358 zyxwvuts Comparison Of Two DOA Tracking Implementations For SDMA zy Wolfgang Utschick, Marco Treiber, Tobias Kurpjuhn, and Josef A. Nossek Lehrstuhl fur Netzwerktheorie und Signalverarbeitung Technische Universitat Munchen zyxw { utschick,treiber,kurpjuhn,nossek} @nws.ei.tum.de ABSTRACT Utilizing adaptive antenna arrays at the base sta- tions of next generation mobile communication systems has been proposed as a promising approach to meet future requirements, e.g. spatio-temporal filtering techniques benefit from the detailed knowledge of the directional channel parameters. Unfortunately, in parameter estimation the computation of the signal subspace often turns out to be the most time-consuming part. However, the computational complexity of the sub- space estimation can be significantly reduced by means of tracking. To this end, we propose a new technique which is merely based on the projector representation of the subspace. The presented results are based on the DSP implementation of two algorithms for tracking the azimuthal and elevation angles of impinging wavefronts in an alternating mobile communications system. I. INTRODUCTION Third and fourth generation mobile radio communica- tions systems will be characterized by a growing number of more sophisticated services including multi-media applications [ l , 2, 31. This leads to increased data rates as well as the asymmetric distribution of the traffic con- cerning the uplink and downlink connection. Adaptive antennas can be deployed to meet the future high spec- tral and quality requirements of the demanding 3G/4G mobile communication systems. Adaptive antennas ex- ploit the inherent spatial diversity structure of the mobile radio channel, provide antenna gain, and enable interfer- ence suppression [4]. The performance of spatio-temporal downlink process- ing is substantially based on channel estimation includ- ing the estimation of spatial characteristics of the mo- bile radio channel. Furthermore channel estimation has to meet real-time requirements in order to cope with the time-varying scenarios in mobile communications sys- tems. To this end, the use of tracking algorithms often offers a crucial alternative to standard methods in pa- rameter estimation and system identification. This work is especially devoted to tracking algorithms for the estimation of signal subspaces. Whereas the tracking of the orthonormal basis of the signal subspace is subject of a number of papers in the past, the immedi- ate tracking of the unique projection matrix of the signal subspace has not been proposed yet. In [5] the author presents a thorough overview of most of the adaptive zyxwvut 0-7803-6465-5100 zyxwvutsrqpon $10.00 zyxwvutsrqp 0 2000 zyxwvutsrqpon IEw algorithms for subspace tracking. Thus, the subspace tracking techniques can be grouped into three families: (1) classical eigenvalue decomposition (EVD) and sin- gular value decomposition (SVD) methods modified for use in adaptive processing, (2) variations of rank-one updating algorithms, and (3) algorithms that consider the eigenvalue/singular value decomposition as a con- strained or unconstrained optimization problem which can be adaptively processed by means of gradient based methods. In this work, we extend the variety of approaches by a tracking method which directly operates on the unique projection matrix onto the signal subspace instead of tracking its orthonormal basis. This new projection based algorithm is motivated by the PAST algorithm [5]. The PAST belongs to the type of methods which con- sider the EVD/SVD as an optimization problem. It has recently been subject of a real-time implementation of algorithms for a GSM base station zyxw [6]. Consequently, the presented simulation results provide a comparison between both algorithms in terms of computational com- plexity and estimation error. 11. SYSTEM STRUCTURE The performance evaluation of the proposed algorithm is based on the DOA (direction of arrival) estimation by means of ESPRIT [7]. Hereby, the signal subspace es- timation has been of particular interest, since this step of ESPRIT-like methods generally turns out to be the most time-consuming part. In order to get close to real systems, a DSP served as the hardware processor. Im- plementation is done on the (26701 32-bit floating point DSP from TEXAS INSTRUMENTS with a clock rate of 150 MHz [8]. The supported very long instruction word technique (VLIW) enables the DSP to address its mul- tiple hardware units simultaneously and to carry out up to eight instructions per clock cycle. The source code is completely written in C and represents an implementa- tion of 2D-Unitary ESPRIT [9]. Thus, two parameters of each of the impinging wavefronts at a uniform rectan- gular antenna array (URA) can be resolved, the azimuth cp and the elevation 6 angles of the DOAs. Since the DSP is not part of a real communication sys- tem, the input data of 2D-Unitary ESPRIT is generated by MATLAB. Thereby the transmission medium is sup- posed to be isotropic and linear. The noise is modeled as a complex, zero-mean white Gaussian process. Un- der the assumptions of narrowband and farfield signals,