2.4 IR of CPMs over Channel Ae ted by Phase Noise
2.4.3 Dete tors for IR Computation over Channels Ae ted
in the MAP sense. Moreover, it is interesting also to evaluate the IR by a
DP-BCJR operatingovera CT orST hannel generation, sin e we derive the
maximumperforman ewe ana hieve,whenemploying adete tionalgorithm
arried out by starting from a single rst order PN model (Wiener or AR1),
overa hannel ae tedby SATMODEPN.Finally,thethird typeof dete tor
is parti ularly suitable to ompute the IR for double-AR1 PN generation at
low baudrate values. Looking to the double-AR1 parameters in Table 2.3 at
64
kBaud, for example,we see that thefast omponents exhibits ab
valuesolose tozerothatinthedete torderivation, we an onsiderits orresponding
PN samples as almost independent. In su h a way, in the dete tion
deriva-tion, we take into a ount both the two AR1 omponent but we also a hieve
a redu tioninthedete tor omplexity.
trialphase value
θ n
,we dene thesampler n (x n , σ n , θ n )
asr n (x n , σ n , θ n ) ,
Z (n+1)T
nT r(t)es ∗ (t − nT ; x n , σ n )e −jθ n dt
= y n (x n , σ n )e −jθ n
(2.36)where the ve tor
y
olle tingy n (x n , σ n )
(2.12) represents a set of su ientstatisti of the re eived signal.
3
Finally we all
r
the ve tor olle ting thesample
r n (x n , σ n , θ n )
for ea h trial symbolx n
,ea h trial CPMstateσ n
, andphasevalue
θ n
,and ea h dis rete-time instant.We now derive the MAP symbol dete tion through FG and SPA. To this
purpose, we rstfa torize the joint distribution
p(x, σ, θ|r)
asp(x, σ, θ|r) ∝ p(r|x, σ, θ)P (σ|x)P (x)p(θ)
(2.37)= P (σ 0 )P (θ 0 )
N −1 Y
n=0
F n (x n , σ n , θ n )T (x n , σ n , σ n+1 )P (x n )p(θ n+1 |θ n )
where
θ
olle tsθ n
samples (n
from0
toN
),T (x n , σ n , σ n+1 )
is an indi atorfun tion, equal to one if
x n , σ n
andσ n+1
satisfy the CPM trellis onstraints andtozerootherwise,andF n (x n , σ n , θ n )
isthebran hmetri fun tion,denedas
F n (x n , σ n , θ n ) = exp
1 N 0 Re n
y n (x n , σ n )e −jθ n o
.
(2.38)Finally,asstated inSe tion 2.4.1,
p(θ n+1 |θ n )
in(2.37) is aGaussian pdf4p(θ n+1 |θ n ) = 1
q 2πσ 2 ∆
exp
− (θ n+1 − θ n ) 2 2σ ∆ 2
.
(2.39)In orderto have a nitestate representation for su h a hannel, we dis retize
thevalues thatsamples
θ n
mayassume. Obviously,weare approximatingthe 3Notethatalsointhis ase, aswellas inthe oherent optimaldete tor,theve tor
y
isobtainedfromasetofjust
M L
lters.4
Notethat,sin ethe hannelphaseisdenedmodulo
2π
,theprobabilitydensityfun tionp(θ n+1 |θ n )
anbeapproximatedasGaussianonlyifσ ∆ ≪ 2π
.x n η f,n+1 η b,n
P (x n ) (σ n , θ n )
P (x n )
(σ n+1 , θ n+1 ) F n T P θ
P e (x n )
Figure2.7: Fa tor graphofDP-BCJR inthe
n
-thtime-interval.real hannel and, by in reasing thenumber of quantization levels, better
ap-proximations anbea hievedatthepri eofagreater hannelstate ardinality.
