Professional Documents
Culture Documents
21 Application of Gfunction in Statistics
21 Application of Gfunction in Statistics
During the past few years Meijer's G-function is extensively used in the
deals with the applications of G-function in the various fields of Statistics. Only
the main results are discussed in the text. Some more results are given in the
exercises at the end of this chapter and more can be found from the references. All
the articles in the list of references do not make use of G-functions directly but
the problems under consideration, in these articles, are such that atleast particu-
lar cases of G-function could be made use of. Some other applications in Physical
Sciences, Engineering and related fields will be discussed in the next chapter.
needs the exact distribution of the test criterion in a form suitable for computing
the exact percentage points. Several statistical tests associated with a multi-
variate normal distribution, along with the exact distributions in particular cases,
of the likelihood ratio test criteria, are available in Anderson [16]. A method of
But none of them proved to be powerful enough to give the exact distributions of
these test criteria in general cases. The methods of Fourier and Laplace transforms
are useful in some of these problems. The method of Mellin transform is success-
fully used by Nair [215] to work out some exact distributions in particular cases.
butions in particular cases by using inverse Mellin transform technique and represen-
of the logarithmic type discussed in Section 5.4. The users of these results are
-190-
Ai-Ani and Jouris [231], Pillai and Jouris [234] and Pillai and Nagarsenker [235]
under consideration in these papers are not solved because their problems are the
cases where the poles of the integrand are not simple except in particular cases.
The exact null and non-null distributions of almost all the multivariate test
criteria, in the general multinormal cases, are obtained in computable forms for
these articles. These problems were open problems since the 1930's and with the
help of a number of techniques several authors have worked out particular cases.
The expansions of Meijer's G-function given in Chapter V yielded the exact distri-
butions of all these test criteria in computable forms and for the general cases.
A detailed account of the different methods, applied to these problems so far, are
Only the dens'ty functions are given in the following subsections.The distribution
functions are available by term by term ~ntegration with the help of Theorem 6.1.1
and hence the discussion is omitted. The density functions are given in series
forms, which are computable, but in particular cases they can be simplified in terms
All the density functions are represented in terms of series involving terms of
the type xa(-log x) b. Hence in order to obtain the distribution functions one needs
x
f u~( -log u)k-ldu = x~+I kE k(k.l). "" (k-r+l) (- log x) k'r
o r=l k(~+l) r
(6.1.1)
In this as well as in later sections we will only give the h-th moment of the
likelihood ratio criterion of the problem under consideration and the exact density
functions. The details of the tests and the methods of deriving these moments may
be seen from any book on Multivariate Statistical Analysis or from Anderson [16] .
In all these problems the (h-l)st moment is nothing but the Mellin transforms of the
density functions, Hence the density functions are available from the inverse Mellin
transforms. Since the density functions exist in all these cases, existence condi-
E(U h) : 11 (6.1.2)
(6ol.2) it is easily seen that by taking the inverse Mellin transform the density
Evidently the G-function in(6.1.3)is in the logarithmic case but this can be eva-
luated by using the technique of Chapter V after identifying the poles. In order
to identify the poles one has to consider four different cases, namely, Case I:
p-even, q-even; Case II: p-odd; q-even; Case III: p-even; q-odd; Case IV: p-odd;
q-odd. It is seen that in Cases I,II and III, E(U s'l) can be written in the
following form:
-b.
E(U s'l) = C ~ (~.j)-aj K (~- ~1 -j) 3 , (6.1.4)
jea jeb
where n P P (n2q+l- j )
c~ = s +'~ + ~ -I and C = II (6.1.6)
j=l p~)
The quantities a,b, a. and b are different for the different cases and these will
J J
be given below.
a.
J
=
{ j j o l 2
p+l
2
2
. p+l
' j = 2
.....
'
p+l + I ..... ~2
2
J - 2 + i, i = 1,2,..., eel
2
(6.1.9)
193-
p-3
:12 . . . . . (6.1.10)
2
l
a. = j :12 ..... 1
]
P ~ - P P + I q+l
2' J - 2 ' 2 ''''' 2 ' (6.1.12)
( 2~ _ l, ] _s~!
- 2 + l, i -
- 1 ,2 .... '
P
2
I ;
b . lj j = 1 2, P- i
] ' ' "'" 2
I ~2 - i,j = q-i + i,
2
i = 1,2, P
.... 2
i ; (6.1.13)
!
Ip-i , p-l. , q+l
I 2 ' J = -~--~ z ..... 2 '
p-3 •
( ~ i, j = ~ 2 1 + i, i = 1,2 ..... (6.1.15)
2 '
- 194-
Pn! j:a!!
2 ' 2 '
For the cases I,II and III the poles are available by equating to zero the various
factors of
a. b
(~ - j ) ] ~ (~ - ~
1 -j) j (6.1.18)
jea jeb
oo a b.
