ML Detection with Blind Prediction for Differential SpaceTime Block Code Systems Seree Wanichpakdeedecha, Kazuhiko Fukawa, Hiroshi Suzuki, Satoshi Suyama Tokyo Institute of Technology 11, Ookayama, Meguroku, Tokyo, 158550, Japan Email: {seree w, fukawa, suzuki, ssuyama}@radio.ss.titech.ac.jp Abstract This paper proposes a maximum likelihood detection (MLD) method for the differential spacetime block code (DSTBC) in cooperation with blind linear prediction (BLP) of fast frequencyflat fading channels. The method, which linearly predicts the fading complex envelope, determines the linear prediction coefficients by the method of Lagrange multipliers, and does not use decisionfeedback nor require information on channel parameters such as the maximum Doppler frequency in contrast to conventional ones. Computer simulations under a fast fading condition demonstrate that the proposed method with a proper degree of polynomial approximation can be superior in BER performance to a conventional method that estimates the coefficients by the RLS algorithm using a training sequence. I. INTRODUCTION Alamouti s spacetime block code (STBC) has attracted much attention because it can achieve full spatial diversity without bandwidth expansion by using two transmit antennas [1. However, STBC requires accurate channel estimation for coherent detection, although it can hardly perform such estimation under fast fading environments. In order to achieve the spatial diversity without channel state information (CSI), differential STBC (DSTBC) was proposed [. In a static channel, however, BER performance of DSTBC is approximately 3 db worse than that of STBC with coherent detection. For improving the performance, some methods including multiplesymbol detection (MSD) were applied to the detection of DSTBC [3, [4. Nevertheless, these schemes assume a timeinvariant channel during each transmission block. To cope with the fast fading, the methods of linear prediction were applied to DSTBC in [5[8. The linear prediction scheme estimates a current complex envelope by using the past ones, and performs the maximumlikelihood sequences estimation (MLSE) to find the sequence that maximizes the log likelihood function. However, this conventional scheme needs information on both the maximum Doppler frequency and signaltonoise ratio (SNR) in order to obtain the optimum prediction coefficients. Another conventional scheme estimates the coefficients by the RLS algorithm using decisionfeedback, which may cause error propagation [9, [10. For solving these problems, this paper proposes blind linear prediction (BLP) for DSTBC that employs the method of Lagrange multipliers to determine the prediction coefficients by using neither the data of decisionfeedback nor information on the channel parameters. b(k) DSTBC Encoder Antenna 1 s(k1),s*(k) Antenna s(k),s*(k1) A. Received signal h 11 (k) h NR(k) Antenna 1 r 1 (k) Antenna N R r NR (k) Symbol Detector Fig. 1 The system model II. SYSTEM MODEL D(k) DSTBC Decoder Fig. 1 shows the system model of DSTBC with two transmit antennas and N R receive antennas. Let T denote the symbol duration. The channel is assumed to be timevarying frequencyflat Rayleighfading, where h (i) is a channel impulse response between the pth transmit antenna (p = 1, ) and the lth receive antenna (l = 1,..., N R ) at discrete time it. The receiver for DSTBC [6 assumes that h (i) is constant during each STBC block (two symbols) as b(k) h (k) = h (k 1), (1) where k is a block index. Note that this assumption is used below except the computer simulation section. It is also noteworthy that it is not satisfied under very fast fading conditions, which causes the receiver considerable intrablock interference [8. r l (i) represents the signal received by the lth receive antenna at discrete time it, and a by1 received signal vector r l (k) is defined as r H l (k) = ( r l (k 1) r l (k) ), () where the superscript H and denote Hermitian transposition and complex conjugation, respectively. On the assumption of (1), r l (k) can be written as r l (k) = D(k)h l (k) + n l (k), (3) where the by1 impulse response vector h l (k) and the by1 noise vector n l (k) are given by h H l (k) = ( h l1(k) h l(k) ), (4) n H l (k) = ( n l (k 1) n l (k) ). (5)
