Professional Documents
Culture Documents
A. K. Rajagopal
Department of Physics and Astronomy
INTRODUCTION
There exists a multitude of frequency distributions for
hydrologic analyses. For example, exponential and
Weibull distributions are often used for frequency
analysis of depth, intensity, duration and number of
rainfall events [Eagleson, 1972; Rao and Chenchagya,
1974, 1975; Richardson, 1982]; gamma distribution for
rainfall-runoff modelling [Nash, 1957; Dooge, 1973;
Singh, 1982a] as well as for flood analysis I-Phien and
Jivajirajah, 1984; Yevjevich and Obeysekera, 1984];
extreme-value (EV) type I distribution and its logarithmic
version for flood frequency analysis [Gumbel, 1958;
Todorovic, 1982; Lettenmeier and Burges, 1981; Singh,
1982b]; Pearson type (PT) 3 distribution and its
logarithmic version [Matalas and Wallis, 1977; Bobee
and Robitaille, 1977; Bucket and Oliver, 1977; Kite, 1977;
Rao, 1980, 1983], as well as lognormal distribution
[Kalinske, 1946; Chow, 1954; Sangal and Biswas, 1970;
Burges et al., 1975; Kite, 1977; Kottegoda, 1977]; etc.
Some of the distributions (e.g., gamma, EV1) have been
derived in standard statistical textbooks, but the
approach of derivation has varied from one distribution
to another except for the Pearsonian family. It is therefore
not clear if there is a unified approach which can be
employed to derive any desired distribution. Such an
approach may have several advantages: (1) It may aid in
understanding the distribution by knowing the type of
information needed for its derivation. (2) It may offer an
alternative method of estimating the parameters
AcceptedJanuary 1986.Discussioncloses August 1986.
0309-1708/86/020091-1652.00
1986 ComputationalMechanics Publications
92
:i(r)dx=l
(1)
(1,~)
= W(z)) 2
!
g(z) dz
dz
I
in which
l[f]=-
Z f~ln f~;
~f~=l
i=a
f u9(z) dz = 1
Often f(x) is n o t k n o w n beforehand, although some of its
properties (or constrainsts) may be known, e.g., moments,
lower and upper bounds, etc. These constraints and the
condition in (1) are generally insufficient to define f(x)
uniquely, but may delineate a set of feasible distributions.
Each of these distributions contains a certain amount of
uncertainty which can be expressed by employing the
concept of entropy.
Entropy was first mathematically expressed by Shannon
(1948a, 1948b). I t h a s since been called the Shannon
entropy functional, SEF in short, denoted as I[f] or I[x]
and is a numerical measure of uncertainty associated with
f(x) in describing the random variable x, and defined as
(4)
(2)
THE P R I N C I P L E O F M A X I M U M ENTROPY
as
(POME)
The P O M E formulated by Jaynes (1961, 1982) states that
'the minimally prejudiced assignment of probabilities is
that which maximizes the entropy subject to the given
information.' Matfiematically, it can be stated as follows:
Given m linearly independent constraints C~ in the form
Ci = f f yi(x)f(x) dx,
i= 1, 2. . . . . m
(5)
~ab
I[f] = -
f(x) In f(x)
dx
(3)
f/
f,"
l[f]=l[9]+
= 119] +
9(z) In 9(z) dz
9(z)ln dx dx
f(x) In dzz dx
aLva(x)]
(6)
~ab
6 ( - I) =
[1 + In f(x)]ff(x)
dx
(7)
et al.
Since -J'~9(x)ln f(x) dx= - ~ f ( x ) In f(x)dx in this
problem, because the first m constraints are the same, we
have
6F(f)=-ff[lnf(x)+l+(ao-1)
+ ~=a
~ ayi(x)]6f(x)dx=O
o - ~=~a.,y~(x)l
f(x)
as given by (6) is
I,--
I,.[f] = -
f(x)
In f(x)
dx = ao +
1.[9]
(14a)
This produces
f(x)= exp[-a
l[glf] = I.,[f] -
aiCi
Im[f ] --I[,[9]
(14b)
n - m
(8)
i=1
I.[9]<~I,,[f ]
where
I.[9] = -
and
I"[f]-I"[g]~'2
n>~m
(9)
g(x) In 9(x) dx
(10)
for
gtxl(Jlxt-olxt?
fo
\
9~f
I[f]=ao+ ~ a,C,
dx~>O (11)
f"g(x)InLf(x)A[
g(x)~ dx
(12)
exp - a o -
ao'i d x = l
i=l
(15)
i=1
I[g [[] =
DERIVATION OF FREQUENCY
DISTRIBUTIONS
resulting in
1
(13)
f(x)
and 9(x),
l[glf ] >>-0
ao=lnffexpl-
~ aLv,ldx
(16)
a o
- - ~-~3a~
=Ci
Inserting (6) for f(x) in the right side of this inequality and
the definitions (8) and (10) we get (9).
