Concurrent Blind Decision Feedback Equalizer

Magno T. M. Silva,
Universidade de S o Paulo - EPUSP - PTC a Av. Prof. Luciano Gualberto, 158, trav. 3 05508-900 - S o Paulo - SP - Brazil a magno@lcs.poli.usp.br

Maria D. Miranda, and Rita Soares

Universidade Presbiteriana Mackenzie - PPGEE Rua da Consolacao, 869 01302-907 - S o Paulo - SP - Brazil a mdm@mackenzie.com.br rita.soares@iqaratelecom.com.br

Abstract Decision Feedback Equalizers (DFEs) play an important role in digital communication systems. They reduce the effect of intersymbol interference, holding a favorable compromise between computational complexity and efcient behavior. When blind adaptation of DFEs is required, algorithms based on the constant modulus criterion may converge to so-called degenerative solutions. Among different algorithms which avoid these undesirable solutions, the one proposed by Szczecinski and Gei is the most efcient. However, in many practical situations, as for example time-variant channels, its good behavior is not always assured. In this paper, it is proposed an improvement of that algorithm based on a concurrent blind deconvolution method. At the cost of a moderate increase in computational complexity, the resulting algorithm presents faster convergence and lower symbol error rate than the previous one. Illustrative simulations considering M -QAM signaling and high denition television (HDTV) channels are presented. Index Terms Decision Feedback Equalizer (DFE), Constant Modulus Algorithm (CMA), Concurrent Algorithms.

I. I NTRODUCTION Decision Feedback Equalizers (DFEs) have been widely used to remove intersymbol interference from the data received through communication channels [1]. They present good performance in difcult environments as, for example, channels with long and sparse impulse response, non-minimum phase, spectral nulls or non-linearities [2]. This behavior can be justied by the interpretation of DFEs as fractionally spaced equalizers which, under certain conditions, are able to achieve the zero-forcing solution [3]. In many practical applications, DFEs must be blindly adapted. In this case, algorithms based on the constant modulus criterion [4] for joint updating of the feedforward and feedback lters may converge to so-called degenerative solutions. This occurs when the signal at the equalizer output is independent of its input [3], [2]. To avoid this problem, many authors consider a decoupled adaptation of the feedforward and feedback lters [5]-[8]. Nevertheless, this adaptation form is less efcient than the joint adaptation, mainly if the channel is time-variant [2].
This work was supported by Fapesp (00/12350-6), CNPq, and Mackpesquisa.

In this scenery, Szczecinski and Gei proposed a new criterion for reliable detection of degenerative solutions [2]. This criterion is based on the Godard cost function with constraint on the feedforward and feedback lters and is minimized by a stochastic algorithm called DFE-CMA-FB (DFE Constant Modulus Algorithm with Feedback). It was shown in [2] that this algorithm avoids degenerative solutions, being more efcient than the previous algorithms proposed in the literature, which impose constraint only on the feedforward lter [3]. In the context of blind adaptation of Linear Transversal Equalizers (LTEs), Castro et al. [9] proposed to operate a Decision-Direct (DD) algorithm concurrently with CMA (Constant Modulus Algorithm) for M -QAM (M -Quadrature Amplitude Modulation) signaling. At the cost of a moderate increase in computational complexity, CMA+DD presents a great improvement in equalization performance over CMA. After that, inspired by this idea, Chen [10] proposed a modication on this scheme based on a novel soft decision-direct (SDD) adaptation. The CMA+SDD, also called bootstrap maximum a posteriori probability (bootstrap MAP) [11], has lower computational complexity and faster convergence rate than CMA+DD. Furthermore, these concurrent algorithms have capability of correcting the phase rotation for M -QAM signaling. Since concurrent algorithms outperform nonconcurrent algorithms for linear equalization, one expects that they can improve the performance of DFEs. Thus, we consider the combination of DFECMA-FB with SDD for joint blind adaptation of the feedforward and feedback lters and M -QAM signaling. The algorithm, called NDEG-CMA-SDD (Non-Degenerative Constant Modulus Algorithm with Soft Decision-Direct), beyond avoiding degenerative solutions, should outperform DFE-CMA-FB. The paper is organized as follows. The blind DFE and Chens concurrent algorithm are revisited in sections 2 and 3 respectively. The proposed NDEG-CMASDD is introduced in Section 4. Thereafter, some simulation results are presented and a conclusion section closes the paper.

