Soft-input, soft-output joint data detection and GRAND: A performance and complexity analysis
Hadi Sarieddeen, Peihong Yuan, Muriel Médard, Ken R. Duffy · 2023
Guessing random additive noise decoding (GRAND) has recently demonstrated maximum-likelihood (ML) decoding performance on efficient, universal silicon realizations. Leveraging input bit-reliability soft information extracted from the channel and noise statistics, GRAND rank-orders and queries noise sequences in non-decreasing likelihood to recover code-words of arbitrary code-book structures. We consider soft-input, soft-output (SISO) GRAND that generates bit-reliability log-likelihood ratios (LLRs) via successive Euclidean-distance computations over a list of noise-recovered words. Noise guessing and list construction follow an ordered reliability bits GRAND (ORBGRAND) mechanism, the guess budget of which controls the performance and complexity trade-offs. The generated LLRs form enhanced a priori information that adapts noise-sequence ordering in a subsequent soft-GRAND iteration. We derive bounds on the achievable rates under per-realization and marginal input soft information and empirically study the achievable rates of SISO-GRAND. We also examine the complexity of the joint data detection and GRAND core, highlighting its superiority to conventional list-based detection schemes. SISO-ORBGRAND can outperform conventional sphere decoding in data detection and LLR generation; the corresponding channel-mismatched rates approximate ML decoding.