Faisal O. Alayyan 1, Raed M. Shubair 2, Yee Hong Leung 3, Abdelhak M. Zoubir 4, and Omar Alketbi 5 1,2,5Khalifa University of Science, Technology and Research (kustar) Sharjah P.O. Box 573 , United Arab Emirates, Email: alayyan.rshubair@kustar.ac.ae.alketbi46@hotmail.com 3 Curtin University of Technology Perth WA 6845, Australia, Email: Y.Leung@exchange.curtin.edu.au 4Signal Processing Group (SPG), TU Darmstadt Merckstrasse 25, 64283 Darmstadt, Germany, Email: zoubir@ieee. org Abstract- This paper considers the problem of channel iden- tification in OFDM-based SIMO systems. In this contribution, we focus on minimum mean square error (MMSE) algorithm, which exploits the system structure and offers superior estimation performance. Simulation results are provided to illustrate this claim and to assess the performance of the proposed algorithms. Index Terms: OFDM, channel identification, equalization, GI, CP, MMSE, LSF, PRo I. INTRODUCTION Orthogonal Frequency Division Multiplexing (OFDM) is a multi- carrier digital modulation technique that facilitates the transmission of high data rates with a limited bandwidth [11]. Due to increase in the normalized delay spread, multi path fading becomes a major concern as systems with high data rate are more liable to intersymbol interference (lSI). Classically, lSI is eliminated by employing a cyclically extended time domain guard interval (01). Thus, each OFDM symbol is preceded by a periodic extension of the symbol itself. This 01 is also known as cyclic prefix (CP) and the system CP- OFDM [11]. Recently, zero-padding OFDM (ZP-OFDM), which pre- pends each OFDM symbol with zeros rather than replicating the last few samples, has been proposed [11]. ZP-OFDM not only has all the advantages of the CP-OFDM, but also guarantees symbol recovery and ensures finite impulse response (FIR) equalization. However, the implementation of a ZP-OFDM system involves transmitter modifications and complicates the equalizer. Adaptive channel equalization techniques have been widely used in communication systems because of their convenience in real-time applications. More recently, a blind adaptive equalization algorithm based on MMSE estimation, which offers a number of attractive properties such as robustness to channel over-estimation errors and low computational complexity, has been introduced [7], [9]. An extension of the MMSE equalizer for fast varying channels was developed in [8]. Although these proposed algorithms have high identification efficiency, they are computationally very intensive and difficult to implement in real time, especially for large sensor array systems. Therefore, these algorithms are not compatible with future 4G wireless communication systems. In Section II, the OFDM-based SIMO system is reformulated. In Section III, we present subspace-MMSE (SS-MMSE) blind channel identification and equalization method, which exploits the system structure and offers superior estimation performance than [7]. It is shown in [7], that the blind channel identification of the MMSE filter results in a performance loss compared to semi-blind approaches. To alleviate this drawback and control non-zero equalization delay, a two-step technique was introduced [7], [8]. However, the two- step technique is not practical for general SIMO channel estimation. For CP-OFDM systems, the estimation methods [11] rely on certain properties of Circulant matrices. Important aspects of the proposed 978-1-4244-3474-9/09/$25.00 2009 IEEE SS-MMSE algorithm include: it requires only a one-step technique to reconstruct the channel vectors. Although this algorithm is prelim- inary and suboptimal in retrospect, it reformulates a traditional chan- nel identification and equalization problem into a block-parameter estimation framework and eliminates the issue of information loss as reported [7], [8], due to the properties of Sylvester matrices [2]. More importantly, it opens up a new direction and leads to the development of a class of data-efficient and fast converging blind equalization techniques, namely the MMSE algorithm, as discussed in Section III. In previous report works [4], MMSE equalization algorithms were developed. Section IV deals with the linear adaptive implementation of the proposed detectors in a time varying environ- ment. In Section V, we propose an alternative detection criterion by using the property restoral (PR). The method partially copes with the redundancy created by the CP to perform blind channel identification by using a least square fitting (LSF) approach. The resulting LSF- MMSE cost function demonstrates robustness to variation of the frequency selective fading. In Section VI, the performance gain of zero-delay and nonzero-delay are discussed. The effectiveness of the proposed algorithms is demonstrated in Section VII through computer simulations. II. SYSTEM MODEL Consider a baseband discrete-time CP-OFDM system, where the transmitted symbols are parsed into blocks of size N: Sk = [sdO) , sdl) , ... , sdN - 1)f and k = 0,1,2 ... , K - 1. The elements of Sk are considered to be independent and identically distributed (i.i.d). We regard these elements to be in the frequency domain. The symbol block Sk is then modulated and converted into time domain using the IFFT matrix F ~ , where FN has entires fn,d = }exp (i2';.;nd) and d, n = 0, ... , N - 1. The data vector Uk = F NSk = [Uk (0) ,Uk (1) , ... , Uk (N - 1)]T is then appneded with a CP 1 of length L, resulting in a size M = N + L signal vector: iiCp,k = Tcpuk. where Tcp is a concatenation of the last L rows of an N x N identity matrix IN (that we denote as Icp) and the identity matrix itself IN, i.e., Tcp = [Ilp, I ~ r. The channel response is denoted by h(r) (l) where l = 0,1, ... , L, r = 1,2, ... , q and q is the number of receiving antennas. To avoid lSI, as indicated previously, the CP length L is selected to be equal to or greater than the largest channel order. Next we introduce some notations and some basic results needed later. The received k-th block at r-th output for n = 0,1, ... , M - 1, is given by L X ~ ~ , k (n) = L h(r) (l) UCp,k (n -l) + V ~ ~ , k (n) (1) 1=0 lThe CP makes the OFDM appear periodic over the time span of interest. where UCp,k (n -l) and the AWON, V 6 ~ , k (n), is assumed to be mutually uncorrelated and stationary. Definmg the following notations XCp,k (n) VCp,k (n) h (l) [ (1) () (2) ( ) (q) ( )] T Xcp,k n , XCp,k n , ... , XCp,k n [ (1) () (2) ( ) (q) ( )] T Vcp,k n 'VCp,k n , ... , VCp,k n [h(l) (l) ,h(2) (l) , ... , h(q) (l) r we can rewrite the input-output relation (1) in vector matrix as L (2) XCp,k (n) = L h (l) UCp,k (n -l) + VCp,k (n) . (3) l=O Let Ho be the qM x N block-lower triangular Toeplitz matrix with first column [hT (0) , ... ,h T (L) ,OT, . .. ,OT] T and HI be the qM x N block-upper triangular Toeplitz matrix with first block-row [0, ... ,0, h (L) , ... ,h (1)]. Based on the aforementioned results and the fact that h(l) = 0, Vl ~ [0, LJ, (3) can be written in the block form as [11] XCp,k where XCp,k lSI ~ Hoiicp,k + H1iicp,k-l +VCp,k H H HoTcpFNsk + H1TCpFNSk-l +VCp,k '--v-" lSI [XlP'k (0) ,Xlp,k (1), ... ,Xlp,k (M -1)r (4) VCp,k [ T T T ]T VCp,k (0) ,VCp,k (1) , ... , VCp,k (M - 1) . (5) To obtain an lSI-free data block, we consider a truncated version of XCp,k. This is done by discarding [Xlp,k (0) ,Xlp,k (1) , ... ,Xlp,k (L - 1) r with the receive-matrix: Rcp = [OqNxqL,IqN]. The resulting received vector can be written as YCp,k where RCpXCp,k H HcpF NSk + nCp,k [ T T T ]T YCp,k YCp,k (0) 'YCp,k (1) , ... , YCp,k (N - 1) (6) (7) [ T T T ]T nCp,k nCp,k (0), nCp,k (1) , ... , nCp,k (N - 1) . (8) Since RcpHl Tcp = 0, lSI inducing matrix HI has been eliminated completely. Note that Hcp = RcpHo Tcp is a block-Circulant channel matrix, with first block-row [h (0),0, ... ,0, h (L), ... , h (1)] and first column [hT (0), ... ,h T (L), ... ,0Tr. III. MMSE-BASED EQUALIZER Consider a 7-delay MMSE equalizer (7 E {O, 1,2, ... , N - 1}). The equalizer weight vector VT corresponding to the desired solution is given by [7], [8], [9] VT = arg min IIUk-T - VHYkl12 = R;;-lhT IIvll=1 (9) where Rn = E (Ykyf!) is the channel covariance matrix of the received data and hT is an n x 1 vector (where n qN) given by: (10) where H (:,7 + 1) denotes