Here, n l (i) is an additive white Gaussian noise at discrete time it with variance σ n and zero mean, and is statistically independent with respect to receive antenna index l. The by STBC symbol matrix D(k) is defined as D H (k) = ( s 1(k) s (k) ), (6) where the by1 transmitted symbol vectors s 1 (k) and s (k) are given by s H 1 (k) = ( s (k 1) s (k) ), (7) s H (k) = ( s(k) s(k 1) ). (8) Here, s(i) represents a complex symbol of N d PSK modulation. s 1 (k) and s (k), which are orthogonal to each other and have unit lengths, correspond to transmitted signals at discrete time (k 1)T and kt, respectively. Note that the power of each transmitted symbol is fixed to 1/ in order to keep the transmit power per symbol constant (= 1). B. DSTBC encoder and decoder The DSTBC encoder in Fig. 1 performs differential encoding with the STBC format, which can be expressed as [ s T 1(k) = A(k)s T 1(k 1) + B(k)s T (k 1) = ( A(k) B(k) ) D(k 1), (9) where the superscript T denotes transposition. The complex vector ( A(k) B(k) ) depends on N d information bits and satisfies the constraint that the elements of s 1 (k) should be a N d PSK constellation symbol. Note that s 1 (0) can be arbitrarily set, e.g. to ( 1/ 1/ ). The DSTBC decoder in Fig. 1 detects ( A(k) B(k) ) from ˆD(k 1) and ˆD(k) which the symbol detector provides as estimated values of D(k 1) and D(k). The detection can be expressed as (Â(k) ˆB(k)) = ŝ T 1(k)ˆD H (k 1), (10) where Â(k), ˆB(k), and ŝ1 (k) are estimated values of A(k), B(k), and s 1 (k), respectively. The derivation of (10) uses the property that D(k) is an unitary matrix. From (Â(k) ˆB(k)), the N d transmitted information bits can be determined. A. Multipath propagation III. CHANNEL MODEL Under the multipath propagation, h (i) can be written as h (i) = d a,d e j(πf DT cos θ,d )i, (11) where a,d and θ,d represent a complex envelope and incident angle of the dth propagation path, respectively. f D is the maximum Doppler frequency and is generally much less than 1. B. Autoregressive (AR) process On the assumption that h (i) of (11) includes M dominant paths, h (i) has M dominant tones and then can be modeled as an AR process of order M. Considering this model and the assumption of (1), h (k) is expressed as h (k) = c,m h [(k m) + v (k), (1) where c,m is equal to an AR parameter multiplied by 1, and v (k) that represents residual propagation paths is assumed to be a negligible whitenoise process. Next, let us consider Taylor series expansion of h [(k m) that is given by ( m) q h [(k m) = h (q) (k), (13) q! where h (q) (k) is the qth derivative of h (k) and is derived from (11) as h (q) (k) = (f DT ) q[ a,d (jπ cos θ,d ) q e j4πf DT k cosθ,d. d (14) Evidently, h (q) (k) is proportional to (f DT ) q and exponentially decays as q increases. Let us assume that h (q) (k) with q > q 1 can be neglected. The integer q 1 is referred to as a degree of polynomial approximation. Substituting (13) into (1) and neglecting h (q) (k) with q > q 1 and the white noise v (k) yield h (k) q 1 [ M c,m ( m) q (q!) 1 h (q) (k). (15) The conditions for any h (q) (k), 0 q q 1, to satisfy (15) are given by c,m = 1, (16) c,m ( m) q = 0, q 1 1, 1 q q 1. (17) Since (16) and (17) are independent of indices l and p, we assume that c,m = c m for all l and p. Accordingly, (1), (16), and (17) can be rewritten in simple vector forms as h l (k) = c m h l (k m) + v l (k), (18) c H b 0 = 1, (19) c H b q = 0, q 1 1, 1 q q 1, (0) where the Mby1 prediction coefficient vector c and the M by1 vector b q are defined as c H = ( c 1 c c M ), (1) b H q = ( ( ) q ( 4) q ( M) q), 0 q q 1. ()
v l (k) is the by1 process noise vector having v (k) as its elements. A. MLSE IV. SYMBOL DETECTION The symbol detector in Fig. 1 performs MLSE on the basis of the linear prediction. The log likelihood function is derived from (3) as follows: First, multiplying the both hand sides of (3) by D 1 (k) results in r 1 (k) r (k) H D n (k) 1st receive antenna nd receive antenna h 1 (k) h (k) B n (k) VA {D(k)} D 1 (k)r l (k) = h l (k) + D 1 (k)n l (k). (3) Since D(k) is an unitary matrix, this can be rewritten as D H (k)r l (k) = h l (k) + ñ l (k), (4) where the by1 vector ñ l (k) is defined as ñ l (k) = D H (k)n l (k). (5) The autocorrelation matrix of ñ l (k) is given by ñl (k)ñ H l (k) = n l (k)n H l (k) = σ ni, (6) r NR (k) N R th receive antenna h NR (k) (a) Structure of symbol detector D H (k)r l (k) c 1 c c M where denotes ensemble average and I is the by identity matrix. It is also evident from (5) that ñ l (k) is statistically independent with respect to both time index k and receive antenna index l. Therefore, the log likelihood function can be expressed as [6 L = l(r l s) = D H n (k)r l (k) ĥ l (k), (7) k where D n (k) is a candidate of