To obtain (I1) we note that
94
#(x)(f(x)_-g(x)']adx
\ g(x) /
1986,
~a~ = var[y,(x)],
~2a o
=--fbO(x)ln[1q f(x)-g(x)ld
x g ~j
>i+21 f f
6~2ao
0 a ~ j = COV[yi(x)yj(x)],
l[glf]= ;f g(x)ln[~]dx
Volume 9, June
(17)
i#j
(18)
(19)
fo
(26)
f ( x ) = e x p [ - a o - a l x - a 2 In x]
(27)
(20)
exp(a)= f o e x p [ - a l x - a 2 In x] dx
i=1
ln xf(x) dx = E [In x]
-,al,X~
) -a~r(1-a2)
(28)
(a 2 -
1) In a 1 + In [F(1 - a2 )]
(29)
Substituting in (27),
f(x) = exp[(1 - a 2 ) I n a I - I n F(1 - a 2 ) - a l x - a 2 In x]
( a 1 )1 - a 2
-F(1 -a2)
(x)- a~exp( - a i x)
(30)
f(x) =
(x)b 1
a
e x p ( - x/a)
(31)
if
exp(-x/a)
f(x)--a- ~
(21)
a 2 = 1 -- b
(33a)
a I = l/a
(33b)
(22)
t3ao=
Oal
~ x e x p [ - a l x - a 2 lnx] dx
~ exp [ - a 1x - a2 In x] dx
= - f~o x e x p [ - a - a l x - a 2 1 n x] dx
n oo
-fo
=-
xf(x) dx
-(b-
l)
fo
lnxf(x)dx+ a
xf(x)dx
o f ( X dx = 1
f o Xf (x) dx = E[ x] = ~
= -E[x] = -2
(23)
t3ao
da2
(34)
~ In x e x p [ - a l x - a 2 In x] dx
~o exp[ - a x - a In x] dx
(24)
= - f o In x e x p [ - a o - a l x - a 2
(25)
= - f o In xf(x)= - E [ l n x]
In x] dx
(35)
e t al.
II
li
II
II
II
II
II
II
II
'W
II
~.
.o
e. o
o.~
=E
II
II
II
i
=
,.Tp ~
.g
~1'~
~lt'q
~lt'q
II
tl
II
II
elle~
Ir
li
II
II
II
II
II
II
II
teq
II
eq
v/
v/
.o~
.~ ~,.N
,.~ "~, ~
I
E
V
II
11
II
~'
"':
-=
II
-~
,~
~-I~:
~'[-~
e-, 0
"
o
....
96
1986,
Volume
9,
June
""
Ieq~
II
-I~
~
I~
"D
I
~"~
~,
It
II
II
II
'~
t'q
II
"~
'~
~-
II
e~
~, ~
~ .~
,~
~ ,~ ...~
~.~
,o,
L.
--~
~,
"
.-~
..~
II
II
II
tl
...g
-.rl~
i
~.
.~
"7.
~ ,
._.o ~
~ ]
~1xl
V
~I~
II
I ~
tt
'~
:~
.-~
~.
II
--I~1e ' ~ - I
~'~
II
--
-d
t-,
e,
>,-=
t~
.=
~.
E K
m
97
e~
II
_~
II
i ~
j _=
II
II
II
_=
~a
II
II
II
"-~
V
o
A
e~
-~
~.
+
x
II
II
-t-
--~
II
V
o
-~
~=
~rd
98
,,.d
~
--.=
"C
I
II
11
~.~
,,...,.
II
,-~
~,
-.=
r"-n
I
+
r~
II
II
.~ ~
~.~_
II
r~
II
"~
"~
,:~
II
~.,~,,~
II
~.;-
II
,~
~,~
o~
,~
o ~
o~
.o
e,l
.-,,T
+
e~
--~
+
I
-
II
.~
II
II
,~
-I
k.
