Noiseless compression using non-Markov models

Adaptive data compression techniques can be viewed as consisting of a model specified by a database common to the encoder and decoder, an encoding rule and a rule for updating the model to ensure that the encoder and decoder always agree on the interpretation of the next transmission. The techniques which fit this framework range from run-length coding, to adaptive Huffman and arithmetic coding, to the string-matching techniques of Lempel and Ziv. The compression obtained by arithmetic coding is dependent on the generality of the source model. For many sources, an independent-letter model is clearly insufficient. Unfortunately, a straightforward implementation of a Markov model requires an amount of space exponential in the number of letters remembered. The Directed Acyclic Word Graph (DAWG) can be constructed in time and space proportional to the text encoded, and can be used to estimate the probabilities required for arithmetic coding based on an amount of memory which varies naturally depending on the encoded text. The tail of that portion of the text which was encoded is the longest suffix that has occurred previously. The frequencies of letters following these previous occurrences can be used to estimate the probability distribution of the next letter. Experimental results indicate that compression is often far better than that obtained using independent-letter models, and sometimes also significantly better than other non-independent techniques.