Orthogonal Space-Time Block Coding over Time-Selective Fading Channels: a PIC Detector for the Gi Systems

F.-C. Zheng School of Electrical Engineering Victoria University of Technology Melboume, Australia (Email: fzheng@ieee.org)
Abstract' - All the orthogonal space-time block coding (0-STBC) schemes are based on the following assumption: the channel remains static over the entire length of the codeword. However, time selective fading channels do exist, and in such case the conventional 0STBC detectors can suffer from a large error floor in the high signal-to-noise ratio (SNR) cases. This paper addresses such an issue by introducing a parallel interference cancellation (PIC) based detector for the coded systems (i = 3 and 4).

A . G. Burr Department of Electronics The University of York York, UK (Email: alister@ohm.york.ac.uk)

inter element interference (IEI). The end effect of all this is an irreducible error floor in the bit error rate (BER) curves in the high signal-to-noise ratio (SNR) region. To suppress such an error floor in the 2-Tx case, an elegant decoder was presented in 141. Since the decoder structure in [4] cannot be used for the 3- or 4Tx case, an effective yet simple zero forcing (ZF) scheme for the 8,and G4 systems was proposed in [6] and 171. This paper proposes an altemative approach to the ZF detector in [6] and [7]. Based on the principle of parallel interference cancellation (PIC), the new detector (termed "PIC detector") offers an even better performance than the ZF detector. The computational complexity of the PIC detector is higher than that of the conventional and the ZF detector, but is still very affordable. Only the 4, coded systems are considered in this paper (the H,coded systems will be addressed in [SI).
11. The System Model for t h e

I. Introduction
Over the past five years, orthogonal space-time block coding (0-STBC) technology has attracted enormous interest due to its high diversity order and low decoding complexity [1]-[4]. The low decoding complexity of 0STBC is directly due to the linear maximum likelihood (ML) decoder at the receiver. The linear ML decoder, however, relies on the so-called "quasi-static channel" assumption: the channel remains static over the length of the entire codeword: 2Ts for the two transmit antenna (2-Tx) STBC (i.e., G? -the Alamouti code [I]), 4T, for the H i systems, and 8T, for the systems (i = 3 for 3-Tx, i = 4 for 4-Tx, and T, is the symbol period). While such an assumption is reasonable in most cases, time selective or fast fading channels do exist in practice, even for the 2-Tx case (see [4][5] and the references therein). In these scenarios, the channel state varies from symbol to symbol. Clearly, the 3- and 4-Tx 0-STBC cases (especially the Gi systems) are much more vulnerable to channel variation than the 2Tx case due to the much longer STBC codeword. The above channel variation will destroy the orthogonality of the channel matrix and therefore cause
*Thiswork was supported by the UK Engineering and Physical
Sciences Research Council (EPSRC) under Grant GRR94428.

G4System

Consider a typical G4 encoded 4-Tx 0-STBC system with 4 transmit (4-Tx) and 1 receive (1-Rx) antennas. A group of four complex symbols, sl, s2,s : ~ and , s4, are passed through a G4 encoder before transmitted over ST,. The encoder output is therefore a 8 x 4 matrix C = [c~,], where c,, is & s ~ , o r*SE (conjugate of s k ) , and is transmitted by Tx j at time i. Also, by letting the channel gain from Tx j to the Rx at time i b e h,(z), the received signal at time i is

,=1

where n., is a complex additive white Gaussian noise (AWGN) with zero mean and a variance of U: (therefore 0 . 5 ~ per : dimension). Also, h,(i) is subject

0-7803-8255-2/04P$Z4l.00 0 2 0 0 4IEEE.

389

to Rayleigh fading but is normalised, i.e., Re[hj(i)], Im[hj(i)] N(0, 0.5).

U:

= 1, or

'

The "quasi-static channel" assumption in all 0-STBC schemes requires that hj(i)= constant over the entire codeword length. This, in the case of G4 and G3 systems, means that the channel remains static over It has been shown in [6][7][14] that even under normal vehicle speeds, the assumption of "quasi-static channel" may not hold for the 4-Tx STBC (e.g., & ' and 8,) systems. As such, this paper assumes that the channel is static over T, only and from one to the next, it is time variant.Clearly, this is a much more general and realistic model with the quasi-static channel now becoming a special case. Perfect channel state information (CSI) is assumed in this paper. For the estimation of CSI, please see [ 5 ] [ 9 ] - [121.

