Adaptive Blind Channel Equalisation

Adaptive blind channel equalisation in chaotic communications by using nonlinear prediction technique
B.-Y. Wang and W.X. Zheng Abstract: In chaotic communications, an ideal channel is often assumed. In practice, channel distortion is inevitable. In particular, in wireless chaotic communications, the channel distortion may be serious and must be compensated. An adaptive blind equalisation algorithm is proposed. The aim of the algorithm is to recover the chaotic signal transmitted through a nite impulse response (FIR) channel. The inherent characteristic of the chaotic signal, that is the high sensitivity to initial conditions, is exploited to formulate the criterion used in deriving the algorithm. The analysis of stability of the proposed algorithm is also provided.
Introduction
A chaotic signal is generated by a deterministic dynamic system, but shows some randomness in its behaviour. Its spectrum has a continuous, broadband nature. In addition, most chaotic signals exhibit highly sensitive dependence on initial conditions. Given two distinct initial conditions arbitrarily close to one another, the trajectories of the chaotic signal starting from these initial conditions diverge and nally become uncorrelated [1]. In many chaotic communication schemes, it is assumed that the transmitter is connected to the receiver by an ideal channel [2]. However, in a realistic communication system, the distortion due to the propagation channel may corrupt the transmitted signal and make it difcult for the receiver to directly recover the transmitted signal. In such situations, channel equalisation is required to compensate the channel distortion [3 6]. In the work of Sharma and Ott [6], a synchronisationbased approach was proposed to compensate for the channel distortions. However, the fact that the conditional Lyapunov exponents of the receiver depend on both the chaotic system and the channel causes some difculties in determining the coupling parameters to reach an approximate synchronisation. Recently, Zhu and Leung [4] proposed an identication approach called minimum nonlinear prediction error (MNPE). The estimation accuracy of the MNPE approach is limited by the truncation error in approximating the innite-order autoregressive (AR) model by a nite-order AR system. In addition to the above batch algorithms, an extended Kalman lter (EKF)-based adaptive equalisation algorithm has been presented [5]. However, in Leung et al. [7], it was observed that
# The Institution of Engineering and Technology 2006 IEE Proceedings online no. 20060009 doi:10.1049/ip-vis:20060009 Paper rst received 17th January and in revised form 17th April 2006 B.-Y. Wang is with the College of Communications and Information Engineering, Nanjing University of Posts and Telecommunications, Nanjing 210003, Peoples Republic of China W.X. Zheng is with the School of Computing and Mathematics, University of Western Sydney, Penrith South, DC, NSW 1797, Australia E-mail: w.zheng@uws.edu.au
the stability of the latter algorithm cannot always be guaranteed. In this paper, an adaptive algorithm is proposed to blindly equalise the channel distortion in chaotic communications. The proposed algorithm is developed by exploiting the inherent characteristic of the chaotic signal, that is the orbits of a chaotic system are very sensitive to initial conditions. The stability of the proposed algorithm is analysed as well. 2 Problem statement
Consider a chaotic communication system, in which the information-bearing chaotic signal is transmitted through a non-ideal channel before it reaches the receiver. Suppose that the chaotic transmitter system is described by xn f xn1 ; . . . ; xnd ; xn [ R m 1
The signal of transmission st is assumed to be stored in the bifurcation parameter of the chaotic system (1). By keeping the bifurcation parameter in the chaotic region, the output of the chaotic system is chaotic and hence occupies a wide bandwidth. In practice, an ideal channel does not exist. The transmitted signal passes through a propagation channel before arriving at the receiver. It was found in the work of Iltis and Fuxjaeger [8] that a typical fading and multipath channel could be modelled by a nite impulse response (FIR) lter. Thus, the received signal would be yn
L1 X l0
al xnl vn
where a [1 a1 aL21]T is the channel coefcient vector 2 and vn is zero-mean white Gaussian noise with variance sn . The objective of blind identication/equalisation is to estimate the channel parameters ak and recover the transmitted signal xn from the received signal yn . In many blind algorithms that deal with non-chaotic signals, a priori information about the statistics of the input signal is exploited. However, instead of the statistics of the chaotic signal, we will exploit the knowledge about the nonlinearity of the chaotic system. That is, the algorithm
IEE Proc.-Vis. Image Signal Process., Vol. 153, No. 6, December 2006
810
to be developed assumes the knowledge of the chaotic map and its parameters. 3 3.1 Adaptive blind identication of FIR systems Related work
3.1.1 EKF-based approach [5]: In order to apply EKF, the equalisation problem has to be reformulated as a nonlinear state-space estimation problem. The channel coefcients will be taken as state variables and modelled as zero-order AR systems. This will lead to the following state equation and measurement equation X n F X n1 wn1 yn where h n X n xn xn1 xnL1 a 1 h iT n1 n1 wn1 01L1 w1 wL 1 h F X n1 f xn1 xn1 xnL2
n1 a1 n a L1 L1 X l0 n a l xnl
where the estimates of the coefcients are updated by using the least mean squares (LMS) algorithm @S a n n1 ^ a ^ a m @a aa ^ n1 !2 6 N N L1 X X X 2 ^ n j S a en a yn aj x
nd n d j0
The coupling vector K should be selected to ensure the synchronisation between the transmitter and the receiver. But a suitable choice would require the knowledge of the unknown channel. Moreover, the slower convergence rate of the synchronisation method is also an issue. 3.2 Proposed approach
3 vn
iT
As the driving signal considered in this paper is not Gaussian, the least square criterion cannot produce an optimal estimator. As is well known, the initialisationsensitivity peculiar to a chaotic signal could be used to design an effective criterion for identifying a linear system. For the accurate value of the input sequence xt , the following equation xt f xt1 ; . . . ; xtd 0 7
n1 aL 1
iT
holds for any instant t. Hence, we dene the nonlinear prediction error (NPE) criterion as follows. Denition: The NPE criterion is dened as ^ t a f x ^ t1 a; . . . ; x ^ td a2 Vt a e2 t a x 8
The EKF algorithm [5] for the system in (3) is summarised as follows ^ n1 ^ njn1 F X X Pnjn1 Fn1 Pn1 FT n1 Qn T 2 1 K n P njn1 GT n1 Gn1 P njn1 Gn1 sn ^ njn1 K n yn H X ^ njn1 ^n X X P n I 2L1 K n Gn1 P njn1 where
Fn1 @F =@X jX X ^ n1 Gn1 @H =@X jX X ^ n1

Qn Efwn wT ng P 21 (n) and H(Xn) L l0 al xn2l is the transfer function of the channel. In applying EKF, the instability problem must be checked, because the condition for the stability cannot be satised in some scenarios [7]. In the work of Zhu and Leung [5], a thorough analysis of the dynamics of EKF was conducted. 3.1.2 Synchronisation method [6]: The coupling synchronisation method receiver of Pecora and Carrol [6] is expressed as ! L1 X n ^ n f x ^ n1 K yn ^j x ^ nj x 5 a
j0
^ t(a) is used to show the dependence of the estimated where x ^ t on the vector a. signal x A vital problem in seeking the optimal solution of (8) is ^ t21 , . . . , x ^ t2d as functions of the known how to represent x quantities. It has a very great effect on the performance of the derived algorithm. In the work of Zhu and Leung [4], ^ t2d are expressed as functions of the obser^t21 , . . . , x x vations. The main idea is that a moving average (MA) system is approximated by an AR system of nite order. Evidently, the truncation error is unavoidable. The simulation results therein indicated that the performance of their approach levels-off and saturate at no more than 40 dB. Although that algorithm achieved a very high accuracy in AR cases, the behaviour of the algorithm in identifying MA systems is not satisfactory. To avoid the performance saturation of the NPE approach, we attempt to nd another way to express ^t21 , . . . , x ^ t2d . From (2), it follows that x xt yt
L1 X l1
al xtl
However, only yt is known in the above equation. A technique used in the recursive nonlinear least square algorithm [9] is very helpful here. During the progress of a recursive algorithm, the estimate of xt can be formed as
L T t1 ^ ut yt utt ; 1 a L T t1 ^ ^ tt ^ t ut x ; x 1 a L T utt 1 ut1 ; . . . ; utL L ^ tt ^ t 1 ; . . . ; x ^ tL T x 1 x
10
where ut s are intermediate variables. More simply, we can replace (10) with
L T t1 ^ ^ tt ^ t yt x x ; 1 a L ^ tt ^ t1 ; . . . ; x ^ tL T x 1 x
11
811
Table 1: Proposed approach for identifying MA systems

^ Initialisation: a
(0)
^ 2L . . . x ^ 21 0; 0L1; x
For t 1, 2, . . . , do
2L T (t21) 2L ^ t yt 2 ( x ^ tt2 ^t ^ t21 , . . . , x ^ t2L]T x ,x 1) a t21 [x
Remark 1: As a gradient-type algorithm, its convergence will depend on the initial value. The proposed algorithm can be initialised with a least square solution based on a very short data record in a blind but batch fashion. 4 Simulation results
^ (t ) a ^ (t21)2 m(@Vt/@a)jaa a ^ (t21) where m is the step size
The derivative @Vt(a)/@a is @Vt a ^ t a f x ^ t1 a; . . . ; x ^ td a 2x @a & ' @ tL ^ a; . . . ; x ^ td a xt1 f x @a t1
12
In this section, simulation results are presented to evaluate the efcacy of the proposed algorithm. In initialisation of the proposed algorithm, the rst block with ten samples is used to produce an initial estimate for the relevant parameters. Moreover, at the beginning of the algorithm implementation, the step size in the adaptive algorithm is set at a small positive number. As the iteration increases, the step size will be changed into another much smaller positive number so as to ensure the best steady-state performance. Example 1: Consider a chaotic communication system, in which the chaotic transmitter is the logistic map xt lxt1 1 xt1 where the bifurcation parameter l is controlled in the chaotic region [3.7 4.0]. The signal of transmission xt is assumed to pass through a third-order FIR channel a [1, 0.45, 2 0.22]. The mean-squared error (MSE) between the true channel coefcients and the estimated channel parameters is exploited to measure the identication performance. Fig. 1 shows the steady-state MSE of the proposed approach under various signal-to-noise ratios (SNRs). For four different SNRs, the algorithm can offer a very accurate data recovery with 200 samples. Fig. 2 provides a comparison between the proposed approach and the synchronisation-based approach [6]. In a large range of SNRs, the proposed algorithm can achieve a 1 2.5 dB improvement in the equalisation error. As the channel order is often unknown, it is necessary to investigate the robustness of the algorithm to order overestimation. In Fig. 3, the algorithm is evaluated when the channel order is overestimated. The observed degradation is less than 3 dB after 300 samples. As the bifurcation l in the chaotic dynamics may be unknown, it is interesting to on the investigate the impact of the mismatch L l 2 l performance. We show the simulation results for the logistic and Chebyshev driving sequences in Fig. 4a and b,
Considering the case where the driving chaotic signal is generated by xt f xt1 Equation (12) can be expanded as & ' @Vt a @ L ^ tt ^ t a f x ^ ^ t1 a x 2 x f x a 1 t1 @a @a & ' @ L ^ tt ^ ^ t1 a x ^ t a f x f x a 13 2x 1 t1 @a With the derived gradient, it is straightforward to implement an LMS-like algorithm. The proposed algorithm is summarised in Table 1. This algorithm requires about 4L multiplication operations at each iteration, so its computational complexity is O(L). Note that the EKF-based algorithm has a computational complexity of O(L 2) [5]. To analyse the stability of the proposed algorithm, as done in Mandic [10] and Mandic and Krcmar [11], we dene an a posteriori error
ap ^ t f x ^ t a ^ t1 a e x t
14
Using (11), we have

L T t ^ x ^ t yt x ^ tt ^ t1 ^ t a ^ t a x 1 a L T t ^ a ^ t1 ^ tt x 1 a
15
After substitution of (13) and (15) into (14), we have

ap L T tL L1 ^ tt ^ t1 f 0 x ^ tt ^ t 1 x je t j jet met x 1 x 2 j
16
Therefore the step-size m should satisfy

L T tL L1 ^ tt ^ t1 f 0 x ^ tt ^ t 1 x j1 mx 1 x 2 j , 1
17
so that je(ap) t j , jetj. Equation (17) imposes a constraint on the step size. One may be concerned about if the second term in the left side of (17) is positive. In order to explain this point, we rewrite that term as
L T t L L T tL1 ^ tt ^ tt ^ t1 f 0 x ^ t2 ^ t1 x x 1 x 1 x
18
The expression in (18) can be approximated by ^ t1 r1 r0 f 0 x 19