the (7 + 1 )-th column vector of H. Then, we can write Rn as (11) Based on the concept of signal subspace used by many eigen-based algorithms in the context of array signal processing, (9) can be written in terms of the signal subspace S as follows VT = SVT (12) where vT is a N-dimensional vector. For the zero-delay MMSE equalizer, we can write h(O) h (L) o o (13) Obviously, CP-OFDM (or even ZP-OFDM) is capable of recovering v = q( L + 1) channel impulse responses for 7 = 0, ... , N - 1. This is precisely the distinct advantage of the proposed algorithm over [7]. Therefore (13) implies that (14) where Rn is a sub-matrix of Rn given by its last r;, = n - v rows. In practice, since the output data vectors are noisy, (14) is solved in the least square sense according to: vo = arg !llin (vf! QSS-MMSEVO) (15) II voll=1 where (16) This quadratic optimization criterion allows unique estimation of v up to a scale factor, and the estimated weight vector is thus obtained as the eigenvector associated with the minimum eigenvalue of QSS-MMSE' Based on (9), channel parameters corresponding to zero-delay ho is estimated as (17) One of the important advantages of subspace-MMSE (SS-MMSE) algorithm is its deterministic property. That is, the channel parameters can be recovered perfectly in the absence of noise, using only a finite set of data samples, without any statistical assumption over input data. Therefore, SS-MMSE algorithm is promising for applications where only a small amount of output data are available. A typical disadvantage of the SS-MMSE algorithm is that they require a computationally expensive EVD of a data covariance matrix to extract the signal eigenvectors S. Partly due to the computational disadvantage of the EVD-based approach, OFDM systems cannot be compatible with 40 communication systems. However, it is possible to merge the advantages of a MMSE-based equalizer operating in OFDM without using signal subspace. Similarly to (14), we write RnVo = O. (18) Therefore, the estimated weight vector is obtained as Vo = arg min (vf! QMMSEVO) IIvll=1 (19) where (20) The resulting algorithm (MMSE) is much more efficient in computa- tion than the SS-MMSE algorithm and quite sensitive to the frequency selective fading. IV. ADAPTIVE IMPLEMENTATION A major drawback of the SS-MMSE algorithm is its high com- putational cost due to the EVD of a large dimensional matrix requiring O(n 3 ) flops of computations for time varying multi path channels. Estimation of time varying multi path communications for equalisation typically involves the use of fast adaptive algorithms having linear complexity. In this section, we introduce a fast adaptive implementation of the SS-MMSE algorithm. The channel covariance matrix Rn is replaced by its recursive version Rn(i) = (3Rn(i - 1) + Yk(i)y{: (i). (21) at the i-th iteration, where Rn (i) is the sample covariance matrix at the data available up to time i and (3 is the forgetting factor (0 < (3 S 1). In the literature, many subspace tracking algorithms have been developed to compute the signal and for noise subs paces recursively, which require O( n 2 N), O( nN 2 ) and O( nN) flops of computation at each update. In our adaptive implementation, we propose to estimate the signal subspace by the efficient subspace tracker a YET another subspace tracker (YAST) [5] which has a complexity of O(nN). The choice of the YAST algorithm is motivated by its remarkable tracking performance compared to other existing subspace tracking algorithms of similar computational complexity such as projection approximation subspace tracking (PAST) [12] and orthogonal PSAT (OPAST) [1]. The second step of our algorithm consists of estimating recursively the N-dimensional vector v as the least eigenvector of matrix Q S ~ - M M S E or equivalently as the dominant eigenvector of its inverse . Based on the development in [7], we introduce in what follows a low computational complexity algorithm by exploiting the shift-invariance property of the temporal data covariance matrices [6]. Now, using (21), we can derive (22) Initialization S(O) Z(O) Rn(O) Vo t(O) C(O) For each time step