Kumaraswamy distributin:

KUMARASWAMY DISTRIBUTIONS: A NEW FAMILY OF
GENERALIZED DISTRIBUTIONS
Pankaj Das
M.Sc. (Agricultural Statistics), Roll No: 20394
ICAR-I.A.S.R.I., Library Avenue, New Delhi-110012
Abstract
Kumaraswamy introduced a distribution for double bounded random processes with hydrological
applications. For any continuous baseline G distribution, G.M. Cordeiro and M. de Castro describe
a new family of generalized distributions (denoted with the prefix "Kw") to extend the normal,
Weibull, gamma distributions, among several well-known distributions. Some special distributions
in the new family such as the Kw-normal, Kw-Weibull, and Kw-gamma distribution are discussed.
We discuss the ordinary moments of any Kw generalized distribution as linear functions of
probability weighted moments of the parent distribution. We also obtain the ordinary moments of
order statistics as functions of probability weighted moments of the baseline distribution. We use
the method of maximum likelihood to fit the distributions in the new class and illustrate the
potentiality of the new model with two application to real data.
Keywords: gamma distribution; Kumaraswamy distribution; moments; normal distribution; order
statistics; Weibull distribution
1. Introduction:
Beta distributions are very versatile and a variety of uncertainties can be usefully modeled by them.
In practical situation, many of the finite range distributions encountered can be easily transformed
into the standard beta distribution. In econometrics, many times the data are modeled by finite
range distributions. Generalized beta distributions have been widely studied in statistics and
numerous authors have developed various classes of these distributions. Eugene et al. (2002)
proposed a general class of distributions for a random variable defined from the logit of the beta
random variable by employing two parameters whose role is to introduce skewness and to vary
tail weight. Following the work of Eugene et al. (2002), who defined the beta normal distribution,
Nadarajaha and Kotz (2004) introduced the beta Gumbel distribution, Nadarajaha and Gupta
(2004) proposed the beta Frechet distribution and Nadarajaha and Kotz (2004) worked with the

beta exponential distribution. However, all these works lead to some mathematical difficulties
because the beta distribution is not fairly tractable and, in particular, its cumulative distribution
function (cdf) involves the incomplete beta function ratio.
Poondi Kumaraswamy proposed a new probability distribution for variables that are lower and
upper bounded. In probability and statistics, the Kumaraswamy's double bounded distribution is a
family of continuous probability distributions defined on the interval [0, 1] differing in the values
of their two non-negative shape parameters, a and b. In reliability and life testing experiments,
many times the data are modeled by finite range distributions.
Eugene et al (2004) and Jones (2004) constructed a new class of Kumaraswamy generalized
distribution (Kw-G distribution) on the interval (0, 1). The probability density function (pdf) and
the cdf with two shape parameters a >0 and b > 0 defined by
-1 -1
( ) (1 - ) ( ) 1- (1 - )a a b a b
f x abx x and F x x  (1)
Where x 
respectively, where 𝑓(𝑥) = 𝑑𝐹(𝑥)/𝑑𝑥 and𝑎,𝑏 > 0 are additional shape parameters to the distribution
F. Except for some special choices of the function 𝐹(𝑥).The associated hazard rate function (hrf)
is
-1
( ) ( )
( )=
1- ( )
a
a
abg x G x
h x
G x
2. Conversion of a distribution into Kw-G distribution:
Let a parent continuous distribution having cdf G(x) and pdf g(x). Then by applying the quantile
function on the interval (0, 1) we can construct Kw-G distribution (Cordeiro and de Castro, 2009).
The cdf F(x) of the Kw-G is defined as
( ) 1 {1 ( ) }a b
F x G x   (2)
Where a > 0 and b > 0 are two additional parameters whose role is to introduce skewness and
to vary tail weights.
Similarly the density function of this family of distributions has a very simple form
1 1
( ) ( ) ( ) {1 ( )}a b
f x abg x G x G x 
  (3)

