Anda di halaman 1dari 18

Submitted to IEEE Transa

tions on Information Theory.

Asymptoti Analysis of Superorthogonal Turbo Codes


1 2
Ola Wintzell , Mi hael Lentmaier , and Kamil Sh. Zigangirov

Department of
Information Te hnology
Lund University
P. O. Box 118
SE-221 00 Lund, Sweden
e-mail: fola, mi hael, kamilgit.lth.se
O tober 24, 2001

Abstra t
In this paper we examine a low-rate turbo oding s heme based on superorthogonal
onvolutional en oders [1℄,[2℄,[3℄. The low-rate oding is suitable for CDMA appli a-
tions. We use the property that the omponent en oders are equivalent to onventional
onvolutional en oders to analyze the asymptoti performan e. We analyze the iter-
ative de oding performan e that an be a hieved when both the ode length and the
number of iterations tend to in nity and present a bound on the iterative limit of the
ode onstru tion. It is shown by asymptoti analysis, that the rate 1/7, 1/15 and 1/31
odes with omponent en oders of memory 3, 4 and 5 have iterative limits below -0.65
dB, -0.88 dB and -0.95 dB, respe tively. Simulations for odes with large permutors
(interleavers) on rm these asymptoti results. The onstru tion is general and an be
done for odes of lower rates as well.

Keywords | Low-density onvolutional odes, low-rate turbo odes, superorthogonal on-


volutional odes, superorthogonal turbo odes, iterative de oding
1 This work was supported in part by the Swedish Foundation for Strategi Resear h - Personal Computing
and Communi ation(PCC) and Eri sson Mobile Communi ations.
2 This work was supported in part by Swedish Resear h Coun il for Engineering S ien es (Grant 98-216).
1 Introdu tion

Low-rate odes have a parti ular appli ation in CDMA systems. There are several di erent
onstru tions that address this area [4℄,[5℄.
Turbo odes [6℄ attra ted a great interest in the oding ommunity, when they were in-
trodu ed in the mid-1990s. Simulation results indi ated an error orre ting performan e of
turbo odes, being loser to the theoreti al limit predi ted by Shannon as observed for any
pra ti al oding system before. Over the last years, generalizations to odes of lower rates
have been presented [1℄,[7℄,[8℄. The majority of theoreti al work about the performan e of
turbo odes is devoted to the investigation of maximum-likelihood de oding, using the on-
ept of uniform permutors [9℄. An analysis of iterative de oding of the original turbo odes
3

and of their onvolutional ode version [10℄ is given in [11℄,[12℄. It is based on a orresponding
analysis of Gallager's low-density parity- he k odes [13℄, given in [14℄ (a losely related work
is presented in [15℄). The results in [12℄ in lude an upper bound on the de oding bit error
probability as a fun tion of the number of iterations and an upper bound on the iterative limit
introdu ed in [14℄. Many ideas used in the analysis in [11℄,[12℄ go ba k to the original work
by Gallager [13℄, who used tree-like graphs to des ribe his probabilisti de oding algorithm
for low-density parity- he k odes. Gallager's algorithm was later extended to a wider range
of odes by Tanner [16℄. A des ription of turbo odes by a Tanner graph is given in [17℄.
In this paper we examine a family of parallel on atenated odes, the superorthogonal
turbo odes [1℄, based on superorthogonal onvolutional odes [2℄,[3℄,[18℄. Iterative de oding
of these odes is shown to result in very good bit-error-rate performan e. Using the property
that ea h omponent en oder is equivalent to a onventional onvolutional en oder, allows us
to derive analyti bounds on the iterative behavior, whi h will give asymptoti performan e
measures of the odes. Simulation results for odes with large permutors agree well with the
analyti al results.
3 The term \interleaver" is usually used in the te hni al literature to des ribe a devi e onsisting of a table
whi h is lled olumnwise by input symbols and read rowwise. Following a suggestion by J. Massey, we will
use the term \permutor" for a general devi e that hanges the order of symbols in the input information
sequen e. The lass of these devi es ontains the onventional interleavers as a spe ial ase

1
Blo k length 4 + +
orthogonal + +
en oder + +
+ +

Figure 1: A memory m = 3 superorthogonal onvolutional en oder.


