Code Annealing and the Suppressing Effect of the Cyclically Lifted LDPC Code Ensemble

TL;DR

This paper introduces code annealing combined with cyclic lifting to design short block length codes with low error floors, providing analytical tools and methods to optimize code performance and suppress error floors effectively.

Contribution

It proposes a novel code design method integrating code annealing and cyclic lifting, with analytical insights into error floor suppression and new construction techniques for low FER codes.

Findings

Cyclic lifting significantly suppresses error floors in code ensembles.

Analytical expressions and bounds for FER error floors are derived.

Regular and irregular codes achieve 10^4 to 10^6 times lower FER than classic codes.

Abstract

Code annealing, a new method of designing good codes of short block length, is proposed, which is then concatenated with cyclic lifting to create finite codes of low frame error rate (FER) error floors without performance outliers. The stopping set analysis is performed on the cyclically lifted code ensemble assuming uniformly random lifting sequences, and the suppressing effect/weight of the cyclic lifting is identified for the first time, based on which the ensemble FER error floor can be analytically determined and a scaling law is derived. Both the first-order and high-order suppressing effects are discussed and quantified by different methods including the explicit expression, an algorithmic upper bound, and an algebraic lower bound. The mismatch between the suppressing weight and the stopping distances explains the dramatic performance discrepancy among different cyclically lifted codes when the underlying base codes have degree 2 variable nodes or not. For the former case, a degree augmentation method is further introduced to mitigate this metric mismatch, and a systematic method of constructing irregular codes of low FER error floors is presented. Both regular and irregular codes of very low FER error floors are reported, for which the improvement factor ranges from 10^6-10^4 when compared to the classic graph-based code ensembles.