1. Iterative Equalization and Decoding
John G. Proakis
COMSOC Distinguished Lecture Tour

2. Conventional Equalization
Soft Output
Hard Output
From Receiver
Filter
Equalizer
Decoder
Possible Equalizer Types:
Linear Equalizer
Decision Feedback Equalizer (DFE)
Maximum A posteriori Probability (MAP) Equalizer
Soft-output Viterbi (MLSE) Equalizer
Possible Decoder Types:
Maximum A posteriori Probability (MAP) Decoder
Viterbi (MLSE) Decoder

3. Turbo Principle / Turbo Coding+
cs(n)
cp1(n) P
cp2(n)
d`(n)
Puncturer and P/S converter
Encoder 1
Encoder 2
d(n)
Turbo Encoder:
Parallel concatenated recursive systematic convolutional encoders
Encoders separated by an interleaver
b(n)
Turbo Decoder:
Two Soft-Input Soft-Output (SISO) decoders separated by interleavers
SISO modules can be
SOVA
MAP
Extrinsic information passed between modules
Le21
P-1
SISO
Decoder 1
Le12 P
xp1
SISO
Decoder 2 xs
xp2 P Ld

4. Serially Concatenated Systems
Serially Concatenated Coding:
Serial concatenated (recursive) convolutional encoders
Encoders separated by an interleaver
Encoder 1
d(n) P
c(n)
c’(n)
b(n)
Encoder 2
Coded Transmission over Multipath Channels:
(recursive) convolutional encoder
Interleaved bits mapped to symbols
Symbols passed through a multipath channel
c(n)
c’(n)
d(n)
Encoder
Symbol
Mapper P
x(n)
Multipath
Channel
r(n)

5. Channel Model / Precoding
Multipath Channel Model:
Received signal:
Rate 1/1 convolutional code D X
h1(n)
h2(n)
hL-1(n)
h0(n) +
w(n)
x(n)
r(n)
Precoded System:
Iteration gain only possible with recursive inner code (channel) [1], [2]
Recursive rate 1/1 precoder is employed before transmission
Most common precoder: Differential encoder D X
h1(n)
h2(n)
hL-1(n)
h0(n) +
w(n)
x(n)
r(n) 
y(n)
y(n-1)

6. Iterative Equalization and Decoding (Turbo Equalizer)
Data bits are convolutionally encoded and interleaved
M-ary PSK modulated signals transmitted through a multipath channel, which is treated as an encoder
Received signals are jointly equalized and decoded in a turbo structure
First proposed by Douillard, et.al [3], where SOVA modules are employed
Bauch extended the idea by employing MAP modules [4]
Convolutional
Encoder P
Symbol
Mapper
Multipath
Channel dn cn
c’n xn rn
MAP
Decoder + P
P-1
Equalizer
Channel
Estimator r
LDe(c’)
LDe(c)
LD(c)
LD(d)
LEe(c)
LEe(c’)
LE(c’) - -

7. Time-invariant Test Channel
Proakis C channel [5]
0.688
0.460
0.227 t
Impulse Response
Frequency Response

8. Low Complexity Alternative Equalizers: DFE and MLSE
Performance of DFE and MLSE over the Proakis C channel [5]
Bit error rate performance [5]

9. Iterative Equalization and Decoding Performance
Iterative equalization and decoding with MAP modules
Recursive systematic convolutional encoder with R=1/2, K=5,
Time-invariant 5 tap channel with a spectral null (Proakis C [5])
Equalizer has perfect knowledge of the channel
Block length 4096
Bit error rate performance of turbo equalizer [4]

10. Hard Iterative DFE and MAP Decoding
Output
Data
Input from
receiver filter
Forward
Filter
MAP
Decoder +
P-1
Symbol
Detector
DFE with
hard input
feedback
Feedback
Filter
Hard encoded
symbols P
During the first pass, symbol detector output is passed to the feedback filter
After the first pass, hard encoded symbol output of the decoder is used in the feedback filter

11. Performance of Hard Iterative DFE and MAP decoder
BPSK modulation
R=1/2, K=7 convolutional coding
Block length 2048
Channel Proakis C

12. Soft Iterative DFE and MAP Decoding
Output
Data
Forward
Filter
MAP
Decoder +
P-1
DFE with
hard input
feedback
Feedback
Filter
Decision
Device
Soft encoded
symbols P
Soft decisions of the decoder is combined with the soft outputs of the DFE:
Hard detected
symbols +
Soft APP
from last
iteration

13. Histogram of DFE Output
The histogram of equalizer estimated output for SNR = 12 dB
The histogram of equalizer estimated output for SNR = 20 dB

14. Modified Soft Iterative DFE and MAP Decoding
Output
Data
Forward
Filter
Conversion
to LLR
MAP
Decoder +
P-1 -
DFE with
hard input
feedback
Feedback
Filter
Decision
Device
Soft encoded
symbols P +
Only extrinsic information is passed to the DFE from the decoder
Variance Estimator Re Im +
Hard detected
symbols +
Soft APP
from last
iteration
Variance Estimator

15. Performance of Soft Iterative DFE and MAP Decoder
Recursive systematic convolutional encoder with R=1/2, K=5,
Time-invariant 5 tap channel with a spectral null (Proakis C [5])
RLS updates at the DFE
Block length 4096
BPSK modulation

