Abstract:

Wireless communication systems with multiple transmit and multiple receive antennas have the potential to provide reliable communication at data rates that are substantially higher than those of the single antenna systems. The core challenge in designing such multiple-input multiple-output (MIMO) systems is to achieve these rates with reasonable computational complexity. A standard transceiver architecture for moving towards that goal is MIMO bit-interleaved coded modulation with iterative "soft" demodulation and decoding (MIMO BICM-IDD). However, the computational cost of the MIMO soft demodulator increases exponentially with the number of bits transmitted per channel use, and hence there is considerable interest in the design of approximate soft demodulation schemes with lower complexity.

In this talk we will develop a computationally-efficient and memory-efficient approach to MIMO soft demodulation, based on semidefinite relaxation, and we will show that this demodulator provides better performance at lower computational cost than the demodulator that is generally considered to provide the best trade-off between performance and complexity.

Semidefinite relaxation techniques have previously been proposed for "hard" demodulation problems for which the output is a vector of binary symbols. However, hard demodulators throw away much of the information that is available in the received signal, and substantially better performance can be obtained via soft demodulation, in which the output is an estimate of the likelihood of each bit.

The development of the proposed demodulator involves several steps, some of which may be of independent interest:

• First, we use the randomization step of the semidefinite relaxation technique to generate a list of candidates from which the likelihoods can be approximated.
• Second, we approximate this randomization step by a collection of independent Bernoulli trials. In combination with other observations, this approximation enables us to reduce the number of semidefinite programs (SDP) that need to be solved to one.
• Third, we demonstrate that since the performance of the demodulator depends, in part, on the richness of the list generated by the randomization step, the SDP should only be solved coarsely.
• Fourth, we show how a combination of incremental updates and hashing leads to an implementation that requires only a fraction of the memory resources required by a conventional implementation.

All four of these observations were required in order to develop a demodulator with better performance and lower computational cost than the existing state-of-the-art demodulator.

This talk is based on work with Mehran Nekuii at McMaster University, and Mikalai Kisialiou and Zhi-Quan (Tom) Luo at the University of Minnesota. (Mikalai is now with Intel, Portland, Orgeon.)