Fig. 1 Error in the equalised data l 4.0
Four curves corresponding to four different SNRs, 20, 30, 40, and 50 dB, respectively, with 200 data samples
where r(0) and r(1) denote the correlation of the estimated chaotic sequence. Generally, r(1) is much smaller than r(0). Therefore it can be reasonably expected that the term in (19) is positive and thus the condition (17) is non-trivial.
812
-3 0 our approach -3 5 E r r o r i n E q u a l i z e d D a ta ( d B ) -4 0 -4 5 -5 0 -5 5 -6 0 -6 5 -7 0 -7 5 30 the approach in [6]
0
=0 =0.2 =0.4
-10
MSE in dB
-20
-30
-40
-50
-60
50
100 150 200 250 300 Number of data samples a
350
400
40
50 SNR (dB)
60
70
0 =1e-4 =0 -10
Fig. 2 Error in the equalised data, l 4.0, 256 data samples

Solid line indicates our approach; dashed line indicates the approach in the work of Sharma and Ott [6]
respectively, where the Chebyshev map is given by xt cosl cos1 xt1 In the case of the logistic driving sequence, the algorithm is very robust to the mismatch. An interesting observation is that the proposed algorithm can still work very well even when the mismatch is rather large. In contrast to the above observation, the algorithm is very sensitive to the mismatch in the bifurcation parameter of the Chebyshev map, as a very small mismatch leads to a greater performance degradation. The main reason for the difference in these two chaotic sequences may be that their dynamics have different sensitivities to the bifurcation parameter. In other words, the Lyapunov exponent of the Chebyshev map is more sensitive to the bifurcation parameter. Example 2: In the second example, we evaluate the performance of the algorithm when some different chaotic sequences are transmitted in the chaotic communication system. Fig. 5 shows the performance of the algorithm when the transmitted signals are generated by the logistic map with four different l. In Fig. 6, we show the performance of the algorithm when the transmitter sends the logistic
0 L L+1 L+3
MSE in dB
-20
-30
-40
-50
-60
100
200
300
400
500
600
700
800
900 1000
Number of data samples b
Fig. 4 Performance of the algorithm when there is a mismatch in the bifurcation parameter
sequence and Chebyshev sequence. In all cases, the stepsize is appropriately selected to guarantee a good steady-state performance. Clearly, the performance of the algorithm relies mainly upon the SNR, rather than upon which type of driving chaotic sequence is transmitted. Example 3: In this last example, we investigate the performance of the algorithm under two time-varying
0
=3.78 =3.86 =3.94 =4.0
-10
-10
-20 MSE in dB
-20 MSE in dB
0 100 200 300 400 500 600
-30
-30
-40
-40
-50
-50
-60
-60
50
Number of data samples
100 150 200 Number of data samples
250
300
Fig. 3 MSE with the overestimated channel order, SNR 50 dB, l 4.0
Fig. 5 MSE for different values of parameter l of the logistic map, SNR 50 dB
813
0 Chebyshev sequence Logistic sequence -10
and Channel B: at 1 0:45 0:01 sinp t=400
0:22 0:02 cospt1=3 =300T Fig. 7a shows the performance of the algorithm when the logistic chaotic sequence is transmitted through Channel A. Fig. 7b corresponds to the case for Channel B. In both cases, the proposed algorithm can offer similar performance to that in the time-invariant channel. 5 Conclusions
-20 MSE in dB
-30
-40
-50
-60
100
200
300 400 500 600 700 Number of data samples
800
900
1000
Fig. 6 MSE for different driving sequences, SNR 50 dB, l 4.0
channels: 8 < 1 0:5 0:22T ; if t , 200 Channel A: at 1 0:45 0:22T ; if 200 t , 400 : 1 0:53 0:19T ; if 400 t
0 SNR=50dB SNR=30dB -10
In this paper, an adaptive algorithm has been proposed to blindly identify and equalise an FIR channel in chaotic communications. The inherent characteristic of a chaotic system, that is the sensitivity to initial conditions, has been exploited to develop our algorithm. With an appropriate step size, the stability of the proposed algorithm can be guaranteed. Compared with the algorithms given in the work of Zhu and Leung [4, 5], the algorithmic simplicity and high estimation accuracy make the proposed approach more attractively applicable in the receiver design of chaotic communication systems. In contrast, however, the algorithms given in Zhu and Leung [4, 5] may be more preferable in the case where the nonlinear function is not differentiable. Finally, an interesting topic for future research is to rigorously show that the expression in (19) is positive and subsequently to determine a constraint for stability of the proposed algorithm on the step size m through upper and lower bound values in terms of the Lyapunov exponent of the chaotic map. 6 Acknowledgment
-20
-30
-40
-50
The authors would like to thank the Editor and three anonymous reviewers for their valuable comments and suggestions that have signicantly improved this paper. This work was supported in part by a research grant from the Australian Research Council and in part by the Research Foundation of the Education Department of Jiangsu Province, Peoples Republic of China (No. 05KJB510089). 7 References
MSE in dB
-60
100
200
600
700
800
a
0 SNR=50dB SNR=30dB -10
-20 MSE in dB
-30
-40
-50
-60
100
200
600
700
800
Fig. 7 Performance of the proposed algorithm in the timevarying channels

a Channel A b Channel B
814
1 Parker, T.S., and Chua, L.O.: Practical numerical algorithms for chaotic systems (Springer-Verlag, New York, 1989) 2 Cummo, K.M., and Oppenheim, A.V.: Circuit implementation of synchronized chaos with application to communications, Phys. Rev. Lett., 1993, 71, pp. 65 68 3 Wang, B.-Y., Chow, T.W.S., and Ng, K.T.: Adaptive identication algorithm for AR system driven by chaotic sequence, Int. J. Bifurcation Chaos, 2003, 13, (4), pp. 963 972 4 Zhu, Z., and Leung, H.: Identication of linear systems driven by chaotic signals using nonlinear prediction, IEEE Trans. Circuits Syst. I, 2002, 49, (2), pp. 170180 5 Zhu, Z., and Leung, H.: Adaptive blind equalization for chaotic communication systems using extended Kalman lter, IEEE Trans. Circuits Syst. I, 2001, 48, (8), pp. 979989 6 Sharma, N., and Ott, E.: Combating channel distortions in communication with chaotic systems, Phys. Lett. A, 1998, 248, (5/6), pp. 347352 7 Leung, H., Zhu, Z., and Ding, Z.: An aperiodic phenomenon of the extended Kalman lter in ltering noisy chaotic signals, IEEE Trans. Signal Process., 2000, 48, (6), pp. 18071810 8 Iltis, R.A., and Fuxjaeger, A.W.: A digital DS spread-spectrum receiver with joint channel and Doppler shift estimation, IEEE Trans. Commun., 1991, 39, (8), pp. 12551267 9 Porat, B.: Digital processing of random signals (Prentice-Hall, Englewood Cliffs, NJ, 1994) 10 Mandic, D.P.: NNGD algorithm for neural adaptive lters, Electron. Lett., 2000, 36, (9), pp. 845846 11 Mandic, D.P., and Krcmar, I.R.: Stability of NNGD algorithm for nonlinear system identication, Electron. Lett., 2001, 37, (2), pp. 200 202
3000
IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 47, NO. 11, NOVEMBER 1999
Adaptive Blind Channel Estimation by Least Squares Smoothing

