L. Fan et al.

: Efficient Robust Adaptive Decision Feedback Equalizer for Large Delay Sparse Channel

449

Efficient Robust Adaptive Decision Feedback Equalizer for Large Delay Sparse Channel
Lingyan Fan, Chen He, Dongjian Wang, and Lingge Jiang
Abstract In this paper, an efficient decision feedback equalizer (DFE) is presented for sparse channels with large delay echoes, which are encountered in many high-speed wireless communication applications. Unlike the conventional DFE, this proposed equalizer considers that the channel is sparse and large number of DFE taps lead to noise accumulation problem. The new method can achieve efficient equalization by setting the active group for equalization taps and clearing the abandoned taps mechanism. As a result, it can both reduce hardware implementation complexity effectively and improve the performance compared to the conventional DFE. Brief analysis about noise accumulation effects in large-scale equalizer and working mechanism description of this effective algorithm exhibits us where the performance gain lies. Simulation results show that this modified DFE exhibits considerable computational savings, faster convergence, and better performance and improved tracking capabilities than the conventional ones. It is quite fit for applications in high-speed systems, such as high definition television (HDTV) and broadband mobile communication1.
Index Terms Complexity reduction, Decision Feedback Equalizer, Large delay sparse channel.
I.

INTRODUCTION

In many digital wireless communication systems, the received signal is often distorted by inter-symbol interference (ISI) due to multi-path channels [1,2]. Decision feedback equalizer (DFE), including feed forward filter (FFF) and feedback filter (FBF), is widely used to reduce ISI as an effective technique for its relatively low complexity and good performance. In high-speed wireless communications systems, such as high definition television (HDTV) terrestrial transmission with transmission symbol rates of 10.76MHz, the ISI has a span of several tens up to hundreds of symbol intervals under the HDTV test channels with the large delay spreads. In order to reduce effectively the ISI component of the received signal, the quite large-scale equalizer is required in the conventional DFE architecture, which implies heavy computational load in hardware implementation.
1 This work was supported in part by natural science foundation of China under Grant No.60372076. Lingyan Fan. is with the Electrical Engineering Department, Shanghai Jiao Tong University, China (e-mail: fanlingyan@ sjtu.edu.cn). Chen He. is with the Electrical Engineering Department, Shanghai Jiao Tong University, China (e-mail: chenhe@sjtu.edu.cn). Dongjian Wang. is with PixelWorks. China. (e-mail: dwang@pixelworks.com). Lingge Jiang. is with the Electrical Engineering Department, Shanghai Jiao Tong University, China.

For the channels of the high-speed system with a few nonnegligible echoes, the channel responses and tap of DFE exhibit sparse behavior. Under large sparse channel, many effective algorithms (including architectures) about DFE by creative researchers have been proposed to reduce complexity and improve efficiency in the last years (some of them are now in wide application) [5,6]. Ian J.Fevrier has proposed some new DFE structure in [5], such as PFE, CFE and TDFE, in which, a new equalization mechanism of subtracting post-echo ISI before feed forward filtering was proposed. Reduction in complexity is achieved due to the small size of the FBF and proper selection of a limited number of FBF coefficients. In [6], another DFE is presented, whose FBF comprises a reduced number of taps due to estimation of channel by a novel method. Both above schemes are done after the channel impulse responses have been estimated. However, there are few discussions that refer to the noise accumulation problem in large scale DFE [10] for large delay sparse channel, which is caused by adaptive updating weigh algorithm or background noise. Due to large scale, noise accumulation is so sever that it affects the performance of the equalization. This problem is investigated specially in this paper. Following the analysis for noise resistant, a novel DFE algorithm is proposed in the paper. It can both improve the performance of system for removing noise accumulation effects in large-scale equalizer and reduce the number of multipliers used in implementation. Besides, the scheme can track the variety of the channel by blind adaptive algorithm, which is not including the estimation for channel. The paper is organized as follows. Sparse channel model is described in Section II. And conventional DFE architecture and problems in it under large delay sparse channel are presented in Section III. Base on noise accumulation effects analysis in Section III, the proposed DFE algorithm is explained in section IV. Simulation results are shown in Section V and conclusions are reached in the final Section VI. II. SPARSE CHANNEL MODEL Assume that the transmitted signal has a symbol period of

