Você está na página 1de 45

Asymptotical Analysis and Comparison of

Two Coded Modulation Schemes Using PSK


Signaling { Part II
Gerd Beyer, Karin Engdahl (Student Member, IEEE),
Kamil Sh. Zigangirov (Fellow, IEEE)y
April 23, 2001

Abstract{ In this, the second part of the paper we analyze a multilevel coded
modulation system using PSK signaling and transmitting over the AWGN
channel. We reduce the analysis of the overall system to the analysis of each
decoding level, considering them as a set of related channels. The concept of
a \genie aided communication system" is introduced to simplify the analysis.
A comparison with the conventional coded modulation system, studied in the
rst part of this work, and some results concerning the synthesis and analysis
of practical multilevel modulation schemes concludes the paper.

Keywords{ Coded modulation, Multilevel modulation, multistage decoding,


PSK-signaling.

 The material of this paper was presented in part at ISIT 2000, Sorrento, Italy, June
25-30 2000.
y The authors are with the Dept. of Information Technology, Lund University, Box 118,
SE { 221 00 Lund, Sweden.

1
I Introduction
In the rst part of this paper [1] we studied conventional coded transmission
over the AWGN channel with q-PSK signaling, q = 2L, L = 1; 2; :::. We
constructed codes (block or convolutional) over the integer ring Zq modulo
q, and then we mapped the q-ary code symbols onto symbols of the alphabet
S (I.1) .
1

In this part of the paper we study multilevel coded modulation systems,


introduced independently by Ungerbck [2] and Imai and Hirakawa [3].
The multilevel scheme enables the usage of the multistage decoder, which
demonstrates performance/complexity advantage over the maximal likeli-
hood decoder of the corresponding conventional coded modulation scheme.
An asymptotical analysis of the multilevel modulation system with multi-
stage decoding and a comparison of this system to the conventional system,
is performed in this part of the paper.
Following Huber and Wachsmann [4],[5] we consider the multilevel sys-
tem as a set of related channels and analyze each channel separately. The
most critical part of this paper is the analysis of decoding error propagation
between the levels. This problem is usually solved by introducing ideal in-
terleaver/deinterleaver pairs at each decoding level [6], or by estimation of
the conditional probabilities of decoding error, given that the decoding on
the previous levels was correct [7]. Our solution to the problem is related to
the second way, but, since the calculation of a conditional error probability,
given that the decoding on previous levels was correct is impossible in most
cases, we operate with conditional probabilities of decoding error at each
level, given that the transmitted sequences at the previous levels are known
to the receiver. Introducing the concept of a \genie aided multilevel sys-
tem" permits us to get asymptotically rigorous bounds for the overall error
1We will refer to the formulas from the rst part of this paper by adding the Roman
number I.

2
probability.

II Multilevel Block Coded PSK signaling


To construct a multilevel rate R = K=N (bits per channel use) length N
block code for the AWGN channel with q-PSK signaling, we rst partition a
binary information sequence of length K
u = u ; u ; :::; uK ;
0 1 1 uk 2 GF(2) (1)
into L subsequences u ; u ; :::; u L of lengths K ; K ; :::; K L , such that
(1) (2) ( ) (1) (2) ( )

K + K + ::: + K L = K . Then each of the subsequences u l , l = 1; 2; :::L,


(1) (2) ( ) ( )

is encoded by a binary rate R l = K l =N length N component block code,


( ) ( )

Figure 1. Let
v l = v l ; v l ; :::; vNl ;
( ) ( )
0
( )
1
( )
1
l = 1; 2; :::; L ; (2)
be the code sequence of the l-th component code. The set of L bits
v L ; v L ; :::; v ;
( ) ( 1) (1)
n n n

one bit from each code sequence v L ; v L ; :::; v , is synchronously mapped


( ) ( 1) (1)

onto a signal point sn 2 S . The sequence


s = s ; s ; :::; sN ;
0 1 1 sn 2 S ;
is transmitted over an AWGN channel with q-PSK signaling. The trans-
P
mission rate (in bits/channel use) is equal to R = Ll R l . The received ( )
=1

sequence r (I.3) is a sum of the transmitted sequence s and the sequence e


of complex Gaussian noise components.
The optimal multistage decoder, Figure 1, consists of a set of L decoders
matched to the codes, used at the corresponding levels of encoding. Each
decoding stage includes a calculation of distances (metrics) to the received
sequence r = s + e from all possible code words on this decoding level. The
3
side information from the previous stages determines, according to the set
partitioning structure, illustrated in Figure 2 for L = 3, the signal set upon
which the metrics are calculated.
To calculate the metrics, knowledge of the apriori distribution of the code
symbols vnl and, correspondingly, the apriori distribution of sn is needed.
( )

As we noted in the rst part of this paper, the capacity C of the AWGN
channel with q-PSK signaling (I.5) is reached for uniform distribution of the
transmitted symbols over the input alphabet S . The symbols of the code
word should also be independent. Analogously, the Gallager function of the
AWGN channel with q-PSK signaling, de ned by formula (I.9), is maximized
for a uniform distribution over S . Consequently, the symbols vnl of the code ( )

words v l , l = 1; 2; :::; L, should be i.i.d. equiprobable binary symbols.


( )

We consider the original AWGN channel with q-PSK signaling as a set


of L related channels. The channel corresponding to the rst level can be
treated as a memoryless channel with binary inputs vn 2 f0; 1g, and con- (1)

tinuous outputs rn, n = 0; 1; :::. Let S and S be the subsets of S , (1)


0
(1)
1

corresponding to transmission of vn = 0 and vn = 1 respectively. Since


(1) (1)

the sequences v ; v ; :::; v L are sequences of i.i.d. equiprobable binary


(2) (3) ( )

symbols, the conditional distribution of sn, given vn = 0 (or vn = 1), is (1) (1)

uniform over S (or S ). The conditional probability density function of


(1)
0
(1)
1

rn, given that vn was transmitted at the rst level, is


(1)

2 X  d (r ; s ) 
E n n
f (rn j vn ) = N q
(1) (1)
exp N ; (3)
sn 2 Svn
0 0
(1)
(1)

where dE (rn; sn) is the squared Euclidean distance between rn and sn. The
conditional probability density function of r ;N = r ; r ; :::; rN , given that [0 ) 0 1 1

v ;N = v ; v ; :::; vN was transmitted is


(1)
[0 )
(1)
0
(1)
1
(1)
1

Y
N 1

f (r
(1)
[0 ;N ) jv [0;N ) ) =
(1)
f (rn j vn ) :
(1) (1)
(4)
n=0

4
Let
u^ = u^ ; u^ ; :::; u^N
(1) (1)
0
(1)
1
(1)
1 (5)
and
v^ = v^ ; v^ ; :::; v^N
(1) (1)
0
(1)
1
(1)
1 (6)
be the decoded information and code sequences at the rst level. The se-
quence v^ determines a sequence of subsets Sv^ = Sv ; Sv ; :::; Sv , to
(1) (1)
(1)
def (1)
(1)
^0
(1)
(1)
^1
(1)
^
(1)

which the symbols of the transmitted sequence s belong, if the decoding at


N 1

the rst level was correct, i.e. if v^ = v . The set Svn of the sequence Sv^ (1) (1) (1)
^
(1)
(1)
(1)

de nes the subsets S and S corresponding to transmission of vn = 0


(2)
0
(2)
1
(2)

and vn = 1, respectively. Then the decoded sequence at the second level


(2)

v^ = v^ ; v^ ; :::; v^N
(2) (2)
0
(2)
1
(2)
1
(7)
determines, together with v^ , a sequence of subsets Sv^ = Sv ; Sv ; :::;
(1) (2)
(2)
def (2)
(2)
^0
(2)
(2)
^1

Sv , to which the symbols of the transmitted sequence belong, if the de-


(2)
^
(2)
N
coding at the rst two levels was correct, i.e. if v^ = v and v^ = v .
1
(1) (1) (2) (2)

In general, the decoded sequences v^ ; v^ ; :::; v^ l , de ne a sequence of (1) (2) ( )

subsets Sv^l l = Svl l ; Svl l ; :::; Svl l , to which the symbols of the transmitted
( )
( )
def ( )
( )
^0
( )
^1
( )
^
( )
( )
N
signal belong, if the decoding at the l rst levels was correct. (Strictly speak-
1

ing, the subsets Svlnl are functions not only of v^nl , but also of the symbols
( )
^
( )
( )

v^nl ; v^nl ; :::; v^n , decoded at the previous levels. We omit the correspond-
( 1) ( 2) (1)

ing subscripts to simplify the notation.)