11 (C~- ~i +v) II (C~- j) j 11
i
(C~- "~ -j)
J
(6.1.19)
v=o jea j6b
where the exponents denote the orders and the quantities a,b,aj and bj are avail-
able from (6.1.7) to (6.1.17). Now by using the results in Chapter V we can w r i t e
T h e o r e m 6.1.2. For cases I, II and III, that is, when p and q are not both odd, the
n+q. 3
2 ~- J b-I
u j b .-i b..-l-v
Z % ( J ) (-log u) ] B W}
jeb v=o v v
(bj- 1)'
0 < u < i, where (6.1.20)
-195-
-a -b
V = (j-t) t ~ ( j - ~ -It ) t
(6.1.21)
tea teb
t~j
-a -h
W = (j + ~ I -t) t ~ (j-t) t , (6.1.22)
tea 'teb
t~j
a
A (r) = ( -i ) r+l r: { Z [ t ]
o tea (j_t) r+l
t~j
b
+ Z [ t ] }, for r > 0 and (6.1.25)
teD (j_ i _t)r+l
(r) I) r+l r: a b
B° = (- [ Z [ t ] + Z [...... t ] I,
t6a i j_t)r+l t6b t)r+l
(~ + t~j (J"
(6.1.26)
Theorem 6.1.3. When p and q are both odd the density function of U is given by,
n ~2 3
(-l) v j + - i + v
-8
f(u) = C { Z [ ] 11 ( i~ - v - t ) t
v=° v~ p(l- v) tea
-196-
n
-b j + -l-j a.-i
j a.-i
x II (-v-t) t] + Z Z ( j )
teb jea (a j-l)' v=o v
3
a.-l-v u2 + - -2 -j b j-i b.-i
~(-log u) j A' V' + Z Z ( j )
v
t~b (bj - i)' v=o v
b .-l-v
(-log u) 1 B' W'] , 0 < u < i, (6.1.27)
v
where A' and B' have the same expressions in (6.1.23) and (6.1.24) with A
v v o
and B replaced by A' and B' respectively and,
o o o
i
r(j- ~) r(j)
v' - v, w' w , (6.1.28)
r(j) r(~ +j)
i
A' = ~(j- ~) - ~(j) + A (6.1.29)
o o
I
B' = ~(j) - ~ ( ~ + j) + B
o o ' (6.1.30)
where ~(') and ~(.,.) are defined in Chapter V. The cummulative distribution
can be easily worked out by using Theorem 6.1.1 and hence the discussion is omitted.
Exact percentage points are computed by using Theorems 6.1.1, 6.1.2 and 6.1.3.
These are available in Mathai [180] . Also the exact percentage points for a
number of likelihood ratio test criteria are computed for the first time by using
-197-
the expansions given in Chapter V. Only a few of them will be given here. A
number of others are available from the references cited at the end of this chapter.
P
II 1~(~ + h)
where q Pi
~ ~ n+l-
_
~ . i
)
=I j =i
cI
P ~n+l-i
j=po+l
In order to use the results of Chapter V and to put (6.1.34) in computable forms
one has to identify the poles of the integrand in (6.1.34). In this connection
one has to consider three different cases. Case I: Po ~ Pl ~ "'" ~ Pr all even
and Pr+l odd such that Pr+l in magnitude is in between Pt and Pt+l for some t j r
and Po ..... Pr' Pr+l exhaust all the po,Pl ..... pq;, Case II: Po ~ Pl ~ "'" ~ Pr
all even and Pr+l ~ Pr+2 ~ "'" ~ Pq all odd with q-r = 2m, m = 0, i .... ,;
Case III: Po ~ Pl ~ "'" ~ Pr all even and Pr+l ~ "'" ~ Pq all odd with q-r=2m+l,
The simplification of the gammas in (6.1.33) is a lengthy process and hence further
discussion is omitted. The details of simplification and the exact density fl(v),
Remark: In the problems discussed in Sections 6.1.1 and 6.1.2 it may be observed
that the garm~as cancel out leaving linear factors in the denominator of the moment
expressions in (6.1.2) and (6.1.33) for a number of cases. Hence the distributions
can also be worked out by using a generalized partial fraction technique developed
This test is described in Anderson ([16], p.262) and the h-th moment of a
E(wh ) FP(2 ) p P ( ~ + h)
= 11 , 0 < ~I < i , (6.1.35)
rp(2+h) j=1
where n and p are positive integers and PP(') = {P(')}P • Therefore the density
0<~I< i, (6.1.36)
where Fp(2 )
C2
IIP r (n+_.~2-" )
j=l
-199-
Case I, p-even: When p is even the poles of the integrand in (6.1.36) are avail-
able from,
2 b.
