Sei sulla pagina 1di 4

820

IEEE SIGNAL PROCESSING LETTERS, VOL. 12, NO. 12, DECEMBER 2005

Turbo DFE Algorithm With


Imperfect Decision Feedback
Vladimir D. Trajkovic, Student Member, IEEE, Predrag B. Rapajic, Senior Member, IEEE, and
Rodney A. Kennedy, Fellow, IEEE

AbstractIn this letter, we propose a new turbo decision feedback equalizer (DFE) algorithm that takes into account decision
feedback error propagation. We derive the exact mathematical
expression for the probability density function (pdf) of soft decision feedback symbols, assuming that soft outputs from channel
decoder are independent identically distributed Gaussian random
variables with known mean and variance. We find a new set
of imperfect turbo minimum mean-square error DFE coefficients based on the derived pdf. The proposed turbo detector
outperforms other turbo equalization algorithms of similar computational complexity in terms of bit-error rate. The achieved
improvement is up to 3.8 dB for severe frequency-selective channels.
Index TermsDecision feedback equalizer (DFE), imperfect
feedback, turbo equalization.

I. INTRODUCTION

LOW-COMPLEXITY turbo equalizer, combining adaptive decision feedback equalizer (DFE) and soft-input
soft-output (SISO) decoding, has been proposed in [1]. A
hybrid turbo equalization scheme, combining maximum-likelihood sequence estimation (MLSE) at the first turbo iteration
and interference cancellers (ICs) at higher turbo iterations,
is analyzed in [2]. The ICs are determined according to the
minimum mean-square error (MMSE) criterion, assuming perfect outputs from the previous iteration. A turbo equalization
scheme combining a linear equalizer (LE) and SISO decoder
has been proposed in [3] and [4]. In the scheme, extrinsic
information from the decoder is used at each following iteration
in order to reduce intersymbol interference (ISI). A modified
version of linear MMSE (LMMSE) in [4] has been derived
in [5], where the correlation between the real and imaginary
parts of complex symbols is taken into account. This technique,
however, is not applicable to noncomplex signaling such as
binary phase-shift keying (BPSK). In [6], it was shown that the
adaptive turbo equalizer employing least-mean-square (LMS)

Manuscript received May 16, 2005; revised July 9, 2005. This work was supported in part by the University of New South Wales and in part by National
ICT Australia (NICTA). NICTA is funded through the Australian Governments
Backing Australias Ability Initiative, in part through the Australian Research
Council. The associate editor coordinating the review of this manuscript and
approving it for publication was Prof. Jonathon A. Chambers.
V. D. Trajkovic and P. B. Rapajic are with the School of Electrical Engineering
and Telecommunications, University of New South Wales, Sydney, NSW 2052,
Australia (e-mail: V.Trajkovic@student.unsw.edu.au; P.Rapajic@unsw.edu.au).
R. A. Kennedy is with the Research School of Information Science and Engineering, The Australian National University and National ICT Australia, Canberra ACT 0200, Australia (e-mail: rodney.kennedy@anu.edu.au).
Digital Object Identifier 10.1109/LSP.2005.859524

Fig. 1. System model.

adaptive algorithm outperforms the MMSE one where the coefficients are obtained according to perfect feedback assumption.
The main contributions of this paper are as follows.
1) We propose a new turbo equalization algorithm, i.e., a
new set of MMSE DFE and MMSE IC coefficients are
obtained without assuming perfect feedback. The derived
turbo equalizer reduces the computational complexity by
two orders of magnitude relative to the schemes proposed
in [3][5].
2) We find the probability density function (pdf) of soft decision feedback symbols assuming that soft outputs from
SISO channel decoder are independent identically distributed (i.i.d.) Gaussian random variables. The pdf is obtained analytically and corroborated by simulations.
3) We show that for time-invariant communication channels
that exhibit severe ISI, the proposed turbo equalizer outperforms the approximate LMMSE in [3] and [4] with
a signal-to-noise ratio (SNR) gain of 0.5 dB, the adaptive turbo equalizer [6] with a SNR gain of 2 dB, and the
conventional MMSE DFE with coefficients obtained assuming perfect feedback [2] with a SNR gain of 3.8 dB.
II. SYSTEM MODEL
The block diagram of the system model is shown in Fig. 1.
The binary information bits are encoded with a convolution code
of rate
. The analysis is also applicable to any code rate
.
The interleaved coded bits are symbol mapped, and the symbols
are transmitted through the ISI channel. At the receiver, the received signal is given by
(1)
where
is the Toeplitz channel impulse response matrix [7,
(4)],
is the input vector of transmitted symbols at time instant , and
denotes interleaving.
is the vector of additive
white Gaussian noise (AWGN) samples with covariance matrix