At the receiver end, the conventional 0-STBC detector assumes that the channel is time invariant over the entire S l ; period. Regardless of any channel variation, the following expression is effectively employed:

r = Hs + n ,

(5)

Xr9.

where H is the estimated channel matrix for the "quasistatic channel":

rs

111. The Conventional 0 - S T B C Detector for

G*

For symbol group s = [SI,s2, s j , sd]'. for 94 [2] is

SI
-s2

the code matrix

32

s:%
- s1

It can be proved that the "linear maximum likelihood (ML)" G, detector in [3] is equivalent to the following two-step procedure [6][7]: -H [Step C l ] Apply linear transform Q = H to the received signal r:
r' = Qr = QHs + v
'

SI
81

-s:$

s1
s2

-SA

(7)

C=

-s*

- ss
s; s; s ;
-s;

si
s;
'

sf
-s;

s ; -s;

(2)

[Step CZ] Carry out the "linear ML" detection:

- -s;

-sj

s;

s; -s;
s;

X7 = arg{ minl[r'];
rmtS

- Ajsrnl*}, i =

1,2,3,4, (8)

s;

Here, v = a n ,

S is the symbol alphabet, [r'Ii the ith

4 j= 1

From Eq.( I), the "manipulated received signal vector" can then be written as r = [ T I , ... , r:,; rf, ..., r=Hs+n, where the channel matrix (ChM) (3)

(i.e., the ith element element of r', and A; = 2xlhj/2 of diagonal QH).

In reality, however, the true physical process in Step C1

is

r' = Qr = QHs

+v .

(9)

Note that in general X V H in Eq.(9) is non-diagonal:

= QH = [&]4x4.

(10)

The conventional 2-step linear detection procedure is truly ML if and only if the channel is truly static over 8Th,in which case H = H and thus $ , ; = 0 for i # j , and 4;t = X i .

390

For a time selective fading channel, however, H # 6 and thus 4 i j # 0 for i # j . Physically, this leads to inter element interference (IEI). The value of the nondiagonal 4ij depends upon the time-selectivity of the channel. When using the above conventional detector (thus the 84 decoder in [3]) for a time selective fading channel, these non-zero &'s are effectively ignored, resulting in extra detection errors in addition to those caused by the AWGN. These extra errors will form an irreducible error floor in thc BER curves in the high SNR region.

Eq.(9). As is well known, PIC (although suboptimum) is an effective yet simple approach in multi-user detection of CDMA (there is a rich literature on PIC for CDMA. See e.g. [ 151 and all the references therein).

/ZnitialisationJ Set iteration number k = 0, and obtain s/O)from the conventional 0-STBC decoder via Eqs. (7) and ( 8 ) : = 2,.

)'IS

[Iteration] For iteration number I; = I , 2, ..., I ,

,'(a)

= r'

- QN0&l)

(13)

IV. The PIC Detector for G4

Although & # 0 (for i # j) in Eq.(lO), it is also true that normally 1dijI << 14ziI.This is because under normal vehicle speeds or Doppler spread, the channel variation over 8T' still tends to be relatively small. As an example, let us consider the following popular AR(I) model for time-selective channels [4]-[ 111:

The symbol detection for the current iteration can then be achieved via a simple least square approach:

h,(i

+ m ) = cY,hj(i) + w3(i+ m ) ,

where i = 1, 2 , 3, 4, S is the symbol alphabet, and [r'(')]i is the ith element of d k )Also, . d'') = [sj"); ( 4 (k) (a)
s2

(11)

, 83

S?