The remainder of this paper is organized as follows: in Se tion 2 we des ribe the re ursive
superorthogonal ode, in Se tion 3 we present the onstru tion of the low-rate turbo ode,
in Se tion 4 we analyze the iterative limits, in Se tion 5 we present simulation results and in
Se tion 6 we summarize the results.

2 Re ursive Superorthogonal Convolutional En oders

Superorthogonal (non-re ursively en oded) onvolutional odes were introdu ed by Zehavi


and Viterbi [2℄,[3℄. The examined en oder onsists of a length m shift register and a length
= 2m orthogonal blo k en oder. The odesymbols generated at ea h time instan e
1

are the bitwise modulo-2 sum of the input symbol, the ontent of the rightmost register
element and a odeword of a length orthogonal blo k (e.g. Hadamard-Walsh, also known
as redu ed Reed-Muller) ode, whi h is determined by the ontents of the (m 1) leftmost
register elements. The en oder has rate R = 1= = 1=2m . An example of a rate 1/4
1

superorthogonal onvolutional en oder of memory m = 3 is presented in Figure 1.


Let the ontent of the shift-register de ne the state of the en oder. We denote ea h pair
of states that only di er in the rightmost register element as a luster. It is evident from
Figure 1, that all outputs asso iated with the two states in a luster are based on one single
orthogonal odeword. Adding the information symbol to the orthogonal odeword ensures,
that the output bits asso iated with the two transitions from a state in the en oder trellis are
omplementary. In the same way, adding the ontent of the last register element makes sure,

2
000 0=0000 000 000 0=0000 000
1=1111 1=1111

001 0=1111 100 001 1=1111 100


1=0000 0=0000
(a) (b)

Figure 2: The 00- ode luster of a superorthogonal ode

that the output bits orresponding to the two transitions into a state are omplementary, as
well. The transitions for the luster 00 are illustrated in Figure 2(a). Moreover, sin e all
odewords of the orthogonal ode are mutually orthogonal, the output bits orresponding to
transitions from di erent lusters are orthogonal. From these properties of the odesymbols
follows that the free distan e of the ode is
dfree = 2m 1
+ (m 1)2m + 2m = (3 + m)2m
2 1 2
: (1)
The regular stru ture of the orthogonal odewords allows a de oder to use the fast Hadamard
transform (FHT), whi h redu es the de oding omplexity to the order of log .
As an be seen in Figure 2(a), a systemati mapping of the super-orthogonal feedforward
onvolutional en oder is impossible, sin e the two transitions into ea h state are asso iated
with the same information symbol. Introdu ing a feedba k loop in the en oder does preserve
the stru ture of the trellis but hanges the information symbol assignment in su h a way, that
a systemati mapping be omes possible. Figure 3 presents a m = 3 re ursive onvolutional
en oder, whi h generates the same ode as the en oder in Figure 1. The orresponding 00
luster is shown in Figure 2(b).
For turbo ode appli ations the \minimal information symbol distan e" is an important
measure. It is de ned as minimal weight of any non-zero information sequen e that generates a
valid odeword. For the original feedforward superorthogonal onvolutional en oder this value
is 1. When the omponent en oders possess this property, a weight 1 information sequen e
will produ e low weight ode sequen es from both the en oders, whi h means that there is no
3
+
+

Blo k length 4 + +
orthogonal + +
en oder + +
+ +

Figure 3: A re ursive memory m = 3 superorthogonal onvolutional en oder.


