Beruflich Dokumente
Kultur Dokumente
atenated
Turbo Codes
Mark Sum Chuen Ho
B.Eng(Hons)
November 2002
Contents
1 Introdu
tion
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5
6
7
7
8
10
11
15
17
18
19
19
22
23
24
25
25
28
29
29
29
30
32
33
36
38
41
41
CONTENTS
ii
3.6 Ee
tive free distan
e and minimum distan
e . . . . . . . .
3.7 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.7.1 Trun
ated union bounds of PCCC . . . . . . . . . . .
3.7.2 Maximum likelihood de
oding and iterative de
oding
3.8 Con
lusion . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4 Interleaver Design
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
6 Con lusion
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
43
46
52
52
53
54
54
55
55
57
65
65
68
69
69
71
72
81
87
91
94
95
95
95
96
96
98
101
101
103
104
105
105
105
107
107
110
114
CONTENTS
iii
119
123
126
127
List of Figures
1.1 Stru
ture of a digital
ommuni
ation system. . . . . . . . . . . . . . . . . .
2.1
2.2
2.3
2.4
2.5
2.6
2.7
2.8
2.9
2.10
2.11
2.12
2.13
iv
1
6
7
8
8
9
14
15
17
18
20
23
24
25
27
27
28
33
34
34
37
43
44
45
LIST OF FIGURES
3.8 Trun
ated union bound and BER performan
e between dierent
riteria of
maximisation for N=4096 and 6 de
oder iterations. . . . . . . . . . . . . .
3.9 En
oder stru
ture for rate 1/n RSC
onstituent
ode (eight state rate 1/3
example). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.10 En
oder stru
ture for rate k/(k+1) systemati
feedba
k
onvolutional
ode
(eight state rate 2/3 example). . . . . . . . . . . . . . . . . . . . . . . . . .
3.11 BER performan
e
omparison between repetition and non-repetition
onstituent
odes for four state rate 1/5 PCCC, N=512. . . . . . . . . . . . .
3.12 BER performan
e
omparison between repetition and non-repetition
onstituent
ode for four state rate 1/5 PCCC, N=1024. . . . . . . . . . . . .
3.13 Trun
ated union bound for rate 1/3 PCCC with 1024 bit interleaver, with
2 state to 128 state rate 1/2
onstituent
odes. . . . . . . . . . . . . . . . .
4.1 Input/output position plot of a 192 bit pseudo-random interleaver. . . . . .
4.2 Input/output position plot of a 192 bit S-random interleaver. . . . . . . . .
4.3 PCCC performan
e with random interleaver and in
reasing S
onstraint,
192 bit interleaver. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.4 PCCC performan
e with random interleaver and in
reasing S
onstraint,
512 bit interleaver. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.5 PCCC performan
e with random interleaver and in
reasing S
onstraint,
2048 bit interleaver. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.6 PCCC performan
e with random interleaver and in
reasing S
onstraint,
4096 bit interleaver. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.7 4 state rate 1/3 SCCC simulation with 192 bit interleaver of dierent S
values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.8 4 state rate 1/3 SCCC simulation with 1024 bit interleaver of dierent S
values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.9 4 state rate 1/3 SCCC simulation with 4096 bit interleaver of dierent S
values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.10 Amount of prote
tion on information bits between random interleaver and
mod-2 interleaver. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.11 Input/output position plot of a 192 bit mod-2 S-random interleaver. . . . .
4.12 Example of a nine element pure mod-3 interleaver. . . . . . . . . . . . . . .
4.13 4 state rate 1/2 turbo
ode with dierent 256 bit mod-2 interleavers, eight
iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.14 4 state rate 1/2 turbo
ode with dierent 1024 bit mod-2 interleavers, eight
iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.15 4 state rate 1/2 turbo
ode with dierent 4096 bit mod-2 interleavers, eight
iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.16 4 state rate 3/5 turbo
ode with dierent 2049 bit mod-3 interleavers, eight
iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.17 Interleave and deinterleave pattern for symmetri
and random interleaver. .
4.18 Input/output position plot of a 192 bit blo
k interleaver. . . . . . . . . . .
v
45
46
46
51
51
52
56
57
58
58
59
59
60
60
61
62
63
64
66
66
67
67
68
70
vi
LIST OF FIGURES
4.19 Input/output position plot of a 192 bit
ir
ular shift interleaver, a = 17. . .
4.20 A 3 4 blo
k heli
al interleaver with interleaving from (a) the upper left
hand
orner and (b) the lower left hand
orner. . . . . . . . . . . . . . . .
4.21 Input/output position plot of a 208 bit blo
k heli
al interleaver. . . . . . .
4.22 Position plot of a 4 3 blo
k heli
al interleaver. . . . . . . . . . . . . . . .
4.23 Input/output position plot of a blo
k heli
al symmetri
interleaver. . . . .
4.24 Input/ouput position plot of a 192 bit
ir
ular symmetri
interleaver with
a = 31. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.25 BER performan
e of 255 bit
ir
ular interleaver with dierent values of a. .
4.26 BER performan
e of 1023 bit
ir
ular interleaver with dierent values of a.
4.27 BER performan
e of 4095 bit
ir
ular interleaver with dierent values of a.
4.28 Input/output position plot of a 192 bit S-symmetri
interleaver. . . . . . .
4.29 Interleaver performan
e
omparison for four state rate 1/3 turbo
ode,
N=1024, six iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.30 4 state rate 1/3 turbo
ode with dierent 400 bit interleavers and eight
iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.31 4 state rate 1/3 turbo
ode with dierent 2048 bit interleavers and eight
iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.32 4 state rate 1/3 turbo
ode with dierent 4096 bit interleavers and eight
iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.33 Comparison between dierent types of symmetri
interleaver with six iterations, N =256. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.34 Comparison between dierent types of symmetri
interleaver with eight
iterations, N =512. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.35 Comparison between dierent types of symmetri
interleaver with eight
iterations, N =1024. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.36 Comparison between dierent types of symmetri
interleaver with eight
iterations, N =4096. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.37 4 state rate 1/2 turbo
ode with dierent 1024 bit interleaver designs and
six iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.38 4 state rate 3/5 turbo
ode with dierent 256 bit interleavers and eight
iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.39 4 state rate 3/5 turbo
ode with dierent 1024 bit interleavers and eight
iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.40 4 state rate 3/5 turbo
ode with dierent 4096 bit interleavers and eight
iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.41 4 state rate 4/6 turbo
ode with dierent 400 bit interleavers and eight
iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.42 4 state rate 4/6 turbo
ode with dierent 2048 bit interleavers and eight
iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.43 4 state rate 4/6 turbo
ode with dierent 4086 bit interleavers and eight
iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.44 4 state rate 1/3 turbo
ode with dierent interleaver sizes and eight iterations.
70
71
72
74
76
77
78
78
79
80
82
82
83
83
84
84
85
85
86
87
88
88
89
89
90
90
LIST OF FIGURES
4.45
4.46
4.47
4.48
4.49
vii
4 state rate 1/3 SCCC simulation with dierent types of 192 bit interleaver.
4 state rate 1/3 SCCC simulation with dierent types of 256 bit interleaver.
4 state rate 1/3 SCCC simulation with dierent types of 512 bit interleaver.
4 state rate 1/3 SCCC simulation with dierent types of 1024 bit interleaver.
4 state rate 1/3 SCCC simulation with dierent types of 4096 bit interleaver.
5.1 BER performan
e of four state rate 1/2 log-MAP de
oder versus de
oder
Eb =N oset, N=1024. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 BER performan
e of four state rate 2/3 log-MAP de
oder versus de
oder
Eb =N oset, N=1024. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3 BER performan
e of four state rate 1/3 PCCC versus de
oder Eb =N oset,
N=1024, six iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.4 BER performan
e of four state rate 1/3 SCCC versus de
oder Eb =N oset,
N=256, eight iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.5 BER performan
e of four state rate 1/3 SCCC versus de
oder Eb =N oset,
N=1024, eight iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.6 BER performan
e of four state rate 1/3 SCCC versus de
oder Eb =N oset,
N=2048, six iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.7 BER performan
e of four state rate 1/3 SCCC versus de
oder Eb =N oset,
N=4096, eight iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.8 f (x) versus Channel Varian
e. . . . . . . . . . . . . . . . . . . . . . . . . .
5.9 Performan
e
omparison between dierent varian
e estimators, four state,
rate 1/3 PCCC, N=384 and six iterations. . . . . . . . . . . . . . . . . . .
5.10 Performan
e
omparison between dierent varian
e estimators, four state,
rate 1/3 PCCC, N=4092 and six iterations. . . . . . . . . . . . . . . . . .
5.11 Performan
e
omparison between dierent varian
e estimators, four state,
rate 1/3 PCCC, N=8192 and six iterations. . . . . . . . . . . . . . . . . .
5.12 Performan
e
omparison between dierent varian
e estimators, four state
rate 1/3 SCCC, N=384 and six iterations. . . . . . . . . . . . . . . . . . .
5.13 Performan
e
omparison between dierent varian
e estimators, four state
rate 1/3 SCCC, N=2048 and six iterations. . . . . . . . . . . . . . . . . . .
5.14 Performan
e
omparison between dierent varian
e estimators, four state
rate 1/3 SCCC, N=8092 and six iterations. . . . . . . . . . . . . . . . . . .
91
92
92
93
93
97
97
A.1 Trun
ated union bound for rate 1/5 PCCC with 1024 bits interleaver, with
2 state to 64 state rate 1/3
onstituent
odes. . . . . . . . . . . . . . . . .
A.2 Trun
ated union bound for rate 1/2 PCCC with 1024 bits interleaver, with
2 state to 64 state rate 2/3
onstituent
odes. . . . . . . . . . . . . . . . .
A.3 Trun
ated union bound for rate 1/2 PCCC with 1024 bits interleaver, with
2 state to 128 state pun
tured rate 1/2
onstituent
odes. . . . . . . . . . .
A.4 Trun
ated union bound for rate 3/5 PCCC with 1024 bits interleaver, with
2 state to 32 state rate 3/4
onstituent
odes. . . . . . . . . . . . . . . . .
98
99
100
100
101
102
106
106
107
108
108
109
119
120
120
121
viii
LIST OF FIGURES
A.5 Trun
ated union bound for rate 1/7 PCCC with 1024 bits interleaver, with
2 state to 32 state pun
tured rate 1/4
onstituent
odes. . . . . . . . . . .
A.6 Trun
ated union bound for rate 1/9 PCCC with 1024 bits interleaver, with
2 state to 16 state rate 1/5
onstituent
odes. . . . . . . . . . . . . . . . .
A.7 Trun
ated union bound for rate 4/6 PCCC with 1024 bits interleaver, with
2 state to 32 state rate 4/5
onstituent
odes. . . . . . . . . . . . . . . . .
B.1 Flow diagram of a mod-k interleaver generator with S distan
e property. .
B.2 Flow diagram of a symmetri
interleaver with S and mod-k properties. . .
121
122
122
124
125
List of Tables
2.1 Rate 1/2 RSC generator polynomials. . . . . . . . . . . . . . . . . . . . . .
3.1 Generator polynomials with dierent number of neighbours. . . . . . . . .
3.2 Eight state rate 2/3
onstituent
odes that were sear
hed a
ording to
dierent
riteria. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3 Best rate 1/2
onstituent
odes. . . . . . . . . . . . . . . . . . . . . . . . .
3.4 Best rate 2/3 pun
tured rate 1/2
onstituent
odes with pun
turing pattern
[1 0. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5 Best rate 1/3
onstituent
odes. . . . . . . . . . . . . . . . . . . . . . . . .
3.6 Best rate 1/4
onstituent
odes. . . . . . . . . . . . . . . . . . . . . . . . .
3.7 Best rate 1/5
onstituent
odes. . . . . . . . . . . . . . . . . . . . . . . . .
3.8 Best rate 2/3
onstituent
odes. . . . . . . . . . . . . . . . . . . . . . . . .
3.9 Best rate 3/4
onstituent
odes. . . . . . . . . . . . . . . . . . . . . . . . .
3.10 Best rate 4/5
onstituent
odes. . . . . . . . . . . . . . . . . . . . . . . . .
3.11 Best rate 5/6
onstituent
odes. . . . . . . . . . . . . . . . . . . . . . . . .
4.1 Input/output index of the heli
al interleaver in Figure 4.22. . . . . . . . . .
5.1 Mean Squared Error for dierent estimators with blo
k size of 384. . . . .
ix
26
42
44
47
48
48
48
49
49
49
50
50
74
109
Glossary
Term
AWGN
BER
BPSK
CC
CWEF
HCCCs
IRWEF
MAP
PCCC
QPSK
RSC
SCCC
SOVA
Denition
interleaver
deinterleaver
additive white Gaussian noise
bit error rate
binary phase shift keying
onstituent
odes
onditional weight enumerating fun
tion
hybrid
on
atenated
onvolutional
odes
input-redundan
y weight enumerating fun
tion
maximum a posteriori
parallel
on
atenated
onvolutional
odes
quadrature phase shift keying
re
ursive systemati
onvolutional
serial
on
atenated
onvolutional
odes
soft output Viterbi algorithm
Page
7
10
12
2
1
7
39
38
39
11
3
1
5
3
11
Summary
Turbo
odes were rst introdu
ed by Berrou et. al in 1993. It is a
lass of
odes whi
h
on
atenate two or more re
ursive systemati
onvolutional
odes (RSC) in parallel, with
the input to ea
h RSC permuted by an interleaver. These
odes
an a
hieve performan
e
very
lose to Shannon's
apa
ity with large interleavers (several kilobits).
The performan
e of turbo
odes is determined by ve dierent fa
tors:
onstituent
en
oder design, interleaver design,
onstituent de
oder algorithm, interleaver size and
number of de
oder iterations. The latter three fa
tors require in
reases in delay or
omplexity in order to gain extra improvement. However, the rst two fa
tors
an a
hieve
improvement in performan
e without in
reases in
omplexity. Due to the presen
e of an
interleaver in the stru
ture of turbo
odes,
onventional design rules for
onvolutional
odes are not suitable. It was found that the performan
e of turbo
odes depends on the
minimum distan
e generated by weight two input sequen
es, the ee
tive free distan
e.
Previous
ode sear
hes involved nding
odes that have maximum ee
tive free distan
e.
Here the
ode sear
h was extended to
odes that maximise ee
tive free distan
e as well
as minimise the number of nearest neighbours. Further
riteria were also used to isolate
the best
ode when more than one
ode has a similar number of nearest neighbours and
the same ee
tive free distan
e.
Another way of improving the performan
e of turbo
odes is to design good interleavers so that maximum performan
e
an be a
hieved for the same blo
k size. Dierent
interleaver generation methods were investigated and their bit error rate performan
e for
parallel and serial s
hemes were
ompared. Symmetri
interleavers, whi
h have identi
al interleave and deinterleave patterns, were also studied. The symmetri
property
an
xi
xii
De
laration
I de
lare that this thesis presents work
arried out by myself and does not in
orporate
without a
knowledgment any material previously submitted for a degree or diploma in
any university; and that to the best of my knowledge it does not
ontain any materials
previously published or written by another person ex
ept where due referen
e is made in
the text.
Mark S. C. Ho
xiii
A
knowledgements
First I would like to thank my supervisors, Dr. Mark Ri
e, Dr. Steven Pietrobon and Dr.
Tim Giles, for their help and
ontributions to the thesis. I would also like to thank all the
sta members at the Institute for Tele
ommuni
ations Resear
h (ITR) at the University of
South Australia for their nan
ial and te
hni
al support whi
h made this study possible.
Espe
ially, I would like to thank Dr. Paul Gray, Dr. Mark Reed, John Buetefuer,
and Paul Petit for all of their dis
ussions and
omments related to this thesis.
Finally, I would like to thank my family in Hong Kong for their support throughout
the ten years that I have been studying in Australia.
xiv
Chapter 1
Introdu
tion
The stru
ture of a digital
ommuni
ation system is shown in Figure 1.1. At the transmitter
side we have the sour
e en
oder
onverting the analog or digital information sour
e to a
sequen
e of binary digits. The sequen
e of binary digits is passed to the turbo en
oder
and the output modulated by the digital modulator before being transmitted a
ross the
physi
al
hannel to the re
eiver. The modulator
an use s
hemes su
h as binary phase
shift keying (BPSK) or quadrature phase shift keying (QPSK). At the re
eiver side the
re
eived signal,
orrupted with noise and interferen
e from the
hannel, is demodulated
by the digital demodulator and its output de
oded by the turbo de
oder to remove any
errors generated due to the
hannel. The output of the turbo de
oder is then
onverted
ba
k to the transmitted information by the sour
e de
oder.
Information
source
Source
encoder
Channel
encoder
Digital
modulator
Channel
Information
sink
Source
decoder
Channel
decoder
Digital
demodulator
The role of the
hannel en
oder/de
oder pair is to prote
t the binary information
generated by the sour
e en
oder by adding redundan
y to the sour
e binary sequen
e.
This added redundan
y
an be used at the re
eiver to
orre
t errors generated due to
noise and interferen
e of the physi
al
hannel.
Dierent
lasses of error
ontrol
odes have been proposed in the last ve de
ades.
