13-Fat Tails-Scaling-Stabl Page 369 Wednesday, February 4, 2004 1:00 PM
Fat Tails, Scaling, and Stable Laws 369
Order Statistics
The behavior of order statistics is a useful tool for characterizing fat-
tailed distributions. For instance, the famous Zipf’s law is an example of
the behavior of order statistics. Consider a sample X 1 , ..., Xn made of n
independent draws from the same distribution F. Let’s arrange the sam-
ple in decreasing order:
Xnn, ≤ ... ≤X 1 ,n
The random variable Xk,n is called the kth upper order statistic. It can
be demonstrated that the distribution of the kth upper order statistic is
k – 1
F –
kn= PXkn<x)= ∑F
r
, ( , ()xFnr()x
r = 0
In addition, if F is continuous, it has a density with respect to F such
that
x
Fkn, = ∫fkn, ()z d Fz()
- ∞
where
f n! k –^1 –
kn= ---------------------------------------F ()F
x nk()x
,
(k – 1 )!(nk– )!
The differences between two consecutive variables in a sample Xk,n
- Xk+1,n are random variables called spacings. In the case of variables
with finite right endpoint xF the zero-th spacing is defined as: X0,n –
X1,n = xF – X1,n. The distribution of spacings depends on the distribu-
tion F. For instance, it can be demonstrated that the spacings of an
exponential random variable are independent, exponential random vari-
ables with mean 1/n for a n-sample. Spacings are a key concept for the
definition of the Hill estimator, as explained later in this section.
Another key concept, which is related to spacings, is that of quantile
transformation. Let X 1 , ..., Xn be IID variables with distribution func-
tion F and let U 1 , ..., Un be IID variables uniformly distributed on the
interval (0,1). Recall that, given a distribution function F, the quantile
function of F, written F←(x), is defined as follows: