- Сообщения
- 2 089
- Реакции
- 145
Zipf-like Distributions
As implicit in the introduction, and in contrast with continuous random variables, in the discrete case a power law in the probability mass function f(n) does not lead to a power law in the complementary cumulative distribution or survival function S(n), and vice-versa. Let us specify our definition for both functions, f(n) = Prob[frequency = n] (as usual), and S(n) = Prob[frequency ≥ n] (changing, for convenience, the usual strict inequality sign by the non-strict inequality). Then, the relation between both is f(n) = S(n) − S(n + 1) andWe consider that the values the random variable takes, given by n, are discrete, starting at the integer value a, taking values then n = a, a + 1, … up to infinity. In this study we will fix the parameter a to a = 1, in order to fit the whole distribution and not just the tail. Then, although for large n and smooth S(n) we may approximate f(n) ≃ −dS(n)/dn, this simplification is clearly wrong for small n. Note that the simplification leads to the implication that a power law in f(n) leads to a power law in S(n), and vice-versa, but this is clearly wrong for small values of n in discrete distributions. The simplification also lies in the equivalence between Eqs (1) and (2), assuming that S(n) is proportional to the rank and inverting Eq (1).