In parti ular, we propose an uniform dis retization, bywhi h samples
θ n
areonsidered belongingto thebelowalphabet
θ n ∈
0, 2π
D , 2 2π
D , . . . , (D − 1) 2π D
(2.40)
where
D
isthenumberofquantizationlevels.Hen e,thepdfp(θ n+1 |θ n )
in(2.39)mustrepla edbythepmf
P θ (θ n+1 |θ n )
,whi h an be hosenfor example as:P θ (θ n+1 |θ n ) =
Z θ n+1 + π
D
θ n+1 − D π
q 1 2πσ ∆ 2
exp
− (x − θ n ) 2 2σ ∆ 2
dx .
(2.41)The FG orrespondingto (2.37) has y les.However, by lustering [7℄the
variables
θ n
andσ n
,weobtain theFGinFig.2.7.Thus, byapplyingto ittheSPAbyusinga non-iterative forward-ba kward s hedule, we obtaintheexa t
a posteriori probabilities
P (x n |r)
ne essary to implement the MAP symboldete tion strategy. With referen e to the messages in Fig. 2.7, the resulting
forwardba kward algorithm is hara terized by the following re ursions and
ompletion:
η f,n+1 (σ n+1 , θ n+1 ) = X
σ n
X
θ n
X
x n
P (x n )F n (x n , σ n , θ n )T (x n , σ n , σ n+1 ) P θ (θ n+1 |θ n )η f,n (σ n , θ n )
(2.42)
η b,n+1 (σ n , θ n ) = X
σ n+1
X
θ n+1
X
x n
P (x n )F n (x n , σ n , θ n )T (x n , σ n , σ n+1 ) P θ (θ n+1 |θ n )η b,n+1 (σ n+1 , θ n+1 )
(2.43)
P e (x n ) ∝ X
σ n
X
σ n+1
X
θ n
X
θ n+1
F n (x n , σ n , θ n )T (x n , σ n , σ n+1 )P θ (θ n+1 |θ n ) η f,n (σ n , θ n )η b,n+1 (σ n+1 , θ n+1 )
(2.44)
with the following initial onditions for the two re ursions:
η f,0 (σ 0 , θ 0 ) = P (σ 0 )/(2π)
andη b,N (σ N , θ N ) = P (σ N )/(2π)
. Itis lear thata betterapprox-imation of the real hannel, whi h is a hieved by in reasing the number of
quantizationlevels
D
,determinesalsoanin reaseddete tion omplexity,sin ethe ardinality ofthe trellisstate
(σ n , θ n )
isproportionaltoD p M L−1
.Finally, we an extend the DP-BCJR algorithm to all ases in whi h the
dis rete-time PNpro ess isaGaussian AR1model(2.25).Insu ha ase, itis
su ient to repla ethepmf
P θ (θ n+1 |θ n )
in(2.41) byP θ (θ n+1 |θ n ) =
Z θ n+1 + π
D
θ n+1 − D π
p 1
2π(1 − a 2 )σ a 2 exp
− (x − a θ n ) 2 2(1 − a 2 )σ a 2
dx .
(2.45)Double-DP-BCJR (D-DP)
Wenow onsiderthedouble-AR1phasenoisemodeldes ribedinSe tion2.4.2.