~~
I
+
~"~
I
I
II
N
+
II
-I.-
~~'
~
",~
'
-~
,~-
I1
II
i ~-
--
r~
77
_~
~
li
II
e~
~
,--,
+
~
_=
I
-~
--=
I
I
e~
II
II
II
_=
-=
+
II
2 ,
"~
[_=
-4-
e.
..7
_=
E~
I
II
..7
II
II
II
.?
.~
..7
.,7
-A
_=
_=
.~
.~~
I
II
_=
+
[.
I
"
_l
e-
100
II
II
+
v
II
( a z - 1)
Oax
a1
(36)
?
~ a - l n al + ~ F ( 1 - a 2 )
~a 2
Let k = 1 - a 2 ,
,--
:--,
#k/~a2 = - 1. Therefore,
Oao
~, + ~
~'z7
ca 2
~k
(37)
II
a 2 -- 1 =.~
a 1
+
or
+
k
1
(38)
In al = qJ(k) - E[ln x]
(39)
--
X,
k =
a 2 -
al
W
t,.,.,a
(40)
e~
._.9.
...
PARAMETER
N
I
"-L
,-.
,.
='=,~
_f
e-
e~
o
.,..~
~"
ESTIMATION
L = l - [ f(xi'O)
(41)
i=l
f ( x ) = Axk e x p [ - i ~ 1 a.J,,(xQ
(19)
In L = ~ In f(xi; O)
(42)
l[f]=-lnA-kE[lnx]+
i=1
N ln f(xi,'O)
---~l l n L = ---~1 ~ ln f ( x i ; O ) = - ~,
i=l
i=l
j=l
t50)
On the other hand,
In L =
i=1
~ a, Z E[yi(xj)]
ln[A~] exp
aiy,lxj)
i=1
(43)
Recall that
= Z l n A + k Z l n x i - Z ai Z yi(xj)
i=1
i=1
i=1
j=l
Multiplying by - ( l / N ) throughout,
i=1
N
~_, f(xi;O)= l
(44)
- ~I l n L = - l n A - k,
'~lnx~
i=l
~
i=1
~a~
~1
yi(xj)
j-
(51)
I [ f ] = - ~ In L
(45)
(46)
Then
N
I[f] = - ~ e e x p ( - e x , ) l n [ e exp(-ex,)]
k=l
= - In + eE[x]
(47)
1%'
(48)
= 1/E[x]
i=1
with
lnL=Nln~-ct
~, f ( x i ; a ) = l
i=l
~ xi
i=1
By maximizing In L,
we obtain
(49)
1-Y~
The difference in the two estimates of ~t given by (48) and
(49) is that the POME method uses expectation ofx or the
pupulation mean, whereas the MLE method uses average
of x or sample mean.
102
f 2 ( x i ; a) ~<l ~ [ f ] ~ N - 1
i=1
V. P. Singh et al.
Table 2. Some pertinent statistical characteristics of annual maximum discharge series for six selected river gaging stations
Area
(km 2)
Mean
~
St.
Deviation
Sx
Skew
coeff,
C~
Kurtosis
coeff.
Ck
38
1 405
238.2
174.5
0.7
2.52
38
1 896
315.7
166.8
0.54
2.77
34
4092
745.1
539.5
0.71
3.03
32
1 890
699.0
223.7
0.41
3.01
36
5 089
1449.7
517.7
0.35
2.55
51
1 659
438.8
159.8
0.71
3.30
Table 3. Parameters of the gamma distribution fated to annual maximum discharge series by MOM, MLE and POME methods
MOM
River Gaging Station
MLE
b
127.85
1.86 '
POME
131.82
1.81
131.82
1.81
88.07
3.59
95.15
3.32
95.15
3.32
390.62
1.91
445.72
1.67
445.72
1.67
71.61
9.76
70.98
9.85
70.98
9.85
184.91
7.84
187.62
7.73
187.62
7.73
58.18
7.54
55.97
7.84
55.97
7.84
Table 4. Parameters of the log-Pearson type (LPT) III distribution fitted to annual maximum discharge series by MOM, MLE and POME methods
MOM
River Gaging Station
Comite River near
Olive Branch, Louisiana
Comite River near
Comite, Louisiana
Amite River near
Amite, Louisiana
St. John River at Nine
Mile Bridge, Maine
St. John River at
Dickey, Maine
Allagash River near
Allagash, Maine
Table 5.