gain in the minimum distan e of the turbo ode from in reasing the size of the permutor. A
re ursive super-orthogonal onvolutional en oder has a minimal information symbol distan e
equal to 2, as any other re ursive en oder. Using these en oders as omponents in a turbo
ode, it is possible to a hieve any arbitrary large minimum distan e by the proper hoi e of
the permutor. A onstru tion pro edure for permutors that gives a logarithmi growth of the
minimum distan e with the permutor length is given in [11℄.
The di erent possible mappings of the orthogonal en oder an be des ribed by the gen-
erator matri es of the ode. One possible generator matrix of the rate 1/4 orthogonal blo k
ode is 0 1
1 1 0 0
G= A : (2)
0 1 0 1
Using the generator matrix G it is possible to immediately draw the superorthogonal on-
volutional en oder from Figure 3 in ontroller anoni al form, like shown in Figure 4. The
ontent of shift-register element i is added to output symbol j , if the entry in the ith row and
j th olumn in G is equal to 1. Using the notation from Chapter 2 in [19℄, a generator matrix
for the superorthogonal onvolutional en oder in Figure 4 is given by
 
G(D) = 1 D D D
1+ + 2 + 3
D D
1+ + 3
D
1+ 3
D D
1+ + 3
D D
1+ 2 + 3
D D
1+ + 3
: (3)
Note, that the olumns in G form all possible ombinations of length m 1 = 2 binary ve tors.
This is a onsequen e of the obvious property, that all the odewords of the orthogonal ode
are mutually orthogonal. Consequently, the n outputs are generated by all di erent possible
4
+
+
+ +
+ + +
+
+ +

Figure 4: A re ursive memory m = 3 superorthogonal onvolutional en oder in ontroller


anoni al form.
delayfree feedforward polynomials of degree m = 3. Independently on whi h feedba k poly-
nomial of degree m = 3 we use, one of the outputs will be generated by the same polynomial,
and hen e be identi al to the input. In other words, the input symbol automati ally appears
at one of the outputs. By adjusting the feedba k polynomial and the generator matrix G to
ea h other it is always possible to obtain a systemati en oder where the information symbol
appears at the rst of the n outputs. In this work we hoose primitive feedba k polynomials,
in parti ular 64 for m = 3, 62 for m = 4 and 45 for m = 5 (in o tal notation).

3 Low-Rate Turbo C odes based on Superorthogonal

Convolutional Codes

The low-rate turbo en oder is illustrated in Figure 5. It onsists of two memory m super-
orthogonal onvolutional en oders (SOCE), des ribed in the previous se tion, and a permutor
 of size N whi h hanges the order of the symbols at its input. Mathemati ally, the permu-
tor an be des ribed by an N  N permutation matrix , having a single one in ea h olumn
and row. A length N information blo k u of N at the input of the permutor is mapped into
the output blo k u0 = u. The rst SOCE if fed by the information blo k u and the se ond
SOCE by the permuted blo k u0 . The outputs of the rst and the se ond SOCE are the parity
he k blo ks v = (v ; : : : ; vN ) and v = (v ; : : : ; vN ), respe tively. Ea h subblo k vi
(1) (1)
1
(1) (2) (2)
1
(2) (1)

5
u u

SOCE 1 v (1)


SOCE 2 v (2)

Figure 5: A memory m = 3 rate R = 1=7 superorthogonal onvolutional turbo en oder.


and vi ontains the 1 = 2m 1 parity he k symbols orresponding to the information
(2) 1

symbols ui and u0i, i = 1; : : : ; N . The information blo k u and the two parity he k blo ks
v and v are then serialized and transmitted over the hannel.
(1) (2)

In this paper we onsider transmission over the AWGN hannel. Let r; r and r be
(1) (2)

the re eived blo k, orresponding to the information symbols and the parity he k symbols
from en oder one and two, respe tively.
The iterative de oder is a standard iterative turbo de oder and operates as follows. First,
using the ve tors r and r , it al ulates the a posteriori probabilities (APP) of the symbols
(1)

from the information blo k u. This is performed by a symbolwise APP omponent de oder,
based on an algorithm known as the BCJR algorithm [20℄. In the se ond iteration, the
se ond omponent de oder operates on the permuted information blo k u0 , making use of the
newly al ulated a posteriori probabilities from the rst iteration, and the parity he k bits
orresponding to the se ond en oder, r . In order to obtain independent new a posteriori
(2)

probabilities the \extrinsi " information from the previous iteration about the symbol at hand
is not used. In the third iteration the a posteriori probabilities of the se ond de oder is used
in the rst de oder. The following iterations, up to the (I 1)th, are arried out analogously.
In the last, I th, iteration, the omponent de oder ombines a posteriori information, whi h
in ludes the extrinsi information from both the de oders and the hannel information, to
make its nal de ision estimating the transmitted information u. In pra ti e, it may be
more onvenient to operate with log-likelihood ratios, whi h has some numeri al advantages
[21℄, rather than to operate with probabilities. Sin e both representations are equivalent we