Ts and the corresponding symbol rate is f s = 1 Ts . After pulse

shaping, the signal will be transmitted through the multi-path channel. Under time-domain delay-line based multi-path model, the impulse response of the channel is given by (1) h(t ) = a (t )e ji , where the ai , i and i represent the attenuation, delay, and phase rotation of the i-th echo, respectively and (t ) is the
i i i

Contributed Paper Manuscript received April 7, 2005

0098 3063/05/$20.00 2005 IEEE

450

IEEE Transactions on Consumer Electronics, Vol. 51, No. 2, MAY 2005

unit impulse response. Under some non-static channel conditions, phase of the echoes may vary with time, which can be represented as i = 2 fd t , f d is the Doppler rotation rate in Hz. The received signal is then match filtered against the (square root Nyquist) transmitted pulse and get complex base-band equivalent received signal as (2) y (t ) = I c(t nT ) + n(t ) ,
n n s

but only less than ten of them are larger than 10dB (relative to the main signal) under given transmission parameters. That is to say, the channel is long spares and has many negligible echoes.

where c (t ) includes effects of pulse shaping filter, multi-path channel and the receiver filter; and {I n } are independent Mary complex symbols with variance I2 = 1 E{| I n |2 } , which are 2 generated every Ts second. The component n(t ) is the independent complex zero mean background noise (always modeled as AWGN noise) whose variant is E[ n(t ) 2 ] 2 = N 0 . And n(t ) is independent of {I n } For high-speed wireless system, the multi-path channel impulse response can be divided into pre-echoes (that reach the receiver earlier than the main signal) and post-echoes (reach later than main signal). Both of them will reflect the DFE taps complex coefficients after convergence. For highspeed wireless channels, such as ATSC test channel [3,4], it often occurs that the pre-echo is strong (which induces a series of large tap coefficients in module) and the post-echo is long and sparse. The features of the pose-echo and pre-echo are present in Fig.1 and Fig.2, respectively.
Fig.2 Channel impulse response with pre-echo at 14 s

III. CONVENTIONAL DFE AND NOISE ACCUMULATION In this section, the conventional DFE is present. Under large sparse channel, conventional DFE faces two problems. One of them is noise accumulation, which is caused by equalization algorithm of the DFE and background noise. If the channel is short, the DFE can work well for the noise is limited in some allowable bound. However, when channel is large and timevariety, the noise accumulates so great that the DFE breakdowns. The other one is heavy computational load per iteration in hardware implementation due to the long span of the introduced ISI by large delay channel. A. Conventional DFE model Typical DFE is composed of a FFF that is a FIR filter used to cancel pre-echoes and near post-echoes and a FBF that is an IIR filter to cancel long post-echoes as shown in Fig.3. yk + M
D D D
D out put FE

f0

f1

f M 1

Ik

Ik

FFF

gN
FBF

g N 1

g1

Fig.1 Channel impulse response with post-echo at 57 s

The echoes delay may not be accurate integer multiple of the symbol period, which acquire the Ts -spaced equalizer to adapt its nearby one or more taps to counteract them. We shall limit discussion in the symbol-spaced equalizer structures. Assuming that symbols are received synchronously by receiver, the output of the receiver filter can be described as following (3) y = I c +n ,
k n k n n k

Fig.3 Conventional DFE architecture

Input vector of DFE is represented as

y , while the output

vector is represented as I . Before slicing, the equalizer output

where yk = y (kTs ) , cn = c(nTs ) , nk = n(kTs ) , and {nk } is independent of {I k } . Due to Fig.1, the HDTV channel will have echoes whose arriving intervals are larger than hundreds of symbol-intervals,

I k of k -time satisfies
Ik =
M 1 i=0

f i yk + M i +

N j =1

g j I k j ,

(4)

where f i and g j are the FFF and FBF coefficients (either real or complex), M and N are the corresponding filter tap

