Phrase elimination in greedy parsing dictionary coders with deferred innovation

Summary form only given. LZ dictionary coders parse the message into successive substrings, each consists two parts, the citation, the longest prefix phrase that has already been accommodated in the dictionary, and the innovation, the symbol immediately following the citation. Suppose the input alphabet set is A and the dictionary D = {p₁, p₂...p_n} is a set of phrases where p_i∈A*, parsed by a greedy-parsing LZ coder. Represented in the form of a dictionary search tree, the process matching a phrase in D with a citation can be viewed as traversing from the root of the dictionary tree by matching consecutive symbols from the input until the mismatching innovation occurs. The dictionary is reduced to D´=D/E. Its phrase index is then encoded by a less redundant code (LRC) with upper bound of codeword length. The expected number of phrases in D´ was estimated. It was also verified with experiments that such estimation is accurate. It was also shown that 3% improvement is typical over LZW coders with LRC and 5% better than standard LZW.