Qing Zhao and Lang Tong, Member, IEEE
Abstract A least squares smoothing (LSS) approach is presented for the blind estimation of single-input multiple-output (SIMO) nite impulse response systems. By exploiting the isomorphic relation between the input and output subspaces, this geometrical approach identies the channel from a specially formed least squares smoothing error of the channel output. LSS has the nite sample convergence property, i.e., in the absence of noise, the channel is estimated perfectly with only a nite number of data samples. Referred to as the adaptive least squares smoothing (A-LSS) algorithm, the adaptive implementation has a high convergence rate and low computation cost with no matrix operations. A-LSS is order recursive and is implemented in part using a lattice lter. It has the advantage that when the channel order varies, channel estimates can be obtained without structural change of the implementation. For uncorrelated input sequence, the proposed algorithm performs direct deconvolution as a by-product. Index Terms Adaptive least squares method, blind channel identication.
I. INTRODUCTION LIND channel equalization has the potential to increase data throughput for communications over time varying channels. To achieve this goal, several requirements must be satised. First, blind channel estimation and equalization algorithms must converge quickly. An important property is the so-called nite sample convergence property, i.e., the channel can be perfectly estimated using a nite number of samples in the absence of noise. This property is especially critical in packet transmission systems where only a small number of data samples are available for processing. Second, the adaptivity of the algorithm is important in tracking the channel variation and maintaining the communication link. Third, low complexity in both computation and VLSI implementation is desired. Although many blind channel estimation and equalization algorithms have been proposed in recent years, few can simultaneously satisfy requirements in convergence speed, adaptivity, and complexity. Deterministic batch algorithms such as the subspace (SS) algorithm [14], the cross relation (CR) algorithm1 [5], [22], the two-step maximum likelihood
Manuscript received June 2, 1998; revised May 11, 1999. This work was supported in part by the National Science Foundation under Contract CCR-9804019 and by the Ofce of Naval Research under Contract N0001496-1-0895. The associate editor coordinating the review of this paper and approving it for publication was Dr. Phillip A. Regalia. The authors are with the School of Electrical Engineering, Cornell University, Ithaca, NY 14853 USA (e-mail: ltong@ee.cornell.edu). Publisher Item Identier S 1053-587X(99)08320-8. 1 This is also referred to as the least squares algorithm, or EVAM.
(TSML) approach [7], the linear prediction-subspace (LP-SS) algorithm [16], and the joint order detection and channel estimation algorithm (J-LSS) [20] converge quickly. Without assuming a specic stochastic model of the input sequence, these methods have the nite sample convergence property. Unfortunately, they suffer from high computation cost, which is usually associated with eigenvalue decomposition, and their adaptive implementations are often cumbersome. On the other hand, recently proposed linear prediction (LP) based algorithms [1], [4] are attractive for their efcient adaptive implementations. The key component of these algorithms is the classical linear predictor that can be implemented recursively both in time and in lter order using, for example, lattice lters. The modular structure of lattice lters makes them good candidates for VLSI implementation. Perhaps the most important shortcoming of these LP-based algorithms is the relatively low convergence speed. Relying on the statistical uncorrelation of the input sequence, these LP-based algorithms fail to have the nite sample convergence property. Consequently, they demand a relatively large sample size for accurate channel estimation, which limits their application in the small data size scenarios. Aiming to satisfy the three design requirements at the same time, we present in this paper a least squares smoothing (LSS) approach to blind channel estimation. Recognizing the isomorphic relation between the input and output subspaces, we rst consider estimating the channel from the input subspace. By projecting the channel output into a particular input subspace , the channel is obtained from the least squares projection is constructed from the channel output by error. When exploiting the isomorphism between the input and output subspaces, this projection leads to least squares smoothing. This geometrical approach to blind channel estimation also provides a simple and unied derivation of different LP-based channel estimators and a clear explanation for the loss of nite-sample convergence property in LP-based approaches. Based on the LSS approach, a new adaptive least squares smoothing (A-LSS) channel estimation algorithm is proposed. Like all deterministic methods, A-LSS preserves the nite sample convergence property with high convergence rate. Similar to linear prediction-based algorithms, the least squares smoothing approach naturally leads to an order and time adaptive implementation that accommodates a wide range of channel variation, both in channel parameters and channel length. In wireless communications, for example, when the channel order suddenly changes due to the addition or loss of reectors, A-LSS simply selects signals from different parts of
1053587X/99$10.00 1999 IEEE
ZHAO AND TONG: ADAPTIVE BLIND CHANNEL ESTIMATION BY LEAST SQUARES SMOOTHING
3001
is the channel impulse response, and is the input samples of the system input and output, sequence. Given and the block row vector as dene the row vector
(2) With
Fig. 1. Single-input
and
similarly dened, we have, from (1) (3)
P -output linear
system.
the lter with neither structural change nor extra computation. Implemented with commonly used basic modules in classical lattice lters, the structure of A-LSS is highly parallel and regular with only scalar operations. This paper is organized as follows. Section II presents the system model and two assumptions used in this paper. The geometrical approach to linear least squares smoothing channel estimation is introduced in Section III. A batch least squares smoothing (B-LSS) algorithm and its connections with existing linear prediction-based approaches are obtained. In Section IV, we present the data structure for the unknown channel order case followed by the adaptive least squares smoothing channel estimation and order detection algorithm. Simulation results are presented in Section V to demonstrate the convergence and performance of A-LSS in channel estimation and tracking. II. PROBLEM STATEMENT A. Notations and Denitions Notations used in this paper are mostly standard. Uppercase and lowercase bold letters denote matrices and vectors , , and denoting transpose, Hermitian, and with MoorePenrose pseudoinverse operations, respectively. Given ( ) is the row (column) space of a matrix , . For a matrix having the same number of columns ( ) as , into (the orthogonal complement of) is the projection of . We dene as the the row space of into . For a set of vecprojection error matrix of , denotes the linear subspace tors . For a vector and a linear subspace spanned by , is the orthogonal projection of into and , and is its projection error. Finally, denote the 2-norm and Frobenius norm, respectively.
Our goal is to estimate C. Assumptions and Properties
from
Two assumptions are made in this paper. The rst one is about channel diversity. A1: Channel Diversity: The subchannel transfer functions are co-prime. do not share common zeros, i.e., A1 is shared by all deterministic blind channel estimation methods. The co-primeness of the subchannel transfer functions ensures that the channel is fully specied (up to a scaling factor) by the noncommon zeros of the channel output. If A1 is not satised, although the noncommon zeros of the subchannel transfer functions can still be identied from the output, the common zeros cannot be distinguished from zeros of the input. Therefore, the channel cannot be identied without knowing the input sequence. The following property, which has also been exploited in [13] and [21], reveals the equivalence between the input and output subspaces under the channel diversity assumption. First, dene the input and output consecutive row (block row) subspaces spanned by vectors as
(4) , we similarly dene and as the Note that for consecutive future data vectors. The subspaces spanned by following property results directly from A1. Property 1: Under A1, there exists a (smallest) such that for any , we have the isomorphic relation between the input and output subspaces (5) Proof: Let
B. The Model The identication and estimation of a single-input -output linear system shown in Fig. 1 is considered in this paper. The system equation is given by From (3), we have
(6)
(7) matrix is the ltering matrix where the with the following block Toeplitz structure:
(1) where output, is the (noiseless) channel is the additive noise, is the received signal, .. . .. . (8)
3002
It has been shown in [9] and [19] that A1 implies the existence such that for any , of a smallest has full column rank. Thus, from (7), we have (9) which leads to (5). Property 1 plays an important role in the smoothing and linear prediction approaches to blind channel estimation. It is this isomorphic relation between the input and output subspaces that makes it possible to identify the channel from the input subspace without the direct use of the input sequence. To ensure channel identiability, the input sequence must be sufciently complex in order to excite all modes of the channel. This requirement is imposed on the linear complexity [2] of the sequence. has A2: Linear Complexity: The input sequence . linear complexity greater than With the denition of linear complexity2 [2], A2 implies that , we have for any . . .
Fig. 2. Projection of
xt+i (i
= 0;
111
; L)
onto
Z.
Consider
consecutive output block row vectors . From (3), we have
. . .
..
..
. (11)
rank
Toeplitz (10)
This implication of A2 will be exploited in Section III as we discuss the necessary number of input symbols for the channel identication by the proposed algorithm. It has been , the necessary and sufcient shown in [18] that when condition for the unique identication of the channel and its has linear complexity greater than input is A1 and that . Because both future and past data are required in the smoothing approach, a stronger condition is assumed in A2 in this paper. III. LEAST SQUARES SMOOTHINGA GEOMETRICAL APPROACH In this section, we introduce a geometrical approach to linear least squares smoothing channel estimation by exploiting the isomorphic relation between the input and output subspaces. The same approach also leads to a least squares derivation of the LP-based algorithms. A. Channel Identication from the Input Subspace The isomorphism between the input and output subspaces given in Property 1 implies that the input subspace can be constructed from the output. The following question arises from this property: Can we identify the channel from the input subspace? If so, by constructing the input subspace from the output, the channel is obtained from the output alone. An answer to this question is presented below.
2 The linear complexity of the vector v with components v ; ; vn01 0 is dened as the smallest value of c for which a recursion of the form c vi = ; n 1) exists that will generate v from j =1 aj vi0j (i = c; its rst c components.
up to a common Suppose that we want to identify . One way to achieve this is scaling factor from all other terms except the ones to eliminate in associated with . Consider projecting into a punctured input subspace that satises the following two conditions: . C1: . C2: Note that A2 ensures . Thus, this punctured input subspace exists. As illustrated in Fig. 2, is the summation of (a vector outside ) and (a vector inside ). The two shaded right triangles immediately suggest (12) Because matrix is independent of , we have the projection error
. . .
. . .
(13)
Note that is a rank-1 matrix whose column and row spaces , respectively. From , the channel are spanned by and can be directly identied. One approach is the least squares tting of the column space of (14) The above optimization can be solved by the singular value or the sample covariance of the decomposition of either projection error sequence. A simpler approach is to take the as an estimate of row ( ) with the maximum 2-norm in . Then, from (13), we have (15)
111
111
3003
introduce two terms: smoothing window size and smoothing order. The smoothing window size is dened as the number of symbols in the current data, and the smoothing order is the number of symbols in the past data. For the case discussed above, the smoothing window size and the smoothing order are and , respectively. The price we paid for avoiding the direct use of input sequence is that more input symbols are required to identify the channel. From Fig. 3, we can see that the projection subspace and the current data span a -dimensional input subspace denoted as , as shown in (18)
Fig. 3. Isomorphism between input and output subspaces.
. . . To gain a better understanding of this geometrical approach to channel estimation, we make the following remarks. is orthogonal to , which is asymptotically 1) When true for an uncorrelated zero-mean input sequence, the converges to (see Fig. 2). In projection error this case, the rank-one decomposition of the projection provides the input sequence as well error matrix as the channel vector , i.e., this geometrical approach to channel estimation performs direct deconvolution as a by-product. into for different are inde2) The projections of pendent of each other and can be carried out in parallel. This property is attractive in algorithm implementation. B. Channel Identication from the Output SubspaceLeast Squares Smoothing 1) Construction of the Input Subspace from the Output: It has been shown in Section III-A that the channel can be into the identied from the projection errors of satisfying C1 and C2. To avoid the direct input subspace from the channel use of the input sequence, we construct output by exploiting the isomorphic relation between the input and output subspaces given in Property 1. Using the denition in (4), conditions C1 and C2 can be written in the form (16) . With the isomorphic relation between the input for any and output subspaces given in (5), we have, for (17) The isomorphism between the input and output subspaces is into illustrated in Fig. 3. The projection of is converted into the smoothing [6], [8] of the current data by the past data and the future . The projection error used to data identify the channel [see (13)] now becomes the smoothing error, which can be obtained from the output alone. Because of the least squares criterion used in the projection, this geometrical approach to blind channel estimation is referred to as the least squares smoothing (LSS) algorithm. Here, we
Toeplitz (19)
The construction of imposes the full-row-rank condition on input the matrix in (19). As a result, we require observation symbols) to be available symbols ( and the linear complexity of the input sequence to be greater [see (10)]. than 2) Batch Least Squares Smoothing Algorithm: We consider here the problem of estimating a channel with order from a observation symbols. The data structure is specibatch of ed, followed by some useful properties. Then, the batch least squares smoothing channel estimation algorithm is presented. observation symbols Given , for a xed smoothing order and the known channel order , dene the overall data matrix as . . .
. . .
(20)
. . . from which we have specied the past data matrix the current data matrix the future data matrix . future-past data matrix In the absence of noise, the overall data matrix following relation with the input signal: . . . ; ; ; has the
Toeplitz (21)
3004
Fig. 4. Batch LSS algorithm.
3) Stochastic Equivalence: The least squares smoothing approach presented above is based on the deterministic modeling of the input sequence. In order to obtain the consistent estimate of the channel in the presence of noise, we consider a stochastic equivalence of the LSS approach. After replacing the projection of data vectors ( ) into a Euclidean space by ) the projection of corresponding random variables ( into a Hilbert space, the geometrical approach presented in Section III-B1 can be adapted to the stochastic framework. as a white Specically, consider the input sequence random process with zero mean and unit variance. Dene the output random vectors
As direct consequences of (21) and Property 1, the relation among the data matrices in (20) and various subspaces is summarized in Property 2. The rank conditions given below are useful in nding the least squares approximation of the noisy data matrices. Property 2: Suppose that the input sequence has linear , , . complexity greater than We have the following properties in the noiseless case: 1) Overall Data Matrix :
(23) of the linear minimum mean square The error sequence based on is then given error (MMSE) estimator of by (24) where and are the covariance matrices dened as (25)
rank 2) Past Data Matrix :
rank 3) Future Data Matrix :
Based on the principle of orthogonality, the error sequence can be considered as the projection error of into . the Hilbert space spanned by the random variables in Similarly to (13), we then have, for the white input sequence , (26)
rank
rank :
and denoting the covariance of With respectively, we have, from (24) and (26)
and
, (27)
4) Future-Past Data Matrix
rank As stated in the above property, the row vectors in the span the -dimensional future-past data matrix projection space . Therefore, the projection error in (13) can be obtained as the least squares smoothing error of the by the future-past data . Specically, in the current data absence of noise, we have . . . . . .
can be directly obtained from . For This implies that and with known a noise sequence independent of , , and second-order statistics, the covariance matrices can be estimated consistently from the observation. The consistent estimate of the channel can then be obtained. C. Connections with Linear Prediction-Based Approaches The main difference between the LP-based approaches and into the LSS approach is the denition of the subspace which the projection is made. Naturally, LP-based approaches include only the past data in the projection space. We now draw connections with existing LP-based methods by using the same geometrical approach to rederive them under the deterministic model of the input sequence. We comment that the LP-based algorithms presented here are not identical to the stochastic versions presented in the literature, although they are closely related. LP-based channel estimators can be classied into one-step and multistep linear prediction algorithms.
(22) from which the batch least squares smoothing channel estimation algorithm (B-LSS) can be derived directly. One possible implementation is summarized in Fig. 4.
3005
1) One-Step LP Algorithm: We consider here the LPbased approach proposed by Abed-Meraim et al. in [1] and [3] to which we refer as the linear prediction-least into squares (LP-LS) algorithm. Instead of projecting , let us consider projecting into , which contains only the past data. We then have, again using the isomorphic relation (28) from which can be obtained directly. By treating as an estimate of , the problem becomes conventional channel estimation with known channel input. The least squares criterion can then be employed to estimate the whole channel, which is the reason that we refer to this one-step LP , which is algorithm as LP-LS. Note that if asymptotically true for uncorrelated zero-mean sequence, we (see also Fig. 2). Hence, this approach have provides consistent estimates for white inputs. Unfortunately, , which causes the for a nite sample size, loss of the nite sample convergence property in the LP-LS approach. 3) Multistep LP Algorithm: In contrast to the LP-LS approach, the multistep LP approach proposed by Gesbert and Duhamel [4] works simultaneously on predictions of several into , we have future data samples. By projecting the prediction error equations
squares smoothing channel estimation and order detection algorithm is then presented. A. Data Structure and Order Detection In contrast with the data structure given in (20) where the smoothing window size is xed, the data structure for an arbitrary smoothing window size is now considered. Suppose that an upper-bound of the channel order is known. For a xed smoothing order and a variable smoothing window size , dene the overall data matrix as . . .
. . .
(32)
. . . (33) , from which we have dened the future data matrix , and the past data matrix . the current data matrix and the future-current data The future-past data matrix are also dened. Compared with the ones dened matrix and , the data matrices in (20), we can see that except for dened here are functions of the variable smoothing window size . We emphasize that in this particular data structure, and are independent of . This property leads to a convenient implementation of A-LSS, which will be discussed in Section IV-B. Similar to the case presented in Section III-B2, when and has linear complexity greater than , in the noiseless case, we have the following relation among the matrices dened in (32) and the various spaces:
(29) For the uncorrelated zero-mean input sequence, we have, asymptotically (30) Interestingly, the above equation was also used by Slock and Papadias in their extension of one-step prediction to -step prediction [17]. However, these equations were not exploited jointly in [17]. Gesbert and Duhamel treated the above as a triangular system and constructed the error differentials . . .
(31) (34)
becomes a familiar problem. The identication of from Note that unlike the one-step approach, the entire channel parameter is identied at once. The multistep LP approach again relies on the uncorrelateness of the input sequence, which, for the same reason as in the LP-LS approach, causes the loss of nite-sample convergence property. IV. ADAPTIVE LEAST SQUARES SMOOTHING CHANNEL ESTIMATION In this section, we develop an adaptive LSS algorithm with unknown channel order. The data structure for a variable smoothing window size is rst specied. The adaptive least
(35) In parallel to the known channel order case in (22), we consider channel identication by a rank-1 decomposition of . Since (36) requires that the orthogonal the rank-1 condition on in has the dimension 1. Comparing complement of the input subspaces in (34) and (35), we can see that this can only be met when . When requirement on , equals ; there is no projection (smoothing)
3006
Fig. 5.
Isomorphism between input and output subspaces for
l= 6 L + 1.
error. The isomorphism between the input and output spaces in this case is shown in Fig. 5(a). The case when is shown in Fig. 5(b). Here, are not in . contributes Since each of these input vectors not lying in , as to the smoothing error, we have the formulation of given in the following theorem. be the least squares smoothTheorem 1: Let ing error. With the data structure specied in (32), we have, in the absence of noise
Fig. 6. Energy of the smoothing error at different smoothing window size l.
. . .
.. ..
. . . . . . . .
(37) Theorem 1 holds the key to the adaptive least squares smoothing channel estimation and order detection algorithm. of the For order detection, we consider the energy . Theorem 1 implies smoothing error dened as is zero when . When that in the noiseless case, , jumps to a value related to the power of the channel and it increases (asymptotically) with , as shown in Fig. 6. Hence, the channel order can be detected by varying from 1 to and comparing the energy of the smoothing error at each with a certain threshold. B. Adaptive Least Squares Smoothing 1) Key Components: When the channel order is unknown, there are three key components involved in the A-LSS approach to channel estimation. First, the smoothing errors at each smoothing window size need to be calculated. Then, based on Theorem 1, the energy of these
