Space-efficient construction of optimal prefix codes

Abstract

Shows that the use of the lazy list processing technique from the world of functional languages allows, under certain conditions, the package-merge algorithm to be executed in much less space than is indicated by the O(nL) space worst-case bound. For example, the revised implementation generates a 32-bit limited code for the TREC distribution within 15 Mb of memory. It is also shown how a second observation-that in large-alphabet situations it is often the case that there are many symbols with the same frequency-can be exploited to further reduce the space required, for both unlimited and length-limited coding. This second improvement allows calculation of an optimal length-limited code for the TREC word distribution in under 8 Mb of memory; and calculation of an unrestricted Huffman code in under 1 Mb of memory.