Analogously to (3) we have the following expression for the conditional
probability density function for rn, given that vnl was transmitted (if we ( )

omit the conditions that the transmitted symbols at the previous levels are
vnl ; :::; vn ).
( 1) (1)

l X  d (r ; s ) 
f l
( )
(rn j vnl ( )
)= 2 exp E n n
(8)
N q N
sn 2 Svlnl
0 0
( )
( )

5
The conditional probability density function of r [0;N ) given v l ;N ; v l ;N ; :::;
( )
[0 )
(
[0
1)
)

v ;n is
(1)

Y
[0 ]
N 1

f (r l
( )
[0;N ) )jvl ( )
[0 ;N ) ) = f l (rn j vnl ) :
( ) ( )
(9)
n=0
The logarithm of the probability density function (9) de nes an optimal
metric for the decoding at the l-th decoding level for the path labeled by the
code sequence v l ;N . It is convenient to normalize it and de ne the distance
( )
[0 )

(metric) between the received sequence r ;N and the code sequence v l ;N as [0 )


( )
[0 )

X
N 1

 l (r
( )
0 [0 )
l
;N ; v ;N ) =
( )
[0 )
 l (rn; vnl ) ;
( )
0
( )
(10)
n=0
where
X  d (r ; s ) 
E n n
l
( )
0 (rn; vnl ( )
) = log 2 exp : N (11)
sn 2 Svlnl
0
( )
( )

The maximum likelihood decoder outputs the sequence that has the minimal
metric.
Given the transmitted sequences at the (l 1)-th, (l 2)-th,... and rst
levels, the channel at the l-th level can be treated as a memoryless channel
with binary input and continuous output. The conditional probability to
receive rn given vnl is given by (8). The value
( )

Z  
C = 2 f l r j v l = 0 log f l (r j v 2l f = (0)r j+v f l=(r0)j v l = 1)
1 l l
~
( ) ( )
l ( ) ( ) ( )
2


( ) ( ) ( ) ( )
r
 2 f l (r j v l = 1)
+ f r j v = 1 log f l (r j v l = 0) + f l (r j v l = 1) dr ;(12)
l( ) l ( )
( ) ( )

2 ( ) ( ) ( ) ( )

where the integration is over the entire complex plane, can be treated as
the e ective capacity of the l-th decoding level of the multilevel system. It
can also be called the maximal achievable rate of the l-th level. The di er-
ence compared to the Shannon capacity consists in that the de nition of the
Shannon capacity includes optimization of the probability distribution of the
transmitted signal over the channel input alphabet. In multilevel modulation
6
schemes the probability distribution of the transmitted signal is usually xed
and chosen to be uniform over the channel input alphabet. In the PSK case
the uniform distribution is optimal, but this is not the case, for example, for
QAM signaling. Therefore the e ective capacity with uniform distribution
over the input alphabet can be less than the Shannon capacity per level. We
will also use the term \e ective capacity" when not only the distribution of
the input signal is xed, but when also the decoding metric is di erent from
the optimal.
The e ective capacity of a multilevel modulation system is de ned as the
sum of the e ective capacities of all levels:
X
L
C~ = ~l C : ( )
(13)
l=1

It can be proved that C~ has the same sense as the Shannon capacity, i.e. for
R < C~ it is possible to achieve reliable communication by letting the code
memory go to in nity and for R > C~ reliable communication is impossible.
We will discuss the proof later. From (I.5), (12) and (13) it follows that for
PSK signaling, see also [4],
C~ = C ; (14)
i.e. the e ective capacity of a multilevel PSK modulation system with opti-
mal multistage decoding coincides with the Shannon capacity of the channel.
To analyze multistage decoding performances of multilevel block codes,
we introduce the concept of a \genie aided communication system" . We 2

will assume that there is a kind genie, that informs the decoder at the l-th
level, which sequences v ; v ; :::; v l were sent on the previous levels.
(1) (2) ( 1)

Actually, we suppose that the genie steps into the decoding process only if
the decoding at the previous levels was erroneous. Then the l-th level channel
output has the conditional probability density function (8). We emphasize
The ideology of a \genie aided communication system" was popular in the 60s, see for
2
example [8]. Now it is practically forgotten.

7
that we operate with conditional probability density functions of the received
signal, given the transmitted sequences at the previous levels, not given that
the decoding at the previous levels was correct. To calculate the last one is
a complicated, if not hopeless problem. We introduce the genie to avoid the
calculation of the conditional density function, given that the decoding was
correct. In this way we reduce the problem of the analysis of the transmission
at the l-th level to the analysis of a conventional binary transmission with a
slightly exotic, symmetrically distributed, continuous output (8).
Obviously, a lower bound for the decoding error probability of the genie
aided communication system is a lower bound also for the decoding error
probability of the original system. This is because the decoder may or may
not use the information from a genie, and hence the system with a genie
can not be worse than a system without it. More surprising is, that the
genie aided communication system has exactly the same probability of block
decoding error (i.e. the probability that the block will have at least one
decoding error) as the system without a genie. The explanation of this is
that if no errors are made, the genie never steps in with its information, and
therefore the system is the same as without a genie. If, on the other hand, an
error is made and the genie steps in, it only helps the decoding on the next
level. The decoding error is already made and the block will be erroneous,
just as it should have been in the case without a genie. I.e. the probability
Pe of a block decoding error is una ected by the presence or absence of the
genie. (The bit error probability Pb is not una ected of course, but it doesn't
matter for asymptotical considerations.)
Similar arguments can be used for nding the e ective capacity of a mul-
tilevel system. Since the genie aided multilevel system has the same decoding
error probability as the original multilevel system, these two systems have
the same e ective capacity, i.e. it is sucient to calculate the level e ective
capacities for the genie aided system. These e ective capacities are given
by (12), and the overall system capacity is equal to the sum of the level
8
capacities (13).
Let Pe l be the probability of a decoding error at the l-th decoding level
( )

for a genie aided system, and let Pb l be the bit error probability at the l-th
( )

level of the same system. Analogously to (I.7)


Pe l  P l  P l ;
( )

l = 1; 2; :::; L ;
( ) ( )
(15)
Kl b ( ) e

and for any nite R l = K l =N , we have


( ) ( )

log P l log P l ( ) ( )
e
EBl
= Nlim
( ) def

!1 N = Nlim
!1 N
b : 2
(16) 2

Let P~e be the block error probability for the multilevel system, i.e. the
probability of at least one decoding error among all L levels. Then
X
L
max
l
Pe  P~e ( ) l Pe l  L  max
( )
l
Pe l : ( )
(17)
l=1
From (17) follows that
E~B = Nlim log P~e = min E l :
def
(18)
2 ( )

!1 N l B

Since the problem of maximizing E~B for a given overall transmission rate R,
is equivalent to the problem of maximizing R for a given E~B , we choose R l , ( )

l = 1; 2; :::; L, such that EB (R ) = EB (R ) = ::: = EBL (R L ) = E~B .


