I

UNIT 16 POINT ESTIMATION

1'1
Structure
ib, 16.1 Introduction
Objectives
16.2 Properties of Estimators
16.3 Methods of Estimation
16.3.1 Method of Moments
16.3.2 Method of Maximum Likelihood
16.4 Summary
-16.5 Solution and Answers
16.6 Additional Exercises

16.1 INTRODUCTION
In Unit 15, you have been introduced to the problem of point estimation and also to
some basic concepts of the theory of point estimation. There we have also discussed
two desirable properties of an estimator, viz., unbiasedness and consistency. In this
unit, the problem of point estimation will be discussed in greater detail. To begin
with, we shall introduce some more concepts. Next, some methods of point
estimation are discussed. In particular, we shall concentrate on two methods of
estimation that are used widely in practice, viz., the method of moments and the
method of maximum likelihood. The first one is easy to implement in practice and
fi.1 the latter leads to estimators with "good" properties.
%

Objectives
After reading this unit, you should be able to*;

'0 list the criteria for the choice of a good estimator

derive estimators by one of the methods discussed
decide which one in a given class of estimators is best according to a given
criterion
assess the goodness or otherwise of any given estimator.

16.2 PROPERTIES OF AN ESTIMATOR

We have already discussed in Unit 15 two properties of an estimator, namely,
unbiasedness and consistency. Let us recall the definitions of unbiasedness and
consistency.

Definition 1: An estimator T (XI, X2, . . . ,Xn), which is a function of the sample

values XI, X2, . . . ,Xn is unbiased for g (8), a known function of the parameter 8, if

where E, denotes the expectation taken when 8 is the parameter and SZ is the
parameter space

Definition 2: An estimator Tn = T (XI, X2, . . . ,Xn) is said to be a consistent

estimator of 8 if
Elements of Statistical In a given problem there might exist more than one unbiased estimator for the same
Inference parameter 8 or the same parametric function g (8). How do we choose among these
unbiasea estimators? As mentioned in Unit 15, one way to choose among various
unbiased estimators for the same parameter is to cdmpare their variances. That is, if
T, (XI, X2, . . . &) and T2 (XI, X2, . . . &) are two unbiased estimators of g (8),
then T1 will be preferred over T2 if

Var, (TI) s Var, (T2) for all 8 E 52 and with strict inequality for at least one 8 E 52 .

This brings us to the concept of uniformly minimum variance unbiased

estimators (UMWE).We have the following definition.

-
Definition 3: For a fixed sample size, n, T T (XI, X2, . . .&) is called a
minimum variance unbiased estimator of g (8) if (i) E, (T) = g (8) for all 8 E SL,
i.e., T is unbiased for g (€I), and (ii) Var, (T) s Var, (T') for all 0 E SL with strict
inequality for at least one 8 E SL, where T' is any other estimator based on
XI, X2, . . .Xn satisfying (i).
How do we locate a minimum variance unbiased estimator in a given problem ?
From definition 3 alone, it may be a very difficult task, if not impossible, to find a
minimum variance unbiased estimator. The following example illustrates this fact.

Example 1: Suppose a random variable X follows a normal distribution with

mean 8 and variance unity, and let XI, X,, . . . XI, be a random sample of size 10
x,
from the population . We know that the sample mean, is unbiased for 8 and so is
XI. Now, Var, (%) - -
1/10, Var, (XI) 1.Therefore, is superior to X1 for
estimating 8 unbiasedly. However, this does not necessarily mean that is thex
minimum variance unbiased estimator of 8. To check whether Ti indeed is the
minimum variance unbiased estimator of 8, it will be necessary to compare the
- variance of with the variances of all other unbiased estimators of 8, which is
clearly an impossible task. One has therefore take recourse to other methods for
locating an unbiased estimator with the smallest variance in the class of all unbiased
estimators.
4
To formalize the concepts, we now consider a population with probability density
function (if the random variable in question is continuous) or probability mass
function (in the discrete case) f (x ;0) where the parameter 8 E 52 C R is a scalar.
The set of all x where f (x ;8) t 0 is called the support off (x; 8). We shall
assume that the support off (x ; 8) is independent of 8. For example, our
discussion will not be applicable to a uniform distribution over the interval (0, €I),
since the support (0,8) is dependent on the parameter 8.
The problem is to estimate a parameter 8 on the basis of the data XI, X2, . . . Xn,
I,'
which is a random sample of size n from f (x ;8). At this stage, it is important to
bring in the notion of a likelihood function. Let XI, X2, . . . & be a random sample
from f (x ; 8) where f (x ;0) is the probability density (or mass) function of a I

random variable X. The joint probability density or mass function of XI, X2, . . . & I:
I
for given 8, is I'
I t

where xl, x2, . . .x, are a realization of XI, X2, . . . & for the given sample. If 8 is
unknown and varies over SZ, I., (8) may be regarded as a function of the variable 8,
and is called the likelihood function of 8.
 We sha!I1 henceforth assume that X is continunu?;and hence f (x ;0) is a probability
density function. The likelihdd function based on the sample XI, X2,. . . X,, is