L. Fan et al.: Efficient Robust Adaptive Decision Feedback Equalizer for Large Delay Sparse Channel

451

number. I k is the output decision. The DFE parameter satisfies the assumption in [5], i.e., the FFF and FBF tap length is greater than or equal to the channel length. Now under the large delay channel, several hundreds of complex multiply operations are required for calculating each I k . B. Noise accumulation The output of the DFE

working properly, it is reasonable to assume that f i , is

2 zero mean as F,i = E[ fi, ] = 0 . However, the variation F ,i

is decided by factors like the equalization algorithm type, the updating step-sizes, background noise level and etc. 2) Secondly, input vector element yk + M i of FFF is a decided value to a specific i . However, it can be seen as 2 satisfying normal distribution y ~ ( Y , Y ) to different samples yi . Then specific product fi , yk + M i is a process satisfying
2 (6) f i , yk + M i ~ (0, yk2+ M i F ,i ) . 3) Thirdly, the products fi , yk + M i are in-dependant to each other for different i .

I k is the estimation of the I k is I k . That is to say, the DFE

N ) compensate the

transmission symbol. Under ideal no-error transmission, equal to transmitted symbol coefficients f i (i = 1 environment. In fact,
M ) and g j ( j = 1

channel multi-path fading, time variety fading and noise in the

I k has always different gap with I k due

Under the above conditions, summation

M 1 i =0

fi , yk +M i

to adaptive algorithm, background noise, and limited length of the DFE. Now the coefficients cannot entirely compensate the fading of channel and noise. So we divided the equalizer output of k -time into more components as follows according to (4).

is also a random process with complicated characters, which is difficult for accurate induction. However, it can be simplified here (from view of engineering) as M 1 (7) f y ~ ( , 2 ) ,
2 2 where F = 0 and F M E[ yi2 ] E[ F ,i ] |i[0, M ) and it is

i =0

i ,

k + M i

Ik = +
where = k

M 1 i =0 N

f i yk + M i +
N

M 1 i =0

( f i f i ) yk + M i
, (5)

g I k j + j =1 j
M 1

(g j g j ) I k j j =1
N

= k + k
i =0

also an approximate normalized process. Here, the channel background noise effects (variants) are included in the received signal components yi . When it comes to the FBE, similar results are that N g I k j ~ ( , 2 ) ,
j =1 j , B B

fi yk + M i +
N j =1

g I k j and j =1 j
N j =1

(8)

k =
=

M 1 i =0 M 1 i =0

( f i f i ) yk + M i + f i , yk + M i +

(g j g j ) I k j

where B = 0 , variant of g j , .

2 2 2 B N E[ I j ] E[ B , j ] | j[1, N ] and B , j is

g j , I k j

In them, f i and g j are the ideal (or theoretic, if exist) solutions for specific multi-path channel. If possible to reach, this ideal coefficients combination will reward us with best performance. However, we can only approach it with help of equalization algorithms in practice. With the ideal solution f i and g j , distortion.

Now back to (5), in the equalizer output I k

= k + k ,

k can be seen as a random transmitted variable for each

sample k . And

can also be seen as a random noise variable,

2 (9) k ~ ( k , dfe ) , 2 2 2 2 dfe satisfy: k 0 , dfe F + B
2

which approximately satisfy where

k is equal to the transmitted signal I k without

and

is the different component between the output

of DFE and transmitted symbol. We call it the filter coefficients noise components, which are caused by equalization updating algorithm or background noise. The feature of

2 2 M E[ yi2 ] E[ F ,i ]|i[0,M ) +N E[I j ] E[ B, j ]| j[1,N ] .

Due to (5) and (9), to most commonly used modulation methods, Symbol Error Ratio (SER) of equalizer output I k satisfies (10) P Q ( 2 ) ,
e

is analyzed as follows.

includes the sum of both the FFF coefficients different

components and the FBF ones. It is quite similar in discussion them. So we only refer to the FFF case in the next part of discussions with some aggressive assumptions, which is helpful for analysis. Now there are three hypotheses that can be used in sparse channel condition as following. 1) First, f i , is regarded as a random process that satisfies

where

k = E[ k ] ; ,
2

dfe

and are constants that is

decided by the modulation way. , and are all larger than zero, which implies that Pe has same trend
2 with ( k dfe ) . From (10), we see that the SER of the DFE is infected