(1) (1) (2) (2) ( ) ( )

P
Then R = Ll R l is the overall transmission rate.
( )
=1

To formulate lower and upper bounds for Pe l we need some additional ( )

de nitions [9]. The Gallager function for the l-th level is (comp. with (I.9))
Z 1 1
 s 1+
l
G (s) = log f (r j 0) s + 2 f (r j 1) s dr ; (19)
l l 1 1
( ) ( ) 1+ ( ) 1+

r 2
2

0 s<1;
where f l (r j ) is de ned by (8) and the integration is over the entire complex
( )

plane. The computational cuto rate is


(1 + z l
o )
( )

Rl ( )
0 =Gl ( )
(1) = log 2 ; 2 (20)
9
where Z 
zol = f l (r j 0)f l (r j 1) dr ;
1
( ) ( ) ( ) 2
(21)
r
is the Cherno bounding parameter for optimal decoding at the l-th level. As
a counterpart to the de nition (I.26) we introduce the expurgation function
for optimal decoding at the l-th level
n h  io ;
Gexl (s) = s 1 log 1 + zol 1<s<1;
1
( ) def
2
( ) s
(22)
and the expurgated error exponent evaluated at R l = 0 ( )

Eexl (0) = slim


( )
G l (s) = 1 log z l : ( ) ( )
(23)
!1 ex 2 o 2

III Bounds for the Block Error Exponent


The following theorems are counterparts of Theorems 1{4 in the rst part of
the paper. We omit the proofs since they are simple reformulations of known
theorems for the channel with binary input and continuous output [10],[11].
Theorem 1 (Sphere-Packing Bound) Consider multilevel transmission
over the AWGN channel with q-PSK signaling. For any block code of rate
R(l) and length N that is used to communicate at the l-th level of a genie
aided system, the decoding error probability is lowerbounded by
n o
EB(l;)sp (R(l) o1 (1))+o2 (1) N
Pe  2 l
( )
; (24)
where
EBl;sp (R l ) = sup G l (s) sR l
( ) ( )
 ( ) ( )
 (25)
s
and G(l) (s) is given by (19). The quantities o1 (1) and o2 (1) go to zero
with increasing N and can be taken as o1 (1) = O1 (log2 N=N ) and o2 (1) =
p
O2(1= N ). 2

10
l can be parametrically repre-
The block sphere-packing error exponent EB;sp ( )

sented as (compare with (I.12))


(El s[G l (s)]0
B;sp = G (s)
( ) (l) ( )

; 0< s< 1: (26)


R(l) = [G(l) (s)]0
Let RBl be the R l at which a straight line in the (R l ; E l )-plane through
( ) ( ) ( ) ( )

the point (0; Eexl (0)) is a tangent to the curve EB;sp


( ) l (R l ). ( ) ( )

Theorem 2 (Straight-Line Bound for Block Codes) Consider multi-


level transmission over the AWGN channel with q-PSK signaling. For any
block code of rate R(l) and length N that is used to communicate at the l-th
level of a genie aided multilevel system, let
l (R(l) ) = E (l) (0) Eexl (0) EB;l sp (RBl ) l
( ) ( ) ( )

EB;sl
( )
ex R : ( )
(27)
RBl ( )

Then, for N ! 1 and any R(l) satisfying


0  R l < RBl ;
( ) ( )
(28)
the decoding error probability Pe(l) is lowerbounded by the inequality
h i
EB;sl
(l)
(R(l) )+o(1) N
Pe l  2
( )
; (29)
where o(1) ! 0, when N ! 1. 2
Theorem 3 (Random Coding Bound for Block Codes) Consider mul-
tilevel transmission over the AWGN channel with q-PSK signaling. For the
l-th level of a genie aided system, there exists a rate R(l) block code of length
N , whose error probability is upperbounded by the inequality
h i
EB;rc
(l)
(R(l) )+o(1) N
Pe < 2 ; (30)
where
l (R l ) = sup G l (s) sR l :
EB;rc
( ) ( )
 ( ) ( )
 (31)
0s1
G l (s) is given by (19) and o(1) ! 0 if N ! 0.
( )
2
11
The random coding error exponent can be represented as (compare with
(I.25))
8 l l
l (R ) =
<R R ; ( ) ( )
0Rl <Rl ; ( ) ( )

EB;rc (32)
( ) 0 0

: EB;sp
l (R l ) ;( ) ( )
R l  R l < C~ l :
( )
0
( ) ( )

Theorem 4 (Expurgated Bound for Block Codes) Consider multilevel


transmission over the AWGN channel with q-PSK signaling. For the l-th level
of a genie aided system, there exists a rate R(l) block code of length N , whose
error probability is upperbounded by the inequality
Pe l < 2 EB;ex (R(l) )N
0  R l < Rexl + o(1) ;
(l)
( )
; ( ) ( )
(33)
where the expurgated block error exponent for the l-th level is de ned by the
parametrical equation
l
EB;ex
( )
= Gexl (s) s Gexl (s) 0 ;
( )
 ( )
 (34)

R l = Gexl (s) 0 + o(1) ;
( ) ( )

(l)
and the expurgation rate Rex is de ned by the equation
( )

Rexl = Gexl (s) 0 s ;
( )
 (35)
=1

o(1) ! 0, when N ! 1. 2
We introduce the following notation for the upper bound of the l-th level
block error exponent
8 l l
< EB;sl(R ) ;
( ) ( )
0  R l < RBl ;
( ) ( )

E Bl (R l ) = : l
( ) ( )
(36)
EB;sp(R l ) ;
( ) ( )
RBl  R l < C~ l
( ) ( ) ( )

Then the upper bound for the overall block error exponent is
l
E~ B (R) = max
l
E B (R l ) ; ( ) ( )
(37)

12
where
X
L
R= Rl ; ( )
(38)
l=1
and R l satis es (36). Analogously, the lower bound of the l-th level block
( )

error exponent is
8 l l
l l
< EB;ex(R ) ;
( ) ( )
0  R l < Rexl ;
( ) ( )

E B (R ) = : l
( ) ( )
(39)
EB;rc(R l ) ;
( ) ( )
Rexl  R l < C~ l ;
( ) ( ) ( )

and the lower bound for overall block error exponent is


E~ B (R) = min
l
E Bl (R l ) : ( ) ( )
(40)

The upper and lower bounds (37) and (40) are optimized, when E Bl (R l ) ( ) ( )

and E Bl (R l ) respectively are constant with respect to l.


( ) ( )

To nd the upper bound for the multilevel block error exponent E~ B (R)
we use the following method. First, for any given value E~ B , nd R l , l = ( )

( ) ( )
P
1; 2; :::; L, such that E~ B = E Bl (R l ). Then Ll R l is equal to the overall
( )
=1

system transmission rate R, corresponding to E~ B . We nd the lower bound


for the multilevel block error exponent E~ B (R) analogously. In Figure 3 the
upper and lower bounds of the block exponents, E Bl (R l ) and E Bl (R l ), are
( ) ( ) ( ) ( )

given for each level, in the case when q = 32, Es=N = 10dB. (We show only 0

the curves for levels l = 3; 4; 5, since R  0 and R  0 at the rst two


(1) (2)

levels.)
In Figure 4 the corresponding bounds E~ B (R) and E~ B (R) for the overall
block error exponent of the multilevel system are given. For comparison
we also present in Figure 4 the upper and lower bounds E B (R) and E B (R),
calculated in the rst part of this paper, for the conventional q-PSK signaling
scheme. One can see that the multilevel block coded modulation scheme with
multistage decoding performs better than the corresponding conventional
scheme.