Fig. 7. Main structure of A-LSS.
smoothing errors is compared with a threshold to detect the channel order. Finally, with the detected channel order , the channel is obtained from . Our goal here is to develop an algorithm to implement the above three components while simultaneously satisfying the requirements in convergence speed, adaptivity, and complexity. Since the LSS approach has the nite sample convergence property and adaptivity is a characteristic feature of linear least squares based algorithms, we next concentrate on the efcient implementation of A-LSS. The main computation cost of the A-LSS algorithm comes at all . Thus, from the calculation of the smoothing errors it is desirable to obtain the smoothing errors efciently, which can be achieved by recursively; a) calculating b) decomposing the smoothing into multistep predictions followed by linear projections;
3007
Fig. 8.
Lattice lter for multistep prediction.
c) sharing the same multistep predictions among the calculation of the smoothing errors at all . Implementing smoothing by multistep predictions enables the use of lattice lters that are modular, computationally efcient, and robust to round-off errors. Details are presented as follows. 2) Key Decomposition: In A-LSS, smoothing is decomposed into multistep predictions and linear projections based on the following basic lemma. and a partitioned matrix Lemma 1: For a matrix with , the least squares projection error of into is given by (38) is equal to the In words, the projection error of into into , where and are projection error of and into , respectively. projection errors of , is the orthogonal complement Proof: In . Therefore of (39) which leads to (40) Based on the partition of apply Lemma 1 to the smoothing error given in (33), we
structure dened in (32), only a single predictor is necessary. This is because from (32), we have (42) and are independent of . where both 3) Main Structure: Based on the decomposition presented above, A-LSS identies the channel in three steps, as shown in Fig. 7. Specically, the multistep predictor generates the , which are independent of . prediction errors In the second step, the smoothing errors at all possible are obtained recursively, and the channel order is detected. Finally, with the detected channel order , the channel is estimated with the approach given in from the smoothing error (15) to meet the requirement in adaptivity. The implementation of these three steps is discussed below. a) Multistep predictions: A lattice lter shown in Fig. 8 is used to obtain the prediction errors of the future and current by the past data . Various adaptive algorithms for data lattice lters can be applied here; see [10], [12], and [15]. Besides the appealing properties mentioned in Section IV-B1, the order-recursive property of lattice lters can be very useful in saving computation cost for the unknown channel-order case. Specically, when the channel order is unknown, the smoothing order (also the prediction order here) has to be . For example, when determined based on the upper-bound , we choose . If is a poor bound of the channel order, may be much larger than necessary, which leads to a higher computation cost. Because of the order-recursive property of the lattice lter, for a detected channel order , only the rst prediction errors ( ) at the th lattice stage are required. This implies that the computation involved in the latter stages and the joint estimation ( ) is saved with no structural change. b) Projections and order detection: From the prediction , the smoothing errors at each smoothing errors
(41) and are the multistep prediction errors where and by the past data , respectively. The of is then obtained by projecting smoothing error into the row space of , which implies that smoothing is decomposed into multistep predictions and linear projections. Although (41) appears to imply that separate linear predictors are required for different , it turns out that with the special data
3008
Fig. 9. Recursive projection and order detection.
window size are obtained in this step for the order detection and channel estimation. recursively (both in and An approach to obtaining in time) by applying the recursive modied GramSchmidt (RMGS) algorithm [11] is presented with a simple example shown in Fig. 9. Suppose that the channel order is upper , and we choose the smoothing order bounded by . From the prediction errors obtained by the multistep predictor, the smoothing errors at are calculated recursively. As shown in Fig. 9, by and , applying Lemma 1 rst to we have (43) where denotes the partitioned matrix and .
in : (45) where has been partitioned into with dened as the rst row block in . This recursion in arises naturally from the data structure given in (32). are obtained, the channel order is detected by Once all is calculated an energy detector. Since the energy of for the channel order detection, the index ( ) of the row with [see (15)] can be easily the maximum 2-norm in obtained. This information will be used in the next step to . estimate the channel from Finally, we want to emphasize that the structure shown in Fig. 9 is particularly attractive for time-varying channels. For example, when the true channel order changes from 2 to 1, the to . The channel with energy detector switches from the new order 1 can be identied without structural change or extra computation. In addition, the modular structure shown in Fig. 9 is suitable for implementation using systolic array processors. c) Recursive channel estimation: With obtained as the by-product of the energy detector in the second step, the
Applying Lemma 1 next to , we have
(44) . Similarly, and which is the th row block of can be obtained. In general, we have the following recursion
3009
Fig. 10.
Recursive channel estimation.
First, this channel allows us to study the location of zeros and how they affect the performance. Second, Hua showed that the CR algorithm [22] and SS algorithm [14], along with the twostep maximum likelihood (TSML) algorithm [7], approach to the Cram erRao lower bound.4 By comparing with CR(SS), we can evaluate the efciency of the algorithms listed in Table I. The second-order channel used by Hua in [7] is specied by a pair of roots on the unit circle. The channel impulse response is given by (48) where are the angular positions of zeros on the unit circle. The relation between the zeros of the two subchannels is , where is the angular distance specied by between the zeros of the two subchannels (common zeros occur when is zero). C. Performance and Robustness Comparison
ALGORITHMS
TABLE I COMPARED IN THE SIMULATION AND THEIR CHARACTERISTICS
channel is estimated from the smoothing error by is an asymptotical the approach given in (15). In addition, estimate of the uncorrelated input sequence. The structure of this recursive channel estimator is illustrated in Fig. 10. V. SIMULATION EXAMPLES A. Algorithm Characteristics and Performance Measure Simulation studies of the proposed A-LSS algorithm as it is compared with existing techniques listed in Table I are presented in this section. We remark that only A-LSS has easy adaptive implementations while still preserving the nite sample convergence property. Algorithms were compared by Monte Carlo simulation using the normalized root mean square error (NRMSE) as a performance measure. Specically, NRMSE was dened as3 NRMSE (46)
was the estimated channel from the th trial, where was the true channel. Noise samples were generated and from i.i.d. zero mean Gaussian random sequences, and the signal-to-noise ratio (SNR) was dened as SNR (47)
The left side of Fig. 11 shows the performance of algorithms listed in Table I against SNR. A well-conditioned channel was , as in [7]. All the deterministic used with algorithms (CR, SS, B-LSS, and A-LSS) had comparable performance. In fact, it was shown in [7] that CR and SS approached the CRB even at a very low SNR for this channel. The LP-based algorithms (MSP and LP-LS) leveled off as because of the loss of nite sample convergence SNR property. We noticed that B-LSS had the same performance as CR (SS), but there was a gap between the performance of A-LSS and B-LSS at low SNR. This gap was caused by the prewindowing problem in A-LSS; several symbols were discarded in the transient stage. Thus, the number of the effective symbols used by A-LSS was less than that used by the batch algorithms, which led to the performance degradation. This performance gap will diminish as SNR or more observation samples become available. An ill-conditioned channel was used to compare the robustness of different algorithms with respect to the loss of , zeros of the channel diversity. With two subchannels were very close to each other. In fact, the condition number of the ltering matrix was around 1.6 104 in this case. The right side Fig. 11 shows that CR and SS performed rather poorly for this ill-conditioned channel. Evidently, B-LSS and A-LSS performed considerably better than all other algorithms. This improvement in robustness is probably because the input subspace may still be well approximated by the output subspace when the channel diversity assumption does not hold. D. Tracking of the Channel Order and Parameters In this simulation, A-LSS was applied to a case when both channel order and channel parameters had a sudden change. The initial channel was given in (48) with . At time , both the channel order and channel parameters were changed by adding zeros
4 In [7], the CRB was derived based on a normalization that differs from the one used in this paper.
The input to a single-input and 2-output linear FIR system was generated from an i.i.d. binary phase shift keying (BPSK) sequence. B. A Second-Order Channel A second-order channel rst used by Hua in [7] is considered here. There are several reasons to consider this channel.
3 The
inherent ambiguity was removed before the computation of NRMSE.
3010
Fig. 11. Performance and robustness comparison (100 Monte Carlo runs, 200 input symbols).
Fig. 12.
Channel order and parameter tracking performance (SNR = 30 dB).
to the two subchannels, respectively. The left side of Fig. 12 shows the energy of the smoothing error before and after the channel variation (the energy of the smoothing error was calculated every 10 symbols). From Fig. 12(a), we can see that before the channel changed, the energy of the smoothing error and was relatively large was around zero at , as predicted by Theorem 1. Fig. 12(b)(d) at 160, are the snapshots of the smoothing error energy at 170, and 180, respectively. We can see that the energy of decreased to zero within 30 samples. At 180, the new channel order can be accurately detected. Consequently, instead of . Neither the channel is estimated from structural change nor extra computation is involved in this process. The NRMSE convergence of A-LSS is shown in the right side of Fig. 12, where A-LSS tracked the channel order and parameters nicely.
D. Order Detection for Multipath Channels In order to evaluate the applicability of the proposed order detection and channel estimation algorithm, we present here
the simulation study of the performance of A-LSS with a multipath channel. The channel we used is a fth-order 4-ray raised-cosine channel as shown in the top left of Fig. 13 with the even and odd samples corresponding to the two subchannels. The upperbound of the channel order was assumed to be 6. The energy of the smoothing error at different smoothing window size is shown in the top right of Fig. 13. We can see that the energy of the smoothing error jumped to a large . Based on this fact, value at the smoothing window size the energy detector estimated the channel order as 1. In the bottom left of Fig. 13, we also plotted the relative increment of the smoothing error energy at two consecutive smoothing window sizes, and we chose maximizing the relative increment as an alternative criterion for the order detection. The same was obtained. As shown in estimated channel order the bottom right of Fig. 13, among all the possible channel orders (from 0 to ), A-LSS gave the best estimate of the channel at the detected order. In fact, as pointed out in [20], it is perhaps not wise to estimate the small head and tail taps in the multipath channel. Instead, it is better to nd the channel order as well as its impulse response that matches the
3011
Fig. 13.
Order detection and channel estimation with multipath channel (SNR = 30 dB, 100 Monte Carlo runs, 200 input symbols).
data in some optimal way. In the case here, with the channel order detected as 1, A-LSS captured the four major taps of the channel impulse response while ignoring the small head and tail taps. VI. CONCLUSION A least squares smoothing (LSS) approach to blind channel estimation based on the isomorphic relation between the input and output subspaces is presented. This approach converts the channel estimation into a linear least squares smoothing problem that can be solved (order and time) recursively. Since the input sequence is modeled as a deterministic signal, this approach preserves the nite-sample convergence property not present in LP-based approaches. Based on the LSS approach, an adaptive channel estimation (A-LSS) algorithm has been proposed. Compared with the batch algorithms such as SS and CR, the A-LSS algorithm is efcient in both computation and VLSI implementation. Because smoothing is computationally more expensive than prediction, A-LSS has higher complexity than the LP-LS and the MSP approaches, which is a price we paid for the nite sample convergence property. The A-LSS algorithm does not assume the channel order as a priori information. The order detector and the recursive property enable A-LSS to detect and track the channel order without structural change or extra computation. Although separate order detection techniques can be applied to the deterministic batch algorithms such as SS and CR, their ability to track the channel order variation demands high implementation cost. The future work involves the theoretical determination of the threshold for the channel order detection. Maximizing the relative increment in the smoothing error energy at two consecutive smoothing window sizes may be considered as an alternative criterion. REFERENCES
[1] K. Abed-Meraim, E. Moulines, and P. Loubaton, Prediction error method for second-order blind identication, IEEE Trans. Signal Processing, vol. 45, pp. 694705, Mar. 1997.
[2] R. E. Blahut, Algebraic Methods for Signal Processing and Communications Coding. New York: Springer-Verlag, 1992. [3] A. K. Meraim et al., Prediction error methods for time-domain blind identication of multichannel FIR lters, in Proc. ICASSP Conf., Detroit, MI, May 1995, vol. 3, pp. 19601963. [4] D. Gesbert and P. Duhamel, Robust blind identication and equalization based on multi-step predictors, in Proc. IEEE Int. Conf. Acoust. Speech, Signal Process., Munich, Germany, Apr. 1997, vol. 5, pp. 26212624. [5] M. L. G urelli and C. L. Nikias, EVAM: An eigenvector-based deconvolution of input colored signals, IEEE Trans. Signal Processing, vol. 43, pp. 134149, Jan. 1995. [6] S. Haykin, Adaptive Filter Theory. Englewood Cliffs, NJ: PrenticeHall, 1996. [7] Y. Hua, Fast maximum likelihood for blind identication of multiple FIR channels, IEEE Trans. Signal Processing, vol. 44, pp. 661672, Mar. 1996. [8] S. Kay, Fundamentals of Statistical Signal Processing: Estimation Theory. Englewood Cliffs, NJ: Prentice-Hall, 1993. [9] S. Y. Kung, T. Kailath, and M. Morf, A generalized resultant matrix for polynomial matrices, in Proc. IEEE Conf. Decision Contr., 1976, pp. 892895. [10] H. Lev-Ari, Modular architectures for adaptive multichannel lattice algorithms, IEEE Trans. Signal Processing, vol. 35, pp. 543552, Apr. 1987. [11] F. Ling, D. Manolakis, and J. G. Proakis, A recursive modied GramSchmidt algorithm for least-squares estimation, IEEE Trans. Acoust., Speech, Signal Processing, vol. ASSP-34, pp. 829836, Aug. 1986. [12] F. Ling and J. G. Proakis, A generalized multichannel least squares lattice algorithm based on sequential processing stages, IEEE Trans. Acoust., Speech, Signal Processing, vol. ASSP-32, pp. 381390, Apr. 1984. [13] H. Liu and G. Xu, Closed-form blind symbol estimation in digital communications, IEEE Trans. Signal Processing, vol. 43, pp. 27142723, Nov. 1995. [14] E. Moulines, P. Duhamel, J. F. Cardoso, and S. Mayrargue, Subspacemethods for the blind identication of multichannel FIR lters, IEEE Trans. Signal Processing, vol. 43, pp. 516525, Feb. 1995. [15] A. H. Sayed and T. Kailath, A state-space approach to adaptive RLS ltering, IEEE Signal Processing Mag., vol. 11, pp. 1860, July 1994. [16] D. Slock, Blind fractionally-spaced equalization, perfect reconstruction lterbanks, and multilinear prediction, in Proc. ICASSP Conf., Adelaide, Australia, Apr. 1994. [17] D. Slock and C. B. Papadias, Further results on blind identication and equalization of multiple FIR channels, in Proc. Int. Conf. Acoust. Speech Signal Process., Detroit, MI, Apr. 1995, pp. 19641967. [18] L. Tong and J. Bao, Equalizations in wireless ATM, in Proc. 1997 Allerton Conf. Commun., Contr. Comput., Urbana, IL, Oct. 1997, pp. 6473. [19] L. Tong, G. Xu, B. Hassibi, and T. Kailath, Blind identication and equalization of multipath channels: A frequency domain approach, IEEE Trans. Inform. Theory, vol. 41, pp. 329334, Jan. 1995.
3012
[20] L. Tong and Q. Zhao, Joint order detection and channel estimation by least squares smoothing, IEEE Trans. Signal Processing, vol. 47, pp. 23452355, Sept. 1999. [21] A. van der Veen, S. Talwar, and A. Paulraj, A subspace approach to blind space-time signal processing for wireless communication systems, IEEE Trans. Signal Processing, vol. 45, pp. 173190, Jan. 1997. [22] G. Xu, H. Liu, L. Tong, and T. Kailath, A least-squares approach to blind channel identication, IEEE Trans. Signal Processing, vol. 43, pp. 29822993, Dec. 1995.
Qing Zhao received the B.S. degree in electrical engineering in 1994 from Sichuan University, Chengdu, China, and the M.S. degree in 1997 from Fudan University, Shanghai, China. She is now pursuing the Ph.D. degree at the School of Electrical Engineering, Cornell University, Ithaca, NY. Her current research interests include wireless communications and array signal processing.
Lang Tong (S87M91), for a photograph and biography, see this issue, p. 2999.
1026
IEEE Transactions on Consumer Electronics, Vol. 50, No. 4, NOVEMBER 2004
Blind Adaptive Channel Equalization Using Multichannel Linear Prediction-Based Cross-Correlation Vector Estimation
Kyung Seung Ahn, Student Member, IEEE, Juphil Cho, and Heung Ki Baik, Member, IEEE
Abstract Blind equalization of transmission channel is important in communication areas and signal processing applications because it does not need training sequences, nor does it require a priori channel information. In this paper, an adaptive blind MMSE channel equalization technique based on second-order statistics is investigated. We present an adaptive blind MMSE channel equalization using multichannel linear prediction error method for estimating cross-correlation vector. They can be implemented as RLS or LMS algorithms to recursively update the cross-correlation vector. Once cross-correlation vector is available, it can be used for MMSE channel equalization. Unlike many known subspace methods, our proposed algorithms do not require channel order estimation. Therefore, our algorithms are robust to channel order mismatch. Performance of our algorithms and comparison with existing algorithms are shown for real measured digital microwave channel.1
Index Terms Blind channel equalization, decision delay, intersymbol interference, prediction error filter.
I. INTRODUCTION Multipath propagation appears to be a typical limitation in mobile digital communication where it leads to severe intersymbol interference (ISI). The classical techniques to overcome this problem use either periodically sent training sequences or blind techniques exploiting higher order statistics (HOS) [1]. Adaptive equalization using training sequence wastes the bandwidth efficiency but in blind equalization, no training is needed and the equalizer is obtained only with the utilization of the received signal. A well-known HOS-based algorithm is constant modulus algorithm (CMA). A main disadvantage of this algorithm is the possibility of local convergence [2]. The fractionally-spaced CMA (FS-CMA) was shown in [3] to guarantee global convergence with finite length equalizers. Since the seminal work by Tong et al. the problem of estimating the channel response of multiple FIR channel driven by an unknown input symbol has interested many researchers in the signal processing areas and communication fields [4], [5].
Kyung Seung Ahn is with the Department of Electronic Engineering, Chonbuk National University, Jeonju, Korea (e-mail: ksahn@chonbuk.ac.kr). Juphil Cho is with the Electronics and Telecommunications Research Institute (ETRI), Daejeon, Korea. Heung Ki Baik is with the Division of Electronics & Information Engineering, Electronics & Information Advanced Technology Research Center, Chonbuk National University, Jeonju, Korea. Contributed Paper Manuscript received August 1, 2004
For the most part, algebraic and second-order statistics (SOS) techniques have been proposed that exploit the structural techniques (Hankel, Toeplitz matrix, et al.) of the single-input multiple-output (SIMO) channel or data matrices. The information on channel parameters or transmitted data is typically recovered through subspace decomposition of the received data matrix (deterministic method) or that of the received data correlation matrix (stochastic method). Although very appealing from the conceptual and signal processing techniques point of view, the use of the aforementioned techniques in real world applications faces serious challenges. Subspace-based techniques lay in the fact that they relay on the existence of numerically well-defined dimensions of the noisefree signal or noise subspaces. Since these dimensions are obviously closely related to the channel length, subspacebased techniques are extremely sensitive to channel order mismatch [6], [7]. The prediction error method (PEM) offers an alternative to the class of techniques above. PEM, which was first introduced by Slock et al. and later refined by Meraim et al. exploited the independent, identically distributed (i.i.d.) property of the transmitted symbols and applies a linear prediction error filter on the received data [7], [8]. The PEM offers great practical advantages over most other proposed techniques. First, channel estimation using the PEM remains consistent in the presence of the channel length mismatch. This property guarantees the robustness of the technique with respect to the difficult channel length estimation problem. Another significant advantage of the PEM is that it lends itself easily to a low-cost adaptive implementation such as adaptive lattice filters. But the delay cannot be controlled with existing one-step linear prediction method [8][11]. These algorithms calculate ZF or MMSE equalizers based on channel identification. In this paper, we propose adaptive blind MMSE channel equalization algorithm that is less sensitive to the equalizer order length. This paper makes three results. First, we estimate the multichannel prediction-based cross-correlation vector between the channel output and the transmitted signal. Estimated cross-correlation vector can then be used to calculate the MMSE blind equalizer. Second, we develop an efficient algorithm to perform this cross-correlation vector tracking. A simple iterative algorithm can be used to find the cross-correlation vector. Third, we derive LMS and RLS type adaptive algorithms for MMSE blind equalization. Most notations are standard: vectors and matrices are boldface small and capital letters, respectively; the matrix transpose, the
0098 3063/04/$20.00 2004 IEEE
K. S. Ahn et al.: Blind Adaptive Channel Equalization Using Multichannel Linear Prediction-Based Cross-Correlation Vector Estimation
1027
v0(n) h0(n) a0(n)
x0(n)
( N + L) 1 , x N (n) and v N (n) are NP 1 vectors as following.