1070-9908/$20.00 2005 IEEE

et al.: TURBO DFE ALGORITHM WITH IMPERFECT DECISION FEEDBACK


TRAJKOVIC

. The received sequence is equalized, deinterleaved, and decoded using the SISO decoder. The decoder finds soft outputs
as log-probability ratios (LPRs) defined as [8]

821

is
be found in Appendix B. The feedback error variance
obtained by numerical evaluation of the second central moment
using the pdf in (4), which yields

(2)
Once the LPRs are calculated, the expectations used at the following iteration are obtained as [1]
(3)
For long sequences, the LPRs at the output of the SISO decoders
can be regarded as Gaussian random variables [9]. Combining
(3) and the Gaussian assumption about LPRs, the pdf of the soft
decision feedback signal at iteration can be found. This pdf is

(9)

Here we estimate the error variance over the whole frame while
typically symbol-wise estimates are used. The FFF at all iterations is fed by the received sequence . The output of the DFE
(IC) at iteration can be represented as
(10)

(4)

, where
and
given that the coded symbol
are the mean value and the variance of LPRs at iteration . The
proof of (4) can be found in Appendix A.
III. MMSE DFE AND IC WITH IMPERFECT FEEDBACK
The DFE consists of two linear filters, namely, feed-forward
filter (FFF) and feedback filter (FBF). In previous work on the
turbo DFE [2], the MMSE criterion has been considered as the
optimization criterion in determining the equalizer coefficients.
Based on the conventional perfect decision feedback assumption, a general expression for both MMSE DFE and MMSE IC
coefficients can be written as
(5)
(6)
where and are FFF and FBF coefficients, respectively, and
is the kth column of the channel matrix
related to the
currently detected symbol.
and
are explained in more
detail in Appendix B. For low SNRs, the feedback error cannot
be neglected, and a new solution that takes into account the error
propagation is found as
(7)
and
(8)
is the variance of the feedback error at iteration .
where
The error is defined as the difference between correct and estimated symbols, i.e.,
. The FFF/FBF coefficients are recalculated once per each turbo iteration, which is
in contrast to [4], where the coefficients are recalculated with
every coded bit at each iteration. The proof of (7) and (8) can

and
At the first iteration, the DFE is implemented and
are calculated using (5) and (6), and
is a vector containing
past hard decisions taken directly from the output of the DFE. At
, ICs are implemented and
and
all higher iterations
are calculated using (7)(9), and
is a vector containing
according to (3).
expectations from the previous iteration
Equations (7) and (8) represent MMSE solution when the feedback error is modeled according to (9). It is the property of the
MMSE solution [10] that it is unique and that all other values
for the FFF and FBF coefficients different from MMSE ones
must produce MSE [obtained according to (19)] that is always
. Consequently, the conventional MMSE DFE with
the coefficients obtained according to (5) and (6) (assuming per, and
fect feedback) produces a MSE that is always
the equality holds if and only if
.
IV. SIMULATION RESULTS
The ISI channel considered here has the impulse response
(also used in
given by
[3] and [5]). The results in Fig. 2 are obtained for BPSK. Information bits are encoded by 16-state recursive systematic convolutional (RSC) channel encoder with the generator polynomials
. The number of FFF taps is 21 at all turbo
given by
iterations, while FBF is ten taps long at the first turbo iteration
and 20 taps at all higher iterations. The soft output Viterbi algorithm (SOVA) [8] has been used in the SISO decoder. Fig. 2
shows BER performance comparison when the DFE coefficients
are determined using the proposed algorithm, the adaptive LMS
algorithm, and the conventional MMSE DFE. The results show
that the proposed detector outperforms both the adaptive and the
conventional detectors after certain number of turbo iterations
delivering a SNR gain of up to 3.8 dB at a BER of
. The
BER results show that the feedback error is better evaluated by
using (9) than using the adaptive LMS algorithm. The reason for
this is the relatively slow convergence speed of the LMS algorithm, especially in the tracking period when the adaptation rule
is not based on known training bits but on the soft decisions (3),

822

IEEE SIGNAL PROCESSING LETTERS, VOL. 12, NO. 12, DECEMBER 2005

Fig. 2. BER comparison for different turbo equalization algorithms for


BPSK: - - coded System (no ISI), - proposed, o - adaptive, - MMSE DFE
(conventional).

Fig. 3. BER comparison for different turbo equalization algorithms for QPSK:
- - coded system (no ISI), - modified LMMSE [5], - exact MMSE LE [3],
- proposed, - approx. MMSE LE [3], - MMSE DFE (conventional).

which for low SNRs can be unreliable. The slow convergence


speed of the adaptive LMS algorithm is a prohibitive parameter that prevents the adaptive LMS turbo equalizer from approaching the BER performance of the proposed detector. Simulations in Fig. 2 regarding the number of turbo iterations are
performed until no more improvement can be achieved. Simulation results shown in Fig. 3 are obtained for the parameters
equivalent to those in [3]. Generator polynomials are given by
. Coded bits are modulated using QPSK. The number
of FFF taps at all iterations is
, while FBF is seven taps
long at the first turbo iteration and 18 taps at all higher iterations.
The maximum a posteriori probability (MAP) [8] algorithm is
used as the decoding algorithm. Fig. 3 shows BER performance
comparison when the DFE coefficients are determined using the
proposed method, the exact and approximate LMMSE used in
[3], modified LMMSE [5], and the conventional MMSE DFE.
Our results show that the proposed method outperforms other

Fig. 4. Computational complexity comparison for different turbo equalization


- proposed, MMSE DFE (conventional), o - adaptive [6],
algorithms:
approx. MMSE LE [3], - modified LMMSE [5], exact MMSE LE [3].

methods of similar computational complexity. It achieves better


BER performance delivering an SNR gain of 0.5 dB relative to
approximate LMMSE and an SNR gain of 3.3 dB relative to
the conventional MMSE DFE (IC). The proposed method exhibits slightly worse performance relative to exact and modified LMMSE, i.e., SNR loss is around 1 dB at a BER of
.
However, the computational complexity comparison in Fig. 4
shows that the proposed algorithm requires around two orders
of magnitude fewer operations than both exact and modified
LMMSEs. The reason is that the proposed algorithm performs
only one inversion of the symmetric matrix in (7) per iteration.
The computational complexity of inversion of the symmetric
matrix can be reduced using Cholesky factorization to
.
The matrix inversion increases the computational complexity
per bit by
, where is the frame size, and it can be neglected for particular parameters. On the other hand, computational complexities of exact and modified LMMSE require one
matrix inversion per bit, which requires an order of complexity
that is
[4]. In addition, from Fig. 4, it can be noted that the
proposed algorithm has the least computational complexity relative to all other algorithms, and its excellent complexity/BER
performance tradeoff makes it a good candidate for practical implementations. The results in both Figs. 2 and 3 are obtained
bits of long random interleavers. Simulation reusing
sults, not presented in this letter, show that shorter frame sizes
cause the degradation in BER performance. However, it is yet
to be investigated whether this is due to short interleaver size or
whether the Gaussian assumption [9] on the LPRs is not sufficiently accurate for short data frames.
V. CONCLUSION
In this letter, we propose a new turbo DFE algorithm that
does not use the conventional assumption about perfect feedback and takes into account the feedback error propagation. We
derived the exact mathematical expression for the decision feedback error pdf, assuming that the soft outputs from a SISO decoder are i.i.d. Gaussian random variables [9]. A new set of
imperfect MMSE coefficients has been found based on the

et al.: TURBO DFE ALGORITHM WITH IMPERFECT DECISION FEEDBACK


TRAJKOVIC

derived pdf. BER performance of the new turbo equalizer has


been improved significantly, delivering an SNR gain of up to
3.8 dB relative to other turbo equalizers with similar computational complexity.