They in
lude Hamming
odes, BCH
odes,
y
li
odes, Reed-Solomon
odes,
onvolutional
odes, et
. All of these
odes are trying to get
loser to the Shannon limit: the
transmission of error free information at the full
apa
ity of the
hannel. The entire
oding so
iety is very ex
ited about the dis
overy of turbo
odes [1, a
lass of
on
atenated
onvolutional
odes that
an a
hieve performan
e very
lose to the Shannon limit. Sin
e
the announ
ement of turbo
odes in 1993, mu
h resear
h has been performed to gain
more understanding of their performan
e, su
h as better
onstituent
odes [2, 3, better
interleavers [4, 5, 6, 7, and simpli
ation of the de
oding algorithm [8, 9, 10.
This thesis is organised as follows: Chapter 2 presents the basi
s of turbo
odes.
We des
ribe the MAP de
oding algorithm [9, 11 and its appli
ation to parallel or serial
on
atenated turbo
odes. A
omparison is made between parallel and serial s
hemes.
We also give some indi
ations about how the performan
e of
onstituent
odes
an ae
t
the performan
e of turbo
odes.
In Chapter 3 we improve the performan
e of parallel
on
atenated turbo
odes by
nding
onstituent
odes that have maximum ee
tive free distan
e as well as minimum
number of nearest neighbours. Additional parameters are also used to nd better
odes
when there is more than one
ode that have identi
al parameters. The bit error rate (BER)
performan
e bounds with two dierent
ode rates are plotted to show that in
reasing the
trellis
omplexity of
onstituent
odes does improve the performan
e of turbo
odes when
maximum likelihood de
oders are used.
In Chapter 4 we look at dierent types of interleavers that
an be used with turbo
odes. We rst present the generation methods for a number of dierent interleaver types.
We then look at a
lass of interleavers that have identi
al interleave and deinterleave pat2
terns,
alled symmetri
interleavers [7. We nd that it is possible that various types of
interleavers, e.g., heli
al interleavers,
an also be symmetri
. In addition to the implementation advantages of symmetri
interleavers, we found that symmetri
interleavers
an
a
hieve better performan
e
ompared to non-symmetri
interleavers. We also investigate
the performan
e of dierent types of interleavers when applied to serial turbo
odes. We
found that, as opposed to the parallel s
heme, a spread interleaver [4 outperforms other
types of interleavers, in
luding symmetri
interleavers.
In Chapter 5 the impa
t of
hannel varian
e mismat
h on the BER performan
e of
serial
on
atenated
onvolutional
odes (SCCC) and parallel
on
atenated
onvolutional
odes (PCCC) was investigated. It was found that serial s
hemes are more sensitive to
hannel varian
e mismat
h when
ompared to parallel s
hemes. We reviewed two dierent
varian
e estimation s
hemes and applied them to the serial s
heme. We found that both
s
hemes work very well with a serial de
oder. The s
hemes also work well for small blo
k
sizes.
Chapter 3
{ We nd
onstituent
odes for PCCCs with maximal ee
tive free distan
e as
Chapter 4
3
Chapter 5
{ We found that SCCCs are more sensitive to varian
e mismat
h
ompared to
We found that both estimators perform quite well when applied to SCCCs.
For PCCC's, the Reed-Asenstorfer varian
e estimator performed better than
the Summer-Stephen varian
e estimator for small blo
k sizes.
Chapter 2
Fundamentals of Turbo
odes
2.1 Introdu
tion
Turbo
odes were introdu
ed by Berrou et. al. in 1993 [1. It is a
lass of
odes whi
h
on
atenate two or more re
ursive systemati
onvolutional (RSC)
odes in parallel, and
where the input to all ex
ept one RSC is rst permuted by an interleaver. These
odes
an
a
hieve a performan
e very
lose to Shannon's
apa
ity with large interleavers (several
kilobits). However, for small interleavers it
an still have amazing performan
e over other
on
atenated
odes.
Sin
e the input to ea
h
onvolutional en
oder is separated by an interleaver, Turbo
odes are en
oded in the form of a blo
k. Ea
h blo
k
onsists of the systemati
un
oded
data and the en
oded data from ea
h en
oder. At the de
oder side, a
omponent de
oder
is used to de
ode information from ea
h
omponent en
oder, with an interleaver and
a deinterleaver so that information
an be ex
hanged between the two de
oders. The
pro
ess of this ex
hange of information is
ontinued via iteration until the reliability of
the information is in
reased to almost one. In this
ase reliability is a measure of the
onden
e of a de
oded bit, ranging from zero (not reliable) to one.
In the following se
tion we dis
uss the idea of iterative de
oding for Turbo
odes.
We also look at some of the issues that ae
t the performan
e of Turbo
odes, su
h as
interleaver size, memory size and number of iterations. We note that the interleaver is a
permutation operation.
5
BPSK or
QPSK
modulator
Turbo
encoder
Channel
AWGN, Rayleigh, etc.
Information
sink
BPSK or
QPSK
demodulator
Turbo
decoder
2.2.1
PCCC en oder
PCCC is a
lass of
ode that
on
atenates two or more RSCs, termed
onstituent
odes (CC), in parallel. The input to one en
oder is not permuted. The input to the other
en
oders are permuted so that the other
onstituent en
oders are fed with interleaved
versions of the input information sequen
e. The Turbo en
oder stru
ture for a rate 1/3
Turbo
ode with two rate 1/2 RSC
onstituent en
oders and one interleaver () is shown
in Figure 2.2. We
an also use pun
turing at the output of the Turbo en
oder to in
rease
- d
s
s
CC1
CC2
0 1
2.2.2
SCCC en oder
An SCCC en
oder is similar to the serial
on
atenated
odes rst proposed by Forney
[12. The main dieren
e is that an interleaver is lo
ated between the output of the outer
en
oder and the input of the inner en
oder. Therefore, the output
odewords of the outer
ode are interleaved,
ompared to PCCCs where only the information bits are interleaved.
The en
oder stru
ture of a rate 1/3 SCCC, using a rate 1/2 outer en
oder and a rate 2/3
7
u1
h0
h1
u2
h2
h0
h0
h1
h1
h2
h2
h2
2
h2
h1
c
(a) Rate 1/2 RSC encoder.
h0
Figure 2.3: Examples of a four state RSC en
oder, (a) rate 1/2 and (b) rate 2/3.
inner en
oder, is shown in Figure 2.4. Although pun
turing is possible with SCCCs, it
O
CC
Outer
d I
CCI nner
z2 + Lx + z10
x
y1
z1 + Lx
6-
?-+
De oder 1
z2
d^
? +
r-
De oder 2
z1 + Lx + z20
of as a soft
orre
tion term that removes errors
aused by noise and interferen
e on the
hannel.
The de
oding of the se
ond soft de
ision de
oder is based on the se
ond set of transmitted symbols (y ) and the soft extrinsi
information from the last soft de
ision de
oder.
This allows the de
oder to make maximum use of all the information available from the
other de
oder. Note that we only need to input y into de
oder 2 to de
ode CC sin
e
the input information
ontains information about d as well.
At the output of the se
ond de
oder we obtain an output that
ontains both the
intrinsi
information from de
oder 1 and extrinsi
information from de
oder 2 (z + Lx +
z 0 ). This information
an either be used to make a hard de
ision or passed ba
k to de
oder
1 for another iteration.
Before another iteration starts, the output from de
oder 2 will go through a pro
ess
that is similar to that of the rst de
oder, where the infomation (Lx + z ) at the input
of the de
oder is subtra
ted from the output to obtain the extrinsi
information (z0 ).
This new information is deinterleaved by the deinterleaver ( ) and fedba
k to the rst
de
oder to
ontinue with the next iteration of de
oding. Deinterleaving
auses z0 to be
renamed z .
Sin
e the information is being re
y
led and improved by the de
oder at ea
h iteration
we expe
t the performan
e of the iterative de
oder to improve at ea
h iteration. As we
an see, an important fa
tor of iterative de
oding lies in the pro
ess and ex
hange of
extrinsi
information between ea
h de
oder to improve the performan
e at ea
h de
oder
iteration. One obvious question is how to obtain the soft output information in the rst
pla
e so that the extrinsi
information
an be generated. In the next se
tion we will look
at how the soft information
an be generated by using a soft de
ision de
oding algorithm.
2
PCCCs and SCCCs, the
omponent de
oder must be able to handle soft information both
at the input and output. This type of algorithm is required so that soft information
an
be passed between ea
h de
oder to improve the performan
e with ea
h de
oder iteration.
Dierent authors have proposed dierent types of algorithms that
an handle soft information, e.g., maximum a posteriori (MAP) [8, 11, 14, log-MAP [9, 10, max-log-MAP
[10, 15 and soft output Viterbi algorithm (SOVA) [16, 17, 18, 19 algorithms.
We rst present the MAP and log-MAP algorithms whi
h give the best performan
e
with iterative de
oding. This is followed by the sub-optimal max-log MAP and SOVA
algorithms (in terms of BER) whi
h give poorer performan
e with iterative de
oding.
2.4.1
The MAP de
oding algorithm was rst proposed by Chang and Han
o
k [14 to
minimise the symbol error probability for an inter-symbol interferen
e
hannel. Later,
Bahl et. al. [11, 20 and M
Adam et. al. [21 presented the algorithm for appli
ation
on the
oded
hannel. Berrou et. al. [1 modied the algorithm presented in [11 for
systemati
onvolutional
odes.
Sin
e then, a number of variations have been proposed. In [8, 9, 10 a simplied MAP
algorithm was proposed whi
h
an redu
e the de
oding
omplexity with no degradation in
performan
e. A sliding window version of the algorithm was also proposed whi
h allows
real time de
oding [8, 15. A simpler MAP algorithm
alled Max-Log-MAP or sub-MAP
was proposed in [10 whi
h performs only a fra
tion worse than the MAP algorithm at
low BERs. The MAP de
oder has also been implemented by a number of resear
hers su
h
as Pietrobon [9 and Barbules
u et. al. [22.
Here we give a des
ription of the MAP de
oding algorithm presented in [9. The job
of the MAP de
oder is, given a priori information of the
oded and un
oded data, to
output a posteriori probabilities of the
oded and un
oded symbols.
We dene Sk and dk as the en
oder state and binary information bit at time k. Ea
h
binary information bit is asso
iated with a state transition from time k to time k + 1.
This
hanges the en
oder state from Sk to Sk .
+1
11
Here we
onsider a rate 1/2 re
ursive systemati
en
oder with an output symbol at
time k whi
h
onsists of un
oded data dk and
oded bit
k . These symbols are modulated
with BPSK and transmitted a
ross an additive white Gaussian noise (AWGN)
hannel.
At the re
eiver we have the re
eived sequen
e
R1N = (R1 ; :::; Rk ; :::; RN )
(2.2)
where Rk = (xk ; yk ) is the re
eived symbol at time k. xk and yk are BPSK modulated
symbols dened as
xk = (2dk 1) + pk
yk = (2
k 1) + qk
where pk and qk are two independent normally distributed Gaussian noise variables with
zero mean and varian
e . If
omplex QPSK modulation is used the total noise varian
e
is 2 . We dene the log-likelihood ratio k asso
iated with ea
h information bit dk as
P r(dk = 0jRk )
(2.3)
k =
P r(dk = 1jRk )
where P r(dk = ijRk ) is the a priori probability of the data bit dk with value i = 0; 1.
The a posteriori probability of a de
oded data bit dk is
2
P r(dk = ijR1N ) =
where
i;m
k
N
i;m
k = P r (dk = i; Sk = mjR1 )
(2.4)
(2.5)
and m is a state index varying from 0 to 2 1, where is the en
oder memory. For
example, = 2 for the en
oder in Figure 2.3a. Combining (2.4) with (2.3) we have the
likelihood ratio as
X
X
k;m = k;m :
(2.6)
k =
0
The de
oder makes a de
ision based on the sign of the likelihood ratio
d^k =
0; k 1 :
1; k < 1
12
P r(RkN+1jdk = i; Sk = m; Rk )
P r(dk = i; Sk = m; Rk )=P r(R1N ):
(2.7)
Sin
e the events before time k do not depend on the observation after time k, the expression P r(Rk jdk = i; Sk = m; RkN )
an be simplied as
1
(2.8)
We dene this as the forward state metri
km. Similarly, the expression P r(RkN jdk =
i; Sk = m; Rk )
an be rewritten as
+1
(i;m)
P r(RkN+1jdk = i; Sk = m; Rk ) = P r(RkN+1jSk+1 = f (i; m)) = kf+1
(2.9)
where f (i; m) is the next state given an input i and
urrent state m. We dene this as
the reverse state metri
km . We dene the bran
h metri
as
0
+1
P r(dk = i; Sk = m; Rk ) = ki;m :
(2.10)
(2.11)
1
X
j =0
) j;b(j;m)
kb(j;m
k 1
1
(2.12)
where b(j; m) is the state going ba
kwards in time from state m given the input j . Figure
2.6 shows a graphi
al representation of the
al
ulation of km.
13
time = k
;m)
0;b(0;m)
PPPP
PPq h
d
1
d
j =1
d
d
d
(11
b
time = k
PPPP
PPPP
PPPP
j = 0PPP
(01
b
;m)
m
k
;b(1;m)
+1
km = P r(RkN jSk = m)
1
X
j =0
(2.13)
(j;m)
kj;mkf+1
:
Figure 2.7 shows a graphi
al representation of the
al
ulation of km. The bran
h metri
ki;m
an be
al
ulated using
ki;m = P r(dk = i; Sk = m; Rk )
(2.14)
where L
= 2= ,
i;m is the
oded bit given data bit dk = i and state Sk = m, ki =
P r(dk = i), xk and yk are the re
eived symbol at time k, and k is a renormalisation
onstant. Substituting (2.14) into (2.11) we
an evaluate (2.6) as
2
k
0
= k1 exp( L
xk )
k
= k exp( Lx )k0
!
(0;m)
=
km exp(L
yk
0;m )kf+1
14
!
(1;m)
km exp(L
yk
1;m )kf+1
(2.15)
time = k
time = k + 1
h
;m
j = 0
d
)
h
d
i
P
PPPP
d
PPPPj = 1
PPPP
P PP
PPPP
;m
f (0;m)
k +1
d
d
d
Ph
f (1;m)
k +1
2.4.2
One way to redu
e the
omplexity of the MAP algorithm is to operate in the logarithm domain. In this way all the multipli
ations and divisions be
ome additions and
subtra
tions [9, 10, 23. One problem of this method is the need to evaluate the sum of
all the exponentials. These operations
an be simplied using the method des
ribed in
[9, 10, where we
an express the log expression as
a
(2.16)
From a rst look this fun
tion may seem too
omplex to implement due to the log and
exponential fun
tion. If we plot the fun
tion ln(1 + e x) versus x we nd that the value
falls exponentially to 10 at x 4, whi
h is a small value that will not make mu
h
dieren
e in the nal output. Therefore it is possible to implement the fun
tion with a
2
15
ln k
Amk =
ln km
Bkm =
ln km
Dki;m =
ln ki;m:
m
k
2v
^1
m=0
1
j =0
1
^
j =0
where
A + Dk + B
0;m
m
k
f (0;m)
k+1
2v
^1
m=0
(1;m)
Amk + Dk1;m + Bkf+1
(2.17)
)
Akb(j;m
+ Dkj;b(1j;m)
1
(2.18)
(j;m)
Dkj;m + Bkf+1
(2.19)
n
^
j = 0
j =0
1
n :
= Kk (zk + L
xk )i L
yk
i;m
where zk = ln k is the log a priori probability and Kk =
(2.15)
an now be expressed as
Lk = zk + L
xk + zk0
ln k
ln k . Equation
0
(2.20)
where zk is the intrinsi
information and zk0 = ln k0 is the extrinsi
information.
It was mentioned in [10, 23 that an eight entry lookup table for log(1 + e x) is
su
ient to guarantee almost ideal performan
e. Figures 2.8 and 2.9 show the ee
t of
lookup table size on BER performan
e using QPSK for a PCCC and SCCC respe
tively.
The maximum entry value was limited to 10. Entries above this value were set to zero. As
16
an be seen from the simulation results the degradation in BER performan
e from using
an eight entry lookup table is very small (<0.1 dB)
ompared to using exa
t values. For
the PCCC
ase, a four state rate 1/3 en
oder (rate 1/2
onstituent
ode g=f7 5g) with a
1024 bit S=16 S-random interleaver was used with eight iterations. For the SCCC
ase,
a four state rate 1/3 en
oder (rate 1/2 outer
ode g=f7 5g and rate 2/3 inner
ode h=f7
3 5g) with a 2048 bits S=20 S-random interleaver was used with eight iterations.
The Log-MAP algorithm will be used as the
onstituent de
oder for the rest of this
thesis with a 32 entry lookup table.
0
10
4 entries
8 entries
16 entries
32 entries
exact
10
BER
10
10
10
10
0.2
0.4
0.6
0.8
1.2
1.4
Eb/No (dB)
Figure 2.8: Comparison of PCCC BER performan
e for dierent lookup table sizes.
2.4.3
Equations (2.17), (2.18) and (2.19)
an be further simplied by letting the fun
tion
V
log(1 + e x) = 0 and repla
ing all the fun
tions by the min fun
tion.
