Professional Documents
Culture Documents
Abstra
t
Analyti
expressions for the exa
t bit error probabilities of rate R = 1=2, memory m = 2
onvolutional
en
oders are derived for a maximum likelihood (ML) de
oder and transmission over the binary symmetri
hannel (BSC). The resulting expressions are rational fun
tions of the
rossover probability of the BSC.
In addition to
lassi
al non-systemati
en
oders without feedba
k we
onsider also re
ursive systemati
en
oders, whi
h be
ame espe
ially important as
omponent en
oders in
on
atenated
oding s
hemes. To
attest the validity of the results, they are
ompared to
omputer simulations. Based on the presented
te
hnique also the bit error probability and the probability distribution of the output log-likelihood ratios
of the Max-Log-MAP algorithm are derived in analyti
form.
I. Introdu tion
Convolutional
odes
an be found in many appli
ations. Their spe
ial stru
ture makes
them simple to en
ode and the Viterbi algorithm [1℄[2℄ provides a very eÆ
ient method
for maximum-likelihood (ML) de
oding. For very good
hannels the performan
e of ML
de
oded
onvolutional
odes
an be estimated quite a
urately with union bound te
h-
niques. However, it gets more and more diÆ
ult to obtain tight bounds on the error
probability when the
rossover probability p of the
hannel gets larger. But this region
is parti
ularly interesting in iterative de
oding s
hemes, where several simple
omponent
de
oders together try to rea
h the
orre
ting
apability of a large overall
ode. Therefore
it
an be of great theoreti
al value to get some insight into the de
oders and analyze the
de
oding performan
e even for
onvolutional
odes with small memory.
That nite Markov
hain theory
an be used for the error probability
al
ulation of a
onvolutional
ode has been observed by Morrissey already in the late 1960's. He
al
ulated
1 This work was supported in part by Swedish Resear
h Coun
il for Engineering S
ien
es (Grant 01-3123).
the exa
t bit error probability of systemati
en
oders without feedba
k for a suboptimum
feedba
k de
oder [3℄. For a memory m = 1 en
oder it has been demonstrated that this
approa
h
an be extended to ML de
oding [4℄. A drawba
k of Morrissey's model is that
the
omplexity grows not only with the
ode memory but also with the de
oding delay.
A method to
al
ulate the event error probability of
onvolutional
odes was developed
by S
halkwijk et. al. [5℄. Their analysis is based on a de
oder operating in the syndrome
former trellis, whi
h
an be shown to be equivalent to a
onventional Viterbi de
oder
[6℄. For
ertain en
oders this de
oding method redu
es the number of possible metri
values, whi
h allows savings in hardware implementations. The Viterbi algorithm is not
only appli
able for
hannel de
oding but also for sour
e en
oding. The sour
e en
oding
performan
e of binary
onvolutional
odes was investigated by Calderbank et. al. [7℄,
also based on some Markov
hain approa
h. The te
hnique from [7℄ was then used by
Best et al. [8℄ to evaluate the exa
t bit error probability, Pb , for
onvolutional
odes with
Viterbi de
oding. An expli
it analyti
expression of Pb has been derived in [8℄ for a simple
memory m = 1
onvolutional en
oder without feedba
k. For
odes with memory m = 2
and m = 3 only numeri
al results have been presented. In the present paper we
al
ulate
and make available
orresponding analyti
expressions for a memory m = 2
onvolutional
ode, using the ideas from [8℄.
Let us
onsider a binary, rate R = 1=2
onvolutional
ode. The information sequen
es
u = : : : u 1 u0 u1 u2 : : : , ut 2 GF (2), are mapped by a
onvolutional en
oder into
ode se-
quen
es v = : : : v 1 v0 v1 v2 : : : , where v t = (vt(1) vt(2) ) and vt(1) ; vt(2) 2 GF(2). The mapping
an be des
ribed by a generator matrix G(D) = (g1 (D); g2(D)), where g1 (D) and g2 (D),
in general, are rational transfer fun
tions. Then the D-transform, v (D), of an arbitrary
ode sequen
e v
an be written as the produ
t v (D) = u(D)G(D), where u(D) is the
D-transform of an input information sequen
e u. The
ode sequen
es are transmitted
over a binary symmetri
hannel (BSC) with
rossover probability p. At the re
eiver a
maximum-likelihood (ML) de
oder, based on the Viterbi algorithm, uses the re
eived se-
quen
e r = : : : r 1 r0 r1r2 : : : , rt = (rt(1) rt(2) ), rt(1) ; rt(2) 2 GF(2), to make an estimate u^
of the information sequen
e u. We are interested in the probability P (^ut 6= ut ) that an
information symbol at some arbitrary time t is de
oded erroneously. Sin
e we assume
innite sequen
es this probability is independent of the time t and will be denoted by
Pb . Wei et. al. [9℄ showed how the te
hnique in [8℄
an be extended to take into a
ount
transmission in blo
k format or nite ba
k-sear
h limit in the de
oder. In the rst
ase
P (^ut 6= ut ) will depend on the time index t and Pb from the innite
ase will form an
upper bound. In the se
ond
ase P (^ut 6= ut ) will depend on the ba
k-sear
h depth. For
ea
h of these parameters the bit error probability has to be
al
ulated individually. All
results that we dis
uss in the present paper
an also be extended to these
ases.
