13-Fat Tails-Scaling-Stabl Page 370 Wednesday, February 4, 2004 1:00 PM
---
370 The Mathematics of Financial Modeling and Investment Management
F←()x = inf{s ∈R: Fs()≥x}, 0 < x < 1
It can be demonstrated that the following results hold:
D
■ F
←
(U 1 )= X 1
D
■ (X 1 ,n, ...,X ) = [F
←
(U 1 ,n), ...,F
←
nn, (Unn, )]
■ The random variable F(X 1 ) has a uniform distribution on (0,1) if and
only if F is a continuous function.
To appreciate the importance of the quantile transformation, let’s
introduce first the notion of empirical distribution function and second
the Glivenko-Cantelli theorem. The empirical distribution function Fn
of a sample X 1 , ..., Xn is defined as follows:
n
1
Fn ()x = ---∑IX( i ≤x)
ni = 1
where I is the indicator function. In other words, for each x, the empiri-
cal distribution function counts the number of samples that are less than
or equal to x.
The Glivenko-Cantelli theorem provides the theoretical underpin-
ning of nonparametric statistics. It states that, if the samples X 1 , ..., Xn
are independent draws from the distribution F, the empirical distribu-
tion function Fn tends to F for large n in the sense that
a.s.
∆n = sup Fn ()x – Fx()→0 , for n → ∞
x ∈R
The quantile transformation tells us that in cases where F is a Pareto
distribution, if we approximate n random draws from a uniformly dis-
tributed variable as the sequence 1,2,...,n, then the corresponding val-
ues of the sample X 1 , ..., Xn will be
1 1 --- 1
,,...,---
1 2 n
which is a statement of the Zipf’s law.
From the quantile transformation, the limit law of the ratio between
two successive order statistics can also be inferred. Suppose that an (infinite)