Multilevel Coding and

Iterative Multistage Decoding

ELEC 599 Project Presentation

Mohammad Jaber Borran

Rice University
April 21, 2000
 Multilevel Coding
A number of parallel encoders
The outputs at each instant select one symbol
q1 K1 N x1
E1 (rate R1)
data bits
from the q2 K2 N x2 Signal
information E2 (rate R2)
M-way Mapping Point
source Partitioning (to 2M-point
of data constellation)

qM KM N xM
EM (rate RM)

M M
1
R   Ri  K i bits/symbol
i 1 N i 1
 Distance Properties
• Minimum Hamming distance for encoder i: dHi ,
Minimum Hamming distance for symbol
sequences
d H  min (d Hi )
i1,, M 

• For TCM (because of the parallel transitions)

dH = 1
• MLC is a better candidate for coded modulation
on fast fading channels
Probability of error for Fading Channels
• Rayleigh fading with independent fading coefficients
Chernoff bound

1 1
Pe (ci , c j )   L
E N0 
L
 d k2 (c i , c j ) 
s
 
k 1 
4

2

d k 0

L’: effective length of the error event (Hamming distance)

dk(ci,cj): distance between the kth symbols of the two
sequences
 Design Criterion for Fading Channels
• For a fast fading channel, or a slowly fading channel with
interleaving/deinterleaving
Design criterion (Divsalar)

max min d H (c i , c j ) L
{c1 ,c 2 ,,c n } i, j
d P (ci , c j )   d k2 (ci , c j )
max min d P (c i , c j ) k 1
{c1 ,c 2 ,,c n } i, j d k2  0

• For a slowly fading channel without

interleaving/deinterleaving
Design criterion max min d E (c i , c j )
{c1 ,c 2 ,,c n } i, j
 Decoding Criterion

• For a fast fading channel, or a slowly fading channel with

interleaving/deinterleaving

L
min  |  k |2 d k2 ( ~
y, ci ) where ~
yk  yk  k
i
k 1

k is the fading coefficient for kth symbol)

– Maximizes the likelihood function
 Decoding

• Optimum decoder: Maximum-Likelihood decoder

• If the encoder memories are 1, 2, …,M,

the total number of states is 2,
where = 1 + 2 + … + M.

• Complexity  Need to look for suboptimum

decoders
• If A and Y denote the transmitted and received
symbol sequences respectively, using the chain
rule for mutual information:

I (Y ; A)  I (Y ; X1 , X 2 ,  , X M )
 I (Y ; X 1 )  I (Y ; X 2 | X 1 )  
 I (Y ; X M | X 1 , X 2 , , X M 1 )

• Suggests a rule for a low-complexity staged

decoding procedure
 Multistage Decoding
X̂1
Decoder D1
X̂ 2
Decoder D2
Y
â

X̂ M
Decoder DM

• At stage i, decoder Di processes not only the

sequence of received signal points, but also
decisions of decoders Dj, for j = 1, 2, …, i-1.
 X̂1 X̂ 2 ˆ
X i 1

...
Y Decoder Di X̂ i

• The decoding (in stage i) is usually

done in two steps
– Point in subset decoding
– Subset decoding

• This method is not optimal in maximum

likelihood sense, but it is asymptotically optimal
for high SNR.
 Optimal Decoding
xi

Pr{a}
M i ( xˆ1 , , xˆi 1 , xi )   fY | A ( y | a)
aA i ( xˆ1 ,, xˆi 1 , xi )  Pr{b}
bA i 1 ( xˆ1 ,, xˆi1 )

– Ai(x1,…, xi) is the subset determined by x1,…, xi

– fY|A(y|a) is the transition probability (determined by
the channel)
 Rate Design Criterion
X̂1
Decoder D1
X̂ 2
Decoder D2
Y
â

X̂ M
Decoder DM

C1  I (Y ; X 1 ) then the rate of the code at

C2  I (Y ; X 2 | X 1 ) level i, Ri, should satisfy

Ri  Ci
C M  I (Y ; X M | X 1 , X 2 ,  , X M 1 )
 3
C
C1
C2
2.5

2
Capacity (bits/symbol)

1.5

0.5

0
-5 0 5 10 15 20
SNR (dB)

Two-level, 8-ASK, AWGN channel

 Rate Design Criterion

Using the multiaccess channel analogy, if optimal

decoding is used,
R2
Ri  I (Y ; X i | { X k }k i )
I(Y;X2|X1)
Ri  R j  I (Y ; X i , X j | { X k }k i , j )
 I(Y;X2)

R
i
i  I (Y ; X 1 ,  , X M )  I (Y ; A)
R1
I(Y;X1) I(Y;X1|X2)
 3

2.5

2
Capacity (bits/symbol)

1.5

C
0.5
C1
C2
I(Y;X1|X2)
0
-5 0 5 10 15 20
SNR (dB)

Two-level, 8-ASK, AWGN channel

 Iterative Multistage Decoding
Assuming – Two level Code
– R1  I(Y;X1|X2)
– Decoder D1: Pr{x1 | xˆ1}
then the a posteriori probabilities are
Pr{a | xˆ1}  Pr{A1 ( x1 ) | xˆ1} Pr{a | A1 ( x1 )}
Pr{a}
 Pr{x1 | xˆ1}
 Pr{b}
bA1 ( x1 )

This expression, then, can be used as a priori

probability of point a for the second decoder.
 Probability Mass Functions

Error free decoding Non-zero symbol

error probability
 3

C
C1
2.5 C2
I(Y;X1|X2)
I(Y;X2|partial X1)
2
Capacity (bits/symbol)

1.5

0.5

0
-5 0 5 10 15 20
SNR (dB)

Two-level, 8-ASK, AWGN channel

 3
C
C1
2.5 C2
I(Y;X1|X2)
I(Y;X2|partial X1)

2
Capacity (bits/symbol)

1.5

0.5

0
-5 0 5 10 15 20 25 30 35
SNR (dB)

Two-level, 8-ASK, Fast Rayleigh fading channel

 0
10
Overall
Encoded
Uncoded
-1
10

-2
Error Probability

10

-3
10

-4
10

-5
10
0 1 2 3 4 5 6 7
SNR per Bit

8-PSK, 2-level, 4-state, uncoded, AWGN channel

 -1
10
Overall
Encoded
Uncoded

-2
10
Error Probability

-3
10

-4
10

-5
10
6 8 10 12 14 16 18 20
SNR per Bit

8-PSK, 2-level, 4-state, uncoded , fast Rayleigh fading channel

 0
10
Overall
First Level
Second Level
-1
10

-2
Error Probability

10

-3
10

-4
10

-5
10
6 8 10 12 14 16 18 20
SNR per Bit

8-PSK, 2-level, 4-state, zero-sum, fast Rayleigh fading channel

 0
10
Overall
First Level
Second Level
-1
10

-2
Error Probability

10

-3
10

-4
10

-5
10
6 8 10 12 14 16 18
SNR per Bit

8-PSK, 2-level, 4-state, 2-state , fast Rayleigh fading channel

 0
10
4-state, zero-sum
4-state, 2-state, 1-iteration
4-state, 2-state, 2-iteration

-1
10
Error Probability

-2
10

-3
10

-4
10
6 8 10 12 14 16 18 20
SNR per Bit

8-PSK, 2-level, fast Rayleigh fading

Higher Constellation Expansion Ratios

• For AWGN, CER is usually 2

– Further expanding  Smaller MSED
 Reduced coding gain
• For fading channels,
– Further expanding  Smaller product distance
 Reduced coding gain
– Further expanding  Larger Hamming distance
 Increased diversity gain
 0
10
TCM, 8-PSK
2-level, 1-iteration, 16-PSK

-1
10

-2
Error Probability

10

-3
10

-4
10

-5
10
0 2 4 6 8 10 12 14
SNR per Bit
 -2
10

-3
10
Error Probability

-4
10

TCM, 8-PSK
2-level, 1-iteration, 16-PSK
2-level, 2-iteration, 16-PSK

-5
10
14 15 16 17 18 19 20
SNR per Bit
 Conclusion
• Using iterative MSD with updated a priori
probabilities in the first iteration, a broader
subregion of the capacity region of MLC scheme
can be achieved.

• Lower complexity multilevel codes can be
designed to achieve the same performance.
• Coded modulation schemes with constellation
expansion ratio greater than two can achieve better
performance for fading channels.
 Coding Across Time
• If channels are encoded separately, assuming
– A slowly fading channel in each frequency bin, and
– Independent fades for different channels
(interleaving/deinterleaving across frequency bins is
used)

 Es h 2 
Pr c  cˆ | h  exp c
2
n  cˆn 
 4 N 0 n 
1
Eh  Pr c  cˆ | h  
Es
 cn  cˆn
2
1
4N0 n
 Coding Across Frequency Bins
• If coding is performed across frequency bins,
assuming independent fades for different
channels (interleaving/deinterleaving across
frequency bins is used)

 E 2
Pr c  cˆ | h  exp s n n n n 
2
h c  cˆ
 4N0 
1
Eh  Pr c  c | h   
ˆ
n 1
Es 2
cn  cˆn
4N0
 0
10
Accross time, 1-iteration
Accross time, 2-iteration
Accross frequency, 1-iteration
-1 Accross frequency, 2-iteration
10
Error Probability

-2
10

-3
10

-4
10
6 8 10 12 14 16 18 20
SNR per Bit

8-PSK, 2-level, 4-state, 2-state