We
onsider the memory m = 2
onvolutional
ode, whi
h is generated by the generator
matrix G(D) = (1 + D + D2 ; 1 + D2 ).2 A
onvolutional
ode
an be en
oded by dierent
generator matri
es, and ea
h generator matrix
an be realized in many ways. Throughout
this paper we assume that the en
oder is realized in
ontroller
anoni
al form [10℄. While
the distan
e properties and the burst (event) error probability are mere
ode properties,
the bit error probability, Pb , does depend on the mapping from the information sequen
es
to the
ode sequen
es and, hen
e, is an en
oder property. We investigate three dierent
en
oders for the
ode, des
ribed by the generator matri
es
1 + D2 1 + D + D2
GA (D) = 1 + D + D2 ; 1 + D2 G B ( D ) = 1; G C (D ) = ;1 :
1 + D + D2 1 + D2
(1)
GA (D) denes a non-systemati
polynomial en
oder, GB (D) and GC (D) dene systemati
en
oders with feedba
k, often referred to as re
ursive systemati
en
oders. If the order of
the
olumns in the generator matri
es, i.e. g1 (D) and g2 (D), is
hanged, one obtains an
equivalent
ode and the bit error probability Pb is un
hanged.
The remainder of this paper is organized as follows. In Se
tion II we introdu
e a
Markov
hain to derive the probability distribution of the
umulative Viterbi metri
s.
Although the prin
ipal steps in this se
tion are already known from [8℄, the
ontent and
our notation will be essential for the understanding of the following se
tions. In Se
tion III
we show how the bit error probability
an be derived for a polynomial en
oder. When the
maximum-likelihood path is not unique the bit error probability will also depend on the
rule the de
oder uses to
hoose among dierent survivor paths. We analyze and
ompare
two dierent de
oding rules in Se
tions III-A and III-B. The previous investigations in
[3℄[4℄[5℄[8℄ were restri
ted to polynomial en
oders without feedba
k. In Se
tion IV we
2 sin
e the generator polynomials
an be represented by the o
tal numbers 7 and 5 this
ode is sometimes referred
to as (7,5)
ode.
O
tober 21, 2002 DRAFT
4
illustrate how the proposed model
an be extended to also take into a
ount re
ursive
systemati
en
oders with feedba
k, whi
h play an important role in iterative de
oding
s
hemes. Then we demonstrate in Se
tion V that the rst de
oding rule, investigated in
Se
tion III-A, is
losely related to
ommon soft-output de
oders and present a way to
derive the bit error probability and the probability distribution of the soft-output values
of the Max-Log-MAP de
oder [11℄. The
al
ulated expressions are reported and dis
ussed
in Se
tion VI and
on
lusions are nally given in Se
tion VII.
When a sequen
e r is re
eived from the
hannel, the ML de
oder sele
ts as its estimate
v^ a
ode sequen
e v whi
h maximizes P (rjv ), i.e.
2
YY 2
XX
v^ = arg max P (rjv ) = arg max P (rt jvt ) = arg max
(i) (i)
log P (rt(i) jvt(i) ) : (2)
v v v
t i=1 t i=1
where q = 1 p, is a
onstant and d(; ) denotes the Hamming distan
e between two
sequen
es. In order to nd the
ode sequen
e v^
orresponding to the minimal metri
= (r) = min
v
d(v ; r) ; (4)
the Viterbi algorithm re
ursively
al
ulates
umulative metri
values by moving forward
through the en
oder trellis. Let t ( ) denote the minimal
umulative metri
up to time t
of all trellis paths that pass at that time through en
oder state . The
omponents of the
ve
tor t+1 = (t+1 (0); t+1 (1); : : : ; t+1 (2m 1))
an be
al
ulated from the
orrespond-
ing ve
tor t of the previous time instant by
t+1 ( ) = min
0
0 (t ( ) + 0 ; (r t )) ;
2S
8 2 S def
= f0; : : : ; 2m 1g ; (5)
where 0 ; (rt ) is the metri
in
rement
orresponding to the trellis transition3 from state
t = 0 to t+1 = . Sin
e the minimization problem is not ae
ted by subtra
tion of
3 The in
rement is dened to be ; (r t ) = 1 if there is no transition between 0 and .
0
t rt = (1; 0) t+1
00 0
t (0) = 0
11 -1
11
t (1) = 1 0
00 0
10 0
t (2) = 1 -1
01
01 2
t (3) = 1
10 -2
Fig. 1. The trellis se
tion at time t,
onsidered in Example 1.
instead of t and keeping in mind that
t (0) = t (0) t (0) = 0 a
ording to this
denition. For given rt and
t at time t the
omponents of
t+1 at the following time
instant are determined by (5) and (6).