do IN [o(n-n)xn I(n-n)x(n-n) a random unitary vector Rn(O)S(O) (tH (O)t(O)) -1 update S( i) using YAST or OPAST algorithms. A(i) N(i) D(i) W(i) B(i) C(i) vo(i) vo(i) t(i) [ IIYk(i)112 -1 -1 0 {3t H (i - l)h (i) [N(i) h(i) 1 1 (32 C(i - I)D(i) (A(i) _ W H (i)D(i))-1 ;2 C(i - 1) + W(i)B(i)WH (i) C(i)vo(i - 1) IIC(i)vo(i - 1)11 S( i)vo (i) (3t(i - 1)+ Yk(i)yt: (i) TABLE I ADAPTIVE BLIND SS-MMSE EQUALIZATION ALGORITHM. where Y k is a sub-vector of Y k given by its last Ii elements. Applying the projection approximation, leads to where W(i) is the N x 2 matrix Rn(i)S(i) ~ Rn(i)S(i - 1) (23) which is valid if the signal subspace S( i) is slowly varying with i. For later reference, consider T(i) = Rn(i)S(i). (24) Substituting (22) into (24) yields T(i) = (3T(i - 1) + Yk(i)y{: (i). (25) Then (16) can be replaced by the following recursion QSS-MMSE(i) = (32QSS_MMSE(i - 1) + D(i)A -l DH (i) (26) where D( i) is the N x 2 matrix (27) and A( i) is the 2 x 2 non-singular matrix -1 ] o . (28) Consider the N x N Hermitian matrix C(i) = QSS-MMSE(i) (29) using the matrix (Schur) inversion [?], we obtian C(i) = ;2 C(i - 1) + W(i)B(i)W H (i) (30) 2QSS-MMSE is a singular matrix when dealing with exact statistics. W(i) = ;2 C(i - I)D(i) and B( i) is the 2 x 2 matrix B(i) = (A(i) _ W H (i)D(i)) -1. (31) (32) The extraction of the dominant eigenvector of C(i) is obtained by power iteration as [7], [8] _. C(i)vo(i - 1) vo(z) = IIC(i)vo(i - 1)11' (33) Thus the MMSE equalizer vector is expressed (up to scalar constant) by vo(i) = S(i)vo(i). (34) A batch-processing implementation of the adaptive blind SS-MMSE equalization algorithm is summarized in Table I. V. LSF-MMSE ALGORITHM Consider Xk = [x[(O),x[(I), ... ,x[(L-l)r to be the k-th restored (equalizer output) OFDM symbol in the frequency domain where xk(m) = [xk 1 )(m),xk 2 )(m), ... ,xk Q )(m)r. Based on the concept of PR algorithm [3], the last L block-samples are identical to the first L block-samples in the symbol, i.e., Xk(m) = xk(N + m), m = 0, ... , L - 1 (35) However, when considering the sample averaged estimate of Rn, the estimate where of the QSS-MMSE is almost surely non-singular matrix due to the estimation (36) errors. According to the previous result, defined in (35), the cost function is expressed as follows K-1 L-1 JLSF-MMSE(g) = L L Ilxk(m) - xk(N + m)112. (37) k=O m=O Using the MMSE concept, (37) can be rewritten explicitly as K-1 L-1 () ""' ""' II H (1) H (2)11 2 JLSF-MMSE g = L...J L...J g Xk - g Xk (38) k=O m=O where = [x[(m), ... ,x[(m - d)]T represents the redundant part of the OFDM symbols and Xk 2 ) = [x[ (N + m), .. . , x[ (N + m - d)]T represents the actual part of the block. Thus, a zero-delay weight vector is formulated as (39) where Rd is a covariance matrix of the (7) with d-window length, and J = Id x n. It is easy to verify that ho is estimated by the following quadratic form I 0 1 2 3 h(/) 1.5 -O.5i O.5i 1.5i TABLE III CHANNEL 2 IMPULSE RESPONSE. Fig. 1. Performance comparison of CP-OFDM estimators using the MMSE, LSF- (40) MMSE and SS-MMSE algorithms: Number of OFDM symbols vs MSE. where QLSF-MMSE = JTRd1 (Xk1) - Xk 2 )) (Xk1) - Xk 2 )) H Rd 1 J. (41) The quadratic optimization criterion allows a unique estimation of the channel up to a complex scaling factor. It is clear that in the LSF-MMSE estimator, the channel impulse response parameter is estimated during time period corresponding to the redundant bands between successive OFDM symbols. VI. EQUALIZATION DELAY Consider the spatiotemporal vector of the signal model (7) as N-1 Yk = L hTuk-T + nk (42) T=O where hT is defined in (10) and represents a sub matrix of H given by the column vectors of indices varying in the range {O, 1, 2, ... , N - I}. Since (43) the performance gain of the nonzero-delay equalizer with (T E {I, 2, ... , N - I}) becomes, in general, negligible once we com- pared to the zero-delay equaliser (T = 0). This condition is not valid for multichannel FIR systems that is proposed in [9] (see [7], [8] for more details), where nonzero-delay equalizers can have much improved performance compared to the zero-delay ones [10]. Unlike the proposed