Sphere decoding (SD) has been proposed as an efficient way to perform maximum-likelihood (ML) decoding of Polar codes. Its latency requirements, however, are determined by its ability to promptly exclude from the ML search (i.e., prune) large parts of the corresponding SD tree, without compromising the ML optimality. Traditional depth-first approaches initially find a "promising" candidate solution and then prune parts of the tree that cannot result to a "better" solution. Still, if this

more »

... te solution is far (in terms of Euclidean distance) from the ML one, pruning becomes inefficient and decoding latency explodes. To reduce this processing latency, an early termination approach is, first, introduced that exploits the binary nature of the transmitted information. Then, a simple but very efficient SD approach is proposed that performs multiple tree searches that perform decreasingly aggressive pruning. These searches are almost independent and can take place sequentially, in parallel, or even in a hybrid (sequential/parallel) manner. For Polar codes of 128 block size, both realizations can provide a latency reduction of up to four orders of magnitude compared to state-of-the-art Polar sphere decoders. Then, a further 50% latency reduction can be achieved by exploiting the parallel nature of the approach.