( F ,i ,

2 F ,i

normal distribution with parameters of ) . Under most conditions that the equalizer is

greatly by the distribution parameter

2 dfe

of the filter

452

IEEE Transactions on Consumer Electronics, Vol. 51, No. 2, MAY 2005

coefficients noise components

. And

2 dfe

is direct

proportion with FFF coefficients M and FBE coefficients N . In discussion of section II, long sparse channel with strong pre-echoes and large delay post-echoes requires the DFE to be equipped with large number of taps. For example, in HDTV system, M and N is up to several hundreds, which lead to 2 that noise component dfe will be large (because M and N are large) according to (10) and the SER will be high. Now the conventional DFE deteriorates or do not work. It should be noted that inductions above are not strict ones and some assumptions are even not quite right under some conditions. However, it is these assumptions that reward us with simplification in analysis and reasonable qualitative (instead of quantitative) conclusions that we need.
IV.

and inactive ones with noise like small coefficients [7]. Only the active tap coefficients are included in output calculation (11) Ik = f y + g I k j ,
i

where and are the active sets for FFF and FBF taps with numbers of elements M 1 and N1 , respectively. Under sparse channel condition, M 1 and N1 can be much smaller than M and N . To a specific tap

k + M i

f i ( g j ), the differential component f i , f i ( g j ) is a

( g j , ) between ideal solution f i ( g j ) and

random variable whose module changes frequently. If f i ( g j ) belongs to the active group, there must be time that it becomes smaller than some of that in the inactive group. Then it is abandoned to the inactive group with flush (noise removal). Abandoned tap (from active to inactive group) is flushed to zero (for each of its components) and is clear for next circles updating. This flushing processing can be seen as a dynamic noise removal procedure and is an inseparable part of algorithm. Thus, the noise coefficients in the active group can periodically cleared and smaller number of active ones take part in the equalizer output calculation ( M1 < M and

THE PROPOSED DFE

A. The new DFE structure As discussed above in section III, the equalizer performance (mainly reflected by SER) has tight relationship with the distribution parameters of

k .

To reduce the SER, we have

2 2 two options: 1) Reduce the level of each F ,i and B , j , which

is decided by adaptive equalization algorithm mainly; 2) Reduce the number of FFF and FBF taps to reduce M and N (however, these parameters decide the equalizer capability, which can not be changed easily. In generally, the FFF and FBF tap length is greater than or equal to the channel length.). The first problem can be solved by appropriate adaptive algorithm, which will be described in subsection B. As for reduce to number of filter taps ( M , N ), we can achieve it by group division based on system requirement and channel condition. In Fig.4, the proposed DFE architecture is present.
yk + M
f0

N1 < N ), which reduces the effects of noise accumulation

greatly.
B. The blind adaptive algorithm used in DFE
2 2 For reducing the level of each F ,i and B , j , appropriate

D
f1

D
D out put FE

f M 1

Ik
Taps gr oup di vi si on Adapt i ve al gor i t hm and coef f i ci ent s updat i ng cont r ol

Ik

gN

g N 1

g1

Fig. 4 The proposed DFE architecture

adaptive algorithm must by applied. There are many adaptive algorithms that can be employed to adjust the filter taps. Theoretic performance bounds of some algorithms are discussed in the contributions as [8,9]. Difference between these algorithms mainly lies in the cost function they used, which ultimately decides the error item used to adjust the equalizer taps to convergence. Selection of adaptive algorithm depends on factors like channel condition, data frame formats, coding mode, system requirements and etc. Actually, the constant modulus algorithm (CMA) and the least mean square (LMS) algorithm are the most popular ones and are employed here in our discussion. 1) CMA algorithm If there is no training sequence, the CMA algorithm is a good blind algorithm, which is always used in initialization of the equalization, especially in the acquiring phase. The error items of CMA is described as