13
IV Multilevel Trellis-Coded PSK Signaling
In the rst part of this paper we considered conventional trellis coded trans-
mission with q-PSK signaling, where the q-ary information sequence u is
rst divided into subblocks of b symbols in each, then the sequence of q-ary
b-tuples is encoded by a rate R = log (q)b=c (bits per channel use) memory 2

m trellis encoder into the code sequence s = s ; s ; :::; si; :::, where si 2 S . 0 1

In this section we consider the following transmission scheme. A bi-


nary information sequence u is rst partitioned into L binary subsequences
u ; u ; :::; u L . The l-th subsequence u l consists of subblocks of length
(1) (2) ( ) ( )

b l , where b l is integer. This can be done by dividing u into blocks of length


( ) ( )

b = b + b + ::: + b L . Then the rst b symbols form a subblock of u ,


(1) (2) ( ) (1) (1)

the next b symbols form a subblock of u , etc. The subsequences


(2) (2)

u l = u l ; u l ; :::; unl ; ::: = u l ; u l ; :::; uil ; ::: ;


( ) ( )
0
( )
1
( ) ( )
0
( )
1
( )
l = 1; 2; :::; L ; (41)
where
unl = unbl ; unbl
( ) ( )
(l)
( )
(l) +1 ; :::; u nl ( )
( +1) (l) b 1
; uil 2 GF (2)
( )
(42)
are encoded by independent binary component trellis codes, of rates R l = ( )

b l =c l (in bits per code symbol) and memories m l , analogously to the block
( ) ( ) ( )

coding case, Figure 1. Let


v l = v l ; v l ; :::; vnl ; ::: = v l ; v l ; :::; vi l ; ::: ;
( ) ( )
0
( )
1
( ) ( )
0
( )
1
( )
(43)
where
vnl = vncl ; vncl
( ) ( )
(l)
( )
(l) +1 ; :::; v nl ( )
( +1) (l) c 1
; vi l 2 GF (2);
( )
(44)
such that
vnl = f (unl ; unl ; :::; unl m)
( ) ( ) ( )
1
( )
(45)
be the code sequence at the output of the l-th encoder. The set of L bits
n o
vi ; vi L
L
( ) ( 1)
; :::; vi (1)
; (46)
14
is synchronously mapped onto the signal point si 2 S . The sequence
s = s ; s ; :::; si; :::
0 1 (47)
is transmitted over the channel. The transmission rate (in bits per channel
P
use) is equal to R = Ll R l . The optimal multistage decoder is analogous
( )
=1

to the block coding case, Figure 1, except that at each level a Viterbi decoder
is used. An optimal metric is de ned by (10).
As in the block coding case, we will use the ideology of a genie aided sys-
tem for analysis of the trellis-coded multilevel system with multistage decod-
ing. The l-th level of the genie aided system, l = 1; 2; :::; L, is characterized
by the burst error probability P l (burst ), the bit error probability P l (bit )
( ) ( )

and the convolutional reliability function (or convolutional error exponent)


ECl de ned as
( )

ECl (R l ) = lim
( ) ( ) def log P l (burst ) = lim log P l (bit ) : (48)
2
( )
2
( )

m l !1 ( )clml m l !1
( )
clml
( ) ( ) ( ) ( )

As we mentioned in the rst part of this paper, it is dicult to nd


the overall burst and bit error probabilities for a nonterminated multilevel
trellis code. This is because of the propagation of errors from other levels.
Therefore, we consider terminated transmission to simplify the analysis of
the error probability, analogously to the case of conventional trellis coding
considered in the rst part of this paper. Then the error propagation is
limited by the frame length. Let us choose the integer N as the frame length
and the integers  ;  ; :::;  L , such that
(1) (2) ( )

( + m )c = ( + m )c = ::: = ( L + m L )c L = N :
(1) (1) (1) (2) (2) (2) ( ) ( ) ( )
(49)
The encoder of the terminated trellis code divides the information sequence
of the l-th level into frames of  l b l -tuples, each frame is followed by a
( ) ( )

m l dummy zero b l -tuple and is encoded into a code word of length N (in
( ) ( )

binary symbols). Then the set (46) of L bits is mapped onto a symbol si . The
value N is chosen such that  l  m l , l = 1; 2; :::; L. Particularly, we can
( ) ( )

15
( )

choose N = maxl m l =R l . Then the transmission rate R  Ll R l , is
( ) 2
P ( )
=1

approximately equal to the rate of nonterminated trellis code transmission.


The multistage decoder decodes the received sequence framewise. The
Viterbi decoder of the rst level decodes the j -th frame u j ; j  , re- (1)
[ (1) ( +1) (1) )

sulting in an estimation of the information sequence u^ j ; j  and the (1)


[ (1) ( +1) (1) )

transmitted code word at the rst level v^ j ; j  l . Then the Viterbi de- (1)
[ (1) ( +1) ( ) )

coder of the second level uses the side information Sv^ to output (1)
(1)
[j (1) ;(j +1) (1) )

u^ j
(2)
[
and v^ j ; j  and so on. The decoding delay at the l-th
;j 
(2) ( +1) (2) ) [
(2)
(2) ( +1) (2) ]

level with respect to the (l 1)-th level is equal to  l b l (in information ( ) ( )

symbols). The frame error probability at the l-th level is Pe l , and the bit ( )

error probability is Pb l . Analogously to conventional trellis coding


( )

log P l log
( )
P l log P l (burst ) = E l :( ) ( )
e b
lim = lim
2
= lim 2
C
2 ( )

m l !1 c l m l
( ) ( )
m l !1 c l m l
( )
m l !1
( ) clml ( ) ( ) ( ) ( ) ( )

(50)
The asymptotical complexity of the decoding at the l-th level is equal to
 l (the number of states in the Viterbi decoder), the probability of error
( )

  l
is Pe l =  l o , where o(1) ! 0 if  l ! 1 and l is the state-
( ) ( )
( ) + (1)
( ) ( )

complexity error exponent for the l-th decoding level.


The asymptotical overall decoding complexity ~ is de ned as the total
number of states in the Viterbi decoders at all decoding levels, i.e.
X
L
~ = l:
( )
(51)
l=1
Since
max
l
 l < ~  L  max
( )
l
l ( )
(52)
the overall complexity is asymptotically de ned by the maximal complexity of
the decoders of the di erent levels, i.e. ~  maxl  l . The overall frame error ( )

probability P~e is de ned as the probability that at least one of the L decoders
makes an error. Since (17) is valid also for the trellis transmission case, the
overall frame error probability is asymptotically equal to the maximal frame
16
error probability of the levels, i.e. P~e  maxl Pe l . Let us de ne the l- ( )

th level state-complexity error exponent l as ECl (R l )=R l and the overall


( ) ( ) ( ) ( )

multilevel system state-complexity exponent ~ as minl l . Clearly, in order to ( )

optimize the complexity-reliability tradeo we have to choose R l and m l , ( ) ( )

l = 1; 2; :::L, such that the decoding complexity and the state-complexity


error exponents of the di erent levels are the same, i.e.
 l = 2b l m l = ~
( ) ( ) ( )
o
1+ (1)
; l = 1; 2; :::L ; (53)
l = ~ ;
( )
l = 1; 2; :::; L : (54)

V Bounds for the Convolutional Error Expo-


nent
The two following theorems establish lower bounds for the l-th level burst
error probability P l (burst ) for a genie aided communication system using
( )

trellis-coded PSK signaling in the AWGN channel.


