You are on page 1of 88

contrast,itisnot really possible to replace the amplifiers in an

undersea system without replacing the fiber as well, and it is far simpler
just to lay a new cable line—fibers, amplifiers, and all. As a
consequence, high-data-rate undersea systems use the
bestavailablefiber.Atpresent,itispossibletoobtainPMDvaluesaslowas

0.02 ps=km1=2, so that intrachannel PMD is not a serious problem in


undersea systems. Moreover, undersea amplifiers are typically
significantly more reliable than terrestrial amplifiers since repairs are
significantly more expensive. A mean time to failure of 20 years is not
exceptional for undersea systems.
Transoceanicsystems,asopposedtoshorterhaulunderseasystems,are
even more specialized. The PDLs in transoceanic amplifiers are
significantly lower than in terrestrial amplifiers. One normally specifies
values of 0.1 dB or less per amplifier, while values as high as 0.5 dB are
not uncommon in terrestrial amplifiers. Amplifier spacings are typically
around 50 km, which allows signal power to be held low while still
maintaining a good signal-to-noise ratio (SNR). Consequently, a
transoceanic transmission line will typically have in excess of
100amplifiers;bycontrast,terrestrialsystemsrarelyhavemorethan10.

Thus, we are led to consider systems with small values of PMD, PDL,
and
PDG,inwhichtheseeffectsaccumulateovermanyamplifiersandlongdistanc
es.
Inthislimit,thepolarizationeffectsdonotleadtomuchpulsedistortion.Inste
ad, they raise and lower the signal and noise power levels and rotate the
polarization state of each wavelength channel as a whole [19, 20]. As a
consequence, the polarization effects do not interact much with
nonlinearity and chromatic dispersion, so that it is possible to calculate
the penalties due to polarization effects separately from other penalties
[16, 17, 19, 20]. The separability of polarization effects is fortunate
because in practice one wants to specify an allowed margin for the
polarization effects, for example, 3 dB, and one wants to
ensurethattheprobabilitythattheactualpenaltywillexceedthisallowedmar
gin, the outage probability, is less than some small number such as 10 —
6
. It is not possible experimentally or through full time-domain
simulations to observe enough fiber realizations to calculate whether a
design’s outage probability is lower than the required specification or
not. However, it is possible to use a reduced model in which one just
follows the Stokes parameters of the signal and the noise of each
wavelength channel (eight numbers per wavelength channel)to
calculate the polarization penalties [16, 17]. One can easily calculate
105realizations from this approach, from which one can extrapolate to
obtain the
outageprobabilityattheallowedmargin.Inthenearfuture,weanticipatetha
tthe importance sampling technique, which has recently been applied to
calculating the penalties due to PMD in a high-PMD system, will allow us
to accurately calculate the outage probability at the allowedmargin.

In Section II of this chapter, we review our notation and the basic


equations that govern light propagation in optical fibers with rapidly
and randomlyvarying birefringence. In Section III, we derive the Stokes
parameter model and present the theoretical and experimental results
that validate it. We conclude with a discussion of applications to
underseasystems.

PROPAGATIONOFPOLARIZEDLIGHTINANOPTICALFIBER
TRANSMISSIONSYSTEM

FiberPropagation
Our starting point is to write the electric field in a single-mode optical
fiber in the form
o0

Eðx;y;z;tÞ ¼
1=2

½U1 ðz;tÞR1 ðx;y;o0 Þ þU2 ðz;tÞR2 ðx;y;o0 Þ]

2e0 c2 bðo0 Þ

~exp½ibðo0 Þz—io0 t] ð1Þ

whichisvalidintheslowlyvaryingenvelopeapproximation[5].Wehavechose
n the z direction to be the propagation direction along the fiber. The
dispersion relation b o is evaluated at the carrier frequency o o0. The
quantity t corresponds to physical time, while x and y indicate the
transverse dimensions, chosen so that ðx; y; zÞ form a right-handed
system. The vector field R1 is the transverse mode profile of the HE11
mode, which includes a small component in the z direction. In the weak-
guiding approximation, which is an excellent approximation for optical
fibers, we may choose R1 so that it is primarily oriented in the x
direction [3]. In that case, designating the unit vector in the z
directionase^ z ,wefindthatR2¼e^ z~R1isorientedprimarilyintheydirection.
ThecoefficientsU1z;tandU2z;t,whicharetheprincipalobjectsofourstudy,
contain all the effects of birefringence. Although, in principle, the
birefringence will lead to slight variations in the fields R1 and R2, in
practice, these variations
aretoosmalltohaveanyobservableeffects.Thefactor½o0 =2e0 c2 bðo0 Þ]has
been chosen so that jU1j þ jU2j corresponds to the optical power in the
weak
guidingapproximation,wheree0isthevacuumdielectricpermittivityandcist
he speed of light inthe vacuum.

Akeypointthatmeritssomeemphasisisthatweareusinganegativecarrie
r frequency,thatis,thefactorexp½ibðo0 Þz—io0 t]appearsinEq.(1),ratherthan
the factor exp io0t ib o0 z that corresponds to a positive carrier
frequency. As Gordon and Kogelnik [21] have pointed out, there is
considerable confusion of notation and nomenclature among
researchers studying polarization effects in
opticalfibers.Muchofthisconfusioncanbetracedbacktothewidespreaduse
of
both positive and negative carrier frequencies among these researchers.
Later in this section, we carefully discuss the current conventions in
nomenclature and notation and how our own choices compare. For the
moment, we note that our nomenclature is completely consistent with
Born and Wolf [22]. Like them, we
useanegativecarrierfrequency,andwedefinerightandleftcircularpolarizati
on, the Stokes parameters, and the Poincare´ sphere in precisely the
same way. One reason for our choice of a negative carrier frequency is
that this convention is overwhelmingly used among researchers who
are studying the impacts of nonlinearity and chromatic dispersion on
optical fiber transmission systems. Ultimately, it will be important to
study polarization effects in combination with other impairments,
rather than in isolation, as is the case with the majority of present-
daywork.Theuseofacommonconventionforthecarrierfrequencywill help
to eliminatemisunderstandings.

We now define the Stokes vectorU U1; U2 z, where thesuperscript

indicates the transpose so that U is interpreted as a column vector. We


also define
theretardedtimet t b0o0z,whereb0o0
dbo=doo¼o0.Theequationgoverning the z evolution
of U now becomes[5]

@U
i —iGUþDBUþiDB0@U—1b00@UþgΣjUj2 U 1UysUÞsUΣ¼0
@z @t 2

@t2

— ð 2 2
3

ð2Þ

where y indicates the conjugate transpose and

DB¼Dbðs3cosyþs1sinyÞ ð3Þ

The sj are the standard Pauli matrices

. 01Σ
s ¼ ; s

. Σ
0—i
¼ ; s

10

0 —1

Σ
ð4Þ

Primesindicatederivativeswithrespecttoo,evaluatedatthecarrierfrequ
ency
o¼o0 ,sothatb00¼d 2 b=do2 jo¼o.TheparameterGaccountsforthefiberloss,
whichispolarizationindependent.Theparameterg¼o0n2=cAeff,wheren
2isthe Kerr coefficient and Aeff is the fiber’s effective area, accounts
for the Kerr nonlinearity. The quantity Db indicates the
birefringence strength, while y
indicatesitsorientation,whichisrapidlyandrandomlyvaryingwithz.Wes
tress
thatthephysicalorientationoftheaxesofbirefringenceinsidethefiberisa
ctually given by y=2, not y [10]. It is useful to define the angle the
way that we have because this angle appears naturally in the Stokes
and Poincare´ sphererepre-
sentationsinwhichthephysicalangularseparationsaremultipliedbytwo.We
are assumingthatyisoindependentsothatDB0Db0s3cosys1siny.Weare also
assuming that birefringence does not contribute to the
chromaticdispersion.

Both assumptions are completely consistent with the experimental


evidence to date. We are also neglecting higher order dispersion and
nonlinearities otherthan the Kerr nonlinearity. It is not difficult to
modify Eq. (2) to include these effects
whenappropriate[5],andwetakeintoaccounthigherorderdispersioninsom
eof the system modeling that we present in SectionIII.

A key assumption that is implicit in Eq. (2) is that the local


birefringence contains no intrinsic helicity. Intrinsic helicity would
appear as a term that is
proportionaltos2.Thisassumptioniswelljustifiedbytheexperimentaldatat
o

dateeveninmoderatelytwistedfibers.Rashleigh[23]showedthatthehelicity

coefficient is quite small, and attempts to induce intrinsic helicity by


twisting the fiber as it is drawn have not proved successful [24]. By
contrast, twisting fiber
onceithasbeendrawnonascalecomparabletothebeatlengthwillleadtoafib
er evolution that in many ways mimics the behavior that is expected in a
fiber with intrinsic helicity even though the local birefringence is linear
[25]. One can also infer that the intrinsic helicity must be small from the
measured ratio of cross- to self-phase modulation. In a fiber with linear
birefringence, the ratio is 2=3. In a fiber with circular birefringence, the
ratio is 2. In a fiber with arbitrary birefringence, this ratio is somewhere
in between [26]. Experiments byBotineau
and Stolen [27] and by Stolen, Botineau, and Ashkin [28] demonstrated
the 2=3 ratio that is expected for linearly birefringent fibers. Later
theoretical work showed that the Kerr coefficient in a fiber with
randomly varying birefringence is expected to be 8=9 as strong in a
fiber with fixed birefringence when the intrinsic birefringence is linear
[29, 30]. The 8=9 factor has been verified experimentally by Buckland
and Boyd [31] and by Chernikov and Taylor [32].

To visualize the evolution of the polarization state of light as it


propagatesin the fiber, it is usefulto define the Stokes parameters,

S0¼Uy U; S1¼Uy s3 U; S2¼Uy s1 U; S3¼—Uy s2 U ð5Þ

The last three parameters together define the Stokes vector S S1; S2; S3 .
We note that S2S S, so that we can define a unit Stokes vector s S=S0.
The set of normalized Stokes vectors defines a sphere, referred to as
the Poincare´ sphere, shown in Fig. 2. The equator corresponds to
linearly polarized light at various orientation angles; the þs1 axis
corresponds to horizontal polarization; the —s1 axis corresponds to
vertical polarization; the þs2 axis corresponds to a 45○orientation; the s2
axis corresponds to a 135○orientation. As the latitude increases on the
sphere, the ellipticity increases. In the upper hemisphere, the light
appears to rotate clockwise as the observor faces an oncoming
beam,while, in the lower hemisphere, the light appears to rotate
counterclockwise. The þs3 axis corresponds to right circular polarization,
while the s3 axis corresponds to left circularpolarization.
FIGURE 2 The Poincare´ sphere.
When light propagates through a medium with fixed birefringence,
its polarization state at a single frequency traces a circle on the Poincare´
sphere around the eigenstates of the fiber, that is, the two polarization
states that propagate without changing in the medium. (Note that every
circle on a sphere has two centers.) Optical fibers are nearly linearly
birefringent, so that the eigenstates are close to the equator of the
Poincare´ sphere. In the case of polarization-
preservingfiber,inwhichthebirefringenceisnearlyfixed,thecircle traced on
the sphere resembles Fig. 3a. However, the random variations of the
axes of birefringence that occur in communications fibers move the
eigenstates randomly on the equator of the Poincare´ sphere. As a
consequence, the polarization state moves randomly over the entire
Poincare´ sphere andeventually covers it uniformly, as shown in Fig. 3b.
It is often naively supposed that the polarization state undergoes a
random walk on the sphere, but the reality is more complex. The
diffusion lengths in the equatorial and azimuthal directions arenot

equal in general and depend on both the beat length LB and the fiber
correlation length hfiber. This dependence is intricate and differs
depending on whether we consider a local frame that is tied to the local
axes of birefringence or a fixed laboratory frame [9]. It turns out that
the usual linear PMD only depends on the equatorial diffusion
coefficient in the local frame [9], which explains whythe

existence of several different diffusion lengths has not been much noted
to date. However, the other diffusion coefficients impact the interaction
of nonlinearity with polarization effects [9, 10] and have been
experimentally observed [33].We
noteaswellthatthelongestlengthscaleforthepolarizationstatetorandomiz
eon the Poincare´ sphere is several kilometers at most, which is short
compared to the
lengthscalesofinterestincommunicationsystems.Thus,itistypicallycorrec
tto assumethatthepolarizationstatesrapidlyrandomizeonthePoincare
´sphere.

Polarization ModeDispersion
Thus far, in our discussion of the polarization evolution, we have
focused on the
evolutionofasinglefrequency.Thedifferentialevolutionofnearbyfrequenci
esis what leads to PMD. To study the evolution due to PMD, we rewrite
Eq. (2) inthe
frequencydomain,neglectingchromaticdispersionandnonlinearity,toobta
in[5]

@U~ ~ ~
.
i —iGUþ DbþDb0 $Σ.s3cosyþs1sinyΣU¼0 ð6Þ
@z

where$ o
o0istheangularfrequencymeasuredwithrespecttothecarrier
frequency,and
U~ðz;$Þ¼
—1

dtexpði$tÞUðz;tÞ ð7Þ

istheFouriertransformofUðz;tÞ.Writingthe2~2identitymatrixasI,wenow

~ z —1~

Σ Ð Σ
defineV¼exp — 0Gðz1 Þdz1 RU,whereR¼cosðy=2ÞIþisinðy=2Þs2isthe
FIGURE 3 Evolution of the polarization state of a single frequency of light in a medium with (a)
fixed birefringence and (b) randomly varying birefringence.

matrixthatwoulddiagonalizetheevolutionofU~,wereitnotforthezvariationof

y. We thus obtain
@V~ ~
.
i þ½ DbþDb0 $Σs3þðyz =2Þs2 ]V¼0 ð8Þ
@z
whereyz¼dy=dz.WenowwriteW~ðz;$Þ ¼T—ðzÞV~ðz;$Þ,whereTsatisfies
1
the equation

@T

i þ½Dbs3þðyz =2Þs2 ]T¼0 ð9Þ


@z
whichcorrespondstoEq.(8)at$¼0.ItfollowsthatW~ðz; $¼0Þisconstant
andthetransformationW~ðz;$Þ ¼T—1 R—1 U~ðz;
$Þmeasurestherelativelyslow evolution of other frequencies relative to
o0. Explicitly, we now find

@W~

0
¯~

wheres¯3

¼ T—1s T.

þDb$s3 W¼0 ð10Þ


@z
In undersea systems, the differential changes in the
polarizationstates in

differentwavelengthchannelsaremoreimportantthanthetimespreadthat
occurs inasinglewavelengthchannel.However,
PMDisdefinedintermsofthisspread.
So,wenowcalculateit.Todoso,wefirstwriteW~ðz;$Þ¼að$ÞA~ðz;$Þ,where

jA~j2¼1andaisreal.WenowdefineamatrixFðz;$Þsuchthat

@A~

i ~

þFA¼0 ð11Þ
@$
From Eqs. (10) and (11), we infer that

@F
iDb0 $s¯

@z

3 F—Fs¯

3Þ þDb0 s¯3

ð12Þ

so that the trace of F is constant as a function of z. Since the


transformation relatingA~z;
$atdifferentvaluesof$isunitary,thematrixFisHermitian. Thus, its
eigenvalues are real and its eigenvectors are orthogonal. We designate
the eigenvectors as Toff T TPMD. The eigenvectors are conventionally
referred to as the principal states, while the difference between the
eigenvalues, 2TPMD, is
referredtoasthedifferentialgroupdelay.Physically,theDGDcorrespondstot
he delay that appears between the components of light that are
launched along the principalstates.

To relate the DGD to the pulse spreading, we now define a new set
of frequency-dependent Stokes parameters

s~ 1¼A~y s3 A~;

s~ 2¼A~y s1 A~;

s~ 3¼—A~ys2 A~

ð13Þ

whichinturnallowsustodefineanewunitStokesvectors~s~1 ;s~ 2 ;s~3inthe


frequency domain and hence a new Poincare´ sphere. We have used a to
distinguishs~ fromthetime-
domainquantitiesSorsthatwedefinedearlierinEq.

(5) et seq., but we stress that s~is not the Fourier transform of either S
or s. The
relationshipbetweens~andeitherSorsisnotsimple,althoughs~becomese
qualto s in the case of a single frequency corresponding to $ 0. Because
all three
quantitiesarereferredtointheliteratureas‘‘S,’’thereadermustpaycareful

Ð
attention. Defining the mean signal time TðzÞ¼ —1tjWðz;tÞj

dt=
1 2 2 1 2 2
Ð Ð
—11jWðz;tÞjdtandthemeansquaresignaltimeTðzÞ ¼ —1tjWðz;tÞj

dt=
2
Ð
—1jWðz;tÞjdt ,2we may define the squared signal spread SðzÞ

Σ Σ
¼T2 ðzÞ— TðzÞ .ByanalogywiththeStokesvector,itisconventionaltowrite

FðzÞ ¼ T

þ ðO s

þOs

—OsÞ ð14Þ

which defines a vector V O1; O2; O3 , whose magnitude equals the DGD
and
thatisreferredtoasthepolarizationdispersionvector.Thegeneralexpressio
nfor the spreading due to PMD is complex [5], but when the variation of
O and s~can
beignoredoverthebandwidthofthesignal,onefindsthat[5,34,35]

S2ðzÞ— S2ðz ¼ 0

1
Vðz;oÞ ~s~ðz;oÞj2 ð15Þ

To determine the length scale on which spreading occurs, we may first


write

s¯3
cosys sinysexpðifs Þ
¼ 16

whichdefinestheanglesysandfs .CalculatingO2 ðzÞato¼o0 ,wefind

O2ðzÞ¼ 8 ð dz

z1

fDb0 ðzÞcos½yðzÞ]g dz

fDb0 ðzÞcos½yðzÞ]g ð17Þ

To make further progress, one must know the autocorrelation function


Cz1 ;z2 Db0z1cosysz1Db0z2cosysz2,whereindicatestheaver- age over an
ensemble of fibers. A wide variety of fiber models all imply [4, 5, 9]
thatCðz1 ;z2 Þ¼1h½Db0 ðzÞ]2 iexpð—jz1—z2 j=hfiberÞ.Usingthisautocorrelation

functioninEq.(17)andnotingthathO1 i ¼hO2 i ¼hO3 i,weconclude

hO2 i¼8h½Db0 ðzÞ]2 ifhfiberzþh2

½expð—z=hfiberÞ —1]g ð18Þ

which is Poole’s classic result [4, 36].