3. Some Special Kw generalized distributions:
3.1. Kw- normal:
The KN density is obtained from (3) by taking G (.) and g (.) to be the cdf and pdf of the normal
2
( , )N   distribution, so that
1 1
( ) ( ){ ( )} {1 ( ) }a a bab x x x
f x
  

   
   
    (4)
where ,x   is a location parameter, σ > 0 is a scale parameter, a, b > 0 are shape
parameters, and (.) and Ф (.) are the pdf and cdf of the standard normal distribution,
respectively. A random variable with density f (x) above is denoted by X ~ Kw-N
2
( , , , )a b  
For µ= 0 and σ = 1 we obtain the standard Kw-N distribution. Further, the Kw-N distribution
with a = 2 and b = 1 coincides with the skew normal distribution with shape parameter equal
to one.
3.2.Kw-weibull:
The cdf of the Weibull distribution with parameters β > 0 and c > 0 is ( ) 1 exp{ ( ) }c
G x x  
for x > 0. Correspondingly, the density of the Kw-Weibull distribution, say Kw-W (a, b, c, β),
reduces to
1 1 1
( ) exp{ ( ) }[1 exp{ ( ) }] {1 [1 exp{ ( ) }] }c c c c a c a b
f x abc x x x x     
        (5)
Here X, a, b, c, β > 0
If c = 1 we obtain the Kw-exponential distribution. The Kw-W (1, b, 1, β) distribution
corresponds to the exponential distribution with parameter β* = bβ.
3.3. Kw-gamma:
Let Y be a gamma random variable with cdf ( ) ( ) / ( )yG y      for y, α, β > 0, where 
(-) is the gamma function and 1
0
( )
z
t
z t e dt
  
   is the incomplete gamma function. The

density of a random variable X following a Kw-Ga distribution, say X ~ Kw-Ga (a, b, β, α),
can be expressed as
1
1 1
( ) ( ) { ( ) ( )}
( )
x
a a b
x xab
ab x e
f x
  
 

  

 
 
   

, x, α, β, a, b >0 (6)
For α=1, we obtain the Kw-exponential distribution.
Figure 1. Some possible shapes of density function of Kw-G distribution. (a) Kw-normal (a, b,
0, 1) and (b) Kw- gamma (a, b, 1, α) density functions (dashed lines represent the parent
distributions)
4. A general expansion for the density function:
Cordeiro and de Castro (2009) elaborate a general expansion of the distribution.
For b > 0 real non-integer, the form of the distribution
1 1
0
{1 ( ) } ( 1) ( ) ( )a b i b ai
i
i
G x G x

 

   (7)
where the binomial coefficient is defined for any real. From the above expansion and formula (3),
we can write the Kw-G density as

( 1) 1
0
( ) ( ) ( )a i
i
i
f x g x wG x

 

  (8)
Where the coefficients are
1
( , ) ( 1) ( )i b
i i iw w a b ab 
   and
0
0i
i
w



5. General formulae for the moments:
The s-th moment of the Kw-G distribution can be expressed as an infinite weighted sum of
PWMs of order (s, r) of the parent distribution G from equation (8) for a integer and for a real
non-integer. We assume Y and X following the baseline G and Kw-G distribution, respectively.
The s-th moment of X, say µ's, can be expressed in terms of the (s, r)-th PWMs
 { }
rs
sr E Y G Y 
of Y for r = 0, 1 ..., as defined by Greenwood et al. (1979). For a integer,
'
, ( 1) 1
0
s r s a r
r
w 

 

  (9)
Whereas for a real non integer the formula
'
, , ,
, 0 0
s i j r s r
i j r
w 
 
 
  