Thus,theaboveDP-BCJRalgorithmderivation anbegeneralizedto the ase
ofa hannelwhosePNmodelisgivenbythesumoftwodis rete-timepro esses
θ (a) n
andθ n (b)
for whi h the following re ursive denitions existθ n+1 (a) = a θ (a) n + v n (a)
(2.46)θ n+1 (b) = b θ n (b) + v n (b)
(2.47)x n η f,n+1
P (x n ) P (x n )
(σ n , θ (a) n , θ n (b) ) η b,n F n T P θ (a) P θ (b) (σ n+1 , θ (a) n+1 , θ (b) n+1 )
P e (x n )
Figure 2.8:FGofDouble-DP-BCJR dete torinthe
n
-th time-interval.where
v (a) n
andv (b) n
areindependent andidenti allydistributedGaussian zero-meanrandomvariablesofvarian e(1−a 2 ) σ a 2
and(1−b 2 ) σ b 2
,respe tively.Also inthis ase, the two phase values an be uniformly dis retized inthedomainand we denote by
D a
andD b
the number of quantization levels forθ n (a)
andθ (b) n
,respe tively.In order to apply the MAP symbol dete tion strategy, we fa torize the
globalpmf
P (x, σ, θ (a) , θ (b) |r)
inawaysimilarto (2.37)andwederivetheFGinFig.2.8.TheSPAallows usto omputethetwo re ursionsand ompletion;
herewereport justtheexpressionfor theforward re ursion:
η f,n+1 (σ n+1 , θ (a) n+1 , θ n+1 (b) ) = X
σ n
X
θ (a) n
X
θ (b) n
X
x n
F n (x n , σ n , θ n (a) , θ (b) n )T (x n , σ n , σ n+1 )
P (x n )P θ (a) (θ n+1 (a) |θ n (a) )P θ (b) (θ n+1 (b) |θ n (b) )η f,n (σ n , θ (a) n , θ (b) n )
(2.48)
where we have dened
F n (x n , σ n , θ n (a) , θ (b) n ) = exp
1 N 0 Re n
y n (x n , σ n )e −j(θ (a) n +θ (b) n ) o
(2.49)
and
P θ (a) (θ (a) n+1 |θ (a) n )
,P θ (b) (θ (b) n+1 |θ (b) n )
arethe pmfsdes ribing the hannelphasetransition, dened asin(2.45).
In su h a ase, the trellis state is
(σ n+1 , θ n+1 (a) , θ (b) n+1 )
and its ardinality resultsD a D b pM L
.Improved-DP-BCJR (I-DP)
Looking at the parameters des ribing the faster PN omponent inTable 2.3,
we see that its
b
parameter is very lose to zero, espe ially for low baudratevalues. In other words, PN samples
θ (b) n
an be onsidered independent from ea h otherand we assumethefollowing model:θ (b) n = v n (b)
(2.50)obtained repla ing
b = 0
in(2.46); hen eP (θ (b) n ) =
N −1 Y
n=0
P θ (b) (θ (b) n ) .
(2.51)where
P θ (b) (θ (b) n )
expression is equivalent to (2.45) witha = 0
. By fa torizing the pmfP (x, σ, θ (a) , θ (b) |r)
, we derive a new dete tion algorithm based onassumption (2.50), whose fa tor graph is provided in Fig 2.9. The forward
re ursion expressionobtained thought theSPAis
η f,n+1 (σ n+1 , θ (a) n+1 ) = X
σ n
X
θ (a) n
X
x n
P (x n )F n (a) (x n , σ n , θ (a) n )T (x n , σ n , σ n+1 )
P θ (a) (θ n+1 (a) |θ n (a) )η f,n (σ n , θ n (a) ).
(2.52)
where the bran h metri is
F n (a) (x n , σ n , θ n (a) ) , X
θ (b) n
F n (x n , σ n , θ (a) n , θ n (b) )P θ (b) (θ n (b) )
(2.53)and
F n (x n , σ n , θ n (a) , θ n (b) )
isprovidedin(2.49).Hen e,we anseethatemployingthe assumption in (2.50), the double-DP-BCJR omplexity is redu ed sin e
η f,n+1 η b,n
P (x n ) (σ n , θ n (a) )
P (x n )
(σ n+1 , θ (a) n+1 ) F n T P θ (a)
P (θ (b) n )
x n P (θ (b) n )
θ (b) n
P e (x n )
Figure2.9: FGof I-DP-BCJRdete torinthe
n
-thtime-interval.here the trellis state is the same of the DP-BCJR, i.e.
(σ n , θ (a) n )
. However,with respe t to the DP-BCJR algorithm we take into a ount a se ond PN
omponent,
θ n (b)
,independent from ea h intervaln
, and so thebran h metriomputation isin reased (see (2.53)).