823

for the matrices

and
. The matrix
is obtained as
. From (10), the error at the equalizer output is
(17)

Combining (16) and (17)

APPENDIX A
PROOF OF (4)

(18)

To prove (4), we start from (3). Due to simplicity of notaand


, respectively.
tion, we use and instead of
Equation (3) becomes

where
MSE is

is a vector of decision feedback error samples. The

(19)

(11)
where

As referenced in Section II and [9], is assumed to comply with


the Gaussian probability distribution with mean and variance
and . From the fundamental theorem [11, p. 130], the pdf of
can be found as

and
.
is a
identity matrix, and
is the average
power of the received symbols, and it can be normalized to 1
without the loss of generality. If feedback errors are assumed to
be i.i.d. random variables with variance , then
. If we find the gradients and set them to , we get

(12)
(20)
are real roots of the equation
is the
where
, and
is Gaussian pdf of . From (11),
derivative of
we find that there is only one real root

and
(21)
From (20) and (21), it is straightforward that

(13)
(22)

and
(14)

and
(23)

Finally, from (12) and taking into account (13) and (14), we find
that the pdf of the random variable is

(15)

APPENDIX B
PROOF OF (5) AND (6)
From (1), the received sequence can be expressed as1
(16)
and
are referred to undecided (uncancelled) and
where
decided (cancelled) symbols, respectively. In this Appendix,
we omit the use of the time index and the iteration index
to simplify notation. For DFE, the matrix
takes the
,
following form:
while for IC, the matrix
is equivalent to
. Since the following analysis
is equivalent for both DFE and IC, we use general notation
1Please observe that interleaving sign is omitted here because the proof does
not depend on whether or not the sequence is interleaved.

REFERENCES
[1] A. Glavieux, C. Laot, and J. Labat, Turbo equalization over a frequency
selective channel, in Proc. Int. Symp. Turbo Codes, Brest, France, Sep.
1997, pp. 96102.
[2] Z.-N. Wu and J. Cioffi, Low-complexity iterative decoding with decision-aided equalization for magnetic recording channels, IEEE J. Sel.
Areas Commun., vol. 19, no. 4, pp. 565665, Apr. 2001.
[3] M. Tuchler and J. Hagenauer, Linear time and frequency domain turbo
equalization, in Proc. Vehicular Technology Conf. 01 Spring, Rhodes,
Greece, May 2001, pp. 14491453.
[4] M. Tuchler, R. Koetter, and A. Singer, Turbo equalization: Principles
and new results, IEEE Trans. Commun., vol. 50, no. 5, pp. 754767,
May 2002.
[5] S. Jiang, H. Sun, and C. S. Leung, Modified LMMSE turbo equalization, IEEE Commun. Lett., vol. 8, no. 3, pp. 174176, Mar. 2004.
[6] V. D. Trajkovic and P. B. Rapajic, Adaptive decision feedback turbo
equalization and multiuser detection, in Proc. IEEE Int. Symp. Spread
Spectrum Applications, Sydney, Australia, Sep. 2004, pp. 540544.
[7] N. Al-Dhahir and J. Cioffi, MMSE decision feedback equalizers:
Finitelength results, IEEE Trans. Inf. Theory, vol. 41, no. 4, pp.
961975, Jul. 1995.
[8] J. Hagenauer, E. Offer, and L. Papke, Iterative decoding of binary block
and convolutional codes, IEEE Trans. Inf. Theory, vol. 42, no. 2, pp.
429445, Mar. 1996.
[9] S. Ten Brink, Convergence behavior of iteratively decoded parallel concatenated codes, IEEE Trans. Commun., vol. 49, no. 10, pp. 17271737,
Oct. 2001.
[10] S. Haykin, Adaptive Filter Theory, 4th ed. Englewood Cliffs, NJ: Prentice-Hall, 2002.
[11] A. Papoulis and U. Pillai, Probability, Random Variables and Stochastic
Processes, 4th ed. New York: McGraw-Hill, 2002.

Potrebbero piacerti anche