g0(n)
s(n) = [ s (n) L s (n L N + 1)]T
s(n)
vP-1(n)
( n D ) s
x N (n) = [xT (n) L xT (n N + 1)]T v N (n) = [ v (n) L v (n)]

and
T T T
(6)
hP-1(n)
Fig. 1 The multichannel representation of a T P -spaced equalizer.
Hermitian, and the Moore-Penrose pseudoinverse are denoted by ()T , () H , and () + , respectively; I P is the P P identity matrix; E[] is the statistical expectation. This paper is organized as follows. In section II, we present models for the channel and blind MMSE equalizer, and we introduce the concepts of signal vectors and matrices. In section III, we discuss multichannel linear prediction problem for crosscorrelation vector estimation. Simulation results and performance comparisons of our proposed algorithms with some well-known existing algorithms are presented in section IV. We conclude our results in section V. II. PROBLEM FORMULATION
Let x(t ) be the signal at the output of a noisy communication channel
where s (k ) denotes the transmitted symbol at time kT , h(t ) denotes the continuous-time channel impulse response, and v(t ) is additive noise. The fractionally-spaced discrete-time model can be obtained either by time oversampling or by the sensor array at the receiver. The oversampled single-input single-output (SISO) model results SIMO model as in Fig. 1. The corresponding SIMO model is described as following
xi (n) = hi (n) s (n k ) + vi (n)

k =0
where P is the number of subchannel, and L is the maximum order of the P subchannel. Let
We represent xi (n) in a vector for as
Stacking N received vector samples into an ( NP 1) vector, we can write a matrix equation as
where H is a NP ( N + L) block Toeplitz matrix, s(n) is
...
x (t ) =
k =
...
aP-1(n)
xP-1(n)
gP-1(n)
0 h(0) L h( L 1) L H= M O M O M . 1) (0) ( L 0 L h L h We assume the following in this paper.
(7)
A1) The input sequence s (n) is zero-mean and white with variance s2 . A2) The additive noise v(n) is stationary with zero-mean and white with variance v2 . A3) The sequence s (n) and v(n) are uncorrelated. A4) The matrix H has full rank, i.e., the subchannels hi (n) have no common zeros to satisfy the Bezout equation [4]. A5) The dimension of H obeys NP > N + L 1 . Consider an FIR linear MMSE equalizer shown in Fig. 1, where gi (n) for i = 0, 1, L , P 1 is the order N equalizer of the ith subchannel and the symbol is estimated from
s(k )h(t kT ) + v(t )
(1)
(n d ) = g H x N (n). s
(8)
According to (5)(7), x N (n) has nonzero correlation with only s (n),L , s (n L N + 1) . Therefore, decision delay d is usually in the interval [0, N + L 1] . For finite SIMO channels, MMSE equalizer of the finite length can be found if assumption A4) holds and the equalizer length N L [8]. An MMSE equalizer minimizes the cost function
L 1
( n d ) . J MMSE = E s (n d ) s MMSE blind equalizer with delay d is given by

2
(9)
(2)
= ai (n) + vi (n), i = 0,L , P 1
2 g MMSE = arg min E g H x N ( n) s ( n d ) . (10) g (n d ) is the solution The minimum MSE filter to estimate s to the Wiener-Hopf equation [12]
H E x N ( n)x N ( n) g MMSE = E x N ( n) s ( n d ) .
(11)
x(n) = [ x0 (n) L xP 1 (n)] h(n) = [h0 (n) L hP 1 (n)]
(3)
Using assumptions A1)A3), we can write the exact correlation matrix x N (n) as
H 2 H 2 R = E x N ( n) x N ( n) = s HH + v I
v(n) = [v0 (n) L vP 1 (n)] .
(12) (13)
where the cross-correlation vector equals (4)

