Linearization of zipfian distribution for chinese characters

Abstract

In this paper, we report our results of least-square fittings to 4 sets of data derived from Chinese characters, namely, character strokes, radicals, characters and words. We have found that fitting using a power series, ie ft versus Rt (f is the frequency of occurrence, R the rank and t is constant) is better than the use of a logarithm series derived from the original simple Zipf's law, ei f R=constant, or log f=c-log R. The dependency of f versus R is found to be of order 5 as we have found that t=0.2. We have also discovered a secondary dependency of f on R of lower order. This secondary dependency can be modeled using a cosine function.