To determine the effects of randomly varying birefringence on a full


time- domain signal, one must solve Eq. (2) using a model for the
variation of DB z . Wai and Menyuk [9] studied two different physical
models of the random variation, one of which took both Db cos y and
Db sin y to be Maxwellian
distributedandtheotherofwhichheldDbfixedandwithyuniformlydistribut
ed. Both models yield nearly identical statistical behavior for the
evolution of the
field’spolarizationproperties.However,bothofthesemodels—
referredtoasfine step models—require resolving the y variations on a
length scale that is small compared to the fiber autocorrelation length.
These step sizes are too small to be useful. In practice, one must take
step sizes that are on the order of kilometers. Marcuse et al. [10]
showed that a coarse step procedure in which one fixes the
birefringenceforalengththatislongcomparedtothemaximumdiffusionleng
th on the Poincare´ sphere and then scrambles the polarization
uniformly on the
sphereforallfrequencieswillyieldthecorrectPMDstatisticsifoneartificially

lowersDb0byafactorð2hfiber =zÞ1=2,wherezisthestepsize.Theyalsoshowed
that this approach does not yield the correct statistics for the nonlinear
fluctua- tions, but these fluctuations are too small to matter in
communication systems to date.

In Section III of this chapter, we describe a reduced model that just


follows one set of Stokes parameters for the signal and the noise in
each of the channels. To
determinetheeffectofPMDoneachsetofStokesparameters,onemustadapt
the coarse step procedure that we just described in the preceding
paragraph. We begin bydefining

Uz;t UðmÞz;texpibðmÞ z ioðmÞ t 19


m¼1

whereUðmÞisthewaveenvelopeofthemthchannelandbðmÞandoðmÞareits
corresponding wavenumber and frequency with b o0 and o0 subtracted,
respectively. We now define the average Stokes parameters for each
channel, writing
ðmÞ
S ¼
t2

h
jUðmÞ ðtÞj2þjUðmÞ ðtÞj2 idt

T t1
t2

ðmÞ
ðh S ¼
jUðmÞ ðtÞj2—jUðmÞ ðtÞj2 idt

T t1
t2

ðmÞ
S ¼

ð h
Re UðmÞ ðtÞUðmÞðtÞidt

1 2

t1

t2
2

ðmÞ
S ¼

ð h
Im UðmÞ ðtÞUðmÞðtÞidt

where T t2 t1 is assumed to be large enough that the channel becomes


statistically stationary. We next define the average Stokes vector of the
mth channelasSðmÞ¼ðS ðmÞ ;S ðmÞ ;S ðmÞ Þ,andwedenoteitsmagnitudebyS ðmÞ.The

degree of polarization of the mth channel

ðmÞ ðmÞ ðmÞ


d ¼S =S
ð21Þ

isbetween0and1.To
applythecoarsestepmethodinanopticalfiberwithPMD, we proceed by
first noting that PMD induces no change in the total power so that
S ðmÞ ðzþzÞ¼S ðmÞ ðzÞ,wherezisthestepsize.Wealsofindthat[16,17],
0 0

SðmÞ ðzþzÞ ¼MðmÞ ðzÞMðmÞ ðzÞSðmÞ ðzÞ ð22Þ

where the subscript j indicates the jth z step in the algorithm. The matrix

ðmÞ
M ¼

1 0 0

0cosðDb0 oðmÞ zÞ —sinðDb0 oðmÞ zÞ

@
0 sinðDb0 oðmÞ zÞ cosðDb0 oðmÞ zÞA

ð23Þ
accounts for the channel-dependent rotation due to the fiber
birefringence and differs for each channel but is the same at each step.
The matrix

cosy sin y cosc

0
j j j
ðmÞ
M ¼

B —
@
sinyjcosfj cosyjcosfjcoscj—sinfjsincj

sinyjsinfj cosyjsinfjcoscjþcosfjsincj

sin y sin c

1
— j j

C
cosyjcosfjsincj—sinfjcoscj A ð24Þ
cosyjsinfjsincjþcosfjcoscj

induces the random rotation at the end of each step that is required by
the coarse step method. We note that it is the same as the well-known
Euler angle rotation
matrix[37].Itisthesameforeachwavelengthchannelbutdiffersateachstep.
At

eachstep,thecosyjarechosenrandomlyfromauniformdistributionbetwee
n—1 and1,whilethefjandcjarechosenrandomlyfromauniformdistribution
between0and2p.ThequantityDb0thatonemustuseinEq.(23)isrelatedtothe
measuredPMDasfollows:First,wenotethatthePMDisdefinedashOi=z1=2 ,

and,assupmingaMaxwelliandistributionoftheDGD,hO2 i ¼ð3p=8ÞhOi2 ,sothat


Db0¼ ðffi3ffiffipffiffi=8Þð1=hfiberÞ
1=2

hOi=z
1=2

fromEq.(18).RecallingthatDb0

must be
reducedinthecoarsestepmethodbythefactorð2hfiber =zÞ1=2,weuse
p 1=2

Db0¼ ðffi3ffiffipffiffi=8Þð2=zÞ

Polarization-Dependent Loss andGain


PMD ð25Þ
In addition to PMD, polarization effects due to polarization-dependent
loss and gain play a major role in undersea systems. They interact with
the PMD in a complex way, and it is not possible to accurately treat any
of these effects in isolation from the others. PDL and PDG are
contributed by the amplifiers, in contrast to PMD, which is mostly
contributed by the optical fiber transmission line. Amplifiers in optical
communication systems typically operate with gain saturation and=or
active gain control elements in order to keep the total output power
nearly constant after every amplifier stage. At the same time, amplifiers
havepolarization-
sensitiveelementslikeisolatorsandWDMcouplersthatinduce polarization-
dependent loss. We are not concerned with the overall polarization-
independentlossinamplifiersbecausewecanassumethattheamplifiersleav
ethe overall gain constant. Thus, in the Jones representation for each
channel m, we maymodeltheeffectofthePDLas

ð
!
U1
! ð
!
¼1 0 U1
ð26Þ
ðmÞ
U
after

a UðmÞ
before

where the second component is in the direction of maximum loss, and a


is related to xPDL, with PDL in decibels, through the relationship xPDL ¼ —
20 log10 a.
Typical values of xPDL in undersea systems are on the order of 0.1 dBper
amplifier,anditisimportanttokeepthisvaluelow[15].ConvertingtotheStok
es representation using Eq. (20), weobtain
ðmÞ

þa2

ðmÞ

1—a2

ðmÞ

S0;after ¼
S0;beforeþ
S1;before
2
ðmÞ

1—a2

ðmÞ

1 þa2

ðmÞ

ð27Þ
S1;after ¼
S0;beforeþ
2
S1;before
2
ðS2

3 after

¼aðS2
m

3 before

where we recall that the Stokes parameters are averaged over time.

PDG is due to polarization hole burning induced by the incoming


signal to an EDFA. The incoming signal saturates the amplifier, lowering
its gain. While the gain in both the signal’s polarization and in the
polarization orthogonal to the signal are lowered, the gain in the signal’s
polarization is reduced slightly more. Thus, the gain in the polarization
orthogonal to the incoming signal is slightly larger than the gain in the
polarization of the incoming signal. The amount of PDG in a single
amplifier is only about 0.07 dB for an EDFA with 3 dB of gain
compression, that is, an amplifier in which the gain is reduced by a
factor of 2by gain saturation relative to the small signal gain. The PDG
becomes larger as the
amplifiergoesdeeperintogaincompression.Themagnitudeofthepolarizati
on

hole burning is proportional to the degree of polarization dpol of the


incoming signal. We may model PDG much like PDL, except that the
direction of
maximumgainisdeterminedbytheexistingsignalinagivensystem.Thus,

writing

ðtotalÞ
S ¼

P
m ¼1

ðmÞ
S ; SðtotalÞ¼

P
m ¼1

SðmÞ ð28Þ
we find that the total degree of polarization is d

pol

¼jSðtotalÞ j=S ðtotalÞandthetotal

stateofpolarizationiss¼SðtotalÞ =jSðtotalÞ j.Wenowwrite

ð
U1

!
ðtÞ
! ð
!
10 —1 U1 ðtÞ
¼R R
ð29Þ

UðmÞ ðtÞ

0g
after

UðmÞ ðtÞ
before

where g is the polarization-dependent gain, normalized to the gain in


the polarization state of the input signal. The value of g is related to
xPDG, with PDG measured in decibels, through the relationship xPDGdpol
20 log10 g [12]. The rotation matrix R is determined by the overall
polarization state of the incoming light since it is this polarization state
that determines the polarization state with reduced gain due to PDG,
while R—1is the inverse of R. The elements of R are related to s through
the relationships s1 ¼ jr11j2 — jr12j2and
s2 is3 2r11r* . Transforming once again from the Jones
representation to the Stokes representation, weobtain
ðmÞ

g2—1

ðmÞ

g2þ1

ðmÞ

ðg—1Þ

ðmÞ

Safter ¼ —
S0;before sþ
2
Sbeforeþ
2
s~ðs~SbeforeÞ
2
ðmÞ

g2þ1

ðmÞ

g2—1
ðmÞ

S0;after ¼
S0;before—
2
s·Sbefore ð30Þ
2

wheresinEq.
(30)istheunitvectorofthetotalpolarizationstatebeforethelight passes
through the PDGelement.

