Training Sequences Design for Symbol Timing Estimation in MIMO Correlated Fading Channels Yik-Chung Wu and Erchin Serpedin Department of Electrical Engineering, Texas A&M University, College Station, TX 77843-3128, USA. Email: {ycwu, serpedin}@ee.tamu.edu Abstract—In this paper, the problem of training sequences de- sign for symbol timing estimation in MIMO channel is addressed. In particular, we consider correlated fading between antennas. The optimal training sequences are derived by minimizing the modified Cramer-Rao bound (MCRB) with respect to the training data. It is found that when the transmit pulse is a root raised cosine pulse and there is no correlation among antennas, the optimal training sequences resemble the Walsh sequences. Furthermore, it is also found that the the impact of not knowing the antenna correlations when designing training sequences is very small. I. I NTRODUCTION Symbol timing synchronization is an important issue in multiple-input multiple-output (MIMO) communication sys- tems, as most of the research works in MIMO systems assume perfect symbol timing information at the receiver [1]-[8]. Symbol timing synchronization in MIMO uncorre- lated flat fading channel was first studied by Naguib et al. [12], where orthogonal training sequences are transmitted at different transmit antennas to simplify the maximization of the oversampled approximated log-likelihood function. This algorithm was extended by the authors in [13] and [14] to increase its estimation accuracy. Recently, the true maximum likelihood (ML) symbol timing estimator in MIMO channel based on training data was derived in [15]. Although the symbol timing estimation algorithms in MIMO channels have been well established, the problem of how to design the training sequences for symbol timing estimation in MIMO channels is not fully addressed. In [13], it was mentioned that the training sequences should be the Walsh sequences with the largest number of sign changes. Unfortunately, this choice of training sequences is based on an intuitive argument, so there is no optimality associated with it. In [14], the training sequences are designed to minimize the contribution of the inter-symbol interference (ISI) term in the approximated log-likelihood function, resulting the so-called “perfect sequences” [16]. However, if the approximated log- likelihood function is not used in the estimation (e.g., for the true ML estimator [15]), the training sequences obtained may not be optimal anymore. In this paper, we try to derive optimal training sequences that are independent of the estimation method. Toward this end, the training sequences are found by minimizing the modified Cramer-Rao bound (MCRB) with respect to the training data. It is found that when the transmit pulse is a root raised cosine pulse and there is no correlation among antennas, the optimal orthogonal training sequences resemble the Walsh sequences. Furthermore, it is also shown that the knowledge of antenna correlations is not important for designing training sequences. II. SIGNAL MODEL Consider a MIMO communication system with N transmit and M receive antennas. At each receiving antenna, a super- position of faded signals from all the transmit antennas plus noise is received. The complex envelope of the received signal at the j th receive antenna can be written as r j (t)= E s NT N i=1 h ij n d i (n)g(t nT ε o T )+ η j (t), j =1, 2, ..., M (1) where E s /N is the symbol energy; h ij ’s are the complex channel coefficients between the i th transmit antenna and the j th receive antenna; d i (n) is the zero-mean complex valued symbol transmitted from the i th transmit antenna; g(t) is the transmit filter with unit energy; T is the symbol duration; ε o [0, 1) is the unknown timing offset and η j (t) is the complex-valued circularly distributed Gaussian white noise at the j th receive antenna, with power density N o . Throughout this paper, it is assumed that the channel is frequency flat and quasi-static. After passing through the anti-aliasing filter, the received signal is then sampled at rate f s =1/T s , where T s T/Q (Q is the oversampling ratio). The received vector r j , which consists of L o Q consecutive received samples (L o is the observation length), can be expressed as (without loss of generality, we consider the received sequence start at t =0) r j = ξ A εo ZH T j,: + η j , (2) where 1 ξ E s /N T , r j [r j (0) r j (T s ) ... r j ((L o Q 1)T s )] T , (3) A ε [a Lg (ε) a Lg+1 (ε) ... a Lo+Lg1 (ε)] , (4) a i (ε) [g(iT εT ) g(T s iT εT ) ... g((L o Q 1)T s iT εT )] T , (5) Z [d 1 d 2 ··· d N ], (6) d i [d i (L g ) d i (L g + 1) ··· d i (L o + L g 1)] T , (7) 1 Notation x T denotes the transpose of x, and x H denotes the transpose conjugate of x. Globecom 2004 81 0-7803-8794-5/04/$20.00 © 2004 IEEE IEEE Communications Society