(0;m)
(1;m)
Lk = min
(Amk + Dk0;m + Bkf+1
) min
(Amk + Dk1;m + Bkf+1
)
m
m
(2.21)
)
)
Amk = min(Akb(0;m
+ Dk0;b(01 ;m) ; Akb(1;m
+ Dk1;b(11 ;m) )
1
1
(2.22)
(1;m)
(0;m)
; Dk1;m + Bkf+1
):
Bkm = min(Dk0;m + Bkf+1
(2.23)
17
10
4 entries
8 entries
16 entries
32 entries
exact
10
BER
10
10
10
10
0.2
0.4
0.6
0.8
1.2
1.4
Eb/No (dB)
Figure 2.9: Comparison of SCCC BER performan
e for dierent lookup table sizes.
This algorithm was rst used for the ISI
hannel [24, 25 and later applied to the
oding
hannel [26, 27. This algorithm has the same maximum likelihood performan
e as the
Viterbi and SOVA algorithms but requires a larger
omputation
omplexity [26. In
turbo de
oders, it performs slightly worse than MAP, but better than SOVA
onstituent
de
oders [26.
2.4.4
SOVA de oder
Traditionally,
onvolutional
odes were de
oded by the Viterbi algorithm. The drawba
k of the Viterbi algorithm is that it
an only generate hard estimates of the de
oded
symbols. This hard estimate will result in performan
e loss when apply to multistage de
oding, su
h as turbo
odes. To over
ome this drawba
k, the Viterbi de
oding algorithm
was modied to generate a soft estimate on the de
oded symbols so that it
an be used
in iterative de
oding. This modied Viterbi algorithm is
alled the SOVA algorithm.
The SOVA algorithm was rst proposed in [16. It is a basi
ally a modied version
of the Viterbi algorithm where the output sequen
e is augmented with soft information.
18
Other s
hemes of the modied Viterbi algorithm were introdu
ed by other authors [16,
17, 18, 19, 28, 29.
The operation of the SOVA algorithm is similar to the hard-output Viterbi algorithm
[30. Rather than output a hard de
ision value of the most likely
ode symbol sequen
e,
it produ
es a soft output based on the likelihood of the
ode symbol sequen
e.
The soft output of the SOVA de
oder at time t is obtained from the dieren
e of the
minimum path metri
among all the paths with symbol 0 at time t and the minimum path
metri
among all the paths with symbol 1 at time t. The sign of the dieren
e is used to
determine the hard estimate. The absolute value is used as the soft output information
for the next de
oding stage.
The input to the SOVA algorithm is identi
al to the Viterbi algorithm. For AWGN,
the bran
h metri
is
al
ulated based on the squared Eu
lidean distan
e between the
re
eived sequen
e and the modulated sequen
e. The path metri
is then
onstru
ted
based on the a
umulated path metri
from the previous state and the bran
h metri
that leads to the
urrent state. The path metri
s are then
ompared, with the minimum
path metri
sele
ted as the survivor. This is performed until the end of the re
eived
sequen
e is rea
hed.
A tra
e ba
k from the end of the sequen
e for the path with the minimum path
metri
is then performed. This is the maximum likelihood sequen
e. At the same time,
the dieren
e between the maximum likelihood path and the
omplementary path is
al
ulated to obtain a soft output for the de
oded symbol at time t. The de
oding
operation is nished when it rea
hes the beginning of the re
eived sequen
e.
A more in-depth explanation of the SOVA de
oding algorithm is given in [31.
PCCC
Figure 2.5 shows the de
oder stru
ture for a PCCC. and are the interleaver
and deinterleaver, respe
tively, whi
h are used to reorganise the soft information between
1
19
the two
onstituent de
oders. The en
oder stru
ture in Figure 2.2 is used for en
oding.
The Log-MAP algorithm des
ribed in Se
tion 2.4.2 is used as the
onstituent de
oder.
Unlike SCCCs, the CC
an be de
oded in any order provided that the
orre
t information is input into the MAP module. The operation of a PCCC is des
ribed in Se
tion
2.3.
Figure 2.10 shows the BER performan
e of a four state rate 1/3 PCCC with a 10,000
bit pseudo-random interleaver and six iterations. Note that ea
h CC should be terminated
0
10
10
iter 1
10
iter 2
BER
10
iter 3
4
10
iter 4
iter 5
10
10
iter 6
7
10
0.2
0.4
0.6
0.8
1
Eb/No (dB)
1.2
1.4
1.6
1.8
Figure 2.10: BER performan
e of a four state rate 1/3 PCCC, N=10,000.
to the zero state. The original turbo
ode [1 s
heme used dual termination, i.e., both CCs
terminated to the zero state. In [32 and [33 only one de
oder was terminated while the
other de
oder was set to the nal value of the forward re
ursion . In [34 it was found
that if either de
oders are not terminated then their respe
tive 's should be initialised
to zero. This
an a
hieve performan
e very
lose to the dual termination s
heme. We
all this the no termination s
heme. In [34 it was found that the dieren
es in BER
20
Original message:
To allow error free
ommuni
ation, error
ontrol
oding is used by adding redundant bits into the transmitted information sequen
e. Turbo
odes are a
lass of
onvolutional error
ontrol
odes whi
h
an a
hieve performan
e
lose to Shannon's
theoreti
al \
oding barrier".
No oding:
D ill u erro free oimqny
atign, erj r bontjml
o$in'$is usee by0add ng"rddtneanp
bits into$the trao3mitvmD inf vmition!seq5en
e. Turbo
Mdds are$a*ahass!of"
onvo
muvion eprnr
kntrol
/d s wh)
h0
an a
hieve"tOrformanse #loSe tm(Tje Shann/n's
th%ozeti
al""
oding ji" iez"* D
illow error freep oiluni
ation, error
ontrol
oling is used by aeding redundant bits
into the transmitted inf rmati n sequen
e. Turbo
odes are a
lass of
onvolutional
error
ontrol
odes whi
h
an a
hieve PorforMan
e
lose tm Shannon's theoreti
al
\
oding barrier".
To allow error free
ommuni
ation, error
ontrol
oding is used by adding redundant bits into the transmitted information sequen
e. Turbo
odes are a
lass of
onvolutional error
ontrol
odes whi
h
an a
hieve performan
e
lose to Shannon's
theoreti
al \
oding barrier".
As we
an see, parts of the text are
orre
ted at ea
h iteration. The
orre
t estimation
of these parts of the text helps the turbo de
oder to
orre
t other parts of the text in
subsequent iterations.
2.5.2
SCCC
Figure 2.11 shows the de
oder stru
ture for an SCCC with the Log-MAP algorithm
as the
onstituent de
oder. Note that the de
oding order of ea
h CC is xed sin
e the
outer
ode is en
oded before the inner
ode. Therefore, the inner
ode must be de
oded
rst before the outer
ode
an be de
oded.
The de
oding of SCCCs is similar to that of PCCCs. The re
eived symbols xI and yI
are fed into the inner de
oder MAP I . At the rst iteration the extrinsi
information zO
is set to 0. After de
oding, the output extrinsi
information L
xI + zI0 is deinterleaved and
input into the outer de
oder as a priori information L
xO + zI . After de
oding we have
output zI + L
xO + zO0 . zI + L
xO is subtra
ted from the output of the se
ond de
oder
and interleaved before input to the inner de
oder as extrinsi
information zO for the next
iteration. Note that MAP O needs to provide outputs for both the data and parity bits,
while MAP I needs to provide outputs for the data bits only.
For the last iteration, a hard de
ision is made on the output of the outer de
oder
zI + L
xO + zO0 . Figure 2.12 shows the BER vs Eb =N performan
e of a four state rate
0
22
x ;y
I
MAPI
Lx
z0
Lx
-
6
?-
r?
Lx
MAPO
?
r
Figure 2.13 shows a
omparison of BER performan
e between a four state rate 1/3
PCCC and SCCC with six de
oder iterations. As we
an see, we have a trade o between
using PCCC and SCCC at dierent Eb =N . At Eb=N ranges between 0 to 0.9 dB, PCCC
performs better than SCCC. After 0.9 dB, the PCCC rea
hes what is known as an \error
oor" or
are, whi
h in fa
t is when the performan
e of the PCCC is limited by the
free distan
e of the turbo
ode. However, the performan
e of the SCCC is still in its
\waterfall" region where the performan
e improves rapidly as Eb =N in
reases. In fa
t,
the
are region of SCCCs is mu
h lower than that of PCCCs. This makes it an attra
tive
alternative to PCCCs if low BER performan
e is required.
0
23
d^
2.5.3
z0
10
iter 1
iter 2
10
iter 3
2
10
iter 4
BER
10
10
iter 5
5
10
10
iter 6
7
10
0.2
0.4
0.6
0.8
1.2
1.4
Eb/No (dB)
Figure 2.12: BER performan e of a four state rate 1/3 SCCC, N=10,000.
10
10
10
PCCC
SCCC
BER
10
10
10
10
10
0.5
1.5
Eb/No (dB)
Figure 2.13: BER performan
e
omparison of four state rate 1/3 PCCC and SCCC,
N=10,000.
2.6.1
SCCC
SCCCs being similar to traditional
on
atenated
odes,
an have an inner and outer
outer
ode with dierent
omplexity. It was observed in [36 that the inner
ode should
have as low
omplexity as possible sin
e the inner
ode is working at very low Eb=N and
has the most
onta
t with the
hannel.
0
2.6.2
PCCC
For PCCCs, it was found in [37 that at a BER range between 10 and 10 , turbo
odes
an give better performan
e with shorter memory
onstituent
odes and a simpler
de
oding
omplexity. However, the high
are
aused by the poor dmin of the
onstituent
ode does not make them suitable for low BER appli
ations.
The BER performan
e of the
onstituent
odes improves as soon as the
omplexity
in
reases. However this improvement in performan
e only o
urs at higher Eb =N where
the improved free distan
e starts to gain advantage. Therefore, in
reasing the performan
e
2
25
g0 g1 dfree d2
2 3 2
4 7 5
8 13 17
16 23 33
32 45 77
3
5
6
7
8
3
6
8
12
20
26
10
s1=4, s2=4
s1=8, s2=4
s1=4, s2=8
1
2
s =8, s =8
3
BER
10
10
10
10
1.2
1.4
1.6
1.8
2
Eb/No (dB)
2.2
2.4
2.6
2.8
Figure 2.14: PCCC BER performan
e with asymmetri
omplexity CC, mixture of four
state and eight state
odes, 192 bit interleaver, six iterations.
1
10
s1=2, s2=2
s1=4, s2=2
1
2
s =2, s =4
s1=4, s2=4
10
BER
10
10
10
10
0.5
1.5
2.5
3.5
Eb/No (dB)
Figure 2.15: PCCC BER performan
e with asymmetri
omplexity CC, mixture of two
state and four state
odes, 192 bit interleaver, six iterations.
27
10
s=2
s=4
s=8
s=16
1
BER
10
10
10
10
Eb/No (dB)
Figure 2.16: BER performan
e of rate 1/2 RSC with memory lengths 1,2,3 and 4.
However, after 2.1 dB, the performan
e starts to
rossover and de
oding the more
omplex
four state
ode rst gives better performan
e. The
ode rate dieren
e is 10 log 3/2 =
1.76 dB, whi
h is
lose to the
rossover point.
28
Chapter 3
Constituent
odes for PCCC
de
oders.
3.1 Introdu
tion
A PCCC is a parallel
on
atenation of two or more
onvolutional
odes. Therefore,
the
hara
teristi
of ea
h
omponent
ode is important in order to improve the performan
e of Turbo
odes. A
ommon way of measuring the performan
e of
onvolutional
odes is to nd their distan
e spe
trum.
Although numerous
ode sear
hes for good
onstituent
odes for PCCCs have been
performed, most of the results are based on maximising the input weight two free distan
e of the
ode. In order to further improve the performan
e we should also minimise
the number of nearest neighbours. In this
hapter we will look at some of the suitable
onstituent
odes that are sele
ted based on a dierent set of
riteria
ompared to other
published results.
se
ond method uses an exhaustive sear
h method based on the Viterbi algorithm. This
involves nding error paths that diverge from the all zero
odeword by extending the
trellis of the
onstituent
onvolutional
ode. This method is more suitable for
omputer
implementation and it
an generate the spe
tral lines of interest e
iently.
Note that the distan
e spe
trum we des
ribe in this se
tion are semi-innite, i.e.,
the distan
e spe
trum
an be extended for as long as the input information sequen
e is
required. This is dierent from the nite distan
e spe
trum that is required by turbo
odes due to the presen
e of a nite length interleaver in the turbo en
oder. The length
of the interleaver limits the length of the distan
e spe
trum required. The pro
ess of how
to apply the semi-innite distan
e spe
trum to turbo
odes is
overed in Se
tion 3.3.
Before we go any further, we will need a brief understanding of the distan
e spe
trum
of
onvolutional
odes and PCCCs.
3.2.1
Convolutional
odes were rst introdu
ed as an alternative to blo
k
odes [39. Dierent de
oding s
hemes su
h as sequential [40, 41, threshold [42 and maximum likelihood
[40, 41 have been used to de
ode
onvolutional
odes.
It is known that the asymptoti
performan
e of
onvolutional
odes for AWGN
an be
improved by maximising the free distan
e of the distan
e spe
trum [40. The free distan
e
of a linear
ode is determined by the smallest non-zero output weight of a
onvolutional
ode's distan
e spe
trum. However, due to the presen
e of an interleaver, it is extremely
di
ult to nd the distan
e spe
trum of a PCCC. To over
ome this, the weight spe
trum
of a PCCC is found by averaging the weight spe
trum over all interleavers,
alled an
\uniform" interleaver [43. The basi
idea is to
ombine the distan
e spe
tra of the two
onstituent
odes to
reate a super spe
tra whi
h
ontains all possible
ombinations of
all the paths. Ea
h element of this super spe
tra is averaged based on the size of the
interleaver and the size of the
ombined error path. We will look in more detail of this
method later in the
hapter. Due to this new te
hnique, it was found in [43, 44 that we
need to alter the desired
hara
teristi
of the
onstituent
onvolutional
odes in order to
30
dw
X df
(3.1)
where N is the interleaver size, dw is the weight of the input information sequen
e that
generates the
ode sequen
e of output
ode weight df , and X is the bran
h gain for an
output
ode sequen
e with weight df .
Note that we
an only minimise this rst term for N ! 1 if dw 2. When dw = 2
we get a multipli
ation fa
tor of N in the nal distan
e spe
trum, whi
h will result in
the distan
e spe
trum being averaged based on the interleaver size. In [44 this pro
ess is
alled spe
tral thinning or interleaver gain in [43. For other spe
trum terms proportional
to N d X d where d0w > dw or d0f > df , the spe
trum terms will have less
ontribution
to the bound as SNR in
reases. Therefore, the main
ontribution at high SNR is from
the rst term of the distan
e spe
trum.
Note that for the
ase when dw = 1 the 1=N fa
tor disappears from (3.1). This means
the size of the interleaver no longer has any ee
t on the spe
tral density of the distan
e
spe
trum. Thus, we want to avoid any
ontribution from input information weight one
odewords.
For non-re
ursive
onvolutional en
oders, it is possible to generate a low weight
output for an input information weight one information word. A weight one information
word input into a non-re
ursive
onvolutional en
oder will always be self-terminating [45
(i.e., remerge with the all zeros path with weight one input information word). Sin
e a
PCCC only permutes the information bits, a weight one input will always generate error
events of low weight for both
onstituent
odes, no matter how it is permuted. Therefore,
onvolutional en
oders with nite impulse responses are not suitable for PCCCs where we
want to use the interleaver to provide us with diversity between ea
h
onstituent en
oder.
On the other hand, re
ursive systemati
onvolutional (RSC) en
oders when using
feedba
k polynomials require a nonzero number of bits to return the en
oder to the all
1
31
zero state. That is, all information words will have a weight that is two or greater.
Sin
e only a small fra
tion of these inputs generate low weight
ode words, the
han
e
that a low weight input for one
onstituent
ode will generate a low weight
odeword for
the other
ontituent
ode is greatly redu
ed. That is, the interleaver is said to breakup
low weight
odewords from one
onstituent
ode to high weight
ode words of the other
onstituent
ode. This basi
ally eliminates the problem of a self-terminating weight one
input information sequen
e that the interleaver
annot breakup.
For re
ursive
odes we need to
onsider weight two or higher self-terminating input
information words whi
h
an produ
e a low weight output
odeword. In [44, it was
said that with the help of an interleaver between ea
h en
oder, most of the low weight
self-terminating input information words will have a higher
han
e of being broken-up
to produ
e an output
odeword that is of high weight at the se
ond en
oder. Also,
all the high weight input information words, after permuting, will have a mu
h higher
han
e of being broken-up than those of lower weight. Therefore, the performan
e mainly
depends on the weight of the output
odeword generated by the input information word
of lowest weight. For the
ase of RSCs, the lowest possible input information weight that
an generate a self-terminating sequen
e is the weight two input information sequen
e.
Hen
e, the problem of nding good
odes for PCCCs lies in nding
onstituent re
ursive
en
oders that have maximum output weight for weight two input sequen
es. We
all these
input sequen
es the ee
tive free distan
e [43 of turbo en
oders, as
ompared to the free
distan
e of a
onvolutional
ode.
3.2.2
32
Aw;j W w Z j
(3.2)
where A(W; Z ) is the number of output weight j
ode words generated by an input
information word of weight w.
For the re
ursive
onvolutional en
oder in Figure 3.1 we
an derive the state diagram
and its trellis as shown in Figure 3.2.
d
D
c
Figure 3.1: En
oder stru
ture for a (2,1,2) RSC
ode with generator polynomials f7,5g .