In systems with a small number of channels, PDG is a pernicious


effect because it leads to excess noise and channel outage [11–13]. We
can reduce this effect by scrambling the polarization state of the
incoming signal to make the degree of polarization, dpol as small as
possible. However, PDL tends to

repolarize polarization-scrambled signals, which then become


susceptible to

PDG. To study the repolarization due to PDL, we consider a simple


example in which there is only a single channel. In the case, it is
possible to calculate the evolution of the probability distribution
function of the degree of polarization, f dpol , analytically using the
methods of stochastic differential equations.These

methodshavebecomeanimportanttoolintheanalysisofopticalfiberswith

randomly varying birefringence since their introduction by Gisin [38]


and by
FoschiniandPoole[39].Theyareoftenconsidereddifficult,butinourviewthe
y are simply an extension of the standard tools of calculus. The currently
available textbooks on this subject are written at the level of rigor that
is common among mathematicians, which makes them difficult to
penetrate for some applied physicists and engineers and may account
for some of the perceived difficulty.
WaiandMenyuk[9]havegivenanelementaryderivationofthebasicapproac
hat thelevelofrigorcommonamongappliedphysicistsandengineers.
There are two versions of stochastic differential calculus that apply
in
differentphysicalcontexts[40].Bothcontextsappearinopticalfibertransmis
sion problems. The first context is one in which the source of the
randomness is varying continually along with the dynamic variables. An
example is the evolution of the polarization dispersion vector in an
optical fiber, the problem that was considered by Gisin [38] and by
Foschini and Poole [39]. In this case, Stratonovich calculus is
appropriate. The second context is when the dynamic evolution is
physically separated from the randomization. In the example thatwe

consider here, in which we study the evolution of dpol, the


randomization of the Stokes parameters, which occurs in the optical
fibers, is separate from the
evolutionofdpol ,whichoccursintheamplifiers.Inthiscase,Itoˆcalculusis
appropriate.
We start by combining Eq. (24) and Eq. (27) to obtain
S0; jþ1¼

S1; jþ1¼

1 a2

S0; jþ
2

1 a2

S0; jþ
2

1 a2

Spol; j cosyj
2

a2
Spol; j cosyj
2

ð31Þ

ðS2þiS3 Þjþ1¼aSpol;jexpðifjÞ

where the cos yj are independent and identically distributed (i.i.d.)


random variablesuniformlydistributedintherange½—
1;1]sothathcos yji¼1=3.
2

Similarly,thefjarei.i.d.randomvariablesthatareindependentofthecosyjand
areuniformlydistributedintherange½0;2p].Thegoalistocalculate f dpol; j ,
where dpol; j Spol; j =S0; j . Only the ratio dpol; j is meaningful since the
difference equations, Eqs. (31), do not take into account the
polarization- independent gain and loss. We will consider the initial
condition dpol;0 0, corresponding to a polarization-scrambled channel.
We first notethat

S2
0; jþ1
Spol;jþ1
¼ a ðS0; j

Spol;j
Þ¼a2ðjþ1Þ ,whichmotivatesonetoreplace

S0; j and Spol; j with xj ¼ S0; j =a jand yj ¼ Spol; j =aj, which satisfy x2— y2¼ 1.

j j

One then obtains the difference equation

1 þ a2

1 — a22

1=2

xjþ1 ¼

xj þ
2a
ðxj—1Þ
2a
cosyj ð32Þ

Since a is small, this difference equation can be approximated by the


stochastic differential equation

dx 2

¼ rx þ ðx
dj

ð33Þ

wherejisnowtreatedasacontinuousvariable,r¼ð1—aÞ2 =2aand

s2¼ð1—a2 Þ2 =12a2 .
We note that Eq. (32) is a forward difference equation, which
follows
physicallyfromtheseparationoftherandomvariationofcosyjandfjinthe
opticalfibersandxintheamplifiers.Consequently,Eq.
(33)shouldbeinterpreted
inthesenseofItoˆ,whichimpliesthattheevolutionoftheprobabilitydistributio
n functionofx;fx ðxÞ,isgovernedbytheFokker–Planckequation[40]

@f @

s2@2

x g 2

Itisusefulnowtochangevariablesfromxtog,wherex¼coshg.Itfollowsthat
dpol tanh g. One then finds that fg g fx x g dx=dg fx x g = sinh g is
gov- erned by the Fokker–Planckequation

@f

þ r—

Σ
1 2 @1
s f

1
— s2g¼0 ð35Þ
which has the solution
2

f ðgÞ ¼

!
g2

ðr=sgÞ—ð1=2Þ

g2

exp —
1=2
ð2s2jÞ Gðb=s2Þ

2s2 j

2s2 j
g g

4 g2

g g

!
g2

’ exp—
p1=2ð2s2jÞ3=2 2s2j

ð36Þ
where G indicates the usual Gamma function. Using f dpol fg g dg=d dpol ,

one may compare Eq. (36) to a direct solution of original difference


equations, Eq. (32). We show this comparison in Fig. 4, where Eq. (32)
was solved with106different realizations. We see that the analytical
theory yields results that are indistinguishable from those yielded by the
original difference equations. The original difference equations do not
have an analytical solution and so could not
haveyieldedtheanalyticalresultdirectly.Thus,ourresultsshowthepoweroft
he methods of stochastic differentialequations.

A thorough discussion of the impact that repolarization has on


undersea systems is given in Section III. However, some simple
conclusions may already be drawn from the simple results that we have
obtained here. Assuming that repolarization of 15% is acceptable, and
allowing an outage probability of

6 <

~
which is about a quarter of the best current value [15]. While this sort of

reduction of xPDL is difficult to obtain, it is worth striving for.

Comments on Notation andNomenclature


As Gordon and Kogelnik [21] have pointed out, there is little uniformity
of notation and nomenclature among researchers studying polarization
effects in optical fibers. Indeed, the same authors (including us) will
switch notation and
nomenclaturefrompapertopaper.Sinceithasbeenaverred[41]thatthislack
of
uniformityisanimpedimenttoatleastexperimentalprogress,somediscussi
onof
ourownnotationandnomenclature,aswellasitsrelationshiptothenotation
and
FIGURE4Comparisonofthedistributionfðdpol ÞobtainedbyMonteCarlosimulationofthe original
difference equations to the theoretically calculated function. The parameters are xPDL ¼ 0:1
dB. (a) j ¼ 100. (b) j ¼ 300. The two approaches yield indistinguishableresults.
nomenclatureofothers,isappropriate.First,asnotedpreviously,weusenega
tive carrier frequencies. By contrast, Gordon and Kogelnik [21] use
positive carrier frequencies, as do Poole and Giles [42]. As noted
previously, our choice of
negativecarrierfrequenciesisdictatedbyadesiretobeconsistentwithnearly
all the theoretical literature that is focused on dispersive and nonlinear
effects in
opticalfibers.ThischoiceisconsistentwithAgrawal[7]aswellaswithBornand
Wolf [22]. We find that if we define the Stokes parameters in the
traditionalway,

as done in this chapter, then s3 ¼ þ1 corresponds to right circular


polarization, ands3 1 corresponds to left circular polarization. For
clarity, we note that
rightcircularpolarizationimpliesthattheelectricfieldvectorofthelightappe
ars to rotate in a clockwise fashion when heading toward the observer.
However, if one uses positive carrier frequencies and defines the Stokes
parameters in the traditional way,thens3 1 corresponds to
right circular polarization. Poole
madethischoiceimplicitlysinceheusesthetraditionalStokesparameters.Th
ere
issignificanthistoricalprecedentforthischoice,whichcoincideswiththechoi
ce
ofShurcliffandBallard[43],astandardreferenceonpolarizedlight.Bycontras
t, Gordon and Kogelnik [21] redefine s3 sothats3 1 corresponds to
right
circularpolarization.Thischoiceappearstobeunprecedented.Wenotethat
most

oftheliteratureonpolarizationeffectsinopticalfibersisindefiniteaboutboth
the definition of the Stokes parameters and the sign of the carrier
frequencies. For most applications, it does not matter; however, it can
be quite important when
makingcarefulcomparisonsbetweentheoryandexperiment[41].

