Lecture Notes AMS312 Spring207 Mar 6th Example: Let Y1 , Y2 ,..., Yn be a random sample from a population with the uniform distribution p.d.f is ˆ2 f ( y; ) 1 0 y . Last time we have shown that both ˆ1 2Y and n 1 Y( n ) are unbiased estimators for , which one is better? n The basic rule is: compare the Var (ˆ1 ) and Var (ˆ2 ) , the smaller, the better. Here is an example to help you understand the rule: Example: Let Y1 , Y2 ,..., Yn be a random sample from N ( , ) . Consider the following estimator 2 for . ˆ1 Y ˆ 2 Y1 Both of them are unbiased: E (Y ) Y 2 N ( , n E (Y1 ) ) and Y1 N ( , 2 ) now Var (Y ) 2 n 2 Var (Y1 ) we can show that Y is a better estimator than Y1 . 95% confidence interval based on Y is [Y 1.96 n , Y 1.96 n ] (suppose is known) 95% confidence interval based on Y1 is [Y1 1.96 , Y1 1.96 ] And we can see that the length of the confidence interval based on Y is 3.92 n while the length of the confidence interval based on Y1 is 3.92 . Now the first length is much smaller than the second length, hence the estimate is more precise. Hence the first estimate is better than the second one. Based on this rule, we can calculate which one is better. EY 0 y dy 2 And EY 0 1 y 2 dy 1 1 3 1 2 * 3 3 1 2 2 2 2 Hence Var (Y ) ( ) Therefore Var (2Y ) 4* 3 2 12 12n Var (ˆ2 ) Var ( n 1 (n 1) 2 Ymax ) Var (Ymax ) n n2 1 Lecture Notes AMS312 Spring207 y FY( n ) ( y ) P(Y( n ) y ) P(max{Y1 , Y2 ,..., Yn } y ) P(Yi y ) ( ) n fY( n ) ( y) d ny n1 FY( n ) ( y) n dy E (Y( n ) 2 ) 0 E (Y( n ) ) 0 ny n 1 n ny n 1 n y 2 dy ydy Therefore, Var (Y( n ) ) n n n 2 n n2 n2 n n 1 n 2 n 2 ( )2 n 2 n 1 n(n 2) Now we see that Var (ˆ1 ) Var (ˆ2 ) hence ˆ2 is a better estimator. If we define 3 2Y1 then E (2Y1 ) and Var (2Y1 ) 4Var (Y1 ) 2 3 . It is bigger than Var (ˆ1 ) . If we define ˆ4 ˆ1 ˆ2 2 hence E (ˆ4 ) the variance is between the variance of ˆ1 and the variance of ˆ2 . Var (ˆ1 ) n 2 The efficiency of ˆ2 in relative of ˆ1 is Var (ˆ2 ) 3 Also study ex5.4.6 in page 389. Section5.5 Minimum—variance estimators. The C-R Lower bound. Theorem: Let Y1 , Y2 ,..., Yn be a random sample from the continuous pdf fY ( y; ) , where fY ( y; ) has continuous first-order and second-order partial derivatives at all but a finite set of points. Suppose that the set of ys for which fY ( y; ) 0 does not depend on . Let ˆ h(Y1, Y2 ,..., Yn ) be any unbiased estimator for .Then Var (ˆ) {nE[( ln fY ( y; ) 2 1 2 ln fY ( y; ) 1 ) ]} {nE[ ]} 2 (A similar statement holds if the n observations come from a discrete pdf, p X (k ; ) ) 2 Lecture Notes AMS312 Spring207 Example: f ( y; ) 1 0 y cannot apply the Cramer-Rao lower bound. Example: If Y1 , Y2 ,..., Yn is random sample for N ( , ) 2 1 Population pdf: fY ( y; , ) e 2 2 ( y )2 y 2 2 For the estimator of : ˆ1 Y (MLE and MOME) ˆ 2 Y1 , ˆ3 Y Y2 ... Yn 1 Y1 Y , ˆ 4 1 n 1 2 Now we calculate the Cramer-Rao lower bound: 1 ln fY ( y; , ) ln[ e 2 ( y )2 2 2 2 ] ln( 2 ) ( y )2 2 2 2 ln fY ( y; , 2 ) y 1 [ 2 ] 2 2 E ( ˆ ) and Var ( ˆ ) n[ E ( 1 2 n )]1 2 Definition (Best estimator (minimum variance estimator for )): Let be an unbiased estimator for if Var ( ) Var ( ) where is any unbiased estimator * * for . is called the best estimator for . * Definition (Efficient estimator for ): If Var ( ˆ ) reaches the Cramer-Rao Lower Bound, then ̂ is called the efficient estimator for . 3