(IEEE Communications Letters, 2023-09-18) Burich, Mariano Eduardo; Garcia-Frias, Javier
We propose novel non-linear graph-based analog codes that directly encode k real-valued source samples into n real-valued samples by using (non-linear) sample-by-sample soft quantization of the input samples followed by a linear transformation on the soft-quantized values. Different from existing analog coding schemes, the proposed analog codes are able to produce additional output symbols in a rateless manner and can be decoded utilizing message passing algorithms dealing with real-valued nodes, achieving a performance close to the theoretical limits.