4.4 Bayesian Algorithms
4.4.4 Bayesian Algorithms based on double-AR1 Phase Noise
Model
All Bayesian algorithms derived until now, are arried out by onsidering a
rstorderphasenoisemodel,i.e.,adis rete-time Wienermodel.Ontheother
hand, sometimes hannels of pra ti al interest are ae ted by a phase noise
whi h anbemodeledasthesumoftworstorderauto-regressivelters.This
is the ase of the SATMODE phase noise, whi h impairs satellite
ommuni- ations hara terized bylow- ost equipment, andwhose double-AR1modelis
des ribed inSe tion 2.4.2. In Chapter 2, numeri al results show that the
de-te tor deriveduntil now, basedonjustasingle-AR1PNgeneration,exhibitsa
signi ant performan e degradation when employed on double-AR1 PN
gen-eration (see Se tion 2.4.4). Thus, it an be useful to derive some Bayesian
algorithms basedon su hdouble-AR1 PN model, asinthe aseof the
D-DP-BCJRalgorithmdes ribedinSe tion2.4.3.Se tion2.4.4,demonstratethatthe
information rate lossa hieved bythese algorithms withrespe t to the
oher-ent ase, is lose to zero. However, the good result in performan e is rea hed
at the pri e of an in reased trellis state omplexity sin e the system state is
omposedbytheunionoftheCPMstate,withthestateofboththetwophase
noise omponents.Thusthesealgorithms annotbetakenintoa ountfor
sys-temsofpra ti alinterest.FinallyinSe tion2.4.3,wealsoproposeanimproved
version of the DP algorithm (the so alled I-DP), suitable when onsidering
double-AR1 PN with a very fast omponent, whi h an be approximated as
jitter,independent fromsampleto sample.ThetrellisoftheI-DPalgorithm is
basedon thefull- omplexityCPMtrellisbasedon theRimoldide omposition
and thus is omposed by
D pM L−1
states. We an redu e the omplexity byapplying the same te hnique des ribed in Se tion 2.4.3 to the MM-DP
algo-rithm of Se tion 4.4.2, whi h is based on the MM de omposition and ounts
just
D
trellis states. The way to obtain the Improved-MM-DP (I-MM-DP) starting from MM-DP is very similar to thatused to derive the I-DPinSe -tion 2.4.3.Inparti ular, rstlywe justneed to hange themean and varian e
of the Gaussian pdf in (4.22) representing phase noise probability transition
p(θ n+1 |θ n )
, sin e the PNθ n
is no more modeled as a Wiener model but asan AR1 pro ess. Se ondly, in order to take into a ount also the se ond PN
omponent
θ f,n
(i.e., the faster omponent), we ompute the bran h metriH n (x n , ψ n−1 , θ n )
in(3.11) byaveraging over su h a omponent. Sin e su h aomponent is a zero mean Gaussian random variable, with varian e
σ 2 f
, thebran h metri be omes:
H n (x n , ψ n−1 , θ n ) = Z
θ f,n
exp (
− θ f,n 2 2σ f 2
)
·
exp ( 1
N 0
Re
" M −2 X
k=0
p k,n a ∗ k,n (x n , ψ n−1 )e −j(θ n +θ f,n )
#)
dθ f,n .
(4.45)wheretheintegralin(4.45) anbeevaluatedasasimplesumbydis retizingthe
faster phase omponent
θ f,n
as done for theslower omponentθ n
.Hen e theI-MM-DP algorithm isobtained byjust in reasing thebran h metri
omple-tionoftheMM-DPalgorithm.However,Se tion 2.4.4showthatimproved-DP
algorithm doesnot get signi ant information rateimprovement with respe t
to the simple DP algorithm. We will verify this statement in theSe tion 4.5,
bynumeri al results.
Dete tion Algorithm Based on Phase Noise Linear Predi tion
In the following, we propose a Bayesian algorithm basedon linearpredi tion
of the phase noise pro ess. In parti ular, we extend the approximate linear
predi tive approa hdes ribed in[71℄,fora generaltime-varyingphasepro ess
θ n
.Weassumeθ n
stationaryanddes ribedbyagivenauto orrelationsequen e of thephasorpro essh n = e jθ n
,denoted byR h (k) = E n
e jθ n+k e −jθn o
.
(4.46)Firstofall,wefo usontheve tor
y
whi h olle tsthesety n (x n , σ n )
ofsuf- ient statisti softhere eivedsignal
r(t)
,for theideal oherent dete tor(seeSe tion4.2).Thesestatisti saregeneratedasindi atedin(2.12),i.e.,byabank
of lters mat hed to all possible length-
T
CPM waveforms{¯s(t; x n , ω n )}
. InthepresentSe tion, wepropose adierent setof su ient statisti s, olle ted
inthe ve tor
r
,obtained proje tingthere eived signalr(t)
overanalternative orthonormal base. In detail, the statisti s are obtained by oversampling there eived signal by an oversampling fa tor
N s
whi h respe ts theanti-aliasingondition;hen e,byoversamplingthesignal
r(t)
in(4.3)att n,k = n T +k T /N s
,we get
r n,k = es n,k (x n , σ n )e jθ n + w n,k , n = 0, 1, . . . , N − 1 k = 0, 1, . . . , N s − 1
(4.47)
where we have dened
r n,k , r(t n,k )
,e s n,k (x n , σ n ) , es(t n,k ; x n , σ n )
andw n,k , w(t n,k )
. In other words,n
is the symbol index, whilek
is the oversampling index.Itistrivialtodemonstratethat{w n,k } n,k
areindependentanduniformly distributed omplexGaussianrandomvariables,withzeromeansandvarian eσ w 2 = 2 N 0 N s /T
.Fromthat onsideration, wendthefollowing expressionfor thepdfp(r|x)
p(r|x) =
N −1 Y
n=0
exp (
− 1 σ w 2
N X s −1 k=0
r n,k − es n,k (x n , σ n )e jθ n 2
)
.
(4.48)We dene
r n,k ′ (x n , σ n ) , r n,k e
s n,k (x n , σ n ) = r n,k e s ∗ n,k (x n , σ n )
(4.49)sothat, repla ing(4.49) in(4.47), when onsidering thetransmitted sequen e
we get
r ′ n,k (x n , σ n ) = e jθ n + w ′ n,k
(4.50)where
w ′ n,k , w n,k s e ∗ n,k (x n , σ n )
isstatisti allyequivalenttow n,k
.Hen e,we anexploitthe re eived samples(4.50) toperform
e jθ n
estimation, basedonlinear predi tion ltering. We derivee j ˆ θ n = P C
i=1 p i r ′ n−i,0 (x n−i , σ n−i )
P C
i=1 p i r ′ n−i,0 (x n−i , σ n−i )
(4.51)
where
C
assumesthemeaningofpredi tionorder and{p i } C i=1
arethepredi toroe ients. The predi tor oe ients an be omputed by solving a
Wiener-Hopf linear system
R p = b
, wherep , (p 1 , p 2 , . . . , p C ) T
is the unknownve tor 10
,
b = [R h (1), R h (2), . . . , R h (C)] T
andR
is a squareC × C
matrix10
Thenotation
(.) T
indi atesthetranspositionoperatorwhoseelementsaredened astheauto orrelationfun tionofthesamples
r n,k ′
,and hasthe followingexpression
R(ℓ, m) =
( R h (|ℓ − m|)
ifℓ 6= m R h (0) + σ w 2
ifℓ = m .
(4.52)
Repla ing(4.51) in(4.48), we derive
p(r|x) ≃
N −1 Y
n=0
exp (
− 1 σ ǫ 2
N X s −1 k=0
|G n (x n , . . . , x n−C , σ n , . . . , σ n−C )| 2 )
(4.53)
where
σ 2 ǫ
is the mean square error predi tion error, whi h an be expressedas[71℄
σ 2 ǫ = R h (0) + σ w 2 − X C
i=0
p i R h (i) .
(4.54)and
G n (x n , . . . , x n−C , σ n , . . . , σ n−C ) , r n,k −es n,k (x n , σ n ) P C
i=1 p i r n−i,0 ′ (x n−i , σ n−i )
P C
i=1 p i r n−i,0 ′ (x n−i , σ n−i ) .
(4.55)
By employing (1.16), we an writethe
s e n,k (x n , σ n )
denitionase s n,k (x n , σ n ) = ¯ s(t n,k ; x n , ω n ) e jπ n
= ¯ s n,k (x n , ω n ) e jπ n
(4.56)where
s(t ¯ n,k ; x n , ω n )
(1.17)isthe omponentoftheCPMwaveform,dependingon just the present symbol
x k
and on the orrelative stateω n
(1.4);π n
(1.5)is the phasestate and we dened
s ¯ n,k (x n , ω n ) , ¯ s(t n,k ; x n , ω n )
. Thus,repla -ing(4.56)and (4.49)in(4.55),we ndthatthe
G n (x n , . . . , x n−i , σ n , . . . , σ n−i )
fun tion be omes
G n (x n , . . . , x n−i , σ n , . . . , σ n−i ) =
(4.57)= r n,k − ¯s n,k (x n , σ n ) P C
i=1 p i r n−i,0 ¯ s ∗ n−i,0 (x n−i , ω n−i ) e j(π n −π n−i )
P C
i=1 p i r n−i,0 s ¯ ∗ n−i,0 (x n−i , ω n−i )
= r n,k − ¯s n,k (x n , ω n ) P C
i=1 p i r n−i,0 s ¯ ∗ n−i,0 (x n−i , ω n−i ) e jπh P i−1 m=0 x n−L−m
P C
i=1 p i r n−i,0 s ¯ ∗ n−i,0 (x n−i , ω n−i )
.