In another notational innovation, Gordon and Kogelnik redefined


the standard Pauli matrices. Doing so simplifies the transformation
between the
StokesandJonesvectors,allowingthemtoreplacethelastthreecomponentsi
n
our Eq. (5) with the expression sj Uysj U. We have not adopted this
notation in

thischapter,anditisourviewthatinmostcasesitisnotworththeconfusionthat

it might lead to, given the long history of the Pauli representation. In
most applications, one picks either the Stokes or the Jones
representation, and it is not necessary to do much transformation back
and forth. However, this notation is a real computational convenience if
one is making many transformations between the two representations.

Finally, we turn to the question of PMD-related notation and


nomenclature. We write the polarization dispersion vector as V. This
choice is consistent with Poole [4], Gisin [38], and most of the rest of
the literature in this field. By contrast, Gordon and Kogelnik use t. We
have used V in this chapter in order to conform with what appears to
be the consensus and to avoid adding to the
notationalconfusion.Thereisnoconsensusastowhatthisvectoriscalled.Poo
le
hashistoricallyreferredtoitasthe‘‘polarizationdispersionvector’’;Gordona
nd Kogelnik refer to it as the ‘‘PMD vector.’’ We have used Poole’s
nomenclature. Since he first described the concept, there is some
precedent for following his naming convention. Moreover, we dislike
the term ‘‘PMD vector’’ because, as noted earlier, PMD is often
confused with DGD. The magnitude of the polariza- tion dispersion
vector has units of time and is a DGD, not a PMD. In closing,we
notethatour
polarizationdispersionvector,likethatofGordonandKogelnikbut
incontrasttothatofPoole,pointstowardtheslowaxis.

REDUCED STOKES PARAMETERMODEL

ModelFormulation
In Sections II.B and II.C, we obtained equations that govern the
evolution of the Stokes parameters in an optical fiber transmission
system with PMD, PDL, and PDG. In this section, we develop these
equations into a system model that will enable us to predict the
penalties due to these effects.

In addition to modeling the PMD, PDL, and PDG, we must account


for the effects of amplified spontaneous emission (ASE) noise. We begin
by introducing
anewsetofStokesparametersforthenoiseðS ðmÞ
m

noise

Þ.Wemusttrackthese

parameters separately from the signal Stokes parameters so that we can


compute

theSNRandultimatelyaQ-
factorforeachwavelengthchannel.BecausetheASE
noiseisunpolarized,eachamplifierwillcausethefollowingchangeintheStok
es parameters,
m

0;noise;after

noise;after

0;noise;before

noise;before

þ2nsp

ðG—1ÞBðmÞ hn

ð37Þ

where nsp is the spontaneous emission factor, G is the amplifier gain, hn


is the
energyofasinglephoton,andBðmÞisthebandwidthofthemthchannel.These
Stokes parameters are affected by the PMD, PDL, and PDG in exactly the
same

way as the signal Stokes parameters and participate in determining the


degree of polarization and the total Stokes parameters. Additionally, if
any part of the gain bandwidth of the EDFA is not included in one of the
optical channels, then this noise energy will participate in the total
energy balance. We may write for this additional portion,
add 0;noise;after

add 0;noise;before

þ2nsp

ðG—1ÞBðaddÞ hn ð38Þ

and we assume that this contribution is unpolarized. This portion is not


included in the example WDM systems that we will present later in this
chapter, but it could be present in some practical systems. We now
write
ðtotalÞ

n n

ðmÞ

ðaddÞ
P P
S0 ¼ S0 þ S0;noise þS0;noise

SðtotalÞ ¼

m¼1
P
m ¼1 n

SðmÞ þ

m¼1
P
m ¼1 n

noise

ð39Þ
and the degree of polarization becomes d

pol

¼jSðtotalÞ j=S ðtotalÞ .Thefinalstepisto

take into account the effect of gain saturation or gain clamping by


assuming that
the total power at the output of the amplifier is fixed at a value S. We
then
renormalizeS ðmÞ ;SðmÞ ;S ðmÞ
ðmÞ
;S
,andS ðaddÞbythefactorS=S ðtotalÞ.

We summarize the complete procedure schematically in Fig. 5. This


procedure is repeated iteratively from amplifier to amplifier.

FromthecalculatedsignalandnoiseStokesparameters,itispossibletode
termineQðmÞ —theso-calledQ-factor—foreachchannelmandfromthatto
infer the penalty due to PDL and PDG in combination withPMD.
Tocalculatethispenalty,wefirstnotethataQðmÞthatwecalculatefromthismode
lisnot
meaningfulbyitselfbecausethemodeldoesnottakeintoaccountdegradatio
ndue to nonlinearity and chromatic dispersion. Whatis meaningful
isthedifferenceDQðmÞbetweentheQðmÞvaluesthatwecalculatewhenPDLand
PDGarepresentandwhentheyareabsentforaspecificrealizationoffiberPM
D.TocalculateQðmÞforaparticularchoiceofPMD,PDL,andPDG,wemust
obtaintheeffectivesignal-to-
noiseratio(ESNR)ofchannelmafterdetectionin

thereceiver.Theopticalsignal-to-noiseratio(OSNR)equalsS ðmÞ =S
0;noise

, but

thereisnosimple,universalrelationshipbetweentheESNRandtheOSNR.It

depends on the modulation format, the signal distortion during


transmission, and the details of the receiver. In the receiver, the signal
will typically pass through a
photodiodedetector,anelectricalfilter,andatime-
domainsamplerwithanarrow
FIGURE 5 Schematic illustration of the modeling procedure.
window.Thereisoftenanonlinearlimiterinthereceiveraswell.Thesefactorsc
ombinetodeterminethereceiverenhancementfactorZthatrelatestheESN
RandtheOSNRforchannelmthroughtherelationshipESNRðmÞ ZOSNRðmÞ . For
standard NRZ transmission, it is usual to approximateZ 2.
ThisvalueisexactforidealsquareNRZpulsesandidealintegrate-and-
dumpreceivers[44,45].Inthefullsimulationsthatwecomparedtothereduce
dmodel,weusedanidealsquarelawdetectortomodelthephotodiode,weus
eda10th-
orderBesselfilterwhosebandwidthequaledthedatarate(5or10GHz),andw
eassumethatthere is an ideal instantaneous sampler in the center of
thetimingwindow.Becauseitisonlypossibletokeepalimitednumberofbitsin
thefullsimulations,thereistoomuchstatisticalvariationifwenumericallyim
plementanidealsampler to obtain accurate results. So, instead we
numerically
calculatetheSNRaftertheelectricalBesselfilter,andwemultiplythatbythee
nhancementfactorZ.

In the work reported in this chapter, we approximated Z Ppeak=Pave,


the ratio of the peak power to the average power in the optical signal
channel just prior to the receiver. For the standard nonreturn-to-
zero (NRZ) modulation

format, this ratio is 2; for the standard return-to-zero (RZ) modulation


format, this ratio is 4; for the chirped return-to-zero (CRZ) of Bergano et
al. [46], in
whichthepulsesreachtheirminimumdurationjustpriortodetection,thisrati
ois approximately 5.3. We recently used noise-free simulations to find
that theactual efficiency factor for the RZ modulation format and our
electrical filter is in the range Z 3:2–3.4, depending on the propagation
distance. We note that in
commercialsystems,wherethebandwidthoftheelectricalfilteris70–
80%ofthe
datarate,wewouldexpectZtobesomewhatlower.Wealsoexpectthattheact
ual enhancement factor for the CRZ format is somewhat less than 5.3.
Because we
ðmÞ
arecalculatingDQ indecibels,theresultsareinsensitivetothechoiceofZ.

We may now use a formula relating the Q-factor to the ESNR,


assuming that the noise is Gaussian distributed [44, 45],
QðmÞ ¼

ESNRðmÞ

pffi ffiffi ffiffiffi ffiffi ffiffiffiffi ffiffiffi ffi ffiffiffi ffiffi ffiffiffiffiffi ffiffi


2 E S N R ðm Þ þ 1
sffi ffiffi ffiffiffi ffi ffiffi ffi
2 B op t

ð40Þ

where Bopt is the optical bandwidth and Belec is the electrical bandwidth.
We are
assumingthatthereceiverisinsensitivetopolarizationandthatthereisnosign
al power in the spaces, so that there is an infinite extinction ratio.
Physically, the electrical detector at the end of the transmission line
receives 2Bopt=Belec noise modes.Therefore,thesignal–
spontaneousbeatnoisepowerS0;sig—sponisgivenby

ðPpeak=PaveÞ

1=2

ðS0 S0;noise Þ

1=2

ðBelec =2Bopt Þ

1=2

, while the spontaneous–spontaneous

beat noise power S0;spon—spon just equals S0;noise. The noise power in the marks is
givenbyS0;sig—sponþS0;spon—spon,whilethenoisepowerinthespacesisjustgiven
byS0;spon—spon.
TheoreticalValidation
Separability of PolarizationPenalties
In order for the penalties due to PDL and PDG to be separable from
the penalties due to nonlinearity and chromatic dispersion, the
nonlinearity and chromatic dispersion must not be allowed to affect the
degree of polarization. In this subsection, we investigate the conditions
under which the impact of nonlinearity and chromatic dispersion on the
degree of polarization of individual channels can safely be ignored.