(10)
We can calculate the moments of the Kw-G distribution in terms of infinite weighted sums of
PWMs of the G distribution. Established power series expansions to calculate the moments of any
Kw-G distribution can be more efficient than computing these moments directly by numerical
integration of the expression.
6. Probability weighted moments:
A general theory for PWMs covers the summarization and description of theoretical probability
distributions, the summarization and description of observed data samples, nonparametric
estimation of the underlying distribution of an observed sample, estimation of parameters and

quantiles of probability distributions and hypothesis testing for probability distributions. (Barakat
and Abdelkader, 2004)
The (s,r)-th PWM of X following the Kw-G distribution, say ,
Kw
s r , is formally defined by
, { ( ) } ( ) ( )Kw s r s r
s r E X F X x F x f x dx


  
(11)
This formula also can be written in the following form
, , , , ,
, , 0 0
( , )Kw
s r r m u v l s m l
m u v l
v
p a b w  
 

  
(12)
Where ,s m l 
is the (s,m+l)-th PMW of G distribution and the coefficients
,
0 0
( , ) ( )( 1) ( 1) ( )( )( )
u
u k mr l kb ma l
r m k m l r
k m l r
p a b
 

  
   
7. Order statistics:
The density  :i nf x of the i-th order statistic, for i = 1,..., n, from i.i.d. random variables X1,... ,Xn
following any Kw-G distribution, is simply given by
  1 1
:
( )
( ) {1 ( )}
( , 1)
i
i
n
n
f x
F x Fx
B n i
f x
i
 
 
 
= 1 ( 1) 1
( ) ( ) [1 {1 ( ) } ]{1 ( ) }
( , 1)
i a b a b n iab
g x G x G x G x
B i n i
   
  
 
(13)
Where B(.,.) denote the beta function and then
  1
0
:
( )
( 1) ( ) ( )
( , 1)
n i
j n i i j
j
i n j
f x
F x
B i n i
f x

  

 
  (14)
After expanding all the terms we get the following two forms

When a= non integer
  , , , 1
0 , , 0 0
:
( )
( 1) ( ) ( , ) ( )
( , 1)
n i v
j n i r t
j u v t r i j
j r u v t
i n
g x
w p a bx G x
B i
f
i n
 
 
 
  
 
 
  
(15)
And when a= integer
  ( 1) 1
,: 1
0 , 0
( )
( 1) ( ) ( )
( , 1)
n i
j n i a u r
j u r i j
r
n
j
i
u
g x
w p abG x
B i n
f x
i
 
   
 
 
 
 
  (16)
Formulae (15) and (16) immediately yield the density of order statistics of the Kw-G distribution
as a function of the density of the baseline distribution multiplied by infinite weighted sums of
powers of G(x). Hence, the ordinary moments of order statistics of the Kw-G distribution can be
written as infinite weighted sums of PWMs of the G distribution. These generalized moments for
some baseline distributions can be accurate computationally by numerical integration as mentioned
before.
8. L moments:
The L-moments are analogous to the ordinary moments but can be estimated by linear
combinations of order statistics. The L-moments have several theoretical advantages over the
ordinary moments. They exist whenever the mean of the distribution exists, even though some
higher moments may not exist. They are able to characterize a wider range of distributions and,
when estimated from a sample, are more robust to the effects of outliers in the data. Unlike usual
moment estimates, the parameter estimates obtained from L-moments are sometimes more
accurate in small samples than even the maximum likelihood estimates (MLEs). The L-moments
are linear functions of expected order statistics defined as
(17)
the first four L-moments are 1 1:1( )E X  , 2 2:2 1:2
1
( )
2
E X X   , 3 3:3 2:3 1:3
1
( 2 )
3
E X X X   
1 1 : 1
0
1( 1) ( 1) ( ) ( )
r
k r
r k r k r
k
r E X    

  

and 4 4:4 3:4 2:4 1:4
1
( 3 3 )
4
E X X X X     . The L-moments can also be calculated in terms of
PWMs given in (12) as
1 1,
0
( 1) ( )( )r k r r k Kw
r k k k
k
 

 