8
In the trellis diagram ea
h path is
onne
ted between a starting state S and ending
state S 0 with an (input information word/output
ode word) pair asso
iate with ea
h
path. For the state diagram, ea
h transition from one state to another has a
ertain input
weight W and output weight Z asso
iated with it. This basi
ally des
ribes the weight
of the input
odeword that
aused the transition and the weight of the output
odeword
generated. Hen
e the transition from state S 0 to S 2 of the rate 1/2 RSC
ode is
aused
by a weight one input W whi
h generates a weight two output Z .
Note that the state diagram is a
losed graph, i.e., it does not
onvey information for
the beginning and ending states whi
h is required in order to analyse error paths. Sin
e
all error paths diverge from the zero state and remerge with the zero state, we
an modify
the state diagram by splitting the zero state to a starting state and ending state (note that
splitting the zero state only works for linear
odes and
odes with
ertain symmetries).
This is
alled a modied state diagram. The modied state diagram of the rate 1/2
ode
is given in Figure 3.3.
2
3.2.3
This method uses the modied state diagram to nd the distan
e spe
trum of the
ode [46 by deriving the state transition matrix A from the state diagram. A state
33
1
S
d/dc
0/00
00
00
S0
1/11
ZW
ZW
1/11
01
01
0/00
1/10
1
S1
10
S2
10
ZW
0/01
0/01
11
11
1/10
Z
S3
Trellis Structure
ZW
State Diagram
Figure 3.2: Trellis stru
ture and state diagram for a (2,1,2) RSC en
oder with generator
polynomials f7,5g .
8
ZW
S3
Z
ZW
ZW
ZW
S0
S2
S1
1
S0
transition matrix
ontains all the parameters going from one state to another. By writing
out the state equations in a matrix form we
an nd the generating fun
tion by solving
the state equation going into the terminal state S 00. Note that the state transition matrix
an be also be solved by using Mason's formula [47.
The state transition equation is given in the following form
x = x + x0
(3.3)
M = 2
xT = (x1 xM )
(x )T = (x xM )
0
0
1
xM
6
xM
= 66 ..
... ...
4
.
xS xS xM
S
x11
x12
x21
x22
00
00
00
1
1
xS1 0
xS2 0
...
xSS 00
0
0
3
7
7
7
5
0
0
where xi is the path gain from state S 0 to state Si and xji is the path gain from state j to i.
is the state transition matrix whi
h des
ribes the distan
e and input weight generated
when going from one state to another. x
ontains the distan
e and input information
weight generated by going from state zero. From the example in Figure 3.3 we obtain the
following state transition matrix and x .
0
2
6
= 64
0
1
0
Z 2W
ZW
0
0 0
ZW 0
0 0
Z
3
7
7
5
6 Z 2W
; x0 = 6
4
0
7
7
5
and x = 64
35
x1
x2
x3
x4
3
7
7
5
By solving A(W; Z ) for x we
an nd the generating fun
tion for our (2,1,2) RSC
ode,
whi
h is
4
Z 5 W 2 (Z + W Z 2 W )
1 2ZW + Z 2W 2 Z 2
= Z 5W 2(W + Z + ZW 2 + Z 2 (3W + W 3) + )
A(W; Z ) = x4 =
= Z W + Z W + Z W + 3Z W + Z W + :
5
The polynomial above shows that we have one path with output distan
e 5 whi
h is
generated by a weight 3 input sequen
e, i.e., Z W , et
. Note that A(W; Z ) is an innite
polynomial, whi
h means that the spe
trum of the
ode is innite. By using the transfer
fun
tion method des
ribed above we
an nd the entire spe
trum of the
onstituent
ode.
However, as we
an see it is not easy to implement. In the next se
tion we will des
ribe
a method that is more suitable for implementation.
5
3.2.4
Dierent to the method in Se
tion 3.2.3, the unidire
tional algorithm [48 makes
use of the trellis instead of the modied state diagram. Essentially, the uni-dire
tional
algorithm is similar to the Viterbi algorithm. Rather than nding the survivor path at
ea
h state of the trellis, all paths are extended until at least one path re-merges with
the all zeros path. The input and output weights of the re-merged paths are stored to
form the weight spe
trum. This method is more suitable for
omputer implementation
sin
e it is similar to the Viterbi algorithm. Here we will brie
y des
ribe the unidire
tional
algorithm.
The algorithm begins by extending all the paths that start from state zero at t=0
that do not terminate at the zero state. We then move to the next trellis se
tion and
extend all the paths that were extended in the last trellis se
tion. If one of the paths
remerges with the zero state, the a
umulated input and output weights of that path are
re
orded. Otherwise, the rest of the paths are extended until some limit is rea
hed. This
limit is usually dened to be when all the metri
s ex
eed some distan
e. Figure 3.4 shows
the unidire
tional algorithm for the rst six trellis se
tions.
36
Input weight
Output weight
S
0/0
1/2
00
00
1/2
01
01
0/0
1/1
10
10
0/1
0/1
1/1
11
11
3/5
2/6
4/6,3/7
3/7(2),4/8
1/4
3/4,2/5
2/5(2),3/6
1/6...
2/3
1/4
1/3
2/4
2/4,3/5
1/5
t=2
t=3
t=4
t=5
0/0
2/8...
00
2/3
01
10
1/2
11
t=0
t=1
37
t=6
t=7
For trellises that have parallel paths this method is more suitable for
omputer implementation. Therefore, this method was sele
ted. To simplify the sear
h the feedba
k
polynomial was limited to only primitive polynomials. This is due to the fa
t that re
ursive
onvolutional
odes will have maximum period if the feedba
k polynomial is primitive
[44. This will give a maximum period of 2 1. Hen
e, h or g has to be primitive. A
list of primitive polynomials for dierent en
oder memory lengths are given in [49.
We denote the minimum output weight of ea
h remerged path,
aused by input weight
i, as di . From Figure 3.4 we have one d = 5 (input weight three output weight ve) path
at t=2 and one d = 6 path at t=3. Note that it is possible to have a multiple number of
remerges that have the same input weight and output weight, su
h as the d = 7 path at
t=6. We
all this the multipli
ity, or the number of neighbours. It is represented by Ni ,
for an input weight i.
0
given in [43. First, we need to dene the input-redundan
y weight enumerating fun
tion
(IRWEF) of an (n; k) systemati
onvolutional
ode C as
AC (W; Z )
w;j
(3.4)
ACw;j W w Z j
where ACw;j denotes the number of output
odewords generated by an input information
word of Hamming weight w with Hamming weight j output parity bits. W and Z are
dummy variables for the information and parity bits. The value of ACw;j
an be obtained
from (3.2). This fun
tion des
ribes the
ontributions of information and parity bits to the
total Hamming weight of the output
odeword, whi
h is (w + j ) for a systemati
ode.
From the IRWEF, we
an obtain the
onditional weight enumerating fun
tion (CWEF)
of the CC
X
1
w AC (W; Z )
C
C
j
Aw (Z ) = Aw;j Z =
:
w!
W w W
j
=1
The nite length of an interleaver ee
tively makes the
onstituent
onvolutional
ode
a blo
k
ode. Therefore, we need to nd the CWEF representation of the
onstituent
onvolutional
ode based on the equivalent blo
k
ode. The equivalent blo
k
ode representation of a
onvolutional
ode is obtained by nding all
ombinations of a set of error
events of the
onvolutional
ode within a length equal to N. The blo
k equivalent CWEF
of the
onvolutional
ode is [43
ACw (Z ) =
Aw;j Z j
with
Aw;j
K [m; nTw;j;m;n
(3.5)
m;n
where mmin m N , 1 n bm=mmin
, mmin is the shortest length error path and
N
m
+
n
K [m; n =
:
n
K [m; n is the expression for the multipli
ity of the output
odewords produ
ed by a single
ombination of n error events with total length m. Tw;j;m;n is the number of
on
atenated
39
paths in the trellis produ
ed by an input information sequen
e of weight w and output
parity weight j with length m and n remerges with the zero state. Note that the n
remerges are subje
t to the
onstraint that they leave the zero state immediately after
remerging.
We
an now obtain the CWEF of a PCCC by using the uniform interleaver
on
ept
we mentioned previously. The uniform interleaver makes the CWEF of the se
ond CC
independent of the rst CC. Therefore, we
an express the CWEF of a PCCC as the
produ
t of the CWEFs of two CCs
ACwp (Z ) =
(3.6)
Finally, the bit error probability of a PCCC based on the uniform interleaver is upper
bounded by [43, 44
Pb
(2n
1)N
2d REb
wd
Q
N
d=df ree
X
N0
(3.7)
where n is the number of
ode bits of the
onstituent
ode and N is the interleaver size.
Also, d = w + j is the Hamming weight of the output
odeword generated by an input
information word of Hamming weight wd and R is the overall
ode rate of the turbo
ode.
The value of wd is obtained from the super distan
e spe
tra
al
ulated from (3.6).
For high Eb=N (bit error rate 10 ), the performan
e is dominated by the rst
term of the bound. Hen
e, we
an simplify (3.7) to
5
Pb
wfree
Q
N
RE
2dfree N b
0
where
dfree=ee
tive free distan
e of the turbo
ode.
wfree=Hamming weight of the information sequen
es
ausing the free
distan
e
odewords.
40
(3.8)
As we
an see from (3.8), in order to minimise Pb we should minimise the Q fun
tion.
Sin
e the Q fun
tion is a negative exponential fun
tion, we should maximise dfree for a
given
ode rate R and Eb=N . For a xed interleaver size N we should also minimise
wfree. For PCCCs, the ee
tive free distan
e is dependent on the free distan
e generated
by weight two input information sequen
es (sin
e the interleaver
annot breakup input
information weight one sequen
es). Therefore, to minimise the bit error probability of
PCCCs we should maximise d and minimise N of the
onstituent
onvolutional
ode.
0
41
min.
not min.
13
15
17
13
07
17
5(1)
4(1)
5(2)
4(2)
dmin
4(1)
4(3)
The se
ond generator polynomial f13,07,17g has a larger number of nearest neighbours at d , dmin and d . The trun
ated union bound and simulation results of both
of these
odes for a rate 1/2 PCCC en
oder with a 400 (S=10) and 4096 (S=25) bit
S-symmetri
interleaver [7 and six de
oder iterations are shown in Figures 3.5 and 3.6,
respe
tively. As we
an see for both interleaver sizes, the
onstituent
ode with a smaller
number of neighbours performs better than the one with more neighbours at low BER;
even-though they have very small dieren
es in distan
e proles. This indi
ates that
the number of nearest neighbours
an have an impa
t on the performan
e of PCCCs,
espe
ially at low BER.
2
The trun
ated union bound in Figures 3.5 and 3.6 in
ludes all distan
e terms up to
approximately twi
e d . This is used for all the bounds generated in this thesis, ex
ept
for
odes where d is greater than 60, where all distan
e terms up to d are used.
2
Figures 3.5 and 3.6 also show a large dieren
e in performan
e between a suboptimum de
oder with a parti
ular interleaver and the trun
ated union bound for a uniform
interleaver. Due to a nite number of terms the bound at low Eb =N does not diverge at
the
ut-o rate. At high Eb=N , the performan
e is limited by the free distan
e at the
0
42
ode. In this
ase, even with sub-optimum de
oding, the parti
ular interleaver
hosen
performs better than the bound at high Eb=N (where the bound is a
urate). Even with
quite dierent performan
e results for dierent interleavers, the ee
t of minimising N
appears to be valid.
0
10
min N
2
not min N2
1
10
10
BER
10
10
10
10
10
10
0.5
1.5
2
Eb/No (dB)
2.5
3.5
Figure 3.5: Trun
ated union bound and BER performan
e of rate 1/2 PCCC using eight
state rate 2/3
onstituent en
oders, N=400 and 6 de
oder iterations.
43
10
min N2
not min N2
1
10
10
BER
10
10
10
10
10
0.2
0.4
0.6
0.8
1
Eb/No (dB)
1.2
1.4
1.6
1.8
Figure 3.6: Trun
ated union bound and BER performan
e of rate 1/2 PCCC using eight
state rate 2/3
onstituent en
oders, N=4096 and 6 de
oder iterations.
rst
ode based on d , the se
ond on dmin and the third d .
2
max.
max.
dmin
max.
dmin
13
15
17
5(1)
4(1)
4(1)
13
11
17
4(1)
4(1)
5(6)
17
07
13
4(3)
4(3)
>10
Table 3.2: Eight state rate 2/3
onstituent
odes that were sear
hed a
ording to dierent
riteria.
These
odes were simulated using the standard PCCC stru
ture with a 400 bit (S=10)
and 4096 bit (S=25) S-symmetri
interleaver and six iterations. The average bound and
simulation performan
e are presented in Figures 3.7 and 3.8. The solid lines represent
bound results and the dashed lines represent simulation results.
As
an be seen from both simulation results, at a BER of 10 the
ode with maximum d performs better then the other two. The maximised dmin
ode is only slightly
6
44
10
Max. d2
Max d
min
Max d
10
10
BER
10
10
10
10
10
10
0.5
1.5
2
Eb/No (dB)
2.5
3.5
Figure 3.7: Trun
ated union bound and BER performan
e between dierent
riteria of
maximisation for N=400 and 6 de
oder iterations.
0
10
max. d2
max. d
min
max. d
3
10
10
BER
10
10
10
10
10
0.2
0.4
0.6
0.8
1
Eb/No (dB)
1.2
1.4
1.6
1.8
Figure 3.8: Trun
ated union bound and BER performan
e between dierent
riteria of
maximisation for N=4096 and 6 de
oder iterations.
45
worse than the rst
ode. The
ode that maximised d performs mu
h worse than the
other two, even-though they have the same dmin and d . Thus, our sele
tion
riteria is to
nd the
ode with the best d , followed by dmin, and then by d .
3
3.7 Results
In [52 a list of the best rate 1/n and k/(k+1) polynomials were given. The sele
tion
riteria for the best polynomials was based on maximising d only. We extend the results
to maximise d as well as minimise N . Extra
riteria were used to separate the best
odes if multiple
odes with identi
al d and N exist. Examples of eight state rate 1/2
and rate 2/3 en
oder stru
tures are given in Figures 3.9 and 3.10, respe
tively.
2
m
+m
g
?g -r 6rg -r 6rg -r
x- +m
r
D
D
D
g
g
g
g
?
?
?
rg - +m g - +m g - +m g - y
- +?m - +?m - +?m- y
+
00
10
01
02
03
11
12
13
20
21
22
23
Figure 3.9: En
oder stru
ture for rate 1/n RSC
onstituent
ode (eight state rate 1/3
example).
rh rh rh rh - y
-y
r
r
r
xh r
h
h
h
??- -
??- -
??- -
??-r
+
+
+
+
y
D
D
D
6h
6rh
6rh
h
x1
13
12
11
23
22
03
21
02
10
20
01
00
Figure 3.10: En
oder stru
ture for rate k/(k+1) systemati
feedba
k
onvolutional
ode
(eight state rate 2/3 example).
Here we list the best generator polynomials for various rate 1/n and rate k/(k+1)
onstituent
odes. Note that for the
ase of rate k/(k+1), the en
oder stru
ture is more
46
ommonly refered to as a systemati
feedba
k
onvolutional
ode. The
ode sear
h was
performed using the unidire
tional Viterbi algorithm with the following order of sele
tion:
d ; N ; dmin ; Nmin ; d ; N . In the following tables m is the memory size of the en
oder.
The trellis size is equal to 2m. Note that g and gi (i = 1; 2; 3; ) are the feedba
k and
feedforward polynomials for rate 1/n
onstituent en
oders. h and hi (i = 1; 2; 3; ) are
the feedba
k and feedforward polynomials for rate k/(k+1)
onstituent en
oders.
The order of
ode sele
tion is dierent from those presented in [2 where the sele
tion
order was d ; N ; d ; N ; d ; N ; d ; N . The goal of our sele
tion is to a
hieve the best
performan
e at high Eb =N , where the performan
e is dominated by the minimum distan
e
of the PCCC
ode. Therefore, asymptoti
ally our
odes should a
hieve the same or better
performan
e than those published. In Tables 3.3 to 3.11, represents
odes that are in
[52 and represents
odes that are in [2. Note that the k/(k+1) en
oder stru
ture in
[2 is dierent to that presented here, but they should have similar distan
e spe
trum
properties. The distan
e spe
trum of the original pun
tured turbo
ode [1 is given in
Table 3.4 () for referen
e.
