01
Spring 2018
Quiz 4 (Take Home)
Due: April 18th, 2018
P
n
xi n
n
f (x1 ; x2 ; :::; xn ; ) = (1 )i=1
Pn
xi n
= (1 )i=1
1
P
n
=u xi ; h (x1 ; x2 ; :::; xn )
i=1
where
P
n
xi n
P
n
u xi ; = (1 )i=1
i=1 1
and
h (x1 ; x2 ; :::; xn ) = 1
n
X
T (X1 ; X2 ; :::; Xn ) = Xi is su¢ cient by the Factorization Theorem. And
i=1
we can write the probability mass function like
= exp x ln (1 ) + ln
1
n
X
From this, we can see that T (X1 ; X2 ; :::; Xn ) = Xi is complete.
i=1
1
0 1
B C
B 1 C
EB
BXn
C
C
@ A
Xi 1
i=1
1
X 1 k 1 n k n
= (1 )
k 1 n 1
k=n
1
X (k 1) 1 n 1 (k 1) (n 1)
= (1 )
n 1 (n 1) 1
k=n
X1
l 1 n 1 l (n 1)
= (1 )
n 1 (n 1) 1
l=n 1
=
n 1
n
X
Since Xi is complete and su¢ cient, the MVUE is
i=1
b= n 1
n
X
Xi 1
i=1
Can you calculate the variance of the MVUE, and can you …nd the Cramer
Rao Lower Bound an unbiased estimator of ? All moments exist (as long as
n 2) due to
1 1
Xn
n 1
Xi 1
i=1
with probability one. So the variance should exist even for n = 2 unlike the
exponential distribution where we had to start at n = 3.
2. Let X1 ; X2 ; ::::; X2n 1 be iid N ( ; 1), where 2 R (n 1). Note this is
an odd number of random variables.
2n
!
X1
b. Determine E X(n) j Xi .
i=1
2
The joint density is
2n 1
!
2n 1 1X 2
f (x1 ; x2 ; ::; x2n 1; ) = (2 ) 2
exp (xi )
2 i=1
2n 1 2n
!
2n 1 1X 2 X1 2n 1 2
= (2 ) 2
exp x + xi
2 i=1 i i=1
2
2n
!
X1
=u xi ; h (x1 ; ::; xn )
i=1
where ! !
2n
X1 2n
X1 2n 1 2
u xi ; = exp xi
i=1 i=1
2
and !
2n 1
2n 1 1X 2
h (x1 ; ::; xn ) = (2 ) 2
exp x
2 i=1 i
2n
X1
That shows Xi is su¢ cient for . We can write the density of X1 as
i=1
1 x2 2
f (x; ) = (2 ) 2
exp + x
2 2
2 2
x 1
= exp + x ln (2 )
2 2 2
This shows that the density is in the exponential class, which implies that
2n
X1
Xi is also complete. Further
i=1
E X =
and by part (a),
2n
!!
X1
E E X(n) j Xi = E X(n) =
i=1
2n
!
X1 2n
X1
Both X and E X(n) j Xi are functions of Xi , and both are unbi-
i=1 i=1
2n
X1
ased for . By completeness of Xi ,
i=1
2n
! !
X1
P E X(n) j Xi =X =1
i=1
3
for all . In other words,
2n
!
X1
E X(n) j Xi =X
i=1
2 1
X
=
n
X1 = X =
Cov X1 ; X p
= = nCov X1 ; X
p1
n
And
2
Cov X1 ; X = E X1 X
2
1+ n 1 2 2
= +
n n
1
=
n
So
1
=p
n
4
From a known theorem (See auxillary notes for proof.) on the bivariate
distribution, it follows that given X = t, X1 has a normal distribution with
mean
X1
X1 + (t X)
X
= +t
=t
and variance
2 2 n 1
X1 1 =
n
So therefore
E Ifx:x cg (X1 ) jX = t
= P X1 cjX = t
0 1
X1 t c t
= P @q q jX = tA
n 1 n 1
n n
r
n
= (c t)
n 1
By the Rao-Blackwell and Lehmann Sche¤e Theorems, the MVUE for is
r
n
b= c X
n 1
4. Let X1 ; X2 ; :::; Xn be iid N ( 1 ; 2) random variables, where 1 2 R and
2 > 0 (both unknown).
