Open Journal of Statistics, 2012, 2, 415-419
http://dx.doi.org/10.4236/ojs.2012.24050 Published Online October 2012 (http://www.SciRP.org/journal/ojs)
Maximum Entropy and Maximum Likelihood Estimation
for the Three-Parameter Kappa Distribution
Bungon Kumphon
Department of Mathematics, Faculty of Sciences, Mahasarakham University, Mahasarakham, Thailand
Received June 27, 2012; revised July 29, 2012; accepted August 11, 2012
ABSTRACT
The two statistical principles of maximum entropy and maximum likelihood are investigated for the three-parameter
kappa distribution. These two methods become equivalent in the discrete case with ,0x

 where 0




1211,0,1,2,kk 

, for the maximum entropy method.
Keywords: Maximum Entropy; Maximum Likelihood; Kapp a Distribution; Lagran ge Multiplier
1. Introduction
Statistical entropy deals with a measure of uncertainty or
disorder associated with a probability distribution. The
principle of maximum entropy (ME) is a tool for infer-
ence under uncertainty [1,2]. This approach produces the
most suitable probability distribution given the available
information as seeks the probability distribution that
maximizes the information entropy subject to the infor-
mation constraints, typically via the method of Lagrange
multipliers. More precisely, the result is a probability
distribution that is consistent with the known constraints
expressed in terms of averages or expected values of one
or more quantities, but is otherwise as unbiased as possi-
ble—i.e. one obtains the least-biased estimate possible on
the given information, maximally noncommittal with re-
gard to missing information.
A family of positively skewed distributions known as
kappa distributions introduced by Mielke [3] and Mielke
and Johnson [4], is very popular for analyzing precipita-
tion data (cf. Park et al. [5], Kysely and Picek [6], Du-
puis and Winchester [7]). Various methods of estimation
for this type of data include the L-moment, Moment, and
Maximum Likelihood (ML) techniques. Many research
papers have shown that the ML is too sensitive to ex-
treme values, especially for small samples although but it
may be satisfactory for large samples, and the final esti-
mate is not always a global maximum because it can de-
pend upon the starting values. The ME can remove this
ambiguity, as various authors have shown—e.g. Hradil
and Rehacek [8], and Papalexious and Koutsoyiannis [9].
Singh and Deng [10] considered the ME method for the
four-parameter kappa distributions, which include the three-
parameter kappa distribution (K3D) introduced by Mei lke
[3]. In this study, we investigate the theoretical back-
ground for parameter estimation by the ME method in
the K3D case. The limitation of its performance com-
pared to the ML method is also discussed.
2. Three-Parameter Kappa Distribution
Let a random variable be denoted by X. The distribution
function of the three-parameter kappa distribution (K3D)
is
1
,, 0
x
fx x





 






 

 

,,
(1)
where



denote the location, scale and shape pa-
rameters respectively (Park et al. [5]), and the corre-
sponding cumulative d istribution function of the K3D is
1
,0.
xx
Fx x






 

 

 

(2)
It is notable that the K3D distribution function, Equa-
tion (1), involves adding the location parameter

to
the two-parameter kappa distribution (K2D), in contrast
to Meilke [3] where only a new shape parameter

is
introduced.
3. The Entropy Framework
3.1. Entropy Measure and the Principle of
Maximum Entropy
The concept of entropy was originally developed by
C
opyright © 2012 SciRes. OJS

B. KUMPHON 416
Ludwig Boltzmann in statistical mechanics. A famous
and well justified measure is the Boltzmann-Gibbs-
Shannon (BGS) entropy
 
ln dxfxx

0
Sf


 (3)
for a continuous non-negative random variable X, where
f
x is the probability density fu nction of X. The given
information used in the principle of maximum entropy
(ME) is expressed as a set of constraints representing
expectations of funct i o ns


j
g
X
1,2,,
j
j n 

0
d1fx x

—i.e.
 
0
d,
jj
EgXg xfx x c




 (4)
ME distributions emerge by maximizing the selected
form of entropy, subject to Equation (4) and the obvious
additional con strain t



1
n
jj
j
. (5)
As precisely mentioned, the maximization is usually
accomplished via the method of Lagrange multipliers,
such that the general solution form of the ME distribu-
tions from maximizing the BGS entropy Equation (3)
(Levine and Tribus, [11]) is

