See discussions, stats, and author profiles for this publication at: [Link]

net/publication/264991805

The Exponentiated Generalized Extreme Value Distribution

Article · January 2006

CITATIONS READS

5 430

2 authors:

Atinuke O. Adebanji Shola Adeyemi

Kwame Nkrumah University Of Science and Technology STATSXPERTS CONSULTING LIMITED
37 PUBLICATIONS   78 CITATIONS    37 PUBLICATIONS   113 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Settlement Census in Somalia View project

REsect: Secondary Resectability of Liver Mets View project

All content following this page was uploaded by Shola Adeyemi on 28 January 2015.

The user has requested enhancement of the downloaded file.

Journal of Applicable Functional Differential Equations (JAFDE)
January-June 2006, Volume 1, No. 1, pp. 41-47

THE EXPONENTIATED GENERALIZED

EXTREME VALUE DISTRIBUTION
Shola ADEYEMI* & Atinuke O. ADEBANJI

ABSTRACT: We introduce an exponentiated generalized extreme value

distribution. This distribution is a two-parameter and a generalized version of the
generalized extreme value distribution. Some of its mathematical properties are
studied.
Key words: Two-parameter generalized extreme value, hazard rate function,
Shannon entropy; e.t.c.
2000 Mathematics Subject Classifications: 62G30.

1. INTRODUCTION
The Generalized Extreme-Value (GEV) distribution of Jenkinson (1955) is often
used to model extremes of natural phenomena such as river heights, sea levels, stream
flows, rainfall and air pollutions in order to obtain the distribution of daily or annual
maxima. Additionally, in reliability context, analogous analyses are performed where
the interest is in sample minima strengths and failure times.
The GEV distribution combines into one the three possible types of limiting
distribution for extremes,as derived by Fisher and Tippett (1928). Let a random
variable X be GEV distributed, then X has the distribution function
1

F(x) = exp[ −(1 − kx ) k ,

k ≠0
= exp [–exp(–x)], k = 0 ...(1.1)
and density function

1 1
f(x) = (1 − kr ) k −1 exp[ −(1 − kx ) k ], k ≠ 0

= exp(–x) exp [–exp(–x)], k = 0 ...(1.2)

1 1
where –∞ < x < for k > 0 and 1 <x <∞ for k < 0. Fisher-Tippett Type I, II and III
k k
distributions correspond to k = 0, k < 0 and k > 0, respectively. In practice, the shape
42 Shola ADEYEMI & Atinuke O. ADEBANJI

1 1
parameter usually lies in the range − < k < [Hoskin et al. (1985)].
2 2

2. THE EXPONENTIATED GENERALIZED EXTREME-VALUE

(EGEV) DISTRIBUTION
Gupta et. al. [1998] introduced the concept of exponentiated distributions when he
obtained, as a generalization of the standard exponential distribution, the exponentiated
exponential (EE) distribution. In particular, the EE distribution is defined by
F(x) = [1 – exp(–λx)]α ...(1.1)
for x > 0, λ > 0 and α > 0, which is simply the α-th power of the cdf of the standard
exponential distribution. The mathematical properties of this EE distribution have
been studied in detail by Gupta and Kundu [2001] and Nadarajah and Kotz [2003].
Adeyemi (2004) has also studied the exponentiated generalized Pareto distribution.
In this paper, we introduce, a two-parameter generalized extreme-value distribution,
herein refered to as the exponentiated generalized extreme value distribution which
generalizes (1.1) in the same way (2.1) generalizes the standard exponential
distribution, and study its properties.
From (1.1), we define the new distribution by the cdf
F(x) = exp [–β(1 – kx)1/k ; k ≠ 0
= exp [–β(exp –x)]; k = 0 ...(2.2)
where β> 0. We refer to (2.2) as the exponentiated generalized extreme-value (EGEV)
distribution. The corresponding density function is
1 1
f(x) = β (1 − kr ) k −1 exp[ − β (1 − kx )2 ]; k ≠ 0
= βexp[–βexp(–x)] exp(–x); k = 0 ...(2.3)
The distribution (2.3) with k = 0 has been studied by Adeyemi and Ojo (2003) as
a generalization of the Gumbel distribution. Therefore, attention will be restricted to
the case k 6= 0 in this study.

