TransmissionMaker
TM
. Fig. 8.2 shows the simulated setup to obtain the
calibration curve. The transmitter consists oI a CW laser Iollowed by the Mach-
Zehnder modulator. The rise time oI the electrical signal (driving the modulator) was
t
r
0.25(bit rate)
-1
. It should be noted that presently Mach-Zehnder modulators are
dominant in RZ transmission.

Fig. 8.2 Simulation conIiguration to obtain the calibration curve, BPF bandpass Iilter. A 3 dB coupler
is used to simpliIy the Iinal expression Ior the indicator
All oI the results presented are Ior a 10 Gb/s 50 RZ signal, with a carrier
Irequency oI 193.1 THz. The bandpass Iilters are chosen such that (8.6) is IulIilled
and (8.7) is maximized. Several possible Iilters were investigated to Iind the optimal
161
conIiguration. For comparison, calibration curves Ior two diIIerent Iilters sets are
shown. The Iollowing Gaussian Iilters were used in the simulation:
- F
1
: f
C1
193.11 GHz, Af5 GHz,
- F
2
: f
C2
193.13 GHz, Af10 GHz,
- F
3
: f
C3
193.125 GHz, Af20 GHz.
Positioning one oI the Iilters at the bitrate Irequency (the Iundamental clock
component) away Irom the carrier helps maximizing (8). The Iilters are oIIset Irom
the carrier Irequency to reduce sensitivity to non-perIect extinction ratio, which
maniIests itselI as power increase in the carrier. The calibration curves were obtained
Ior a constant signal power and the ASE power was varied to obtain a given OSNR.
No other impairments were present at this point. Fig. 8.3 presents the calibration
curves AK(OSNR) on a log-log scale, Ior Iilter pairs (F
1
, F
2
) and (F
2
, F
3
). The ASE
power Ior OSNR calculation is measured in a 0.1 nm bandwidth.

Fig. 8.3 Example calibration curves Ior the presented method. Filters f
C1
193.11 THz, Af
1
5 GHz,
f
C2
193.13 THz, Af
2
10 GHz, f
C3
193.125 THz, Af
3
20 GHz.
It should be noted that the calibration curves in the log-log scale are quite
linear and the slope is close to 1. This conIirms the validity oI the simpliIied analysis
presented above. It can be seen that the curve shapes depend on the Iilter choice and,
Ior example, some choices may yield smaller a dynamic range. The Iilters were
optimized using a simple trial and error procedure. For Iurther assessment the Iilter
pair (F
1
, F
2
) is chosen as it has been veriIied that measurement errors due to other
signal impairments are less Ior this conIiguration. For this pair oI Iilters, the dynamic
range is large, and is more than 25 dB Ior the OSNR range oI 30 dB.
162
As the calibration curve is obtained Ior the idealized, noise-only case, one
could expect that the presence oI other signal impairments could cause errors in the
OSNR measurements. In particular, nonlinear eIIects in the Iibre can change the
signal power spectrum and thereIore introduce errors in the method. The inIluence oI
the most signiIicant spectrum-changing phenomena is investigated in next section.
Since the method is modulation speciIic, diIIerent calibration curves would be
necessary Ior diIIerent modulation Iormats. Since the OSNR indicator AK is
maximized when one oI the Iilters captures no signal power, and the other captures as
much signal power as possible, the method can be expected to perIorm well also with
narrow-spectrum Iormats or Iormats with suppressed spectrum components.
8.3. Robustness analysis against other signal impairments
Since the method is based on the modulation spectrum change due to the
presence oI ASE, it is reasonable to expect that other spectrum-changing phenomena
could inIluence the OSNR indicator. BeIore detailed analysis is presented, it is worth
pointing out that the method proposed is insensitive to most linear eIIects, including
attenuation, chromatic dispersion, polarisation mode dispersion (PMD) and other
polarisation related issues (although the applicability oI OSNR is limited in the
presence oI large polarization dependent losses), as they do not change the optical
spectrum. This is another advantage oI the technique, as other methods are oIten
sensitive to PMD |7| or to chromatic dispersion |6|.
Amongst the spectrum-changing phenomena, the most important role is
generally played by Iibre nonlinearity. It is well known that nonlinear eIIects,
particularly selI-phase modulation (SPM), can signiIicantly change the optical
spectrum in the Irequency range oI interest |1|. ThereIore, this issue has been
thoroughly investigated by simulation. There is also a host oI other phenomena which
can change the spectrum. The most important oI these are considered here, including
the eIIect oI reduced extinction ratio, modulator chirp and diIIerent electrical drive
rise time. Finally, the cases when other pulse shapes (or diIIerent transmitters) are
operated in the network and when the carrier Irequency is not perIectly aligned with
the ITU-T grid are also considered (although the latter does not change the shape oI
the spectrum). To estimate the robustness oI the method against each oI these
impairments, the impairment oI interest was introduced into the setup in Fig. 8.2 and
then the OSNR indicator AK(OSNR) was measured. In this case, AK carries not only
163
inIormation due to OSNR but also a contribution due to the other eIIect, which can
possibly lead to erroneous measurement. To quantiIy the inIluence oI the additional
signal impairment, this impaired AK is compared with the idealized case, which is
obtained according to (5) and shown as the calibration curve in Fig. 8.3. This
procedure gives OSNR measurement errors due to each impairment, Ior the purposes
oI presentation deIined as OSNR(AK) - actual OSNR (in dB).
8.3.1. Impact of SPM in fibre
Although in theory some oI the impairments mentioned above could be
avoided in an optimized optical link, the Iibre nonlinearity will always be present and
this eIIect can be only reduced when the signal power is low. ThereIore the inIluence
oI Iibre nonlinearity will be discussed Iirst. The simulated system consists oI Iive, 100
km long, SMF Iibre spans with ampliIiers aIter each span to compensate Ior
propagation losses. Typical SMF parameters were used: dispersion parameter D 17
ps/kmnm, attenuation parameter A 0.2 dB/km, nonlinear index n
2
2.610
-20

m
2
/W. To observe the dependence oI the errors on the launched power, two cases
were simulated. These were with average input signal power oI 1 mW and 5 mW,
which corresponds to, respectively, 4 mW and 20 mW peak power, Ior 50 RZ
signal. The results are given in Figs. 8.4 and 8.5, respectively.

Fig. 8.4 OSNR measurement error against propagation length Ior true OSNR values oI 10, 15, 20 and
25 dB, average input power 1 mW.
164

Fig. 8.5 OSNR measurement error against propagation length Ior true OSNR values oI 10, 15, 20 and
25 dB, average input power 5 mW.
As seen in Figs. 8.4, 8.5, the Iibre nonlinearity does indeed change the OSNR
indicator, as expected. However, the errors are within acceptable limits, in most
practical cases. Although an error oI 1.3 dB (Fig. 8.5) is quite signiIicant, this
accuracy should still be acceptable. For 1 mW input power (Fig.8.4), the errors are
negligible, being less then 0.15 dB. It is interesting to note that, as Iound Irom
simulations, the errors do not accumulate along the Iibre. Instead they Iollow a
pseudo-oscillatory pattern and appear to be bounded. The reason Ior this behaviour is
that the main eIIect oI SPM is to redistribute power between diIIerent Irequency
components and this eIIect is similar Ior the two observed spectral intervals so the net
inIluence is reduced, when calculating the ratio K
N
Irom (8.4). The higher simulated
launch power is quite large, and even higher powers are seldom used in linear
transmission due to possible nonlinear penalty and due to limitations oI the EDFA
power budget when shared between large number oI channels. For the higher launch
power case considered here, the nonlinear eye-opening penalty would be about 1 dB
|17|. Although the bounded evolution oI our OSNR metric along the link is worth
noting, it may be expected that large nonlinearity would Iinally lead to signiIicant
monitoring errors. ThereIore, the method may be oI limited use Ior very long highly
nonlinear links (where oIten conventional out-oI-band monitoring technique can be
used, when only several channels are transmitted), but is well suited Ior metro
networks (especially considering its possible cost-eIIectiveness). The errors are
smaller Ior low OSNR values and this region is usually the most critical. It should be
165
noted that the dependence on OSNR is due to the Iact that a large ASE power can
Iurther 'damp possible changes in K
N
.
8.3.2. Robustness against degraded extinction ratio
The modulation spectrum depends to some extent on the extinction ratio r oI
the modulator. As mentioned beIore, the main eIIect oI a reduced extinction ratio is to
increase DC power in the carrier. Although external modulators usually provide a
good extinction ratio, it is still oI interest to check how a non-perIect extinction
changes the measurement. ThereIore, cases were investigated where the extinction
ratio oI the Mach-Zehnder (MZ) modulator varied by 10 dB. It was Iound that in this
case errors can be slightly larger than 1 dB Ior OSNR25 dB (Fig. 8.6). The error is
mainly due to the Iact that the increase oI the optical power in the carrier is detected
even iI the Iilters are oIIset Irom the carrier Irequency due to the limited steepness oI
the Iilters. This suggests a possible way to minimize the errors by using Ilat-top Iilters
with sharp slopes (please note that relatively gentle Gaussian Iilters are used in
simulations). However, it should be emphasized that r5 dB (the lower limit in
Fig.8.6) is not a very realistic value Ior a Mach-Zehnder modulator. The errors Ior r in
the range Irom 10 dB to 20 dB are no larger than 0.5 dB.
Robustness against reduced extinction ratio
-1.5
-1
-0.5
0
5 10 15 20
extinction ratio [dB]
O
S
N
R

e
r
r
o
r

[
d
B
]
OSNR=25 dB

Fig. 8.6 OSNR measurement error against reduced extinction ratio r Ior true OSNR25 dB. The
reIerence value was obtained Ior r15 dB.
8.3.3. Robustness against modulator chirp
Another eIIect which can change the spectrum is Irequency chirping in
modulators. A MZ modulator can be made chirpless, but is oIten operated with some
166
chirp to improve the signal dispersion properties or to lower the drive voltage.
ThereIore, the case when modulator is chirped was considered and the chirp
parameter , (deIined as the ratio oI phase changes to intensity changes oI the
modulator`s output signal, normalized to intensity) was varied Irom -1 to 1, which
covers the range typical Ior MZ modulators |18|. The inIluence oI the modulator chirp
is shown in Fig. 8.7. As the errors are less than 0.2 dB, they are negligible.
Robustness against moduIator chirp
-0.2
-0.1
0
0.1
-1 -0.5 0 0.5 1
chirp parameter
O
S
N
R

e
r
r
o
r

[
d
B
]
10 dB
15 dB
20 dB
25 dB
OSNR

Fig. 8.7 OSNR measurement error against the modulator chirp. o varied Irom -1 to 1, Ior true OSNR
values oI 10, 15, 20 and 25 dB
8.3.4. Robustness against electrical drive characteristic
The reIerence calibration curve was obtained Ior an electrical rise time oI
t
r
0.25(bit rate)
-1
. As the bandwidth oI the driving electronics can slightly vary in
real applications, this would be also reIlected in the driving signal rise time.
Fluctuations in the electrical signal driving the MZ modulator will translate into
changes in the optical spectrum. To quantiIy this dependence, the scenarios were
simulated when t
r
changes Irom 0.5 to 1.5 times its nominal value. The errors are
presented in Fig. 8.8. In this case the errors are also below 0.2 dB Ior OSNR20 dB,
thus they are negligible.
167
Robustness against non-perfect eIectricaI drive rise time
-0.4
-0.3
-0.2
-0.1
0
0.1
2/16 3/16 4/16 5/16 6/16
rise time (normaIized to bit period)
O
S
N
R

e
r
r
o
r

[
d
B
]
10 dB
15 dB
20 dB
25 dB
OSNR

Fig. 8.8 OSNR measurement error against the electrical signal rise time, Ior true OSNR values oI 10,
15, 20 and 25 dB.
8.3.5. Sensitivity to different pulse shapes
Although the method is, strictly speaking, designed Ior use with a speciIic
pulse shape, one could envisage that in a transparent network diIIerent pulse shapes
may be used Ior RZ transmission. At the moment MZ modulators are dominant. There
is some interest, however, in using electro-absorption modulators (EAM) Ior RZ
transmission |19| and mode-locked lasers are also common in many labs. The main
diIIiculty in comparing the perIormance oI the method against other pulse shapes
stems Irom the Iact that, Ior typical driving conditions, pulses Irom MZ modulators
have a strictly deIined width, whereas Ior example there is no such limit on the pulses
Irom mode-locked lasers. ThereIore, it is questionable what rule should be adopted to
Iacilitate comparison with the reIerence pulses. The rule adopted here is based on the
signal power. It is required that the pulses have the same average and peak power as
the reIerence pulses. Three cases were considered, when the pulses are Gaussian,
Irom a chirpless EAM and a chirped EAM (,0.7). The errors are presented in Table.
8.1.
Table 8.1 OSNR measurement error Ior diIIerent pulse shapes, Ior true OSNR oI 10, 15 and 25 dB
OSNR |dB| Gaussian EAM, , 0 EAM, , 0.7
10 1.99 0.07 -0.06
15 2.67 0.24 -0.44
25 -2.41 -2.99 -2.34