K (cz - ~2 + j) j" I[ (C~ - ~I P2 + j) J (6.1.37)
j=l j=l
by equating to zero the various factors in (6.1.37). Hence when p is even, f2(~l)
__ n 1 P'+"J
n p+j 7-7"2
f2(c~I) = C 2 ~ I Zj [0~12 2 Ma" (,~I) + ~°I Mb. (,~i)} ,
J J
aj for (6.1.38) is a.j = j,j = 1 ' .... ~ - I and b j for (6.1.38) is given in
(6.1.37). The notation in (6.1.39) will be retained throughout the remaining sub-
A(t) = ~t+l
a. log B at y = 0 . (6.1.40)
a.
J ~ t+l ]
#Y
Similarly A "t'(
~ is available from B b Hence in the following sections only
b.
] ]
l ~
aj,b.,Ba.] and Bb. will be listed. The method of obtaining A it) from B
a. a.
] ] ] J
for (6.1.38) will be illustrated here. In this case
Now take the logarithmic derivative of (6.1.42) and evaluate at y = 0 then one gets,
P-I
A (r) • r _> I = ( - 1 ) r + t r ' ( 2 Z ~ ( r + l , ~ - j - ~i - k )
k=o
i" 1
P ~(r+l, ~2 "j) + [ _ _ _ _ L + 2 +...+ (-I) r+l
2 (_ j+l)r+l (_j+2)r+l
P-I
+ (~+1) + (j+2) +...+ 2 ]}; (6.1.43)
I r+l 2r+l (~ -J-z)"
..r+l
-~k(.), where ~ and zeta functions are defined in (5.3.2) and (5.3.12) respectively.
F(~-j) I~-j-l)...r(2)
Bb • = . . . . i P -i
J (.i)J-I(__2)0-2... (__j+l) ip ~ (I + ~2 -j)j [(3 __j)(5 __j)2...(~2 " ~ .j)2 ]
R R R ~-I R
= ~(_1)2 (_~)~ ...(~_j)2 (~_j_1)2 ...(_~+1~1E~2 (~ +j j)
Case II, p-odd: When p is odd the density function is of the form
- = -n+1
-~+ j - ! _ n-l
~-=+ j
f2(~l) : C2 ~ i Z. [~i 2 2 Ma.(el) + ~I 2 2 2 ~.(~i) ]
J ] ]
(6.1.46)
where 0 < o I < i, aj and bj have the same forms as in (6.1.38) with p replaced
by p + l a n d p - 1 r e s p e c t i v e l y . Ba. a n d Bb. a r e g i v e n b e l o w .
J ~
I 3
P(~2 - ~ -j) r ~ - ~ -j)...r(2)
Bb . = e-_!
J p-I i 3 2 .(-j D-2)~ 2
I~ 2 (~2-J)[('l)J-l(-2)J-2"''(-J+l)][('J+ ~)(-J+'2 ) ""
p-I
3 2 ~22) 2 ]-I
• .. (-J+D][(-j+ -I
2 )(-j+ 7) ...(-j+ ]
Again A(r) and A(r) are available by the procedure described in (6.1.43)
a. b,
] ]
In the multinormal case the h-th moment of a criterimn W 2 for testing the
hypothesis that the diagonal elements are equal given that the covariance matrix
n
E(W h) = php pp (~ + h) P(n~2 ) , (6.1.49)
p,O ~2
y,P)
f3(~2) = C~ ~ I H [ -- [
1,0 pP n
(~,I), (~,I),. • ., (~,1), (~,I) ..... (2,1)
-203-
n + I, n 2 n p-I
p-l,0 2 p ~+~ ..... 2 + 2
= C~ ~2 "I g [~21
p-l,p-I
n n n
2 ' 2 .... ' 2
Qo
n + v
C3 ~2 Z ~2 Ma.('~2 ), 0 < ~2 < i . (6.1.50)
v=o J
where Ma (~2) has the same form as in (6.1 • 39) with aj. = p-i
3
r (n-~2) r(~)
' , C3 = _
(6.1.51)
C3 pp(2 ) p(np-l)/2 (2~)@ pP(2 )
p-I
A (r) , r > I, = (-i) r+l r' {(p-I) ~(r+l,l) - Z ~(r+l,-v+-i)
a -- j=l P
J
+ (p-I) [ i +...+ I ]} .
(.l)r +I (_v) r+l (6.1.53)
The hypothesis that the covariance matrix ~ is of the form 021 where a 2 is
D
test. This test is a combination of the tests given in 6.1,3 and 6.1.4. The h-th
php p ( ~ ) P P(n+-'~21-' + h)
E(W h) 71 (6.1.54)
P(~ + ph)j=l p(~)
(~ ,P)
l,p 7 n-2 , 1 ) ,
(--~- ,1) . . . . ]
n-2 j
p-l,0 ~--~- + p ' j = 1,2 ..... p-I
C4G
p-l,p-I
[0~
In-2 _ j j = 1,2 .,p-i
] ,
2 2 . . . .