0
exp
f
x
 gx







,1,2,,jn
, (6)
where j

, are the Lagrange multipliers
linked to the constraints in Equation (4) and 0

is the
multiplier linked to the additional constraint Equation
(5).
3.2. Justification of the Constraints
Samples are drawn from positively skew or heavy-tailed
distributions, located on the right far from the mean. Sta-
tistically, such values are considered to be outliers and
consequently strongly influence the sample moments.
The logarithm function is applied to the data set to
eliminate the influence of extreme values. The maximum
entropy distribution is uniquely defined by the chosen
constraints, which normally contain information from
observations or theoretical considerations. Thus in geo-
physical applications for example, important prior char-
acteristics of the underlying distribution should be pre-
served—e.g. a J-shaped, Bell-shape or heavy-tailed dis-
tribution. The constraints should also be chosen based on
the suitability of the resulting distribution in regard to the
empirical evidence. More details on appropriate con-
straints are discussed in [11]. In this study, we choose a
single constraint to express the features of the distribu-
tion given the empirical evidence.
3.3. The Estimation of Maximum Entropy
There are four steps in the ME method to estimate the
objective distribution—viz.
1) Specification of appropriate constraints;
2) Construction of the Lagrange multipliers;
3) Derivation of the entropy function of the distribu-
tion; and
4) Derivation of the relation between the Lagrange
multiplier and the constraints.
Step 1 Specification of Appropriate Constraints.
Taking the natural logarithm, from (1) we have

1
lnlnln x fx

 

 

 

 

 
 



(7)
To establish the entropy as expressed in Equation (3),
multiply Equation (7) with
f
x
(0, ) and integrate over
the entries space

to obtain

0
1
ln lnd
x
Sfxx

 

 




  
 
 


 (8)
which is to be subject to the constraints

0
lnd ln
xx
fx xE







 





 


  





0
d1fx x

(9)
. (10)

Step 2 Construction of the Lag r ange Multipliers.
From Equation (6)

01
explnx fx


 






  




 

, (11)


, where 01

 are the Lagrange multipliers. Substituting
Equation (11) into Equat i o n ( 10) we have
01
0
explnd 1
xx


 






  


 




, (12)
such that




 


 


1
1
1
11
01
0
11
00
11
11
0
11 11
11
0
expexp lnd
dd
1d
11d (13)
xx
x
x
zz z
uu u
uu u








 






 











 





 







 
  

 


 





B. KUMPHON 417

On setting x z


 



such that


11
dd
x
zz
 


, and z
u

zu such that dd


.
Since from Equation (13) we require

0
exp 0  

112,   0,1,2, ,kk implying

01 211   , 0,1,2,kk.
Consequently,






11
1
11 11
00
1
11
1
exp 1
1
1
 






 




11
d
1
.
1
uuu













(14)
Then on taking logarithms we have
  
01
1
ln1 1 ln
lnln 1
1
ln 1
11
.

 

 
 
 
 
 
 





Step 3 Derivation of the Entropy Function of the Dis-
tribution.
Substituting Equation (14) into Equation (11) gives
 

1
1
11
x
1
1
1
1
1
1
fx



























 


and again taking the natural logarithms, we have

11
1
1
1 ln
,

1
1
ln( )ln1ln
11
ln ln
ln
fx
x










 

 


 
 
 
 

 
 

 

and hence from the definition of entropy, Equation (3),



11
1
1
1 ln
.

1
1
ln 1ln
1
lnln 1
ln
S
x
E











 











 


 


(15)
Step 4
Derivation of the Relation between the La-
grange Multipliers and Constraints.
Let 1
a

 such that 2 dda1,


 1 1b


1
ddb
such
that


, and 11
1c 


 such that
1
1
c





and 2
1c



.
Since
 
ln tt
t


 is the digamma function, it
follows that

2
11
ln a





 ,
 
1
1
ln 1b 

 
,

12
11
ln 1c

 

 

 ,
and

1
1
1
ln 1c



 


,,
.
There are four parameters in Equation (15)—viz.

 and 1

. To maximize Equation (15), we need
to set the following partial d erivative to zero:
  
1
12
1
1
22
1
11
ln 1ln
111
1ln
11ln
1111
ln 1
11ln
S
xx
Ex
ca
xx
Ex





 

 




 
 




 

 

 














 




  
(16)
















 






,






















 
11
11 1
1
ln1ln 1
lnln 1
=lnln
0
S
x
E
x
cb E



 



 
 

 

 



 
 
 











 


 




B. KUMPHON
OJS
418
  
222
111
lnln 1cbE  

,
x









 



 






(17)

11
1,
xx
E
xx












 








1
SE

 







 








(18)
and
1
11SE

  
  








 







1
xx
E
xx







 














(19)
Assuming 11




, Equations (16) and (17) yield
 
22 2
2
1211
ln 11ln
111
ln 11
ba
Sx x
EE
x
xx
CEE x
 




 
  

 
  


 
  

    
 

 


  

 
 



   



 
  

 
 
 




x 




 










ln
x










 


 




(20)


where
  
2
Cb
12
a 



,,
is a constant.
The parameter estimation for the K3D (i.e. of


0) from the ME for

 and


012110,1,kk   

2,, are obtained from
Equations (18)-(20), respectively. By the definition of
expectation of random variable


x
EYyPY y

all
assume that


PY y equal to 1, thus Equation (18),
11
,
i
i
x
x









 
 