168
The errors calculated were always within 3 dB. This is however quite an
abstract scenario as the use oI MZ modulators is predominant in practical networks
when RZ transmission is employed. It should be noted that direct modulation is rarely
used Ior RZ transmission, due to the large chirp.
8.3.6. Sensitivity to channel frequency misalignment
Finally, there is another eIIect which should be mentioned, which does not
change the shape oI the spectrum, but can still introduce errors to the method. This is
carrier Irequency oIIset Irom the ITU-T grid. Although the carrier Irequency should
be strictly controlled in DWDM systems, in practice some Irequency driIt can occur.
This will inevitably inIluence the readout oI the method iI the Iilters are Iixed to the
nominal channel Irequency. However, this is an issue common to most OPM schemes
(including OSNR measurement) utilizing narrowband optical Iiltering. Fig. 8.9
presents the eIIect oI the carrier misalignment on the OSNR indicator.
It is clear that a carrier Irequency oIIset may have a negative eIIect on the
accuracy oI the proposed method. The plot oI error vs. oIIset is not symmetrical since
the Iilters were oIIset Irom carrier wavelength to reduce the sensitivity to the
extinction ratio. In practice, one may sacriIice the robustness against extinction ratio
in Iavour oI better resilience to carrier misalignment iI such an impairment is more
likely to occur in the network. From Fig. 8.9 it can be noted that Ior this particular set
oI Iilters, only about 2 GHz carrier driIt can be tolerated to maintain accuracy,
although tolerance to positive carrier shiIt is much larger. Most DFB lasers have
Irequency temperature coeIIicient oI about 12.5 GHz/C |20| which, with the
temperature control within 0.1C, translates into the Irequency driIt oI about 1.2 GHz,
which is within the tolerance oI the method. Moreover, wavelength locked
transmitters are becoming more popular and can provide Irequency stability better
than 0.5 GHz |21|. However, in some scenarios, the channel wavelength is not
controlled well enough. In these cases the method would have to be combined with a
wavelength locking oI the Iilters. This requires Iilter tunability but has the advantage
that a single Iilter can be used, instead oI two. A tunable Iilter is naturally available iI
the method is implemented as an upgrade in systems already utilizing narrowband
OSAs Ior monitoring. Alternatively, the unsymmetrical characteristics oI the curves in
Fig. 8.9 suggest that it might be useIul to intentionally calibrate the method with the
positive oIIset oI 1-2 GHz to reduce the sensitivity to the negative oIIset.
169
InfIuence of carrier frequency offset
-1
0
1
2
3
4
5
6
7
8
-4 -3 -2 -1 0 1 2 3 4
offset [GHz]
O
S
N
R

e
r
r
o
r

[
d
B
]
10 dB
15 dB
20 dB
25 dB
OSNR

Fig. 8.9 OSNR measurement error against carrier Irequency oIIset, Ior true OSNR values oI 10, 15, 20
and 25 dB
8.4. Summary and conclusions
In this chapter, a novel OSNR monitoring method has been presented. The
method is based on the assessment oI the modulation spectrum. The main advantages
oI the method are its simplicity and cost-eIIectiveness. The robustness oI the method
has been assessed through numerical simulations. Since the method is based on the
assessment oI the optical spectrum, it is insensitive to most linear optical phenomena
such as chromatic dispersion or PMD. However, the principle oI the method also
implies that it is sensitive to eIIects changing the optical spectrum. The sensitivity oI
the method to the most signiIicant such eIIects has been investigated. The robustness
analysis against modulator chirp, Iibre nonlinearity, electrical signal rise time,
extinction ratio degradation and carrier Irequency oIIset has been presented. In most
cases, it is Iound that the method is robust and eIIective, although the sensitivity to
Iibre nonlinearity suggests that it may be better suited Ior use in quasi-linear links. As
with most other techniques based on narrowband Iiltering, the sensitivity to channel
misalignment cannot be avoided and in some cases might require Iilter tunability,
which can, however, reduce the number oI required optical Iilters. Although the
analysis was presented here Ior a 50 RZ signal, the method is applicable also (aIter
suitable recalibration) to other modulation Iormats.
170
8.5. References
|1| G. P. Agrawal, Lightwave Technologv. Telecommunication Svstems.
Hoboken: Wiley, 2005.
|2| D. C. Kilper and W. Weingartner, "Monitoring optical network perIormance
degradation due to ampliIier noise," J. Lightw. Technol., vol. 21, pp. 1171-
1178, 2003.
|3| M. P. Dlubek, A. J. Phillips, and E. C. Larkins, "Simple and Cost-EIIective
OSNR Monitoring Method based on Modulation Spectrum Assessment,"
Europ. Conf. Netw. Opt. Commun. NOC 2007, vol. 12, pp. 168 - 175, 2007.
|4| M. P. Dlubek, A. J. Phillips, and E. C. Larkins, "Method Ior OSNR
Monitoring Based on Modulation Spectrum Assessment," accepted for
publication in IET Optoelectronics.
|5| W. Chen, R. S. Tucker, X. Yi, W. Shieh, and J. S. Evans, "Uncorrelated beat
noise measurement Ior optical signal-to-noise ratio monitoring". Europ. Conf.
Opt. Commun. ECOC 2005, vol. 4, pp. 971-972, 2005.
|6| C. Dorrer and X. Liu, "Noise monitoring oI optical signals using RF spectrum
analysis and its application to phase-shiIt-keyed signals," IEEE Photon.
Technol. Lett., vol. 16, pp. 1781-1783, 2004.
|7| J. H. Lee, D. K. Jung, C. H. Kim, and Y. C. Chung, "OSNR monitoring
technique using polarization-nulling method," IEEE Photon. Technol. Lett.,
vol. 13, pp. 88-90, 2001.
|8| C. Xie, D. C. Kilper, L. Moller, and R. RyI, "Orthogonal-Polarization
Heterodyne OSNR Monitoring Insensitive to Polarization-Mode Dispersion
and Nonlinear Polarization Scattering," J. Lightw. Technol., vol. 25, pp. 177-
183, 2007.
|9| Z. C. Zhenning Tao, Libin Fu, Deming Wu, Anshi Xu,, "Monitoring oI OSNR
by using a Mach-Zehnder interIerometer," Microwave Opt. Technol. Lett., vol.
30, pp. 63-65, 2001.
|10| J. C. Li, K. Hinton, S. D. Dods, and P. M. Farrell, "Enabling ASON Routing
via Novel Signal Quality Metrics," Opt. Fiber Commun. Conf. OFC 2007, pp.
1-3, 2007.
171
|11| G. Maier, C. Busca, and A. Pattavina, "Multi-domain routing techniques with
topology aggregation in ASON networks," Int. Conf. Opt. Netw. Design
Model. ONDM 2008, pp. 1-6, 2008.
|12| H. Suzuki and N. Takachio, "Optical signal quality monitor built into WDM
linear repeaters using semiconductor arrayed waveguide grating Iilter
monolithically integrated with eight photodiodes," Electron. Lett., vol. 35, pp.
836-837, 1999.
|13| S. Curtis, "ROADMs drive an agile networking revolution," fibersvstems
Europe, vol. 12, pp. 17-19, 2008.
|14| F. Audet, "Commissioning ROADMs," EXFO Application Note,
http://documents.exIo.com/appnotes/anote169-ang.pdI, accessed June 2008.
|15| W. Moench and J. Larikova, "In-Service Measurement oI the OSNRin
ROADM-based Networks," JDSU White Paper,
http://www.jdsu.com/product-literature/ecocwpIoptmae.pdI, accessed
June 2008.
|16| E. Ip and J. M. Kahn, "Power spectra oI return-to-zero optical signals," J.
Lightw. Technol., vol. 24, pp. 1610-1618, 2006.
|17| J. Kissing, S. Vorbeck, M. Plura, and E. Voges, "Design oI high data rate
optical transmission systems with selI phase modulation and noise," Electric.
Eng. (Archiv fur Elektrotechnik), vol. 84, pp. 123-128, 2002.
|18| H. Kim and A. H. Gnauck, "Chirp characteristics oI dual-drive. Mach-Zehnder
modulator with a Iinite DC extinction ratio," IEEE Photon. Technol. Lett., vol.
14, pp. 298-300, 2002.
|19| W-J. Choi, A. E. Bond, J. Kim, J. Zhang; R. Jambunathan, H. Foulk, S.
O'Brien, J. Van Norman, D. VandegriIt, C. Wanamaker, J. Shakespeare, H.
Cao, "Low insertion loss and low dispersion penalty InGaAsP quantum-well
high-speed EAM modulator Ior 40-Gb/s very-short-reach, long-reach, and
long-haul applications," J. Lightw. Technol., vol. 20, pp. 2052-2056, 2002.
|20| R. Ramaswami and K. Sivarajan, Optical Networks, 2 ed: Morgan KauImann,
2002.
|21| J. Kim, B. S. Choi, H. Yun, S. H. Oh, J.-H. Lee, H. Ko, K.-S. Choi, S. Park, J.
T. Moon, and M.-H. Park, "Thermally controlled wavelength locker integrated
in widely tunable SGDBR-LD module," IEEE Photon. Technol. Lett., vol. 16,
pp. 2430-2432, 2004.
172
9. Nonlinear evolution of Gaussian ASE noise in zero-
memory nonlinear fibre
In Chapter 4, it was explained that the Q-Iactor method is based on the
assumption that the electrical noise pdI at the input to the threshold device is
Gaussian. Therein and in Chapter 7 it was also discussed that this assumption is
Iundamentally incorrect in direct detection systems, where the squaring operation
perIormed by optical receiver produces beat noise terms, having resultant pdIs
diIIerent Irom Gaussian. In this case, the more realistic electrical pdIs are given by the
(non-central) chi-square distribution, derived under the assumption that optical ASE
noise is a complex Gaussian random variable |1, 2|. This assumption about the nature
oI the ASE noise produced by EDFA ampliIiers is widely used, as it results in a
simple description and agrees very well in all practical cases with the more rigorous
quantum-mechanical description |3|. The validity oI this assumption has been also
conIirmed experimentally by means oI the linear (coherent) sampling |4|. In Chapter
7, the reasons Ior the sensible accuracy oI the Q-Iactor in direct detection systems,
despite it being based on essentially wrong assumptions, have been discussed.
However, the Q-Iactor method, due to its simplicity, is oIten used to evaluate
the quality oI the signal using other modulation Iormats and detection schemes as well
|5, 6|. It may be expected that the perIormance oI the Q-Iactor method depends on the
actual pdI oI the electrical noise in given system conIiguration. For example, it was
shown that in diIIerential phase shiIt keyed (DPSK) systems, the accuracy oI the Q-
Iactor is much worse than in direct detection systems due to the output noise shape
|7|. On the other hand, in coherently detected amplitude shiIt keyed (ASK) systems,
the output electrical noise retains the Gaussian nature oI the ASE noise (although
coherent ASK systems are not very common, they recently receive some renewed
attention as ways to increase spectral eIIiciency are sought |8|). In this case, it could
be intuitive to use the Q-Iactor unquestioningly, as it may be expected to be strictly
valid. ThereIore, the question oI validity oI the Q-Iactor in such ASK systems has
received much less interest than Ior example in DPSK systems (although, admittedly,
the importance oI the DPSK systems may have been also a Iactor).
In this chapter we show that, in the presence oI signiIicant Iibre nonlinearity,
the Q-Iactor loses its accuracy in coherent ASK systems. Moreover, it is discussed
173
that the presence oI this nonlinearity obliterates the well-known sensitivity advantage
oI the coherent detection over the direct detection. The analysis presented in this
chapter is based on approximating the nonlinear Iibre as a zero-memory nonlinear
(ZMNL) device |9|. This approximation allows obtaining analytical expressions Ior
statistical moments (particularly Ior kurtosis) oI the ASE noise propagating in a
ZMNL Iibre. The expressions Ior the moments and, in particular, the evolution oI
kurtosis, investigated later in this chapter, clearly show that the noise loses its
Gaussian character and they may be used as a metric Ior the deviation Irom Gaussian
nature. This aspect oI our research has resulted in publication |10|. As an extension oI
the ZMNL Iibre model, Iormulae describing the nonlinear evolution oI the ASE pdIs
have been also obtained. These expressions can be directly used Ior quantitative
analysis and BER calculations. The shapes oI the evolving pdIs are very similar to
those reported in the literature based on numerical simulations solving the Iull
nonlinear Schroedinger equation describing Iield propagation in optical Iibre |11, 12|.
This part oI work presented here resulted in paper |13|.
9.1. Propagation of light in optical fibre
Propagation oI light (and all electromagnetic phenomena) in the most general
case is described by the celebrated Maxwell equations |14|. However, because oI their
generality, Maxwell`s equations are also diIIicult to solve and even a numerical
solution is computationally expensive and not straightIorward. ThereIore in practice,
several physical and mathematical approximations or transIormations are applied, to
obtain simpler models better suited Ior the description oI particular system. These
models should obviously take into accounts the physical properties oI the medium, in
which the light propagates. In the case oI optical Iibre, the generally accepted model
is based on the wave propagation equation known as the nonlinear Schroedinger
equation (NLSE) |14, 15|. The derivation oI NLSE is Iairly lengthy |14| and the
rigorous derivation is even longer |15|. Since this equation is very well known to the
optical communications community, its derivation is not presented here and the
interested reader is directed to one oI the numerous derivations Iound in the literature
(see, Ior example, |14, 15|). Instead, the Iinal, most commonly encountered, Iorm oI
this equation is given |14|
U U
T
U
U i
:
U
i
2
2
2
2
2 2