Our starting point is the Manakov equation, written in the form [9,
10]
@U 1

00@ U
2
8 2

i — b
@z 2

þ gjUjU¼0 ð41Þ
@t2 9
This equation may be derived from Eq. (2) by averaging over the rapidly
and randomly varying birefringence and neglecting all fluctuating terms.
Physically, Eq. (41) holds in the limit of low PMD. We also neglect the
polarization-
independentgainandlosssinceithasnoeffectonourresultsexcepttoeffectiv
ely renormalize the distance over which nonlinearity acts [16, 19, 20].
We will also assume that the dispersion between channels is large since
we anticipate that the effect of nonlinearity and chromatic dispersion
on the degree of polarization of individual channels will be negligible in
thislimit.

SubstitutingEq.(19)intoEq.(41),weobtain

@UðmÞ

100@2 UðmÞ 8

n
m 2 m 8
q 2 m

i — b
@z 2

þ gjUð
@t2 9
Þ ð
jU
Þ
g
q¼1;6¼m

jUðÞ jUð
Þ
¼0 ð42Þ

where we have neglected the four-wave mixing terms, consistent with


our assumption that the dispersion between the channels is large. We
now find that dS0=dz ¼ 0 and that

ðmÞ
dS

8 g t2

¼i

.
ðmÞ ðmÞ* ðmÞ* ðmÞ P
½U U þU U ]

ðqÞ ðqÞ* ðqÞ* ðqÞ


½U U —U U ]

—½UðmÞ UðmÞ—U U iP
Σ
ðqÞ ðqÞ* ðqÞ* ðqÞ
½U U þU U ] dt ð43Þ
where we have used the definitions of the Stokes parameters in Eq.
(20). In a highly dispersive system, the channels with q m rapidly pass
through channel m in the time domain. Consequently, the evolution of
the mth channel is only
affectedbytheaveragevariationintheotherchannels.So,wecaneffectivelyt
reat
theseotherchannelsascontinuouswaves.Wethusmakethesubstitution

ðqÞ ðqÞ* ðqÞ* ðqÞ


U U —U U !

ðqÞ ðqÞ* ðqÞ* ðqÞ


½U U —U U ]dt ð44Þ
from which we conclude

ðmÞ n
dS 8
1

ðmÞ

ðqÞ

ðmÞ

ðqÞ

P
¼ g ½S2S3—S3S2] ð45Þ
dz 9
WecanfindsimilarexpressionsfordS ðmÞ =dz,anddS ðmÞ =dz,sothatwefinally

obtain
2 3

dSðmÞ 8 m
n
q

¼ gSð
dz 9
Þ P
~ SðÞ ð46Þ

The effect of dispersion does not appear in Eq. (46); only the
nonlinearity appears, and the equations are analogous to the equations
that govern nonlinear rotation of continuous-wave beams [47].
However, the large local dispersion is critical because it must be large
enough so that each channel appears as a continuous-wave background
to its neighbors. It is an immediate consequenceof Eq. (46), referred to
as the mean field model, that the Stokes parameters of a single-channel
system do not evolve. Moreover, regardless of the number of channels,
the polarization of each channel simply rotates, so that the degree of
polarization is notchanged.

Although the mean field model is nonlinear, a complete analytical


solution can be found. This result is intrinsically significant because the
number of large- dimensional nonlinear systems for which exact
solutions can be found is limited. However,
theformissomewhatcumbersomeandisnotpresentedhere.Itmaybe found
elsewhere [16,20].

The mean field model has been validated by simulating NRZ signal
transmission with dispersion management [16, 19, 20]. The NRZ signal
was polarization scrambled using synchronous phase modulation, as
described by Bergano and Davidson [48]. Polarization scrambling of the
optical carrier is achieved by differential modulation of the optical
phases of two polarization
ðmÞ ðmÞ
stateswithasinusoidalsignal,U ðtÞ¼AðtÞexp½ifðtÞ]andU ðtÞ ¼

A2 ðtÞexp½if2 ðtÞ],wheref1 ðtÞ ¼d1þa1cosðoph tþc1þp=2Þandf2 ðtÞ ¼d2þa2cosðoph


tþc2þp=2Þ.Here,oneletsA1 ðtÞ ¼c1 HðtÞand A2 t c2H t , where c1 and c2 are
constant coefficients. One sets H t 1 in the time slots of the marks and H t
0 in the time slotsof the spaces, except

when making a transition from a space to a mark or a mark to a space,


in which case the transition is smoothed over 5% of the pulse rise and
fall times using a hyperbolic tangent function. Choosing a1 ¼ 3:307 and
a2 ¼ 0:903, one findsthat the difference nearly equals j0;1, the first zero
of the zeroth Bessel function. With this choice and setting c1 ¼ c2, an
ideal square pulse is depolarized. The sum a1 þ a2 was chosen to be
consistent with Bergano and Davidson [48]. The phase modulation
frequency oph corresponds to the bit rate, c1 and c2 describe the
relativephasesbetweenthephasemodulationandthedatabits,andd1 andd2

denotearbitraryoffsets.Byvaryingc1;c2;d1;d2;c1,andc2,onecanadjust

the initial degree of polarization to any desired value.


In Fig. 6, we show the evolution of the Stokes parameters of two
channels
spaced1nmapart.InFig.6a,weshowthepredictionofthemeanfieldmodel.In
Fig. 6b, we show the evolution in a dispersion map that uses a span of
normal dispersion fiber at D1 ¼ —2 ps=nm-km, followed by a span of
anomalous dispersionfiberatD217ps=nm-
km,wherethedispersionsareatacentral

wavelength of 1.58 mm. The third-order dispersion was 0.07 ps=nm2-


km in both spans, from which the dispersion at any particular
wavelength could be deter- mined. The length of the map was 1000 km,
the bit rate was 5 Gb=s, and the power in each channel was
approximately 0.4 mW. In Fig. 6c, the dispersions were multiplied by 10.

Comparing Figs. 6a and 6b, one finds that visible quantitative


differences existbetweenthepredictionsoftheManakovmodel,Eq.
(41),andthemeanfield model, Eq. (46), that only disappear when the
dispersion becomes quite large as
showninFig.6c.Nonetheless,theStokesparametersstilloscillatearoundthei
r
FIGURE 6 Evolution of the Stokes vector components as a function of distance in a 5-Gbps
system. The dispersion map length is 1000 km, and the channel spacing is 0.5nm. The solid lines are
the Stokes components of channel 1; the dashed lines are the Stokes components of channel 2. (a)
Stokes model result. (b) Manakov model result, D1 ¼ —2 ps=nm-km, D2 ¼ 17 ps=nm-km.(c)

Manakovmodelresult,D1¼—20ps=nm=km,D2¼170ps=nm-km.Othersimulationparameters

arel¼1550nmforchannel1,l¼1550:5nmforchannel2;c1¼0andc2¼0:7pforchannel1,
c1¼0andc2¼0:7pforchannel2;thepeakpowerinthe1-polarizationis0.24mWforchannel1
and0.2mWforchannel2;thepeakpowerinthe2-polarizationis0.2mWforchannel1and0.24mW for
channel2.
initial values in Fig. 6b, although with somewhat different frequencies
and amplitudesthaninFig.6a.Therearenolong-termdriftsintheStokes
parameters
fromthepredictionsoftheManakovmodel.Thus,wewouldanticipatethatth
ere is little change in the degree of polarization, and this prediction is
borne out in Fig. 7, where we show dpol for each channel over 10,000
km. The change is only
about0.02.Inparticular,onefindsthatifdpol¼0initiallyforbothchannels,

which is obtained by setting A1 A2, then the channels undergo little


repolar- ization.

Wang [16] and Wang and Menyuk [20] carried out extensive
parameter studies to determine the limits of validity of the mean field
model. They found that as they increase the number of channels, the
predictions of the mean field model agree better with the Manakov
model because the presence of multiple channels leads to better
averaging over the different channels. As the data rate increases, the
predictions of the mean field model again agree better with the
Manakov model assuming that the channel spacing scales
proportionately. As noted earlier, adding realistic polarization-
independent gain and loss makes no difference because it merely
rescales the equations. One also finds that adding amplitude
modulation so that the equations are RZ rather than NRZ makes no
significantdifference.Whenthechannelspacingincreases,thepredictionso
fthe mean field model agree better with the Manakov model and when
the channel
spacingdecreases,thepredictionsdeteriorate.Atachannelspacingofabout

0.3nm for the 5-Gbps system that Wang and Menyuk considered, the
predictions become unacceptably poor. Similarly, reducing the map
length leads to worse averaging and deterioration of the predictions of
the mean field model. Below about 200 km, the discrepancies
becomeunacceptable.