Theorem 5 (Viterbi-Type Lower Bound) Consider multilevel transmis-
sion over the AWGN channel with q-PSK signaling. For any trellis code of
rate R(l) = b(l) =c(l) and memory m(l) , that is used to communicate at the
l-th level of a genie aided multilevel system, the burst error probability is
lowerbounded by the inequality
h i
EC;sp (R(l) )+o(1) c(l) m(l)
0  R l < C~ l ;
(l)
l
P (burst ) > 2
( )
; ( ) ( )
(55)
(l)
where EC;sp (R(l)) is the convolutional sphere-packing exponent, de ned by the
parametric equation
l = G l (s)
EC;sp ( ) ( )

0<s<1: (56)
R l = G l (s)=s
( ) ( )

G l (s) is the Gallager function (19) and o(1) ! 0 if m l ! 1.


( ) ( )
2

17
Let RCl be the rate at which EC;sp
( ) l (R l ) = E l (0).
ex
( ) ( ) ( )

Theorem 6 (Straight-Line Bound for Trellis Codes) Consider multi-


level transmission over the AWGN channel with q-PSK signaling. For any
trellis code of rate R(l) = b(l) =c(l) and memory m(l) , 0  R(l) < RC(l) , that is
used to communicate at the l-th level of a genie aided system, the burst error
probability is lowerbounded by the inequality
h i
Eex
( l)
(0)+o(1) c(l) m(l)
P l (burst )  2
( )
; (57)
(l)
where Eex (0) is the zero rate expurgated error exponent (23) and o(1) ! 0 if
m ! 1. 2
The two following theorems establish upper bounds for the l-th level burst
error probability P l (burst ) of a genie aided communication system with
( )

q-PSK signaling in AWGN. They are proved for periodically time-varying


trellis codes [9]. The construction of codes consists of two phases. First, we
construct at each level a periodically time-varying binary convolutional code,
and then we map the L-tuples (46) onto the signal point si 2 S . We assume
that the period T l of the l-th level code increases faster than the memory
( )

m l of the code, for example, T l = log  l .


( )
 ( ) ( ) 2
2

Theorem 7 (Random Coding Bound for Trellis Codes) Consider


multilevel transmission over the AWGN channel with q-PSK signaling. For
the l-th level of a genie aided multilevel system, there exists a rate R(l) =
b(l) =c(l) trellis code of memory m(l) , whose burst error probability is upper-
bounded by the inequality
8
>
<2
h i
R(0l) +o(1) c(l) m(l)
; 0Rl <Rl ( ) ( )

l
P (burst ) < >
( )
0
(58)
:2 h
EC;sp
(l)
i
(R(l) )+o(1) c(l) m(l)
; Rl
( )
0  R < C~ l ;
( )

where R0(l) is the computational cuto rate of the l-th level (20), EC;sp
(l)
(R(l) )
is the convolutional sphere-packing error exponent (56), and o(1) ! 0, if
m(l) ! 1. 2
18
Theorem 8 (Expurgated Bound for Trellis Codes) Consider multi-
level transmission over the AWGN channel with q-PSK signaling. For the
l-th level of a genie aided multilevel system, there exists a rate R(l) = b(l) =c(l),
0  R(l) < R0(l) , trellis code of memory m(l) , whose burst error probability is
upperbounded by the inequality
P l (burst ) < 2
( ) EC;ex
( l)
(R(l) )c(l) m(l)
; (59)
where the expurgated convolutional error exponent is de ned by the paramet-
rical equation
(l) (l)
EC;ex = Gex (s) + o1(1) ;
(l)
1<s<1: (60)
R = Gex (s)=s + o2(1) ;
(l)

(l)
Here Gex (s) is the expurgation function (22) of the l-th level, o1(1) ! 0,
o2 (1) ! 0 when m(l) ! 1. 2
We introduce the following notations for the upper and lower bounds of
the l-th level convolutional error exponent
8 l
l l
< Eex (0) ; 0  R l < RCl ; ( ) ( ) ( )

E C (R ) = : l
( ) ( )
(61)
EC;sp(R l ) ; RCl  R l < C~ l ; ( ) ( ) ( ) ( ) ( )

8 l l
< EC;ex(R ) ; 0  R l < R l ;
( ) ( ) ( ) ( )

E Cl (R(l) )
0
( )
= : EC;sp
l (R l ) ; R l  R l < C~ l :
( ) ( )
(62)
( ) ( ) ( )
0

From (50) and Theorems 5 { 8 follows that the frame error probability
and the bit error probability asymptotically satisfy the bounds
E Cl (R l )  lim
( ) ( ) Pe l = lim Pb l = E l (R l )  E l (R l ); l = 1; 2; :::; L:
( ) ( )
( ) ( ) ( ) ( )

clml clml
( ) C ( ) C ( ) ( )

(63)
From (53){(62) also follows that
8E
l l l l l
< R (l)
ex (0)
; 0  R l < RCl ;
( ) ( )

(R ) = E C (R )=R = :
( l)
( ) ( ) def ( ) ( ) ( )
(64)
sl ( )
; RCl  R l < C~ l ;
( ) ( ) ( )

19
where s l satis es the equality R l = G l (s l )=s l and
( ) ( ) ( ) ( ) ( )

l (R l ) = E Cl (R l )=R l = s l ;
( ) ( ) def ( ) ( ) ( ) ( )
(65)
where s l satis es the equality R l = Gexl (s l )=s l for 0  R l < R l and
( ) ( ) ( ) ( ) ( ) ( ) ( )
0

R = G l (s l )=s l for R l  R l < C~ l .


( ) ( ) ( ) ( )
0
( ) ( )

To nd the bounds for the overall state-complexity error exponent we use


the same method, as we used for nding the bounds for the overall block
error exponent. To ful ll conditions (53) and (54) we choose R l and m l ( ) ( )

such that
b m = b m = ::: = b L m L ;
(1) (1) (2)
(66) (2) ( ) ( )

and such that


E C (R )c m = E C (R )c m = ::: = E CL (R L )c L m L ;
(1) (1) (1) (1) (2) (2) (2) (2) ( ) ( ) ( ) ( )
(67)
if we are interested in the upper bound of the overall state-complexity error
exponent ~ or
E C (R )c m = E C (R )c m = ::: = E CL (R L )c L m L ;
(1) (1) (1) (1) (2) (2) (2) (2) ( ) ( ) ( ) ( )
(68)
if we are interested in the lower bound. From (64), (66) and (67) we have
(R ) = (R l ) = ::: = L (R L ) = ~ :
(1) (1) (2) ( ) ( ) ( ) def
(69)
The sum of the rates R l , that satisfy (69), gives the overall transmission
( )

rate, corresponding to the upper bound ~ of the overall state-complexity


exponent ~. Analogously, from (65), (66) and (68) we have
(R ) = (R ) = L (R L ) = ~ :
(1) (1) (2) (2) ( ) ( ) def
(70)
The sum of the rates R l , that satisfy (70), gives the overall transmission rate
( )

corresponding to the lower bound ~ of the overall state-complexity exponent


~.
In Figure 5 the curves l (R l ) and l (R l ), l = 2; 3; 4; 5, for q = 32 and
( ) ( ) ( ) ( )

Es=N = 10dB are given (for l = 1, R  0). Using the principle described
0
(1)

20
above, we get upper and lower bounds for the overall state-complexity error
exponent from these curves, and overall transmission rate R = Ll R l .
P ( )
=1

The result ~ = ~(R) and ~ = ~(R) is presented in Figure 6, where the bounds
(R) and (R) for conventional transmission over the AWGN channel with
q-PSK signaling are also given for comparison.