6
will not distinguish between them. For a more detailed des ription of turbo de oding and
log-likelihood algebra we refer the reader to other papers su h as [6℄,[21℄. Below we will
analyze the performan e of the onstru ted low-rate turbo ode based on superorthogonal
onvolutional odes.

4 Iterative Limits Analysis

The iterative limit, whi h was de ned in [14℄ in onne tion with an investigation of Gallager's
low-density parity- he k odes [13℄, is a fundamental limit spe ifying when error free trans-
mission is possible with a parti ular ode onstru tion ( ode family) and iterative de oding.
In this paper the ode onstru tion is given by the superorthogonal turbo odes, as introdu ed
in the previous se tion.
For maximum likelihood de oding algorithms the de oding omplexity per bit, C , is b

growing exponentially with the blo k length N , whi h makes them impra ti al for large values
of N . A ru ial feature of iterative de oding algorithms, like the standard turbo de oding
algorithm, is, that the omplexity C is growing linearly with the number of iterations I ,
b

independently on N . Although this feature is essential, it is not suÆ ient to ompletely


hara terize an iterative de oding algorithm in general. Consider a de oder that in the rst
iteration pi ks two arbitrary distin t odewords. Based on the re eived sequen e it hooses
the more probable odeword and keeps it in memory. In the se ond iteration it pi ks a
new one of the remaining odewords and ompares it with the memorized. Again it only
keeps the more probable odeword in memory. After he king all odewords, whi h requires
I = 2N iterations, this de oder will have found the most probable odeword. The de oding
omplexity, C , of this maximum likelihood de oder is a linear fun tion of the number of
b

iterations I . To disregard su h pra ti ally irrelevant de oding algorithms, we demand from


an iterative de oding algorithm, that
 the de oding omplexity per bit, C , is a linear fun tion of the number of iterations I
b

 the bit error probability, Pb, goes to zero at least exponentially with the number of

7
iterations I , when the size of the permutor and I go to in nity, if the signal-to-noise
ratio per bit, Eb=N , is larger than some limit .
0

The value of depends on the stru ture of the onsidered ode and the de oding algorithm.
The iterative limit, denoted by SNR , is the in mum of over all superorthogonal turbo
0

odes with xed omponent en oder memory m and all iterative de oding algorithms. We
present an upper bound SNR on SNR . Sin e the Shannon limit gives a lower bound on
0 0

SNR , this also proves the existen e of the iterative limit. It would be interesting to nd a
0

more tight lower bound on SNR . 0

In our analysis of iterative de oding we make use of results given in the paper [11℄, whi h
deals with graphs des ribing turbo odes. Conditions are formulated, under whi h the a
posteriori probabilities, al ulated in the de oding pro ess, stay \ orre t" during the de oding
iterations. In the following we summarize the results that are relevant in the ontext of this
paper.
Assume, that the omponent de oders operate on a sliding window of size 2 + 1, i.e.
to de ode a parti ular symbol only the losest information symbols ( to the left and  to
the right) are involved, together with the orresponding parity- he k symbols. This permits
a graph representation of the turbo odes, analogous to the representation of low-density
parity- he k odes in [14℄. The iterative de oding of any information symbol un an then be
des ribed by a tree-like graph, alled (un,)- lan, whi h illustrates how di erent symbols are
involved in the de oding of un during the iterations.
In an (l ; )-nondegenerated turbo ode, a ording to the de nitions in [11℄, all symbol
0

nodes in this tree up to the l th generation represent di erent information symbols (the graph
0

des ribing the ode has no y les of length 2l or less). For su h a ode the a posteriori prob-
0

abilities, al ulated by an iterative de oder, are \ orre t" during l iterations. The following
0

theorem gives a lower-bound on the degeneration level l of turbo odes, analogously to the
0