| o

c
c
+ =
c
c
(9.1)
174
Equation (9.1) describes propagation oI a linearly polarized electric Iield U(:,T) in
optical Iibres. Here, : is the longitudinal dimension and the time T is measured in the
retarded Irame, moving with group velocity. The terms involving o and
2
are linear
operators, describing Iibre attenuation and group velocity dispersion, respectively.
The term involving y is responsible Ior the nonlinearity. Typically, the main source oI
nonlinearity in optical Iibre is the Kerr eIIect, leading to the dependence oI the
reIractive index on optical power. Equation (9.1) is typically adequate to describe the
propagation oI pulses practically encountered in optical communications, in some
special cases, however, the addition oI higher order terms, especially those
responsible Ior third order dispersion and Raman scattering, may be necessary |14|.
Also, the eIIect oI PMD is not included in (9.1). Nevertheless, this equation is
commonly used in the literature to describe the pulse evolution, as it captures the most
relevant physical phenomena |16-20|.
Even iI (9.1) is derived under various assumptions and approximations, it is
still a complicated second order nonlinear partial diIIerential equation. Hence, its
analysis is diIIicult, as it cannot be solved Ior the general case and can only be
analytically solved Ior some speciIic cases with the help oI the inverse scattering
method. In general, solving (9.1) requires numerical methods. ThereIore, to simpliIy
matters and to obtain some physical insight into various phenomena which it
describes, it is Irequent to introduce Iurther simpliIications into (9.1) |14|. In
particular, this approach is common when analytically investigating the nonlinear
interaction between the signal and ASE noise in optical Iibre |19-22|, as otherwise the
stochastic partial diIIerential equation obtained aIter U(:,T) is taken as random needs
to be evaluated using Monte-Carlo methods, i.e. it needs to be multiply numerically
solved, which is very computationally expensive. BeIore the applied simpliIication is
discussed, it is worth reviewing the main purpose oI reIs. |19-22|. Since the paper oI
Gordon and Mollenauer |18|, it is known that the nonlinear coupling between the
signal and ASE noise produces nonlinear phase noise (NLPN), which is also known
as the Gordon-Mollenauer eIIect. This nonlinear phase noise may be detrimental in
phase shiIt keyed transmission, thereIore |19-22| sought to derive the pdI oI NLPN
and to calculate the BER in optical DPSK systems with NLPN.
The simpliIication oI (9.1) used in |19-22| consists in neglecting the
dispersion term. This is motivated by the reasoning that the main physical
phenomenon (coupling between the signal and the noise) is still included in the
175
resultant equation and that dispersion would mainly aIIect the magnitude oI this eIIect
by reducing the eIIective nonlinearity. Additionally, in |19-22| it was assumed that
the link is very long and that the ampliIiers are spaced so densely compared to the
characteristic length scales oI the system that the noise can be considered as
distributed along the link. These assumptions allow transIorming the NLSE into the
Ito stochastic diIIerential equation, so that the system can be modeled as one-
dimensional random walk (Wiener process) |20, 22|. Using these assumptions, Ho
|22| and Mecozzi |23| were able to derive the pdI oI NLPN, which has obvious
applications to calculate BER in phase shiIt keyed systems.
In this chapter, the inIluence oI the Gordon-Mollenauer eIIect on the evolution
oI quadratures oI the ASE noise is investigated. This evolution is important Ior
understanding the perIormance oI coherently detected ASK signals in nonlinear Iibre.
Instead oI transIorming (9.1) into Ito stochastic equation, an approximation oI the
optical Iibre as a ZMNL system is used. This allows deriving simple analytical
expressions Ior the nonlinearly transIormed moments oI the ASE quadratures and,
Iurthermore, Ior the evolving joint pdI oI the two quadratures.
9.1.1. ZMNL model of nonlinear fibre
To derive the ZMNL model, it is assumed that U(:, T) is an electrical Iield at
the output oI a linear optical ampliIier, i.e. it contains an incoherent component (ASE)
and may contain the coherent signal as well. Only the input ampliIier is assumed in
the link (contrary to the assumption about the distributed noise used, Ior example, in
|20|). Since nonlinear propagation is oI interest here, the dispersion term, responsible
Ior pulse spreading, is neglected, similarly to the approach in the works brieIly
reviewed above. This reduces (9.1) to a stationary equation, which can describe
propagation oI a CW wave (although it neglects some important physical eIIects, like
modulation instability). II the noise is propagated along with signal pulses, this
simpliIication physically means that the worst case scenario is considered, when the
signal power contributes most strongly to the nonlinearity. Although strictly speaking
no-dispersion Iibre does not exist, this simpliIication is approximately valid Ior Iibres
with very low dispersion (e.g. dispersion shiIted Iibres in the 1550 nm region) or
situations when nonlinear length is much shorter than dispersion length (e.g. highly
nonlinear Iibres) thereIore this simpliIication corresponds to the highly nonlinear
regime. In this regime, the propagation equation thus takes the Iorm
176
.
2
2
U U U i
:
U
i
o
=
c
c
(9.2)
From the system theory point oI view, this approximation transIorms the
nonlinear system with memory into a much simpler ZMNL system |9|. The complex
input Iield U
0
U(0, T) is given by the signal P
S
and zero mean complex Gaussian
noise (ASE), with the two initial quadratures U
0R
, U
0I
(without the loss oI generality,
the input signal is assumed to be real). P
S
is easiest thought oI as a constant power but
it can also be alternatively regarded as a peak signal pulse power, since there is no
time dependence in (9.2) as a consequence oI neglecting dispersion. The input Iield is
thereIore:
). , 0 ( ) , 0 ( )) , 0 ( ( ) , 0 (
0 0
2 1
T iU T U T P T U
I R S
+ + = (9.3)
Using the standard normalized amplitude A(:, T), where
), , ( )
2
exp( ) , (
0
T : A
:
P T : U
o
= (9.4)
where P
0
is an arbitrary characteristic power, it is possible to obtain the solution to
(9.2) in an easy way. The power P
0
is usually taken as a peak power in the case oI
well deIined pulses, when spectral evolution due to selI-phase modulation is analyzed
|14|. However, when the ASE Iield is considered, the natural and more convenient
quantity is the noise power in a quadrature (quadrature noise variance), as we are
particularly interested in the statistical evolution oI noise. In this case, the normalized
process A(0, T) corresponding to U(0, T) is
I R
iA A M T A
0 0
) , 0 ( + + = (9.5)
where M
2
P
s
/P
0
is ratio oI the input signal power to the noise power. Since ASE is a
zero-mean process and P
0
is equal to the variance oI each noise quadrature, A(0, T) is
normalized complex Gaussian with the noise quadratures A
0R
ReA(0,T)}-M,
A
0I
ImA(0, T). The time argument is dropped oII to simpliIy the notation. The joint
pdI oI these two Gaussian variables is given by |24|:
,
2
) (
exp
2
1
) , (
2
0
2
0
0 0
0
|
|
.
|

\
|
+
=
I R
I R A
A M A
A A p
t
(9.6)
AIter substituting (9.4) into (9.2), the Iinal simpliIied Iorm is obtained
. ) exp(
2
0
A A P : i
:
A
o =
c
c
(9.7)
177
This equation has been used in the past to analyze the spectral evolution oI a
signal in nonlinear Iibre |14| and can be easily solved to obtain
)) , ( exp( ) , ( T : i i T : A
NL L
| | + = (9.8)
where
L
arctan (A
0I
/A
0R
) is the initial linear phase,
2 1 2
0
2
0
) (
I R
A A + = is the initial
magnitude and
NL
(:, T) is the nonlinear phase accumulated during the propagation.
The nonlinear phase
NL
is given by
) ( ) (
2
0
2
0
: D A A
I R NL
+ = | (9.9)
and coeIIicient D(:) describing the total accumulated nonlinearity is purely
deterministic and given by
) 1 ( ) (
0
:
e P : D
o
o

= (9.10)
Eq. (9.8) thereIore gives the transIormation Iunction oI the ZMNL Iibre and shows
the main advantage oI the ZMNL approach, as the moments and pdI oI A(:, T) can be
now calculated using standard methods Irom probability theory |9, 24|. This equation
will be later used to derive Iormulae describing the evolution oI the moments and pdIs
oI the quadratures.
9.2. Propagation of purely incoherent field in ZMNL fibre
BeIore the more interesting case oI signal and ASE noise interplay is
considered, let us Iirst brieIly consider the case when ASE propagates alone, i.e. when
U(0, T) represents ASE noise only. In this case, U(0, T) is drawn Irom a circular
Gaussian process, so that the quadratures are independent, zero-mean Gaussian
variables with equal variance o
2
and A(0, T) is a similar process with normalized
variance. For the initial Gaussian noise A(0, T), the magnitude is well known to have a
Rayleigh distribution and the phase is uniIormly distributed over the interval (-r, r|
|24|. It is easy to notice Irom (9.8), that the magnitude is preserved in the case
considered here, whereas the phase depends on D(:). This suggests that in direct
detection schemes, when electrical signal is proportional to the squared magnitude,
the statistics oI the received signal in this case do not change, even though the
quadratures are no longer Gaussian (as shown later in this chapter). The Kerr
nonlinearity still indirectly inIluences BER in such systems, mainly by changing the
noise spectrum |25|. This eIIect is beyond the scope oI the current model. As the
178
nonlinear phase is given by the sum oI two squared Gaussian variables (9.9), it has a
(central) chi-squared distribution |1, 2, 24|. For two degrees oI Ireedom, this reduces
to the exponential pdI p
NL
(
NL
)
,
) ( 2
exp
) ( 2
1
) (
|
|
.
|

\
|
=
: D : D
p
NL
NL NL
|
| (9.11)
where
NL
0. It can be proven that
L
and
NL
are independent random variables, just
as and
L
are independent. ThereIore, the pdI oI the total phase in (9.8) is given by
the convolution oI the pdIs oI the two contributions. This gives
NL L
| | | + = (9.12)
) ( ) ( ) ( | | |
NL L
p p p - = (9.13)

>
< s
<
=

t | t
t | t t
t |
|
| t t
t |
) ( 2 ) ( 2 ) ( 2 1
) ( 2 1
) ( ) 2 (
) 1 ( ) 2 (
0
) (
: D : D : D
: D
e e e
e p (9.14)
and the total phase might seem to depend on the accumulated nonlinearity at the Iirst
glance. Closer inspection reveals that, iI the nonlinear phase is normalized again to
the standard interval (-r, r|, the uniIorm distribution is reproduced, due to the Iorm oI
the Iunctions describing the rising (-r, r| and Ialling (~r) slopes oI p(). ThereIore,
the Kerr nonlinearity does not change the statistics oI the magnitude and phase oI the
noise, when the noise propagates alone. Later, it is shown that the quadratures also
remain uncorrelated, hence the Kerr nonlinearity does not break the Gaussian property
oI ASE noise, when no coherent signal is present.
9.3. Propagation of moments of the quadratures in ZMNL fibre
Equation (9.8) gives the transIormation Iunction oI ZMNL in terms oI
magnitude and phase, which are natural quantities to investigate in, respectively,
direct detection and phase shiIt keyed systems. In coherent ASK systems, the more
useIul description would be in terms oI quadratures. Since this aspect oI the nonlinear
ASE evolution seems to have received much less attention, it will be investigated in
this chapter. In what Iollows, explicit expressions Ior the moments oI the quadratures
are given |10|. These permit the estimation oI the deviation Irom Gaussian pdI on the
basis oI higher order moments.
179
9.3.1. Derivation of expressions for moments
From (9.8), the ZMNL transIormation in terms oI quadratures can be
theoretically derived by introducing the transIormation pair
,
) ( )) , ( sin( ) , (
) ( )) , ( cos( ) , (
0
0