Example 1: Let us determine
t+1 , if
t = ( 1; 1; 1) and rt = (1; 0). Figure 1 illus-
trates this situation in the trellis se
tion at time t. The in
rement values 0 ; (rt ) are
equal to 0, 1, 1, and 0 for the transitions labeled by
ode symbols 00, 01, 10, and 11,
respe
tively. The minimal values into the states t+1 = 0; 1; 2; 3 are 1, 0, 1, and 2,
respe
tively, resulting in
t+1 = (1; 0; 1).
A key observation is, that the number of dierent metri
ve
tors
t that
an o
ur at
arbitrary times t when pro
eeding through the trellis is nite. For the simplest
onvolu-
tional
ode with memory m = 1, whi
h is studied in [8℄, there are only ve dierent
t ,
namely the s
alar values 2; 1; 0; 1; 2. In
ase of the
onsidered rate R = 1=2
ode with
memory m = 2,
t has three
omponents taking values between 3 and 3. The total
number of dierent
t is equal to #
= 31. Be
ause of the linearity of the
ode the set of
metri
ve
tors is independent of the transmitted sequen
e.
In the following analysis we assume, without loss of generality, that the all-zero sequen
e
v = 0 is transmitted. A
ording to the
hannel model, the elements of the re
eived se-
quen
e r are Bernoulli distributed random variables. As a
onsequen
e, also the
ompo-
nents of
t are random variables. To obtain the probability distribution of the
umulative
Viterbi metri
s
t at some arbitrary time t we
an, like in [8℄,
onstru
t a Markov
hain
that has the dierent possible values of
t , denoted by
1 ,
2 ,. . . ,
#
, as states. This
Markov
hain is des
ribed by a transition matrix M , where the element mj;k at row j and
olumn k is equal to the probability to
ome from state
t 1 =
j to state
t =
k . This
probability, whi
h is independent of the time t, is given by
X
P (
t =
k j
t 1 =
j ) = P (
t =
k ; r t 1 j
t 1 =
j ) ; (7)
rt 1
= M T ; 1 + 2 + + # = 1 ; (8)
For en
oders with polynomial generator matri
es the set of en
oder states at time t + 1
an be partitioned into two subsets, S0 and S1 ,
orresponding to information symbol
ut = 0 and ut = 1, respe
tively. In a realization in
ontroller
anoni
al form this follows
immediately from the fa
t that the input is fed dire
tly into the shift register dening the
en
oder states. Consequently, the bit error probability Pb = P (^ut 6= ut ) is equal to the
probability, that the ML de
oder
hooses a
ode sequen
e whose path in the trellis passes
at time t + 1 through a state belonging to the wrong subset. For the en
oder des
ribed
by GA (D), this is the
ase if the minimal metri
, given in (4),
orresponds to a path
passing through a state t+1 2 S1 = f2; 3g.
An ML de
oder needs to
onsider
omplete path metri
s to be able to sele
t the estimate
ode sequen
e v^. The
umulative metri
values
t ( ) take into a
ount only the part of
the
ode sequen
es up to time t. For this reason we introdu
e
orresponding values
~t ( )
O
tober 21, 2002 DRAFT
7
for the remaining part of the paths after time t, and obtain
def
t ( ) =
t ( ) +
~t ( ) = vmin d(v ( 1;t 1℄ ; r( 1;t 1℄) + vmin d(v [t;1) ; r[t;1)) ; (9)
:t = :t =
where denotes an oset
aused by the normalization in (6). Then t ( ) is the minimal
omplete metri
of all paths that pass at time t through en
oder state t = . The value
~t ( ), denoting the minimal metri
of all paths that start at time t in state t = ,
an
be
al
ulated by a ba
kward re
ursion, equivalent to (5) and (6) in reverse dire
tion. The
probability distribution of
~ t = (~
t (1);
~t(2); : : : ;
~t (2m 1))
an be obtained with the
method des
ribed for
t in Se
tion II if the transition matrix M is transposed. In
ase
of the
onsidered memory m = 2
ode we
an make use of some symmetry properties:
moving in reverse dire
tion through the en
oder state diagram is equivalent to ex
hanging
states = 1 and = 2. Hen
e,
P
~ t = (~
1 ;
~2 ;
~3) = P
t = (~
2 ;
~1 ;
~3 ) ; (10)
i.e. the distribution of
~ t follows dire
tly from the distribution of
t , derived in the previous
se
tion, and does not require any additional
al
ulations. It
an be seen in (9) that the
ve
tors
t and
~ t are
ompletely independent, as they result from non-interse
ting parts
of the
ode sequen
es. To obtain the distribution of t we
an go through all
ombinations
of
t and
~ t and
ompute
X
P( t = )= P (
t =
)P (
~t =
~ ) : (11)
+
~=
methods involve randomness to assure that the bit error probability is independent of the
transmitted sequen
e.