I,, (0) = f (x,; 0) f (x2; 0 ) . . . . f (x,; 0). ,

Suppose g ( 4 ) is an estimator of 0 such that

II -
unbiased for 0, then B (8) 0. Now,
Ee[g(X)l<m.Let

B(8) = E,[g(X_)l-8
B (0) is called the bias of the estimator g ( 5 ) in estimating 0. Clearly, if g ( X_ ) is

and assuming f (x ;0) to be differentiablew.r.t. 0.

The function ln L, (0) is called the score function based on the obserVations
d0
XI,X2, . . . Xn. Now, since f (x ;0) is a density function, we have
\

for all 0.
f
For brevity, we write the above equation as

where A is that part of R, where,,I (8) is positive.

We now assume the (1) and (2) can be differentiatedw.r.t. 0 under the integral sign.
Then.

and

where B' (0) - d

,B(Q)
 Elemmb of StatbtIcat Making use of the relation
Inference

a
d
4 (8) - ( -&1.I.,,, (8)) 4 (8). we can write (3) and (4) alternatively as
I
I

I
-1nL,(8)
I L,@dx_ - 0 (5)

d
and $g(x)z[lnL,(~)]~(~)dx_-l+B'(~) (6)
A

respectively.
..
Since L, (8) is the joint density of XI, X2, . ,X,, when 8 is the parameter, the
relations (5) and (6) may be written in terms of expectations, as

and

Combining (7) &d (8) we have .

d
g(X)delnL,,(0)
I = 1+B1(0) (8)

inequality states that for any two random variables U and V

The ~auchy-~chwarz
withE(~~)<m,E(~~)<ol, n

( 10)

with equality if and only if U and V are linearly related.

Let U - d
g(Z)-e,V = ~ l n k ( 8 )

Then from (10) we have

2
where I,, (8) = E,[$ Lo 4 (@)I. I,, (8) is called the Fbhainformation in the
sample (X,, X2, . . . ,X,,). The equality (11) is known as the Cramer-Rao
1
inequality.
 It can be shown that Point Eatlmatlon

I,, (9) = nI1 (8)

where I,, (8) is the Fisher information contained in one observation. The inequality
(11) can then be written alternatively as

where, we write I (8) in place of I1 (8) for simplicity.

Ifg(X)isunbiasedfor8,thatis,ifEe(g(~)) - 8,then
E,[~(x)-~]' -
Var,[g(X)]andB(8)
for an unbiased estimator g ( X ) of 8, we have
-
0andhenceB1(8) - O.Thus,

The lower bound l/(nI (8)) to the variance of an unbiased estimator g ( X_ ) of 8, is

-
called the Cramer Rao lower bound. Thus if the regularity conditions assumed
earlier hold, the variance of an unbiased estimator g ( 5 ) of 8 cannot be smaller
than l/(nI (8)) and hence if an unbiased estimator of a 8 has variance equal to
l/(nI (8)). It is the minimum variance unbiased estimator of 8.
If g ( X_ ) is an unbiased estimator of 6 (8), a known function of 8, the Cramer-Rao .
. inequality takes the form

Var,[g(X)I 2 [~(8)12/lnl(8)} (14)

I
We can now define an efficient estimator. ,

Definition 4: An unbiased estimator g ( 5 ) of 6 (8) is said to be efficient in the

2
Cramer-Rao sense if its variance is equal to the lower bound [ 6' (8) ] /[ nI (8) ]
where n is the sample size and I (8) is the Fisher information in a single observation.
It is also a uniformly minimum variance unbiased estimator (UMVUE) of 6 ( 8 ) in
the sense that it has the smallest variance uniformly for all 8 E GI in the class of all
unbiased estimators.

Note that it is possible that there exists a uniformly minimum variance unbiased
estimator for 6 (8) but the vakiance of this estimator does not attain the Cramer -
Rao lower bound.