In convention DFE, number of filter taps ( M , N ) is decided by channel length, but the feature of channel is not be considered, such as sparseness. Actual number of coefficients employed in I k calculation can be reduced by group division for taps, which brings us two gifts: improving SER performance and reduce complexity. Considering the sparse property of the multi-path channel in high-speed system, in receiver the DFE filter taps are divided into two groups: active ones with relative large coefficients

ek = I k ( I k ) ,
4 2

(12)

where is the Godards constant that is decided by

= E[ I k ] E[ I k ] .
2) LMS algorithm The LMS algorithm always has good convergence and is with high adaptability, but it need rich training information. After updating the tap coefficients of DFE by the CMA

L. Fan et al.: Efficient Robust Adaptive Decision Feedback Equalizer for Large Delay Sparse Channel

453

algorithm, the symbol error ratio of slicer output is low enough that we can use slicer output to substitute the training symbols. It is called DD-LMS mode. The error items of LMS is described as

ek = I k I k .
f k +1 = fk + 1 ek d y* d (y * d yT d ) k k k
* k d * T k d k d

(13)

active/inactive flags. To reduce the complexity of module comparison, the taps can be divided sub-groups and only comparison of one sub-group is performed for each updating circle. If no exchange occurred for one circle, the comparison results will be added to the next circles comparison progress until exchange occurs.
c) Cleaning the abandoned tap coefficients

With both CMA and LMS algorithms, FFF and the FBF are updated by the error item as following, respectively. , (14)

g k +1 = g k + 2 ek d I (I I ) where f k and g k represent vector of FFF tap and FBF tap,

respectively. d is the symbol delay of the updating processing that is set for hardware implementation convenience. 1 and

2 are the step-sizes adopted for the updating procedure.

They may adopt various values within different taps region with considering both convergence speed and convergence precision. Scalars (y k d y k d ) and (I k d I k d ) are two normalization items, which entitles the adaptive algorithm with finite feasibility in the equivalent updating step-sizes under different signal level according to some contributions.
C. The realization of the proposed DFE We know that reducing the number of filter taps ( M , N ) resists noise; besides, it can reduce the complexity in hardware implementation for using less multiplier. Number of expensive multipliers needed in (11) is obviously much smaller than that in (4). This is quite important when equalizer scale is large. Why not enjoy the free lunch with only definite cost of additive control module that is quite cheap compared to multipliers. Now the problem lies in designing a smart robust mechanism for active taps selection. 1) The selection of active taps Based on analysis above, the proposed DFE can be realization by the following steps under large delay sparse channel.
a) Initialization
* T
* T

Abandoned tap (from active to inactive group) is flushed to zero (for each of its components) and is clear for next circles updating. This flushing processing can be seen as a dynamic noise removal procedure and is an inseparable part of algorithm. Being benefit form the cleaning process, the noise in DFE taps can be resisted. 2) The updating mechanism of the filter coefficients Due to (14), the updating of the filter coefficients needs also large calculation every symbol. There are M and N taps to be updated in FFF and FBE, respectively. However, not all filter coefficients must be updated, special in good channel condition. So the filter coefficients are divided into U groups. Every symbol, only one or two groups are updated to track the channel. How to divide the groups for taps and how to update are decided by the channel condition. For example, if there are pre-estimation of channel feature, the filter coefficients responding to main path are updated per symbol time, but the other coefficients are updated per several symbol time.
D. Performance analysis of the proposed algorithm 1) Noise removal ability If a specific tap is with small ideal solution f i ( g j ), the

differential component f i , ( g j , ) is a random variable whose module changes frequently. If f i ( g j ) belongs to the active group, there must be time that it becomes smaller than some of that in the inactive group. Then it is abandoned to the inactive group with flush (noise removal). When vestigial noise occurs (for dynamic channel) in adaptive updating, similar is the case. Thus, the noise coefficients in the active group can periodically cleared and smaller number of active ones take part in the equalizer output calculation ( M1 < M and

Due to experiment results, about 15%~30% of total filter taps are set to active sets randomly. This percent selection is based on the scale of the equalizer and channel condition. Larger number of filter taps and sparser channel admit lower percentage.
b) Adaptive selection of taps

N1 < N ), which reduces the effects of noise accumulation

greatly. 2) Complexity reduction In the design and realization of the DFE, the main calculation burden focuses on the updating of the filter coefficients and calculation of the filter output. So we can reduce the complexity of the DFE by two methods discussed above. Number of expensive multipliers needed in (11) is obviously much smaller than that in (4). This is quite important when equalizer scale is large. Why not enjoy the free lunch with only definite cost of additive control module that is quite cheap compared to multipliers. In Table I, the number of complex multiplication is compared, and U > 2 .