2
1+
+
3 +
4 +
2
+
6+
dmin
3(1)
3(1)
5(6)
6(1)
5(1)
6(1)
13
17
6(1)
8(1)
7(3)
23
37
6(1)
12(1)
8(3)
23
33
7(2)
12(1)
23
35
7(2)
75
53
8(1)
67
45
45
7(1)
8(1)
6(1)
9(9)
8(1)
6(1)
10(14)
8(1)
7(1)
8(4)
7(1)
10(6)
12(1)
7(1)
8(3)
7(1)
10(8)
20(1)
9(1)
8(1)
9(7)
10(5)
8(3)
20(1)
10(2)
8(1)
8(2)
10(5)
77
8(2)
20(1)
9(3)
8(1)
11(14)
8(1)
45
67
8(3)
20(1)
8(2)
8(1)
10(6)
10(5)
147
101
7(1)
36(1)
13(1)
8(1)
7(1)
10(2)
147
135
10(12)
36(1)
12(1)
10(1)
10(6)
10(4)
155
117
10(11)
36(1)
10(1)
10(1)
10(4)
10(3)
313
275
10(1)
68(1)
15(1)
12(3)
11(6)
10(1)
47
3
4
2
dmin
2(1)
2(1)
6(4)
9(4)
3(1)
4(3)
3(1)
4(1)
5(1)
13
15
4(3)
5(2)
4(2)
4(1)
6(10)
6(1)
6(2)
23
33
5(7)
6(1)
5(5)
6(19)
5(2)
6(3)
23
35
4(1)
6(1)
4(1)
6(13)
6(5)
6(1)
23
31
4(1)
7(2)
4(1)
6(3)
6(1)
37
21
4(4)
4(3)
5(2)
4(1)
5(2)
6(2)
67
45
5(1)
11(2)
6(2)
6(4)
5(1)
7(6)
147
165
6(7)
19(2)
6(2)
6(2)
6(2)
6(1)
147
117
5(1)
19(2)
7(2)
5(1)
6(1)
7(2)
367
255
6(3)
35(2)
8(1)
7(2)
6(3)
7(4)
Table 3.4: Best rate 2/3 pun
tured rate 1/2
onstituent
odes with pun
turing pattern
[1 0.
m
1+
+
5
+
6
dmin
5(1)
8(2)
8(1)
8(1)
10(1)
12(1)
14(1)
7(1)
10(1)
7(1)
8(1)
9(1)
10(1)
17
10(3)
14(1)
10(2)
10(1)
14(6)
14(2)
17
17
8(1)
14(1)
11(3)
8(1)
13(9)
10(1)
23
25
37
11(1)
22(1)
11(1)
12(2)
15(4)
12(1)
23
33
37
10(1)
22(1)
12(1)
10(1)
12(1)
14(2)
23
27
35
11(1)
22(1)
11(1)
12(2)
13(3)
14(2)
67
45
57
13(2)
38(1)
15(1)
14(2)
13(2)
14(1)
67
45
51
11(1)
38(1)
17(2)
16(6)
11(1)
14(2)
155
117
127
15(2)
70(1)
15(1)
18(1)
15(1)
16(3)
147
101
115
11(1)
70(1)
23(1)
14(1)
11(1)
16(2)
5(1)
13
15
13
2+
3+
1+
+
3 +
4
+
2
dmin
6(1)
6(1)
1
1
1
1
1
1
1
1
7(1)
7(1)
10(2)
10(1)
10(1)
14(2)
16(1)
18(1)
13
11
15
17
12(1)
20(1)
12(1)
14(1)
18(2)
20(3)
23
33
37
25
15(1)
32(1)
15(1)
16(2)
17(1)
18(1)
23
27
35
37
14(1)
32(1)
16(1)
14(1)
18(3)
18(1)
23
33
27
37
14(1)
32(1)
16(1)
14(1)
16(1)
20(3)
23
35
33
37
14(1)
32(1)
16(1)
14(1)
16(1)
20(3)
75
45
57
67
18(3)
56(1)
20(1)
18(2)
18(1)
22(2)
67
51
45
71
15(1)
56(1)
23(1)
20(3)
15(1)
20(2)
dmin
8(1)
8(1)
12(2)
12(1)
12(1)
16(1)
18(1)
20(1)
13
11
15
16
17
15(1)
24(1)
15(1)
16(1)
23(2)
22(1)
23
25
27
35
37
19(1)
42(1)
19(1)
20(1)
23(3)
20(1)
23
27
33
35
37
18(1)
42(1)
20(1)
18(1)
20(1)
24(1)
1+
+
3 +
4 +
5+
6+
2
dmin
2(1)
2(1)
6(4)
9(4)
3(1)
4(3)
3(1)
4(1)
5(1)
6(1)
13
15
17
4(1)
5(1)
4(1)
5(1)
6(6)
6(1)
23
35
27
5(3)
8(2)
5(3)
6(9)
7(29)
6(1)
23
35
33
5(4)
8(2)
5(3)
6(10)
5(1)
6(1)
75
51
73
6(11)
12(2)
6(3)
6(3)
6(5)
8(54)
147
133
135
6(5)
20(2)
8(4)
6(3)
6(2)
8(22)
147
101
131
6(2)
20(2)
7(3)
6(1)
7(10)
6(1)
1+
3
4
2
5+
dmin
5(6)
3(3)
3(3)
4(7)
5(16)
6(41)
3(3)
3(3)
4(7)
5(16)
6(44)
4(5)
4(1)
4(3)
4(1)
5(7)
6(19)
35
4(1)
5(1)
4(1)
5(7)
5(3)
6(13)
31
35
4(1)
5(1)
4(1)
5(7)
5(3)
6(11)
21
35
37
4(1)
5(1)
4(1)
5(9)
5(4)
6(12)
23
31
32
37
4(7)
6(3)
4(5)
4(2)
6(69)
6(33)
75
53
57
67
5(8)
9(2)
5(2)
5(5)
5(1)
6(7)
45
51
63
71
5(9)
9(2)
5(2)
5(5)
5(2)
6(6)
2(3)
2(3)
3(6)
3(6)
13
11
15
17
23
25
33
23
27
23
49
7(1)
dmin
2(2)
2(2)
2(1)
2(1)
2(1)
16
17
17
11
35
27
23
35
23
4(35)
6(453)
4(26)
5(76)
6(238)
2(1)
3(7)
4(17)
5(54)
6(205)
3(1)
4(2)
3(1)
4(8)
5(42)
6(179)
3(1)
4(3)
3(1)
4(7)
5(41)
6(177)
37
31
4(7)
5(2)
4(4)
4(3)
5(14)
6(86)
21
37
31
4(8)
5(2)
4(4)
4(4)
5(22)
6(82)
35
33
37
31
4(11)
5(2)
4(6)
4(5)
5(12)
6(82)
23
25
27
31
35
4(7)
5(2)
4(4)
4(3)
5(18)
6(78)
75
53
57
67
71
4(2)
7(2)
4(1)
4(1)
5(10)
6(38)
2(6)
2(6)
13
11
15
13
15
23
3(9)
7(16)
3+
4(1)
dmin
13
11
12
15
17
3(4)
3(1)
3(3)
4(23)
5(154)
6(873)
23
21
31
32
35
37
4(24)
5(4)
4(11)
4(13)
5(86)
6(441)
45
16
41
51
63
73
4(11)
6(2)
4(6)
4(5)
5(46)
6(284)
50
10
sim. g=[7 5 7]
sim. g=[7 5 5]
bound g=[7 5 7]
bound g=[7 5 5]
2
10
BER
10
10
10
10
0.5
1.5
Eb/No (dB)
2.5
Figure 3.11: BER performan
e
omparison between repetition and non-repetition
onstituent
odes for four state rate 1/5 PCCC, N=512.
1
10
sim. g=[7 5 7]
sim. g=[7 5 5]
bound g=[7 5 7]
bound g=[7 5 5]
2
10
BER
10
10
10
10
0.5
1.5
Eb/No (dB)
2.5
Figure 3.12: BER performan
e
omparison between repetition and non-repetition
onstituent
ode for four state rate 1/5 PCCC, N=1024.
51
3.7.1
Figure 3.13 shows the trun
ated union bound for rate 1/3 PCCCs with rate 1/2
onstituent en
oders from 2 to 128 states. As we
an see, the performan
e of PCCCs
in
reases as we in
rease the
omplexity of the
onstituent
odes. Although the trun
ated
union bounds are only a
urate for Eb =N >2dB, it is a good indi
ation that PCCCs with
in
reased
omplexity and maximum likelihood de
oding do perform better asymptoti
ally
ompared to PCCCs with lower
omplexity. Figures A.1 to A.7 in Appendix A give the
trun
ated union bounds for other
ode rates, whi
h were found to have similar properties.
0
10
10
10
10
BER
10
10
10
12
10
s=2
s=4
s=8
s=16
s=32
s=64
s=128
14
10
16
10
18
10
20
10
3
Eb/No (dB)
Figure 3.13: Trun
ated union bound for rate 1/3 PCCC with 1024 bit interleaver, with 2
state to 128 state rate 1/2
onstituent
odes.
3.7.2
From the average BER bound results given in this
hapter we
an see that in
reasing
the
omplexity of the
onstituent de
oder
an improve the performan
e. A note of
aution,
the average BER bound is based on maximum likelihood de
oding whi
h may not hold
52
53
Chapter 4
Interleaver Design
4.1 Introdu
tion
Interleaving is a way of improving the performan
e of an error
orre
tion s
heme in
a bursty
hannel. It has the job of spreading out long bursts of errors. The design of
interleavers in turbo
odes is important sin
e it has the job of redu
ing the
orrelation
between neighbouring bits. This redu
es the bit error rate (BER) at ea
h de
oder iteration. In this
hapter we will look at dierent properties and types of interleavers for the
appli
ation of turbo
odes.
Note that an interleaver is dened by its \type" and the \properties" it possesses.
An interleaver \type" des
ribes the stru
ture that is inherent to the way the interleaver
is generated. On the other hand, the \properties" are generi
parameters whi
h des
ribe
useful or undesirable features of the interleaver. For ea
h interleaver type it is possible
to have none, one or more properties whi
h
ould result in dierent BER performan
e or
implementation
omplexity.
Pun
turing was rst applied to
onvolutional
odes [53 as a way of in
reasing the
ode rate of existing rate 1/2 de
oders. However, pun
turing
an degrade the performan
e
of turbo
odes due to uneven prote
tion of the information bits
aused by the random
interleaver. This problem
an be redu
ed by using a mod-k interleaver whi
h ensures
uniform prote
tion a
ross every bit in the sequen
e. For example, let I; J = f0; 1; 2; :::; N
1g. A normal interleaver maps I to J, i.e., : I ! J . A mod-k interleaver ensures that
54
S distan e Property
An interleaver with the spread or S distan
e property [4 will, after interleaving, separate all neighbouring elements at least S interleaver index distan
e apart, i.e., S =min(j(i)
(j )j; j (i) (j )j), for all i; j 2 I , ji j j = 1. The performan
e of turbo
odes improves as the S distan
e in
reases. However, it is more di
ult to sear
h for interleavers
with large S sin
e the proportion of interleavers de
reases with in
reasing
onstraint S .
Usually, a randomly sele
ted interleaver with good performan
e
an be generated with
p
S < b N=2
, where N is the size of the interleaver [4. These interleavers are
alled
S-random. It is usual to have some algorithm whi
h
an generate S-random interleavers
[4.
The input/output position plot of a 192 bit S-random interleaver with S = 1 and
S = 8 are given in Figures 4.1 and 4.2. As we
an see from Figure 4.2, the elements
1
55
are more evenly spread
ompared to a random interleaver with S = 1. For PCCCs,
where
onvolutional en
oders are used as
onstituent
odes, shorter error bursts o
ur
more frequently than longer error bursts. Hen
e, an S-random interleaver should be
able to breakup these shorter sequen
es more ee
tively, in
reasing the free distan
e or
de
reasing the number of nearest neighbours. This results in better performan
e.
200
180
160
Output position
140
120
100
80
60
40
20
20
40
60
80
100
120
Input position
140
160
180
200
200
180
160
Output position
140
120
100
80
60
40
20
20
40
60
80
100
120
Input position
140
160
180
200
Mod-k Property
Pun
turing
an be used to in
rease the
ode rate of an existing
ode while maintaining
the same en
oder and de
oder stru
ture. The obje
tive of a pun
turing pattern [41 is to
show whi
h of the output bits in ea
h of the output streams are to be pun
tured before
transmission. A zero represents a bit that is going to be deleted from the output stream,
while a one represents a bit that is going to be transmitted.
It was found that most pun
tured turbo
odes will
ause uneven error prote
tion on
the information bits after interleaving [5. Uneven error prote
tion is
aused by the fa
t
that, after interleaving, ea
h systemati
bit of a turbo
odeword may not have an equal
number of parity bits asso
iated with it. With uneven pun
turing, there may be long
57
10
S=1
S=6
10
BER
10
10
10
10
0.5
1.5
2
2.5
Eb/No (dB)
3.5
4.5
Figure 4.3: PCCC performan
e with random interleaver and in
reasing S
onstraint, 192
bit interleaver.
1
10
S=1
S=5
S=10
S=13
2
BER
10
10
10
10
0.5
1.5
Eb/No (dB)
Figure 4.4: PCCC performan
e with random interleaver and in
reasing S
onstraint, 512
bit interleaver.
58
10
S=1
S=10
S=20
2
10
BER
10
10
10
10
0.2
0.4
0.6
0.8
1
Eb/No (dB)
1.2
1.4
1.6
1.8
Figure 4.5: PCCC performan
e with random interleaver and in
reasing S
onstraint, 2048
bit interleaver.
interleaver size=4096
10
S=1
S=10
S=25
2
10
BER
10
10
10
10
10
0.2
0.4
0.6
0.8
1
Eb/No (dB)
1.2
1.4
1.6
1.8
Figure 4.6: PCCC performan
e with random interleaver and in
reasing S
onstraint, 4096
bit interleaver.
59
10
N=192(S=1)
N=192(S=6)
1
10
BER
10
10
10
10
10
0.5
1.5
2.5
3.5
Eb/No (dB)
Figure 4.7: 4 state rate 1/3 SCCC simulation with 192 bit interleaver of dierent S values.
0
10
N=1024(S=1)
N=1024(S=10)
N=1024(S=16)
1
10
10
BER
10
10
10
10
10
0.2
0.4
0.6
0.8
1
Eb/N0 (dB)
1.2
1.4
1.6
1.8
Figure 4.8: 4 state rate 1/3 SCCC simulation with 1024 bit interleaver of dierent S
values.
60
10
N=4096(S=1)
N=4096(S=10)
N=4096(S=25)
1
10
BER
10
10
10
10
0.2
0.4
0.6
0.8
1.2
1.4
Eb/N0 (dB)
Figure 4.9: 4 state rate 1/3 SCCC simulation with 4096 bit interleaver of dierent S
values.
strings of parity bits that are fully pun
tured. This results in very weak prote
tion of
these data bits,
ausing a de
rease in performan
e for these data bits
ompared to those
data bits that have long sequen
es of parity.
To redu
e the problem of uneven error prote
tion the interleaver is designed to ensure
that after pun
turing, the parity bits are uniformly spread a
ross all the information bits.
For example, in Figure 4.10 we have a rate 1/2 pun
tured turbo
ode with pun
turing
pattern 10 and 01 for the parity bits of the rst and se
ond en
oder. The pun
turing
pattern 11 is used for the information bit stream output.
In Figure 4.10 p() are the resultant parity bits after pun
turing with dierent interleaver patterns applied , namely u, (u) and mod (u). The shaded bits represent
parity bits from even index positions while blank bits represent parity bits from odd
index positions.
As we
an see, random interleaving
an
ause uneven pun
turing of the parity bits.
That is, after en
oding some information bits will have two parity bits, while other in2
61
formation bits will have no parity bits; e.g., bit 1 has no parity bit while bit 2 has two
parity bits. However, the mod-2 interleaver, mod (u), ensures that ea
h parity bit is
pun
tured on
e and only on
e, thus ensuring uniform prote
tion a
ross all information
bits (ignoring edge ee
ts at the start and end of the blo
k).
2
1111
0000
0000
1111
0000
0000
1111
even:1111
0000 odd:
1111
0000
1111
p(u)
111
000
000
111
000
111
0
000
111
000
111
p( (u))
puncture:
0000
1111
0000
0000
1111
11111
2
3
0000
1111
1111
0000
0000
1111
000
111
0000
1111
000
111
0000
000
111
0000
1111
000
111
31111
0
2
0000
1111
000
111
0000
1111
000
111
after
puncturing
000
111
0000
000 1111
1111
000
111
0000
1111
2
30000
0
1
p( mod-2 (u)) 111
000
111
0000
1111
111
000
000
111
000
111
0
000
111
000
111
111
000
000
111
000
111
2
000
111
000
111
000
111
000
111
000
111
000
111
2
000
111
000
111
000 1111
111
0000
3
3
Figure 4.10: Amount of prote
tion on information bits between random interleaver and
mod-2 interleaver.
The
on
ept of a mod-2 interleaver was rst applied to turbo
odes in [5 as an \oddeven" interleaver, where improved performan
e was a
hieved for pun
tured rate 1/2 turbo
odes. A mod-2 interleaver ensures that the even parity from the rst
onstituent
ode
and odd parity from the se
ond
onstituent
ode are only interleaved in the even and odd
positions, respe
tively. Figure 4.11 illustrates the input/output position plot of a mod-2
interleaver. In this gure, 2 represent even positions and represent odd positions.
Note that we
an extend the mod-2 restri
tion to mod-k where k is any positive
integer. It
an be used in appli
ations where k-1 parity bits are pun
tured from ea
h
onstituent
ode. We
all this a pure mod-k interleaver where the modulus rule applies
to all elements of the interleaver. Therefore a pure mod-k interleaver has i mod k = j
mod k, where the interleaver maps i ! j .
62
200
180
160
Output position
140
120
100
80
60
40
20
20
40
60
80
100
120
Input position
140
160
180
200
Figure 4.11: Input/output position plot of a 192 bit mod-2 S-random interleaver.