The Fisher information I (8) can be shown to be equal to

-Ee[-$hf(~;€l)]~

This is sometimes computationally simpler compared to the formula

E0 [ -&III (8) 12, given earlier.

.
Example 2: Let XI, X2, . . ,X, be a random sample from a normal population with
unknown mean p and variance unity. The density function of a normal random
variable with mean p and variance unity is
Elements of Statistical
Infemnce

and thus

Hence I (p) = -E = 1and the Cramer - Rao lower bound is

I/[ nI (p) ] - l/n. Now, we mow that = n- 2 Xi is unbiased for p and

Var, ( ) - i- 1
l/n. Therefore, Var, ( ) attains the Cramer-Rao lower bound and %
is the U M W E of p. It can be shown that there is only one such UMWE, that is, 51
is the unique UMVUE of p.

Example 3: Let for n r 3, XI, X2, . . . ,Xn denote a random sample of size n from a

Poisson population with parameter A. Then, f - n-I Z Xi is inbiased for A, and

z) - Mn. Now,
i-1
Var, (

d
-lnf(x;A)
dA
- -1+x/l,

d2
and -lnf(x;A) = -xA2
dA2

Therefore, I (A) - [ -$- Eh ln f (x ;A) ] = E (x)A2 = A-.'

So that the Cramer-Rao lower bound is l/[nI (A)] - Mn. Since

X
Varh ( ) = Mn, is the UMVUE of A.

E l ) Let XI, X2, . . . ,Xnbe independent Bernoulli random variables, that is,
XI, X2, . . . ,X, are independent random variables with P (Xi 1) p,
P(Xi = Q) 1-pfori = 1,2,..., n.Showthatif
- -
-
5

S X1 +X2 + . . .+Xu,S/n isthe UMWEof p.

The next example demonstrates that a uniformly minimum variance unbiased

estimator for a parameter might exist but the Cxamer-Rao lower bound is not
attained.
Example 4: Let X be a Poisson random variable with parameter 0 and suppose we
-
wish to estimate 6 (0) e- 0 on the basis gf a sample of size one. Consider the
estimator
 - 0, otherwise.

:Then, E0 [ T (X) ] - 1.P0 [X - 01 - e-', so that T (X) is unbiased for e-e Also,

[ [ T (x) } ] - E0 [T (X) ] = e-e and hence

2
~ uE,t

VarO(T(X)) - e
-0
-e
-20
- e
-0
(1-e-').

.'
Now, the probability mass function of X is
P

andthus,lnf(x;Q) - f ( x ; 0 ) e-e8X/x!
-0+xln0-ln(x!),
-

I I
Also, 6 (0) - e- ',so that 6' (0) =
d
6 (0) - - ce.Hence, the Cramer-Rao
t lower bound to the variance of T (X), using (14), is
0 /

-
But Yare [ T (X) ] e- (1 - e- ') > 0e-20 for 0 > 0. Thus, T (X), though unbiased
- -
for 6 (0) e e, has a variance larger than the Cramer-Rao lower bound. However,
'Irf
it can be shown that T (X) is the only unbiased estimator of 6 (8)' e and hence -
Is the U M W E of e-e.
?4#
We now bring in another important concept, namely; that of sufficient statistic and
touch upon it briefly. Let X be a random variable having probability density (or,
mass) function f (x ; 8) and XI, X2, . . . ,Xn be independent observations on X that
is, let XI, X2, . . . ,X, be a random sample from a mulation with density (mass)
function f (x ; 8). The joint distribution of ( XI, X2, . . . ,Xa ) clearly depends on 0.
Is it possible to find a statistic (a furiction of ( XI, X,, . . . ,X, ) which contains all
the "information" about 0? Such a question becomes relevant when we want to
summarize the available data, because storing large bodies of data is expensive and
might give rise to errors of recording etc. Moreover ,it is unnecessary if we are able
to summarize the data without losing any "information". A statistic containing all
information about 0 is called a sufficient statistic. We give below a praise
definition.
 -

Elements or Stattstlcal A statistic T = T ( XI, X2, . . . ,Xn ) is said to be a sufficient statistic for the
Inference
parameter 8 if the conditional distribution of ( XI, X2, . . . ,X, ) given T does not
depend on 8.

From the above definition, it is clear that if there is a sufficient statistic for 0, then
since the conditional distribution of XI, X2, . . . ,X,, given the sufficient statistic is
independent of 8, no other function of the observations can have any additional
information about 8,given the sufficient statistic.