 
(18)
In particular, 1 1,0 2 1:1 1:0 3 1:2 1:1 1:0, 2 , 6 6Kw Kw Kw Kw Kw Kw
             
9. Other measure:
9.1. Mean deviations:
Mean deviation denotes the amount of scatter in a population. This is evidently measured to some
extent by the totality of deviations from the mean and median. Let X ∼ Kw-G (a, b). The mean
deviations about the mean (δ1(X)) and about the median (δ2(X)) can be expressed as
1 ' ' '
1 1 1 1 1( ) ( ) 2 ( ) 2 ( )X E X F T        and '
2 1( ) ( ) 2 ( )X E X M T M     Where
'
1 ( )E X  , M = median, '
1( )F  is come from pdf and ( ) ( )
z
T z xf x dx

 
9.2. Extreme values:
If  1 ··· /nX X X n   denotes the sample mean from i.i.d. random variables following (2),
then by the usual central limit theorem ( ( )) / ( )n X E X Var X approaches the standard normal
distribution as n →∞ under suitable conditions. Sometimes one would be interested in the
asymptotics of the extreme values Mn = max(X1... Xn) and mn = min(X1,...,Xn).
First, suppose that G belongs to the max domain of attraction of the Gumbel extreme value
distribution. Then by Leadbetter et al. there must exist a strictly positive function, say h(t), such
that
1-G(t +xh(t))
lim = exp(-x)
1-G(t)t
for every x ∈ (−∞,∞).So, it follows by Leadbetter et al.(1987)
that F also belongs to the max domain of attraction of the Gumbel extreme value distribution with
lim Pr{ ( ) } exp{-exp(- )}n n n
n
a M b x bx

  

10. Estimation of parameter:
Let γ be the p-dimensional parameter vector of the baseline distribution in equations (2) and (3).
We consider independent random variables X1,..., Xn, each Xi following a Kw-G distribution with
parameter vector θ = (a,b, γ). The log-likelihood ( ) function for the model parameters
obtained from (3) is
1 1 1
( ) {log( ) log( )} log{ ( ; )} ( 1) log{ ( ; )} ( 1) log{1 ( ; ) }
n n n
a
i i i
i i i
n a b g x a G x b G x   
  
         
The elements of the score vector are given by
1
( 1) ( ; )( )
log{ ( ; )}{1 }
1 ( ; )
an
i
i a
i i
b G xd n
G x
da a G x




  


1
( )
log{1 ( ; ) }
n
a
i
i
d n
G x
db b



  
And
1
( ; ) ( ; )( ) 1 1 ( 1)
[ {1 }
( ; ) ( ; ) ( ; ) 1
n
i i
a
ij i i i
dg x dG xd a b
d g x d G x d G x
 
      


  


These partial derivatives depend on the specified baseline distribution. Numerical maximization
of the log-likelihood above is accomplished by using the RS method (Rigby and Stasinopoulos,
2005) available in the gamlss package in R. Since numerically the maximum likelihood estimation
of the parameters of the Kw-G distributions is much simpler than the estimation of the parameters
of the generalized beta distributions, we recommend to use Kw-G distributions in place of the
second family of distributions.
We can compare non-nested Kw-G distributions by penalizing over-fitting using the Akaike
information criterion given by AIC =
^
2 ( ) 2 *p  , where p* is the number of model
parameters. The distribution with the smallest value of AIC (among all distributions considered)
is usually taken as the best model for describing the given data set. This comparison is based on
the consideration of a model that shows a lack of fit with one that does not.

11. Relation to the Beta distribution: The density function of beta distribution is defined as
1 11
( ) ( ) ( ) {1 ( )}
( , )
a b
f x g x G x G x
B a b
 
 
Assume that ,a bX is a Kumaraswamy distributed random variable with parameters a and b. Then
,a bX is the a-th root of a suitably defined Beta distributed random variable. More formally, Let 1,bY
denote a Beta distributed random variable with parameters 1  and b  . One has the following
relation between ,a bX and 1,bY .
1/
, 1,
a
a b bX Y
With equality in distribution.
1 1 1 1/
, 1, 1,
0 0
{ } (1 ) (1 ) { } { }
a
x x
a a b b a a
a b b bP X x abt t dt b t dt P Y x P Y x  
         
