Iterative Multiuser Detection for Convolutionally Coded Asynchronous DS-CDMA

1
Iterative Multiuser Detection for Convolutionally
Coded Asynchronous DS-CDMA

• 9th IEEE International Symposium on
• Personal, Indoor, and Mobile Radio Communications
• Boston, MA September 9, 1998
• Matthew Valenti and Brian D. Woerner
• Mobile and Portable Radio Research Group
• Virginia Tech
• Blacksburg, Virginia

2
Introduction

• Performance of multiple access systems can be
  improved by multiuser detection (MUD).
• Verdu, Trans. Info. Theory 86.
• Implemented with Viterbi algorithm, complexity
  O(2K).
• Optimal MUD is too complex for large K.
• Suboptimal approximations
• Decorrelator, MMSE,DFE, PIC, SIC, etc.
• Most studies on MUD concentrate on the uncoded
  performance.
• Here we consider the effects of coding.
• We propose a receiver structure that approximates
  joint MUD and FEC-decoding.
• The algorithm allows for asynchronous users and
  fading.

3
MUD for Coded DS-CDMA

• Practical DS-CDMA systems use error correction
  coding (convolutional codes).
• Soft-decision decoding outperforms hard-decision
  decoding (by about 2.5dB).
• However, the optimal MUD passes hard-decisions to
  the channel decoder!
• Therefore it is possible for the coded
  performance of a system with MUD to be worse than
  the coded performance without the MUD.
• If MUD and FEC are to be used,the interface
  should be improved.
• The decoder for turbo codes gives insight on how
  to improve this interface.
• Use soft-decisions and feedback.

4
Relation to Other Work

• T. Giallorenzi and S. Wilson
• Optimal joint MUD/FEC-decoding
• Trans. Comm. Aug. 1996
• Uses a super-trellis.
• High complexity O(2WK)
• Suboptimal approaches.
• Trans Comm. Sept. 1996
• Separate MUD and Channel decoding.
• Soft values passed from MUD to channel decoder.
• No feedback used.
• See also P. Hoehers paper at ICUPC 93.

5
Relation to Other Work

• M. Reed, C. Schlegel, et al
• Feedback from FEC-decoder to MUD
• Similar to the decoder for turbo codes.
• Synchronous DS-CDMA
• One-shot detector.
• Convolutional codes
• Turbo Code Symp 97, ICUPC 97
• Turbo codes
• PIMRC 97
• Close to single-user bound for K5 users and
  spreading gain of N7.
• AWGN channel

6
Relation to Other Work

• M. Moher
• Feedback from FEC-decoder to MUD.
• Multiuser systems with high signal correlation.
• FDMA with overlapping signals.
• Random interleaving.
• Synchronous systems
• Trans. Comm., July 1998
• Asynchronous systems
• Comm. Letters, Aug. 1998
• Close to single user bound for K5,10 and
  ?0.6,0.75
• K-symmetric channel.
• AWGN

7
Turbo Codes and Iterative Decoding

• A turbo code is the parallel concatenation of two
  convolutional codes.
• An interleaver separates the code.
• Recursive Systematic Convolutional (RSC) codes
  are typically used.

RSC Encoder 1
Data
Output
interleaver
RSC Encoder 1
8
Turbo Decoding

• A turbo decoder consists of two elementary
  decoders that work cooperatively.
• Soft-in soft-out (SISO) decoders.
• Implemented with Log-MAP algorithm
• Feedback.
• Each decoder produces a posteriori information,
  which is used as a priori information by the
  other decoder.
• Iterative

A priori probability
A priori probability
SISO Decoder 1
SISO Decoder 2
Received Data
Estimated Data
9
Serial Concatenated Codes

• The turbo decoder can also be used to decode
  serially concatenated codes.
• Typically two convolutional codes.

n(t) AWGN
Outer Convolutional Encoder
Inner Convolutional Encoder
Data
interleaver
Turbo Decoder
interleaver
APP
Inner SISO Decoder
Outer SISO Decoder
Estimated Data
deinterleaver
10
Turbo Equalization

• The inner code of a serial concatenation could
  be an Intersymbol Interference (ISI) channel.
• ISI channel can be interpreted as a rate 1 code
  defined over the field of real numbers.

n(t) AWGN
(Outer) Convolutional Encoder
ISI Channel
Data
interleaver
Turbo Equalizer
interleaver
APP
(Outer) SISO Decoder
SISO Equalizer
Estimated Data
deinterleaver
11
Turbo Multiuser Detection

• The inner code of a serial concatenation could
  be a MAI channel.
• MAI channel can be thought of as a time varying
  ISI channel.
• MAI channel is a rate 1 code with time-varying
  coeficients over the field of real numbers.
• The input to the MAI channel consists of the
  encoded and interleaved sequences of all K users.