16.3 METHODS OF ESTIMATION

In this Section, we shall discuss some common methods of finding estimators. We
concentrate on two useful and commonly used methods, namely, the method of
moments and the method of maximum likelihood.

16.3.1. Method of Moments

The method of moments for estimation of paraqeters is often used mainly because
of its simplicity. The method consists in equating sample moments to population
moments and solving the resulting equations to obtain the estimators.
..
Let X1, X2, . ,Xnbe a random sample from a population with distribution
-
function depending on a k-dimensional parameter 9 (81, e2,. . . . . ,%). Let

be the r-th sample moment. Suppose & = E (X ' ) exists for r = 1, <. . . ,k. The
mgthod of moments involves solving the equation
m,' = %' (el, E)2, . . . . . ,eL)., 1s r s k.

In order to estimate the k components of J€ one clearly needs to equate at least k

sample moments to k population moments. However, which of the k moments are
4
to be equated is not specified. In practice, one generally takes the first k moments .
The method is now illustrated by some examples.

Example 5: Let XI, X2, . . . ,X,,be a random sample from a normal population with
-
mean p and variaace u 2. Here, the parameter 9 (p, o 2 ) is 2-dirhensional. In
.order to obtain the method of moments estimators of p and u 2, we equate the first
I
two sample moments to the corresponding population moments, that is,

mll - n-I
n
Exi
i- 1
--
XisequatedtoE(X) - p
*
andm,' - - EX:
n
-
1
n

i 1
'

isquatedto~(x2) - $+a2.

The first of these two equations gives X as an estimator of p ;

A -
second, using p = X, we have an estimator of a as
2
- ?. From the
Observe that p
A

" 2
- -X is an unbia
estimator of p but & is not unbiased for a 2.
However, both p and a are co~.,-.ent estimators of p and a 2 respectively.
A
.. - . I

Polnt fithation
I

Example 6: Let XI, X2,. . . ,X,be a random sample from a uniform distribution
with density function

- 0, elsewhere.

Then, ktl - E (X ) - (a+ $)/2, E ( x2) - -

kt2 (a2 + a$ + p2)/3.

Instead of equating mt1 to ptl and m i to pt2,we may as well equate mtl to pP1

and m2 - n-I 2 (Xi - zf to fi - Var (X). It is easy to see that

II

i-1

Thus, the equations to be solved are

and n-I 2 ( x i - Kf - ($ - aY/12

The.solution of these equations give us the method of moments estimators of

a a n d $as

E2) Let XI, X2, . . . ,Xn be a random sample from a Poisson distribution with
parameter A. Obtain two estimators of A using the method of moments.
E3) Let XI, X2, . . .,X, be a random sample of size N from a binomial population
with parameters n and p, both unknown. Obtain the method of moments
estimators of n and p.

As we have mentioned earlier, the method of moments is useful in practice because

of its simplicity. The properties of such estimators are not established in general and
have to be investigated separately for each estimator. Another method, which gives
'efficientw estimators for large sahples, under some reasonable conditions, is the
method of maximum likelihood. We study this method in the following subsection.
 Elements of Statistical
I Idereace 16.3.2 Method of Maximum Likelihood
I
I
To appreciate this method of estimation, it is perhaps best to start with an example.
Example 7: Let XI, &, .. .,Xnbe a random sample of size n fiom a Poisson
population with parameter 8. The likelihood function, based on the observations
'XI, ...,%is

The method of maximum likelihood consists in choosing as an estimator of 8 that

value of 8 (say 8,)which maximizes the likelihood function L,,
(8).go is called the
maximum likelihood estimator of 8. Obviously, 80 depends on the observed sample
XI, &, .. .,&. In order to find a maximum likelihood estimator of 8, we have to
find the value of 8, at which 4 (8)is maximum over the interval (0,a),as 8 > 0
here. Now,

It is known that In L, (8)attains its maximum at a point if and only if L, (8)

attains its maximum at go Now,

d
Therefore, -ln 4 (8) 1
de
- $ = 0 provided $ - -
n-I
n

i 1
Xi.