1'.

is another i.i.d. complex Gaussian random where wj(i) variable having zero mean and variance CT; and being statistically independent of hj(i). Also, am= E[hj(i)hj(i m)]= Jo(27rmf,1TS), where f d is the Doppler frequency and Jo(.) is the 0th order Bessel function of the first kind. If the fading paths originate sufficiently far away from the receiver, a,,, can he assumed to be the same for all the transmitter antennas. To illustrate the dominance of &, similarly to [14], we introduce the following p factor:

As in a CDMA PIC detector, the above procedure comprises two components: tentative symbol estimation (TSE) and tentative interference subtraction (TIS). With the iterations progressing, the TSE will contain fewer and fewer errors, making the TIS more and more accurate. Once the IEI related errors have been eliminated, Eq.(14) becomes a linear ML procedure. This explains why the PIC detector potentially offers a much better performance than the corresponding ZF detector. As to the number of iterations, our simulations have shown that I = 3 iterations normally deliver most of the gain.

where (1A((, denotes the Frobenius norm of A, and matrix CPND comprises the non-diagonal elements of CP: aND = CP - diag(CP). Clearly, the above p factor reflects the ratio of the squared magnitude of the nondiagonal elements to the squared magnitude of the diagonal elements in CP, and its value is dependent upon the Doppler spread of the channel. The relation between p and fdTs is demonstrated in Fig. 1, which is obtained using 40,000 H and 6 realisations. It is easy to see that for the normal range of Doppler spreads we always have II << 1. Based on the above observation, we can now apply the principle of parallel interference cancellation (PIC) to

[Quasi-static channels and algorithm evolution] When the channel is indeed static over the entire 8T, period (i.e., H = E), the Initialisation stage (i.e. the conventional 0-STBC detector) will give the optimum solution. In such a special situation, as CPND = 0 and r$ii = Xi, the iterations in Eqs. (13) and (14) will not alter the optimum initial solution. To this extent, the PIC detection algorithm represents an evolution of the conventional 0-STBC detector. [Algorithm complexi@] Compared with the conventional &decoder, the main computation increase for the PIC detector is from the calculation of matrix CP and the extra iterations in Eqs. (13) and (14). It is easy to show that the total increase involves

0-7803-8255-2/04/$20.M) Oux)4 IEEE.

391

[32 A 4 (3 A . l ) I ] complex number (CN) multiplications/symhol and [28 A[ + 311 CN additions/symbol, where M is the modulation level and I is the number of iterations. Therefore, the computational complexity of the PIC detector is higher than that of both the conventional and the ZF decoders [6][7]. As I is normally small ( 5 3), however, the extra computation in the PIC detector is still moderate (compared with the full ML search over all 4 symbols, whose complexity is O(M')), and can well he justified by the enormous performance improvement.

+ + +

the same observation also applies to the GJ coded 3-Tx systems.

VI. Conclusions
This paper has presented a PIC based detector structure for the & ' 0-STBC systems over time-selective fading channels. While the conventional & ' detectors under such conditions tend to suffer from a considerable irreducible error floor in the high SNR cases, the PIC detector shows no error floor at all. The PIC'Srelatively higher computational cost can he justified by its enormous performance gain.
0.08

[The G j encoded systems] By setting IL~(Z) = 0 in Eq.(l) and all the other related equations, the above PIC detector can directly be used in the E, encoded systems.

. .

V. Simulations
The G4 system under 16-QAM (Gray encoded) modulation and the time selective fading channel in Eq.(lI) are employed. The signal to noise ratio (SNR) at the receiver is defined as SNR = (MT~zE.9)/~: = MTES/u: (since U% = l ) , where E , is the Tx power at each antenna, and A& = 4 (for the & ' system). Also, the UMTS symbol rates are considered: 7: = SF/(3.84 x 10') seconds with

f = 2 GHz.
fdTe'lOO

Five vehicle speeds are simulated V = 0 (quasi-static channels), 70, 100, 130, and 160 km/h (leading to different cy, values) for SF = 128 (or equivalently V = 0, 35, 50, 65, and 80 km/h for SF=256). These correspond to the f d T s values of 0, 0.0043, 0.0062, 0.0080, and 0.0099 (see Fig. 1 for the p values). For the case o f f & = 0.0099, the BER details of the PIC for I = 0 (i.e., the conventional decoder), 1, 2, and 3 arc shown in Fig. 2. Clearly, two or three iterations deliver most of the performance gain even for such a high speed case. For all the other speeds, only the BER results for I = 3 are plotted in Fig.3. For comparison, the results of the conventional G4 detector are shown in Fig. 4. It is easy to see that the PIC detector exhibits no error floor while the conventional G4 detector does. Most importantly, the BER degradation of the PIC caused by channel variation is very small indeed within the considered f,,T, range. The penalty, however, is a relatively higher computational complexity. In addition, our simulations (not shown here) have indicated that

Fig. 1 The value of printed as "mu" here) with respect to f,jT6.