12
System Diagram
multiuser interleaver
Convolutional Encoder 1
interleaver 1
MAI Channel
MUX
n(t) AWGN
Convolutional Encoder K
interleaver K
Turbo MUD
multiuser interleaver
APP
Bank of K SISO Decoders
SISO MUD
multiuser deinterleaver
Estimated Data
13
MAI Channel Model

• Received Signal
• Where
• ak is the signature waveform of user k.
• ?k is a random delay (i.e. asynchronous) of user
  k.
• Pki is received power of user ks ith bit
  (fading ampltiude).
• Matched Filter Output

14
Optimal Multiuser Detection Algorithm Setup

• Place y and b into vectors
• Place the fading amplitudes into a vector
• Compute cross-correlation matrix

15
Optimal MUD Execution

• Run Viterbi algorithm with branch metric
• where
• Note that most derivations of the optimal MUD
  drop the p(b) term.
• Here we keep it.
• The channel decoder will provide this value.
• The algorithm produces hard bit decisions.
• Not suitable for soft-decision channel decoding.

16
Soft-Output MUD

• Several algorithms can be used to produce
  soft-outputs (preferably log-likelihood ratio).
• Trellis-based.
• MAP algorithm
• Log-MAP, Robertson et al, ICC 95
• OSOME, Hafeez Stark, VTC 97
• SOVA algorithm
• Hagenauer Hoeher, Globecom 89
• Non-trellis-based.
• Suboptimal, reduced complexity.
• Linear decorrelator, MMSE.
• Subtractive (nonlinear) DFE, SIC, PIC.

17
Simulation Parameters

• K5 users
• Power controlled (same average power).
• N7 (processing gain), code-on-pulse.
• Random spreading codes.
• Convolutional Code
• Constraint length 3.
• Rate 1/2.
• Interleaving
• 24 by 22 block interleaver (L528).
• Log-MAP decoding.
• Both MUD and channel decoder.
• 3 iterations.

18
Simulation Results AWGN Channel

• After the second iteration, performance is close
  to single-user bound for BER greater than 10-4.
• For BER less than 10-4, the curves diverge.
• This behavior is similar to the BER floor in
  turbo codes.
• Only a slight incremental gain by performing a
  third iteration.
• The extra processing for the third iteration is
  not worth it.

19
Simulation ResultsRayleigh Flat-Fading Channel

• Fully-interleaved Rayleigh flat-fading.
• i.e. fades are independent from symbol to symbol.
• After second iteration, performance is close to
  the single-user bound.
• The curves do not diverge as they did for AWGN.
• Why?
• The instantaneous received power is different for
  the different users.
• Therefore the MUD has one more parameter it can
  use to separate signals.

20
Conclusion

• A strategy for iterative MUD/FEC-decoding is
  proposed.
• Based on the concept of turbo processing.
• Similar to other researchers work, but the
  algorithm is generalized to allow
• independently faded signals
• code and bit asynchronism.
• Proposed strategy was illustrated by simulation
  example.
• Significant performance gain by performing 2
  iterations.
• When signals are independently faded, the
  algorithm exploits the differences in
  instantaneous signal power.

21
Future Work

• The study assumes perfect channel estimates.
• The effect of channel estimation should be
  considered.
• The estimator could be incorporated into the
  feedback loop.
• The proposed strategy is still very complex
• O(2W2K) per iteration.
• Future work should consider the use of reduced
  complexity multiuser detectors.
• This structure could also be used for TDMA
  systems.
• TDMA only a few strong interferers, small K.
• Highly correlated signals, can take advantage of
  this system.
• Can use observations from multiple base stations.
• See our work at VTC, ICUPC, and Globecom CTMC.