= + =
= + =
A g T : T : A
A g T : T : A
I NL L I
R NL L R
| |
| |
(9.15)
which is an inverse transIormation to the one used in the Iirst place in (9.8). The
nonlinearly transIormed moments can be calculated directly Irom the input statistics
and the transIormation Iunction, without the need Ior the output pdI |9, 24|. For the n-
th output moment the general transIormation relation is
| |
| |
,
) , ( ) ( ) , (
) , ( ) ( ) , (
0 0 0 0 0 0
0 0 0 0 0 0

=
=
}}
}}
I R I R A
n
I
n
I
I R I R A
n
R
n
R
dA dA A A p A g T : A E
dA dA A A p A g T : A E
(9.16)
where E|.| denotes statistical expectation, g
R
(A
0
), g
I
(A
0
) are given by the appropriate
transIormation Irom (9.15) and p
A0
(A
0R
, A
0I
) is the joint Gaussian density (9.6). Eqn.
(9.16) is diIIicult to solve in Cartesian coordinates (A
0R
, A
0I
) Ior the Kerr nonlinearity
and it is easier to work with polar coordinates (,, ). The transIormation is given by
,
) sin(
) cos(
0
0

=
=


I
R
A
A
(9.17)
so that the respective expressions Ior the n-th moments oI the in-phase and quadrature
components become
| |
| |
,
) ( sin
2
) ( cos
2
0
2
cos 2
2
1
0
2
cos 2
2
1
2 2
2 2

+ =
+ =
} }
} }
+

+
+

+

t

t
t
t
t
d d e D A E
d d e D A E
M M
n
n
n
I
M M
n
n
n
R
(9.18)
where the dependence oI moments on the propagation length through D(:) is
understood. The integration over phase is given by
, )) cos( exp( ) ( cos
2

t
t
d M D
n
}

+ (9.19)
(or similar expression with sine Iunction) and the power oI the term cos
n
(.) needs to
be reduced in order to proceed. The power reduction Iormulae Ior arbitrary n Iollow
directly Irom the theorem that cos(nx) can be expressed in terms oI single angle x as
180
T
n
cos(x), where T
n

is the Chebyshev polynomial |26|. As n increases, the Iormula
becomes lengthy, but otherwise the calculation is straightIorward. AIter the power oI
the cosine term in (9.19) has been reduced to 1, it may be observed that the integral
(9.19) can be expressed in terms oI the sum oI modiIied Bessel Iunctions oI the Iirst
kind I
k
(Mp). The order k depends on the particular decomposition oI the term cos
n
(.).
In this chapter, we consider propagation oI the Iirst Iour moments. These are later
used to calculate kurtosis (a statistic related to the Iourth moment), as the kurtosis is a
very powerIul indicator oI a non-Gaussian pdI. AIter the integration over phase, the
Iirst 4 (ordinary) moments oI the real and imaginary components are given by
, ) cos( )
2
exp( ) ( | |
0
2
2 2
1
2
}

+
=

d D
M
M I A E
R
(9.20a)
, ) sin( )
2
exp( ) ( | |
0
2
2 2
1
2
}

+
=

d D
M
M I A E
I
(b)
, )
2
exp( )| 2 cos( ) ( ) ( |
2
1
| |
2 2
0
2
2 0
3 2

d
M
D M I M I A E
R
+
+ =
}

(c)
, )
2
exp( )| 2 cos( ) ( ) ( |
2
1
| |
2 2
0
2
2 0
3 2

d
M
D M I M I A E
I
+
=
}

(d)

d
M
D M I D M I A E
R
)
2
exp( )| 3 cos( ) ( ) cos( ) ( 3 |
4
1
| |
2 2
2
3
0
2
1
4 3
+
+ =
}

(e)

d
M
D M I D M I A E
I
)
2
exp( )| 3 sin( ) ( ) sin( ) ( 3 |
4
1
| |
2 2
2
3
0
2
1
4 3
+
=
}

(I)
, )
2
exp( )| 4 cos( ) (
) 2 cos( ) ( 4 ) ( 3 |
8
1
| |
2 2
2
4
0
2
2 0
5 4



d
M
D M I
D M I M I A E
R
+
+
+ =
}

(g)
. )
2
exp( )| 4 cos( ) (
) 2 cos( ) ( 4 ) ( 3 |
8
1
| |
2 2
2
4
0
2
2 0
5 4



d
M
D M I
D M I M I A E
I
+
+
=
}

(h)
The intimate relationship between the moments oI the two components is
immediately visible. It should be noted that the Iormula Ior each moment consists oI
terms oI the Iorm
181
}

0
2 2
) 2 exp( ) cos( ) ( d nD M I
n
m
(9.21)
or alternatively the cosine term in (9.21) is replaced with sine. BeIore solving this
integral, it may be noted that simple substitution transIorms (9.21) into a Fourier
cosine transIorm oI a kernel, which can be treated as a generalized Rician pdI
(although the area under this kernel is not normalized to 1). This integral always
exists as the asymptotics oI the modiIied Bessel Iunction are no Iaster than
exponential Iunction. This means that every moment oI arbitrary order exists aIter the
propagation in nonlinear Iibre. ThereIore, the pdI could (in principle) be reconstructed
Irom the moments |9, 24|. To proceed, it is necessary to express (9.21) as
}

0
2
) exp( ) ( d M I
n n
m
, (9.22)
where inD
n
= 2 / 1 and the real or imaginary part oI the solution is taken, as
appropriate. Integral (9.22) can be Iound in tables |27|. Using 6.631.1 Irom |27|, with
a well-known relation between the (ordinary) Bessel Iunction oI the Iirst kind and the
modiIied Bessel Iunction oI the Iirst kind |26|, the Iormal solution oI (9.22) is
)
4
; 1 ;
2
1
(
! 2
) 2 1 2 1 2 1 (
) exp( ) (
2
1 1
) 1 ( 2 1 1
0
2
n
m n
n
n
n
n n
m
M
n
m n
F
n
m n M
d M I

+
+ +

+ + I
=
+ + +

}
(9.23)
where I(.) is the gamma Iunction and
1
F
1
(a; b; x) is a conIluent hypergeometric
Iunction |26|. In the general case, the Iormal solution given by (9.23) is not much
more useIul than the integral representation (9.21), since both require numerical
evaluation (although the conIluent hypergeometric Iunction is implemented in many
mathematical packages). However, since only integer moments are considered here,
1
F
1
(a; b; x) simpliIies considerably as both parameters m and n are integers, so that it
could be expressed in terms oI elementary, or at least Iamiliar, Iunctions.
Additionally, the sum (mn) is always odd, which results in particularly simple
expressions. In the special case
1
F
1
(a; a; f(x)), simpliIies to exp(f(x)) |26|. For other
identities, the reader should consult |26| or other reIerences on higher special
Iunctions. Using (9.20) and (9.23), the Iour moments can be expressed analytically as:
)
2
exp( )
4
exp(
4
Re | |
2
1
2
2
1
M M M
A E
R

)
`

=

(9.24a)
182
)
2
exp( )
4
exp(
4
Im | |
2
1
2
2
1
M M M
A E
I

)
`

=

(b)

+ + =

2
2
2
0
2
0
2
2
4
3
2
2
2
4 4
0
2
2
0
2
2
16
)
4
(
4
Re | |

M
M
M M
M
R
e
M e
e e
M e
A E (c)

+ =

2
2
2
0
2
0
2
2
4
3
2
2
2
4 4
0
2
2
0
2
2
16
)
4
(
4
Re | |


M
M
M M
M
I
e
M e
e e
M e
A E (d)
2
4
4
3
3
4 4
1
2
3
1
3
2
3
2
1
2
1
2
16
)
8
(
2
3
Re
4
1
| |
M
M M M
R
e e
M
e e
M M
A E

+ =


(e)
2
4
4
3
3
4 4
1
2
3
1
3
2
3
2
1
2
1
2
16
)
8
(
2
3
Im
4
1
| |
M
M M M
I
e e
M
e e
M M
A E

=


(I)
2
4
5
4
4
4 4
2
2
4
2
2
4 4
0
2
4
2
0
4
3
0
4
2
4
2
2
2
2
2
0
2
0
2
0
2
32
) 3
4
(
2

)
2 32
(
3
Re
8
1
| |
M
M M M
M M M
R
e e
M
e e
M M
e e
M
e
M
A E

+ + +
+ +

=




(g)
2
4
5
4
4
4 4
2
2
4
2
2
4 4
0
2
4
2
0
4
3
0
4
2
4
2
2
2
2
2
0
2
0
2
0
2
32
) 3
4
(
2

)
2 32
(
3
Re
8
1
| |
M
M M M
M M M
I
e e
M
e e
M M
e e
M
e
M
A E

+ +
+ +

=




(h)
From the moments (9.24a-h), the central moments m
kR,kI
are calculated Irom the
expressions
| | | |
2 2
2
X E X E m = (9.25)
| | 2 | | | | 3 | |
3 2 3
3
X E X E X E X E m + = (9.26)
| | 3 | | | | 6 | | | | 4 | |
4 2 2 3 4
4
X E X E X E X E X E X E m + = (9.27)
Mixed moments, describing the relationship between the two components A
R,I

can be derived in an analogous way. The Iirst mixed (joint) moment (linear
correlation) is thereIore given by
183
2
4
3
2
2
2
2
2
8
Im
2
1
| |
M
M
I R
e e
M
A A E

(9.28)
and the correlation coeIIicient r is obtained Irom
I R
I R I R
m m
A E A E A A E
r
2 2
| | | | | |
= (9.29)
ThereIore, r0 when D(:) =0 as the inspection oI (9.24a, b) and (9.28) reveals, so that
the two quadratures become anticorrelated. However, when M0, no correlation
between the quadratures is developed, which, together with considerations given in
Section 9.2, shows that in this case Kerr nonlinearity does not break the Gaussian
nature oI ASE. These Iindings will be conIirmed when an extension oI the ZMNL
model presented here, allowing calculation oI the joint pdI oI the two quadratures, is
given in later in this chapter. In the general case, the common assumption oI the
independence oI the ASE quadratures is not valid in nonlinear Iibre. Similar Iormulae
exist Ior higher moments and mixed moments. In the next section, based on the
Iormulae derived here, the propagation oI kurtosis is investigated, to clearly show
deviations Irom Gaussian character oI the ASE.
9.3.2. Propagation of kurtosis
As mentioned above, a convenient measure oI the deviation Irom Gaussian pdI
is the kurtosis k, which is deIined as the ratio oI the Iourth central moment to variance
squared
. 3
2
2 4
= m m k (9.30)
The shiIt by 3 normalizes the kurtosis to 0 Ior Gaussian variables with higher
values indicating a distribution that is more sharply peaked (leptokurtic), and lower
values indicating that the distribution is more Ilat-topped (platykurtic) |9|. ThereIore,
to investigate the possible deviations Irom Gaussian character, the kurtosis is
proposed here. Eqns. (9.24-9.27, 9.30) allow computation oI the kurtoses oI the two
quadratures. BeIore the general case is discussed, it is worth emphasizing again that
when M0 (noise only case), none oI the moments (including the linear correlation)
considered here are changed during the propagation. For combined noise and signal
propagation, several plots are shown below.
184
-2
-1
0
1
2
3
4
5
6
7
8
0 10 20 30 40 50 60 70 80 90 100
Iength [km]
kurtosis
Po=0.8 mW,X
Po=0.8mW,Y
Po=1.0mW,X
Po=1.0mW,Y
Po=1.5mW,X
Po=1.5mW,Y