Even though the trellis bran
hes, sele
ted by the
onsidered de
oder at ea
h time t,
always lie on some ML path, they not ne
essarily form one
ontinuous trellis path. Stri
tly
speaking, the algorithm denes thus a symbol-by-symbol de
oder rather than a sequen
e
de
oder. Sin
e this
an only happen if the survivor is not unique, we still
onsider this
de
oder as a Viterbi (ML) de
oder.
where the path splits into two possible dire
tions, one of them is sele
ted at random. This
rule is sometimes
alled the
oin-
ip tie-breaking rule. It is worth to mention, that a
de
oder based on this rule is dierent from a de
oder that
hooses among all maximum
likelihood sequen
es with equal probability.
The analysis of this algorithm would be the same as for de
oding rule 1, if the sele
ted
survivor path would pass at time t + 1 with equal probability through all surviving states,
i.e. en
oder states with minimal t+1 ( ). In reality these states are passed with dierent
probabilities, depending on the path sele
tion rule, and to determine if a bit error o
urred
at time t, it is no longer suÆ
ient, like in Se
tion III-A, to observe the ve
tor t+1 . Some
additional knowledge is required about the survivor path, sele
ted by the de
oder. Re
all,
that a bit error at time t o
urs, if the sele
ted survivor path passes through a state
t+1 2 S1 . Hen
e, Pb
an be expressed as
X XX
Pb = P (ts+1 = ) = P (ts+1 = j
t+1 =
)P (
t+1 =
) ; (13)
2S1 2S1
where P (ts = ) denotes the probability, that the sele
ted survivor path passes at time t
through state . The
onditioned probabilities P (ts+1 j
t+1 )
an be
al
ulated re
ursively.
For any = 0; : : : ; 2m 1 and
i , i = 1; : : : ; #
, we
an write
X
P (s = j
=
i ) = P (s = j
=
i ; r )P (r ) : (14)
r
The probability P (s = j
; r )
an be expressed as a fun
tion of the probabilities
P (s +1 j
+1 ). As mentioned before, for given r and
at time > t, the ve
tor
+1 is
determined by (5) and (6).
Example 2: Consider the
ase
= ( 1; 1; 1), r = (1; 0) and
+1 = (1; 0; 1),
whi
h is illustrated in Figure 2. The sele
ted survivor path will pass one of the transitions
( ; +1 ) = (1; 0), (1; 2), (2; 1), (3; 1), (3; 3). Observe, that there is a tie at state +1 = 1.
From this follow the probabilities
P (s = 0j
; r ) = 0
P (s = 1j
; r ) = P (s +1 = 0j
+1 ) + P (s +1 = 2j
+1 )
1
P (s = 2j
; r ) = P (s +1 = 1j
+1 )
2
1
P (s = 3j
; r ) = P (s +1 = 1j
+1 ) + P (s +1 = 3j
+1 ) :
2
O
tober 21, 2002 DRAFT
10
r = (1; 0)
+1
00 1
(0) = 0
+1 (0) = 0
11 0
11
(1) = 1 1
+1 (1) = 1
00 1
10 1
(2) = 1 0
+1 (2) = 0
01
01 3
(3) = 1 10 -1
+1 (3) = 1
Corresponding expressions
an be obtained for other
, r and
+1 and substituted into
(14). We will obtain a system of equations of the form
m 1 #
2X X
P ( s
= j
=
) =
k
0 ;
j P (s +1 = 0 j
+1 =
j ) ; (15)
0 =0 j =1
to obtain the probabilities P (ts+1 j
t+1 ). The bit error probability follows nally from (8)
and (13).
The equation system is a generalization of equations (14) and (15) in [8℄ for a memory
m = 1 en
oder. It's nature is similar to that of system (8) for the steady state distribution
of the Markov
hain in Se
tion II, but (15), (16) are more diÆ
ult to solve (124 instead of
31 independent equations for the memory m = 2
ode). Equations (15) and (16) do not
depend on the information mapping in the en
oder trellis and are therefore invariant over
the three en
oders given in (1).
O
tober 21, 2002 DRAFT
11
When the generator matrix is not polynomial but
ontains rational generators, like
GB (D) and GC (D), there exists a feedba
k loop in the realization and as a
onsequen
e
the information bit does not dire
tly enter the shift register. Then the set of en
oder states
annot be partitioned into subsets
orresponding to ut = 1 and ut = 0, as we assumed in
Se
tion III, and it is no longer possible to determine if a bit error o
urs by only knowing
the state t+1 whi
h is passed by the sele
ted survivor path.
Assume we know rt and
t and
al
ulate the resulting ve
tor
t+1 using (5) and (6).
