IEEE Communications 2002 Bucharest, Romania, December, 5-7, 2002 A FAST ML-BASED RECEIVER FOR MIMO RICIAN FADING CHANNEL L. COLLIN 1 , P. ROSTAING 2 , O. BERDER 2 , G. BUREL 2 1 IRENav Ecole Navale, BP 600, 29240 Brest collin@ecole-navale.f 2 LEST-UMR CNRS 6165 6 Av. Le Gorgeu, BP 809, 29285 Brest C´ edex Philippe.Rostaing@univ-brest.fr Abstract: We derive a fast maximum likelihood based (MLB) decoder for a multi-input multi-output (MIMO) Rician fading channel with additive white Gaussian noise (AWGN). The basic idea is to take profit of the Rician channel structure to significantly reduce the search of the optimum vector of symbols by the ML crite- rion. When channel diversity is low, we obtain bit error rates (BER) which are very close to the BER of the maximum likelihood (ML) optimum decoder. Comparisons in terms of BER for Quadrature Amplitude Modulation (QAM) are performed for the MLB, ML and OSIC (Ordered Successive Interference Cancella- tion) decoders via simulations. Finally, the ratio of computational complexities (depending of the number of transmitters and on the constellation), between the ML and the MLB is presented to show the interest of the proposed approach. I INTRODUCTION Narrowband MIMO transmission systems receive increased attention for a few years, due to their abil- ity to provide large spectral efficiencies over rich scattering transmission channels. Spatial multiplexing systems, known as V-BLAST (Bell Laboratories Layered Space-Time) architecture, have been proposed recently, and first laboratory experiments have shown that spectral efficiencies as high as 20bits/s/Hz can be obtained [3]. The basic model of a narrowband MIMO transmission system is: r = Ha + n (1) where H is the channel matrix, and r, a, and n respectively stand for the n R -dimensional received vector, the n T -dimensional transmitted vector (the entries of which are the symbols), and the noise vector. The objective of the receiver algorithm is to estimate a when r and H are known (in practice, H is estimated using a training sequence). We also assume 1 E[aa H ]=(p 0 /n T ) I nT with p 0 the total transmitted power, E[nn H ]= σ 2 n I nR with σ 2 n the noise variance and E[an H ]=0. It is well known that the optimal method is the maximum likelihood (ML). However, this method is difficult to use in many applications because it requires a large computation time. In this paper, we propose a faster algorithm whose performances, in terms of bit error rates (BER), are close to the ML performances when the low diversity hypothesis on the channel model is true. The main limitation to MIMO is the low spatial diversity. This occurs typically for correlated fading MIMO channel [5]. Otherwise, uncorrelated matrix channel containing random entries with similar and non-zero average values could be poorly conditioned, particularly when standard deviations of the matrix entries are not too large with respect to the average values. A typical example is the Rician channel model. This model will be considered throughout the paper. 1 The superscript H denotes transpose conjugate and I nT is the n T -dimension square identity matrix.