PROCEEDINGS OF THE INTERNATIONAL WORKSHOP ON TELECOMMUNICATIONS

II. T HE BLIND DFE A communication system with a typical DFE model is depicted in Figure 1 [12]. The signal a(n), assumed independent, identically distributed, and non Gaussian, is transmitted through an unknown channel whose model is constituted by an FIR (Finite Impulse Response) lter H(z) and additive white Gaussian noise (n). The received signal u(n) is ltered by an FIR feedforward lter F (z) with Mf taps. The past decisions are fed back and ltered by an FIR feedback lter B(z) with Mb taps. Then, a linear combination of the lters outputs enters to the decision device. The blind DFE must mitigate the channel effects and recover the signal a(n) for some delay d .
(n )
a (n )

TABLE I S UMMARY OF DFE-CMA-FB. Initialize the algorithm by setting: f (0) = [ 0 0 1 0 0 ] b(0) = 0, g(0) = 0, Eyf (0) = 0

For n = 1, 2, , compute: yf (n) = f H (n 1)u(n) yb (n) = bH (n 1) (n) a y(n) = yf (n) + yb (n)

a e(n) = (R2 |y(n)|2 )y(n)

Eyf (n) = Eyf (n 1) + (1 )|yf (n)|2

g(n) = g(n 1) + (1 )u(n)yf (n)

c(n) = b(n 1)

Eyf (n)

H (z )
Channel

u (n )

Feedforward filter

(n) = o (c(n)) f (n) = f (n 1) + [e (n)u(n) + (n)g(n)] b(n) = [1 (n)]b(n 1) + e (n) (n) a

F (z)
y f (n )

y(n )

Decision device

a (n d )

z 1 Delay
yb (n )

B( z)
Feedback filter

A common problem of blind equalization based on the constant modulus criterion is the random phase rotation [13]. In general, the output of the equalizer is given by y(n) = ej a(nd )+(n), with [0, 2[ and complex noise sample (n). The phase rotation must be compensated before the decision device. The literature contains several blind phase recovery algorithms [13], but for simplicity we assume in this paper the modied Phase Tracking Algorithm (PTA) [2], [12], which provides the following phase update (n + 1) = (n) + Im(y (n) (n d )) a being y (n) = y(n)ej(n) , the step-size, and Im(v) the imaginary part of v. III. C HEN S CONCURRENT ALGORITHM In the concurrent scheme for blind adaptation of LTEs, proposed by Castro et al. [9], to avoid error propagation due to incorrect decisions, the updating of the Decision-Direct coefcients is executed after that of CMA. Its adjustment is made only if the CMA adaptation has achieved a successful updating with high probability [9], [10]. In the Chens concurrent scheme [10], this adaptation form is also considered but with a Soft DecisionDirect (SDD). The use of SDD instead of DecisionDirect makes Chens concurrent algorithm more efcient than the original one proposed in [9]. Moreover, the adaptation of the CMA and SDD coefcients occurs simultaneously. In the sequence, this algorithm is revisited. Assume that the LTE coefcient vector is w(n) = wc (n) + wd (n), (3)

Fig. 1.

A typical DFE model.