3. THE SHAPE
The first derivative of logef (x) for the EGEV distribution is

d log e f ( x ) k −1 1
−1
= + β (1 − kr ) k ...(3.1)
dx 1 − kx
The Exponentiated Generalized Extreme Value Distribution 43

Standard calculations based on (3.1) show that f(x) exhibits a single mode at x = x0
1 d log e f ( x )
with f(–∞) = 0 = f   , where x0 is the solution of = 0 . Possible shapes
k dx
of f(x) for some selected values of β and k can easily be obtained from (3.1).

4. HAZARD RATE FUNCTION

f ( x)
The hazard rate function defined by h(x) = is an important quantity
1 − F ( x)
characterizing life phenomena. For the EGEV distribution, h(x) takes the form

1
β exp[ − β (1 − kx ) k ]
h(x) = 1 1 ...(4.1)
(1 − kr ) 2 [1 − exp{− β (1 − kx ) k }]
The first derivative of logeh(x) with respect to x is

1 1
1
k −1 −1 β (1 − kx ) k exp[ − β (1 − kx ) k ]
d log h( x ) + β (1 − kx ) k +
= 1 − kx 1 ...(4.2)
dx
k [1 − exp{− β (1 − kx ) k }]
For k < 0 and x ≥ 0, h(x) is increasing and it is decreasing for k > 0 and x < 0.

5. THE MOMENTS
The r-th moments of the EGEV distribution is obtained from
1
E(1 – kX)r =
∫ k (1 − kx )r
−∞
f ( x )dx ...(5.1)

We obtain the r-th moments in a recurssive formula as

r
r 
∑  j  (−k ) j µ j = β − kr Γ ( kr + 1) ...(5.2)
j =0  
In particular, the first four moments of the EGEV distribution are given as

1
µ = [1 − β − k Γ( k + 1)] ...(5.3)
k
44 Shola ADEYEMI & Atinuke O. ADEBANJI

β −k
µ2 = 2
[Γ(2k + 1) − Γ (k + 1) + β k ] ...(5.4)
k

β −k
µ3 = 3
[4 β k − 15Γ (k + 1) + 6Γ (2k + 1) − Γ(3k + 1)] ...(5.5)
k

β −k
µ4 = 4
[58Γ (k + 1) − 14Γ (2 k + 1) + 4Γ (3k + 1) − Γ(4 k + 1) − 13β k ]
k
...(5.6)
From (5.3) - (5.6), the coecients of skewness and kurtosis can be easily calculated
for selected values of β and k.
We also obtained the probability weighted moments of a continuous random
variable X with distribution function given by (2.2) which are the quantities
Mp,r,s = E[Xp {F(X)}r {1 – F(X)}s] ...(5.7)
for real p, r, and s [Greenwood (1979)]. For the EGEV distribution, we have

1 s
 s  (−1) j βΓ (kp + 1)
Mp,r,s = k {1 − [ ∑   kp +1
]} for s ∈ Z + ⊂ ℜ ...(5.8)
j =0  j  [ β (r + j + 1)]

The result (5.7) generalizes those obtained by Dupuis (1996) since for p = 1 and
s = 0, we have

1 βΓ (k + 1)
M1,r,0 = k {1 − }; r > − 1 ...(5.9)
[ β (r + 1)]k +1
which when β = 1, reduces to the result of Dupuis (1996). Also by setting r = 0 in
(5.9) with β = 1, we obtain another result of Dupuis (1996).

6. MAXIMUM ENTROPY PROPERTY

The entropy of a one-dimensional random variable with density function f(x) is
defined as
H(X) = E [–logf (X)] ...(6.1)
and measures the uncertainty associated with the distribution of X Jaynes (1957). A
straight forward calculatiobn shows that the entropy of X ~ EGEV (β,k,µ,σ), a
parametrized EGEV distribution, is
The Exponentiated Generalized Extreme Value Distribution 45

β
H(X) = 1 – (1 – k)Γ′(1) – (2 – k) log   ...(6.2)
σ 
The concept of entropy has been successfully applied in a variety of fields
including statistical mechanics, statistics, stock market analysis, queuing theory,
image analysis, reliability estimation [Ayebo and Kozubowski (2003)] and the
reference therin. Finding the maximum entropy distribution is more or less a general
inference procedure.
For any random variable with pdf as in (2.2) satisfies

1
 x − µ k β
E [log  1 + k  ] = (1 − k ) (Γ′(1) − log ) ...(6.3)
 σ  σ
Since
1

∫
H(X) = – k f ( x ) log f ( x ) dx = H ( X EGEV )
−∞

Therefore, the EGEV distribution makes Shanon entropy the maximum.

