Large-Alphabet Semi-Static Entropy Coding Via Asymmetric Numeral Systems

Abstract

An entropy coder takes as input a sequence of symbol identifiers over some specified alphabet and represents that sequence as a bitstring using as few bits as possible, typically assuming that the elements of the sequence are independent of each other. Previous entropy coding methods include the well-known Huffman and arithmetic approaches. Here we examine the newer asymmetric numeral systems (ANS) technique for entropy coding and develop mechanisms that allow it to be efficiently used when the size of the source alphabet is large—thousands or millions of symbols. In particular, we examine different ways in which probability distributions over large alphabets can be approximated and in doing so infer techniques that allow the ANS mechanism to be extended to support large-alphabet entropy coding. As well as providing a full description of ANS, we also present detailed experiments using several different types of input, including data streams arising as typical output from the modeling stages of text compression software, and compare theproposed ANS variants with Huffman and arithmetic coding baselines, measuring both compression effectiveness and also encoding and decoding throughput. We demonstrate that in applications in which semi-static compression is appropriate, ANS-based coders can provide an excellent balance between compression effectiveness and speed, even when the alphabet is large.