The summary of DFE-CMA-FB is shown in Table I. Its detailed derivation can be found in [2]. This algorithm avoids degenerative solutions and provides a joint blind adaptation of F (z) and B(z). The following notation and denitions are used: ()T and ()H stands respectively for transpose and complexconjugate transpose of a vector and () for complexconjugate; f and b are the coefcients of the feedforward and feedback lters respectively. The vectors u(n) and a(n) are dened by u(n) = [u(n) u(n 1) u(n Mf + 1)]T , (1) a(n) = [(nd 1) a(nd Mb )]T , a (n) is the Lagrange multiplier, = o > 0 if c(n) > 0, (x) = {1 if x 0; 0 if x < 0} is the step function, is a forgetting factor, the step-size, and a R2 E{|a(n)|4 }/E{|a(n)|2 } the expected modulus, being E{} the expectation operator [2], [4]. It is usual to assume o = 2 and = 0.95.

(2)

PROCEEDINGS OF THE INTERNATIONAL WORKSHOP ON TELECOMMUNICATIONS

being wc and wd coefcient vectors updated by CMA and SDD algorithm respectively and that the equalizer output is y(n) = wH (n 1)u(n). After successful convergence, the equalizer output can be modelled by y(n) a(n d ) + (n) (4)

where (n) = R (n)+jI (n) is approximately a complex Additive White Gaussian Noise (AWGN) with 2 2 E{R (n)I (n)} = 0 and E{R (n)} = E{I (n)} = . Thus, the a posteriori probability density function (p.d.f.) of y(n) can be approximated by
K K

p(y(n))

k=1 l=1

|y(n) akl |2 Pkl exp 2 2

(5)

where akl takes the value from the M -QAM symbol set dened by A = {akl = (2kK 1)+j(2lK 1), 1 k, l K} M and Pkl the a priori probability with K = of akl . For equiprobable symbol transmission Pkl = 1/M, 1 k, l K. Let the complex plan be divided into M/4 regular regions. As indicated in Figure 2, each region Ai contains four symbols, that is Ai = {aim , m = 1, 2, 3, 4}, being i = 1, 2, . . . , M/4. If the equalizer output is within the region Ai , a local approximation to the a posteriori p.d.f. of y(n) is

L-stage process. The identication of the region Ai does not cause a signicant increase in computational complexity since it requires only 2(L1) comparisons. Thus, the algorithms complexity is always equivalent to the minimum complexity of the 4-QAM case [11]. Considering for example 64-QAM, the complex plan is divided in 64/4 = 16 regions. Each region must contain 4 symbols as shown in Figure 2. Suppose that the equalizer output is 5.3 + j5.7. The rst stage consists in classifying the output into the 4 quadrants of the complex plan. In the example the output is in the rst quadrant because its real and imaginary parts are positive. At the second stage, the output must be classied into other 4 regions and at the last stage (n) can be calculated using the four symbols aim , m = 1, 2, 3, 4 of the nal region Ai . The concurrent CMA+SDD has a computational complexity that lies between those of CMA and the concurrent CMA+DD [10]. Due to soft-decision direct nature, the step-size range which ensures the convergence of CMA+SDD is larger than that of CMA. Moreover, as shown in [11], CMA+SDD does not suffer from a serious phase rotation as CMA does.

Im
7

ai 2

Ai

ai1

equalizer output
5

ai 3

ai 4

pi (y(n))

|im (n)|2 1 exp 8 2 m=1

(6)
3

where im (n) = aim y(n) and the a priori probability has been set to 1/4. The SDD algorithm maximizes log of the local a posteriori p.d.f criterion given by JM AP = E{ log[i (y(n))]}. p (7)

Re
1 3 5 7

Using a stochastic gradient algorithm to adjust wd (n), we arrive at wd (n) = wd (n 1) + d (n)u(n) where |im (n)|2 im (n) exp 2 m=1 |im (n)|2 exp 2 m=1
4 4

(8)

Fig. 2. Illustration of local regions for SDD adaptation and 64QAM (rst quadrant of the complex plan).