Gallager-Margulis theorem [13℄,[22℄ for low-density parity- he k odes:


Theorem 1 [11℄ There exists an (l0 ; )-nondegenerated turbo ode with permutor size N ,

8
su h that
l0 
1 (log N 3 log (log N ) 4) 1 ; (4)
2 log (2)
2
2 2 2

where is some onstant that does not depend on N .

The proof of this theorem is onstru tive, and as a onsequen e we an onstru t a turbo
ode suitable for I = l independent de oding iterations for any arbitrary large number l
0 0

and any , if N ! 1. This allows us to onsider in the following only iterative de oding of
(l ; )-nondegenerated odes where all iterations I = l are independent.
0 0

In prin iple, sin e all probabilities involved in the de oding are independent, it would be
possible to exa tly al ulate the bit error probability Pb by re ursively pro eeding upwards
through the tree. One problem that ompli ates the situation is, that we annot resort to
numeri al methods only, if we want to obtain a rigorous bound on the iterative limit. Instead
we need some analyti expression for { or at least bound on { the de oding performan e,
that gives insight into the asymptoti behavior of the de oder when the permutor size N and
the number of iterations I = l go to in nity. On the other hand it is diÆ ult to nd su h
0

an expression or bound whi h is simple enough to handle and still so a urate that it gives
reasonably tight results. To ope with this problem, we introdu e, like in [11℄,[12℄,[14℄, a
two-phase iterative de oding algorithm, whi h permits an asymptoti analysis by ombining
numeri al and mathemati al methods.
The rst de oding phase (the rst I iterations) is identi al to the usual turbo de oding
1

algorithm as des ribed in Se tion 3 and is analyzed using numeri al methods. There is no
known feasible method to dire tly al ulate the output distributions of the omponent APP
(BCJR) de oder. Therefore, we use Monte Carlo methods, like in [23℄, to re ursively estimate
the probability density fun tions of the log-likelihood ratios zn of the symbols un after the
individual iterations, going level by level upwards in the de oding tree.
In the se ond de oding phase we use a di erent algorithm, whi h we all max-path APP
algorithm. Like in the rst de oding phase the omponent de oders operate on a sliding
window of size 2+1 around the symbol to de ode. Consider the nth se tion of the omponent
de oder trellis, orresponding to the information symbol un. There are 2m transitions in this
9
se tion that orrespond to un = 0 and 2m transitions that orrespond to un = 1. Instead
of al ulating the a posteriori probability for symbol un, like in the rst de oding phase, the
omponent de oder of the se ond phase determines in ea h iteration the probability of the
most probable path in the trellis orresponding to un = 0 and the probability of the most
probable path orresponding to un = 1. The turbo de oder operates on these two probabilities
in the same way as in the rst phase on the outputs of the BCJR de oder. The results from
the rst de oding phase are used as input in the rst iteration of the se ond phase.
The omponent de oder in the se ond phase an be seen as a suboptimal version of the
BCJR algorithm and hen e the two-phase algorithm gives a pessimisti result ompared to an
algorithm that uses the rst de oding phase during all iterations. The advantage of the se ond
phase is, that it an be analyzed by generalizing the on ept of path weight enumerators to
the de oding tree. By applying methods, that are ommon for investigating onventional
onvolutional odes, su h as the Bhatta haryya bound and the Viterbi bound on the bit error
probability, we an then obtain an upper bound on the bit error probability of the two-phase
algorithm for an arbitrary large number I = l of iterations.
0

In the next se tion we present results of our analysis of the two-phase algorithm and
ompare the re eived bound on the iterative limit with omputer simulations.