VI Suboptimal Multistage Decoding


In the previous section, we analyzed the multistage decoder which uses a
Viterbi decoder with optimal decoding metric at each decoding level. In this
section we will analyze a suboptimal (Viterbi type) decoder that uses the
following suboptimal nearest neighbor metric (compare with (10))
X
n
nnl (r[0;n]; v[0(l);n]) =
( )
nnl (ri; vi l ) ;
( ) ( )
(71)
i=0
where
nnl (ri; vi l ) = min l dE (ri; si):
( ) ( )
(72)
si 2 Sv l ( )
( )
i

In all other aspects the decoder is analogous to the multistage Viterbi de-
coder, that uses the optimal metric. The nearest neighbor metric is often used
in multistage decoding of multilevel modulation schemes (see [6], [12]{[14]).
In [13] we proved that the e ective capacity of multistage decoding with this
metric practically coincides with the channel capacity. In this section we will
limit our analysis to the practically most interesting range of transmission
rates, up to the computational cuto rate. Our goal is to present a lower
bound nn for the overall state-complexity error exponent of the multilevel
Viterbi decoder with nearest neighbor metric.
Let us analyze the Viterbi decoder on the l-th level, using the suboptimal
nearest neighbor metric (71). We assume again that a kind genie informs the
decoder which sequences v ; v ; :::; v l were sent at the previous levels, if
(1) (2) ( 1)

the decoding on at least one level was erroneous. In Appendix A the Cherno
21
bounding parameter znnl for l-th nearest neighbor decoding is described. The
( )

values of zol and znnl are given in the Table I, and since they depend only on
( ) ( )

q0 = q=2l , they are presented as functions of this parameter. The relative


1

di erence (znnl zol )=zol does not exceed 0:003.


( ) ( ) ( )

Analogously to (20){(23), we get that the computational cuto rate for


Viterbi decoding with nearest neighbor metric at the l-th level is equal to
l
Rnnl = log 1 +2znn ;
( )
( )
2 (73)
the expurgation function is
n h  io ;
Gnnl (s) = s 1 log 1 + znnl 1<s<1;
1
( ) def
2
( ) s
(74)
and the convolutional expurgated error exponent is de ned by the paramet-
rical equation (comp. with (60))
Ennl = Gnnl (s) + o (1) ;
( ) ( )

1<s<1:
1
(75)
Rl
( )
= Gnnl (s)=s + o2 (1) ;
( )

Theorem 9 (Nearest Neighbor Expurgated Bound for Convolutional


Codes) Consider multilevel transmission over the AWGN channel with q-
PSK signaling. For the l-th level of a genie aided multilevel system using
Viterbi decoding with nearest neighbor metric, there exists a rate R(l) =
b(l) =c(l) trellis code of memory m(l) , whose burst error probability is upper-
bounded by
Pnnl (burst )  2 Enn c m R l  Rnnl ;
(l) (l) (l)
( )
; ( ) ( )
(76)
(l)
where Enn (R) is de ned by (75). 2
From (76) follows, analogously to (50){(54), that
Pb;nn <  l
  ( ) nn
( l)
+o(1)
; (77)

22
where  l = 2b l m l is the decoding complexity on the l-th level, o(1) ! 0
( ) ( ) ( )

when  l ! 1, and
( )

nnl = snnl ; ( ) def ( )


(78)
where snnl satis es the equality
( )

R l = Gnnl (snnl )=snnl :


( ) ( ) ( ) ( )
(79)
Then, analogously to (69), we get the lower bound of the overall state-
complexity error exponent nn that corresponds to the overall transmission
P
rate R = Ll R l , when R l , l = 1; 2; :::; L satis es the equalities
( ) ( )
=1

nn (R ) = nn (R ) = ::: = nnL (R L ) = nn :
(1) (1) (2) (2) ( ) ( )
(80)
The overall computational cuto rate is de ned by the equality
X
L
Rnn = Rnnl :( )
(81)
l=1
Since znnl  zol , the function nn(R), 0  R  Rnn, practically coincides
( ) ( )

with the function ~(R).

VII Some Numerical Results and Simulations


The main goal of this paper is the comparison of two transmission schemes
over the AWGN channel with q-PSK signaling: the conventional coded mod-
ulation system and the multilevel coded modulation system with multistage
decoding. In this section we will compare performances of these two systems
when both are using trellis codes with Viterbi decoding. We assume that
both systems have the same asymptotical state-complexity  and the same
bit error probability Pb , and consequently the same state-complexity error
exponent . We compare rst the relative asymptotic gain (for  ! 1) in
signal-to-noise ratio Es=N (in dB) of the multilevel system compared to the
0

conventional one. (Since the transmission rates of the systems are the same
the coding gain in signal-to-noise ratio per bit Eb =N is the same.) 0

23
To compare these two systems we rst calculate the upper ~ = ~(R)
and the lower ~ = ~(R) bounds for the state-complexity error exponent
~ = ~(R) of the multilevel system for a given signal-to-noise ratio Es=N , 0

as it was described in the end of Section V. An example of such bounds for


q = 32 and Es=N = 10dB is presented in Figure 6. Then for any given value
0

^ we calculate two values R and R, such that ^ = ~(R) and ^ = ~(R).


Now consider the conventional trellis coded modulation system, and cal-
culate for which signal-to-noise ratio (E s=N ) we have ^ = (R). Then we
0

nd for which signal-to-noise ratio (E s=N ) we have ^ = (R). Analogously,


0

(E s=N ) corresponds to ^ = (R) and (E s=N ) corresponds to ^ = (R).


0 0

The di erence (in dB) between (E s=N ) and Es=N gives an upper (opti-
0 0

mistic) bound for the coding gain of the multilevel system corresponding
to transmission rate R. The di erence between (E s=N ) and Es=N gives0 0

a lower (pessimistic) bound for the coding gain corresponding to transmis-


sion rate R. In addition to the lower and upper bounds for the coding gain
we calculate a realistic estimation of the coding gain, the di erence between
(E s=N ) and Es=N , corresponding to transmission rate R. The realistic es-
0 0

timation corresponds to the coding gain of the multilevel system over the
conventional system, given that their decoding error probabilities satisfy the
expurgated bounds.
In Figure 7 our results are presented for Es=N = 10dB and q = 32. We
0

note that for rates R = 0 and R = R~ the upper bound, the lower bound and
0

the realistic estimation coincide. The corresponding gains are 1:83 dB and
0:35 dB. For all rates 0  R < 0:9R~ the realistic estimation gives more than
0

1dB gain for the multilevel system. In Table II the gains are presented for
di erent signal-to-noise ratios Es=N at rate R = 0 and at the computational
0

cuto rate R = R = R~ .
0 0

The curves l (R), l = 1; 2; :::; L can be used for choosing parameters


( )

in the quasioptimal multilevel system whose asymptotical performances are


guaranteed by the expurgated bound. Suppose that we would like to con-
24
struct a system whose state-complexity error exponent is asymptotically not
less than ^. Choosing R l , l = 1; 2; :::; L, such that ^ = l (R l ), we nd
( ) ( ) ( )

the l-th level transmission rate R l = b l =c l , l = 1; 2; :::; L, of the multilevel


( ) ( ) ( )

system. To choose the memories of the component codes we use (53),


b m = b m = ::: = b L m L
(1) (1) (2) (2) ( ) ( )
(82)
If the components codes satisfy the expurgated bound (59), the constructed
code have a guaranteed asymptotically quasioptimal complexity-reliability
tradeo . This construction method of asymptotically quasioptimal multi-
level systems can be used for construction of practical systems with mod-
erate complexity. In Appendix B we give an example in which we use this
construction method.