IV. T HE PROPOSED ALGORITHM Motivated by the good behavior of Chens concurrent algorithm for blind adaptation of LTEs and M -QAM signaling, we propose a concurrent algorithm for blind adaptation of DFEs. The algorithm, named NDEG-CMA-SDD (Non-Degenerative CMASDD), operates concurrently the SDD algorithm with DFE-CMA-FB. Thus, one expects that it avoids degenerative solutions as DFE-CMA-FB and acquires the good behavior of concurrent algorithms. Its operations consist basically of those shown in Table I added to SDD operations. The algorithm is summarized in Table II. Note that wf and wb are respectively the feedforward and feed-

(n) =

(9)

To ensure a proper separation of the clusters, the parameter should be lower than the half of the minimum distance between two neighbouring symbol points [10]. As indicated in (9) the SDD adaptation needs to compute only 4 exp( ) at each iteration, which can be implemented through a table. The task of M QAM equalization, where M = 22L , is achieved with

PROCEEDINGS OF THE INTERNATIONAL WORKSHOP ON TELECOMMUNICATIONS

back coefcient vectors updated by the SDD algorithm (Equation (8)) and f and b the coefcient vectors updated by DFE-CMA-FB (Table I). Table III shows the computational complexity of the algorithms. The complexity of NEG-CMA-SDD is compared to those of DFE-CMA+PTA and DFECMA-FB+PTA. Although the conventional CMA for blind adaptation of DFEs, called DFE-CMA [2], does not avoid degenerative solutions, it is considered here for comparison. The NDEG-CMA-SDD complexity shown in Table III was evaluated for 4-QAM signaling. In the case of M -QAM signaling, to identify the region Ai , it is necessary to add 2(L 1) comparisons. It is relevant to note that the computational complexity of the proposed algorithm is moderately higher than that of DFE-CMA-FB.
TABLE II S UMMARY OF NDEG-CMA-SDD. Initialize the algorithm by setting: f (0) = [ 0 0 1 0 0 ], wf (0) = g(0) = 0, Eyf (0) = 0

input signal are normalized to have unit power, i.e. 2 2 a = u = 1.

TABLE III C OMPUTATIONAL COMPLEXITY OF THE ALGORITHMS PER
WEIGHT UPDATE .

Op.

DFE-CMA +PTA
8Mf +8Mb +19 8Mf +8Mb +7

DFE-CMA-FB +PTA
16Mf +12Mb +28 14Mf +10Mb +11

NDEG-CMASDD
20Mf +16Mb +36 20Mf +16Mb +25 1 4

exp

a) 1 0.5 hk 0 0

b(0) = wb (0) = 0
5 10 k 15 20 25

For n = 1, 2, , compute: yf (n) = f (n 1) + wf (n 1)

H H

u(n)
b) 10 Magnitude (dB) 0 10 20 0 Phase (rad) 10 20 30

yb (n) = [b(n 1) + wb (n 1)] a(n) y(n) = yf (n) + yb (n)

a e(n) = (R2 |y(n)|2 )y(n)

Eyf (n) = Eyf (n 1) + (1 )|yf (n)|2

g(n) = g(n 1) + (1 )u(n)yf (n)

0.5

c(n) = b(n 1) (n) = o (c(n))

Eyf (n)

f (n) = f (n 1) + [e (n)u(n) + (n)g(n)] b(n) = [1 (n)]b(n 1) + e (n) (n) a Identify the region Ai making 2(L 1) comparisons and compute: |im (n)|2 im (n) exp 2 m=1
4 4

0.5 Nomalized frequency

(n) =

exp
m=1

|im (n)|2 2

Fig. 3. a) Impulse response and b) frequency response of the HDTV channel used in simulations.

wf (n) = wf (n 1) + d (n)uf (n) wb (n) = wb (n 1) + d (n) (n) a

V. S IMULATION R ESULTS A channel of high denition television (HDTV) terrestrial metropolitan transmission is assumed [1]. In this case, the channel presents large echoes fairly close to the main path. Such a scenario is common where large buildings are close to the receiver, e.g. in New York [1]. The coefcients and frequency response of the channel are depicted in Figure 3. In the simulations, the transmitted signal and the equalizer