5 Results of the Analysis and Computer Simulations

After the numeri al analysis of the rst de oding phase, we obtain the estimated probabil-
ity density fun tions f (z) and f (z) of the log-likelihood ratios zn after the I th iteration,
0 1 1

onditioned on un = 1 and un = 0, respe tively. In Figure 6 (a) and (b) we present f (z) 0

for di erent I for hannels with Eb=N = 0:66 dB and Eb =N = 0:65 dB, respe tively.
1 0 0

Observe, that in the s enario of Figure 6 (b) the average log-likelihood ratio zn is in reasing
with I , while it gets stu k after some I in the ase of the slightly inferior hannel in Fig-
1 1

ure 6 (a). From these density fun tions we an al ulate the Bhatta haryya parameter of the

10
0.3 0.3
0.25 0.25
0.2 0.2
f0(z)

f0(z)
0.15 0.15
0.1 0.1
0.05 0.05
0 0
0 10 20 30 40 50 0 10 20 30 40 50
z z

(a) Eb =N0 = 0:66 dB for I1 =1,50 and 100 (b) Eb =N0 = 0:65 dB for I1 =1,50,70,75 and
(from left to right). 78 (from left to right).

Figure 6: The distribution of the log-likelihood ratio zn for the memory m = 3 ode at
Eb =N = 0:65 and -0.66 dB.
0

information symbols after the rst de oding phase,


Z 1 Z 1
DI1 =
def
(f (z)f (z))
0 1
1
2 dz = (f (z)f ( z))
0 0
1
2 dz : (5)
1 1
In Figure 7 we present DI , evaluated for the same hannels as in Figure 6, as a fun tion of
1

the number of iterations I for the memory m = 3 omponent en oder. (The Bhatta haryya
1

parameter, DI , is also an upper bound on the bit error probability Pb after the rst de oding
1

phase.)
To analyze the se ond de oding phase, we introdu e the fun tion

F (Z ) =
1 T (X; Y; Z )
; (6)
V
X Z Y =1
X=V

where T (X; Y; Z ) denotes the extended path enumerator [19℄ of the omponent en oders. The
dummy variables X , Y and Z generate the total weight, the length and the information weight
of ode sequen es, respe tively. V = exp( RE =N ) denotes the Bhatta haryya parameter
b 0

of the AWGN hannel.

11
0
10
−1
10
−2
10

1
I
D
−3
10
−4
10

20 40 60 80 100
I1

Figure 7: The Bhatta haryya parameter as a fun tion of the number of iterations I for the 1

memory m = 3 ode at Eb =N = 0:65 and -0.66 dB, from bottom and up. The ir le marks
0

D de ned in Theorem 2 for Eb =N = 0:65.


r 0

Lemma 1 [12℄ Let FV (Z ) satisfy the onditions


8
>
>
>
>
FV (Z )jZ =0 = 0 ;
>
<
= FV0 (Z )
def <1 ; (7)
>
>
>
Z =0
>
Z FV (Z ) ! 1 if Z ! 1 :
>
:1

Then there exists a positive root Z =Z r to the equation

FV (Z ) = Z : (8)
The onditions in (7) require, that V does not ex eed a ertain value, depending on the
omponent en oders.
Now we are ready to formulate the following
Theorem 2 [12℄ Consider a sequen e of (l0 ; )-nondegenerated superorthogonal turbo odes,
where l0 ! 1 and  ! 1. Suppose these odes are used for the transmission over an
AWGN hannel with signal-to-noise ratio per bit Eb =N0 , su h that the extended path weight
enumerator satis es (7). Let the two-phase algorithm des ribed in Se tion 4 be used for
de oding, having a total number of iterations I =l 0 and I1 < I iterations in the rst phase.

12
3mSOC 4 5
R 1/7 1/15 1/31
SNR 0:65 0:88 0:95
0

SNR 1:15 1:39 1:49


Shannon

D 0.095 0.11 0.13


r

Table 1: Values of the upper bound SNR and the orresponding Shannon limit, SNR
0 Shannon ,
and D . r

def
Then, if DI1 < V Z r =D r , where Z r is a positive root to equation (8) the bit error probability
Pb for the given sequen e of odes goes to zero at least exponentially with I , when I and N
go to in nity.