VIII Conclusion
In this paper we compared a conventional coded modulation system with a
multilevel coded modulation system both for block and trellis coding. We also
compared multilevel systems in the case of optimal and suboptimal decoding
and proved that they have practically the same performances. An example
of the energy gain of the multilevel systems as a function of the transmission
rate is presented in Figure 7 for the AWGN channel with 32-PSK input. In
Table II the energy gain (in dB) for zero transmission rate (R = 0) and for
the computational cuto rate R are presented.
0

We introduced the concept of a \genie-aided multilevel system". It pro-


vides asymptotically rigorous bounds on the decoding error probabilities.
To synthesize the multilevel coding system we have developed a method of
asymptotically quasioptimal choice (maximizing the overall rate for a given
guaranteed decoding error probability and a given decoding complexity) of
code parameters (transmission rates and block length / code memory) for
di erent levels. Surprisingly enough, the same principle of synthesis of mul-
tilevel systems, that was used in the asymptotical case, when the code length
25
/ code memory goes to in nity, can also be applied for the synthesis of prac-
tically interesting codes of reasonable code memory size.

Appendix A The Cherno Bounding Parameter


It was proven in [13] that the Cherno bounding parameter znnl for l-th ( )

nearest neighbor decoding can be calculated as


znnl = min
( )

' l () ;0
( )
(83)
where ' l () is the generating function of the random variable
( )

 l (r) = nnl (r; 0) nnl (r; 1) ;


( ) def ( ) ( )
(84)
where r = x + | y. We have [13] in the q-PSK case that
rE 
 (r) = 2 2 Ns x + x cos  l + y sin  l ;
l ( )
(85) ( ) ( )

where  l = 2=q0, q0 = q=2l . The conditional probability density function


( ) 1

p
to receive x + | y given that the transmitted symbol is s = Es and that
0  y  x tan  l , 0  x  1 is
( )

q0
8 rE !
1X 2
1
< 1 2
l
fq0 (x; y) = 
( )
exp : 2 x 2 Ns cos 2 l i ( )

i=0
!9
0

1
rE = 2

2 y 2 Ns sin 2 l i ; : ( )

Then ' l () can be calculated as


( )

Z1 x Z  tan (l)

' l () =
( )
exp  l (r) fq0l (x; y) dy dx :
( ) ( )
(86)
x=0 y=0
From (86) follows [13] that
0 1
 E B r2 X0
Z
  A ()  C
q
2
1
( l)

' l () = exp Ns @1 +  A()Q( A()) exp 2 dA ;


2
( )

0
i=0 =0
(87)
26
where
rE rE rE
A() = 2 N cos(2 i )+2 2 N  cos( ) 2 2 Ns  cos ; (88)
s l ( ) s l ( )

0 0 0

R1 p
and Q(x) = exp ( t =2) = 2dt.
2

Appendix B An Example of the Construction of a


Multilevel Code
Here we give an example where we use the construction method described in
Section VII.
Example: Suppose that we would like to construct a multilevel system for
q = 8, Es=N = 10dB. Let us choose ~ = 2:3. In Figure 5 the functions
0

(R ) (left solid line), (R ) (left dashed line), (R ) (left dash-


(3) (3) (2) (2) (1) (1)

dotted line) are presented. Then R  0:25, R  0:8 and R  0:98


(1) (2) (3)

correspond to ~ = 2:3. Let us choose R = 1=4, R = 4=5, and R = 1


(1) (2) (3)

(uncoded transmission), the overall rate of the system R = 2:05 and m = 4, (1)

m = 1.
(2)

We chose the following codes: for the rst level the code with genera-
tor polynomials (in octal representation) (46; 52; 66; 76) and for the second
level the punctured rate 1=2 code with generator polynomials (23; 35) of the
mother code and punctuation matrix
1 0 1 0
P= 1 1 0 1 : (89)

Simulation results of this system for Es=N ranging from 3dB to 10dB
0

are presented in Figure 8. For comparison simulation results for the rst two
levels of the genie aided system are given. We also given the exact value of
the bit error probability for the last level. We can see, that for Es=N = 10dB 0

the overall bit error probability and the bit error probabilities for the levels of
the genie aided system are very close. This was our goal in the construction

27
of the multilevel system. Since the parameters of the system were optimized
only for signal-to-noise ratio Es=N = 10dB, the system naturally has worse
0

performances for lower signal-to-noise ratios.


Since simulations at higher signal-to-noise ratios becomes more and more
time consuming, we have used the Viterbi owchart diagram [9],[11], to get
an upper bound for the bit error probabilities Pb at each level:

Pb < 1b @T (W; L; I )
@I W z ; (90)
LI
= o
= 1
= 1

where T (W; L; I ) is the extended path enumerator of the convolutional code


and zo is the Cherno bounding parameter for Viterbi decoding with optimal
metric (21). (For W = znn the bounds practically coincide with those given
in Figure 8.) In Figure 8 these upper bounds of Pb for the rst two levels in
the range of signal-to-noise ratios Es=N from 10dB to 13dB are presented.
0

The result of the calculation of Pb for the last level is also presented. 2

28
References
[1] G. Beyer, K. Engdahl, and K. Sh. Zigangirov, "Asymptotical analysis of
two coded modulation schemes using PSK signaling, Part I," submitted
to IEEE Transactions on Information Theory, Jun. 1999.
[2] G. Ungerbck, "Channel coding with multilevel/phase signals," IEEE
Transactions on Information Theory, vol. IT-28, pp. 55{67, Jan. 1982.
[3] H. Imai and S. Hirakawa, "A new multilevel coding method using error-
correcting codes," IEEE Transaction on Information Theory, vol. IT-23,
pp. 371{377, May 1977.
[4] J. Huber, "Multilevel codes: Distance pro les and channel capacity," in
ITG-Fachbereicht 130 (conf. rec.), 1994, pp. 305{319.
[5] J. Huber and U. Wachsmann, "Capacities of equivalent channels in mul-
tilevel coding schemes," Electronics Letters, vol. 30, pp. 557{558, no. 7
Mar. 1994.
[6] Y. Kofman, E. Zehavi, and S. Shamai, "Performance analysis of a mul-
tilevel coded modulation system," IEEE Transactions on Communica-
tions, vol. 42, pp. 299{312, Feb./Mar./Apr. 1994.
[7] H. Herzberg, "On the spectrum of distances of a multilevel code, decoded
by a multistage decoder," IEEE Transactions on Information Theory,
vol. IT-43, pp. 1736{1739, Sep. 1997.
[8] J. M. Wozencraft and I. M. Jacobs, Principles of Communication Engi-
neering, Wiley, 1965.
[9] R. Johannesson and K. Sh. Zigangirov, Fundamentals of Convolutional
Coding, IEEE Press, Piscataway N.J., 1999.

29
[10] R. G. Gallager, Information Theory and Reliable Communication, Wi-
ley, New York, 1968.
[11] A. J. Viterbi and J. K. Omura, Principles of Digital Communication
and Coding, McGraw Hill, 1979.
[12] K. Engdahl and K. Sh. Zigangirov, "On the calculation of the error
probability for a multilevel modulation scheme using QAM-signaling,"
IEEE Transactions on Information Theory, vol. IT-44, pp. 1612{1620,
Jul. 1998.
[13] K. Engdahl and K. Sh. Zigangirov, "The Cherno bounding parameter
for a multilevel modulation scheme using PSK-signaling," International
Journal of Electronics and Communication (AEU)  , pp. 18{24, no. 1
1999.
[14] K. Engdahl and K. Sh. Zigangirov, "A comparison analysis of hexago-
nal multilevel QAM and rectangular multilevel QAM," Probl. Peredach.
Inform., vol. 34, pp. 39{50, no. 4 Oct./Nov./Dec. 1998.