In any real system, it is important to carry out a validation effort like


theone just described here prior to validating the full Stokes model
presented inSection

III.A by comparing it to full system simulations including PDL and PDG. In


order for the Stokes model to yield useful results, the changes in any
channel’s degree of polarization induced by nonlinearity, chromatic
dispersion, and intrachannel PMD must be negligible.

FIGURE 7 Evolution of the degree of polarization as a function of distance. Parameters are the
same as those in Fig. 6b.
Stokes ModelValidation
We now present a validation of the Stokes model described in
Section III.A.
WecomparetheStokesmodeltoafullmodelthatincludestheeffectsofPDLan
d PDGaswellasPMDforbothsingle-channelandeight-
channelWDMsystemsat a data rate of 10 Gbps per channel. In the WDM
studies, a channel spacing of 1 nm was used. All the results presented
in this subsection used an RZ format. Additional work that uses a data
rate of 5 Gbps per channel and studies the NRZ
andCRZformatsmaybefoundin[16].Theresultsaresimilar.

The full model is based on Eq. (2). These studies [16, 17] used a
periodic dispersion map that consisted of one section of a single-mode
fiber whose dispersionD1atl0¼1:55mmis16ps=nm-
kmandwhoselengthis264km,and anothersectionofdispersion-
shiftedfiberwhosedispersionD2atl01:55mm is2ps=nm-
kmandwhoselengthis33km.Inbothsections,thedispersionslope
was0.07ps=nm2 -km.Channelsforwhichll0hadpre-andpost-dispersion
compensation, split equally, to compensate for the excess dispersion. In
the WDM simulations, each channel was filtered using a 10th-order, 60-
GHz optical
Besselfilterattheendofthetransmissionline.Allsimulationsincludedsquari
ng

in the photodetector and a 10th-order, 10 GHz electrical filter. The


simulations used the standard coarse step method [9], described in
Section II.B, to include PMD, and used Eqs. (27) and (30), presented in
Section II.C, to include PDLand PDG. The simulations used standard
Monte Carlo methods to include ASE noise [16]. Each set of parameters
was studied using 20 different realizations of the ASE noise and the
fiber. However, the bit string was the same in all 20 cases in order to
avoid Q-variations due to the pattern dependences in the limited strings
of64bitsperchannelthatitwaspossibletokeepinthesimulations.

For each set of parameters, the decision level in the full model
simulations was empirically set to obtain the best OSNR. The OSNR was
computed in the timedomain,aftertheBesselfilter,bycalculatingðI1—
I0 Þ=I0 ,whereI1isthe average current in the marks and I0 is the average
current in the spaces. After determining the OSNR for each of the 20
realizations, Wang [16] and Wang and Menyuk [20] found the
corresponding Q values using Eq. (40) aftermultiplying
the OSNR by Z to obtain the ESNR. The choice of Z was the same as for
the
reducedmodel.FromtheQvalues,WangandMenyukcouldthencalculatethe

meanhDQðmÞ iandthestandarddeviationsðmÞforcomparisontotheStokes

model.Giventhelargerandomvariationofthethesignal–
spontaneousbeatnoise
fromrealizationtorealization,whichleadstosignificantvariationsinDQðmÞfrom
realization to realization, 20 realizations is not really sufficient.
Moreover, with only 64 bits per channel, significant pattern
dependences arose. The number 20 was chosen due to computational
limitations that make running a significantly larger number of cases
impractical [16, 20]. Thus, a comparison of the Stokes
modeltothefullmodelshouldbeviewedasademonstrationofconsistency,n
ota complete check of the Stokesmodel.
The Stokes model does not suffer from these computational
limitations, which is why it was developed in the first place. In the
comparisons with the full model, Wang [16] and Wang and Menyuk [20]
used 2000 realizations. The applications presented in Section III.D used
as many as 105realizations. For this reason, it is our view that the Stokes
model is at least as reliable as full simulations for determining the
combined effects of PMD, PDL, and PDG.

We first compare the full model to the Stokes model in the simple
casewhen the pulse modulation format is RZ. The pulses are the same as
in SectionIII.B.1,

except that UðmÞ ðtÞ ¼AðtÞcosðot=2þp=2Þ and UðmÞ ðtÞ ¼AðtÞ

cos opht=2 p=2 , so that the pulses are amplitude modulated but
unchirped. We show the results for DQ as a function of the PDL in Figs.
8 and 9 for a single channel system, setting the PMD 0:1 ps=km1=2and
the PDG 0:0 and 0.06 dB, respectively. The agreement between the two
models is quite good. The PDL values that were compared are 0.1, 0.2,
.. . ; 0.6 dB. We note that when sQ 1,

the expected deviation of the Q-factor from its mean in the full
simulationmodel is approximately 1= 19 0:23 because there are only 20
realizations at each value of PDL. Thus, the deviation between the full
model and the Stokes model lies within the expected statistical error of
the full model. We note that the difference between the two models is
systematic rather than random because the
fullmodelyieldedeitherhigherorlowervaluesthantheStokesmodelforboth
FIGURE 8Comparison of the signal degradation as a function of PDL in the Stokes model and in
thefullsimulationmodel,wherePMD¼0:1ps=km1=2andPDG¼0:0dB:(a)hDQi.(b)sQ .Solidlines
indicatetheStokesmodelanddashedlinesindicatetheaverageofthefullsimulationmodel.
FIGURE 9 Comparison of the signal degradation as a function of PDL in the Stokes model and in
thefullsimulationmodel,wherePMD¼0:1ps=km1=2andPDG¼0:06dB:(a)hDQi.(b)sQ.Solid lines indicate
the Stokes model and dashed lines indicate the average of the full simulation model.

DQ and sQ in every plot as we varied the PDL. This systematic deviation


is due to the use of the same fiber realizations and the same bit pattern
for all 20 realizations. Wang [16] and Wang and Menyuk [20] found that
the choice of the fiber realization is more significant than the pattern
dependences for both the RZ simulations presented here and NRZ
simulations. However, they also found that pattern dependences
become more important for CRZ simulations.

Comparing Figs. 8 and 9, it is apparent that PDG adds a substantial


penalty tothesingle-
channelsystemsalmostindependentofthePDL.WhenthePDLis

0.6dBbutthePDGis0dB,hDQiisunder2dB.Bycontrast,whenthePDGis

0.06dB,hDQiisconsistentlyabove2dBregardlessofthePDLandalmost
reaches 4 dB when the PDL is 0.6 dB. However, sQ increases only slightly
with nonzero PDG. One finds similar results with the NRZ format;
however, polarization scrambling substantially reduces the effect of the
PDG, as expected [16].

We turn next to a comparison of the Stokes and full models with an


eight- channel RZ system. We show the comparison in Fig. 10 when the
PDG is 0 dB.
ComparisontoFig.8showsthatthedegradationDQðmÞisalmostthesameas
withasinglechannel,butthesðmÞvaluesarelarger.WithPDGincluded,weshow
the comparison in Fig. 11. In contrast to the single-channel system, the
effect of PDG is negligible. Again, an NRZ system yields similar results
[16].
FIGURE 10 Comparison of the signal degradation as a function of PDL in the Stokes model and
inthefullsimulationmodel,wherePMD¼0:1ps=km1=2andPDG¼0:0dB:(a)hDQi.(b)sQ .Solid
linesindicatetheStokesmodelanddashedlinesindicatetheaverageofthefullsimulationmodel.
FIGURE 11 Comparison of the signal degradation as a function of PDL in the Stokes model and
inthefullsimulationmodel,wherePMD¼0:1ps=km1=2andPDG¼0:06dB:(a)hDQi.(b)sQ.Solid
linesindicatetheStokesmodelanddashedlinesindicatetheaverageofthefullsimulationmodel.
ExperimentalValidation
We now compare the Stokes model to experiments that were carried
out in a recirculating loop configuration described by Carter et al. [49].
The recirculating loop is a little over 100 km long. Recirculating loops are
a simple and efficient way to study long-haul transmission systems.
However, it has long been known that the behavior of recirculating
loops that are shorter than about 500 km can differ significantly from
real transmission systems. Yet, this behavior has only recently been
characterized [16]. In this section, we compare the evolution of the
degree of polarization in the experimental system to the predictions of
theStokes model. We present cases in which a single-channel 10-Gbps
pseudo-random signal is propagating and cases in which there is no
initial signal, and the light in the recirculating loop grows from ASE
noise. The work presented in this section not only serves to validate the
Stokes model, but it also gives insight into the polarization evolution in
short recirculatingloops.