In the process of receiving the signal, taps need exchange between active and inactive groups to track the variety of channel. During coefficients updating, some taps in the active group will be smaller (in module) than some in the inactive one. So we get the smallest and largest tap in the active and inactive groups, respectively. If the former is smaller than the later, the smaller active tap will be substituted by the larger inactive tap and abandoned to the inactive group. This exchange is easily performed by setting the corresponding

454 TABLE I COMPARISON OF THE NUMBER OF COMPLEX OR REAL MULTIPLICATION Conventional DFE The proposed DFE

IEEE Transactions on Consumer Electronics, Vol. 51, No. 2, MAY 2005

2M + 2 N
M 1 + N1 + 2( M + N ) / U

Log10(Symbol Error Ratio)

3) Robustness The in-group comparison, exchanging and flushing processes operate during all DFE working phase. When the channel is fast fading (with dynamic echoes) or the background noise is severe, exchanging and flushing operation occurs frequently during each circle of comparison, which can adapt to the change of channel condition. However, it rarely occurs if the channel is static and is with high signal-to-noise ratio (SNR), which entitles the algorithm with high robustness and stability.

the proposed DFE has steady performance. But in single static channel, the conventional DFE can work with unsteady performance, when Doppler rate is up to 100Hz, it cannot work. That is, noise accumulation is so large in fast timevarying environment. At the same time, the reduced DFE can work well for filter coefficients noise is resisted by reducing the number of taps coefficients and flushing processing.
0

-1

-2

V. SIMULATION RESULTS In this section, the proposed DFE has been tested for different sparse channels, including single echo channel and ATSC R2.1 channel. Some simulation results are described below. The simulated experiments consider ATSC system based on 8-VSB modulation, 10.76M-Hz symbol-rate on 6MHz bandwidth. Due to the channel delay, the conventional DFE has M = 256, N = 512 taps, while the active taps of the proposed DFE is set as M1 = 64 and N1 =112 for FFF and FBF, respectively.
A. Single echo channel In the subsection, the performance of the proposed DFE under AWGN and single echo multi-path condition is shown. And the single path is static or dynamic echo with following parameters in Table II.
TABLE II COMPARISON OF THE NUMBER OF COMPLEX MULTIPLICATION Single static echo Attenuation (dB) Delay (us) Doppler rate (Hz) 3 20 0.05 Single dynamic echo 8 3 100

-3 AWGN Theoritic AWGN Conv-DFE AWGN New-DFE Static echo Conv-DFE Static echo New-DFE Dynamic echo Conv-DFE Dynamic echo New-DFE

-4

-5 14

16

18

20

22

24

26

28

SNR (in dB)

Fig.5 SER performance comparison

15

10
Dynamic echo Conv-DFE

5 Mean Square Error(dB)

Static echo Conv-DFE

-5
Dynamic echo New-DFE

-10
Static echo New-DFE

-15

200

400

600 800 1000 Number of symbol( X 828)

1200

1400

1600

Fig.5 shows the SER performance comparison for conventional full size DFE and the reduced DEF under different signal and noise ratio (SNR). This figure demonstrates that the new structure can improve the system performance of about 1-2dB under both AWGN and single static echo multi-path channel. When the channel is with dynamic echo of 100Hz Doppler rate, the conventional DFE breakdowns, but the proposed DFE has good performance all the time. In Fig.6, initial convergence performance is compared between the conventional DFE and the new DFE under single static and dynamic echo conditions. It can be shown that the conventional DFE deteriorates in the time-varying environment of both 0.05Hz and 100Hz Doppler rate, while

Fig.6 Initial convergence MSE curves

The tap coefficients of the FFF and FBF of the conventional DFE and the proposed DFE are shown under single static and dynamic channel in Fig.7 (a) (b) and Fig.8 (a) (b), respectively, when the DFE are steadily working. From them, we observe that besides the dominant taps corresponding to each multipath, there are many nonzero taps coefficients that should be zero and be negligible in FFE and FBE when full size DFE is used. Based on analysis above, they are introduced by adaptive algorithm and background noise. These nonzero taps coefficients influence the performance of the DFE as noise. However, the reduced DFE has lower noise taps coefficients due to its structure and cleaning mechanism.