As an example, we present a nine element mod-3 interleaver, mod (u), in Figure
4.12. Ea
h shade represents a dierent remainder of modulus 3. As we
an see, ea
h
element in the input sequen
e u is interleaved a
ording to their modulus 3 value. Therefore, after interleaving, the element addresses with modulus 3 values stay in the modulus
3 positions.
A mod-k interleaver
an be
onstru
ted from k length N=k interleavers. The data is
written into a k by N=k matrix with ea
h row permuted by ea
h length N=k interleaver.
The data is then read by
olumn. i.e., (i) = ki k + i mod k where j ; 0 j k 1
is a length N=k interleaver. We have that (i) mod k = i mod k proving that the
onstru
tion generates a mod-k interleaver.
3
mod
In the pure mod-k interleaver we restri
t all elements of the input sequen
e by the
mod-k rule. The only goal of the mod-k interleaver is to ensure the unpun
tured parity bit
stays with the information bit. Therefore we
an relax the restri
tion of mod-k interleavers
63
u mod 3 =
u
mod-3
111
000
1
2
000
111
000
111
000
111
000
111
7
2
000
111
000
111
1111
0000
0000
1111
0000
1111
0
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
0000
1111
6
( u )1111
0000
1111
0000
1111
0000
1111
1111
0000
1
2
0000
1111
0000
1111
000
111
000
111
000
111
000
111
000
111
000
000
111
000
111
3
4
5
6 111
7 8
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
000
111
000
0 111
1
8 111
3 111
4 5
000
111
000
111
000
000
111
000
111
000
111
000
111
000
000 111
111
000
111
111
000
000
111
000
111
0
000
111
000
111
000
111
64
to a pure mod-3 interleaver. The
ow diagram for the generation of mod-k S-random
interleavers is given in Appendix B.
4.2.3
Symmetri Property
10
S=8 Srandom
S=8 Smod2
1
10
BER
10
10
10
10
10
0.5
1.5
Eb/No (dB)
2.5
Figure 4.13: 4 state rate 1/2 turbo
ode with dierent 256 bit mod-2 interleavers, eight
iterations.
0
10
S=16 Srandom
S=16 Smod2
1
10
BER
10
10
10
10
10
0.2
0.4
0.6
0.8
1
Eb/No (dB)
1.2
1.4
1.6
1.8
Figure 4.14: 4 state rate 1/2 turbo
ode with dierent 1024 bit mod-2 interleavers, eight
iterations.
66
10
S=26 Srandom
S=26 Smod2
1
10
10
BER
10
10
10
10
10
0.2
0.4
0.6
0.8
1
Eb/No (dB)
1.2
1.4
1.6
1.8
Figure 4.15: 4 state rate 1/2 turbo
ode with dierent 4096 bit mod-2 interleavers, eight
iterations.
1
10
S=20 Srandom
S=20 Random Smod3
S=20 Pure Smod3
2
10
BER
10
10
10
10
Eb/No (dB)
Figure 4.16: 4 state rate 3/5 turbo
ode with dierent 2049 bit mod-3 interleavers, eight
iterations.
67
Interleave
Deinterleave
Symmetric Interleaver
1
Interleave
Deinterleave
Random Interleaver
Figure 4.17: Interleave and deinterleave pattern for symmetri
and random interleaver.
4.3.1
Blo k Interleaver
The blo
k interleaver is one of the frequently used interleaver types [4, 41, 58, 59.
It is implemented in many wireless
ommuni
ation systems (GSM, IS-95) to break up
burst errors that will be en
ountered by the de
oder. Its operation involves writing the
information bits into the rows of an m n matrix and reading the bits in the
olumn
dire
tion of the same matrix (m and n are the number of rows and
olumns, respe
tively).
That is (i) = m(i mod n) + i div n, i 2 I; mn = N; i div n = bi=n
. Deinterleaving
is performed simply by writing the interleaved bits in the
olumn dire
tion and reading
them in the row dire
tion. That is (j ) = n(i mod m) + j div m; j 2 J . This type
of interleaver has been widely used for most
on
atenated
ode systems where the outer
ode (usually a Reed-Solomon
ode) are blo
k interleaved before being en
oded by the
inner
ode to break up the long bursts of errors generated by the inner de
oder.
Figure 4.18 plots the input and output positions of a 192 bit (16 12) blo
k interleaver. We
all this an input/output position plot; it gives a good visual pi
ture of an
1
68
interleaver pattern.
4.3.2
The output of the
ir
ular interleaver is based on the following
ongruen
e modulo
n relation
(j ) = (aj ) mod N:
(j ) is the output position of an interleaved element, j is the input position of the element,
p
and a < b 2N
is the step size whi
h is
hosen to be relatively prime to the interleaver
size N . The step size determines the separation between two neighbouring elements after
interleaving. A detailed dis
ussion of this interleaver stru
ture is given in [4. Figure 4.19
shows the input/output position plot of a 192 bit
ir
ular shift interleaver with a = 17.
4.3.3
The blo
k heli
al interleaver [60 writes ea
h input data element in the row dire
tion
of an m n matrix, where m has to be
hosen relatively prime to n. Interleaving starts
from the upper left hand
orner element, in
rementing the index of ea
h dimension by
one with modulo m and n in the row and
olumn dire
tion, respe
tively, i.e.,
ri+1 = ri + 1 mod m
i+1 =
i + 1 mod n
i = 0; 1; ; N
where ri and
i are the row and
olumn indi
es for the ith bit. Their initial values are
r0 = 0
0 = 0:
This is similar to the \simile" interleaver proposed in [61. Figure 4.20(a) illustrates the
interleaving pro
edure of the blo
k heli
al interleaver. Another way of interleaving is to
69
200
180
160
Output position
140
120
100
80
60
40
20
20
40
60
80
100
120
Input position
140
160
180
200
180
160
Output position
140
120
100
80
60
40
20
20
40
60
80
100
120
Input position
140
160
180
200
Figure 4.19: Input/output position plot of a 192 bit
ir
ular shift interleaver, a = 17.
70
start from the lower left hand
orner, Figure 4.20(b), i.e.,
ri+1 = ri
1 mod m
i+1 = i + 1 mod n
0 = 0:
First sequence
Second sequence
Third sequence
10
11
10
11
(a)
(b)
Figure 4.20: A 3 4 blo
k heli
al interleaver with interleaving from (a) the upper left
hand
orner and (b) the lower left hand
orner.
Figure 4.21 shows the input/output position plot of a 208 bit (16 13) blo
k heli
al
interleaver.
4.3.4
Pseudo-random Interleaver
200
180
160
Output position
140
120
100
80
60
40
20
0
20
40
60
80
100
120
Input position
140
160
180
200
Figure 4.21: Input/output position plot of a 208 bit blo k heli al interleaver.
4.3.5
Symmetri Interleavers
As dis
ussed in Se
tion 4.2, interleavers with symmetri
properties only require a
single interleaver pattern to perform both interleaving and deinterleaving. Below is a list
of interleavers with symmetri
properties that
an be used for PCCCs and SCCCs with
dierent blo
k sizes.
Blo k Symmetri
we have
(i) = n(i mod n) + i div n
( (i)) = n(fn(i mod n) + i div ng mod n) + fn(i mod n) + i div ng div n
73
10
10
11
11
im ; in f (im ; in )
0,0
0,1
0,2
1,0
1,1
1,2
2,0
2,1
2,2
3,0
3,1
3,2
0,0
1,1
2,2
3,0
0,1
1,2
2,0
3,1
0,2
1,0
2,1
3,2
f (i)
0 0
1 4
2 8
3 9
4 1
5 5
6 6
7 10
8 2
9 3
10 7
11 11
74
The se
ond part of the expression, (n(i mod m) + i mod n) mod n is equivalent to i mod
n sin
e it is just a multiple of n plus i mod n. For the rst part, we
onje
ture that
(n(i mod m) + i mod n) mod m is equal to (i div n) when m = (n + 1). We have
g (i) = [n(i mod (n + 1)) + i mod n mod (n + 1)
75
(4.2)
180
160
140
Output position
120
100
80
60
40
20
20
40
60
80
100
Input position
120
140
160
180
To prove that this will produ
e a symmetri
interleaver, we use the interleaver
generating equation of the
ir
ular interleaver
f (f (i)) = a(ia mod N ) mod N = i
a(ia mod N ) = jN + i
jN + i
ia mod N =
a
jN + i
ia = kN +
a
2
) (ak + j )N + i = ia
if l = ak + j
then lN + i = ia
lN
+ 1 = a2
i
now let l = im ) mN + 1 = a2 :
nding the pair of a and N that satises the property in (4.2). For a given blo
k size of
p
N bits, we need to nd the multiple m, su
h that mN + 1 results in an integer value.
Sin
e we are working in mod N ,m = 0; 1; 2; ; N 1. In other words, we want to nd
all the m whi
h results in an integer value of a that satises
mN + 1 = a (integer):
(4.3)
For example, to nd the step sizes of a 192 bit symmetri
interleaver will require us to
p
nd (0; 1; 2; ; 190; 191) 192 + 1 su
h that the result is an integer value
a = 1 , 31 , 65 , 95 , 97 , 127 , 161 , 191.
Figure 4.24 shows the input/output position plot a 192 bit symmetri
ir
ular interleaver
with a = 31. The .m Matlab le used to nd all values of a given a blo
k size N is given
in Appendix C.
180
160
140
Output position
120
100
80
60
40
20
20
40
60
80
100
Input position
120
140
160
180
Figure 4.24: Input/ouput position plot of a 192 bit ir ular symmetri interleaver with
a = 31.
Figures 4.25, 4.26 and 4.27 show the BER performan
e of the four state rate 1/3
PCCC with
ir
ular interleavers of dierent a values.
77
10
10
BER
10
10
10
10
10
0.5
1.5
2.5
Eb/No (dB)
3.5
4.5
Figure 4.25: BER performan
e of 255 bit
ir
ular interleaver with dierent values of a.
1
10
10
BER
10
10
10
10
0.5
1.5
Eb/No (dB)
2.5
Figure 4.26: BER performan
e of 1023 bit
ir
ular interleaver with dierent values of a.
78
10
10
BER
10
10
10
10
0.5
1.5
Eb/No (dB)
2.5
Figure 4.27: BER performan
e of 4095 bit
ir
ular interleaver with dierent values of a.
Pseudo-Random Symmetri
180
160
140
Output position
120
100
80
60
40
20
20
40
60
80
100
Input position
120
140
160
180
six dierent interleavers with the same properties. Although there are no obvious improvements in performan
e between S-random and S-symmetri
interleavers; the absen
e
in degradation between the two types of interleavers still make S-symmeti
interleavers
an attra
tive alternative.
Finally, Figures 4.33, 4.34, 4.35 and 4.36 show a
omparison between dierent types
of symmetri
interleavers for a rate 1/3 four state PCCC and dierent interleaver sizes.
All results were averaged over six dierent interleavers with the same properties.
As we
an see from the gures, the S-symmetri
interleaver outperforms all the other
symmetri
interleaver types due to its improved distan
e property.
10
Random
Srandom
Ssymmetric
2
10
BER
10
10
10
10
0.2
0.4
0.6
0.8
1
Eb/No (dB)
1.2
1.4
1.6
1.8
Figure 4.29: Interleaver performan
e
omparison for four state rate 1/3 turbo
ode,
N=1024, six iterations.
1
10
Random interleaver
S=10 Srandom interleaver
S=10 Srandom symmetric interleaver
2
10
BER
10
10
10
10
0.5
1.5
Eb/No (dB)
2.5
Figure 4.30: 4 state rate 1/3 turbo
ode with dierent 400 bit interleavers and eight
iterations.
82
10
Random interleaver
S=20 Srandom interleaver
S=20 Srandom symmetric interleaver
2
10
BER
10
10
10
10
0.1
0.2
0.3
0.4
0.5
Eb/No (dB)
0.6
0.7
0.8
0.9
Figure 4.31: 4 state rate 1/3 turbo
ode with dierent 2048 bit interleavers and eight
iterations.
1
10
Random interleaver
S=25 Srandom interleaver
S=25 Srandom symmetric interleaver
2
10
BER
10
10
10
10
10
0.2
0.4
0.6
0.8
1
Eb/No (dB)
1.2
1.4
1.6
Figure 4.32: 4 state rate 1/3 turbo
ode with dierent 4096 bit interleavers and eight
iterations.
83
10
10
BER
10
10
10
10
0.5
1.5
2
Eb/No (dB)
2.5
3.5
Figure 4.33: Comparison between dierent types of symmetri
interleaver with six iterations, N =256.
2
10
BER
10
10
10
1.5
2.5
Eb/No (dB)
Figure 4.34: Comparison between dierent types of symmetri
interleaver with eight
iterations, N =512.
84
10
10
BER
10
10
10
10
10
0.5
1.5
Eb/No (dB)
2.5
Figure 4.35: Comparison between dierent types of symmetri
interleaver with eight
iterations, N =1024.
1
10
10
BER
10
10
10
10
0.5
1.5
Eb/No (dB)
2.5
Figure 4.36: Comparison between dierent types of symmetri
interleaver with eight
iterations, N =4096.
85
10
Block
Helical
Random
Srandom
Smod2
Ssym.
Ssym. mod2
10
P(e)
10
10
10
10
10
0.2
0.4
0.6
0.8
1
Eb/No (dB)
1.2
1.4
1.6
1.8
Figure 4.37: 4 state rate 1/2 turbo
ode with dierent 1024 bit interleaver designs and
six iterations.
types.
Figures 4.41, 4.42 and 4.43 show the BER performan
e of dierent mod-4 interleavers
with interleaver size 400, 2048 and 4086 bits. A four state rate 4/6 PCCC with pun
turing
pattern (0 0 0 1) and (1 0 0 0) for the rst and se
ond
onstituent
ode is simulated for 8
iterations. For interleaver size of 400 bits all interleaver types have similar performan
e.
For interleaver size of 2048 bits the pure S-mod-4 interleaver performs better than the
other three interleaver types. The S-mod-4 interleavers with symmetri
properties performs slightly better than those without symmetri
properties. For interleaver size of 4086
bits, all interleaver types give very similar performan
e, with the pure S-mod-4 interleaver
performing better at Eb=N below 2.5 dB.
Interleavers with mod-k properties do improve performan
e with pun
turing, as
an
be seem from mod-2 interleavers. However, when high amounts of pun
turing are applied
(above mod-3 pun
turing) the performan
e gain from using mod-k interleavers is very
small. Simulation results also show that in most
ases the dieren
e in performan
e
0
86
between pure mod-k interleavers and random mod-k interleavers are not very signi
ant.
Therefore, the added randomess does not seem to result in a better interleaver.
1
10
10
BER
10
10
10
10
10
4
Eb/No (dB)
Figure 4.38: 4 state rate 3/5 turbo
ode with dierent 256 bit interleavers and eight
iterations.
87
10
10
BER
10
10
10
10
10
Eb/No (dB)
Figure 4.39: 4 state rate 3/5 turbo
ode with dierent 1024 bit interleavers and eight
iterations.
1
10
10
10
10
10
10
10
Eb/No (dB)
Figure 4.40: 4 state rate 3/5 turbo
ode with dierent 4096 bit interleavers and eight
iterations.
88
10
10
BER
10
10
10
10
10
0.5
1.5
2
Eb/No (dB)
2.5
3.5
Figure 4.41: 4 state rate 4/6 turbo
ode with dierent 400 bit interleavers and eight
iterations.
0
10
10
BER
10
10
10
10
10
0.5
1.5
Eb/No (dB)
2.5
Figure 4.42: 4 state rate 4/6 turbo
ode with dierent 2048 bit interleavers and eight
iterations.
89
10
10
BER
10
10
10
10
10
0.5
1.5
Eb/No (dB)
2.5
Figure 4.43: 4 state rate 4/6 turbo
ode with dierent 4086 bit interleavers and eight
iterations.
1
10
N=192 (S=6)
N=1024 (S=16)
N=4096 (S=25)
N=8192 (S=64)
10
BER
10
10
10
10
10
0.5
1.5
2
Eb/No (dB)
2.5
3.5
Figure 4.44: 4 state rate 1/3 turbo
ode with dierent interleaver sizes and eight iterations.
90
10
N=192 Rand.
N=196 Block
N=192 Circ.
S=9 Srand.
S=9 SSym.
10
BER
10
10
10
10
1.5
2.5
3.5
Eb/No (dB)
Figure 4.45: 4 state rate 1/3 SCCC simulation with dierent types of 192 bit interleaver.
91
10
10
BER
10
10
10
10
10
0.5
1.5
2
2.5
Eb/No (dB)
3.5
4.5
Figure 4.46: 4 state rate 1/3 SCCC simulation with dierent types of 256 bit interleaver.
1
10
N=512 Rand.
N=529 Block
N=512 Circ.
S=13 Srand.
S=13 SSym.
10
BER
10
10
10
10
1.5
2.5
3.5
Eb/No (dB)
Figure 4.47: 4 state rate 1/3 SCCC simulation with dierent types of 512 bit interleaver.
92
10
10
BER
10
10
10
10
10
0.5
1.5
2
Eb/No (dB)
2.5
3.5
Figure 4.48: 4 state rate 1/3 SCCC simulation with dierent types of 1024 bit interleaver.
0
10
10
BER
10
10
10
10
0.5
1.5
2.5
3.5
Eb/No (dB)
Figure 4.49: 4 state rate 1/3 SCCC simulation with dierent types of 4096 bit interleaver.