MLE
POME
0.063
171.0
- 5.6
0.062
173.5
- 5.6
0.063
171.0
- 5.6
0.223
7.4
3.9
0.114
32.0
1.9
0.093
42.3
1.6
0.288
9.8
3.5
0.197
23.5
1.7
0.156
33.6
1.1
0.062
29.1
4.7
0.067
26.8
4.7
0.062
29.1
4.7
0.101
14.1
5.8
0.071
31.1
5.0
0.062
37.6
4.9
0.051
52.6
3.3
0.053
50.3
3.3
0.051
52.6
3.3
Parameters of the extreme value type (EV1) I distribution fitted to annual discharge series by MOM, MLE and POME methods
MOM
MLE
b
POME
b
0.0074
160.0
0.0078
159.0
0.0075
161.0
0.0077
241.0
0.0074
238.0
0.0074
238.0
0.0024
502.0
0.0024
498.0
0.0024
502.0
0.0057
598.0
0.0053
593.0
0.0054
591.0
0.0025
1 220.0
0.0023
1 200.0
0.0023
1 200.0
0.0080
367.0
0.0078
365.0
0.0078
365.0
103
OBSERVATIONS
M L E , POME
;/
f/
500
U
o
MOM
600
//
,.
/ ,,/
E 4OO
v
w
a: 300
u
~ 2ooi
100
REDUCED
;
VARIATE
2 0'
9 9'
O0$1rR1TIONS
MLE, POIml[
600
g
3
~/J/
/,
roOM
3o0
/ ;
iio
~'p"
~ /
s. f
"
zoo
/ *
I00
Oo
I
,o
I
20
i
30
,io
REDUCED
O,i
,. . S. ,.0
ZO ~0 4'0 SO 6 0'
PROBABILITY
,'o
:o
;o
'
'
:o
VARIATE
E N T R O P Y AS A C R I T E R I O N F O R G O O D N E S S
O F FIT
It is plausible to employ entropy to evaluate goodness of
fit, and consequently delineate the best parameter
estimates of a fitted distribution. This can be
accomplished as follows. For a given sample compute the
entropy and call it observed entropy. To this end, we may
104
9Z'S
SEF Difference
SEF of
sample
(2)
MOM
(3a)
POME
(3b)
MLE
(3c)
MOM
[(2)-(3a)]
3.592
3.162
3.166
3.166
0.430
0.426
0.426
3.664
3.343
3.349
3.349
0.321
0.315
0.315
3.397
2.946
2.977
2.977
0.451
0.420
0.420
3.412
3.144
3.143
3.143
0.268
0.269
0.269
3.611
3.325
3.326
3.326
0.286
0.285
0.285
3.781
3.544
3.540
3.540
0.237
0.241
0.241
POME
[(2)-(3b)]
MLE
[(2)-(3c)]
SEF Difference
SEF of
sample
(2)
MOM
(3a)
POME
(3b)
MLE
(3c)
MOM
[(2)-(3a)]
3.592
3.155
3.153
3.155
0.437
0.439
0.437
3.664
3,219
3.283
3.283
0.445
0.381
0.381
3.397
2.883
2.935
2.926
0.514
0.462
0.471
3.412
3.123
3.123
3.123
0.289
0.289
0.289
3.611
3.320
3.320
3.323
0.291
0.291
0.288
3.781
3.541
3.540
3.541
0.240
0.241
0.240
POME
[(2}--(3b)]
MLE
[(2)-(3c)]
SEF Difference
SEF of
sample
(2)
MOM
(3a)
POME
(3b)
MLE
(3c)
MOM
[(2)-(3a)]
3.592
3.105
3.095
3.103
0.487
0.497
0.489
3.664
3.340
3.348
3.348
0.324
0.316
0.316
3.397
2.919
2.920
2.919
0.478
0.477
0.478
3.412
3.122
3.139
3.135
0.290
0.273
0.277
3.611
3.313
3.329
3.329
0.298
0.282
0.282
3.781
3.535
3.542
3.542
0.246
0.239
0.239
POME
[(2)-(3b)]
MLE
[(2)-(3c)]
CONCLUSIONS
ACKNOWLEDGEMENTS
105
106