In more detail, the experimental system is a dispersion-managed


recirculat- ingloopthatcontains100kmofdispersion-
shiftedfiberwithanormaldispersion of 1:1 ps=nm-km at 1551 nm and 7
km of standard fiber with an anomalous dispersion 16.7 ps=nm-km at
1551 nm. The entire loop comprises one period of
thedispersionmap.ThePMDofthefiberisbelow0.1ps=km1=2.Asingle2.8-nm
bandwidth optical filter and five EDFAs are in the loop. The polarization
evolution inside the loop was investigated using a commercial
polarization analyzer, the HP 8509B [16]. By sampling the Stokes
parameters as a function of the propagation, we can determine the
degree of polarization as a function of the propagation time or,
equivalently, distance. In Fig. 12, we show the evolution of the degree of
polarization. The different curves correspond to different values of the
BER measured at 20,000km, and we obtained these different values of
the final BER by using different settings of the polarization controllers.
The polarization evolution inside the loop is closely correlated to the
final BER. The signal is highly polarized when the BER is less than 10 —
10
at 20,000 km. When the polarization controllers are set so that the
final BER increases, the signal depolarizes increasingly withdistance.
FIGURE 12 Evolution of the degree of polarization corresponding to different BERs. (a) BER¼
10—9. (b) BER¼ 10—6. (c) BER¼10—2.
To measure the PDL in the loop, it was necessary to open the
recirculating loop into a 107-km straight-line experiment and then
measure the PDL of the entire line. Varying the polarization state of the
input signal so that it covers the entire Poincare´ sphere, one measures
the output power as a function of the output polarization state. The
difference between the maximum and minimum power equals the PDL.
In the experiments presented here, the total PDL of the loop equaled
0.35 dB.

To compare the Stokes model to the experiments, one must first


modify the model to take into account the periodicity of the loop. This
periodicity is important because the PDL contributions are no longer
random but repeat with the same period as the loop. Using the
modified model, one finds that when the PDL equals 0.45 dB, the results
of the model are in exact agreement with the experiments, as shown in
Fig. 13. The agreement between the model and the experiment is
acceptable because the error in measuring the PDL is expected to equal
approximately 0.1 dB, and the open loop did not contain the switches
and couplers that were used in the closed-loop experiments.

In Fig. 13, we also show the effect of reducing the PDL, keeping the
setting of the polarization controller in the Stokes model that yields the
lowest BER. As
thePDLbecomessmaller,thenoiseplaysanincreasinglyimportantrole,leadi
ng to an increased depolarization. When the PDL equals 0.01 dB, the
degree of polarization falls below 0.5. The repolarization of the noise
when there is no signal also becomes smaller as the PDL decreases.
When the PDLequals

0.01 dB, one finds that the degree of polarization after 27,000 km just
equals 0.2, as shown in Fig.13b.
FIGURE 13 Evolution of the degree of polarization with (a) signalþnoise and (b) noise only. The
experimental results are shown as stars. The theoretical curves correspond, in order of decreasing
degree of polarization, to PDLs of 0.45, 0.25, 0.15, 0.05, and 0.01 dB.
Applications to TransoceanicSystems
We now apply the Stokes model to the problem of
calculatingtheoutageprobabilityintransoceanicsystemsassumingasystem
marginforpolarizationeffectsofeither2.5or3.0dB.Thecalculationspresente
dhereused105realizationsforeachchoiceofparameters,and,whennecessar
ytocomputetheoutageprobability,aGaussiandistributionwasfittedtothetai
lofthenumericallydetermined probability distribution function.
Acceptable outageprobabilitiesaretypically around 10—6, corresponding
to a little more than half a minuteper
year.ThenumberofWDMchannelsintransoceanicsystemshasgrownrapidlyi
nrecentyears.WhiletheeffectofPMDonasinglechannelistypicallysmallinun
dersea systems, where the PMD is usually quite low, the PMD
doesrotatethepolarization states of the different channels with respect
to one
another.Inotherwords,thePMDchangestheangularseparationofthechann
elsonthePoincare
´sphere.AsaconsequenceoftheinteractionofthePMDandthePDL,differentc
hannels will undergo different amounts of loss when they pass
throughadevicewithPDL.Becausethegainsaturationorgainclampinginthea
mplifiersistunedtoeffectivelyrestorethetotalsignalpowerinallchannels,so
mechannelsgainpowerattheexpenseofothers.Thiseffectleadstoarandomw
alkinthepowerofeachchannelandcancauseoneormorechannelstofade.We
presentresultsherethat show this mechanism is the primary cause of
fading insystems
withmorethanapproximately10channels,incontrasttosingle-
channelsystemsinwhich

PDG is the primary cause of fading.

Wefirstconsiderasysteminwhichthechannelspacingandtheopticalfilte
r bandwidth equal 0.6 nm, with other system parameters set as
follows: PMD 0:1 ps=km1=2, PDL 0:0 dB, and PDG 0:06 dB. Figure 14
shows that as the number of channels increases, the importance of
PDG decreases as expected from the argument in the
precedingparagraph.

Next, we consider a system in which the PDG equals zero, leaving


only the
effectsofPMDandPDLinthemodel.Inthisexample,thechannelspacingis
1.0nm,andtheopticalfilterbandwidthis0.5nm.ThePMDequals0.1ps=km1=2
,
andthePDLequals0.1dBineachopticalamplifier.Byincreasingthenumberof
channels,oneobtainstheresultshowninFig.15.IfDQallowed,theallowed

FIGURE 14 The degradation and variance of the Q-factor as a function of the number
of channels.
FIGURE 15 Outage probability as a function of the number of channels. The solid line is for

DQallowed ¼ 2:5 dB; the dashed line is for DQallowed ¼ 3:0 dB.

degradation level for any single channel, is set equal to 2.5 dB, then the
outage probability dramatically increases from 6:5 10—13in the case of a
single channel to 3:0 10—4when there are many channels. With only
three channels, the outage probability already exceeds 10 —5. If we raise
DQallowed to 3.0 dB, then the maximum outage probability falls to 2:3 10
—6
, a decrease of more than 2 orders ofmagnitude.

When the amplifier spacing increases from 33 to 45 km and then to


50 km, the average value of Q decreases due to the additional ASE noise
that is addedto
thetotalsignal.However,theoutageprobabilitydecreasesbecausethenumb
erof PDL elements along the transmission line is reduced, as shown in
Fig. 16. When the number of channels is 40, the outage probability
drops from 3:0 10—4to 1:3 10—5and 2:8 10—6, respectively. So, when
one designs a WDM system
andchoosestheamplifierspacing,onehastotakeintoaccountbothnoise-
FIGURE 16Outage probability as a function of the number of channels. Amplifier spacing equals

(a) 45km and (b) 50km. The solid line is for DQallowed¼2:5dB; the dashed line is for

DQallowed ¼ 3:0 dB.


induced and polarization-induced penalties. If the PDL is the same in
each
amplifier,thenashortamplifierspacingwillintroducelessnoisebutwillincrea
se the outage probability. By contrast, a long amplifier spacing will
introduce more noisebutwilldecreasetheoutageprobability.

Figure14showsthattheeffectofPDGbecomesinsignificantwhentherea
re more than approximately 10 channels in a WDM system. To further
investigate this issue, one may add a PDG of 0.07 dB to the case shown
in Fig. 15. We show
theseresultsinFig.17.Insteadofasmalloutageprobabilitywhenthenumber
of
channelsissmall,onefindsthattheoutageprobabilitypeaksatasmallnumber
of
channelsandthendecreasestoitsfinalvalue.Thedramaticincreaseintheout
age probability when the number of channels is small is due to the
faster growth of ASE noise that is induced. The outage probability then
decreases as the number of channels becomes larger because the PMD
between the channels leads to an averaging of the polarization states so
that the degree of polarization for the total signal is nearly zero, and the
PDG leads to nearly no excess noise growth. When the number of
channels equals 40, the outage probability is 2:2 10—4, which is actually
smaller than the corresponding value of 3:0 10—4when there is no PDG.
The reason for this paradoxical decrease is that the PDG tends to
compensatefortheeffectsofPDLonchannelsthatexperienceexcessloss.

ACKNOWLEDGMENTS

Sections II.B, II.C, and all of III are based on the Ph.D. dissertation of Dr.
Ding
Wang.Oneofus(CRM)isalsogratefultoDrs.H.SunnerudandF.Bruye`refor
making their Ph.D. dissertations available to him. The insights in both
disserta- tions were useful. We are grateful for financial support from
the Air ForceOffice of Scientific Research, the Defense Advanced
Research Projects Agency, the Laboratories for Physical Sciences and
Telecommunications Sciences at the Department of Defense, the
Department of Energy, and the National Science Foundation. We are
grateful to the submarine systems group, then at AT&T Bell
Laboratories, for arranging for some early financial support for the
development
oftheStokesmodel.Inparticular,wethankPeterRungeandFrankKerfootfor

FIGURE 17 Outage probability as a function of the number of channels, where PDG¼ 0:07 dB.
The solid line is for DQallowed ¼ 2:5 dB; the dashed line is for DQallowed ¼ 3:0 dB.
encouragingthismodel’sdevelopment.Finally,wearegratefultoCienaCorpora- tion and Science
Applications International Corporation for recent support that has allowed us to better validate the
models presented here and broaden their range ofapplications.

You might also like