93
94
Chapter 5
Varian
e estimation for PCCCs and
SCCCs
5.1 Introdu
tion
In analysing wireless
hannels we usually assume that we know exa
tly the varian
e
of the
hannel. However, in real appli
ations this is usually not the
ase. The varian
e
needs to be derived from the data that was re
eived and hen
e is only an approximation.
There remains a question on how will a non-a
urate varian
e ae
t the performan
e of
an iterative de
oder, for both PCCCs and SCCCs.
In [63, 64 a study of varian
e variation on the performan
e of PCCCs was performed.
It was found that the penalty due to the performan
e for PCCCs was small if the varian
e
estimated was below the a
tual value of the varian
e, i.e. estimate true . However, the
ee
t on the performan
e of SCCCs has not been previously studied. In this
hapter we
will rst look at the ee
t of dierent varian
e osets on the performan
e of PCCCs and
SCCCs and see if we
an apply the varian
e estimation methods previously studied for
PCCCs.
2
5.2.1
2 = (2R
Eb
No
) :
1
5.2.2
It was found in [63 and [64 that the varian
e mismat
h fa
tor does not signi
antly
ae
t the performan
e of PCCCs if the estimated varian
e is less than the true noise
varian
e. This is similar to that of the log-MAP de
oder in the last se
tion sin
e it is an
96
10
0 db
1 db
2 db
3 db
4 db
BER
10
10
10
0
Eb/No offset
Figure 5.1: BER performan
e of four state rate 1/2 log-MAP de
oder versus de
oder
Eb =N oset, N=1024.
0
10
0 db
1 db
2 db
3 db
4 db
BER
10
10
10
0
Eb/No offset
Figure 5.2: BER performan
e of four state rate 2/3 log-MAP de
oder versus de
oder
Eb =N oset, N=1024.
0
97
integral part of the PCCC de
oder. We
an explain this by looking at the stru
ture of
a PCCC de
oder. Varian
e estimation is used at both
onstituent de
oders, where the
hannel information is being fed into the de
oder. This
an help distribute the errors
between the two de
oders. Hen
e, the performan
e of a PCCC de
oder with an S=16
S-random interleaver is similar to a MAP de
oder with varian
e mismat
h, as we
an see
in Figure 5.3.
0
10
10
BER
10
10
10
10
10
Eb/No=0 dB
Eb/No=0.5 dB
Eb/No=1 dB
Eb/No=1.5 dB
4
0
Eb/No offset (dB)
Figure 5.3: BER performan
e of four state rate 1/3 PCCC versus de
oder Eb =N oset,
N=1024, six iterations.
0
5.2.3
Figures 5.4, 5.5, 5.6 and 5.7 show the performan
e of a four state rate 1/3 SCCC
versus varian
e mismat
h for six iterations with log-MAP
onstituent de
oder and Srandom interleaver sizes of 256 (S=8), 1024(S=16), 2048(S=20) and 4096(S=26) bits,
respe
tively. As we
an see from the results, the same
annot be said about SCCCs,
where both under estimates and over estimates do not perform very well. This might be
98
due to the fa
t that the pro
ess of varian
e estimation is only used at the inner de
oding
stage. Hen
e, varian
e mismat
h errors only appear at the rst de
oding stage,
ompared
to PCCCs where the errors are more evenly spread a
ross both de
oders. Also note that
the de
oder a
hieves better performan
e if the Eb =N is somewhere between -0.5 to -1 dB
away from the true Eb =N . We
annot nd a satisfa
tory explanation for this ee
t whi
h
ould be a fo
us for future investigation.
0
10
Eb/No=0 dB
Eb/No=1 dB
Eb/No=2 dB
Eb/No=2.5 dB
BER
10
10
10
10
0
Eb/No offset (dB)
Figure 5.4: BER performan
e of four state rate 1/3 SCCC versus de
oder Eb =N oset,
N=256, eight iterations.
0
99
10
BER
10
10
10
Eb/No=0 dB
Eb/No=0.5 dB
Eb/No=1 dB
Eb/No=1.5 dB
10
0
Eb/No offset (dB)
Figure 5.5: BER performan
e of four state rate 1/3 SCCC versus de
oder Eb =N oset,
N=1024, eight iterations.
0
10
10
BER
10
10
0 dB
0.5 dB
1 dB
1.2 dB
1.4 dB
10
10
0
Eb/No offset (dB)
Figure 5.6: BER performan
e of four state rate 1/3 SCCC versus de
oder Eb =N oset,
N=2048, six iterations.
0
100
10
10
BER
10
10
10
Eb/No=0 dB
Eb/No=0.5 dB
Eb/No=1 dB
10
0
Eb/No offset (dB)
Figure 5.7: BER performan
e of four state rate 1/3 SCCC versus de
oder Eb =N oset,
N=4096, eight iterations.
0
[64, 9 proposed a way of estimating the varian
e dire
tly from the magnitude of the
re
eived symbol rn by nding the ratio between E (rn) and E (jrnj) , whi
h is a fun
tion
of Es=N . We have
2
E (rn2 ) = Es + 2
and
E (jrnj) =
e
2
(Es =2 2 )
+ Es erf
101
q
Es
2 2
i
1 + E2
q
h
q
i
+ E2 erf E2
s
q
(Es =2 2 )
e
Es
= f 2 :
2
As we
an see, it is not easy to nd the varian
e from the ratio by using this equation.
A pra
ti
al approa
h is to nd the ratio and its related
hannel varian
e for the
operational range of the de
oder and express it as a lookup table. By using this lookup
table we
an nd the varian
e of the
hannel by obtaining the ratio from the re
eived
data. Another way, as suggested in [64, is to nd an nth order polynomial that
an
approximate the ratio E (rn)/E (jrnj) for a xed
hannel varian
e range that is related to
the de
oder's Eb =N operating range.
2
1.55
true
estimated
f(x)
1.5
1.45
1.4
1.35
0.4
0.6
0.8
1.2
1.4
1.6
Channel Variance
1.8
2.2
2.4
where the
oe
ients of the third order polynomial were obtained from the polyt fun
tion in Matlab [65. First the fun
tion f (x) is analysed for the
hannel varian
e range
that is equivalent to an Eb =N range from -2dB to 5 dB. The Matlab fun
tion polyt is
then used to nd the polynomial that ts the fun
tion f (x) in the least squares sense.
The
oe
ients of the polynomial generated by Matlab is then used in the estimator to
estimate the
hannel varian
e given the re
eived E (rn)/E (jrnj) ratio. The Matlab le
used to generate the
oe
ients is given in Appendix D.
An advantage of this method is that it
an provide an estimate of the varian
e before
the de
oding of ea
h blo
k. This allows the de
oder to operate at a varian
e that is quite
lose to the true varian
e. The estimated varian
e at ea
h blo
k is averaged with the last
estimate using
~j = ^j + (1 )~j
(5.1)
0
where ^j is the estimated varian
e of the
urrent blo
k j , ~j is the varian
e of the last
blo
k , and
an be varied to adjust the averaging response, whi
h is 0.99 for this
ase.
2
5.3.2
In [63 an estimator that uses the turbo de
oder hard de
ision output to estimate
the noise varian
e was proposed. It makes use of the fa
t that, given enough Eb =N ,
the hard-de
ision output of the turbo de
oder will be very
lose to the symbol a
tually
transmitted. Hen
e, we
an make use of this fa
t to a
urately estimate the varian
e of
the
hannel by using the equation
0
N
X1
^j2+1 =
(rk;j d^k;j )
k=0
(5.2)
where rk;j is the re
eived symbol at time k in blo
k j , d^k;j is the hard de
ision output of
the turbo de
oder after an integer number of iterations and ak is the fading amplitude of
103
a fully interleaved, frequen
y non-sele
tive, slowly fading Rayleigh
hannel. Note that the
estimated varian
e is for the next blo
k of the turbo de
oder. The varian
e is averaged
at ea
h de
oded blo
k using
~j2+1 = ^j2+1 + (1 )~j2 :
(5.3)
A disadvantage of this estimator is that it has a one blo
k delay. That is, the estimator
has to wait for the hard de
ision of the rst blo
k before the estimated varian
e
an be
used for the next blo
k. However, this
an be over
ome by using a simple estimator to
nd a
lose estimate of varian
e before the rst iteration starts.
5.3.3
An assumption of the Reed estimator is that it assumes the de
oder knows exa
tly
the magnitude of the transmitted amplitude. This is not true in a real
ommuni
ation
systems where the amplitude of the signal is dependent on the signal to noise ratio due to
the auto gain
ontrol of the re
eiver amplier. At high SNR the amplitude of the signal
is
lose to the gain level of the auto gain
ontroller. At low SNR, the
ombination of
noise and signal results in the signal amplitude being redu
ed by the auto gain
ontroller.
Without knowing the SNR, this amplitude
annot be determined.
Here, we propose a new estimator whi
h is similar to the Reed estimator, but we estimate the amplitude of the re
eived symbol from the hard de
isions of the turbo de
oder.
The varian
e,
an be estimated by
2
^j2+1 = E [(Xj
^j )2
= E [Xj ^j
2
where E[ is the sample average and ^j is the mean amplitude of the estimated re
eived
data blo
k j
1 NX ^
^j =
dk;j rk;j
N
1
k=0
104
rk;j is the re
eived symbol at time k in blo
k j and d^k;j is the hard de
ision output of the
turbo de
oder after an integer number of iterations. E [Xj2 is estimated by
N
X
1
E [Xj '
(rk;j ) :
1
N k=0
PCCC
Figures 5.9, 5.10 and 5.11 present a performan
e
omparison between dierent varian
e estimators for a four state rate 1/3 PCCC with S-random interleaver sizes of 384
(S=11), 4096 (S=25) and 8192 (S=35) bits, respe
tively. Six PCCC de
oder iterations
were performed before a hard de
ision was made.
As we
an see, the Summer, Reed and New Reed methods perform very well for
all interleaver sizes. There is no dieren
e in BER performan
e for all three estimator
s
hemes. Note that the Reed estimator
annot be implemented in pra
tise sin
e the a
tual
signal amplitude is not known.
5.4.2
SCCC
Figures 5.12, 5.13 and 5.14 present performan
e
omparisons between dierent varian
e estimators for a four state rate 1/3 SCCC and six iterations. Three dierent pseudorandom interleaver sizes were simulated: 384, 2048 and 8192 bits. For SCCCs the
performan
e of the estimators are quite similar to the PCCC. For the three blo
k sizes
simulated all three varian
e estimators perform quite well. However, the Reed estimator
does perform a fra
tion of a dB worse than the other two estimator at low Eb/No.
It also appears from the simulations that the performan
e of the Summer estimator
degrades for low Eb =N . This
ould be due to the fa
t that the outer de
oder does not
0
105
10
Exact
Summer
Reed
New Reed
2
10
BER
10
10
10
10
0.2
0.4
0.6
0.8
1
Eb/No (dB)
1.2
1.4
1.6
1.8
Figure 5.9: Performan
e
omparison between dierent varian
e estimators, four state,
rate 1/3 PCCC, N=384 and six iterations.
1
10
Exact
Summer
Reed
New Reed
10
BER
10
10
10
10
10
0.2
0.4
0.6
0.8
1.2
1.4
Eb/No (dB)
Figure 5.10: Performan
e
omparison between dierent varian
e estimators, four state,
rate 1/3 PCCC, N=4092 and six iterations.
106
10
Exact
Summer
Reed
New Reed
2
10
BER
10
10
10
10
0.1
0.2
0.3
0.4
0.5
Eb/No (dB)
0.6
0.7
0.8
0.9
Figure 5.11: Performan
e
omparison between dierent varian
e estimators, four state,
rate 1/3 PCCC, N=8192 and six iterations.
dire
tly use the re
eived data, depending entirely on the output of the inner de
oder.
This varian
e estimation error from the inner de
oder will thus be amplied by the outer
de
oder.
10
Perfect estimation
Summer
Reed
New Reed
1
10
BER
10
10
10
10
0.5
1.5
2.5
Eb/No (dB)
Figure 5.12: Performan
e
omparison between dierent varian
e estimators, four state
rate 1/3 SCCC, N=384 and six iterations.
0
10
Perfect estimation
Summer
Reed
New Reed
1
10
BER
10
10
10
10
0.5
1.5
Eb/No (dB)
Figure 5.13: Performan
e
omparison between dierent varian
e estimators, four state
rate 1/3 SCCC, N=2048 and six iterations.
108
10
Perfect estimation
Summer
Reed
New Reed
10
BER
10
10
10
10
10
0.1
0.2
0.3
0.4
0.5
Eb/No (dB)
0.6
0.7
0.8
0.9
Figure 5.14: Performan
e
omparison between dierent varian
e estimators, four state
rate 1/3 SCCC, N=8092 and six iterations.
Estimator =1.191612 =1.062025 =0.946531
Summer 1.536258e-02 5.144830e-03 1.350778e-03
Reed
8.565042e-03 1.387372e-03 2.243528e-03
New Reed 1.214932e-02 4.511558e-03 3.999086e-03
Table 5.1: Mean Squared Error for dierent estimators with blo
k size of 384.
2
We found that SCCCs are more sensitive to varian
e mismat
h
ompared to PCCCs. Two
dierent varian
e estimators that were applied to PCCCs were reviewed and applied to
SCCCs to test its performan
e. It was found that the Reed estimator made an unrealisti
assumption that the signal amplitude is known. We
orre
ted this in the New Reed
estimator. We found that the Summer estimator, Reed estimator and the New Reed
estimator perform very well for both large and small blo
k sizes.
109
Chapter 6
Con
lusion
6.1 Summary of results
In Chapter 3
riteria for good CCs for PCCCs were investigated. Previous
ode
sear
hes were based on nding CCs that maximised ee
tive free distan
e. Ee
tive free
distan
e is the minimum free distan
e that is generated by an input weight two
odeword.
Based on the uniform bound we found that in addition to maximising the ee
tive free
distan
e, we should also minimise the number of nearest neighbors. The number of nearest
neighbours is determined by nding the number of times input weight two error paths
remerge with the zero state. We also found that for some CCs whi
h have high
ode
rate, there are very likely multiple generator polynomials having identi
al ee
tive free
distan
e and number of nearest neighbours. Additional
riteria, namely dmin ; Nmin; d
and N were used to nd the best CC. We found some
odes that are identi
al to those
published in [43 as well as some
odes that should perform better sin
e we maximised
both ee
tive free distan
e and minimum distan
e.
3
We also looked at three dierent ways of nding the distan
e spe
trum of CCs. The
rst method uses the modied state diagram and Mason's formula to nd the generating
fun
tion of the
onvolutional
ode. The se
ond method uses the state transition matrix
to solve the state equation going into the terminal state. This results in a generating
fun
tion of the
onvolutional
ode. The third method uses the unidire
tional algorithm
whi
h is similar to the Viterbi algorithm. This last method is easier to program sin
e it
110
involves systemati
extension of paths from ea
h state of the trellis se
tion to nd error
paths that remerges with the zero state.
Using the unidire
tional algorithm the best CCs were found for rate 1/2, 1/3, 1/4,
1/5, 2/3, 3/4, 4/5, 5/6 and pun
tured 1/2. From the trun
ated uniform bound of these
odes, we found that the BER performan
e with maximum likelihood de
oding is improved
by in
reasing the memory of the CC.
In Chapter 4 we investigated dierent types of interleavers that
an be used for
PCCCs and SCCCs. First, we looked at the generation method of two types of blo
k
interleaver: row-
olumn blo
k interleaver and blo
k heli
al interleaver. These interleavers
are based on writing ea
h element row by row and reading ea
h element a
ording to a
dierent pattern. We then looked at
ir
ular interleavers, whi
h are based on the
ir
ular
shift of ea
h elements with step size a. This type of interleaver is more random
ompared
to blo
k interleavers and relatively easy to generate in real time. We also looked at
pseudo-random interleavers whi
h are generated randomly. Although it
an give better
performan
e, they are quite hard to generate in real time without the use of large lookup
tables.
We also looked at various interleaver properties that
an be used to improve the
performan
e of PCCCs. First, we looked at the S distan
e property. This ensures an
S distan
e separation between the
urrent elements and S neighbouring elements. This
property, if
ombined with pseudo-random interleaving, performs quite well with both
PCCCs and SCCCs. The next property, the mod-k property,
an be used to redu
e the
problem of uneven error prote
tion
aused by parity pun
turing of PCCCs. The property
ensures that every information bit, after pun
turing, will be uniformly prote
ted. This
property, if
ombined with pseudo-random interleaving,
an outperform a pseudo-random
interleaver for pun
tured PCCCs.
The last property we investigated is the symmetri
property. An interleaver with this
property will have identi
al interleave and deinterleave patterns. This property has the
advantage of a simpler implementation sin
e only one interleave sequen
e is required. A
111
quadrati
interleaver is one type of interleaver that
an have this symmetri
property. It
is also possible to generate symmetri
interleavers using the
ir
ular, blo
k, blo
k heli
al
and pseudo-random interleavers.
We found that if we
ombined the S-distan
e, mod-k and symmetri
properties for
PCCCs, we
an a
hieve better performan
e
ompared to interleavers with only one property. For SCCCs, the symmetri
property was found to redu
e performan
e for some
interleaver sizes.