10

15

20

25 SNR

30

35

40'

45

Fig. 2 The BER performance of the PIC & ' detector with respect to the number of iterations.

10

15

20

25

30

3 5

40

45

SNR .

Fig. 3 The BER performance of the PIC Gd detector

for I = 3.

[2] V. Tarokh, , H. Jafarkhani, and A.R. Calderhank. "Spacetime block codes from orthogonal designs," IEEE Trans. Inform. Theow, vol. 45, no. 5, pp. 1456-67, July 1999. [3] V. Tarokh, H. Jafarkhani, and A.R. Calderhank, "Spacetime block coding for wireless communications: Performance results," IEEE J. Selected Areas in Commir., vol. 17,no. 3,pp. 451460,March 1999. [4] T. A. Tran and A.B. Sesay, "A generalised simplified ML decoder of orthogonal space-time block code for wireless communications over time-selective fading channels," Proc. ofIEEE VTC2002-S, pp. 1911-5, May2002. [5] Z. Liu, X. Ma, and G.B. Giannakis, "Space-time coding and Kalman filtering for time-selective fading channels," IEEE Trans. Commu., vol. 50, no. 2, pp.183-6, Feb. 2002. [6] F.-C. Zheng and A.G. Burr, "Receiver design for orthogonal space-time block coding for four transmit antennas over time-selective fading channels," Proc. of IEEE Globecom 2003, San Francisco, USA. [7] F.-C. Zheng and A.G. Burr, "A zero-forcing detector for orthogonal space-time block coding over time-selective fading channels: the Gi systems," submitted to IEEE
Trans. Wireless Commir.

10

15

XI

25 SNR

30

35

40

45

Fig. 4 The BER performance of the conventional

Ed

detector.

References
[l] A.M. Alamouti, "A simple transmit diversity techniques for wireless communications," IEEE J. SAC, vol. 16, no.

8, pp.1451-8, October 1998.

[SI F.-C. Zheng and A.G. Burr, "Orthogonal Space-Time Block Coding over Time-Selective Fading Channels: a PIC Detector for the Hi Systems," Proc. o f IEEE ICC 2004, Paris, France. [9] R.A. Iltis, "Joint estimation of PN code delay and multipath using the extended Kalman filter," IEEE Trans. Commu., vol. 38, no. 10, pp.1677-85, 1990. [IO] A. W. Fuxjaeger and R. A. Iltis, "Adaptive parameter estimation using parallel Kalman filtering for spread spectrum code and Doppler tracking," IEEE Transactions on Commu., vol. 46, no. 5, pp. 2227-2230, June 1994. [ l l ] M.K. Tastsanis, G.B. Ginannakis, and G. Zhou, "Estimation of fading channels with random coefficients," SignalProcessing, vol. 53, no. 213, pp. 21 1-229, 1996. [12] M. Enescu and V. Koivunen, "Time-varying channel tracking for space-time block coding," Proc. o f IEEE VTC 2002-S, Birmingham, AL, USA, May 2002. [13] 0. Tirkkonen, A. Boariu, and A. Hottinen. "Minimal non-orthogonal Rate 1 space-time block code for 3+ Tx antennas," Proc. of IEEE 6th ISSSTA, NJIT, NJ, USA, pp. 429-432, 6-8 September 2000. [I41 F.X. Zheng and A.G. Burr, "Space-time block coding for four transmit antennas over time-selective fading channels: Orthogonal or non-orthogonal design?'?", Proc. ofIEEE VTC2003-F, Orlando, USA. [IS] D. Divsalar, M.K. Simon, and D.Raphaeli, "Improved parallel interference cancellation for CDMA," IEEE Trans. Commu., vol. 46, no. 2, pp.258-268, Feb. 1998.

0-7803-8255-2/04/$20.00

IEEE.

393