LLR-based attention-weighted projection decoding for sparse vector coded short-packet transmissions

This paper presents an enhanced sparse vector coding (SVC) scheme tailored for short-packet communication over noisy wireless channels. Conventional SVC often uses random (Gaussian, Bernoulli) or structured (partial Hadamard) spreading/projection matrices with baseline decoding based on naive projection scores to recover the sparse signal support, which may suffer from reliability degradation under fading conditions. To address this, we introduce an enhanced decoding approach leveraging LLR-based attention-weighted projection, which dynamically reweights received signal measurements according to their reliability, giving more importance to more reliable measurements. Simulation results over both Rayleigh flat fading and 3GPP TDL-C/TDL-D frequency-selective channels show that the proposed LLR-based attention-weighted projec tion decoding outperforms baseline projection decoding with both structured and random projection matrices. Moreover, we extend our analysis to a short-packet (3–11 bits) transmission framework over the 3GPP PUCCH. The results highlight that the proposed SVC scheme using a partial Hadamard spreading matrix combined with LLR-based attention-weighted projection decoding outperforms 3GPP RM codes under optimal ML decoding conditions for very low BLER targets, making it a promising short-packet coding candidate for ultra-reliable low-latency 6G short-block-length UL/DL channels.