Let us now introdu
e an auxiliary ve
tor it+1 = (it+1 (0); it+1 (1); : : : ; it+1 (2m 1)), where
ea
h element it+1 ( ) 2 f0; 1; g denotes the value of information symbol ut
orresponding
to the surviving bran
h into state t+1 = . If there is a tie, this value might not be
spe
ied uniquely. In this
ase it+1 ( ) is set to ''. Sin
e for polynomial en
oders the
information bit ut determines the state t+1 the information assignment is
onstant and
has the form it+1 = (0; : : : ; 0; 1; : : : ; 1). For en
oders with feedba
k it+1 is dierent for
dierent transitions
t ; rt !
t+1 . The probabilities P (
t+1 ; it+1 j
t )
an be obtained in
a similar way as before (
f. (7)). Then we
an
al
ulate the joint probability for the pair
t+1 ; it+1 ,
X
P (
t+1 ; it+1 ) = P (
t+1 ; it+1 j
t )P (
t ) ; (17)
t
where
t is distributed a
ording to the steady distribution, following from (8). In this
manner the states of the Markov
hain are, at time t + 1, split a
ording to the dierent
possible information assignments it+1 . Like before, we
an
ombine the pre
eding and
forth
oming parts of the trellis paths and obtain
X
P( t+1 = ; it+1 ) = P (
t+1 =
; it+1 )P (~
t =
~ ) ; (18)
+~
=
and the bit error probability Pb is equal to the sum of P ( t+1 = ; it+1 = i) over those
and i that indi
ate an error.
Let us rst
onsider the de
oding rule from Se
tion III-A. We dene the sets M and Me
in the same way as before, however, Me depends now on both and i, i.e. Me = Me ( ; i).
Furthermore, we introdu
e the subset M ( ; i) of M( ), whi
h
orresponds to states
with symbols i that, due to ties, are unspe
ied and marked by ''. Then the probability
O
tober 21, 2002 DRAFT
12
that a bit error o
urs,
onditioned on t+1 = and it+1 = i, is equal to the fra
tion
( ; i) = jMe j+1
jMj
=2jM j
, and the bit error probability is given by
X
Pb = ( ; i)P ( t+1 = ; it+1 = i) : (19)
;i
The same ideas
an be applied to the se
ond de
oding rule from Se
tion III-B. We
obtain
X
Pb = P (^ut 6= ut j
t+1 =
; it+1 = i)P (
t+1 =
; it+1 = i) ; (20)
;i
where
X 1 X
P (^ut 6= ut j
t+1 ; it+1 ) = P (ts+1 = j
t+1 ) + P (ts+1 = j
t+1 ) ; (21)
2S1 (it+1 )
2 2S (i
t+1 )
and S1 (i), S (i) denote the sets of en
oder states for whi
h i( ) = 1 and i( ) = ,
respe
tively.
These derivations of the bit error probability Pb for
onvolutional en
oders with rational
generator matri
es in
lude the polynomial en
oders,
onsidered in Se
tion III, as a spe
ial
ase.
V. Soft-Output De oders
In situations where the de
oder output is further pro
essed, whi
h is the
ase in iterative
de
oding s
hemes, the performan
e
an essentially be improved with an algorithm that
delivers soft output information. The optimal reliability values are obtained by a posteriori
probability (APP) de
oders, whi
h are usually realized by a forward-ba
kward (BCJR)
algorithm, proposed by Bahl et.al. [14℄. Unfortunately, the de
oding
omplexity of APP
de
oders is mu
h larger than for Viterbi de
oding. Be
ause of this fa
t, suboptimal soft
output de
oding algorithms are very often used in pra
ti
e. One su
h algorithm is the
Max-Log-MAP algorithm [11℄, whi
h originally was introdu
ed as an approximation to
the BCJR algorithm. On the other hand, this algorithm
an be shown to be equivalent
to the modied soft-output Viterbi algorithm (SOVA) [12℄, a modi
ation of the Viterbi
algorithm that delivers soft-output values.
Aside from the redu
ed
omplexity
ompared to optimal APP de
oders, the similarity
to the Viterbi algorithm makes the Max-Log-MAP algorithm also simpler to analyze. For
an analysis of the iterative de
oding performan
e of turbo
odes, given in [15℄, a max-path
algorithm was
onsidered, whi
h delivers the same output as the Max-Log-MAP algorithm,
but operates with probabilities instead of log-likelihood ratios.
Using the results from Se
tion III-A, it is possible to derive the probability distribution
of the soft-output values of the max-path (Max-Log-MAP) algorithm. The max-path
de
oding algorithm
al
ulates for ea
h time t the probability of the most probable trellis
path,
orresponding to ut = 0, and the probability of the most probable trellis path
orresponding to ut = 1. The de
ision on symbol ut is then made in favor of the more
probable trellis path. Let us express the output of the algorithm in terms of the log-
likelihood ratios
maxv:ut=0 P (rjv )
t = log = max log P (rjv ) max log P (rjv ) : (22)
maxv:ut=1 P (rjv ) v:ut=0 v :ut =1
t = vmax log P (rjv ) vmax log P (rjv ) = min t+1 ( ) min t+1 ( ) ;
:u =0
t :u =1 t 2S0 (it+1 )[S (it+1 ) 2S1 (it+1 )[S (it+1 )
(23)
where S0 (i), S1 (i) and S (i) denote the sets of en
oder states for whi
h i( ) = 0,
i( ) = 1 and i( ) = , respe
tively4 . With (23) and (18) the probability distribution of
the soft-output values of the max-path (Max-Log-MAP) algorithm be
omes
X
P (t = ) = P( t+1 = ; it+1 = i) ; (24)
( ;i)=
where
( ; i) = min ( ) min ( ) : (25)
2S0 (i)[S (i) 2S1 (i)[S (i)
From (22) follows dire
tly, that the hard de
ision output of this algorithm is the same as
for the standard Viterbi algorithm, provided that the maximum-likelihood path is unique.