7. ESTIMATION OF PARAMETERS
By introducing location and scale parameters to the EGEV distribution, we consider
their maximum likelihood estimation. The log-likelihood for a random sample x1,
x2, ..., xn from (2.2) is
1
1  n  x −µ  n
 x − µ k
logL(β, µ, σ, k) = n logβ – n log σ +  − 1  ∑ log  1 − k i  − β ∑  1− k i
 k  i =1  σ  i =1  σ 
...(7.1)
The first order derivative of (7.1) with respect to the four parameters are:

1
∂ log L n n  x − µ k
= − ∑ 1 − k i
σ 
...(7.2)
∂β β i =1 

1
k −1 β n
 x − µ k
∂ log L n
 x −µ  σ
− ∑ 1 − k i σ 
∂µ
= σ
∏ log  1 − k i σ  i =1 ...(7.3)
i =1
46 Shola ADEYEMI & Atinuke O. ADEBANJI

1
∂ log L β n
 x − µ k 1  n
∂σ
=
k
∑  1 − k i σ  − n  k − 1 − σ
i =1

and

 1  x − µ 
n  i 
∂ log L n  x −µ   k  σ 
βψ ( µ , σ , k ) − 2 log  i  −
∂k
= k  σ  x −µ
1− k i
σ
...(7.5)
1
∂  x − µ k
where ψ(µ, σ, k) = ∑ i =1  1 − k i
n

∂k  σ 

Setting these expressions to zero and solving them simultaneously yields the
maximum likelihood estimates of the four parameters. Unfortunately, but for β, all
other derivatives are nonlinear and can only be solved numerically for application
purposes. These are the areas of our next study, getting the best estimation method
for the parameters and their applications.

8. CONCLUSION
In this paper, we have introduced the exponentiated generalized extreme-value
distribution. This distribution is more flexible than the ordinary generalized extreme-
value distribution and can be used to suit many application purposes. The aim of
this paper is to study some mathematical properties of the distribution. Some other
properties are yet to be looked into. Characterizing properties of the distribution and
applications are other areas to be studied.

REFERENCES
[1] Adeyemi, S. The exponentiated generalized Pareto distribution. Submitted.
[2] Adeyemi, S. and M.O. Ojo A generalization of the Gumbel distribution. Kragujevac
Journal of Mathematics 25 (2003), pp. 19 - 29.
[3] Ayebo, A. and T.J. Kozubowski An asymetric generalization of Gaussian and Laplace
laws. (2003) to appear in Journal of Probability and Statistical Science.
[4] Dupuis, D.J. Estimating the probability of obtaining nonfeasible estimates of the
The Exponentiated Generalized Extreme Value Distribution 47

generalized extreme-value distribution. J. Statist. Comput. Simul. 56, (1996), 23-38.

[5] Fisher, R.A. and Tippett, L.H.C. Limiting forms of the frequency distribution of the
largets and the smallest member of a sample. Proc. Cambrdge Phil. Soc. 24, (1928),
180-190.
[6] Greenwood, J.A., Landwetor, J.M., Matalas, N.C. and Wallis, J.R. Probability
weighted moments: Definition and Relation to parameters of several distributions
expressable in inverse form. Water Resources Res., 15, (1979), 1049-1054.
[7] Gupta, R.C., Gupta, R.L. and Gupta, R.D. Modelling failure time data by Lehman
and Weibull distributions. Comm. Statist. Theory and Method 27, (1998), 887-984.
[8] Gupta, R.D. and D. Kundu exponentiated exponential family: an alternative to gamma
and Weibull distributions. Biometrical Journal, 43, 117-130.
[9] Hosking, J.R.M., Wallis, J.R. and Wood, E.F. Estimation of the generalized
extremevalue distribution by the method of probability weighted moments.
Technometrics, 27 (1985), 251-261.
[10] Jaynes, E.T. Information Theory and Statistical Mechanics I. Phys. Rev. 106, (1957),
620-630.
[11] Nadarajah, S. and S. Kotz The exponentiated Fretchet distribution InterStat. December
2003, 1, pp. 1-7.
[12] Nadarajah, S. and S. Kotz On the exponentiated exponential distribution 2003, to
appear in Statistica.

Shola ADEYEMI
Department of Mathematics,
Obafemi Awolowo University Ile-Ife, Nigeria.
e-mail: sholadeyemi2003@[Link]

Atinuke O. ADEBANJI
Department of Mathematical Sciences,
University of Agriculture, Abeokuta, Nigeria
e-mail: tinuadebanji@[Link]

View publication stats