16. Performance of Soft Iterative DFE and MAP Decoder
Recursive systematic convolutional encoder with R=1/2, K=5,
Time-invariant 5 tap channel with a spectral null (Proakis C [5])
RLS updates at the DFE
Block length 4096
QPSK modulation

17. Iterative Linear MMSE Equalization and Decoding
SISO Linear MMSE Equalizer [6]:
Known channel:
Received signal:
Likelihood ratio for MMSE estimator output, :
Channel matrix:
MMSE estimator output:
where,

18. Iterative Linear MMSE Equalization and Decoding
Steps to compute symbol estimates with the Linear MMSE equalizer:
Soft output calculation assuming Gaussian distributed estimates:

19. Performance of SISO MMSE Linear Iterative Equalizer
Recursive systematic convolutional encoder with R=1/2, K=5,
Time-invariant 5 tap channel with a spectral null (Proakis C [5])
Equalizer has perfect knowledge of the channel
Block length 4096

20. Comparison of System Performances
Recursive systematic convolutional encoder with R=1/2, K=5,
Time-invariant 5 tap channel with a spectral null (Proakis C [5])
BER results after 6 iterations
Block length 4096

21. Experimental Study of Iterative Equalizers
Channel
Probe
Training
Symbols
Information
Dead
Time

22. Joint MMSE Equalization and Turbo Decoding
Le12
P-1
Le12
MAP
Decoder 1 P
xp1
Demapper
&
S/P
Converter
MAP
Decoder 2 xs
xp2 Ld
Lp2
Lp1 P
x(n) ^ P
P-1
s2(n)
e(n)
Adaptive
Algorithm +
e-jq(n)
y(n)
Forward
Filter x +
x(n) ~
Decision
Device
Feedback
Filter
x(n)
Training
Symbols

23. Decision Feedback Equalizer (DFE)
Soft output of the DFE:
RLS algorithm is used to track channel variation:
Noise variance estimate:

24. MAP Decoding Maximize a posteriori probability:
Decision variable written in the form of log-likelihood ratio:

25. MAP Decoding (BCJR Algorithm)
State transition probability:
where
channel
value
a priori
information
extrinsic
information

26. MAP Equalizer
depends on the channel trellis defined by hl(n) with 2(L-1) states
If xn-l for J<l<L-1 is known
Number of states is reduced to 2(J-1)

27. Per-Survivor Processing
Discarded Paths
Survivor Paths 00 01 1 1 1 1 1 10 1 1 11 1 1 1
n n n n n
Path metric:
Survivor path:

28. Each survivor path has a separate channel estimator
hl (n) + -
Adaptive
Algorithm
Each survivor path has a separate channel estimator
The input to the channel estimator, , is the estimates within the survivors
RLS algorithm is employed

29. Channel Estimator
Initial channel estimate is based on the correlation of the preamble
RLS algorithm is employed to track the channel
Noise variance estimate:

30. Experimental Results DFE results for transducer 7.
Channel impulse response estimate for transducer seven obtained using the channel probe
DFE results for transducer 7.
Eye Pattern - Filter coefficients
PLL phase estimate - Bit error distribution

31. Experimental Results
Channel impulse response estimate for transducer seven obtained using adaptive channel estimator
Comparison of received signal with the estimated received signal based on the channel estimate

32. Experimental Results
Sparse Channel with multipath delay in the order of 200 symbols
Length of the DFE or channel estimator filters cannot cover the channel
Sparse processing is needed

33. Results of DFE Turbo Decoder

34. Results of Iterative DFE Turbo Decoder

35. Results of Iterative DFE MAP Decoder

36. Results for Iterative Map Equalizer Turbo Decoder

37. Results for Iterative MAP Equalizer MAP Decoder

38. Conclusions
Due to error propagation in the DFE, turbo decoder cannot provide performance improvement beyond the second iteration  Error Floor
Joint DFE and turbo decoding adds an additional loop to the system and lowers the error floor
Joint channel estimator and iterative equalizer is able to decode packets with low SNR, which cannot be decoded with the DFE
Tail cancellation is an effective way to reduce the computational complexity of the MAP equalizer
If the channel is sparse, although the DFE filter lengths are short, the DFE is able to provide enough information to the turbo decoder
A sparse DFE can be used to improve the performance of the DFE/MAP Decoder and the DFE/Turbo Decoder

39. References
[1] S. Benedetto, et.al., “Serial concatenation of interleaved codes: Design and performance analysis,” IEEE Trans. Info. Theory, vol. 42, pp , April 1998
[2] I. Lee, “The effect of a precoder on serially concatenated coding systems with ISI channel,” IEEE Trans. Commun., pp , July 2001
[3] C. Douilard, et.al., “Iterative correction of intersymbol interference: Turbo-equalization,” European Transactions on Telecommunications, vol. 6, pp , Sep.-Oct. 1995
[4] G. Bauch, H. Khorram, and J. Hagenauer, “Iterative equalization and decoding in mobile communications systems,” in Proc. European Personal Mobile Commun. Conf., pp
[5] J. Proakis, Digital Communications, McGraw-Hill Inc., 2001
[6] M. Tuchler, A. Singer, and R. Koetter, “Minimum mean squared error equalization using a priori information,” IEEE Trans. Signal Proc., vol. 50, pp , March 2002