=o . n P -l+j
. . . . ~ n p 3 + j
--
2 2 2 2 2
C4 [ % ~ M a .(~) + Z co Mb (~) ] ,
j =i 3 j =I J
(6.1.55)
C~ 4 = • C4 =
PP p r(n+__~.L~ l-p np-I P p.n+l-j
" 2 ) (2~) 2 p 2 71 [ 2 )
j =I j =i
(6.1.56)
- 205-
2 2
1
II P(r) !I P(-j--~ +r)
=
r=2 r=l p-3.
, j = 1,2 .....
S ,
j-i . p-i r 2 '
J
11 (-r) j'r II P(-j+ ~2 + p)
r=l r=l
p-1
2 i p-I P i
11 P(-j- -~ + r) [ II P ( - j + ~ + r ) ] -
r =i r =i P
p-i D - I -1
[(-1)(-2) ..(-j+ P-~ )1 2 [(-j+~-~-' -l) 2 ...(-j+l) ]
• 2
for j = p-._._~l p + l
2 ' 2 .... ;
p-1 p-i
2 j 2
I
r(r) 1I P ( - j + ~ + r )
r =2 r =i p-3
BB. = • j = 1,2,..., 2 ;
J j-i j-r p-i i r)
II (-r) II P(-j+ ~ + ~2 +
r =i r =i P
2 1 p-i i -i
P(-j+~+r) [ ~r(-j+7+e2+p)]
r=l r=l
, (6.1.57)
j= P-~ e!!
2 ' 2 ....
- 206-
(j,j = 1,2, P 2
"''' 2
a . =
J
l
j,j = 1,2 ..... 22 ,
b.
]
j o 2 2 + 1,22 + 2 . . . .
p-2
P -l-j 2
211 i
F<r) I[ F(-j- -~ + r)
r =2 r=l
B
8.
= j = 1,2,. i -2 ;
°''2
3 j-i P -2
[ 71 (-r) J-r] 2 ~'(-j+ 22 + p )
r =I
(_ j+ .p.~21) p-1
r =I
r~p/2
p-2
2 1
II F(-j- -~ +r)
r =I
J- ~2 +I P -i P -2 p-i
II (-r)] 2 (- j+ p -2"
L 2) II (- j+r) r iI r(-j+ 22 + _r)
r =I r =I r=l P
r~ p/2
for j_>~2- i.
£-l-j
2111 F(r) 22 H-i r ( - j + ~ +r)
r=2 r=l
m b .=
j = 1,2,.., ~ - 2 ;
• " 2
J
[j-1
71 (-r) j-r]
.
(-J+22' p-i i r)
~I P(-j+ i + 22 +
r =i r =I P
r~ p/2
-207-
P- 1
2
1
II F(-j+ ~ +r)
r=l
P-2
j- ~ + I i -i
2 )p~l_ i r
(-r)] 2
l[ (-j+r)r) (-j+ ~2 r=l F(-j+i +~2 + p)
r =i r=l
r~ p/2
(6.1.58)
for j ~ ~ -I, with the restriction that when j = p/2 the factor (-j+ ~) is to
other tests are given in Mathai [172]. Some of these fall under the category of
H-functions but these are reducable to G-functions and hence the techniques of
Chapter V are applicable. There are some other test criteria for which the distri-
butions do not fall in the category of G-functions and the exact distributions for
these are not yet a v a i l a b l e for the general cases. One such test is the one for
testing that the mean vector equals a given vector and the covariance matrix equals
available for the distribution of the likelihood ratio criterion for testing this
distributions are the distributions of statistical test criteria when the null
in computable forms are needed for comparison of tests and for studying various
ratio criteria for testing hypotheses on multinormal populations, were not avail-
able for most of the problems till the 1960's. A breakthrough is achieved in this
-208-
direction with the help of the theory of Zonal Polynomials and Hypergeometric Func-
tions of matrix arguments developed by several authors of which the main articles
are James ([122], [123], [124], [125]), Herz [117] and Constantine [67]. Mathai
and Saxena [199] defined and developed the theory of the G-function with a matrix
argument.
of G-functions by several authors which include Khatri and Srivastava [136], Pillai,
AI-Ani and Jouris [231], Pillai and Jouris [234] and Pillai and Nagarsenker [235].
In these articles, as remarked earlier, the problems under consideration are not
solved because the problems under consideration are the log~rithmic cases and hence
of multivariate criteria are obtained for the general cases and for the first time
which defines a representation of the real linear group GL(m) in the eector space
where
and the summation is over all partitions of k into not more than m parts. Each
Vk, K contains a unique one dimensional subspace invariant under the orthogonal
group 0(m). These subspaces are generated by the Zonal Polynomials, ZK(A) w h i c h
symbols.
argument Z is defined as
(al)K-.-(ap)K CK(Z)
F .. " bl,...,bq; Z) = Z (6.2.6)
P q (a I, .,ap, k=o K
(bl)K..,(bq)K k~
with p < q+l or p = q+l and IIZII < i, w h e r e IIZII is a suitable "norm", for
k = k I + k 2 + . . . + k m,
The parameters a i and bj are all complex numbers and none of the b.j is an integer
By using the above concepts the non-null distributions can be worked out in terms
direction see Mathai [172]. For the purpose of illustration we will discuss one
problem here.
butions Np(O,El) and Np(O,%2) respectively, where Np(B,E) stands for a p-variate
multinormal density with mean vector ~ and covariance matrix E. Then S I = XX'
and S 2 = YY' are independent and have the Wishart distributions Wp(ni,%i),
i = 1,2,. Let 0 < fl < "'" < f < ~ be the eigenvalues of,
P
I% I - % %21 = 0, (6.2.9)
degrees of freedom and covarisnce matrix %. Then a criterion W for testing the
hypothesis
~) /~ = I , ~ > O, known
P
is given by
P
W = ]I (l-ei) = lip- Ell (6.2.10)
i=l
where Ip is an identity matrix of order p, A = diag (~l,...,~p),
E 1 = diag (e I , .... ep), e i = ~fi/(l+Sfi) , i = 1,2 ..... p and I(" )[stands for
the determinant of (-). The h-th moment of W is available in Pillai, Ai-Ani and
-211-
Jouris ([231]) aS~
nI
n2 -~- n nl n -i
Pp(~) Pp(~- + h) 15 } 2FI(~, ~-- ; ~ + h, Ip - ( S A ) )
E(W h) =
n2
Pp (~---) P (@ + h) (6.2.11)
P A
p(p-l)/4 P
rp(U) = ~r lI P(u- ) . (6.2.12)
i=l
By taking the inverse Mellin transform of (6.2.11) the density function can be
written as,
nl i-I
n2 -i p,O ~ - + k. - -- i = I,. .,p
-- i 2 ' "
g(00) = A K co 2 G [ col ]
P'P _(i-I) , i = 1,2,...,p
2
(6.2.13)
0 < ~I < I and where A K stands for the expression
-nl n nl )-I
r e ( 2 ) I 5A[ T ~ ('2)K (2-)K CK(Ip-(5'%
AK = Z % (6.2.14)
n2 k=o K k '
Pp(~- )
The density function in (6.2.13) can be put into computable form by expressing the
and by using the notations in (6.1.39) to (6.1.43) we can write gl(~) in then2folio{-
wing form where for convenience g(~) in (6.2.13) is written as g(~) = A K m - ~ gl(~)
and in gl(~) we will write n I as q for convenience. For Cases I,II and III to be
discussed below,
_ ~ +j . ~ ! +j
2 2 2
gl(m) = Z ~ M (~) + Z m M b (m) (6.2.15)
a,
jea J jeb 3
0 < co < I, where aj,bj,Ba, and Bb. are listed below for the different values of Pl
3 J
and q = n I .
2
L..,.
U
II II
m o
m
i u.~. i i..~.
I
u,.. II
e.. II I
II u.... II II
It da "
II II
A II
IA
+
+ v
o. +
i +
+
I +
+ i +
i ~c?~- ~ 1'-o i
+
v i
v i i-o
+
II 1
q- e.,
+ ~
t~
II
I
+ +
+ +
I
+ 2
+ + ~
~ +
+ , +
v
!3
f~
,....= i'-i
i'-I
.o
rl
Ir
i II
u..h
e~
"7,
,::," I',~ I~ ~t-a ,....,.
d~
i i I_.~,
I L.,~, o
e~ {._~.
iI i'~" II
p.a 0
II
it,.,> ill L...~. r..~ {"El -
II IA I1 II i~
"" ~ "
+ ~i~ bo ,.a
+ v
+ ~
+ ~
!
b~ ro
II v + + ~
+
~ + ~ + ,,.e>
I
• I
+
i
i +
2
+ +
H. +
-214-
bj
i
|p+l
j, j = 1,2 ..... P21 ,
. p+l q+l _
I 2 ' J = , .... 2 T Kp_ I ,
i = 1,2 . . . . P-I - I
" 2
-a -b
B
a.
= K (-j+t) t II (-j- ~1 + t) t
; (6.2.20)
3 tea teb
t~j
-a -b
BD. = ~I ( ~i - j + t) t ii (-j+t) t (6.2.21)
J tea teb
t~j
i
F(~2 + 2 -J +~2 + kl)} respectively and
q+l + k2+v
(-I) v r(1) ~ 2 -a
I t
Rv v~ P( + kl-k 2- v ) tea
-b t
x ~ (- ~ - P - kf- ~ + t)
2 2 (6.2.23)
teb
-215-
(6.2.23).