30
LIST OF FIGURE CAPTIONS
Fig. 1: System description of the L-level coded modulation scheme using
2L-PSK signaling.
Fig. 2: Set partitioning of 8-PSK signal constellation.
Fig. 3: Upper E Bl (R l ) (solid lines) and lower E Bl (R l ) (dashed lines)
( ) ( ) ( ) ( )

bounds of the block error exponent of a genie aided multilevel sys-


tem; q = 32, Es=N = 10dB. The enumeration of the levels l = 3; 4; 5
0

is from bottom to top.


Fig. 4: Comparison of upper E B (R) (correspondingly E~ B (R)) and lower
E B (R) (correspondingly E~ B (R)) bounds for the overall block error ex-
ponent for conventional and multilevel modulation; q = 32, Es=N = 0

10dB.
Fig. 5: The upper l (R l ) and lower l (R l ) bounds for the l-th level
( ) ( ) ( ) ( )

state-complexity error exponent, l = 2; 3; 4; 5; q = 32, Es=N = 10dB.0

Fig. 6: Comparison of upper (R) (correspondingly ~(R)) and lower (R)


(correspondingly ~(R)) bounds of the overall state-complexity error
exponents for conventional and multilevel modulation schemes; q = 32,
Es=N = 10dB.
0

Fig. 7: The optimistic (upper), pessimistic (lower) and realistic estimation


of the coding gain of multilevel modulation scheme compared to the
conventional modulation scheme.
Fig. 8: Study of bit error probability for a three-level modulation system,
whose parameters are optimized for Es=N = 10dB, see example. The
0

solid line corresponds to the simulation of the whole system in the


range of Es=N from 3 to 10 dB. The dashed curve corresponds to
0

the rst level of the system { simulation results in the range 3{10dB of
31
Es=N , and upper bound for Pb in the range 10{13dB. The dash-dotted
0

curve corresponds to the second level of genie aided system { simulation


results in the range 3{10dB of Es=N and upper bound for Pb in the
0

range 10{12dB. The dotted curve corresponds to the exact calculation


of Pb for the third (uncoded) level of the genie aided system.

32
LIST OF TABLE CAPTIONS
Tab. I: Values of the Cherno bounding parameters zo and znn as functions
of the signal-to-noise ratio Es=N .
0

Tab. II: The asymptotical gain of multilevel 32-PSK signaling in compar-


ison to conventional 32-PSK signaling for R = 0 and R = R~ as a
0

function of the signal-to-noise ratio Es=N for the multilevel scheme.


0

33
LIST OF FOOTNOTES
1 We will refer to the formulas from the rst part of this paper by adding
the Roman number I.
2 The ideology of a \genie aided communication system" was popular in the
60s, see for example [8]. Now it is practically forgotten.

34
BIOGRAPHIES
Gerd Beyer was born in Aalen, Germany, in 1973. He received the M.Sc.
degree in computer science from the University of Ulm, Germany, in
1999. Working at his master thesis, during 1998-1999, he spent six
months at the Department of Information Technology at Lund Univer-
sity, Sweden. Since November 1999 he has been working as consultant
at Beck et al. in Munich, Germany.
Karin Engdahl (S'96) was born in Sweden in 1971. She received her
M.Sc. degree in Electrical Engineering from Lund University, Sweden,
in 1995. Since then she is with the Dept. of Information Technology,
Lund University, where she is working towards a Ph.D.
Kamil Sh. Zigangirov (M'95-SM'99-F'2001) was born in the USSR in
1938. He received the M.S. degree in 1962 from the Moscow Institute
for Physics and Technology and the Ph.D. degree in 1966 from the
Institute of Radio Engineering and Electronics of the USSR Academy
of Sciences, Moscow.
From 1965 to 1991 he held various research positions at the Institute
for Problems of Information Transmission of the USSR Academy of
Sciences, Moscow, rst as a Junior Scientist, and later as a Main Sci-
entist. During this period he visited several universities in the United
States, Sweden, Italy, and Switzerland as a guest researcher. He orga-
nized several symposia on information theory in the USSR. In 1994 he
received the Chair of Telecommunication Theory at Lund University,
Lund, Sweden. His scienti c interests include information theory, cod-
ing theory, detection theory, and mathematical statistics. In addition
to papers in these areas he has published a book on sequential decoding
of convolutional codes. With R. Johannesson he coauthored the text-
book "Fundamentals of Convolutional Coding" (Piscataway, NJ: IEEE

35
Press, 1999).

36
u Encoder 1 v
(1) (1)

u Encoder 2 v
(2) (2)

u Partition
of
2L-PSK s AWGN r Multistage u^
information p mapper decoder
p
p
u L Encoder L v L
( ) ( )

Figure 1:

uuu
u u
v = 0
(1) uuu
HHvH = 1
(1)




 HHHj
eue ueu
u u e e
eue ueu
v =0
(2)
@ v =1(2)
v =0(2)
@v = 1
(2)

@@
R e @@R
eue e e uee e
e u
e e u u e e e e
eue eee eeu uee

Figure 2:

37
8
upper bounds
lower bounds
7

5
EB (R )
(l)

4
(l)

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
(l)
R

Figure 3:

38
12
conventional modulation
multilevel modudation

10

8
EB(R)

0
0 0.5 1 1.5 2 2.5 3
R

Figure 4:

39
10

6
γ(l)(R(l))

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
(l)
R

Figure 5:

40
10

5
γ

multilevel modulation
1
conventional modulation

0
0 0.5 1 1.5 2 2.5 3
R

Figure 6:

41
5

2
gain [dB]

realistic estimation
−1
lower bound
upper bound

−2
0 0.5 1 1.5 2 2.5 3
R

Figure 7:

42
0
10

−1
10

−2
10

−3
10

−4
10

−5
Pb

10

−6
10

−7
10

10
−8 simulated overall performance
level 1
level 2 of genie aided sys.
10
−9 level 3 of genie aided sys.

−10
10
3 4 5 6 7 8 9 10 11 12 13
Es/N0 in [dB]

Figure 8:

43
Es=N
0 q optimal zo suboptimal znn  znn=zo
1 2 0.284 0.284 1.000
1 4 0.743 0.745 1.003
1 8 0.988 0.989 1.001
1 16 1.000 1.000 1.000
1 32 1.000 1.000 1.000
1 64 1.000 1.000 1.000
3 2 0.136 0.136 1.000
3 4 0.564 0.566 1.003
3 8 0.962 0.964 1.002
3 16 1.000 1.000 1.000
3 32 1.000 1.000 1.000
3 64 1.000 1.000 1.000
5 2 0.042 0.042 1.000
5 4 0.345 0.346 1.002
5 8 0.896 0.898 1.003
5 16 0.999 1.000 1.000
5 32 1.000 1.000 1.000
5 64 1.000 1.000 1.000
8 2 0.002 0.002 1.000
8 4 0.080 0.080 1.000
8 8 0.661 0.662 1.002
8 16 0.989 0.990 1.001
8 32 1.000 1.000 1.000
8 64 1.000 1.000 1.000
10 2 0.000 0.000 1.000
10 4 0.013 0.013 1.000
10 8 0.420 0.420 1.001
10 16 0.952 0.953 1.001
10 32 1.000 1.000 1.000
10 64 1.000 1.000 1.000
Table I:

44
Es=N [dB] gainjR [dB] R~ gainjR R [dB]
0 =0 0 = ~0

0 0:85 0:70 0:5


1 0:95 0:85 0:51
3 1:16 1:20 0:5
5 1:37 1:59 0:47
8 1:66 2; 16 0:41
10 1:83 2:51 0:35
15 1:7 3:37 0:3
Table II:

45