 
   
 

 
 
 
 
 
 
1
n
i
x
Ex




 


 


 






 (21)
and apply this assumption to Equations (19) and (20).
4. The Maximum Likelihood Estimation
From Equation (1), the log-likelihood function can be
written as
1
ln, ,
1
ln ln
ni
i
L
x
n




 
 






  
 



 (22)
where i
x
is the i-th value of the random variable
X
and is a sample size. Multiply with –1 and differen-
tiating Equation (22) partially with respect to each pa-
rameter, we obtain the MLE by equating each of the fol-
lowing partial derivatives to zero:
n

1
1
ln 1
i
n
ii
x
Lx



















 












 (23)
1
1
ln
i
n
ii
x
n
Lx



 




 


 
 



 






 (24)

B. KUMPHON 419
21
1
1
ln ln
11
1l
ni
i
n
ii
x
n
L
x





 








 









 
n
ii
xx 
 



 


 


 









,0x




(25)
By Equation (21), a comparison of the equations of the
ME and the MLE immediately reveals that Equation (18)
is equivalent to Equation (23), Equation (19) to Equation
(24) and Equation (20) to Equation (25), where


and

01211kk  0,1,2,. Consequently,
the two methods become equivalent for discrete random
variables.
5. Conclusion
A positive skewness distribution, the three-parameter ka p pa
distribution, is considered. Parameter estimation by the
maximum likelihood method requires a certain cutoff in
the parameter space or a best starting value, for otherwise
the solution may appear under-determined instead of a
unique answer (there can exist a concave set). The prin-
ciple of maximum entropy is another tool to address this
problem under constraints that show the characteristic of
the distribution given the empirical evidence, using the
method of Lagrange multipliers. For

01211kk  ,0x0,1,2,, and

 the
principle of maximum entropy method is equivalent to
the maximum likelihood method for the discrete case.
REFERENCES
[1] E. T. Jaynes, “Information Theory and Statistical Me-
chanics,” Physical Review, Vol. 106, No. 4, 1957, p. 620.
doi:10.1103/PhysRev.106.620
[2] E. T. Jaynes, “Prior Probabilities,” IEEE Transactions on
Systems Science and Cybernetics, Vol. 3, No. 4, 1968, pp.
227-241. doi:10.1109/TSSC.1968.300117
[3] P. W. Mielke, “Another Family of Distributions for De-
scribing and Analyzing Precipitation Data,” Journal of
Applied Meteorology, Vol. 12, No. 2, 1973, pp. 275-280.
doi:10.1175/1520-0450(1973)012<0275:AFODFD>2.0.C
O;2
[4] P. W. Mielke and E. R. Johnson, “Three-Parameter Kap pa
Distribution Maximum Likelihood Estimates and Likeli-
hood Ratio Tests,” Monthly Weather Review, Vol. 101,
No. 9, 1973, pp. 701-707.
doi:10.1175/1520-0493(1973)101<0701:TKDMLE>2.3.C
O;2
[5] J. S. Park, S. C. Seo and T. Y. Kim, “A Kappa Distribu-
tion with a Hydrological Application,” Stochastic Envi-
ronmental Research and Risk Assessment, Vol. 23, No. 5,
2009, pp. 579-586. doi:10.1007/s00477-008-0243-5
[6] J. Kysely and J. Picek, “Probability Estimates of Heavy
Precipitation Events in a Flood-Prone Central-European
Region with Enhanced Influence of Mediterranean Cy-
clones,” Advances in Geosciences, Vol. 12, 2007, pp. 43-
50. doi:10.5194/adgeo-12-43-2007
[7] D. J. Dupuis and C. Winchester, “More on the Four-Pa-
rameter Kappa Distribution,” Journal of Statistical Com-
putation and Simulation, Vol. 7, No. 2, 2001, pp. 99-113.
doi:10.1080/00949650108812137
[8] Z. Hradil and J. Rehacek, “Likelihood and Entropy for
Statistical Inversion,” Journal of Physics: Conference Se-
ries, Vol. 36, 2006, pp. 55-59.
[9] S. M. Papalexious and D. Koutsoyiannis, “Entropy Base
Derivation of Probability Distributions: A Case Study to
Daily Rainfall,” Advances in Water Resources, 2012, in
Press. doi:10.1016/j.advwatres.2011.11.007
[10] V. P. Singh and Z. Q. Deng, “Entropy-Based Parameter
Estimation for Kappa Distribution,” Journal of Hydro-
logic Engineering, Vol. 8, No. 2, 2003, pp. 81-92.
doi:10.1061/(ASCE)1084-0699(2003)8:2(81)
[11] R. D. Levine and M. Tribus, “The Maximum Entropy
Formalism,” MIT Press, Cambridge, 1978.