Fig. 9.1 Analytical evolution oI kurtosis against propagation length, M
2
10
Fig. 9.1 shows the evolution oI the kurtosis oI the in-phase component (X) and
the quadrature component (Y) along the Iibre, Ior diIIerent noise powers. M
2
(P
S
/P
0
)
was constant and equal to 10. Fig. 9.2 presents the dependence oI the kurtosis on M
2
.
The noise power was kept constant at P
0
1.0 mW and signal power was varied. The
propagation length was : 50km. Figs. 9.1 and 9.2 suggest that nonlinear noise
evolution exhibits threshold characteristics and the nonlinear behavior oI the pdI
appears at some input power. In Fig. 9.2, the threshold power is about 8 mW. Below
this value, the noise approximately retains its Gaussian character. As apparent Irom
Fig.9.1, the deviation Irom the Gaussian statistics can be very signiIicant even Ior
moderate input powers in the case oI the in-phase component. This suggests that this
eIIect should be taken into account when calculating BER in nonlinear coherent links.
Later in this chapter, BER degradation in coherent ASK systems due to this eIIect is
examined. The physical origin oI this nonlinear noise evolution is the same as the
origin oI the Gordon-Mollenauer eIIect |18, 19, 21, 28| and in Iact this phenomenon
should be understood as the maniIestation oI the Gordon-Mollenauer eIIect Ior the
noise quadratures.
185
-2
-1
0
1
2
3
4
5
6
7
8
0 2 4 6 8 10 12 14
M^2 [mW/mW]
kurtosis
X component
Y component

Fig. 9.2 Analytical dependence oI the kurtosis on M
2
; P
0
1.0 mW, : 50km
These results can be compared with the results in |12, 28|, where the evolution
oI pdIs in Iibre with small dispersion was investigated based on simulations. In that
work however, the nonlinear eIIects appeared aIter a longer propagation length due to
reduction in the peak power. The probability contours Irom the joint pdI oI the two
quadratures shown in |13, 14| are very similar in shape to the contours Iound Ior
ZMNL Iibre Irom Monte Carlo simulations (the shape oI probability contours in
ZMNL Iibre is discussed in more detail later in this chapter, when the analytical
expressions Ior the pdIs oI the quadratures are derived). This suggests that it should
be possible to decouple the dispersion and nonlinearity such that the physical system
can be represented as ZMNL Iibre, Iollowed by an optical Iilter where the ZMNL
block would model the 'eIIective nonlinearity and the Iilter would model the
dispersion.
9.3.2.1.Verification using Monte-Carlo simulations
To veriIy the validity and correctness oI the analytical expressions (9.24),
extensive Monte-Carlo simulations were perIormed based both on the ZMNL
transIormation Iunction (9.8) and on the Iull numerical solution oI the NLSE, using
commercial simulation package VPItransmissionMaker, which utilizes the split-step
Fourier method to solve NLSE (9.1). In the approach based on (9.8), a sample set oI
100 000 samples was generated using Gaussian random number generator and the
transIormed set was calculated using (9.8). Moments oI the transIormed set were
computed.
186
Noise only case: To veriIy the observation that Kerr nonlinearity does not
break the Gaussian property in this situation, ASE only was propagated (using both
approaches explained above) and the kurtoses oI the output set were computed as
explained above. Even Ior very large ASE power oI P
0
20 mW, the kurtosis oI both
quadratures consistently maintained the value oI 0 strongly suggesting Gaussian
character oI the noise. The deviations Irom this value were very small (less than 0.03)
and typical Ior Monte Carlo simulations.
Noise and signal case: The procedure here is identical to the one described
above. Very good agreement with the analytical results has been Iound. As an
example, Fig. 9.3 corresponds to Fig. 9.1, and shows the evolution oI the kurtoses
with the propagation length. The parameters in the simulation were identical to those
used to obtain Fig. 9.1 Irom analytical expressions. Additionally,
VPItransmissionMaker simulations have also been perIormed to check the
consistency oI the model with well established numerical methods. As seen in Fig.
9.4, simulation results again agree with analytical values. Fig. 9.5 presents the
dependence oI kurtosis on M
2
determined using Monte Carlo simulations and
corresponds to Fig. 9.2. The corresponding Iigures are very similar. ThereIore, it can
be concluded that the analytical expressions (9.24) are correct.
-2
-1
0
1
2
3
4
5
6
7
8
0 10 20 30 40 50 60 70 80 90 100
Iength [km]
kurtosis
Po=0.8mW,X
Po=0.8mW,Y
Po=1.0mW,X
Po=1.0mW,Y
Po=1.5mW,X
Po=1.5mW,Y

Fig. 9.3 Evolution kurtosis against propagation length, conditions correspond to Fig. 9. 1. Monte-Carlo
simulations
187

Fig. 9.4 Evolution kurtosis against propagation length, conditions correspond to Fig. 9. 1.
VPItransmissionMaker simulations

Fig. 9.5 Dependence oI the kurtosis on M
2
; conditions correspond to Fig. 9. 2. Monte-Carlo simulations
Kurtosis is oIten used as an indicator oI the relative weight oI the tails oI the
pdI. In many physical cases, more sharply peaked distributions (larger kurtosis) have
heavier weights and Ilat-topped pdIs have lighter tails (an extreme example is the
uniIorm distribution without any tails at all) |29|. ThereIore, it may be expected that
the Kerr nonlinearity has signiIicant impact on BER in coherent ASK systems, as
their perIormance directly depends on the pdIs oI the quadratures (in particular, on the
pdI oI the in-phase noise component). This impact is quantiIied in Section 9.4.2.
188
Direct calculation also shows that the skewness (deIined as the ratio m
3
/m
2
3/2
)
oI the pdIs also changes so that the pdIs are no longer symmetrical. However, since a
skewness oI 0 is not unique to a Gaussian pdI, as many pdIs are symmetrical,
(although a skewness diIIerent Irom this value does indicate that a pdI is not
Gaussian), this moment was not closely investigated in this work (Ior the sake oI
mathematical rigor, it should be mentioned that there exist other distributions with
kurtosis oI 0, but they are rare and tend to be unphysical |29|). The third moments
E|A
R,I
3
| were still necessary to calculate the kurtosis Irom (9.27).
9.4. Propagation of pdfs of the quadratures in ZMNL fibre
Although investigations oI the evolution oI the statistical moments presented
in Section 9.3 are very instructive to show the deviations Irom Gaussian character
and, as shown, the pdIs oI the quadratures may be reconstructed Irom the moments
(Ior example by using expansions similar to the Laguerre expansion used in Chapter
7) as the moments are Iinite, in most cases it is desirable to have a closed analytical
expression Ior the pdI. In this section, such an expression Ior the joint pdI oI the
quadratures is derived. Furthermore, this expression is later used to examine the
inIluence oI the Iibre nonlinearity on BER in coherent ASK systems. It can be noted
that the expressions Ior moments derived in Section 9.3 are suIIicient to estimate the
BER using Q-Iactor. It will be however shown that such an estimation may be very
inaccurate.
9.4.1. Evolution of the pdfs of the ASE quadratures
Our starting point is again equation (9.8), describing the ZMNL Iibre
transIormation Iunction, and (9.6) describing the input Gaussian statistics. As the Iirst
step, the pdI oI A
0
(given by (9.6)) is transIormed into polar coordinates Ior magnitude
p and phase . In this representation, it is given by
|
|
.
|

\
| +
=
2
cos 2
exp
2
) , (
2 2
L
L
M M
p
L
|
t

|
|
(9.31)
Eq. (9.31) gives the joint pdI Ior p,
L
. Now, the joint pdI oI the magnitude p and the
total phase


L

NL
is obtained Irom the transIormation

+ =
=
L
D | |

2
,
(9.32)
189
The argument oI D has been dropped to simpliIy notation, but the dependence oI D on
: through (9.10) is understood. The Jacobian oI the transIormation (9.32) is J1 and
the system (9.32) has only one solution. Thus, the joint pdI oI p, is
|
|
.
|

\
| +
=
2
) cos( 2
exp
2
) , (
2 2 2
D M M
p
|
t

|
|
(9.33)
To obtain the joint pdI oI the two quadratures, another transIormation should be used

=
=
|
|
cos
, sin
R
I
A
A
(9.34)
The Jacobian oI this transIormation is J
2 2
I R
A A + so Iinally the joint pdI oI the
two quadratures is
( )
|
|
.
|

\
|
+ + + + +

=
2
) sin( ) cos( 2
exp
2
1
) , (
2 2 2 2 2 2 2
D A D A A D A D A A M M A A
A A p
R I I R I R R I
I R A A
I R
t
(9.35)
The evolution oI the initially Gaussian joint pdI along the ZMNL Iibre is
shown in Figs.9.6a-d. The contours oI constant probability in the A
R
, A
I
plane are
clearly deIormed Irom their initial circular shapes (Fig.9.6a, hence the name circular
Gaussian noise) and assume progressively more complicated shapes. It can be
observed that with signiIicant nonlinearity, the contours cannot be approximated by
ellipses. Hence, the assumption that the noise retains its Gaussian character but simply
develops correlation between the quadratures is not valid. The contours are plotted Ior
P
0
1.5 mW, M
2
10 and the nonlinear coeIIicient y2 W
-1
km
-1
. Note that (9.35)
predicts that ASE travelling alone (M0) does not change its Gaussian character,
which was already observed when the evolution oI moments was discussed. It is also
worth mentioning that even iI (9.35) predicts that the evolution oI the pdIs is periodic
in D, this observation cannot be used to compensate Ior the eIIect oI the accumulated
nonlinearity as in practice D is always much smaller than the Iirst period.
190

Fig. 9.6 Evolution oI the joint pdI along the Iibre: a) D0, b) D0.02, c) D0.04, d) D0.06. In each
case P
0
1.5 mW and M
2
10.
The analytical results obtained Ior ZMNL agree well with the contours
obtained Irom Monte Carlo simulations in |11|. Although it was argued in |11| that
the presence oI dispersion signiIicantly reduces the strength oI the nonlinear
interaction between the signal and noise such that the resultant noise pdI can still be
approximated as non-circular Gaussian (with induced correlation between the
quadratures), the results reported in |12| suggest that in many cases this is an
oversimpliIied approach. In such cases, the analytical model proposed here may be
Iitted to the simulated data (and thereIore may be used to estimate the accumulated
eIIective nonlinearity in the presence oI dispersion), signiIicantly reducing the
number oI required simulated samples. Note that the shape oI the joint pdI depends
both on D and M. For higher values oI M (which is equivalent to a scaled signal to
noise ratio), typical Ior digital communications, the pdI assumes a crescent-like shape,
very similar to the one reported in |12|, where it was Iound based on numerical
simulations.
The Iirst order pdI oI each quadrature, oIten oI more interest Irom the practical
point oI view (certainly in coherent ASK systems) can be obtained Irom (9.35) aIter
integrating out the other component
191
( )
( )

|
|
.
|

\
|
+ + + + +

=
|
|
.
|

\
|
+ + + + +

=
}
}


R
R I I R I R R I
I A
I
R I I R I R R I
R A
dA
D A D A A D A D A A M M A A
A p
dA
D A D A A D A D A A M M A A
A p
I
R
2
) sin( ) cos( 2
exp
2
1

) (
2
) sin( ) cos( 2
exp
2
1

) (
2 2 2 2 2 2 2
2 2 2 2 2 2 2
t
t
(9.36)
The integration in (9.36) is possible to be carried out analytically by applying
the Jacobi-Anger expansion |26|, which expands the Iormula oI the type
exp(xcos(B)) in an inIinite series containing Bessel Iunctions (analogous expansion
Ior exp(xsin(B)) exists as well). However, the resultant expressions are given in terms
oI an inIinite sum oI inIinite sum oI terms containing Bessel Iunctions and are
diIIicult to handle. ThereIore, in practice the Iinal pdIs can be obtained by numerical
integration. To check the validity oI the derivation presented above, the pdIs Irom
(9.36) are compared with normalized histograms Irom Monte-Carlo simulations. The
histograms are generated in a way described in Section 9.3.2.1, i.e. a set oI 100 000
Gaussian samples was propagated according to (9.8).
The output histograms oI the two quadratures are shown in Figs. 9.7a and
9.7b. The histograms were again obtained Ior P
0
1.5 mW and M
2
10. The
propagation length was : 50 km, which corresponds to the case D0.06 in Fig. 9.6d
and the attenuation was o0.046 km
-1
(0.25 dB/km). To check the validity oI the
analytical expressions (9.36), Figs. 9.8a and 9.8b present the analytically calculated
pdIs Ior the same conditions. It should be noted that the simulated and analytical pdIs
are very similar. As expected Irom Fig. 9.6, the pdIs are Iar Irom Gaussian, with the
real component more peaked and the imaginary component wider, with increased
variance. Both pdIs are asymmetric. This shows that the Kerr nonlinearity indeed
destroys the Gaussian nature oI the noise.
192

Fig. 9.7 Simulated histograms oI the two quadratures. a) Real component b) Imaginary component.

Fig. 9.8 Calculated pdIs oI the two quadratures. a) Real component b) Imaginary component.
9.4.2. Influence of Kerr nonlinearity on BER in ASK systems
As already mentioned, the Kerr nonlinearity has diIIerent impacts on direct
detection and coherent ASK systems. In this section, the negative impact on the
coherent systems is discussed in detail. However Iirstly, it is brieIly shown that
statistics oI optical power are not changed by Iibre nonlinearity, which explains why
direct detection systems are largely immune to Gordon-Mollenauer eIIect.
9.4.2.1.Direct detection systems
It can be observed by calculating appropriate moments Irom (9.24) that the
signal power may be transIerred to the variances oI the quadratures due to the Kerr
nonlinearity. The visible broadening oI the pdIs in Fig.9.7b and 9.8b is a result oI this
power transIer. ThereIore, it might be somewhat surprising that this increased noise
193
power does not deteriorate the perIormance oI direct detected ASK systems.
However, it may be Iormally shown that when direct detection is used, no
perIormance degradation is observed in the regime considered here. This can be
expressed mathematically by suitably transIorming the pdIs oI the quadratures. In
direct detection systems, the quantity oI interest is the optical power, related to the
quadratures by PA
R

2
A
I

2
. Proceeding in a way similar to that utilized to derive
(9.35), the joint pdI oI the power P and the total phase is given by:
|
|
.
|

\
|
+
=
2
) cos( 2
exp
4
1
) , (
2
DP P M M P
P p
I
|
t
|
|
(9.37)
The marginal pdIs Ior phase and power can be obtained by integrating out the
other component. The pdI oI phase is very sensitive to the accumulated nonlinearity D
but the power pdI is not, despite the term DP in Iormula (9.37). In Iact, it can be
easily shown, with the help oI the identity |30|
) ( 2 )) sin cos ( exp(
2 2
0
2
0
v x B I d x v B + = +
}
t o o o
t

(where B is an arbitrary constant) that the pdI oI the optical power is given by the
noncentral chi-squared statistics
) (
2
exp
2
1
) (
0
2
P M I
M P
P p
I
|
|
.
|

\
| +
=
t
|
(9.38)
Equation (9.38) is a standard well-known expression Ior the pdI oI the squared
envelope in a narrowband Gaussian noise |30| and is not dependent on D. This shows
that the Kerr nonlinearity does not directly aIIect the perIormance oI the direct
detection systems, even though it may signiIicantly increase the variances oI the two
quadratures. Indeed, the result just Iormally proven can be deduced Irom (9.8), as
already mentioned. The increased individual quadrature variances do not cause BER
degradation because the quadratures are no longer independent. Note however, that iI
additional optical signals (like relative intensity noise |31|) are present at the input oI
the photodetector, the resultant pdI is expected to be diIIerent, due to the mixing
between other optical signals. It has been however reported that Kerr nonlinearity
does change the BER to some extent also in direct detection systems |12|. This is
mainly due to the Iact nonlinear propagation also changes the spectrum oI the ASE
194
and ampliIies some Irequency components. This eIIect is not visible in our derivation
since matched Iilters are implicitly assumed. Additionally, in the presence oI
dispersion, phase Iluctuations may be converted into amplitude Iluctuations.
9.4.2.2.Coherent amplitude modulated systems
The situation is dramatically diIIerent in the case oI coherently detected ASK
systems. Although direct detection systems are Iar more prevalent, coherent detection
is still considered and being researched since higher sensitivity and higher spectral
eIIiciency can be achieved using this scheme |8, 32|. In particular, optical quadrature-
amplitude modulation (QAM) is receiving renewed attention as it promises very high
spectral eIIiciency |8|. ThereIore, it is oI interest to investigate how the Iibre
nonlinearity changes the transmission quality in this case. When coherent
(synchronous) detection is used, the detected sample is just one quadrature oI the
complex ASE noise (when a perIect detector is assumed, i.e. thermal and shot noises
are neglected) |30|. When no nonlinearity is present, the noisy signal obviously obeys
Gaussian statistics, as already discussed. In the presence oI the Kerr nonlinearity, our
investigations showed that this is not longer the case. BeIore calculating the BER in
the presence oI nonlinearity, it should be noted that signal power is coupled to the
orthogonal quadrature through the Kerr eIIect, as the optical wave propagates. This
simply amounts to rotating the coordinate Irame, such that both quadratures are non-
zero mean. This should not be conIused with the parametric gain eIIect, where the
signal power acts as a pump and power is transIerred between the signal and noise,
resulting in increased noise variances. ThereIore, beIore coherent detection can be
applied, it is convenient to rotate the coordinate system such that the whole signal is
projected along one axis (the noise has to retain its zero-mean character). The rotation
transIormation is given by |24|

+ =
+ =


cos sin '
, sin cos '
I R I
I R R
A A A
A A A
(9.39)
where
R
A' ,
I
A' are quadratures in the rotated Irame and the angle oI rotation is the
angle the mean signal makes with the real axis in the old Irame. Using (9.35) and
(9.39) and aIter some algebra, the joint pdI Ior the quadratures in the new Irame is
given by
195
|
|
.
|

\
|
+
|
|
.
|

\
|
+ + +
=
2
) ' ' sin( ' 2
exp
2
) ' ' cos( ' 2 ' '
exp
2
1
) ' , ' (
2 2
2 2 2 2 2
' '

t
D A D A MA
D A D A MA M A A
A A p
R I I
R I R R I
I R A A
I R

(9.40)
The pdIs Ior the single quadratures are, as usual, obtained by integrating out
the other quadrature. Fig. 9.9 presents BER in the coherent detection system as a
Iunction oI the signal to noise power (M) beIore the ZMNL Iibre. Total nonlinearity D
is used as a parameter. PerIect extinction ratio is assumed i.e. pdI oI the zero` bit is
given by the Gaussian distribution (as noise alone is not nonlinearly transIormed) and
pdI oI the one` bit is nonlinearly transIormed. The optimum threshold was
numerically calculated, such that the pdIs oI zero` and one` bits intersect. It can be
seen that the presence oI nonlinearity signiIicantly degrades the BER, especially at
low BER. This eIIect is most visible Ior high M values, when the nonlinear terms in
(9.40) become more important. This is in agreement with Fig. 9.2, which shows that
the nonlinear evolution oI kurtosis depends on M. It should be noted that the BER
degradation is visible even iI the joint pdI contours do not exhibit apparent deviations
Irom Gaussian shapes. This cannot be attributed to an increased variance (Irom the
parametric gain), as in all cases shown in Fig. 9.9 the detected noise variance was
actually slightly smaller than the input variance (variance oI the other quadrature,
discarded in the process oI coherent detection, was however ampliIied), so this BER
degradation is entirely due to the nonlinearly transIormed pdI. This shows that the use
oI Gaussian approximation Ior the BER (the Q-Iactor), when even moderate
nonlinearity is present, would lead to signiIicant BER estimation errors (since the
variance was slightly reduced, BER curve estimated using the Q-Iactor would lie just
below the case D0 in Fig. 9.9). Fig. 9.9 conIirms that the Kerr nonlinearity may
signiIicantly degrade the perIormance oI coherently detected ASK systems. This is in
contrast with the direct detection systems which, in the regime investigated in this
chapter, are insensitive to this nonlinearity.
196

Fig. 9.9 Degradation oI the BER perIormance in coherent detection systems due to the Kerr
nonlinearity
9.5. Summary
In this chapter, the inIluence oI the Kerr nonlinearity on direct detection and
coherent ASK was investigated. A ZMNL Iibre approximation was adopted to be able
to derive analytical Iormulae Ior evolution oI the moments and pdIs oI the ASE
quadratures. The analytical expressions have been compared with Monte-Carlo
simulations to make sure they are correct. It was shown that Kerr nonlinearity has
deleterious eIIect in coherent ASK systems. Interestingly, the BER penalty is not
accompanied by the increase in the noise variance but it is caused by the change oI
shape oI the pdI oI the in-phase quadrature. This shows that the application oI the Q-
Iactor method to estimate BER in coherent ASK links may lead to large estimation
errors, thereIore one needs to know signal history to be able to make a judgment iI the
Q-Iactor method is credible and applicable.







197
9.6. References
|1| D. Marcuse, "Calculation oI bit-error probability Ior a lightwave system with
optical ampliIiers and post-detection Gaussian noise," J. Lightw. Technol.,
vol. 9, pp. 505-513, 1991.
|2| P. A. Humblet and M. Azizoglu, "On the bit error rate oI lightwave systems
with optical ampliIiers," J. Lightw. Technol., vol. 9, pp. 1576-1582, 1991.
|3| G. Einarsson, Principles of Lightwave Communications. New York: Wiley,
1996.
|4| C. Dorrer, "Monitoring oI optical signals Irom constellation diagrams
measured with linear optical sampling," J. Lightw. Technol., vol. 24, pp. 313-
321, 2006.
|5| B. Slater, S. Boscolo, V. K. Mezentsev, and S. K. Turitsyn, "Comparison oI
BER estimation methods in numerical simulation oI 40 Gbit/s RZ-DPSK
transmission," Conf. Lasers Electro-Opt. CLEO 2007, pp. 1, 2007.
|6| W. Shieh, Q. Yang, and Y. Ma, "107 Gb/s coherent optical OFDM
transmission over 1000-km SSMF Iiber using orthogonal band multiplexing,"
Opt. Express, vol. 16, pp. 6378-6386, 2008.
|7| G. Bosco and P. Poggiolini, "On the Q Iactor inaccuracy in the perIormance
analysis oI optical direct-detection DPSK systems," IEEE Photon. Technol.
Lett., vol. 16, pp. 665-667, 2004.
|8| J. Hongo, K. Kasai, M. Yoshida, and M. A.-N. Nakazawa, M., "1-Gsymbol/s
64-QAM Coherent Optical Transmission Over 150 km," IEEE Photon.
Technol. Lett., vol. 19, pp. 638-640, 2007.
|9| S. Primak, V. Kontorovitch, and V. Lyandres, Stochastic Methods and their
Applications to Communications. Stochastic Differential Equations Approach:
John Wiley & Sons, 2005.
|10| M. P. Dlubek, A. J. Phillips, and E. C. Larkins, "Nonlinear Evolution oI
Gaussian ASE Noise in ZMNL Fiber," J. Lightw. Technol., vol. 26, pp. 891-
898, 2008.
|11| P. Serena, A. Orlandini, and A. Bononi, "A Parametric Gain Approach to
PerIormance Evaluation oI DPSK/DQPSK Systems with Nonlinear Phase
Noise," pp. 129-136, in Optical Communication Theorv and Techniques, E.
Forestieri, Ed., 2005.
198
|12| E. Vanin, G. Jacobsen, and A. Berntson, "Correlated noise generation in high-
speed transmission Iiber links," Europ. Conf. Netw. Opt. Commun. NOC 2007,
vol. 12, 2007.
|13| M. P. Dlubek, A. J. Phillips, and E. C. Larkins, "Evolution oI the probability
density Iunctions oI Gaussian ASE noise in zero-memory nonlinear Iiber,"
accepted for publication in Opt. Fiber Technol.
|14| G. P. Agrawal, Nonlinear Fiber Optics, 4 ed: Academic Press, 2007.
|15| C. R. Menyuk, "Application oI multiple-length-scale methods to the study oI
optical Iiber transmission," J. Eng. Math., vol. 36, pp. 113-136, 1999.
|16| J. Garnier, L. Videau, C. Gouedard, and A. Migus, "Propagation and
ampliIication oI incoherent pulses in dispersive and nonlinear media," J. Opt.
Soc. Am. B, vol. 15, pp. 2773-2781, 1998.
|17| J. P. Gordon and H. A. Haus, "Random walk oI coherently ampliIied solitons
in optical Iiber transmission," Opt. Lett., vol. 11, pp. 665-667, 1986.
|18| J. P. Gordon and L. F. Mollenauer, "Phase noise in photonic communications
systems using linear ampliIiers," Opt. Lett., vol. 15, pp. 1351-4, 1990.
|19| K.-P. Ho, "PerIormance degradation oI phase-modulated systems due to
nonlinear phase noise," IEEE Photon. Technol. Lett., vol. 15, pp. 1213-1215,
2003.
|20| A. Mecozzi, "Long-distance transmission at zero dispersion: combined eIIect
oI the Kerr nonlinearity and the noise oI the in-line ampliIiers," J. Opt. Soc.
Am. B, vol. 11, pp. 462-469, 1994.
|21| A. Mecozzi, "Limits to long-haul coherent transmission set by the Kerr
nonlinearity and noise oI the in-line ampliIiers," J. Lightw. Technol., vol. 12,
pp. 1993-2000, 1994.
|22| K.-P. Ho, "Asymptotic probability density oI nonlinear phase noise," Opt.
Lett., vol. 28, pp. 1350-1352, 2003.
|23| A. Mecozzi, "Probability density Iunctions oI the nonlinear phase noise," Opt.
Lett., vol. 29, pp. 673-675, 2004.
|24| A. Papoulis, Probabilitv, random variables, and stochastic processes, 3 ed:
McGraw-Hill Book Co., 1991.
|25| G. Bosco, A. Carena, V. Curri, R. Gaudino, P. Poggiolini, and S. Benedetto,
"A novel analytical method Ior the BER evaluation in optical systems aIIected
by parametric gain," IEEE Photon. Technol. Lett., vol. 12, pp. 152-154, 2000.
199
|26| M. Abramowitz and I. A. Stegun, Handbook of Mathematical Functions with
Formulas, Graphs, and Mathematical Tables, 10 ed: U.S. Department oI
Commerce, 1972.
|27| I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series, and Products, 6
ed: Academic Press, 2000.
|28| E. Vanin, G. Jacobsen, and A. Berntson, "Nonlinear phase noise separation
method Ior on-oII keying transmission system modeling with non-Gaussian
noise generation in optical Iibers," Opt. Lett., vol. 32, pp. 1740-1742, 2007.
|29| A. Stuart and J. K. Ord, Kendalls Advanced Theorv of Statistics. London:
Charles GriIIin & Co., 1987.
|30| J. G. Proakis, Digital Communications, 4 ed. New York: McGraw Hill, 2000.
|31| I. Jacobs, "Dependence oI optical ampliIier noise Iigure on relative-intensity-
noise," J. Lightw. Technol., vol. 13, pp. 1461-1465, 1995.
|32| S. Camatel and V. Ferrero, "Homodyne coherent detection oI ASK and PSK
signals perIormed by a subcarrier optical phase-locked loop," IEEE Photon.
Technol. Lett., vol. 18, pp. 142-144, 2006.



200
10. Conclusions and future work
This thesis has been concerned with the subject oI optical perIormance
monitoring in Iuture transparent optical networks. The Iirst Iour chapters serve as an
introduction to various aspects oI this wide subject, explaining the need Ior new
monitoring methods in optical networks and clariIying why, as the complexity and
transparency in optical networks increase, OPM methods appear to be best suited to
IulIil the Iuture requirements Ior seamless and Ilexible signal health and network state
monitoring.
There are a host oI diIIerent physical quantities possible to monitor using
OPM methods. The trade-oII between the required monitoring capabilities and the
cost and complexity oI the monitoring subsystem is delicate and, in some cases, the
Iinal judgement may depend on the particular network scenario. However, probably
the two most important signal quality metrics are the BER and OSNR. This is due to
the Iact that the BER is the ultimate quality measure Ior any digital signal. The
importance oI OSNR stems Irom its intimate relation with BER in linear systems and
Irom the special position optical ampliIiers hold in transparent optical networks.
ThereIore, in this thesis we have mainly concentrated on these two aspects oI OPM,
which is BER estimation and OSNR monitoring.
As repeatedly emphasised throughout this thesis, BER estimation without the
Iull signal detection is not an easy task. Optical sampling, when coupled with suitable
data processing, is seen as a promising method, which could be utilized Ior bitrate-
transparent BER monitoring. The data processing typically amounts to generating the
signal histograms and then estimating BER based on them. ThereIore, Chapter 6 has
presented experimental investigations oI the Ieasibility oI using FWM-based optical
sampling in SOAs Ior signal quality monitoring. Experimentally determined
perIormance oI such an optical sampling system has been discussed and compared
with the theoretical model. Comprehensive noise analysis has been given, which
suggests that the perIormance oI the investigated sampling scheme may not be
suIIicient Ior practical applications requiring high accuracy and Ilexibility. The
discussion presented in Chapter 6 has been preceded with more basic experimental
investigations oI the perIormance and behaviour oI SOAs in general and FWM in
SOAs in particular. These investigations have been given in Chapter 5 and resulted in
201
suggesting an alternative physical explanation oI the phenomenon oI carrier density
pulsations mediating FWM and in proposing a novel simple method Ior optical signal
extinction ratio improvement.
The theoretical aspect oI BER estimation has also been examined within the
scope oI this thesis. In Chapter 7, a novel optical signal quality metric has been
proposed. The proposed metric is based on Laguerre expansion and statistical
moments. Full derivation oI the new metric is presented. The perIormance oI the new
metric is assessed through simulations and its advantages and limitations are
discussed. The method is compared to the currently most popular method, the Q-
Iactor. It has been shown that in direct detection systems our method oIIers certain
beneIits as compared to the Q-Iactor approach. In particular, an accurate threshold
estimation is also possible using our method, which, as discussed in Chapter 7, is
useIul in some applications. The method makes use oI the Iirst three statistical
moments oI the optical signal, which makes it well suited to apply in tandem with an
optical sampling system. Moreover, the method can also Iind use in numerical
simulations oI optical links, as the brute Iorce approaches to estimate the BER Irom
Monte Carlo simulations are prohibitively computationally expensive.
The subject oI OSNR monitoring has been addressed in Chapter 8, where an
alternative OSNR monitoring method has been proposed. As mentioned in Chapter 8,
as the complexity oI optical networks increases, the need Ior in-band OSNR
monitoring becomes more and more apparent. We have proposed to tackle this
problem by making use oI the knowledge oI the signal bit rate and modulation Iormat.
These two parameters determine the optical signal modulation spectrum and thereIore
any deviations Irom this spectrum due to the presence oI ASE noise can be detected
by measuring the ratio oI the powers in two spectral slices taken Irom the signal. The
major advantages oI the method are its simplicity and cost-eIIectiveness. The
robustness oI the method against other signal impairments changing the signal optical
spectrum has been assessed through simulations. The method appears to be
reasonably robust against other spectrum changing phenomena investigated in
Chapter 8.
In Chapter 9, the issue oI BER estimation and modelling was explored Irom
another angle. In this chapter, the evolution oI the quadratures oI ASE noise in ZMNL
Iibre has been thoroughly investigated. The analytical expressions Ior the evolution oI
statistical moments and pdIs oI ASE quadratures have been derived. Our
202
investigations show that ASE looses its Gaussian character when propagating
alongside a deterministic signal in nonlinear optical Iibre but not when it propagates
alone. This has a direct impact on the BER in coherent ASK systems. It has been
shown that the nonlinearity has a signiIicant negative inIluence on the perIormance oI
such systems. This is because it leads to a nonlinearly transIormed pdI oI the ASE
noise, with Iatter tails. Moreover, it has been shown that application oI the Q-Iactor in
coherent ASK systems requires a dose oI caution. Even iI the Q-Iactor method may be
expected to be strictly valid in linear ASK links, the presence oI Iibre nonlinearity
may result in large estimation errors when attempting to use the Q-Iactor. As Iound in
Chapter 9, this is because, despite the Iact that the tails oI the pdIs are Iatter, the
variance oI the in-phase component is slightly reduced in the process oI nonlinear
noise evolution.
10.1. Future work
There are several exciting research possibilities, which can be pursued to
improve and Iurther develop the topics investigated in this thesis. As regards optical
sampling using SOAs, it was mentioned in the summary oI Chapter 6, that large
XPM-induced spectral shiIts combined with an optical Iilter may also be used Ior
sampling purposes. Although the attainable spectral shiIt may be too small Ior very
large bandwidth signals, it is oI interest to experimentally evaluate the capabilities oI
this conIiguration and to compare it with the perIormance oI the sampling system
utilizing FWM, presented in Chapter 6. To improve the quality oI SOA-based
sampling systems, various interIerometric gates seem to oIIer better perIormance than
FWM-based sampling. Among them, SOAs in Mach-Zehnder conIigurations may be
the most promising. ThereIore, one research possibility is to perIorm a similar system
characterization as presented in Chapter 6 but Ior Mach-Zehnder-based sampling
gates. Mach-Zehnder conIiguration may also be interesting in the context oI the
extinction ratio enhancement method presented in Chapter 5. The extinction ratio
improvement method has been experimentally validated Ior relatively low bit rate
(4 Gb/s) signals. In practice, it is oI much greater interest to be able to optically
regenerate very Iast signals, above 10Gb/s. Thus, experimental evaluation oI the
speed limits oI the method is important. It is diIIicult to predict what the maximum
signal speed is Ior the simple regenerating conIiguration presented in Chapter 5.
However, it may be expected that placing the SOA within an interIerometer may
203
substantially improve the system response time and extend the applicability oI the
method by making use oI both XPM and FWM eIIects.
The novel BER estimation method, proposed in Chapter 7, should be
evaluated in a real link. This would give an indication about the robustness oI the
method against other eIIects impacting the BER. More analytical work is required to
accommodate the level splitting eIIect in the Laguerre method. Although in principle
this extension oI the method is not too complicated and should be equivalent to the
corresponding extension oI the Q-Iactor method (discussed in Chapter 4), the
accuracy oI the extended method should be evaluated. Additionally, it would be
interesting to evaluate the metric using numerical Monte Carlo simulations, i.e. by
numerically solving the NLSE and simulating other link components. This may be
done Ior example by using the JPItransmissionMaker simulator. It is especially
interesting to check the perIormance oI the method in long-haul nonlinear links, as
ASE may lose its Gaussian character in such cases (although, as shown in Chapter 9,
this has only a limited impact on direct detection systems). Also, it is worth
investigating the potential oI the method as a metric Ior Iacilitating link design.
Typical design approach is to maximize the Q-Iactor but, as discussed previously, this
does not guarantee the optimum perIormance Ior direct detection systems. ThereIore,
a Monte Carlo-based perIormance comparison oI the links designed Ior maximum Q-
Iactor and Ior optimizing the Laguerre metric is oI interest. Furthermore, as explained
in Chapter 7, one oI the advantages oI the new metric over the Q-Iactor method is that
the new metric correctly predicts the optimum threshold, whereas the Q-Iactor and
Gaussian noise model Iail when an accurate prediction oI the threshold is required, Ior
example Ior turbo-code decoders. ThereIore, the perIormance oI soIt decoders
employing the proposed BER model should be investigated. Finally, it would be very
interesting to combine the Laguerre expansion method with high quality optical
sampling system. The FWM-based optical sampling system investigated in this thesis
probably does not oIIer suIIicient sampling quality, but iI Mach-Zehnder-based
sampling system is built, the possibility to evaluate the joint applicability oI this
tandem (monitoring system BER estimation method) is certainly exciting.
A similar remark about the need Ior practical evaluation applies to the OSNR
monitoring method proposed in Chapter 8 as it is oI importance to check how the
method perIorms in practice against various impediments. Some more investigations
204
are needed to check the robustness oI the method against cross phase modulation and
linear incoherent crosstalk. Also, the calibration curves and perIormance oI the
method Ior other modulation Iormats can be investigated. II the method is combined
with some other technique to measure the OSNR, the optical power ratio initially
proposed to determine the OSNR may be instead used Ior modulation Iormat
recognition. As the automatic modulation recognition would be a useIul Ieature in
optical networks, the potential oI this application should be explored in more detail.
The Iindings oI Chapter 9 can be supplemented by an alternative signal quality
metric, which can be used in coherent ASK systems to IaithIully estimate the signal
quality even in the presence oI signiIicant Iibre nonlinearity. The method preIerably
ought to be oI comparable simplicity as the Q-Iactor and should be based on statistical
moments, as these are the quantities most easily derived Irom histograms. It seems
reasonable that the method should make use oI kurtosis, as this moment has been
shown to be a metric Ior accumulated nonlinearity. It should be noted that moments
are the quantities most readily available in practical scenarios. The derived Iormulae
Ior moments and pdIs oI the ASE quadratures can be a good starting point to derive
the new metric. Also, the analytical model presented in Chapter 9 should be extended
to include the orthogonally polarized ASE component. Even iI the nonlinear
interaction between orthogonally polarized waves is signiIicantly reduced compared
to the parallel polarization case, it is still Iinite and the model would be greatly
improved iI this interaction was incorporated. Direct comparison oI the analytically
derived joint pdI Ior the ASE quadratures with the numerically calculated counterpart
at the end oI a long-haul nonlinear link with dispersion should be explored. This
comparison with simulations would veriIy iI it is possible to model such a link as a
ZMNL Iibre Iollowed by a linear Iilter without unacceptable loss in accuracy. The
BER estimation Ior ASK systems presented in Chapter 9 does not include the eIIect oI
spectral evolution due to Iibre nonlinearity. ThereIore, another signiIicant
improvement oI the work presented in this thesis would be to take this eIIect into
account in the analytical model.
205
Appendix A. Derivation of formulae for optically sampled
moments
A derivation oI the optically sampled signal-spontaneous and spontaneous-
spontaneous noise variances has been given in |1|, where it was discussed that scaling
oI the optically sampled moments is necessary to obtain a meaningIul BER estimation
based on the Q-Iactor. The same consideration applies to any other BER estimation
technique employing optically sampled moments. The Q-Iactor method makes use oI
only the Iirst two moments, but other BER models may require higher moments. (One
example is given in Chapter 7, where the BER estimation method presented requires
the Iirst three moments). ThereIore a simpliIication and generalization oI the method
presented in |1| is summarized here. The approach presented here allows the relation
oI any arbitrary optically sampled moment to its electrically sampled counterpart.
To begin with, let us consider the optical sample a
C
(t) gated out Irom a noisy
signal Iield a
S
(t)
) ( )) ( ) ( ( ) ( t f t a t a t a
ASE S C
+ = q (A1)
where a
ASE
(t) is a complex circular Gaussian ASE noise. The photocurrent generated
by an ideal integrator detector is
t t f t a t a
T
R
i
ASE S
d ) ( ) ( ) (
2
2
}