More details about the analysis of the se ond phase an be found in [12℄.
The riti al value D an be evaluated for di erent signal-to-noise ratios Eb =N with help
r 0

of the extended path enumerator T (X; Y; Z ) of the omponent en oders. In Figure 7 the
orresponding D is shown by a ir le. Note that at Eb =N = 0:66 dB, DI will not go
r 0 1

below D within 100 iterations.


r

By evaluation of DI for di erent SNRs and omparison with the respe tive riti al values
1

D , we obtain an upper bound SNR on the iterative limit of the ode. In Table 5 we present
r 0

the upper bound SNR and the orresponding Shannon limit SNR
0 and the riti al value
Shannon

of the Bhatta haryya parameter D for the orresponding signal-to-noise ratio.


r

Computer simulations for random permutors of size 10 were arried out in order to
5

evaluate the performan e of the evaluated turbo ode in the AWGN hannel. For ea h re eived
blo k the de oder ran 200 iterations (100 de odings in ea h omponent de oder) before making
its de ision. The results of the bit error rate (BER) as a fun tion of the signal-to-noise ratio,
Eb =N , is depi ted in Figure 8 for the odes with omponent en oder memory m=3, 4, and
0

5. The verti al solid lines represent the upper bound SNR on the iterative limit.
0

13
0
10
−1
10
−2
10
−3
10

BER
−4
10
−5
10
−6
10

−1.1 −1 −0.9 −0.8 −0.7 −0.6 −0.5


E /N [dB]
b 0
Figure 8: Bit-error rates, dashed and dotted lines, and SNR , verti al solid lines, as fun tions
0

of Eb =N for the memory 3, 4 and 5 omponent odes, from right to left.


0

6 Con lusions

We have shown that superorthogonal en oders are equivalent to onvolutional en oders and
established that these low-rate en oders are suitable for iterative de oding. By an asymptoti
analysis we have shown, that for omponent en oder memories m = 3, 4 and 5, orresponding
to ode rates R = 1=7, 1/15 and 1/31, error free ommuni ation is possible at about 0.5 dB
away from the Shannon limit. Simulation results are in good agreement with the asymptoti
limits.
The low-rate ode an be utilized in a CDMA appli ation and the orthogonal odewords
allow redu ed omplexity in the de oder. The onstru tion is general and odes of lower rate
an be onstru ted as well.
The very re ent results in [8℄ present the onstru tion of turbo-Hadamard odes. These
odes have lower omplexity and are shown, by simulations, to perform better than the
superorthogonal turbo odes. It would be interesting to generalize the analysis in this paper
to s hemes with more than two omponent en oders and rate b= omponent odes, where
b > 1. Then the iterative limits of the turbo-Hadamard odes ould be derived as well.

14
Referen es

[1℄ P. Komulainen and K. Pehkonen, \Performan e evaluation of superorthogonal turbo


odes in AWGN and at Rayleigh fading hannels," IEEE Journal on Sele ted Areas in
Communi ations, vol. 16, pp. 196{205, Feb 1998.

[2℄ A. J. Viterbi, \Orthogonal tree odes for ommuni ation in the presen e of white gaussian
noise," IEEE Trans. on Communi ations, vol. COM-15, pp. 238{242, 1967.
[3℄ E. Zehavi and A. J. Viterbi, \On new lasses of orthogonal onvolutional odes," in
Pro eedings of Bilkent Conferen e in New Trends in Communi ation, Control, and Signal
Pro essing, (Ankara, Turkey), 1990.
[4℄ A. Viterbi, \Very low rate onvolutional odes for maximum theoreti al performan e of
spread-spe trum multiple-a ess hannels," IEEE Journal on Sele ted Areas in Commu-
ni ations, vol. 8, pp. 641{649, May 1990.