The performance of NDEG-CMA-SDD is compared to those of DFE-CMA-FB+PTA and DFE-LMS (DFELeast Mean Square) for the DFE training with Mf = 11, Mb = 17 and 16-QAM signaling. Figure 4 shows the measurement of the symbol error rates (SER). As expected, the supervised DFE-LMS algorithm without error propagation (WEP) has the best performance. NDEG-CMA-SDD presents a performance close to DFE-CMA-FB+PTA for signal-to-noise ratio (SNR) below 19 dB but outperforms it for SNR > 19 dB becoming close to DFE-LMS for SNR > 22 dB. Thus, the proposed algorithm presents a behavior close to DFE-CMA-FB+PTA for lower SNR and outperforms it for higher SNR.

PROCEEDINGS OF THE INTERNATIONAL WORKSHOP ON TELECOMMUNICATIONS

0 1 log (SER) 2 3 4 5 6 15 DFELMS (WEP) DFECMAFB NDEGCMASDD 17 19 21 SNR (dB) 23 25

CMA-FB+PTA presents error bursts (Figure 6-c) while DFE-LMS (WEP) (Figure 6-a) and NDEG-CMA-SDD (Figure 6-b) have faster recuperation.
4 2 a) 0 2 4 4 2 b) 0 2 4 4 0 1000 2000 3000 4000 0 1000 2000 3000 4000

Fig. 4. Plot of decimal logarithm of SER for Mf = 11, Mb = 17, 16-QAM, channel of Figure 3, d = 10, = 103 , d = 102 , = 102 , = 0.06, o = 1, and = 0.95.

Figure 5 shows the decision based Mean Square Error MSE = E{|y(n)(nd )|2 } of NDEG-CMAa SDD and DFE-CMA-FB+PTA for SNR = 22.5 dB and 16-QAM signaling. To facilitate the visualization, the MSE signals were ltered by a moving-average lter with 32 coefcients. In this case, the algorithms were experimentally adjusted to reach the same steadystate MSE. Note that NDEG-CMA-SDD has faster convergence rate than DFE-CMA-FB+PTA.
10 NDEGCMASDD 12 MSE (dB) 14 16 18 20

10

2 c) 0 2 4 1 d) 0.8 0.6 0 1000 2000 iterations 3000 4000 0 1000 2000 3000 4000

DFECMAFB

Fig. 6. Output errors, 16-QAM, SNR=20 dB, Mf = 11, Mb = 17 a) DFE-LMS ( = 102 , d = 5), b) NDEG-CMA-SDD ( = 2 103 , d = 2 102 , = 0.06, o = 1, = 0.95) c) DFECMA-FB+PTA ( = 5 104 , = 0.06, o = 1, = 0.95, = 5 103 ), d) Absolute root values of the dynamic multipath channel.

2 iterations

4 x 10

6
4

Fig. 5. Decision based MSE for DFE-CMA-FB+PTA ( = 2.5 104 , = 0.06, o = 1, = 0.95, = 2.5 103 ) and NDEG-CMA-SDD ( = 103 , d = 102 , = 0.06, o = 1, = 0.95), Mf = 11, Mb = 17, SNR = 22.5 dB, 16-QAM, and channel of Figure 3; average over 50 experiments.

Figure 7 shows the measurement of the symbol error rates for the dynamic multipath channel. Due to error bursts DFE-CMA-FB+PTA presents high values of SER. The DFE-LMS algorithm shows the best performance followed by NDEG-CMA-SDD. However, the latter suffers with low values of SNR. For SNR < 17.5 dB its performance becomes worse than that of DFE-CMA-FB+PTA. As previously observed, the good behavior of the proposed algorithm is obtained for high values of SNR.
0 1 log10(SER) 2 3 4 5 6 15 DFELMS (WEP) DFECMAFB NDEGCMASDD 20 SNR (dB)
Fig. 7. Plot of decimal logarithm of SER for Mf = 11, Mb = 17, 16-QAM, and dynamic multipath channel.

