Hybrid Multi-Kernel Construction of Polar Codes

Abstract

A new family of multi-kernel polar codes are proposed. Instead of using identical matrices in each stage, assigning different constituent matrices to different virtual channels based on their properties will further improve polarization performance and error rate performance. Matrices with the same block-length can be classified based on Bhattacharyya parameters and a measure M (G) is proposed to compare their polarization performance. Construction framework and graph representation of hybrid multi-kernels are explicitly given. Numerical illustrations show that hybrid multi-kernels have better performance than previous multi-kernels.