Low Complexity Detectors For Bit-Interleaved Coded Modulation Over Flat-Fading MIMO Channels

Michael Samuel
Advisor: Michael P. Fitz

Tuesday, September 29, 2009 at 1:00pm
Engr IV Room 67-124

Abstract:
Bit-interleaved coded modulation over MIMO channels provides high spectral efficiencies and good performance. Although the interleaver is very beneficial in i) alleviating deep fades and ii) introducing independence among the coded bits, it renders an optimal receiver impossible to realize. For this reason, a two-stage suboptimal receiver is always assumed that consists of a cascaded detector and decoder. While the decoder is quite standard in most cases, the detector involves a great deal of a trade-off between complexity and performance. For this goal, two complexity-reduction techniques are considered in this thesis.

The first technique focuses on the 2 x 2 channel. The complexity reduction is done by using a space-time block code with the so-called multi-strata structure that guarantees fast decodability. For any channel matrix, the upper left block of the R matrix (after QR-decomposition) is always a scaled identity. Since the original multi-strata code has inferior performance compared to the Golden code, a search was done to find a new code whose minimum determinant is 2.3874 with the QPSK constellation and shows a performance degradation of less than 0.5 dB compared to the Golden code. For the case of iterative decoding, a very interesting case is shown where the multi-strata code is better than spatial multiplexing both in performance and complexity.

In the second technique, a generalization of the Schnorr-Euchner algorithm is used to build two kind of iterative soft detectors. The original Schnorr-Euchner algorithm can be easily extended to more general objective functions - i.e., that are not necessarily a squared Euclidean norm - but that can be decomposed in a certain form. This is exploited in iterative detectors to include the input a priori probabilities. The first kind is called the most-contributive-2N detectors. It searches efficiently for the most contributive N lattice points in the two log-sum-exponents of the bit LLRs. The special case of N = 1 yields a max-log approximation. The second kind is a pragmatic iterative detector. It looks for a subset of the lattice points and uses it as a representative of the whole lattice. Unlike former attempts, it shrinks the sphere radius during the search in order to end it soon. The search radius is controlled via a radius backlash update strategy which confines the search to within a certain distance (backlash) from the best found point thus far. The advantages of the two new kind are that they i) take into consideration the input a priori probabilities ii) do not require initial radius estimation and iii) their search step is void of square root and division operations. The case of non-perfect channel estimates is also considered. For this, an expression of the a posteriori probabilities is first derived under the assumption of uncorrelated channel estimation errors. Then, it is shown how to extend the most-contributive-2N detector to operate in this case. The performance gain with dense constellations is very significant.