L. Fan et al.: Efficient Robust Adaptive Decision Feedback Equalizer for Large Delay Sparse Channel
1 0.5 Tap coef. 0 -0.5 -1 Tap coef. 1 0.5 0 -0.5 -1

455

FFF-real FFF-imaginary

FFF-real FFF-imaginary 50 100 150 Tap Index of FFF 200 250

50

100

150

200

250

Tap Index of FFF 0.6 0.4 Tap coef. 0.2 0 -0.2 -0.1 -0.4 FBF 0 50 100 150 200 250 300 350 400 450 500 -0.2 0 50 100 150 Tap coef. 0.1 0 FBF 0.2

200

250

300

350

400

450

500

Tap Index of FBF

Tap Index of FBF

(a) Conventional DFE coefficients

1 FFF-real FFF-imaginary

(b) The proposed DFE coefficients Fig.8 Under single dynamic echo

0.5

-0.5

B. ATSC channel Similarly, we investigate the proposed DFE under ATSC channel, and give simulation results under R2.1 3#. The channel is described as followers in Table III.
100 150 200 250

Tap coef.

-1

50

TABLE III ATSC R2.1 3# ENSEMBLES Path 1 Path 2 14 -1.8 125 Path 3 14 0.15 80 Path 4 4 1.8 45 Path 5 8 5.7 5Hz Path 6 12 35 90

Tap Index of FFF 0.6 0.4 Tap coef. 0.2 0 -0.2 -0.4 0 50 100 150 200 250 300 350 400 450 500

FBF

Attenuation (dB) Delay (us) Phase or Doppler

0 0 0

Tap Index of FBF

(b)The proposed DFE coefficients Fig.7 Under single static echo

1 0.5 Tap coef. 0 -0.5 -1

Under the ATSC R2.1 3# channel condition, the SER performance comparison of the conventional DFE and the proposed DFE are shown in Fig.9. Fig.10 shows initial convergence performance of the conventional DFE and the new DFE under ATSC channel condition. We see that our approach provide faster and steadier initial convergence than the conventional DFE. Due to noise accumulation in the full size DFE, MSE will periodically jump 0dB.
0

FFF-real FFF-imaginary 50 100 150 Tap Index of FFF 200 250

-1
Log10(Symbol Error Ratio)

0.2 0.1 Tap coef. 0 -0.1 FBF -0.2 0 50 100 150 200 250 300 350 400 450 500

-2

-3

-4

Tap Index of FBF

(a)The conventional DFE coefficients

-5 14

AWGN Theoritic AWGN Conv-DFE AWGN New-DFE ATSC R2.1 3# Conv-DFE ATSC R2.1 3# New-DFE

16

18

20

22

24

26

28

SNR (in dB)

Fig.9 Initial convergence MSE curves

456
15

IEEE Transactions on Consumer Electronics, Vol. 51, No. 2, MAY 2005

VI. CONCLUSION
ATSC R2.1 3# Conv-DFE ATSC R2.1 3# New-DFE

10

Mean square Error(dB)

-5

-10

In this paper, a DFE algorithm is presented, in which a group division based noise removal mechanism is carried out with considering long delay sparse multi-path character. Both performance and complexity of the DFE are improved at the same time. Analysis and simulation results show that this new DFE can both remove (at least relieve) the noise accumulation effects in large-scale equalizers (needed in high-speed wireless communication systems) and reduce the implementation complexity effectively. Besides, it has better SER performance and initial convergence than the conventional DFE. Benefit from its superiority to conventional DFE, this new structure DFE can be used with good applications in systems like HDTV and wideband mobile communication system.
3000