Remark I: There are several test criteria associated with the test criterion dis-
cussed in this section. Two of them are the likelihood ratio test criteria asso-
ciated with Wilks criterion for testing regression and for testing independence
when there are only two subvectors. The exact distributions of these are available
Remark 2: The exact non-null distributions of most of the test criteria mentioned
in Section 6.1 are still open problems. From the discussion in Section 6.2 it
should be remarked that once the non-null moments of these criteria are available
Remark 3: In a recent series of papers Davis has employed the method of differen-
tial equations in deriving the exact distributions and he obtained some particular
cases. For a paper of this category see Davis [77]. It should be remarked that
whenever the density functions are representable in terms of G-functions the dif-
ferential equations can be written down directly from (1.5.1) which may be noticed
from the different problems considered in Sections 6.1 and 6.2. The method of dif-
was employed for the first time by Nair [215] and from this paper it can be noticed
that the method is not a powerful one. The most powerful technique, available so
far, for tackling these types of problems, is the one given in Chapter V and
and Sons, New York contains an up to date literature in this field. Some applica-
tions of characterization theorems may be seen from Csorgo and Seshadri [75].
-216-
tions to the various concepts in Statistics and Information Theory. Recent develop-
ment in this direction are available from the report Mathai and Rathie [192].
But a G-function is useful to show the converse that a particular property is not s
unique property of a certain distribution and thus a G-function helps to get counter
h(x) = ~
P(C~) xC~-I e -~x , ~ > 0, ,~ > 0, 0 < x < oo (6.3.1)
oo
it -CZ
= (1- T ) (6.3.2)
If XI,...,X n is a simple random sample from this gamma population, that is, a set
= (X 1 +...+ X n)/n
is given by
@_(t) [@x (~)]n it -nCz
x = = (i- ~ ) (6.3.3)
follows from the properties of Fourier transforms, by comparing (6.3.2) and (6.3.3)
it is seen that the sample mean X again has a gamma distribution with parameters
and ~ scaled by the sample size n. That is, the new parameters are nC~ and n6 . It
is natural to ask whether this property is a unique property of the gamma distribu-
tion or not. By using some properties of Special Functions it can be shown that this
for y > O, v > O, p > 0 and hl(Y,V,p,a ) = 0 elsewhere, then the sample mean has
the probability law hi(y,nv,np,na), where n is the sample size and Iv(. ) is the
modified Bessel function of the first kind which is defined in Chapter II.
Therefore,
Remark: By using a G-function one can construct more general classes of distribu-
tions enjoying the same property. Similar properties are investigated by Mathai
and Saxena [195]. Due to the presence of a number of parameters, the various
functions appearing in this and in the remaining sections are all assumed to be non-
It is a well known fact that the ratio of two independent standard normal
natural to ask the converse question that if the ratio of two independent variates
standard normal? Several authors have given counter examples, see Laha [145].
There are also two other ratio distributions frequently used in Statistical
Inference. They are the Student-t and F-statistics. Mathai and Ssxena [197]
developed a general technique for giving counter examples for such problems. These
counter examples belong to very general classes of distributions and wider classes
can be constructed by using G-functions. We will state one such theorem here.
functions
D(a.+l+v)x v a.+v+l a.+v+2
Mai,v(x ) = l t -x 2)
Z-~(v+l)F(ai) 2FI ( 12 ' 2 ~ v+l; , (6.3.8)
for x > 0, i = 1,2, then XI/X 2 has the F-distribution for all values of v such
BI-I
h2(x ) = P(~lq~2 ) x
, x > O, ~i > O, ~2 > O.
r(~l)r(~2) (l+x)Bl+~2
(6.3.9)
This is a slightly modified form of the density. It is easy to see that (6.3.8)
is a density, that is, M (x) > 0 for all x and f M (x)dx = I for
ai,v o ai~v
i = 1,2,. This follows from Erd~lyi, A.et.al ([86], I.P. 336 (3)). Now the
ratio XI/X 2 should satisfy the L.H.S. of (6.3.10). From Erd~lyi, A. et.al([86],
co t J (ut) a.-i
v l -t
f t e dt = M (u) , (6.3.11)
a.
o r ( a .1) I,V
a.+l
t Jv(ut) ai-i e_t/ad t
f t = M (au)a I (6.3.12)
o F(ai) ai,v
a2-1
P(al+a2) a , for all v. (6.3.13)
r(al)F(a2) (l+a) ~Ita2
Note: For convenience the Parseval property of the Hankel transform will be
co
f t Jv(pt) Pl(t)dt = @l(p )
o
a nd
then
oo
f t @l(t) P2(t)dt = / p '#l(p) ~$~2(P)dp,
o o
Remark: From the nature of Theorem 6.3.2 it can be seen that wider classes can
above and further references to the literature in this topic are given in Mathai
authors have investigated the posterior distributions of random variables when one
or more parameters have prior distributions. Several such practical problems can
Theory, for example see Karlin ([133], p.21). Instead of dealing with particular
conditional and unconditional distributions one can use generalized Special Func-
Mathai and Saxena [134]. One such theorem will be given here.