(4.58)
Sin ethe orrelative stateis omposedbythe
L − 1
most re ent pastsymbol,we understand that the fun tion
G n
just depends on the present symbolx n
andonasetof
L − 1 + C
pastsymbolsgivenby(x n−1 , . . . , x n−(L−1+C) )
.Thus,bydeninga newsystemstate
µ n , x n−1 , . . . , x n−(L−1+C)
(4.59)
we have
G n (x n , . . . , x n−i , σ n , . . . , σ n−i ) ≡ G n (x n , µ n ) ,
(4.60)and the pdf in(4.53) anbe written as
p(r|x) ≃
N −1 Y
n=0
V n (x n , µ n )
(4.61)where
V n (x n , µ n ) , exp (
− 1 σ 2 ǫ
N X s −1 k=0
|G n (x n , µ n )| 2 )
(4.62)
Finally, we an derive the MAP symbol dete tion strategy by marginalizing
p(x, µ|y)
(whereve torµ
olle tsµ n
elements,withn
from0
toN − 1
)byFGand SPA. Indetail, byBayesrule
p(x, µ|r) ∝ p(r|x, µ) P (µ|x) P (x)
= P (µ 0 )
N −1 Y
n=0
V n (x n , µ n )I(x n , µ n , µ n+1 )P (x n )
(4.63)where
I(x n , µ n , µ n+1 )
is thetrellis indi ator fun tion, equal to one ifx n
,µ n
,and
µ n+1
satisfythe trellis onstraints and to zero otherwise. It is lear that fa torization (4.63) is equivalent to the fa torization in (2.14) of the optimaloherent dete tor, where here the state denition
µ n
repla es the CPMstateσ n
(and thus, also the trellis indi ator fun tion is dierent), and where thebran h metri
V n (x n , µ n )
repla esF n (x n , σ n )
.Thus, forward re ursion,ba k-ward re ursion and ompletion stage are performed as des ribed in (2.17),
(2.18), (2.21), respe tively. From (4.59), we know that the BCJR algorithm
we have derived isbased on atrellis of
M L−1+C
states. So,it is lear that itsomputational omplexity rapidly in rease with the parameters
L
andC
. Inthe present work,we do not address the problemof omplexity redu tionfor
the des ribed algorithm, but we think that a possible solution is represented
byredu ed-sear h te hniques(des ribed inSe tion3.6).
Predi tor Coe ients for Wiener and double-AR1 PN Models
In order to ompute the predi tor oe ients
p
by the Wiener-Hopf linear systemR p = b
, we need to assume a statisti al model for the phase noiseθ n
.Fromsu h model we an derive an expressionforR h (k)
dened in(4.46),ne essaryfor the omputation ofmatrix
R
(4.52).Ifwemodelthephasenoiseasadis rete-timeWienerpro essofin remental
varian e
σ ∆ 2
,it iseasy to verifythatR h (k) = e −|k|σ 2 ∆ /2 .
(4.64)Ontheotherhand,whenwemodelthePNbythedouble-AR1modeldes ribed
inSe tion 2.4.2,we an derive
R h (k)
asfollows.Firstof all,were allthatthehara teristi fun tion ofanyrandom variable
β
isgiven byγ β (t) = E n
e jtβ o
.
(4.65)Hen e,wenotefrom(4.46),that
R h (k)
isequivalenttothe hara teristi fun -tionoftherealrandomvariables(θ k+n −θ n )
, omputedfort = 1
.Thedis rete-timephasenoisepro ess
θ n
weare onsidering,isthesumoftwoGaussianrstorder auto-regressive pro esses,
θ a,n = a θ a,n−1 + v a,n θ b,n = b θ b,n−1 + v b,n
where
v a,n
andv b,n
aretwo Gaussian zero-mean random variables of varian eσ v,a 2
andσ v,b 2
,respe tively.θ a,n
andθ b,n
arestill Gaussian randomvariables ofvarian e
σ 2 a
andσ b 2
,respe tively.Thus,it is easy to prove that(θ k+n − θ n )
isstillaGaussianvariable,withmeanzeroandvarian e
2[R θ (0) − R θ (k)]
,whereR θ (k)
isthe auto orrelation sequen eR θ (k) = E{θ n θ n+k } .
(4.66)Thus, sin e the hara teristi fun tion of a zero-mean Gaussian random v
ari-able
β
withvarian eσ 2
is [72℄γ β (t) = e −σ 2 /2
(4.67)we derive that
R h (k) = e [R θ (k)−R θ (0)] .
(4.68)Finally,sin e the auto orrelationfun tion
R θ (k)
isthe sum of theauto orre-lationfun tionsof its AR1 omponents