!
2
1 (x 1)
f (x; 1; 2) = (2 2)
2
exp
2 2
2 2
x 1 1 1
= exp + x ln (2 2)
2 2 2 2 2 2
From this, one can see by inspection that this is in the two parameter expo-
nential class. That implies
n
X n
X
T1 = Xi ; T2 = Xi2
i=1 i=1
5
are jointly su¢ cient for ( 1 ; 2) and complete. Now consider this transfor-
mation.
T1
V1 = X =
n
T12
T2 2 n
V2 = S =
n 1
This is a one-to-one transformation because we can solve for (T1 ; T2 ). By
a theorem we discussed in class, it follows that X; S 2 are also su¢ cient and
complete for ( 1 ; 2 ). This pair is more useful for part (b) than the …rst suggested
pair.
p
b. Using (a), …nd the MVUE for 1 2.
(n 1) S 2 d 2 d n 1
= n 1 = Gamma ;2
2 2
We need to …nd E (S). (And we’ve done this exercise before in Chapter 8.)
(n 1) S 2
t=
2
p
2 p
S=p t
n 1
r Z1 n2 1 t
2 t exp 2
E (S) = n 1 dt
n 1 2 2
n 1
0 2
r n n Z1 n2 1 t
2 22 2 t exp 2
= n n dt
n 1 2 n2 1 n 1 22 2
2 0
r
2 n p
2
= n 1 2
n 1 2
So by independence,
r !
n 1 n 1 p
2
E X n S = 1 2
2 2
Since r n 1
n 1 2
X n S
2 2
6
p
is a function of X; S 2 , and it’s unbiased, it must be the MVUE for 1 2.
n
!
n 1X n
f (x1 ; x2 ; :::; xn ; ; ) = exp xi + Ifx:x g x(1)
i=1
n
!
X
= u x(1) ; xi ; ; h (x1 ; x2 ; :::; xn )
i=1
where
n
! n
!
X 1X n
n
u x(1) ; xi ; ; = exp xi + Ifx:x g x(1)
i=1 i=1
and
h (x1 ; x2 ; :::; xn ) = 1
n
!
X
So X(1) ; Xi are jointly su¢ cient for ( ; ). To show completeness,
i=1
consider the one-to-one transformation
V1 = nX(1)
and
n
X n
X1
V2 = Xi nX(1) = (n j) X(j+1) X(j)
i=1 j=1
n
!
X
If we can show (V1 ; V2 ) is complete, then it follows that X(1) ; Xi is
i=1
also complete. Now by the independent spacings theorem,
d
(n j) X(j+1) X(j) = exp ( )
7
d
n X(1) = exp ( )
So
P nX(1) v1 = P n X(1) v1 n
v1 n
=1 exp
and
1 v1 n
gV1 (v1 ) = exp
1 v1 n v2n 2
exp v2
1 v1 n v2n 2
exp v2
= n
exp
(n 1)
E (h (V1 ; V2 )) = 0
Written in terms of a double integral, this is
Z1Z1 v2n 2
exp v2
1 v1 n
h (v1 ; v2 ) n exp dv2 dv1 = 0
(n 1)
n 0
or
Z1Z1
v1 v2
h (v1 ; v2 ) exp v2n 2
exp dv2 dv1 = 0
n 0
8
or
Z1 v2n 2 exp v2
h (n ; v2 ) n 1 dv2 = 0
(n 1)
0
h (n ; v2 ) = 0
for all v2 > 0. Since this argument worked for any 2 R, it follows that
h (v1 ; v2 ) = 0
for v1 2 R and v2 > 0. Putting this all together, we can conclude that
P (h (V1 ; V2 ) = 0) = 1
for all 2 R and > 0, and
! thus we have that (V1 ; V2 ) are complete. In
n
X
turn, this means X(1) ; Xi are complete, since it’s a function of (V1 ; V2 )
i=1
(one-to-one correspondence).
E X(1) =
n
E X(1) = +
n
E X = +
Multiply the top equation by negative one and add to obtain
n 1
E X E X(1) =
n
9
So the MVUE for is
b= n n
X X(1)
n 1 n 1
Now multiply the top equation by n to obtain
E X nE X(1) = (n 1)
b= n 1
X(1) X
n 1 n 1
10