-15

500

1000

1500

2000

2500

Number of symbols(X828)

Fig.10 Initial convergence MSE curves

REFERENCES
J.G.Proakis, Digital Communications, 3rd ed. New York: McGraw Hill 1995 [2] T.S.Rappaport, Wireless Communications: Principles and Practice. Englewood Cliffs, NJ: Prentice-Hall, 1996 [3] ATSC Recommended Practice: Receiver Performance Guidelines (A/74, Jun. 2004) and ATSC digital television standard, Revision B with Amendment 1 (A/53B, Sept. 2001). To: www.atsc.org [4] Advanced Television Technology Centre (ATTC): Evaluation of ATSC 8-VSB Receiver Performance in the Presence of Simulated Multipath and Noise (doc. #99-04A, Sept. 1999). To: www.attc.org [5] I.J. Fevrier, S.B. Gelfand and M.P. Fitz, Reduced complexity decision feedback equalization for multipath channels with large delay spreads, IEEE Trans. on Commu, Vol.47, Is. 6, pp. 927 937, Jun. 1999. [6] Rontogiannis, A.A.; Berberidis, K., Efficient decision feedback equalization for sparse wireless channels, IEEE Trans. on Wireless Communications, vol. 2, Is. 3, pp. 570-581, May 2003. [7] M.R.Asharif, K.Murano and M.Hatori, TV ghost canceling by LMSRAT digital filter, IEEE Trans. on Consumer Electronics, vol.27, No.4, pp.588-604, Nov.1981 [8] M.Ghosh, Blind decision feedback equalization for terrestrial television receivers, Proc.IEEE, vol.86, pp.2070-2081, Oct.1998. [9] Wu Wen-Rong and Tsuie Yih-Ming, An LMS-based decision feedback equalizer for IS-136 receivers, IEEE Trans. on Vehicular Technology, vol. 51, Is. 1, pp.130 143, Jan. 2002. [10] Dongjian Wang, Lingge Jiang, Chen He and Lingyan Fan, On design of noise resistant complexity reduced decision feedback equalizer for large delay sparse eches, ICC 2005, submitted for publication. [1] Lingyan Fan was born in Gansu, China, in 1979. She is pursuing Ph.D. degree in department of Electronics engineering, Shanghai Jiao Tong University. And her current research interests are intelligent information processing and communication signal processing. Chen He obtained the B.E. degree and the M.E. degree from South-east University of China in 1982 and 1985, respectively. In 1994, he received the Ph.D. degree from Tokushima University of Japan in electronics system. Now, he is a professor of Shanghai Jiao Tong University of China. His current research interests are wireless communication systems and intelligent information processing. Dongjian Wang was born in Jiangsu, China, in 1979. He received the B.S. degree from Nanjing University of Post and Telecommunication, and M.E. from Shanghai Jiao Tong University. He is currently DSP engineer of PixelWorks. China. His research interests include communication signal processing and image communication. Lingge Jiang received the B.E. degree in Radio Engineering from South-east University of China in 1982, and the M.E. and Ph.D. degrees in Electrical Engineering from Tokushima University of Japan, in 1993 and 1996, respectively. Now, she is a professor of Shanghai Jiao Tong University of China. Her current research interests are wireless communication systems, intelligent information processing and chaotic theories and its applications in communication.

Fig.11 (a) and (b) show the tap coefficients of the FFF and FBF of the conventional DFE and the proposed DFE respectively.
1 0.5 Tap coef. 0 -0.5 -1 FFF-real FFF-imaginary

50

100

150

200

250

Tap Index of FFF 0.4 0.2 Tap coef. 0 -0.2 -0.4 0 FBF

50

100

150

200

250

300

350

400

450

500

Tap Index of FBF

(a) Conventional DFE coefficients

1 0.5 Tap coef. 0 -0.5 -1 FFF-real FFF-imaginary

50

100 150 Tap Index of FFF

200

250

0.4 0.2 Tap coef. 0 -0.2 -0.4 0

FBF

50

100

150

200

250

300

350

400

450

500

Tap Index of FBF

(b) The proposed DFE coefficients Fig.11 Under ATSC R2.1 #3 channel