D(k) and ĥ l (k) is an estimate of the impulse response vector h l (k). B. Blind linear prediction (BLP) 1) Applying linear prediction to MLSE: Let us discuss the estimation of ĥ l (k) by linear prediction. Considering (18) and (4), ĥ l (k) can be approximated as ĥ l (k) = c m D H n(k m)r l (k m). (8) Substituting (8) into (7) yields the branch metric given by B n (k) = DH n(k)r l (k) c m D H n(k m)r l (k m). (9) A block diagram of the symbol detector using this branch metric is shown in Fig. (a). For all l, r l (k) is multiplied by D H n(k) which the Viterbi algorithm (VA) processor provides. The resultant is fed into the linear predictor that generates ĥ l (k) following (8) and of which structure is depicted in Fig. (b). Squaring D H n(k)r l (k) ĥ l (k) and combining the resultants with respect to l yield the branch metric, which is passed into the VA processor. The VA processor searches the sequence of D n (k) that maximizes the log likelihood function of (7) by VA, and + (b) Structure of linear predictor Fig. Symbol detector outputs the resultant as the detected signal ˆD(k). VA considers sequences of D n (k) as those of states, and can effectively find the maximum likelihood sequence on a trellis diagram. In this case, the state at discrete time kt is expressed as {s 1n (k 1), s 1n (k ),..., s 1n (k M)} where s 1n (k) is a candidate of s 1 (k). Therefore, the numbers of states and state transitions are equal to MN d and N d, respectively. Fig. 3 shows the trellis diagram for BPSK (N d = 1) with M =. ) Prediction coefficients: Conventional methods determine the prediction coefficients c by solving YuleWalker equation [7 or using the RLS algorithm [9 which require information on and SNR, or detected symbols. The proposed blind method can determine the coefficients without detected and training symbols and any channel information as follows: A by1 prediction error vector e l (k) is defined as e l (k) = D H (k)r l (k) c m D H (k m)r l (k m) = [ h l (k) c m h l (k m) [ + ñ l (k) h l (k) c m ñ l (k m). (30) When D n (k) = D(k) for all k, the branch metric of (9) is given by B n (k) = e l (k). (31) Since e l (k) can be considered a virtual noise, minimizing e l (k) results in the optimal BER performance. To de
s(k),s(k1),s(k),s(k3) 1,1,1,1 1,1,1,1 1,1,1,1 1,1,1,1 1,1,1,1 1,1,1,1 1,1,1,1 1,1,1,1 1,1,1,1 1,1,1,1 1,1,1,1 1,1,1,1 TABLE I Simulation condition Packet format 18 symbols Constellation mapping BPSK, QPSK Receive antennas 1 branch Channel Fast flat Rayleigh fading 1 1 TABLE II Proper value of the degree of the polynomial approximation (q 1 ) corresponding to prediction order (M) M q 1 1 3 1 4 1, 1,1,1,1 1,1,1,1 1,1,1,1 1,1,1,1 Fig. 3 Trellis diagram for BPSK with the prediction order of (M = ) crease e l (k) as much as possible, the proposed method forces the first term of (30) to be zero, and then minimizes the mean squared norm of the second term. This is because the first term can be zero due to the property that h l (k) is a deterministic process, whereas the second term cannot vanish due to the property that ñ l (k) is a stochastic process. After forcing the first term to be zero, el (k) is expressed as e l (k) = σ n ( 1 + c m ). (3) Minimizing (3) is equivalent to c m = c H c min. (33) The conditions for the first term to vanish are given by (19) and (0). Thus, the above minimization is equivalent to minimizing (33) under the constraints of (19) and (0). By applying the method of Lagrange multipliers, this problem becomes equivalent to obtaining c that minimizes the following cost function f(c): f(c) = c H c + λ 0 (1 c H b 0 ) + λ 0(1 b H 0 c), for q 1 = 0 c H c + λ 0 (1 c H b 0 ) + λ 0(1 b H 0 c) q 1 λ q c H b q q 1 λ qb H q c, for q 1 1 q=1 q=1 (34) where {λ q } is a set of Lagrange multipliers. By taking the derivatives of f(c) with respect to c and {λ q }, and equating its results to zero, the optimal c is derived as c = c = q 1 (B 1 q 1 ) q+1,1 b q, (35) where (B q1 ) iu = b H i 1 b u 1, i, u = 1,..., (q 1 + 1). Note that the optimal c does not depend on any channel characteristics and varies according to q 1. It is also noteworthy that the solution with M = and q 1 = 1 is equivalent to linear prediction coefficients for SISO in [11. A. Simulation condition V. COMPUTER SIMULATIONS A series of computer simulations was conducted to evaluate the performance of BLP. The simulation condition is summarized in Table I. For comparison, the conventional methods, the solution of YuleWalker equation using information on SNR and, and