+ =
q
(A2)
The n-th moment oI i is given by
( )
n
k k k ASE k ASE k S k ASE k S k S
n
n
t t f t a t a t a t a t a t a
T
R
i |
.
|

\
|
+ + + |
.
|

\
|
=
}
d ) ( ) ( ) ( ) ( ) ( ) ( ) (
2
2
* *
2 q
(A3)
where t
k
is the dummy time variable. Depending on the order n, the number oI
resulting terms grows accordingly. Fortunately, most oI these terms average to zero,
as explained below. Since the integration and statistical averaging are linear
operations, their order can be exchanged (obviously, the dummy variable in each oI
the n single integrals needs to be diIIerent to be able to use a common n-Iold
integration sign). As the signal Iield is deterministic, it can be pulled out oI the
ensemble averaging. As a result, the terms oI the general Iorm
) ( ) ( ... ) ( ) (
* *
q ASE p ASE l ASE k ASE
t a t a t a t a (A4)
206
need to be evaluated, where the exact structure oI each term Iollows Irom basic
algebra. The number oI multiplicative terms in (A4) equals at most n and is reduced
when the Iactor oI ,a
S
,
2
is present in Iront oI the ensemble averaging (aIter pulling it
out as the signal is deterministic). To evaluate these terms, it is most convenient to
make use oI the Gaussian moment theorem Ior the complex circular Gaussian process
|2|. This theorem states that (A4) averages to zero unless the number oI ordinary and
conjugate terms is equal. (This is a consequence oI the independency oI the noise
quadratures.) Furthermore, when (A4) does not vanish, it can be expressed in terms oI
the second order ASE moments, which is the physically accessible inIormation, as the
second moment is equivalent to the ASE spectrum. From the Gaussian moment
theorem, (A4) can be decomposed in the non-vanishing case to |2|


=
+
x
r ASE m ASE q ASE ASE p ASE ASE
m ASE m ASE m ASE ASE
t a t a t a t a t a t a
t a t a t a t a
) ( ) ( ) ( ) ( ) ( ) (
) ( ... ) ( ) ( ... ) (
* *
2
*
1
2
*
1
*
1
, (A5)
where the summation is over m! possible permutations producing pairings oI the Iorm
) ( ) (
*
r ASE p ASE
t a t a . For instance, Ior the Iirst nontrivial case when m2, there are Iour
possible pairing grouped in two terms:
) ( ) ( ) ( ) ( ) ( ) ( ) ( ) (
) ( ) ( ) ( ) (
3
*
2 4
*
1 4
*
2 3
*
1
4
*
3
*
2 1
t a t a t a t a t a t a t a t a
t a t a t a t a
ASE ASE ASE ASE ASE ASE ASE ASE
ASE ASE ASE ASE
+
=
(A6)
Similarly, the case m3 results in 6 terms, each oI them being a product oI
three second order terms. The second order temporal moment denoted as R(t
1
-t
2
) is the
noise autocorrelation. AIter some tedious algebra, the general Iormula Ior the third
order central moment becomes
3 2 1 3
2
2
2
1
2
3 1 2 1
3
3 2 1 3
2
2
2
1
2
3 2 3 1 2 1
3
3
d d d ) ( ) ( ) ( ) ( ) ( 6
d d d ) ( ) ( ) ( ) ( ) ( ) ( 2
t t t t f t f t f t t R t t R P
T
R
t t t t f t f t f t t R t t R t t R
T
R
S
opt
}}}
}}}
|
.
|

\
|
+ |
.
|

\
|
=
q
q

(A7)
The Iinal expression depends on the Iorm oI R(t
1
-t
2
), which is the Fourier
transIorm oI the ASE noise power spectrum. Following |1|, Gaussian shaped ASE
spectrum and sampling gate shape are assumed. With this assumption, the solution to
(A7) is
207
( )( )
|
|
.
|

\
|
A + A +
A
+
A +
A
=
2 2 2 2
2
3
3
2
2 2
3
3
3
3 3 3
3
1 3 1
3
3 1
2
t t
t
t
t
q
o o
o
ASE S
o
o
ASE opt
B B
B
T
S P
B
B
T
S R
(A8)
which, Ior the cases oI practical interest Ior optical sampling, can be approximated as

|
|
.
|

\
|
A +
A
+
A +
A
=
2 2
2
3
3
2
2 2
3
3
3
3 3 3
3
2 1
3
3 1
2
t
t
t
t
q
o
o
ASE S
o
o
ASE opt
B
B
T
S P
B
B
T
S R (A9)
II necessary, higher order moments can be calculated in the same way. Although the
algebra becomes lengthy, the analytical apparatus remains the same.



















A.1. References
|1| M. Westlund, H. Sunnerud, M. Karlsson, and P. A. Andrekson, "SoItware-
synchronized all-optical sampling Ior Iiber communication systems," J.
Lightw. Technol., vol. 23, pp. 1088-1099, 2005.
|2| J. W. Goodman, Statistical Optics, Wiley, 2000.
208
Appendix B. Derivation of the variance of the ASE-signal
beating in the presence of temporal gating in optical domain
In this Appendix, the derivation oI the ASE-signal beating variance
2
sig A
o is
given. Since ASE-signal beating is created only when the optical samples are present
at the receiver, whereas the receiver integration time is much longer than the sample
duration, the standard expression |1, 2| Ior this variance cannot be used. The
derivation presented here is in many aspects related to the derivation presented in |3|,
where the aim was to compare the optically sampled Q-Iactor with the Q-Iactor
measured using standard methods (the diIIerence between electrically measured and
optically sampled Q-Iactor is discussed in Chapter 4).
The squaring operation in (6.5) produces the Iollowing ASE-signal beating
terms
( )
}

- -

+ =
2 /
2 /
2 1 2 1
) ( ) ( ) ( ) ( ) ( ) ( ) (
T
T
ASE ASE sig A
dt t a t A t f t a t A t f
T
R
t i (B1)
where o q ) ( ) ( ) (
0 1
t h t f t f - = and
2 0
) ( ) ( ) ( o t h t A t a
ASE ASE
- = . The variance oI
(B1) can be calculated Irom
( )
( )
}
}

- -

- -

+
+ =
2 /
2 /
2 1 2 1
2 /
2 /
2 1 2 1
2
) ( ) ( ) ( ) ( ) ( ) (
) ( ) ( ) ( ) ( ) ( ) ( ) (
T
T
ASE ASE
T
T
ASE ASE sig A
ds s a s A s f s a s A s f
T
R
dt t a t A t f t a t A t f
T
R
t o
(B2)
where angle brackets denote statistical averaging. AIter perIorming the multiplication
operation, there are Iour terms containing the random a
ASE
(t) Iield: ) ( ) ( s a t a
ASE ASE
-
,
) ( ) ( s a t a
ASE ASE
-
, ) ( ) ( s a t a
ASE ASE
- -
and ) ( ) ( s a t a
ASE ASE
. Among them, the last two
terms average to zero, under assumption that the ASE is a complex circular Gaussian
noise, i.e. its quadratures are independent and identically distributed (see also
Appendix A). Although this is not necessarily true with regards to the optical noise
which has traversed a length oI nonlinear Iibre (as discussed in Chapter 9), Ior the
analyzed setup and Ior practical noise and signal power levels this assumption is
valid. Simple algebraic manipulations yield the Iollowing Iormula

209
| |
} }

- -

+ =
2 /
2 /
2 /
2 /
1 1 2
2
2
2
) ( ) ( ) ( ) ( ) ( ) (
T
T
T
T
ASE ASE ASE ASE sig A
dtds s a t a s a t a s f t f P
T
R
o (B3)
Since ASE noise at the SOA output is white, the two ensemble averaging
terms in (B3) are equivalent to the optical Iilter autocorrelation Iunction R(t-s), which
is assumed here to be stationary, which gives:
} }

=
2 /
2 /
2 /
2 /
1 1 2
2
2
2
) ( ) ( ) ( 2
T
T
T
T
sig A
dtds s t R s f t f P
T
R
o (B4)
In the experiments reported in Chapter 6, the Santec tunable bandpass Iilter
was used. This is a Gaussian shaped Iilter, thereIore its autocorrelation is also
Gaussian. Typically, the 3 dB FWHM Iilter bandwidth
dB
f
3
A is known or measured
using an optical spectrum analyser. This can be translated into the Gaussian width f A
by using 355 . 2
3dB
f f A = A . Finally, the Iilter autocorrelation Iunction R(t - s) is given
by ) 2 ) ( exp( 2 ) (
2 2
f ASE
s t f S s t R t t A = , where
1 2 2 2
) 4 (

A = f
f
t t and
f S P
ASE ASE
A = t 2 is the ASE noise power aIter the Iilter, copolarized with signal.
The sampling Iunction f
1
(t) depends on the shape oI the sampling pulses. The
Pritel soliton laser used in the experiments produces sech
2
pulses, which Ior practical
purposes can be very well approximated by the Gaussian Iunction. A Gaussian
spectrum Iiltered through a Gaussian Iilter retains its Gaussian character, so the pulse
shape aIter the Iilter remains Gaussian, oI reduced peak power and broadened
duration. For input FWHM pulse width
p
t , the Gaussian pulse width
is 355 . 2 /
p G
t t = . The input pulse ) 2 exp(
2 2
G
t P t is smeared by the Gaussian Iilter
(assuming that the Iilter centre Irequency is the same as the pulse carrier Irequency)
to ) 2 exp( 1
2 2 2 2
t t t t P
G f
+ , where
2 2 2
f G
t t t + = . As intuitively expected, Iiltering
through a Iilter much wider than the input pulse bandwidth has negligible eIIect on
the pulse shape. The pulse energy is obviously conserved by a lossless Iilter. For
simplicity, Iilter dispersion has been neglected in deriving the output shape. In most
cases the dispersion would result in slightly longer pulse widths but this also depends
on the initial pulse chirp, which Ior the Pritel laser is not known. In practice the
inIormation about the dispersion is not readily available, as most manuIacturers
provide only power spectrum.
210
The assumption oI Gaussian Iilter and pulse shape results in a simple and
practical analytical expression Ior
2
sig A
o . In this case, (B4) simpliIies to
f G
f G
f G
ASE sig A
P P
T
R
t t
t t
t t
oo
q
t o
2 2
2 2
2
2
2
2
2
5 2
2
4
+
+
=

(B5)
where the integration limits were extended to inIinity under the assumption that
incident pulses are much shorter than the integration time. (B5) constitutes the
Iormula Ior the variance oI ASE-signal beating in the optical sampling system based
on FWM in SOAs and is used in Chapter 6 to characterize noise properties oI this
type oI optical sampling system.














B.1.References
|1| G. Agrawal, Lightwave Technologv. Telecommunication Svstems. Hoboken:
Wiley, 2005.
|2| N. A. Olsson, "Lightwave systems with optical ampliIiers," J. Lightw.
Technol., vol. 7, pp. 1071-1082, 1989.
|3| M. Westlund, H. Sunnerud, M. Karlsson, and P. A. Andrekson, "SoItware-
synchronized all-optical sampling Ior Iiber communication systems," J.
Lightw. Technol., vol. 23, pp. 1088-1099, 2005.

S-ar putea să vă placă și