If there exist several survivor paths that disagree upon symbol ut , the log-likelihood ratio
is equal to t = 0 (minimal reliability). In this
ase, a de
ision is usually made for u^t = 0
4 Equation (23) is true if the base of the logarithm equals q=p. Otherwise a
onstant fa
tor will appear in the
term on the right hand side.
and u^t = 1 with equal probability, sin
e t = 0 delivers no information about the symbol
ut .
The bit error probability Pb of the max-path (Max-Log-MAP) de
oder
an be
al
ulated
with equation (19), when ( ; i) is set to
8
>
>1; if ( ; i) < 0
>
>
<
( ; i) = 1=2; if ( ; i) = 0 (26)
>>
>
>
:0; otherwise.
Surprisingly, the resulting bit error probability is slightly higher than the ones of the
Viterbi algorithm, given in (12),(13),(19) and (20). Consequently, for transmission over
the BSC, the max-path (Max-Log-MAP) de
oder is worse than the Viterbi de
oder in
terms of the bit error probability. This ee
t, whi
h is
hara
teristi
for a dis
rete
hannel,
is theoreti
ally interesting and
an be
onrmed by
omputer simulations.
Applying the methods des
ribed in Se
tions II{IV we evaluated the bit error probabili-
ties as fun
tions of the
rossover probability p of the BSC for the en
oders with generator
matri
es GA,GB and GC , given in (1). We derived the probabilities for the two de
oding
rules, introdu
ed in Se
tion III-A and III-B, and for the Max-Log-MAP de
oder. The
resulting expressions are rational fun
tions in the variable p. Figure 3 shows the
urves of
these exa
t bit error probabilities for all three en
oders. Results of
omputer simulations
with the se
ond (
oin-
ip) de
oding rule are also presented and mat
h perfe
tly with the
orresponding
urves. Through the
omplete range of the
hannel
rossover probability p
the se
ond de
oding rule results in a slightly better bit error performan
e than the rst
de
oding rule or the Max-Log-MAP de
oder. However, the dieren
e is so small that the
urves are diÆ
ult to distinguish in the gure. We want to stress on
e more, that all the
onsidered de
oding rules ensure that the derived probabilities are valid independently
of the transmitted sequen
es, even though the all-zero sequen
e is used as referen
e in
the analysis. The dependen
e of the performan
e on the de
oding rules raises the still
unsolved question how to nd an optimal survivor sele
tion rule, giving the best possible
result. Note also, that the Viterbi algorithm in general is not optimal in the sense of
−1
10
−2
10
Pb
B
PSfrag repla
ements 10
−3
−4
10 A
C
−5
10 −1
10
p
Fig. 3. The exa
t bit error probabilities Pb for the en
oders with generator matri
es GA , GB and GC for
the rst (dashed) and se
ond (solid) de
oding rule and the Max-Log-MAP de
oder (dotted). The points
marked by dots
orrespond to simulation results.
k fk hk k fk hk
0 6 35 -909044077180391296 313229432403674880
1 -72 36 2417988471991882560 -275042638680130944
2 768 37 -5330207094877951104 -153021977548253952
3 265 -4872 38 10295173207183127104 1410859165841021568
4 -351 27084 39 -17867563322317690624 -4097687548943430144
5 -1981 -114288 40 28247006369981566976 8851064408146699776
6 43537 426132 41 -41013528951384722432 -16114717950327183360
7 -285281 - 1347828 42 54981081370326788864 25848346636450576896
8 1299366 3538206 43 -68285806284195823616 -37298529692365197312
9 - 4959709 -7283652 44 78751150567728106496 48973288757912229888
10 17059854 7979868 45 -84449551149776898048 -58907722481064726528
11 -58312320 23650476 46 84269456602031430656 65181387751386765312
12 221005066 -209292570 47 -78263767733638082560 -66509732270121345024
13 -920731628 967965564 48 67632157238602244096 62665601363281059840
14 3823571126 -3557957562 49 -54343652041446424576 -54544451406933688320
15 - 14727932596 11231234940 50 40557700180785000448 43846560611492241408
16 51078738901 -30859274556 51 -28072204850764808192 -32523561374632574976
17 -157943395547 72470887548 52 17985550123648270336 22227746940363177984
18 434026993772 -135511955922 53 -10640924019658096640 -13967785331540951040
19 -1057453043540 146188834860 54 5796710625752629248 8048655559373586432
20 2268958822042 247858143180 55 -2897387685058379776 -4238512129990066176
21 -4182839660275 -2238249813240 56 1323178513955356672 2031363725532463104
22 5987394598045 9187203123972 57 -549289584889954304 -881523356935913472
23 -2925500798458 -29356207455480 58 205999648093765632 344231188706623488
24 - 24406816376597 80346002993016 59 -69266281384116224 -120038930264358912
25 143941851696404 - 192777484284072 60 20686532953243648 37029049345769472
26 -563599817927379 400411866456582 61 -5422810214170624 -9984956763734016
27 1832188505707872 -676541843021712 62 1228899082043392 2317917521707008
28 -5164854573769100 722404687735464 63 -235938273820672 -453967607758848
29 12680943146670664 556270890548112 64 37324322766848 72954817806336
30 -26697619822281292 -6051612119253816 65 -4671850676224 -9238474653696
31 45968866059359024 21818446713585504 66 433858805760 864496386048
32 -55089943581085088 -57753599060602272 67 -26575110144 -53150220288
33 2281961809544704 124682392316170368 68 805306368 1610612736
34 237382529550233168 -222544666659852960
TABLE I
The
oeffi
ients of the polynomials f (p) and h(p) in the numerator and denominator of
Pb (p) of en oder A.