-221-
density function
and f(xlp) = 0 elsewhere, where ~ and ~ are additional parameters. Let the
I[ II P(l+aj)
g(p) = j=k+l 'l~(-bJ) j=v+l k,v a I ..... ar
~G [kp I
k P(l+bj)
If V P(-aj)
II w'$ bl,. .,bs
j =l j =i
(6.4.2)
for p > O. The parameters in the G-function are assumed to be such that the G-
function under consideration exists and is real and non-negative. The uncondi-
r+s
where k+v > ('-7--~), R(bj + I + P~-) > 0, j = i ..... k. (6.4.3)
Now (6.4.3) follows from (6.4.4) and Erd$1yi, A. et. al ([86], II, p.419 (5)).
-222-
Some useful particular cases of Theorem 6.4.1 are given below for different
(a) Exponential
1,0 - ~ -i
X e"p~ = X G [Xp[0] k ~x ~-I ( ~ + ~ ) ~
0, i
(b) GaTm~a
6~-I X~ e-p~ I
P z~ ~ I~(~ +~) x~-l(~+~) -a(l+ ~)
r(a) P(~)
(c) Beta
r(~) r(B)
_~)~-i
Wells, Anderson and Cell [354] considered the distribution of the product of
two independent stochastic variables one of which is a central Raleigh variable and
problem in Physical Sciences. There are several papers on the ratio, product and
linear combinations of stochastic variables and all these problems come from
Physical, Natural and Biological Sciences and from Economics. The literature in
-223-
this direction is vastand a short list of references is available from Miller [213].
stochastic variables having some specified distributions. Mathai and Saxena [196]
where the density of each one is defined in terms of a product of two H-functions.
After analysing the structural set up of densities the authors have noticed that
almost all the ratio and product distributions which are used in statistical litera-
ture will come as special cases from the results obtained in Mathai and Saxena [196].
A large number of other product and ratio distributions are also available from the
results in this article. The technique can also be extended to the product and
stochastic variables there is a large number of papers. Mathai and Saxena [198]
one has a density defined in terms of H-functions with different parameters and ob-
tained the exact distribution. The results of this article cover almost all distri-
ture and many more in this category. Since the H-function is the most generalized
of results are available from the above mentioned articles. In Section 6.5, we are
not giving any specific result in order to save space. These results are readily
and the techniques of Special Functions are useful and powerful in tackling problems
come across some statistical techniques of deriving new results in the theory of
Special Functions. This may he seen from Mathai and Saxena ([200],[201]) and
articles of the author'S in this line are given in the list of references.
-224-
From the discussions in this chapter it is evident that G-functions are appli-
few years G-functions will be applied to many other topics in Statistics as well.
may be seen from the recent articles by Pillai and Jouris [234], Pillai and
Nagarsenker [235] and Khatri and Srivastava [136]. Two more articles in these lines
EXERCISES
see Section 5.4 of Chapter V. These appear in some of the following problems.
6.1. Show that f(u) in (6.1.3) reduces to the following forms for particular
cases.
_m -2 3 ! 1 1 1 _t
i) p = 3,q = 3; C u2 [(l-u) 2 -3u 2 sin-l(l-u) 2 + 3u log[u -2 + u 2 (l_u)2 ]]
1
C = P(n+2) r ( 7n+l
)/[ P(n-l) F(~ -I)3~ 2 ];
n i
-- -2
ii) p = 3,q = 4: C u2 [l-u 2 + 8u2(l-u) - 6u log u ],
n_2 5 i 3 i I
iii) p = 3,q = 5: C[u 2 (l-u) 2 - 45u(l-u) 2 + (30 u 2 - ~--
15 u2)sin-l(l-u) 2
1 I i
+ (30u- ~15
- u2)log [ u-2 + u-2 (l-u) 2] ,
I
= P(n+4) I~(~)/[90 2 r(n-l) r(2 -i)] ;
-225-
n_2 i 3 5
(iv) p = 3,q = 6: C u 2 [1-16 u 2 - 65 u + 16 J - 65 u 2 - 16 u 2 + u 3
n_ 2 7 3
(v) p = 3,q = 7: C u 2 .1855
[(l-u) 2 -{----~)u(l-u)
~ - I05. (1-20
(-~--) ~ + 8u 2)
i i I i
u-~ sin-l(l-u) -j + ( ~ ) ( 8 - 2 0 u + u2)u log({l+(l-u)2]u -2 ,
.n+5.
c = r(n+6) r~-T;l[2 r(7) P(~) r(n-l) r(~ -1)l ;
n 3
(vi) p = 3,q = 8: e u~ -2 [l_(@)u~(l - u 3 )-(----~)
1428 u(l-u 2) + 896 u2(l-u)
4
-u - 84u(l-5u + u2)log u] ,
[I n
C = P (n+7) ~ + 3)/[2 P(8) r(5) p(n-l) (~ -i)];
n-5 i 3 5
(vii) p = 4,q = 4: C u 2 [1-15 u 2 -80u + 80u 2 + 15u 2 - u 2
3
-30(u + u 2) log u] ,
3
- 3 0 ( 3 u + 8u 2 + 3u 2) log u] ,
5 7 3 5
+ 1099u 2 + 35u3-u 2 - 210(u+5u 2 + 5u 2 + u2)log u],
3 3 5
-8624 u 2 (l-u)-420(u + 8u 2 + 15u 2 + 8u 2 + u3)log u],
n-5 9 1 7
(xi) p = 4,q = 8: C u. .2. .[l_u
. 2 . 63(u2- - 4
u4)-(5104 - ~)(u-u )
3 5
29988 (u 2 - u)-28244(u2-u 2) - 252(3u
3 5 7
+ 35u 2 + 105u 2 + 105u 2 + 35u 3 + 3u2)log u] ,
6.2. Show that the density function fl(v) in (6.1.34) reduces to the following
forms in particular cases. [Most of these are worked out by Consul (1967,[69]).
(n-Pl-3)/2
(ii) q =I C v l-vl/2) pI-I
all Pl
Pl
P(pl ) P(~--+ I) ]
(n-Pl-5)/2 1 2 2PI'I
(iv) q =i C v (l-v / ) 2Fl(Pl-2,Pli2Plil-vl/2 )
all Pl
Po = 4 C = P(n-3) P(n-l)/[2 P(2Pl) r(n-Pl-I ) r(n-Pl-3)]
(n-P2-3) P2 -I/2 P2 P2
(v) q =2 C v (l-v) 2FI(--~ ,--~ ; P2+i/2; l-v)
all P2
C : p2 (~)/[ p(~ _ P2 n-P2-1
Po = Pl = I --~ ) P ( ~ ) P(P2+I/2)]
Po = Pl = 2
-228-
6.3. Show that for p = 3 the density function f2(~l ) in (6.1.36) can be put
n_ 2 i
2 ~ i i 3
c ~i (1-~i) 2F1(~ ' ~; ~; I-~1)
6.4. Show that for p = 3 the density in (6.1.50) reduces to the form
n
2 I
C ~22 2F i(~ ' ~; i;i-~2)
6.5. Show that the density function in (6.1.55) reduces to the following form
n 2 3
(i) p = 3: C 2 (i_~0)2 5 7 5
2FI(~, ~" ~, i-~),
q~ (3n+I)/2
C = 2n+l ~(o_~ )/[ F(n-l) F(@ -i)3 ] ;
n-5 1 7 I
(ii) p = 4: C o0 2 (1_ 2)2 2FI(I, 3 9 i_~o2 ) •
~;
6.6. Obtain the exact densities of (det S), W 2 and W 3 from the following moment
expressions.
t Pm(t+ ~)
(i) E [(det S) ] - (det 2%) t exp (tr ~) IFI(I+ 2; 2 ~ g)'
rm(~)
-229-
where Z,g,S are positive definite matrices, n,m are positive integers, (det S)
denotes the determinant of S and tr(.) denotes the trace (-) = the sum of the
P (h+ t
(ii) E(W2 h) = P 7) PP( ) iFl(h; h + ~ ; -~ ),
t n
Pp(7) Pp(h+ 3)
O< W 2 < i;
Pp(2 ) pp(n--~2 + h) n
n n n
(iii) E(W3 h) = Jlp_p2j ~ 2FI(7 ' 7; ~ + h; p2),
pp(n-~2 ) P (2 + h)
P
6.7. Show that the following densities have the gsr~na property discussed in
Section 6.3.1 .
I i
- -v -~ - (~+~)3
(i) (~_~) 2 e -px x v e
I i[ ] ,x > O,
-V -V y-~
p((2) . (p"~) ( p"l't3
~_>8, v > 0 ;
x
v(~-~3) -v e-px x -I e-(C~-~3)~
(ii) i I -2v Iv [(C~q)x)] , x> O,v > O, ~ _ > 8 ;
[(p+cz)g + (pa~3)~]
-px 2v-2m-i ~m
(iii) e 2 e
iFl(2V,2v-2m;(~-~)x), x > O, (v-m) > O;
p(2v-2m)(p-~)2m(p-B)-2v
-230-
-px 2m+2v -I
(iv) e x 2 2
1 -C~ x
IF2(v; m+ V, m+ v+--2 "-- 4 )'
p(2m+2v )p-2m(p2 + 2 ) - v
x> O, m+ v>O .
6.8 Show that the ratio of any two independent stochastic variables having the
2
x
r('~ + l)x v e- 2-- IFI ( ~; v+l; ~x-2 ), x > 0 .
Nv(x) = F(v+l) (2~) I/2 2 v/2
6.9. Evaluate the exact density of PlXl +o..+ PnXn if X 1 .... ,X n are independent
~i "I -aix i
(i) CI x i e , x i > 0, ~i > 0, a i > 0, ~i > 0 ;
(x.
~i -I -a.x. l
i i
(ii) C2 x i e , x.l > O, 5.1 > 0, a.l > 0
c~.-1
I -a.x.
i I
(iii) C3 x i e , x i > 0, c~i > O, a i > 0 ;
- a ,x°
i i
(iv) C4 e , x. > 0, a. > 0,
i l