Jones (2008) explored the background and genesis of the Kw distribution and, more importantly,
made clear some similarities and differences between the beta and Kw distributions. For example,
the Kw densities are also unimodal, increasing, decreasing or constant depending in the same way
as the beta distribution on the values of its parameters. He highlighted several advantages of the
Kw distribution over the beta distribution: the normalizing constant is very simple; simple explicit
formulae for the distribution and quantile functions which do not involve any special functions; a
simple formula for random variate generation; explicit formulae for L-moments and simpler
formulae for moments of order statistics. Further, according to Jones (2008), the beta distribution
has the following advantages over the Kw distribution: simpler formulae for moments and moment
generating function; a one-parameter sub-family of symmetric distributions; simpler moment
estimation and more ways of generating the distribution via physical processes.
12. Application:
We illustrate the superiority of some new Kw-G distributions proposed here as compared with
some of their sub-models. We give two applications (uncensored and censored data) using well-
known data sets to demonstrate the applicability of the proposed regression model.

12.1. Censored data:
In this section we present an example with data from adult numbers of Flour beetle (T. confusum)
cultured at 29°C presented by Cordeiro and de Castro (2009).
Table 1 gives AIC values in increasing order for some fitted distributions and the MLEs of the
parameters together with its standard errors. According to AIC, the beta normal and Kw-normal
distributions yield slightly different fittings, outperforming the remaining selected distributions.
Notice that for the beta normal distribution the variability in the estimates of a, µ and σ is
appreciably greater.
The fitted distributions superimposed to the histogram of the data in Figure 3 reinforce the result
in Table 1 for the gamma distribution. The beta normal and the Kw-normal distributions are almost
indistinguishable. This claim is further strengthened by the comparison between observed and
expected frequencies in Table 2. The mean absolute deviation between expected and observed
frequencies reaches the minimum value for the Kw-normal distribution.

Figure 3. Histogram of adult number and fitted probability density functions.
Table 2. Observed and expected frequencies of adult numbers for T. confusum cultured at 29°C
and mean absolute deviation (MAD) between the frequencies

Based on the values of the LR statistic (Section 8), the Kw-gamma and the Kw-exponential
distributions are not significantly different yielding LR = 1.542 (1 d.f., p-value = 0.214).
Comparing the Kw-gamma and the gamma distributions, we find a significant difference (LR =
6.681, 2 d.f., p-value = 0.035).
12.2. Uncensored data (voltage):
Here, we compare the results of Nadarajaha et al (2011). They fits of some distributions to a data
set which gives the times of failure and running times for a sample of devices from a field-tracking
study of a larger system. At a certain point in time, 30 units were installed in normal service
conditions. Two causes of failure were observed for each unit that failed: the failure caused by an
accumulation of randomly occurring damage from power-line voltage spikes during electric storms
and failure caused by normal product wear. The required numerical evaluations were implemented
using the SAS procedure NLMIXED. Table 3 lists the MLEs (and the corresponding standard
errors in parentheses) of the parameters and the values of the following statistics for some fitted
models: AIC (Akaike information criterion), BIC (Bayesian information criterion) and CAIC
(Consistent Akaike information criterion). These results indicate that the Kw-Weibull model has
the lowest AIC, CAIC and BIC values among all fitted models, and so it could be chosen as the
best model.
13. Conclusion:
Following the idea of the class of beta generalized distributions and the distribution by
Kumaraswamy, we define a new family of Kw generalized (Kw-G) distributions to extend several
widely-known distributions such as the normal, Weibull, gamma and Gumbel distributions. For
each distribution G, we can define the corresponding Kw-G distribution using simple formulae.