multichannel FIR system [9], OFDM-based SIMO system does not necessarily need nonzero-delay equalizer. VII. NUMERICAL EXAMPLES I 0 1 2 3 hell 3.5+2.7i -O.75+0.5i -1.0-0.25i O.25+0.0i TABLE II CHANNEL I IMPULSE RESPONSE. In this section, we illustrate the merits of our proposed algorithms. symbols drawn from the BPSK constellation. The data plotted were averaged over 200 independent runs. The FIR channels used in the simulation are depicted in Tables II and III. Note that the frequency response of channell involves no frequency nulls. In contrast, the frequency response of channel 2 has a null at frequency bin 2. For OFDM-based SIMO systems (one-transmit and two-receive antenna), channels 1 and 2 are deployed, whereas channel 1 is used for OFDM-based SISO systems (one-transmit and one-receive antenna). To remove the effect of lSI, a CP of length equal to the channel length was inserted. Next, we test our algorithms in different scenarios. Simulation Example 1: Figure 1 shows the proposed performance of the algorithms at SNR=20 dB as a function of the number of OFDM symbols. We see that the performance of the MMSE and SS- MMSE estimators improve with increasing the number of OFDM symbols. Furthermore, large number of OFDM symbols is required -OO.!--7,c" -----o",---7,c" ----c!",---;,cw .. --!". $NR(de) Fig_ 2. Performance comparison of CP-OFDM estimators using the MMSE, LSF- MMSE and SS-MMSE algorithms: SNR vs MSK The figure of merit is the achievable BER averaged over the OFDM Fig_ 3. Performance comparison of CP-OFDM estimators using the MMSE, LSF- data frame compromising of N = 20 samples. We used the signal MMSE and SS-MMSE algorithms: SNR vs BER. Fig. 4. Perfonnance comparison of LSF algorithms via MMSE and ZF equalizers: SNR vs BER. " '. " ;-10
-"":'..:'-h
Fig. 5. Convergence of the adaptive SS-MMSE equalization algorithm. to obtain good channel estimates by the MMSE, and SS-MMSE estimators, In contrast, the LSF-MMSE estimator does not need large number of OFDM symbols and is able to identify the channel coefficients with much smaller number of OFDM symbols. Note that the performance of the LSF-MMSE algorithm is independent of the number of OFDM symbols as long as the identifiability condition is fulfilled. Additionally, as can be seen the SS-MMSE estimator is much more robust to noise and also gives superior performance at large numbers of OFDM symbols. Note that even allowing 1000 OFDM symbols, the MMSE estimators are unable to attain the performance of SS-MMSE estimator that it attains with only 100 OFDM symbols. Simulation Example 2: In this simulation study, we fixed the number of symbols to 200 and varied the SNR from 5-50 dB. Figure 2 shows the MSE as a function of SNR for the estimators. We observe that the performance of the three estimators improves with increasing SNR. In comparison with the MMSE estimator, the LSF-MMSE estimator performs better. Furthermore, the proposed LSF-MMSE performs well for low SNR, whereas a further increase yields a slight improvement in comparison to the other algorithms. High SNR is required to obtain good channel estimates by MMSE estimator, which results in the same performance of the LSF-MMSE estimator. Among the three estimators, the SS-MMSE estimator is much more robust to noise and gives superior performance at high SNR. Simulation Example 3: For the channel estimates obtained in Simu- lation example 2, we investigate the performance of the corresponding multi-channel equalizer. For each SNR value we took the average of the channel estimates over all Monte Carlo runs and computed the corresponding H. Figure 3 shows the BER as a function of the SNR for 0-20 dB. For the purpose of comparisons, the performance of the equaliser with perfect channel estimation is considered. Similar to the previous example increasing SNR yields a significant improvement in the equalizers performance. For low SNR, the system with LSF- MMSE equalizer achieves a performance very similar to that of the system with perfect channel estimation. Simulation Example 4: Figure 4 shows the overall BER performance of the proposed spatial diversity LSF based on MMSE