Abstract
This letter proposes a non-binary polar-coded quantization (NB-PCQ) scheme that achieves near-optimal rate-distortion performance for both continuous-valued Gaussian sources and discrete-valued multilevel sources. Specifically, we use the non-binary polar code (NBPC) soft-input decoder as the vector quantization (VQ) encoder for source compression, while the NBPC encoder as the VQ decoder for source reconstruction. Further, we investigate the optimized design of both reconstruction constellations and transition probabilities. Numerical results validate the efficiency and superiority of the proposed scheme. For a Gaussian source, with parameters $(R,N,q)=(2,2^{10},16)$ , NB-PCQ exhibits a distortion gap of 0.89 dB from the theoretical bound, significantly outperforming the BPCQ baseline by 2.31 dB.