[5℄ P. Frenger, P. Orten, and T. Ottosson, \Code-spread CDMA using maximum free dis-
tan e low-rate onvolutional odes," IEEE Trans. on Inform. Theory, vol. IT-48, pp. 135{
144, Jan 2000.
[6℄ C. Berrou and A. Glavieux, \Near optimum error- orre ting oding and de oding: Turbo
odes," IEEE Trans. on Communi ations, vol. COM-44, pp. 1261{1271, O t. 1996.
[7℄ W. Leung, K. Ho, and L. Ping, \Approa hing low rate Shannon limit using on atenated
Hadamard odes," Ele troni s Letters, vol. 36, pp. 45{46, Jan 2000.
[8℄ L. Ping, W. Leung, and K. Y. Wu, \Low-rate turbo-Hadamard odes."
Submitted to IEEE Trans. on Inform. Theory, available online at
http://www.ee. ityu.edu.hk/~liping/resear h/journal, 2001.

[9℄ S. Benedetto and G. Montorsi, \Unveiling turbo odes: Some results on parallel on ate-
nated oding s hemes," IEEE Trans. on Inform. Theory, vol. IT-42, pp. 409{428, Mar h
1996.
15
[10℄ K. Engdahl and K. S. Zigangirov, \On the theory of low-density onvolutional odes I,"
Problemy Pereda hi Informatsii, vol. 35, no. 4, pp. 12{27, 1999.

[11℄ D. Truha hev, M. Lentmaier, and K. S. Zigangirov, \Some results on erning design
and de oding of turbo- odes," Problemy Pereda hi Informatsii, available online (in En-
glish) at http://www.it.lth.se/mi hael/Publi ations/Publi ations.html, vol. 37,
pp. 6{23, July-September 2001.
[12℄ M. Lentmaier, D. Truha hev, and K. S. Zigangirov, \On the theory of low-density on-
volutional odes II," Problemy Pereda hi Informatsii, available online (in English) at
http://www.it.lth.se/mi hael/Publi ations/Publi ations.html, vol. 37, pp. 15{
35, O tober-De ember 2001.
[13℄ R. G. Gallager, Low-Density Parity-Che k Codes. Cambridge, Mass.: MIT Press, 1963.
[14℄ K. S. Zigangirov and M. Lentmaier, \Mathemati al analysis of an iterative algorithm for
low-density ode de oding," Problemy Pereda hi Informatsii, vol. 36, pp. 35{46, O tober-
De ember 2000.
[15℄ T. Ri hardson and R. Urbanke, \The apa ity of low-density parity- he k odes under
message-passing de oding," IEEE Trans. on Inform. Theory, vol. IT-47, pp. 599{618,
February 2001.
[16℄ R. M. Tanner, \A re ursive approa h to low omplexity odes," IEEE Trans. on Inform.
Theory, vol. IT-27, pp. 533{547, September 1981.

[17℄ N. Wiberg, Codes and de oding on general graphs. PhD thesis, University of Linkoping,
April 1996.
[18℄ A. J. Viterbi, CDMA Prin iples of Spread Spe trum Communi ation. Addison-Wesley,
1995.
[19℄ R. Johannesson and K. S. Zigangirov, Fundamentals of Convolutional Coding. Pis at-
away, N.J.: IEEE Press, 1999.
16
[20℄ L. R. Bahl, J. Co ke, F. Jelinek, and J. Raviv, \Optimal de oding of linear odes for
minimizing symbol error rate," IEEE Trans. on Inform. Theory, vol. IT-20, pp. 284{287,
Mar h 1974.
[21℄ J. Hagenauer, E. O er, and L. Papke, \Iterative de oding of binary blo k and onvolu-
tional odes," IEEE Trans. on Inform. Theory, vol. IT-42, pp. 429{445, Mar h 1996.
[22℄ M. A. Margulis, \Expli it group-theoreti onstru tions for ombinatorial designs with
appli ations to expanders and on entrators," Problemy Pereda hi Informatsii, vol. 24,
no. 1, pp. 51{60, 1988.
[23℄ T. Ri hardson and R. Urbanke, \Thresholds for turbo odes," in Pro eedings of 2000
IEEE Intern. Symposium on Inform. Theory, p. 317, 2000.

17

Anda mungkin juga menyukai