MAP Decoding Algorithm for Extended Turbo Product Codes over Flat Fading Channel Changlong Xu, Ying-Chang Liang, and Wing Seng Leon Institute for Infocomm Research 21 Heng Mui Keng Terrace, Singapore 119613. Abstract— In this paper, an efficient MAP iterative decoding algorithm is proposed for extended turbo product codes over flat fading channel. The overflow avoidance of extrinsic information and the stop criteria of iterative decoder are also discussed. Simulation results are presented to verify the performance and show the efficacy of the proposed decoding algorithm. be x = (x0 , . . . , xi , . . . , xN ) with xi ∈ {−1, 1} by mapping the bit 0 to symbol 1 and bit 1 to symbol −1. After transmission over a flat fading channel, the received signal y = (y0 , . . . , yi , . . . , yN ) may be expressed as I. I NTRODUCTION where α = diag (α0 , . . . , αi , . . . , αN ) is a (N +1)×(N +1) diagonal matrix with αi , the unit-variance i.i.d. fading coefficient, and g = (g0 , . . . , gi , . . . , gN ) is the zero-mean AWGN vector, with each of its elements having variance N0 . If the receiver has perfect knowledge of the fading channel or channel state information (CSI), then coherent detection may be performed by eliminating the phase rotations caused by the channel. Thus, the phase recovered signal may be written as r = yαH (2) Because of low complexity and high data rate, turbo product codes (TPC) have already been adopted in many wireless standards, such as IEEE 802.16, satellite communication systems and digital storage systems. Extended TPC can be obtained by appending an additional column and one row of single parity check bits to conventional TPC. The minimum Hamming distance of extended TPC is much larger than that of original TPC, thus, almost all the TPC used in practical systems are extended TPC. In 1994, Pyndiah proposed an iterative decoding algorithm based on the Chase algorithm for TPC [1]. However, the reliability factors and weighting factors in the algorithm are optimized by experiment. It is not convenient to implement the algorithm since the values of two factors vary with the sizes of TPC and the number of iterations. Several iterative decoding algorithms are presented for twodimensional binary block and convolutional codes in [2]. But the structure of the two dimensional binary block codes is not same as that of TPC defined in IEEE 802.16, and the extended block codes are not considered either. In this paper, we propose a maximum a posteriori probability (MAP) decoding algorithm for extended TPC over flat fading channel. We also investigate the effects of the extrinsic information having excessive large magnitudes or overflowing, and the stopping criteria for the iterative decoder. The performance comparison between the MAP and the Chase algorithms are presented by computer simulations. y = xα + g The output of the MAP decoder is defined as the a posteriori log-likelihood ratio for a transmitted “+1” and a transmitted “-1” in the information sequence L(x̂k ) = log x∈C,xk =+1 x∈C,xk =−1 = log P (x|r) P (x|r) N x∈C,xk =+1 j=0 N x∈C,xk =−1 j=0 = log P (xj |rj ) P (xj |rj ) P (xk = +1) P (rk |xk = +1) + log P (xk = −1) P (rk |xk = −1) N P (xj |rj ) + log II. MAP D ECODING A LGORITHM FOR EXTENDED TPC OVER FLAT FADING C HANNELS x∈C,xk =+1 j=0,j=k N x∈C,xk =−1 j=0,j=k The encoding and decoding process of TPC are the same for all columns and all rows. For simplicity and brevity, we only need to consider the encoding and decoding algorithm for any one row. We assume that the source information is an extended binary linear block codeword in C(N, K) and expressed as c = (c0 , . . . , ci , . . . , cN ) with ci ∈ {0, 1}, and cN is the extended single parity check bit. Thus, the transmit signal with binary phase-shift keying (BPSK) modulation may 1­4244­0785­0/06/$20.00 (1) P (xj |rj ) = L(xk ) + Lc rk + Le (xk ) (3) where L(xk ) and Le (xk ) are a priori information and an extrinsic information of bit xk , respectively. Lc = 4αk Es/N 0 is the reliability value of the channel. To calculate the extrinsic information, we define the following log-likelihood ratio 2182 L(xi |ri ) = log P (xi = +1|ri ) = L(xk ) + Lc rk . P (xi = −1|ri ) (4) Using the defined log-likelihood, the a priori probability P (xi |ri ) can be expressed as e±L(xj |rj ) 1 + e±L(xj |rj ) −L(xj |rj )/2 e eL(xj |rj )xj /2 . (5) = 1 + e−L(xj |rj ) = So the extrinsic information can be rewritten as Q Q P Q P Q N x∈C,xk =+1 j=0,j=k N x∈C,xk =−1 j=0,j=k −L(xj |rj )/2 −L(xj |rj ) 1+e e e −L(xj |rj )/2 −L(xj |rj ) 1+e N = log e x∈C,xk =+1 j=0,j=k N x∈C,xk =−1 j=0,j=k L(xj |rj )xj /2 eL(xj |rj )xj /2 eL(xj |rj )xj /2 eL(xj |rj )xj /2 −1 . If N − K > K, the computing complexity in Equation (6) can be reduced by using the dual code C of the extended code C [3]. In this way, the Equation (3) can be expressed as L(x̂k ) = L(xk ) + Lc rk + 1+ log 1− N i=2 j=0,j=k i=2 −2 10 −3 10 (7) 2N +1−K 2N +1−K 10 (6) BER Le (xk ) = log P P We use Monte Carlo simulations are used to evaluate the effectiveness of the proposed MAP decoding scheme for extended TPC. The square extended TPCs are constructed by using extended Hamming codes (8, 4, 4), (16, 11, 4),(32, 26, 4) and (64, 57, 4) as the component codes. Perfect CSI is assumed and the number of iterations is set to 4. We also included the simulation results of the Chase decoding algorithm for extended TPC in [4] for comparison purposes. The number of the test patterns is 16 and the value of the scaling factor α is fixed as 0.5 for all the iterations. The modulation used is QPSK. (−xik ) tanh N j=0,j=k L(xj |rj ) 2 tanh (1−xij ) 2 L(xj |rj ) 2 −4 10 (1−x ) ij 2 −2 (8,4) CHS (8,4) MAP (16,11) CHS (16,11) MAP (32,26) CHS (32,26) MAP (64,57) CHS (64,57) MAP −1 0 1 2 3 4 5 SNR(dB) where x represents the codewords of the dual code C , and the index i = 1 denotes the all “+1” codeword. It is easy to realize the iterative decoder for extended TPC after implementing the soft-input and soft-output decoder of the component codes. However, the overflow problem of extrinsic information and the correct stop criteria may affect the decoding performance significantly. We suggest the following solutions to these problems: 1) Overflow Avoidance The magnitude of extrinsic information may become excessively large and overflow with the increased number of iterations, especially for high signal to noise ratio (SNR). It will consequently degrade the decoding performance. To circumvent the overflow, we set z = 0.9999z, when z = tanh(L(xj |rj )/2) > 0.9999. 2) Stopping Criteria The assumption that the received signal and the extrinsic information are statistical independent may not be valid with increasing number of iterations. This will cause performance degradation, if the decoding procedure is not terminated at an appropriate time. In this paper, we use the syndrome values of the of the component codes as an indication to terminate the decoding process. That is, if the values of the syndrome are all zeros, the decoder will, otherwise, the iterative decoding procedure will continue. Fig. 1. The TPC performance comparison between the Chase and MAP decoding algorithm with QPSK under AWGN channel. −1 10 −2 10 BER P (xj = ±1|rj ) III. S IMULATION R ESULTS −3 10 −4 10 −5 10 0 (8,4) CHS (8,4) MAP (16,11) CHS (16,11) MAP (32,26) CHS (32,26) MAP (64,57) CHS (64,57) MAP 2 4 6 SNR(dB) 8 10 12 Fig. 2. The TPC performance comparison between the Chase and MAP decoding algorithm with QPSK under flat fading channel. The performances of the Chase and proposed MAP decoding algorithms for extended TPC transmitted over AWGN 2183 channel is shown in Figure 1. Signal-to-noise ratio (SNR) is defined as the ratio of symbol energy to noise variance. The results show that the TPC performance is degraded with increased code rate for both the proposed MAP and Chase decoding algorithms. For square TPC constructed by extended Hamming codes (8, 4, 4) and (16, 11, 4), the performances of Chase algorithm are better than that of MAP algorithm. However, the situation is reversed for square TPC constructed by extended Hamming codes (32, 26, 4) and (64, 57, 4). The reason is that the Chase decoding algorithm with 16 test patterns is more efficient for smaller sizes of extended Hamming codes. Figure 2 displays performances of the proposed MAP and Chase algorithm to decode extended TPC transmitted over a flat fading channel. It is also observed that the performance of MAP algorithm is superior to the Chase algorithm for extended TPC with high code rates. R EFERENCES [1] R. Pyndiah, A. Glavieux, A. Picart, and S. Jacq, “Near optimum decoding of product codes,” in Proc.of GLOBECOM, vol. 1, San Francisco, CA, Nov. 1994, pp. 339–343. [2] J. Hagenauer, E. offer, and L. Papke, “iterative decoding of binary block and convolutional codes,” IEEE Trans. Inform. Theory, Vol. 42, pp. 429445, March 1996. [3] G. Battail, M. C. Decouvelaere, and P. Goldlewski, “Replication decoding,” IEEE Trans. Inform. Theory, vol. IT-25, pp.332-345, May 1979. [4] Changlong Xu, Wing Seng Leon and Ying-Chang Liang, “Chase decoding algorithm for shortened turbo product codes in flat fading channels,” in Radio and Wireless Symposium, 2006 IEEE, San Diego, CA, Juan. 2006, pp. 15 - 18 IV. C ONCLUSIONS In this paper, we propose an efficient MAP iterative decoding algorithm for extended turbo product codes over flat fading channel. The performance comparison between the proposed MAP algorithm and Chase algorithm is investigated by simulation results. They show that the performance of MAP algorithm is significantly better than that of Chase algorithm for extended TPC with high code rates. 2184