the RLS algorithm using N T S training symbols instead of detected ones, were also evaluated. Note that the RLS algorithm using only detected symbols [9, [10 is obviously inferior to that using the training in BER performance, owing to the error propagation. B. of BLP Fig. 4 shows the effects of on BER performance of BLP with BPSK and M = 3. It can be seen that increasing q 1 can improve the tracking performance whereas it degrades the BER performance under low SNR. This is because larger q 1 makes the approximation of (15) more accurate but increases the number of constraints of (0), which enhances the noise power of (3) and causes the degradation in low SNR region. Note that the BER performance of BLP with q 1 = at high SNR degrades due to the intrablock interference. Considering this tradeoff, the proper values of q 1 from = to 5 10 are listed in Table II. C. Comparison with other methods Figs. 5 (a) and (b) show average BER performance versus average E b /N 0 of BLP and the conventional methods with = 0.0 for BPSK and QPSK, respectively. The result with the solution of YuleWalker equation can be considered as a lower bound of the other methods. It can be seen that BLP is superior to the RLS method using the training in BER performance when average E b /N 0 is greater than 0 db and that the BER of BLP is very close to the lower bound. It is also found that the BER with QPSK is worse than that with
M = 3 BPSK = 0.0 q 1 = 1 10 Known SNR and RLS BLP, q 1 =1 N TS = 4 16 3 0 10 10 (a) Average E b /N 0 = 10dB M = 3 BPSK 0 5 10 15 0 5 30 35 40 E b/n 0 (db) (a) BPSK = 0.0 10 q 1 = 0 10 Known SNR and RLS N TS = 4 16 3 1 BLP, q 1 = 1 10 (b) Average E b /N 0 = 40dB Fig. 4 of blind linear prediction (BLP) versus with the polynomial approximation degree q 1 as a parameter 0 5 10 15 0 5 30 35 40 E b/n 0 (db) (b) QPSK Fig. 5 versus average E b /N 0 of proposed and conventional methods with the prediction order M = BPSK. The reason is that the fading fluctuation often causes two bits error per symbol with QPSK while it causes at most one bit error per symbol with BPSK. VI. CONCLUSION ML detection with BLP for DSTBC on a fast fading channel has been proposed. BLP can determine the linear prediction coefficients without neither decisionfeedback nor information on the channel parameters in contrast to the conventional ones. Computer simulation has shown that the BER performance of BLP with a proper degree of the polynomial approximation is superior to that of the conventional RLS method using the training method. REFERENCES [1 S. M. Alamouti, A simple transmit diversity technique for wireless communications, IEEE J. Select. Areas Commun., vol. 16, no. 8, pp. 14511458, Oct. 1998. [ V. Tarokh and H. Jafarkhani, A differential detection scheme for transmit diversity, IEEE J. Select. Areas Commun., vol. 18, no. 7, pp. 11691174, July 000. [3 D. Divsalar and M. K. Simon, Multiplesymbol differential detection of MPSK, IEEE Trans. Commun., vol. 38, no. 3, pp. 300308, Mar. 1990. [4 C. Gao and A. M. Haimovich, Multiplesymbol differential detection for spacetime block codes, in Proc. Conf. Information Sciences and Systems, Princeton, Mar. 00. [5 F. Adachi, Adaptive differential detection using linear prediction for Mary DPSK, IEEE Trans. Veh. Technol., vol. 47, no. 3, pp. 909918, Aug. 1998. [6 E. Chiavaccini and G. M. Vitetta, Further results on Tarokh s spacetime differential technique, in Proc. IEEE ICC 0, vol. 3, pp. 1778178, Apr. 00. [7 P. Tarasak, H. Minn, and V. K. Bhargava, prediction receiver for differential spacetime block codes with spatial correlation, IEEE Commun. Lett., vol. 7, no. 11, pp. 543545, Nov. 003. [8 P. Tarasak, H. Minn, and V. K. Bhargava, Improved approximate maximumlikelihood receiver for differential spacetime block codes over Rayleighfading channels, IEEE Trans. Veh. Technol., vol. 53, no., pp. 461468, Mar. 004. [9 R. Schober and W. H. Gerstacker, Decisionfeedback differential detection based on linear prediction for MDPSK signals transmitted over Ricean fading channels, IEEE J. Select. Areas Commun., vol. 18, no. 3, pp. 39140, Mar. 000. [10 R. Schober and Lutz H.J. Lampe, Noncoherent receivers for differential spacetime modulation, IEEE Trans. Commun., vol. 50, no. 5, pp. 768777, May 00. [11 K. Fukawa and H. Suzuki, Recursive least squares adaptive algorithm maximum likelihood sequence estimation: RLSMLSE an application of maximum likelihood estimation theory to mobile radio, IEICE Transaction BII, vol. J76BII, no. 4, pp. 014, Apr. 1993.