In Chapter 5 we looked at varian
e estimation methods for SCCCs. It was found
in [63, 64 that the BER performan
e degradation of PCCCs is small if the estimated
varian
e is below the a
tual varian
e value. We studied the ee
t on performan
e for
SCCCs if there are mismat
hes in varian
e estimation. We found that SCCCs are more
sensitive to varian
e errors. If the estimated Eb =N is 0.5 to 1 dB above the true Eb =N
the performan
e starts to degrade. However, if the estimated Eb =N is 0.5 to 1 dB below
the true Eb=N , in some
ases we a
hieve better performan
e. This is a strange ee
t that
we
ould not explain. For estimated Eb=N greater than 1 dB below the true Eb =N , the
performan
e qui
kly degrades.
Two dierent varian
e estimation s
hemes that were proposed were investigated for
PCCCs and SCCCs. We had to modify one of the s
hemes to take into a
ount that the
re
eived signal amplitude is usually not known. We found that both s
hemes perform
quite well for SCCCs, in some
ases even better than the true varian
e
ase. This might
due to the fa
t that under estimation works better for SCCCs.
0
Re
ently, there is a new type of interleaver based on the Golden ratio whi
h
an lower
the
are of the PCCCs for bit error rate and frame error rate. It would be interesting
to see how this interleaver stru
ture
an be integrated with the mod-k and symmetri
properties so that it
an further improve performan
e.
The ee
t of a
hieving better SCCC performan
e if the varian
e is slightly over
estimated requires further study. It is not
lear why this would happen to SCCCs sin
e
it does not appear for PCCCs. If we
an explain this behaviour we might be able to nd
a better SCCC de
oder.
113
Bibliography
[1 C. Berrou, A. Glavieux, and P. Thitiumjshima, \Near Shannon limit error
orre
ting
oding: Turbo
odes," in Intl. Conf. on Commun., (Geneva, Switzerland), pp. 1064{
1070, May 1993.
[2 S. Benedetto, R. Garello, and G. Montorsi, \A sear
h for good
onvolutional
odes
to be used in the
onstru
tion of turbo
odes," IEEE Trans. on Commun., vol. 46,
pp. 1101{1105, Sept. 1998.
[3 M. S. C. Ho, S. S. Pietrobon, and T. Giles, \Improving the
onstituent
odes of
turbo en
oders," in Pro
eeding 1998 IEEE Global Tele
ommuni
ations Conferen
e,
(Sydney, Australia), pp. 3525{3529, November 1998.
[4 S. Dolinar and D. Divsalar, \Weight distribution for turbo
odes using random and
nonrandom permutations," JPL TDA Progress Report, pp. 56{65, Aug. 1995.
[5 S. A. Barbules
u and S. S. Pietrobon, \Interleaver design for turbo
odes," Ele
troni
Letters, vol. 30, pp. 2107{2108, De
. 1994.
[6 S. Crozier, J. Lodge, P. Guinand, and A. Hunt, \Performan
e of turbo-
odes with
relative prime and golden interleaving strategies," in Pro
eeding Sixth International
Mobile Satellite Conferen
e, (Ottawa), pp. 268{275, June 1999.
[7 M. S. C. Ho, S. S. Pietrobon, and T. Giles, \Interleavers for pun
tured turbo
odes,"
in Pro
eedings APCC/ICCS '98, vol. 2, (Singapore), pp. 520{524, Nov. 1998.
[8 S. S. Pietrobon and A. S. Barbules
u, \A simpli
ation of the modied Bahl de
oding
algorithm for systemati
onvolutional
odes," in Int. Symp. on Inform. Theory and
its Appli
ations, (Sydney, Australia), pp. 1073{1077, Nov. 1994.
[9 S. S. Pietrobon, \Implementation and performan
e of a turbo/MAP de
oder," Int.
J. of Satellite Commun., vol. 16, pp. 23{46, Jan./Feb. 1998.
[10 P. Robertson, E. Villebrun, and P. Hoeher, \A
omparison of optimal MAP de
oding algorithms operating in the log domain," in Pro
. IEEE Int. Conf. Commun.,
(Seattle, USA), pp. 1009{1013, June 1995.
114
[11 L. Bahl, J. Co
ke, F. Jelinek, and J. Raviv, \Optimal de
oding of linear
odes for
minimizing symbol error rate," IEEE Trans. Inform. Theory, vol. IT-20, pp. 284{287,
1974.
[12 D. G. Forney, Jr., Con
atenated
odes. Cambridge, MIT Press, 1966.
[13 S. Benedetto, D. Divsalar, G. Montorsi, and F. Pollara, \Serial
on
atenation of interleaved
odes: Performan
e analysis, design, and iterative de
oding," IEEE Trans.
Inform. Theory, vol. 44, pp. 909{926, May 1998.
[14 R. W. Chang and J. C. Han
o
k, \On re
eiver stru
tures for
hannels having memory," IEEE Trans. Inform. Theory, vol. IT-12, pp. 463{468, O
t. 1966.
[15 A. J. Viterbi, \An intuitive justi
ation and a simplied implementation of the MAP
de
oder for
onvolutional
odes," IEEE J. Sele
ted Areas Commun., vol. 16, Feb.
1998.
[16 J. Hagenauer and P. Hoeher, \A Viterbi algorithm with soft-de
ision outputs and
its appli
ations," in Pro
. IEEE GLOBECOM, (Dallas, USA), pp. 1680{1686, Nov.
1989.
[17 J. Hagenauer, P. Robertson, and L. Papke, \Iterative (\TURBO") de
oding of systemati
onvolutional
odes with the MAP and SOVA algorithms," in Pro
. of the
ITG Conf. \Sour
e and Channel Coding", (Muni
h, Germany), pp. 21{29, 1994.
[18 Y. Li, B. Vu
eti
, Y. Sato, and Y. Furuya, \A soft-output Viterbi algorithm," in
Mobile and Personal Commun. Systems Conferen
e, (Adelaide, Australia), pp. 223{
231, Nov. 1992.
[19 C. Berrou, P. Adde, E. Angui, and S. Faudeil, \A low
omplexity soft-output viterbi
de
oder ar
hite
ture," in IEEE Int. Conf. on Communi
ations, (Geneva), pp. 737{
740, May 1993.
[20 L. Bahl, J. Co
ke, F. Jelinek, and J. Raviv, \Optimal de
oding of linear
odes for
minimizing symbol error rate," in Pro
. IEEE Int. Symp. on Information Theory,
(Asilmoar, CA), p. 90, May 1972.
[21 P. L. M
Adam, L. R. Wel
h, and C. L. Weber, \M.A.P. bit de
oding of
onvolutional
odes," in Pro
. IEEE Int. Symp. on Information Theory, (Asilomar, CA), p. 91, May
1972.
[22 S. A. Barbules
u, R. Chang, and S. Yaghmour, \Turbo
ode
-QPSK modem for
Intelsat digital servi
es," in Pro
eedings of 2nd International Symposium on Turbo
Codes and Related Topi
s, (Brest, Fran
e), pp. 487{490, 4-7 September 2000.
115
[36 P. K. Gray, Serially
on
atenated trellis
oded modulation. PhD thesis, Univ. of South
Australia, Mar
h 1999.
[37 J. Sun, W. Qi, W. H. Mow, and K. H. Li, \Turbo
odes with short memory
omponent
odes," in Pro
eedings APCC/ICCS '98., vol. 2, (Singapore), pp. 530{534, 1998.
[38 J. Y. Couleaud, \High gain
oding s
hemes for spa
e
ommuni
ations," ENSICA
Final Year Proje
t Report, University of South Australia, Sept. 1995.
[39 P. Elias, \Coding for noisy
hannel," IRE Conv. Re
., no. 4, pp. 37{47, 1955.
[40 S. Lin and D. J. Costello, Jr., Error Control Coding: Fundamental and Appli
ations.
Upper Saddle River, New Jersey 07458: Prenti
e-Hall, 1983.
[41 S. B. Wi
ker, Error Control Systems for Digital Communi
ation and Storage. Upper
Saddle River, New Jersey 07458: Prenti
e Hall, 1995.
[42 J. L. Massey, Threshold De
oding. Cambridge, Mass.: M.I.T. Press, 1963.
[43 S. Benedetto and G. Montorsi, \Unveiling turbo
odes: Some results on parallel
on
atenated
oding s
hemes," IEEE Trans. Inform. Theory, vol. 42, pp. 409{429,
Mar
h 1996.
[44 L. C. Perez, J. Seghers, and D. J. Costello, Jr., \A distan
e spe
trum interpretation
of turbo
odes," IEEE Trans. Inform. Theory, vol. 42, pp. 1698{1709, Nov. 1996.
[45 D. Divsalar and P. Pollara, \Turbo
odes for PCS appli
ations," in Pro
eedings of
IEEE ICC'95, pp. 54{59, June 1995.
[46 J. G. C. Clark and J. B. Cain, Error-Corre
tion Coding for Digital Communi
ations.
New York: Plenum Press, 1981.
[47 S. Mason and H. Zimmermann, Ele
troni
Cir
uits, Signals, and Systems. New York:
Wiley, 1960.
[48 L. R. Bahl, C. D. Cullum, W. D. Frazer, and F. Jelinek, \An e
ient algorithm
for
omputing free distan
e," IEEE Trans. Inform. Theory, vol. IT-18, pp. 437{439,
May 1972.
[49 W. W. Peterson and E. J. WeldonJr, Error-Corre
ting Codes. Cambridge: The MIT
Press, 1961.
[50 D. Divsalar and F. Pollara, \Hybrid
on
atenated
odes and iterative de
oding,"
JPL TDA Progress Report, pp. 1{23, Aug. 1997.
[51 S. S. Pietrobon, R. H. Deng, A. Lafane
hhre, G. Ungerboe
k, and D. J. Costello, Jr.,
\Trellis-
oded multidimensional phase modulation," IEEE Trans. Inform. Theory,
vol. 36, pp. 63{89, Jan. 1990.
117
[52 D. Divsalar and F. Pollara, \On the design of turbo
odes," JPL TDA Progess Report,
pp. 99{121, Nov. 1995.
[53 J. B. Cain, G. C. Clark, Jr., and J. M. Geist, \Pun
tured
onvolutional
odes of rate
(n-1)/n and simplied maximum likelihood de
oding," IEEE Trans. Inform. Theory,
vol. IT-25, pp. 97{100, Jan. 1979.
[54 J. Hokfelt, T. Maseng, and O. Edfors, \Assessing interleaver suitability for turbo
odes.," in To appear in Nordi
Radio Symposium 1998, (Saltsjobaden, Sweden),
O
t. 1998.
[55 Small World Communi
ations, \ Iterative De
oding of Parallel Con
atenated Convolutional Codes", Jan. 1999. Appli
ation Note Version 1.4.
[56 O. Y. Takeshita and D. J. Costello, Jr., \New
lasses of algebrai
interleavers for
turbo-
odes," in Pro
eedings of IEEE International Symposium on Information Theory, (MIT, Cambridge, MA USA), p. 419, Aug. 1998.
[57 A. S. Barbules
u and S. S. Pietrobon, \Interleaver design for three dimensional turbo
odes," IEEE International Symposium on Information Theory, p. 37, Sept. 1995.
[58 T. S. Rappaport, Wireless Communi
ations Prin
iples and Pra
ti
e. NJ, USA: Prenti
e Hall, 1996.
[59 J. G. Proakis, Digital Communi
ations. M
Graw-Hill Book Co., 3 ed., 1995.
[60 D. T. Chi, \A new blo
k heli
al interleaver," in Pro
eedings of MILCOM 1992, (San
Diego, CA), pp. 799{804, O
t. 1992.
[61 A. S. Barbules
u and S. S. Pietrobon, \Terminating the trellis of turbo
odes in the
same state," Ele
troni
s Letters, vol. 31, pp. 22{23, Jan. 1995.
[62 S. Dolinar, D. .Divsalar, and F. Pollara, \Code performan
e as a fun
tion of blo
k
size," TMO Progess Report, vol. 42, May 1998.
[63 M. Reed and J. Asenstorfer, \A novel varian
e estimator for turbo-
ode de
oding,"
in Pro
. ICT '97, (Melbourne, Australia), pp. 173{178, April 1997.
[64 T. A. Summers and S. G. Wilson, \SNR mismat
h and online estimation in turbo
de
oding," IEEE Transa
tions on Communi
ations, vol. 46, pp. 421{423, April 1998.
[65 The Mathworks, In
., The Student Edition of Matlab. New Jersey, USA: Prenti
e-Hall
International Editions, 1992.
118
Appendix A
Performan
e bounds of PCCC
This appendix
ontains the performan
e bounds of PCCC with rate 1/2, 1/3, 2/3, pun
tured 1/2, 3/4, 1/4, 1/5 and 4/5 CC.
0
10
s=2
s=4
s=8
s=16
s=32
s=64
10
10
10
BER
10
10
10
12
10
14
10
16
10
18
10
20
10
5
Eb/No (dB)
10
Figure A.1: Trun
ated union bound for rate 1/5 PCCC with 1024 bits interleaver, with
2 state to 64 state rate 1/3
onstituent
odes.
119
10
s=2
s=4
s=8
s=16
s=32
s=64
10
10
10
BER
10
10
10
12
10
14
10
16
10
18
10
20
10
5
Eb/No (dB)
10
Figure A.2: Trun
ated union bound for rate 1/2 PCCC with 1024 bits interleaver, with
2 state to 64 state rate 2/3
onstituent
odes.
0
10
s=2
s=4
s=8
s=16
s=32
s=64
s=128
10
10
10
BER
10
10
10
12
10
14
10
16
10
18
10
20
10
5
Eb/No (dB)
10
Figure A.3: Trun
ated union bound for rate 1/2 PCCC with 1024 bits interleaver, with
2 state to 128 state pun
tured rate 1/2
onstituent
odes.
120
10
s=2
s=4
s=8
s=16
s=32
10
10
10
BER
10
10
10
12
10
14
10
16
10
18
10
20
10
5
Eb/No (dB)
10
Figure A.4: Trun
ated union bound for rate 3/5 PCCC with 1024 bits interleaver, with
2 state to 32 state rate 3/4
onstituent
odes.
0
10
s=2
s=4
s=8
s=16
s=32
10
10
10
BER
10
10
10
12
10
14
10
16
10
18
10
20
10
5
Eb/No (dB)
10
Figure A.5: Trun
ated union bound for rate 1/7 PCCC with 1024 bits interleaver, with
2 state to 32 state pun
tured rate 1/4
onstituent
odes.
121
10
s=2
s=4
s=8
s=16
10
10
10
BER
10
10
10
12
10
14
10
16
10
18
10
20
10
5
Eb/No (dB)
10
Figure A.6: Trun
ated union bound for rate 1/9 PCCC with 1024 bits interleaver, with
2 state to 16 state rate 1/5
onstituent
odes.
0
10
s=4
s=8
s=16
s=32
10
10
10
BER
10
10
10
12
10
14
10
16
10
18
10
20
10
5
Eb/No (dB)
10
Figure A.7: Trun
ated union bound for rate 4/6 PCCC with 1024 bits interleaver, with
2 state to 32 state rate 4/5
onstituent
odes.
122
Appendix B
Interleaver generator
ow diagram
Figure B.1 shows the
ow diagram for the generation of mod-k S-random interleavers.Figure
B.2 shows the
ow diagram for the generation of symmetri
interleaver with distan
e S
and mod-k properties .
123
No
num_
array
size>0?
No
interl_
array
size=N?
Yes
Yes
Randomly pick a number from the
num_array. Decrease size of
num_array by 1.
END
Check the S-distance property
of the newly generated number.
Set the flag if pass.
Yes
number
rejected?
Empty reject_array.
No
Figure B.1: Flow diagram of a mod-k interleaver generator with S distan
e property.
124
No
num_array
size>0?
No
interl_array
size=N?
Yes
Yes
Get the address of the first nonempty element in interl_array.
END
Randomly pick a number from
the num_array.
Empty reject_array.
Yes
number
rejected?
No
If reject_array size>0,
put all the reject numbers
into num_array.
Figure B.2: Flow diagram of a symmetri
interleaver with S and mod-k properties.
125
Appendix C
Matlab
ode for Chapter 4
This is the Matlab
ode used to nd the a parameters for the
ir
ular symmetri
interleaver.
N=4096;
m=(1:N)-1;
a=sqrt((m*N)+1);
a integer=a(nd((a-
oor(a))==0));
126
Appendix D
Matlab
ode for Chapter 5
This is the Matlab
ode used to generate the
oe
ients for the third order polynomial
used by the Summer estimator.
N=50000;
R=1/3; %
ode rate
data=2*
eil(rand(1,N)-0.5)-1; %generate data and add noise
noise=randn(1,N);
db=[-2:1:5;
varian
e=1./(R*2*10.^(db./10));
std=sqrt(1./(R*2*10.^(db./10)));
rat=0;
%evaluate the ratio of the re
eived data
for i=1:max(size(std))
rx=data+(noise*std(i));
rat(i)=mean(rx.^2)/(mean(abs(rx)).^2);
end
%nd the
oe
ient of third order polynomial using polyt
P=polyt(rat,varian
e,3);
plot(varian
e,rat,P(1).*rat.^3+P(2).*rat.^2+P(3).*rat+P(4),rat,'s-.');
127
grid on
xlabel('Channel Varian
e');
ylabel('f(x)');
legend('true','estimated');
128