formulas, whi
h
an be useful if one is only interested in lower ranges of the
hannel
rossover probability p, where the terms of lower order dominate. The rst few
oeÆ
ients
of the series expansion
an be found in Table II. The
oeÆ
ient in the lowest order term
of the series expansion determines the asymptoti
behavior when p tends to zero. For very
small p it is known that the union bound
1 X
X 1
Pb < i n(w; i) Pw ; (27)
w=dfree i=1
delivers a good approximation of the bit error probability. Here dfree denotes the free
distan
e of the
ode, n(w; i) denotes the number of paths with total weight w and infor-
mation weight i, and Pw is the pairwise error probability, that a path with weight w is
O
tober 21, 2002 DRAFT
17
de
oding rule 1
265 p3 943 p4 651 p5 42277 p6 456331 p7
A 6 + 2 2 2 6 +O p8
de oding rule 2
Max-Log-MAP de
oder
89 p3 753 5 151489 7
A 2 + 490 p4 2 p 22210 p6 2 p +O p8
35057 p6 33637 p7
B 82 p3 + 262 p4 2743 p5 2 + 2 +O p8
143 p3 4049 5
C 2 + 301 p4 2 p 19226 p6 614 p7 + O p8
TABLE II
Series expansion of the bit error probabilities for the different en
oders and de
oding
rules.
more probable than the transmitted all-zero path. It is
ommon to upper bound the pair-
p
wise error probability by Pw < W w with the Bhatta
haryya parameter W = 4p(1 p).
Using the weight enumerator fun
tion of the en
oder the union bound (27)
an then be
expressed in a
losed form (Viterbi bound) [16℄. A tightened version of this bound
an
be obtained by making use of the fa
t that P2i 1 = P2i , i 1 (van de Meeberg bound)
[17℄. For the
onsidered
ode, only paths of weight w = dfree = 5 and w = 6
ontribute to
the leading
oeÆ
ient of the series expansion. Therefore, to estimate this
oeÆ
ient the
exa
t pairwise error probability Pw
an be used instead of the Bhatta
haryya bound. The
estimates resulting from the dierent variants of the union bound (27) are shown in Ta-
ble III. A
omparison of Table II and Table III shows, that the union bounds are suÆ
ient
for an estimation of the order of the leading term but do not give the exa
t value of the
oeÆ
ient itself. It
an also be seen that the van de Meeberg bound and even more the
different en oders.
Viterbi bound suer from not using the exa
t values of the pairwise error probabilities Pw .
In the region of high
rossover probabilities p it is mu
h harder to nd tight bounds on
the performan
e of a
onvolutional
ode [18℄. Observe that in this region the polynomial
en
oder, whi
h performs best for low values of p, is worse than the re
ursive systemati
en
oders. It has been shown in [19℄ that also in the general
ase, as the
hannel gets
worse, re
ursive systemati
en
oders outperform polynomial en
oders. This fa
t makes
the re
ursive systemati
en
oders very attra
tive in iterative de
oding s
hemes, where the
omponent
odes have to
ope with
hannels
lose to or even beyond their
apa
ity limit.
In prin
iple all the methods
an be applied to
odes with higher memory. Unfortunately,
the
omplexity of the
al
ulations grows dramati
ally with the memory of the
omponent
ode and we were not able to derive
orresponding results for
odes with m 3. For
typi
al examples of en
oders with memory m = 3 and m = 4 the number of Markov
hain
states be
omes #
= 471 and #
= 25641. Another extension would be to investigate
soft-de
ision de
oding by
onsidering a dis
rete memoryless
hannel (DMC) with more
than two output values. But also in this
ase the number of metri
ve
tors will in
rease
dramati
ally. A method to analyze a memory m = 2 en
oder for a
ontinuous additive
white Gaussian noise (AWGN)
hannel was proposed in [20℄. The evaluation of the bit error
probabilities, whi
h is also based on a forward and ba
kward re
ursion,
an unfortunately
only be performed numeri
ally and also requires quantization for the representation of the
underlying probability density fun
tions.