In order to investigate the efcacy of the algorithms in tracking dynamic multipath we consider a channel whose coefcients are given by [ 1 0 0 0 0 0 0 0 0 0 0 0.5 sin(2f nT )], where f = 5 Hz, T = 0.0929 ms, and n is the time index [1]. Figure 6 shows errors at the output of decision devices by considering 16-QAM signaling, Mf = 11, and Mb = 17. The SNR was set to 20 dB and the step-size of each algorithm was experimentally adjusted to obtain the best tracking performance. The absolute values of the channel roots are shown if Figure 6-d. The channel has 11 zeros homogeneously distributed in a circle. Thus, all the zeros have the same absolute values. The channel becomes more difcult to equalize when the zeros are close to the unit circle. In this case, DFE-

25

30

PROCEEDINGS OF THE INTERNATIONAL WORKSHOP ON TELECOMMUNICATIONS

VI. C ONCLUSION We have proposed an improvement of DFE-CMAFB for blind adaptation of DFEs and M -QAM signaling based on a concurrent method which uses a Soft-Decision Direct algorithm. At the cost of a moderate increase in computational complexity, we showed by means of simulations that the resulting algorithm presents good tracking performance and lower symbol error rate than DFE-CMA-FB for high values of SNR. R EFERENCES
[1] M. Ghosh, Blind decision feedback equalization for terrestrial television receivers, Proceedings of the IEEE, vol. 86, pp. 2070-2081, Oct. 1998. [2] L. L. Szczecinski and A. Gei, Blind decision feedback equalisers, how to avoid degenerative solutions, Signal Processing, vol. 82, pp. 1675-1693, 2002. [3] C. B. Papadias and A. Paulraj, Decision-feedback equalization and identication of linear channels using blind algorithms of the bussgang type, Proceedings of ASILOMAR-29, pp. 335340, Pacic Grove, Oct. 1995. [4] D. N. Godard, Self-recovering equalization and carrier tracking in two dimensional data communication system, IEEE Transactions on Communications, pp. 1867-1875, Nov. 1980. [5] R. A. Casas et al., Blind adaptation decision feedback equalizers based on constant modulus algorithm, Proceedings of ASILOMAR-29, vol. 1, pp. 698-702, Pacic Grove, Oct. 1995. [6] L. Tong and D. Liu, Blind predictive decision-feedback equalization via the constant modulus algorithm, Proccedings of ICASSP97, pp. 3901-3904, Munich, Apr. 1997. [7] J. Labat, O. Macchi, and C. Laout, Adaptive decision feedback equalization: can you skip the trainning period?, IEEE Transactions on Communications, v. 46, pp. 921-930, July 1998. [8] R. A. Casas et al., On initialization strategies for blind adaptive DFEs, Proceedings of WCNC99, vol. 2, pp. 792796, New Orleans, Sept. 1999. [9] F. C. C Castro, M. C. F. Castro, and D. S. Arantes, Concurrent blind deconvolution for channel equalization, Proceedings of ICC2001, vol. 2, pp. 366-371, Helsink, 2001. [10] S. Chen, Low complexity concurrent contant modulus algorithm and soft directed scheme for blind equalization, IEE Proceedings - Vision, Image, and Signal Processing, v. 150, pp. 312-320, Oct. 2003. [11] S. Chen, T. B. Cook, and L. C. Anderson, A comparative study of two blind FIR equalizers, Signal Processing, v.14, pp.18-36, 2004. [12] J. Proakis, Digital Communications, McGraw-Hill, New York, 1995. [13] A. Haidary and M. Nasiri-Kenari, A family of simple blind phase recorevy algorithms, Proc. of ISPACS2000, vol. 1, pp. 24-28, Honolulu, 2000.