2
-
In order to verify whether L,(8)is indeed maximum at 0 go, we compute the
second derivative of In L,(8) at 8 = go and check whether it is negative . Here,

d2
and clearly, 7 in L,(0) I a < 0. This shows that 4 ( 8 )is maximized at
do

8 - 8, =
i-1
X/n. Since there is a unique maximum for L,( 8 )and the maximum

is attained at 8 - -8, dl $ Xi, go is the maximum likelihood e s e a t o r of 8.

i-1

We next consider an example where the parameter 8 is a vector instead of a scalar

as in Example 7.
Example 8: Suppose XI, X2, . .. ,Xn is a random s h p l e from a normal population
with mean p and variance a2,both unknown. The likelihood function is

,,I I , 02, = (hinn.

exp 1- 2 . 1
2n2:
- pf (021-nfl
 Thus,

where c is a constant independent of p and o '. The partial derivatives of Lo 4 @

2
w.r.t.panda are

Equating these two partial derivatives to zero, we get the likelihood equations.
These equations have unique solutions

The verification of the fact that these solutions actually maximize the likelihood
G
function is left to the reader. Hence, and 6 are the maximum likelihood
2
estimators of p and a respectively.
E4) Let XI, XQ,. . . ,X,, be a random sample from a population with density
function

I = 0, elsewhere.
j Find the maximum likelihood estimator of 8.
I
In the case of a scalar parameter, the likelihood function is a function of one
variable (as in the case of Example 7) and if this function is twice differentiable in
the domain of its definition, then one can use the methods of Calculus to find the
maximum. However, if the parameter 8 is a vector parameter, the likelihood
function is a function of several variable and finding the points of maxima of such
functions might be difficult in general. In such cases, special methods, depending
lon the problem on hand are needed. Of course, it is possible that the likelihood
function may not be differentiable at all and in that case also, we might have to
resort to special techniques. The following example is an illustration of such a
situation.
Example 9: Let XI, X2, . . . ,X,,be a random sample from a uniform distribution
with density function

The likelihood function is

- 0, elsewhere.

- 0, otherwise
We can write the likelihood function alternatively as
L, (8) - 8-', if 0 r xfn,r 8

Where x(,) is the largest observation in the sample. The derivative of L, (8) does not
vanish and hence, we cannot use the methods of Calculus to g ~a tmaximum
likelihood estimator. However, L,(8) attains its maximum at 8 = x(,) and x(,) is
the unique maximum likelihood estimator of 8.

There is another way to look at the same problem. Since 4 (8) €I-', 0 r Xi s 0

-
is an ever-decreasing function of 8, the maximum can be found by selecting 8 as
...
small as possible. Now, 8 r X, for i 1,2, . . ,n and in particular, 8 a x(,).
Thus, 4 (8) can be made no larger than l/x;,) and the unique maximum likelihood
estimator of 8 is xfn).
Are maximum likelihood estimators unbiased and unique h every situation?
The answer to both the above questions is in the negative. That maximum
likelihood estimators need not be unbiased is demonstrated by making an appeal to
n
Examples 8 and 9. In Example 8, we had seen that 6' (Xi - z)2is the
i-1
maximum likelihood estimator of a 2, the variance of a normal population with
unknown mean p. Clearly, this estimator is not unbiased for a 2. Again, in Example
9, it was demonstrated that x(,), the largest observation in the sample is the
maximum likelihood estimator of 8. But, it can be shown that
Eo (4)) = nW(n + I), so that xo is not unbiased for 8.

To see that maximum likelihood estimabr need not be unique, consider the
following example. a

Example 10: LRt XI; X., . . . ,& be a random sample from a uniform distribution

I :, :1
over 8--
L

likelihood is
8 + - ,where8isunknowi;and8ES2 = ( x : - o o s x < m ] . ~ h e
4

- 0; otherwise ;

- 0, otherwise.

Thus, ;
,(8) attains its maximum provided

8-1/2 s min(Xl.. . . . . ,%)

and 8 + 1/2 r max (XI3. . . . .,x n ) 9

or, when
8 s min (XI, . . . . , . q)+ 1/2
 and . . . . .,&)- 1/2.
8 r max (XI,
This means that any statistic T (XI, . . .,X,,)satisfying
m a x i - 1 / 2 r T(X1,....., &)sminX,+1/2
i

is a maximum likelihood estimator of 8. In fact, for 0 < a < 1,

T(X,,X2, .. .. .,X,,) - (maxXi-1/2)+a(mainXi-mqX;+ 1 )

1 1 I

lies in the interval max Xi - 1/2 r T s min Xi + 1/2. Thus, for any a, O< a < 1,
i i

- -
the above estimator is a maximum likelihood estimator of 8. In particular, for
a 1/2, we get an estimator T1 ( rnax X; + min X, )/2 and for a = 1/3, we
i i
get the estimator T2 = ( 4 max Xi + 2 min X; - 1 )/6.
i i

Both TI and T2 are maximum likelihood estimators of 8.

Are thcr e any "good" properties of maximum likelihood estimators?

Before we attempt to answer this question, we introduce the concept of asymptotic

efficiency. An estimator Tn based on a sample of size n for a parameter 8 is said to
be asymptotically efllcient if lim n Var (Tn) = 1A (8) where I (8) is the per
n+-
observation (Fisher) information. Recall that the Cramer-Rao lower bound to
Var, ('T,) k
is I/[ nI (8) under some regularity conditions.
a c

I ,
4
The important properties of maximum likelihood estimators are that under certain
, regularity conditions, these estimators are
(i) Consistent
(ii) Asyinptotically efficient
(iii) Asymptotically normal with mean 8 and variance I/[ nI (8) 1.
The third property says that for large samples, the distribution of the maximum
likelihood estimato& of 8 is approximately normal with mean 8 and variance
I/[ a1 (8) 1.
The exact statements of the above results and their proofs are beyond the scope of
this course and are therefore not given here.
&*. -
6
-"(I
$
s
d
?
15.4 SUMMARY
i

In this unit, we have

1. discussed some prbperties that an estimator should preferably possess, like
unbiasedness, consistency and efficiency,
2. derived the Cramer-Rao lower bound to the variance of an estimator and
demonstrated the use of this bound in finding minimum variance unbiased
estimators,
3. discussed two commonly used methods of estimation, namely, the method of
moments and the method of maximum likelihood.
16.5 SOLUTIONS AND ANSWERS
E l ) Here, the probability mass function of the random variable, X is

f(7';p) - px(l-Pf-X,~ - 0,l.

Ther&forelnf(X;p) - Xlnp+(l--X)ln(l-p),
d
-Inf(X;p)
dp
- Wp-(1-X)/(l-p),

since E, (X) = p. Therefore, the Crarner-Rao lower bound to the variance is

p (1 - p)/n. Let S = XI + X,+ . . .+ X,. Then S/n is unbiased for p. Also,

var, (S/n) - 2
n-2
i-1
Var, (x,) = n- 2 [ np (1 - P) I = P (1 - p)/n. Hen-
S/n is the U M W E of p. %

E2) Here XI, X2, . . . .Xnis a random sample from a Poisson distribution with
parameter A. Hence E (Xi) h for i - -
1,2, . . . ,n. Equating the sample
mean to the population mean leads to the following equation:

which gives a moments estimator of A as fi = R - ~ ~ a isince

n,
E ( x2) = A2 + A, equating the second sample moment a'bout zero, viz.,
. n-I 2- q2
i 1
to the corresponding population moment yields the equation

Since A > 0, a unique positive solutidn of the above equation gives the second
moments estimator of h as

E3) We are given that XI, X2, . . . . XNis a random sample from a binomial
population with parameters n and p, both unknown. We know that if X has a
binomial distribution with parameters n and p, then
E, (XI - np, Varp (X) - np ( 1 - p).
 equate the fmt two sample moments K1 z-
N

1 1
X, and K' 2
N

i-1
to the first
Pdnt EdWtlon

two population moments, the following equations result:

The first of these gives = m n as an estimator of p, where is an estimator

of n. Using this estimator in the second equation and solving for n gives .

E4) Here XI, X2, . . . .X,,is a random sample from a population with density
function

- 0, elsewhere
Therefore, the likelihood function is

d
and dB ln 4 (8) - -dB
n
+ 2 X,/e2.
x- 1

d
Equating -In L, (8) to zero, gives on solving for 8,
de

which is negative at 8
estimator of 8.
-6-R Hence is the maximum likelihood
Elements of Statistical
Inference

I. Let XI, X2, . . . . Xn be a random sample from a distribution with density

function

= 0, otherwise.

Show that R = n- ' 2 X, is unbiased for and Var, ( St ) = e2/n.

Does Vare ( ST ) attain the Cramer-Rao lower bound.?

2. Let XI, X2, . . . . Xn be a random sample from a normal population with mean
zero and variance a 2. Construct an unbiased estimator of a as a function of

. You are given that if X is normal with mean zero and variance a 2 ,
i

3. Let XI, X2, . . . . X, be a random sample from a distribution having finite mean

p and finite variance a 2. Show that T ( XI, X2, . . . . X, ) =

n (n + 1)
i i x i
i- 1
is unbiased for p.
4. Let XI, X2, . . . .Xn be a random sample of size n from a distribution with
probability density function

4
- 0, elsewhere.
Obtain a maximum likelihood estimator of 8.