Finally we dis
uss some extension of the
onsidered model that we see as an interest-
ing topi
for forth
oming investigations. When reliability information is passed from one
omponent de
oder to another in an iterative de
oding s
heme, it is important that the
de
oding algorithm
an make use of a priori information about the symbols to de
ode.
Consider a systemati
en
oder and assume that for information symbols the a priori prob-
ability pa = P (ut = 1) is known. To take this information into a
ount, equation (3)
an
be modied to
qq q
v^ = arg min dinf (r; v ) log a + dpar (r; v ) log ; (28)
v pa p p
where qa = 1 pa , and dinf , dpar denote the Hamming distan
es
ounting only information
or parity-
he
k symbols, respe
tively. This
orresponds to a modi
ation of the Viterbi
de
oder su
h that the metri
in
rement for information symbols is larger than for parity-
he
k symbols. If pa is known, in prin
iple the error probability of this algorithm
an be
al
ulated with the same model as for the analysis of the usual Viterbi de
oder. Unfor-
tunately, the number of dierent Markov
hain states will depend on the ratio between
log(qa q )=(pa p) and log q=p, whi
h makes a general
al
ulation of Pb (pa ; p) as a fun
tion in
two variables impossible, at least with this simple variant of the
onsidered model. If this
ratio is not rational it is not
lear that the number of states #
is nite, and hen
e even
a numeri
al
al
ulation might not be possible for arbitrary
ombinations pa and p. For
an exa
t analysis of iterative de
oding it would be very interesting to nd some modi
a-
tion of this method that
ir
umvents this problem. We see this as a dire
tion for future
resear
h.
We have derived analyti
expressions for the bit error probabilities of dierent rate
R = 1=2 memory m = 2
onvolutional en
oders. The results for
lassi
al en
oders with-
out feedba
k were
ompared to re
ursive systemati
en
oders. All
onsidered en
oders
generate the same
onvolutional
ode but lead to dierent bit error probabilities. The bit
error probability does also depend on the rule the de
oder uses to sele
t among multiple
survivor paths. We
onsidered two dierent de
oding rules. Additionally we derived the
bit error probability and soft-output value distribution of the Max-Log-MAP de
oder. All
these results are expressed as rational fun
tions of the
rossover probability of the BSC.
The resulting expressions are made available online and
an be used as a referen
e.
Referen es
[1℄ A. J. Viterbi, \Error bounds for
onvolutional
odes and an asymptoti
ally optimum de
oding algorithm",
IEEE Trans. Inform. Theory , vol. IT{13, no. 2, pp. 260{269, Apr. 1967.
[2℄ G. D. Forney, Jr., \The Viterbi algorithm", Pro
. IEEE , vol. 61, pp. 268{278, 1973.
[3℄ T. N. Morrissey, Jr., \Analysis of de
oders for
onvolutional
odes by sto
hasti
sequential ma
hine methods",
IEEE Trans. Inform. Theory ,vol. IT-16, pp. 460{469, July 1970.
[4℄ T. N. Morrissey, JR., \A Markovian analysis of Viterbi de
oders for
onvolutional
odes", in Pro
. Nat. Ele
-
Trans. Inform. Theory , vol. IT-24, no. 5, pp. 552{562, Sept. 1978.
[7℄ A. R. Calderbank, P. C. Fishburn, A. Rabinovi
h, \Covering properties of
onvolutional
odes and asso
iated
latti
es", IEEE Trans. Inform. Theory , vol. IT-41, no. 3, pp. 732{746, May 1995.
[8℄ M.R. Best, M.V. Burnashev, Y. Yanni
k, A. Rabinovi
h, P.C. Fishburn, A.R. Calderbank, and D.J. Costello,
Jr., \On a te
hnique to
al
ulate the exa
t performan
e of a
onvolutional
ode", IEEE Trans. Inform. Theory ,
vol. IT{41, no. 2, pp. 441{447, Mar
h 1995.
[9℄ L. Wei, T. Aulin, and H. Qi, \On the ee
t of trun
ation length on the exa
t performan
e of a
onvolutional
ode", IEEE Trans. on Inform. Theory , vol. 43, no. 5, pp. 1678{1681, Sept. 1997.
[10℄ R. Johannesson and K. Sh. Zigangirov, Fundamentals of Convolutional Coding , IEEE Press, Pis
ataway, N.J.,
1999.
[11℄ P. Robertson, E. Villebrun, and P. Hoeher, \A
omparison of optimal and sub-optimal de
oding algorithms
in the log-domain", Pro
. ICC , Seattle, WA, pp. 1009{1013, June 1995.
[12℄ M. Fossorier, F. Burkert, S. Lin, and J. Hagenauer, \On the equivalen
e between SOVA and Max-Log-MAP
de
odings", IEEE Communi
ations Letters , vol. 2, pp. 137{139, May 1998.
[13℄ J. Hagenauer and P. Hoeher, \A Viterbi algorithm with soft-de
ision outputs and its appli
ations", Pro
.