2 E x N ( n) s ( n d ) = s H(d ).
x(n) = s (k )h(n k ) + v (n).

k =0
L 1
x N (n) = Hs(n) + v N (n)
(5)
where H (d ) denotes the (d + 1)th block column of the channel convolution matrix H [12]. Based on assumption that s (n) the unit variance, the blind MMSE equalizer with decision delay d is given by
g MMSE = R + H (d ).
(14)
1028

H H H J MMSE = 1 g MMSE H2 H2 g MMSE + g MMSE Rg MMSE
(n d ) is the solution The minimum MSE filter to estimate s to the Wiener-Hopf equation [12]
III. PROPOSED ALGORITHMS
A. Multichannel Linear Prediction Consider the noise-free case. For convenience, we can rewrite (5) as
(25) (26) (27)
Then, minimizing J MMSE yields
g MMSE = R + H 2 = R + H (d )
From (23), we know that
H 2 H 2 H E f (n)f (n) = s H 2 H 2 = s H (d )H (d )
s1 (n) x N (n) = Hs(n) = [ H1 H 2 H 3 ] s 2 (n) s3 ( n)
(15)
where H1 is of dimension NP d , the NP 1 vector H 2 is the (d + 1)th column of H , and the last part of H is denoted by H 3 with dimension NP ( N + P d 1) [13]. An NP 1 multichannel linear prediction error vector can be obtained as [13]
Then, the rank of the matrix E[f (n)f H (n)] is one. Therefore, H (d ) is the singular vector corresponding to the largest singular value of matrix E[f (n)f H (n)] . Substituting this estimated cross-correlation vector into (26), MMSE blind equalizer is obtained. If the decision delay d equals channel length, then we can obtain channel coefficient since H 2 is just the channel coefficient vector. B. Adaptive Implementation We propose the adaptive algorithms for updating the multichannel linear prediction error filter coefficients. Two multichannel linear prediction problems are required in estimating the MMSE blind equalizer. We are required to compute the multichannel prediction matrices and to estimate the multichannel prediction errors and in (16) and (20). In order to fast convergence, we can use the RLS algorithm to update the multichannel linear prediction as following
z
x N ( n) I P1 f 1 ( n) = = H 1s 1 (n) x M (n d )
(16)
where P1 is an NP MP matrix. x N (n) is defined in (6), and x M (n d ) is defined as
x( n d ) . (17) x M ( n) = M x ( n d ) The optimal P1 is obtained by minimizing as following cost function

E f 1 (n)f 1H (n) . J1 = tr (18) Letting the partial derivative of (18) with respective to P1 equals to zero as following
J1 H = E x N (n)x M (n d ) + P1x M (n d )x M (n d ) =0. (19) P1
We get as
H x M (n d )x M P1 = E (n d ) E [ x N (n)x M (n d )] (20) Consider another multichannel linear prediction error problem
Compute multichannel linear prediction error vector:
f 1 (n) = x N (n) P1 (n 1)x M (n d ) f 2 (n) = x N (n) P 2 (n 1)x M (n d 1).

z
(28)
Compute Kalmna gain:
K 1 ( n) = K 2 ( n) =
z
1Q1 (n 1)x M (n d )
H 1 + 1x M (n d )Q1 (n 1)x M (n d )
1Q 2 (n 1)x M (n d 1)
H 1 + 1x M (n d 1)Q 2 (n 1)x M (n d 1)
(29)
Update inverse of the correlation matrix: (30)
H Q1 (n) = 1 Q1 (n 1) K1 (n)x M (n d )Q1 (n 1) H Q 2 (n) = 1 Q 2 (n 1) K 2 (n)x M (n d 1)Q 2 (n 1) .
x N ( n) s1 (n) f 2 ( n) = . (21) = H 1 H2 I P 2 x M (n d 1) s 2 ( n) The proof is again provided in [13]. Compared with (18), (19), and (20), we can know that the optimal P 2 is obtained as following
H x M (n d 1)x M P2 = E (n d 1)
Update multichannel linear prediction coefficients matrix:

H P1 (n) = P1 (n 1) + f 1 (n)K1 ( n) H P 2 (n) = P 2 (n 1) + f 2 (n)K 2 (n).
(31)
E [ x N (n)x M (n d 1) ]
(22)
Compute another prediction error vector:
f (n) = f 1 (n) f 2 (n).
(32)
In order to consider H 2 , we can compute
f (n) = f 2 (n) f 1 (n) = H 2 s 2 (n).

The transmitted signal with decision delay d is written as
H H2 [0 H 2 0] s(n). H H2 H2
(23)
s(n d ) =
(24)
The term (0 < 1) is intended to reduce the effect of past values on the statistics when the filter operates in nonstationary environment. It affects the convergence speed and the tracking accuracy of the algorithm [14]. From the covariance matrix of f (n) in (27), its estimation of adaptive manner is given by
Substitute (8) and (24) back to (9), we get
F(n) = F (n 1) + f (n)f H (n).
(33)
1029
To update the correlation estimates recursively, we can use the sample estimator
H R ( n) = n k x N ( k ) x H N ( k ) + x N ( n) x N ( n) k =0 n 1
2 a11 e1 + a a1 Fe(0) e(1) = = Fe(0) 2 a11 e1 + a a1
( )e
2 1
NP + L + aa 1
( )e )
NP 1
NP
Fe(0)
(34)
= R (n 1) + x N (n)x (n).
H N
e(2) =
M
Defining B(n) = R (n) and using the matrix inversion lemma (e.g., [14], p. 565A) with (34), we can write
Fe(1) = Fe(1)
( )
2 1
a2 a1
e2 + L + Fe(1)
aNP a1
( )
aNP a1
NP 2 1
e NP
)
e NP
B(n) = 1 B(n 1) 1K (n)x H N ( n )B (n 1)

and
(35)
e ( n) =
Fe(n 1) = Fe(n 1)
a11 e1 +
( )
2 n 1
e2 + L +
( )
NP n 1
Fe(n 1)
).
(42) (43)
K ( n) =
B(n 1)x M (n) . 1 + 1 x H N ( n)B ( n 1) x N ( n)

1
(36)
Because of the ordering of the eigenvalues, as n
Notice that (35) and (36) do not require a matrix inverse problem and thereby provide a recursive method for computing B(n). From (26), we can use the following equation to obtain the MMSE-equalizer estimation
e( ) = a1e1
g(n) = B(n)H (d ).
(37)
which is the eigenvector of F corresponding to the largest eigenvalue. The eigenvalue itself is found by a Rayleigh quotient,
The above multichannel linear prediction problems can be computed by an LMS algorithm. The first one can be updated by
e H (n)Fe(n) 1 . e( n )
(44)
f 1 (n) = x N (n) P1 (n 1)x M (n d ) f 2 (n) = x N (n) P 2 (n 1)x M (n d 1).

The second one is updated by
H P1 (n) = P1 (n 1) + 1f 1 (n)x M (n d )
(38)
It should be noted that the MMSE equalizer is designed for transmitted symbol recovery at specific decision delay. Thus, different decision delay can result in different performance. A recursive form to get best decision delay is discussed in [12] and [22]. To get best decision delay choice, [12] proposes the minimizing MSE is given by
P 2 (n) = P 2 (n 1) + 2 f 2 (n)x H M ( n d 1).

C. Cross-Correlation Vector Tracking Algorithm
(39)
J MMSE (d ) = 1 H H (d )R + H (d ).
(45)
A simple iterative algorithm known as the power method can be used to find the largest eigenvector and its associated eigenvalue [15], [23]. Let F be a matrix with possibly eigenvalues ordered as 1 > 2 3 L NP , with corresponding eigenvectors e 1 , e 2 , L , e NP . Let e(0) be a normalized vector that is assumed to be not orthogonal e(1) . The vector e(0) can be written in terms of the eigenvector as
If the transmitted symbols have constant modulus (CM) property, which is practical case in digitally modulated signal such as QAM or PSK, the best decision delayed blind equalizer can be determined by the following CM index [22]:
2 H . J CM (d ) = (46) g d xN (n) 1 where g d is d-delay blind equalizer. The blind equalizer having the smallest J MMSE or J CM value will be considered as the best decision delayed blind equalizer. In many practical channels, it has been observed [1] that selecting d ( N + L) 2 results in good performance. 2
e(0) = a 1e1 + a 2 e 2 + L + a NP e NP
(40)
for some set of coefficient a i , where a 1 0 . We define the power method recursion by
IV. SIMULATION RESULTS

In this section, we use computer simulations to evaluate the performance of the proposed algorithm. We compare the performance of the proposed algorithm with some existing algorithms. The SNR is defined to be at the input to the equalizer as shown in Fig. 1.
2 P 1 E j = 0 a j ( n) (47) SNR = . 2 P 1 E v n ( ) j j =0 As a performance index, we estimate the residual ISI over 50 independent Monte Carlo runs. The residual ISI is defined as
Fe(n) e(n + 1) = . Fe(n)

Then
(41)
1030
1.2 1 0.8 0.6 0.1

Real part
0.2
Imaginary part
SNR=20dB
0.8 0.6
0.2 0 -0.2 0 5 10 15 20
Magnitude
0 5 10 15 20
0.4
0.4
-0.1
0.2 0 -0.2 0
Fig. 2 Real and imaginary part of the shortened real measured channel impulse response.
MSE and CM curves
MSE CM
10
12
14
16
10
Tap SNR=30dB
Normalized Value
0.8 0.6
Magnitude
-2
10
-1
0.4
0.2 0
10
10
12
14
-0.2 0 2 4 6 8 10 12 14 16
Decision Delay
Fig. 3 The best decision delay choice rule.
Tap
ISI =
where
2 P 1 2 2 q(n) max q(n) E j = 0 a j ( n) n 2 max q(n) n
(48)
Fig. 4 Magnitude of the estimated channel under SNR = 20dB and 30dB at 50 trials, the original channel (solid line), the averaged estimated standard deviation (dotted line), and the mean value of 50 estimates (square symbol)
10
1
ISI curves, SNR=25dB

LMS, =0.001 LMS, =0.003 RLS, =0.950 RLS, =0.995
q (n) = gi ( j )hi (n j ).
i =0 j = 0
P 1 L 1
(49)
10
The source symbols are drawn from a 16-QAM constellation with a uniform distribution. The noise is drawn from white Gaussian distribution at a varying SNR. The simulated channel is a length-16 version of an empirically measured T -spaced 2 ( P = 2) digital microwave radio channel with 230 taps, which we truncated to obtain a channel with L = 8 . The Microwave channel chan1.mat is founded at http://spib.rice.edu/spib/ microwave.html. The shortened version is derived by linear decimation of the FFT of the full-length T 2 -spaced impulse response and taking the IFFT of the decimated version (see [16] for more details on this channel). The overall channel impulse response h(n) is shown in Fig. 2. We consider the performance of the proposed RLS and LMS algorithms for blind MMSE equalizer. Let the equalizer order N = 8, and let the delay be d = 7 . In Fig. 3, we show the J MMSE and J CM versus decision delay d under SNR =
ISI
10
-1
10
-2
10
-3
500
1000
1500
2000
Samples
Fig. 5 Comparison of residual ISI for the proposed algorithm under SNR = 25dB.
30dB. Fig. 4 shows 50 estimates of the cross-correlation vector under SNR = 20dB and 30dB, respectively. We have discussed earlier that we can obtain channel coefficient vector if delay equals to channel length. From this figure, we can
10
1
1031
ISI curves
SNR=15dB SNR=20dB SNR=25dB
10
ISI curves, SNR = 25dB

CMA FEPF BPEF Halford Our LMS Our RLS
10
10
ISI
10
-2
ISI
0 500 1000 1500 2000
10
-1
10
-1
10
-2
10
-3
10
-3
500
1000
1500
2000
Samples
Samples
Fig. 6 ISI curves of the proposed RLS algorithm for different SNRs
Residual ISI versus different equalizer order
SNR = 20dB SNR = 25dB 10
0
Fig. 9 ISI comparisons of existing algorithms and the proposed algorithm under SNR = 25dB.
10
10
-1
10
-2
10
-3
10
11
12
Equalizer order
Fig. 7 Residual ISI curves versus the different equalizer order under SNR = 20dB and SNR = 25dB.
Unequalized
2 1
algorithm after 2000 samples for several equalizer orders under SNR = 20dB and 25dB. From this figure, we can conclude that the exact order estimation is not needed in the proposed algorithm. Fig. 8 shows the received constellation and the equalized constellation at SNR = 25dB for 1000 samples. We compare the performance of the proposed algorithm with some existing algorithms: the CMA of [17], least-squares lattice (LSL) one-step forward prediction of [9] (denotes FPEF), LSL one-step backward prediction of [9] (denotes BPEF), and RLS equalizer of [18] (denote Halford). Let the all equalizer orders be N = 8 for fair comparison. Fig. 9 shows the ISI curves for the proposed LMS and RLS algorithms and existing algorithms at SNR = 25dB. It is shown that the proposed algorithms perform better than the others.
Residual ISI
After equalization
V. CONCLUSION
This paper presents blind MMSE channel equalization using multichannel prediction-based cross-correlation vector estimation. A block-type algorithm is developed using the correlation matrices of the received signal. Then, we have developed the RLS and LMS type algorithm for obtaining the multichannel linear prediction error. Furthermore, we have used the iterative power method for tracking the crosscorrelation vector. Our proposed algorithms do not require the exact channel order estimation and are robust to channel order mismatch in nature of linear prediction characteristics. Moreover, cross-correlation vector estimation algorithm can be used for partial or complete channel estimation via delay control. Simulation results show that the proposed algorithms outperform many existing algorithms.
-2 -2 0 2
-1 -1 0 1
Fig. 8 Scatter plots before and after equalization for the proposed RLS algorithm under SNR = 25dB.
obtain the channel estimation from the cross-correlation vector with delay corresponding to the channel length. The performance of the proposed RLS and LMS algorithms are shown in Fig. 5. It is found from the results that the proposed RLS and LMS algorithms and achieve sufficiently low ISI after 1000 samples. Fig. 6 shows the ISI curves of the proposed RLS algorithm under SNR = 15dB, 20dB, and 25dB. Fig. 7 presents the performance of the proposed RLS
REFERENCES
[1] [2] J. G. Proakis, Digital Communication, 4th ed. New York: McGraw Hill, 2000. Z. Ding, R. A. Kennedy, B. D. O. Anderson, and C. R. Johnson, Jr., Illconvergence of Godard blind equalizers in data communication systems, IEEE Trans. Signal Processing, vol. 39, no. 9, pp. 13131327, Sept. 1991.
1032 [3] [4] [5] Y. Li and Z. Ding, Global convergence of fractionally-spaced Godard (CMA) adaptive equalizers, IEEE Trans. Signal Processig, vol. 44, no. 4, pp. 818826, Apr. 1996. L. Tong, G. Xu, and T. Kailath, Blind identification and equalization based on second-order statistics: A time domain approach, IEEE Trans. Inform. Theory, vol. 40, no. 2, pp. 340349, Mar. 1994. L. Tong, G. Xu, B. Hassibi, and T. Kailath, Blind channel identification based on second-order statistics: A frequency-domain approach, IEEE Trans. Inform. Theory, vol. 41, no. 1, pp. 329334, Jan. 1995. E. Moulines, P. Duhamel, J.F. Cardoso, and S. Mayrargue, Subspace methods for the blind identification of multichannel FIR filter, IEEE Trans. Signal Processing, vol. 43, no. 2, pp. 516525, Feb. 1995. K. Abed-Meraim, E. Moulines, and P. Loubaton, Prediction error method for second-order blind identification, IEEE Trans. Signal Processing, vol. 45, no. 3, pp. 694704, Mar. 1997. C. B. Papadias and D. T. M. Slock, Fractionally spaced equalization of linear polyphase channels and related blind techniques based on multichannel linear prediction, IEEE Trans. Signal Processing, vol. 47, no. 3, pp. 641653, Mar. 1999. J. Mannerkoski and D. P. Taylor, Blind equalization using leastsquares lattice prediction, IEEE Trans. Signal Processing, vol. 47, no. 3, pp. 630640, Mar. 1999. J. Mannerkoski, V. Koivunen, and D. P. Taylor, Performance bounds for multistep prediction-based blind equalization, IEEE Trans. Signal Processing, vol. 49, no. 1, pp. 8493, Jan. 2001. D. T. M. Slock and C. B. Papadias, Further results on blind identification and equalization of multiple FIR channels, in Proc. ICASSP, 1995, pp. 19641967. J. Shen and Z. Ding, Direct blind MMSE channel equalization based on second order statistics, IEEE Trans. Signal Processing, vol. 48, no. 4, pp. 10151022, Apr. 2000. X. Li and H. Fan, Direct estimation of blind zero-forcing equalizers based on second-order statistics, IEEE Trans. Signal Processing, vol. 48, no. 8, pp. 22112218, Aug. 2000. S. Haykin, Adaptive Filter Theory, 3rd ed. Englewood Cliffs, NJ: Prentice-Hall, 1996. G. H. Golub and C. F. Van Loan, Matrix Computations, 3rd ed. Baltimore, MD: Johns Hopkins Univ. Press, 1996. J. Endres, S. D. Halford, C. R. Johnson, and G. B. Giannakis, Simulated comparisons of blind equalization algorithms for cold startup applications, Int. J. Adaptive Contr. Signal Process., vol. 12, no. 3, pp. 283301, May 1998. T. R. Treichler, I. Fijalkow, and C. R. Johnson, Fractionally spaced equalizers: how long should they be?, IEEE Signal Processing Mag., vol. 13, no. 3, pp. 6581, May 1996. G. B. Giannakis and S. D. Halford, Blind fractionally spaced equalization of noisy FIR channels: direct and adaptive solutions, IEEE Trans. Signal Processing, vol. 45, no. 9, pp. 22772292, Sept. 1997. B. Yang, Projection approximation subspace tracking, IEEE Trans. Signal Processing, vol. 43, no. 1, pp. 95107, Jan. 1995. P. Common and G. H. Golub, Tracking a few extreme singular values and vectors in signal processing, Proc. IEEE, vol. 78, no. 8, pp. 13271343, Aug. 1990. D. Gesbert and P. Duhamel, Robust blind channel identification and equalization based on multi-step predictors, in Proc. ICASSP, 1997, pp. 36213624. H. Luo and R. Liu, Blind equalizers for multipath channels with best equalization delay, in Proc. ICASSP, 1999, pp. 25112514. T. K. Moon and W. C. Stirling, Mathematical Methods and Algorithms for Signal Processing. Upper Saddle River, NJ: Prentice-Hall, 2000.
IEEE Transactions on Consumer Electronics, Vol. 50, No. 4, NOVEMBER 2004 Kyung Seung Ahn (S02) received the B.S., and M.S. degrees in the Department of Electronic Engineering from Chonbuk National University in 1996 and 1998, respectively. He is now working toward to Ph.D. degree in Department of Electronic Engineering from Chonbuk National University. His current interests, which include signal processing and communication theory, are currently focused on the field of blind channel identification, equalization, OFDM, broadband multiple antenna systems, and wireless communications. Juphil Cho was born in Jeonju, Korea, in 1970. He received the B.S. degree in Information and Telecommunication Engineering from Chonbuk National University, Korea, in 1992, and M.S. and Ph.D degrees in Electronic Engineering from Chonbuk National University, in 1994 and 2001, respectively. Since December 2000, he has been with Electronics and Telecommunications Research Institute (ETRI). Currently, he is senior member of engineering staff at Mobile Telecommunication Research Division. His current research interests are 4G mobile communications, OFDM systems, and signal processing technology in communications. Heung Ki Baik (S80-M83) received the B.S., M.S., and Ph.D. degrees in the Department of Electronic Engineering from Seoul National University in 1977, 1979, and 1987, respectively. From Jan. 1990 to Dec. 1990, he was with the Electrical Engineering, University of Utah as a visiting scholar. He has been with the Division of Electronics and Information Engineering, Chonbuk National University, Jeonju, Korea, since 1981 and is currently a professor there. He is a staff of Electronics and Information Advanced Technology Research Center at Chonbuk National University. His research interests are in signal processing with emphasis on adaptive and nonlinear signal processing. He is a member of the KICS, IEEK, ASK, and IEEE.
[6] [7] [8]
[9] [10] [11] [12] [13] [14] [15] [16]
[17] [18] [19] [20] [21] [22] [23]
182
Channel Estimation Standard and Adaptive Blind Equalizaion

Yuang Lou, Member, IEEE
Abstract - In this paper, we present a novel concept of channel estimation standard (CES) and apply a new CES error criterion to the process of an adaptive blind equalization. It is shown that the establishment of the CES contributes to the development of a practical communication scheme for approaching to the capacity of a high SNR band-limited channel without using a preamble signal training. I. INTRODUCTION HANNONs capacity formula of a band-limited S Gaussian channel indicates the possible performance gain of a digital communication system. Even though a coded-modulation scheme contributes to the significant improvements of a communication system, its basic requirement of an ideal Gaussian channel is hard to reach in practice. Thus, the technique of a combined adaptive equalization and coding scheme is highly desirable in order to combat the IS1 degradaand therefore to achieve a higher performance tion [l] gain. It motivates our research on a new CES error criterion such that the higher performance gain of a communication system can be accessed through an IS1 environment without resorting to a preamble signal training. The block diagram of a digital communication system is shown in Fig. 1. {Ik} stands for a transmitted data sequence. The overall communication system consists of a transmission channel {h} for which the samples of an equivalent low-pass impulse response { h l } are unknown, and an adaptive blind equalizer { c } [5] with a coefficient set { c l } . { n k } rep-
resents a noise process and {yk} is an observable signal sequence at the input port of {c}. {ik) is the sequence of equalized signal at the output port of the equalizer { c } . The adaptive equalizer { c } is followed by a decision device, which slices the equalized signal {fk} and makes the final decision {ik}. The purpose of an adaptive blind equalization process is to recapture the input data sequence { I k } with the minimum IS1 without resorting to a preamble signal training [2 - 51. Once the IS1 is minimized or in general the characteristics of the overall communication system (3) are under control, a combined coding technique can then be applied. Since the characteristics control of an overall communication system is a key factor to a system performance, we shall concentrate our discussion on this topic. The remainder of this paper is organized as follows. A novel concept of channel estimation standard (CES) is proposed in section 1 1 .A convex CES cost function 1 is then constructed. In section 111, a new CES error criterion is developed, analyzed, and applied to an adaptive blind equalization process. Compared computer simulations of the CES adaptive blind equalizers are then presented. Section IV concludes this paper.
P I
{s}
Paper approved by Nikolms A. Zervos, the Editor for Transmission Systems of the IEEE Communications Society. Manuscript received April 13,1992; revised November 25, 1993; May 23, 1994. T h i s research was supported in part by SGA Foundation for Research Development and in part by NASA Research Grant NAG 1-734. Yuang Lou was with the Department of Electrical and Computer Engineering, Northeastern University, Boston, MA 02115, USA, and is now with the Department of Radio Systems, GTE Laboratories, 40 Sylvan Road, Waltham, MA 02154, USA. IEEE Log Number 9410046.
0090-6778/95$4.00
S ( t ) = h(t)6 3 c(t)
Fig. 1. Block diagram of a communication system
1 1 . CHANNEL ESTIMATION STANDARD

At the beginning of this section, the assumptions and definitions used in the CES development are presented.
0 1995 IEEE
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 43, NO. 2/3/4, FEBRUARY/MARCH/APRIL 1995
183
(1) We assume that the input data { I k } to a transmission channel is a QAM signal sequence. The { I k } sequence is ergodic with E(&} =E 0, m d E ( I p I z } = &bk kt Here, the superscript * stands for the complex conjugate, & = E(I Ik 1') is the average power of transmitted signah, and 6 k k ' denotes the Kronecker delta function. (2) The noise process nk is assumed to be additive, white Ganssisn with zero mean and variance U : . It 8s asmmed that {ak} is independent of { I k ) . (3) We assume that the transIlltssion channel { h} is linear, and received signal samples { Y k } are elements of a stationary PFWBS. In the following developments, we define a time update average at the kth iteration such that
impulse response in the unit of input symbol duration T, and { $ k ( m ) } is the estimation set o f the actual impulse response samples {S(m)} at the Cth iteration. Through the signal reeeption process, an adaptive equalizer is self adjusted in order to control the characteristics of the {$} system by minimizing the 1 cost function. Thus, a CES adaptive blidd equalization directly supports the basic requirements In the development of a TCM coded-modulation scheme. Under the situation where we can assume that E { . i i j k } is a reasonable approximation of E{I&} [5], we formulate the estimation of {B(m)} at time t = kT such that
(5)
for k < N o b , where Nob is the finite length of an observation window in time. When k 1 Nob, the time update average with a sliding window is defined as
for m = 0, 1, .,N- 1. The {gk(m)}estimation of ( 5 ) is then substituted into (4) to complete the design of a CES cost function. The convvexity property of the CES cost function 1 is shown in [5].
..
<
>k
=<
>k-1
{'k)
- {'k-N,,c)
Nob
*
(2)
III. CES ADAPTIVE BLIND EQUALIZATION

Since a CES error measure relies on the statistical properties of a transmitted signal instead of a reliable signal reference, it forms the theoretical foundation of an adaptive blind equalization process. When we apply the CES concept and the 1 cost function to an adaptive blind equalization, we realize that, in order to minimize the noise influence and the residual IS1 simultaneously, futither improvements on the 1 cost function is of significant importance [4, 51. Combining the CES cost function el directly with a decision-directed (D-D) MSE cost function, we have
N - 1
Thus, the time update average over a random signal process is still a random signal process. Since the adaptive equalizer {c} and the transmission channel {h} are both linear, the overall communication system 3 = h @ c is also linear. Here, @ denotes the convolution. In the follodring analysis, we assume that is time invariant. Thus, the samples of a system characteristics under the 1/T sampling rate can be expressed as
(3)
where the summation indexed by I goes over the length of rui adaptive equalizer {e}. The concept of a novel estimation standard - channel estimation standard (CES) is illustrated as follows. The CES is the basis o f an error measure to an adaptive aystem, which i s d&ed by the difference between the ideal characteristics and the estimated characteristics, say the system impulse response, of an overall communication system. Thus, a CES cost function 1 can be constructcad as
= 2E{(1-0)
cls(m)-Sl,(m)1'
+a
I .fk-fk 1').
m=O
(6) Parameter a in (6) is a pre-selected constant within the range of 0 5 Q 5 1. Even though 2 shows the merits in its simple design, it presents a weak point of using the D-D symbol { f k ) . The process_of minimizing 2 is affected by the e m r from the (I&}sequence, which results in a larger SNR loss and a slower convergence speed. In order t o improve the dwign of 2, we propose a new CES error criterion e such that
N - 1
= 2E{(1-a)
I S(m)-8k(m) 1'
m=O
+a
I z k - f k 1')
(7)
In (4), (S(m)} is the ideal impulse response samples of the { S I system, N is the length of the { J } system
where a Statistical reference
z k
is set up as [5]
184
t
0
0
Im Ik
0 0
TABLE A
IMPULSE RESPONSE OF TRANSMISSION CHANNEL I
Re
Euclidean Distance
Channel I Impulse Response
MO)
h(1) h(2) h(3) h(4) h(5) h(6)
-0.00467736-jO.0037418 0.003968872+j0.028064163 -0,02245133-jO.09728 0.798893185fj0.48644 -0.20393292+j0.25538 0.045838-jO.06922 -0.01496+j0.0187
TABLE B
Fig. 2. 16-point QAM constellation & its decision regions
IMPULSE RESPONSE OF TRANSMISSION CHANNEL I1
for 0 < a 5 1. Several positive points of 21, are illustrated as follows. (1) Due to the application of Zk, the c error criterion is an efficient utilization of the CES error measure and the Euclidean distance measure. The minimization of c does not rely on the correctness of D-D symbol jk. (2) Whenever the c error measure is minimized, i.e., > k ( m ) -+ S ( m ) ,Z k is converged to f k . Thus, the mode transition is smooth between a converged CES process and the process of an MSE equalization. The convexity property of the c error criterion and the existence of its unique global minimum are investigated as follows. Define the gradient of a real quantity with respect to a complex vector C = ( C - L , . ,q, .,CL)T as
Proof: Since the D-D symbol jk is obtained by the rule of a minimum Euclidean distance on the basis of ik, the distance between any two adjacent reference points in the {fk} set makes the local property of {jk} to be a constant with respect to when Lemma I is true. Q.E.D.
..
..
Theorem I: Under the condition of Lemma I, the c error criterion is a convex U [6] function in the vector C space. Its global minimum is unique. given at C =
cpt
Proof: By Lemma I, we have the gradient of c with respect to C as
and
(9)
(10) for a finite length adaptive equalizer of 2L 1 adjustable complex coefficients. Here, the superscript stands for a vector transpose.
Lemma I: When a signal point i k falls far away from the predetermined decision boundaries, we have the decision symbol f k as a highly nonlinear functioc of the equalizer tap weight vector C with V c -I k = 0.
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL.43, NO. 2/3/4, FEBRUARY/MARCH/APRIL 1995
185
-6.
00
Fig. 3(a) Eye pattern of received signal (Channel I with 16-point Q.4M at 30 dB/symbol)
Fig. 3(b) Eye pattern of received signal (Channel I1 with 16-point QAM at 30 dB/symbol)
Here, the superscript t stands for the conjugate transpose, and y k = ( y k + ~ , ...,Yk,. .. , Y ~ - L )represents ~ the vector of received signal samples. Since the matrix in v~[(o,c)'] is Hermitian and positive definite, the conv&ty of c with respect to C is shown. The global minimum of c can only be obtained at vcr - =0 when C is equal to Copt.
Q.E.D.
Least-mean square algorithms of the c and 2 adaptive blind equalization are developed and discussed in [4, 51. Compared computer simulations between the r and 2 adaptive blind equalizers are shown here for a full response signaling system, i.e., S(0) = 1 and S ( m ) = 0 for all m # 0. The weighting factor is set at a = I and the constellation of a transmitted 2' . QAM signal IS shown in Fig. 2. Characteristics of the two complex channel models are given in Table A and Table B, respectively. At SNR = 30 dB/symbol or equivalently at SNR = 23.9 dB/bit, the eye patterns of received signals from both channel I and channel I1 are totally blurred. By the comparison of Fig. 3(a) and Fig. 3(b), we see that the IS1 distortions from channel I1 is much worse than that of channel I such that a severe IS1 results in a homogeneously blurred eye pattern but not only in a tilted square [5]. Fig. 4 shows the comparison of Monte Carlo simulations. The lower bound of the error rate is obtained from the theoretical calculation of a 16-point &AM data transmission over a Gaussian channel [l],
Three types of adaptive equalizers are under the simulation. The triangle points are for a training MSE equalizer. The dotted and crossed points are from the processes of a CES adaptive blind equalization where the 6 and 2 error criteria are applied, respectively. The comparison shows that there are a small amount of SNR losses between the 6 and 2 processes over the channel I transmission where the IS1 distortion is relatively light. The differences shown in the output SNR become much larger between the c and 2 processes over the channel I1 transmission where the IS1 distortion is much more severe. From these compared computer simulations, we see that the convergence of a CES adaptive blind equalization sets the characteristics of an overall communication system 3 under control such that the average error rate in { f k } and the undesirable IS1 distortions axe minimized simult aneously.
w.SUMMARY A N D CONCLUSION
In this paper, a novel concept of channel estimation standard (CES) is illustratd which defines an error measure between the ideal chatacteristics and the estimated characteristics of an overall communication system. It has been shown that a converged CES error measure supports the basic assumption in the development of a TCM coded-modulation scheme. In the paper, we illustrate the CES concept via its ap-
186
0,
4.00
8.00
12.00
16.00
20.00
24.00
28-00
Signal-to-Noise R a t i o ( S N R / b i t )
Fig. 4. Monte Carlo simulation on 16-point QAM systems
plication to an adaptive blind equalization process.
A CES cost function is defined and its convexity is

shown. The superior performance of a CES adaptive blind equalizer supports our theoretical arguments.
V. ACKNOWLEDGEMENT
The author would like to acknowledge the comments and support from John G. Proakis, M. Schetzen, C.L. Nikias and H. Raemer.
REFERENCES
[l] Proakis, John G., Digital Communications, Second Edition, McGraw-Hill Book Company, 1989. [2] Godard, D.N., Self Recovering Equalization and Carrier Tracking in Two Dimensional Data Communication Systems, ZEEE Transaction on Communications, vol. COM-28, No.11, pp.18671875, November 1980. [3] Lou, Y., Adaptive Blind Equalization Criteria and Algorithms for QAM Data Communication Systems, Proceedings of the 9th International Phoeniz Conference on Computers and Communications, pp.222-229, March 1990. [4] Lou, Y., Comparison of Adaptive Blind Equalizers, ZEEE International Conference on Acoustics, Speech, And Signal Processing, San Francisco, CA, March 1992. [5] Lou, Y., Adaptive Blind Equalization Technique for QAM Data Communications, Ph.D. Dissertation, Department of Electrical and Computer Engineering, Northeastern University, Boston, MA 02115, June 1992. [6] Schetzen, M., The Volterra and Wiener Theories of Nonlinear Systems, Reprinted Edition, John Wiley & Sons, Inc., pp.504509, 1989.
I _ -

Adaptive Blind Channel Equalisation

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Adaptive Blind Channel Equalisation

Uploaded by

Copyright:

Available Formats

Adaptive blind channel equalisation in chaotic communications by using nonlinear prediction technique

Fn1 @F =@X jX X ^ n1 Gn1 @H =@X jX X ^ n1

Table 1: Proposed approach for identifying MA systems

^ (t ) a ^ (t21)2 m(@Vt/@a)jaa a ^ (t21) where m is the step size

The derivative @Vt(a)/@a is @Vt a ^ t a f x ^ t1 a; . . . ; x ^ td a 2x @a & ' @ tL ^ a; . . . ; x ^ td a xt1 f x @a t1

Using (11), we have

After substitution of (13) and (15) into (14), we have

Therefore the step-size m should satisfy

The expression in (18) can be approximated by ^ t1 r1 r0 f 0 x 19

-3 0 our approach -3 5 E r r o r i n E q u a l i z e d D a ta ( d B ) -4 0 -4 5 -5 0 -5 5 -6 0 -6 5 -7 0 -7 5 30 the approach in [6]

100 150 200 250 300 Number of data samples a

Fig. 2 Error in the equalised data, l 4.0, 256 data samples

Number of data samples b

Number of data samples

100 150 200 Number of data samples

0 Chebyshev sequence Logistic sequence -10

and Channel B: at 1 0:45 0:01 sinp t=400

300 400 500 600 700 Number of data samples

Fig. 6 MSE for different driving sequences, SNR 50 dB, l 4.0

300 400 500 Number of data samples

300 400 500 Number of data samples

Fig. 7 Performance of the proposed algorithm in the timevarying channels

Adaptive Blind Channel Estimation by Least Squares Smoothing

1053587X/99$10.00 1999 IEEE

similarly dened, we have, from (1) (3)

Our goal is to estimate C. Assumptions and Properties

consecutive output block row vectors . From (3), we have

Fig. 4. Batch LSS algorithm.

rank 2) Past Data Matrix :

rank 3) Future Data Matrix :

4) Future-Past Data Matrix

Isomorphism between input and output subspaces for

Fig. 7. Main structure of A-LSS.

Lattice lter for multistep prediction.

Fig. 9. Recursive projection and order detection.

Applying Lemma 1 next to , we have

Recursive channel estimation.

TABLE I COMPARED IN THE SIMULATION AND THEIR CHARACTERISTICS

inherent ambiguity was removed before the computation of NRMSE.

Channel order and parameter tracking performance (SNR = 30 dB).

IEEE Transactions on Consumer Electronics, Vol. 50, No. 4, NOVEMBER 2004

0098 3063/04/$20.00 2004 IEEE

v0(n) h0(n) a0(n)

( N + L) 1 , x N (n) and v N (n) are NP 1 vectors as following.

s(n) = [ s (n) L s (n L N + 1)]T

x N (n) = [xT (n) L xT (n N + 1)]T v N (n) = [ v (n) L v (n)]

Fig. 1 The multichannel representation of a T P -spaced equalizer.

xi (n) = hi (n) s (n k ) + vi (n)

We represent xi (n) in a vector for as

where H is a NP ( N + L) block Toeplitz matrix, s(n) is

0 h(0) L h( L 1) L H= M O M O M . 1) (0) ( L 0 L h L h We assume the following in this paper.

s(k )h(t kT ) + v(t )

( n d ) . J MMSE = E s (n d ) s MMSE blind equalizer with delay d is given by

= ai (n) + vi (n), i = 0,L , P 1

x(n) = [ x0 (n) L xP 1 (n)] h(n) = [h0 (n) L hP 1 (n)]

v(n) = [v0 (n) L vP 1 (n)] .

where the cross-correlation vector equals (4)

x(n) = s (k )h(n k ) + v (n).

x N (n) = Hs(n) + v N (n)

IEEE Transactions on Consumer Electronics, Vol. 50, No. 4, NOVEMBER 2004

(25) (26) (27)

Then, minimizing J MMSE yields

s1 (n) x N (n) = Hs(n) = [ H1 H 2 H 3 ] s 2 (n) s3 ( n)