Achieving Near MAP Performance With an Excited Markov Chain Monte Carlo MIMO Detector

Abstract

We introduce a revised derivation of the bitwise Markov Chain Monte Carlo (MCMC) multiple-input multiple-output (MIMO) detector. The new approach resolves the previously reported high SNR stalling problem of MCMC without the need for hybridization with another detector method or adding heuristic temperature scaling factors. Another common problem with MCMC algorithms is the unknown convergence time making predictable fixed-length implementations problematic. When an insufficient number of iterations are used on a slowly converging example, the output log likelihood ratios can be unstable and overconfident. Therefore, we develop a method to identify rare slowly converging runs and mitigate their degrading effects on the soft-output information. This improves forward-error-correcting code performance and removes a symptomatic error floor in bit error rate plots. Next, pseudo-convergence is identified with a novel way to visualize the internal behavior of the Gibbs sampler. An effective and efficient pseudo-convergence detection and escape strategy is suggested. Finally, the new excited MCMC (X-MCMC) detector is shown to have near maximum-a-posteriori performance even with challenging, realistic, and highly-correlated channels at the maximum MIMO sizes and modulation rates supported by the 802.11ac WiFi specification, $8\times 8$ MIMO 256 quadrature amplitude modulation.