equalizer for a SNR range of 2-10 dB. In order to check the spatial diversity gain, we also simulated the LSF-MMSE without diversity (i.e., SISO). In addition, the LSF algorithm based on the ZF equalizer [3], is applied to both SISO and SIMO systems. The following comments can be made: The ZF equalizers suffer from a performance penalty in all cases compared to the MMSE equalizers. Antenna diversity results in performance improvement. Specif- ically, see the performance gap between the SIMO and SISO receivers employing MMSE and ZF equalizers. Channel 2 possesses a frequency null, but both algorithms for the SIMO and SISO systems prove robust against this. Simulation Example 5: We now examine the tracking performance of the proposed SS-MMSE algorithm in a slowly time varying environment at SNR=15 dB. Figure 5 shows the MSE performance against the number of OFDM symbols for YAST and OPAST algorithms. As we can see, in this experiment, YAST outperforms OPAST. VIII. CONCLUSION This paper introduces blind channel identification and equalization algorithms based on the MMSE equalizer for OFDM-based SIMO systems. Obviously, the SS-MMSE equalizer obtains better results than MMSE and LSF-MMSE equalizers, but we must remark the difference in computational cost. The SS-MMSE equalizer requires an EVD of an n x n autocorrelation matrix, thus requiring O(n 3 ) floating point operations. Based on the adaptive algorithm, linear computational complexity is obtained, which only requires O(nN) flops. Thes proposed algorithms estimates ito without extracting the signal subspace. In CP-OFDM system, any column vector it.,. of the channel matrix H contains sufficient information. Therefore, one- step approach is satisfied and meets the requirements of the wireless communication systems. REFERENCES [1] K. Abed-Meraim, A. Chkeif and Y. Hua, "Fast orthogonal PAST algorithm," IEEE Signal Processing Letters, Vol. 7, pp. 60-62, March 2000. [2] K. Abed-Meraim, P. Loubaton and E. Moulines, "A subspace algorithm for certain blind identifications problems," IEEE Transactions on Signal Processing, vol. 43, no. 2, pp. 499-511, March 1997. [3] F. O. Alayyan, K. Abed-Meraim and A. Zoubir, "Blind equalization in OFDM systems exploiting guard interval redundancy;' in Proceedings of 39th Asilomar Conference on Signals. Systems and Computers, pp. 697-700, November 2005. [4] F. O. Alayyan, K. Abed-Meraim and A. Zoubir, "Blind MMSE Channel Identification And Equalization Algorithms for OFDM Systems," In Proceedings of the 9th International Symposium on Signal Process- ing and its Applications (ISSPA '07), Sharjah, United Arab Emirates (U.A.E.), February 2007. [5] R. Badeua, B. David and G. Richard, "YET another subspace tracker," in Proceedings of the IEEE International Conference on Acoustics. Speech, and Signal Processing (ICASSP), vol. 4, pp. 329-331, March 2005. [6] C. E. Davila, "Efficient, high performance, and subspace tracking for time-domain data," IEEE Transactions on Signal Processing, vol. 48, pp. 3307-3315, 2000. [7] I. Kacha, K. Abed-Meraim and A. Belouchrani, "A Fast adaptive blind equalization algorithm robust to channel order over-estimation errors," IEEE Sensor Array and Multichannel Signal Processing Workshop, pp. 148-152, July 2004. [8] I. Kacha, K. Abed-Meraim and A. Belouchrani, "A new blind MMSE equalizer for MIMO systems;' IEEE 16th International Symposium on Personal, Indoor and Mobile Radio Communications, pp. 1024-1028, July 2005. [9] I. Kacha, K. Abed-Meraim and A. Belouchrani, "Fast Adaptive Blind MMSE Equalizer for Multichannel FIR Systems," EURASIP Journal on Applied Signal Processing, vol. 2006, no. 14827, pp. 1-17, June 2006. [10] J. Shen and Z. Ding, "Direct blind MMSE channel equalisation based on second-order statistics," IEEE Transactions on Signal Processing, vol. 48, pp. 1015-1022,2000. [11] Z. Wang and G. B. Giannakis, "Wireless Multicarrier Communications, where Fourier meets Shannon," IEEE Signal Processing Magazine, pp. 29-48, May 2000. [12] B. Yang, "Projection Approximation Subspace Tracking;' IEEE Trans- actions on Signal Processing, vol. 43, no. 1, pp. 95-107, January 1995.