We show how some mathematical properties of the Kw-G distributions are readily obtained from
those of the parent distributions. The moments of the Kw-G distribution can be expressed explicitly
in terms of infinite weighted sums of probability weighted moments (PWMs) of the G distribution.
The same happens for the moments of order statistics and PWMs of the Kw-G distributions.
We discuss maximum likelihood estimation and inference on the parameters. The maximum
likelihood estimation in Kw-G distributions is much simpler than the estimation in beta generalized
distributions. Further, we can easily compute the maximum values of the unrestricted and restricted
log-likelihoods to construct likelihood ratio statistics for testing nested models in the new family
of distributions. An application of the new family to real data is given to show the feasibility of
our proposal.
References:
Azzalini, A. (1985). A class of distributions which includes the normal ones. Scandinavian Journal
of Statistics. 12:171-178.
Barakat, H. M. and Abdelkader, Y. H. (2004). Computing the moments of order statistics from
nonidentical random variables. Statistical Methods and Applications. 13:15-26.
Barlow, R. E. and Proschan, F. (1975). Statistical theory of reliability and life testing: probability
models. Holt, Rinehart and Winston, New York, London.
Cordeiroa, Gauss M. and Castrob, Mario de (2009). A new family of generalized distributions.
Journal of Statistical Computation & Simulation. 00: 1-17.
Eugene, N., Lee, C., and Famoye, F. (2002). Beta-normal distribution and its applications.
Communications in Statistics. Theory and Methods. 31:497-512.
Fletcher, S. C. and Ponnambalam, K. (1996). Estimation of reservoir yield and storage distribution
using moments analysis. Journal of Hydrology. 182: 259-275.
Greenwood, J. A., Landwehr, J. M., Matalas, N. C. and Wallis, J. R. (1979). Probability weighted
moments - definition and relation to parameters of several distributions expressable in
inverse form. Water Resources Research. 15:1049-1054.

Hosking, J. R. M. (1990). L-moments: analysis and estimation of distributions using linear
combinations of order statistics. Journal of the Royal Statistical Society. Series B. 52:105-
124.
Jones, M. C. (2004). Families of distributions arising from distributions of order statistics (with
discussion).Test. 13:1-43.
Jones, M. C. (2008). Kumaraswamy's distribution: A beta-type distribution with some tractability
advantages. Statistical Methodology. 6:70-81.
Kumaraswamy, P. (1980). Generalized probability density-function for double-bounded random-
processes. Journal of Hydrology. 462:79-88.
Leadbetter, M.R., Lindgren, G. and Rootzén, H. (1987). Extremes and Related Properties of
Random Sequences and Processes. Springer, New York, London.
Nadarajaha, S. and Gupta, A. K. (2004). The beta Frechet distribution. Far East Journal of
Theoretical Statistics. 14:15-24.
Nadarajaha, S. and Kotz, S. (2006). The beta exponential distribution. Reliability Engineering &
System Safety. 91: 689-697.
Nadarajaha, S., Cordeirob, Gauss M. and Ortegac, Edwin M. M. (2011). General results for the
Kumaraswamy-G distribution. Journal of Statistical Computation and Simulation. iFirst:
1-29.
Rigby, R. A. and Stasinopoulos, D. M.(2005). Generalized additive models for location, scale and
shape (with discussion). Applied Statistics. 54:507-554.
R Development Core Team. (2009). R: A Language and Environment for Statistical Computing.
R Foundation for Statistical Computing. Vienna, Austria.
Sundar, V. and Subbiah, K. (1989). Application of double bounded probability density-function
for analysis of ocean waves. Ocean Engineering. 16:193-200.
Seifi, A., Ponnambalam, K. and Vlach, J. (2000). Maximization of manufacturing yield of systems
with arbitrary distributions of component values. Annals of Operations Research. 99:373-
383.

Stasinopoulos, D. M. and. Rigby, R. A. (2007). Generalized additive models for location scale and
shape (GAMLSS) in R. Journal of Statistical Software. 23:1-46.

Kumaraswamy distributin:

More Related Content

What's hot (17)

Similar to Kumaraswamy distributin: (20)

Recently uploaded (20)

Kumaraswamy distributin: