Homework 3 Prob 2.4.2 p x X e X , 0, X 0 p p n* , l* n l* * n* e l* n* 1 1. find expectation and variance E ( ) 0 n l* * n* e l* n* 1d , n l * * e l* n* 11d , 0 n* 1 l* l* n* 11 d . e n* 0 E ( ) 1 U e U n* 11 dU , l*n* 0 Let U l* . We have 1 n* 1 l*n* n *. l* n l* * E ( 2 ) 2 n* 0 n*n* , l*n* (ANS) e l* n* 1d 1 l * l* n* 2 1 d . e l*n* 0 Again, let U l* . We have E ( 2 ) Hence, the variance is given by 1 l*2 n* 0 e U U n* 2 1 dU , n* 2 (n* 1)( n* ) l*2 n* l*2 (n* 1)( n* )n* l*2 n* n 2 n* . * l*2 , var E 2 E 2 , n*2 n* l*2 2. find p x X First, we need to find p x X n*2 n *. l*2 l*2 (ANS) p x X p x X p n* , l* d , 0 e n l* * X n* 0 e l* n* 1d , n l* * 0 n* l* Let U l* X . Hence, we have X n* e l* X l* X n* 11 d , n px X l* * e U U n* 11 dU 0 n* l* X n n l* * n* 1 n*l* * . n* l* X n* 1 l* X n* 1 n* 1 Using the Bayes’ rule, we obtain p x X e p x X p n* , l* px X X n l* * n* , e l* n* 1 , n n*l* * l* X n* 1 1 e l* X l* X n* 11 . (ANS) n* 1 The PDF of the posterior probability is of the same form as the prior probability. Hence, the mean square estimate is given by ˆms E X 2 E ˆms E ˆ2ms 2ˆms 2 , n* 1 l* X (ANS) E X E ˆ2ms 2ˆms 2 X , since Ea( x, y) E x E y a( x, y) X . It follows that 2 E ˆms E X ˆ2ms 2ˆms E X E 2 X , because ̂ms is a function of X only and does not depend on when X is given. The above equation can be reduced more and we have 2 E ˆms E X ˆ2ms 2ˆms ˆms E X E 2 X , 2 E ˆms E 2 E X ˆ2ms . Hence, we obtain 2 n 1 2 2 n* n* ˆ , E ms E X * l*2 l* X 2 n*2 n* n* 1 EX 2 2 l* l* X n*2 n* n*2 n* l*2 n * 0 l* 12 X 2 l* X n* 1 n n * * n n*l* * 12 l* * n n 3 0 l* X * n*2 n* n* n* 12 , n* 2l*2 l*2 n n 1 = * * , n* 2l*2 l*2 dX , dX , (ANS) 3. N observation case N p x X p x X i p n* , l* d , i 0 i 1 N e N X i i 1 0 0 n* N l* n* n l* * n* e l* n* 1d , e l* l n* 1d , where l is defined in the text page 142. Using the same procedure as the single observation case, we obtain, n px X 0 l* l l* * n* N n* e l* l l* l n* N 1 d l* l , n l* * n* N l* l n* N n* . Using the Bayes’ rule, we obtain p x X p x X p n* , l* p x X N N X i e i 1 n l* * n* , e l* n* 1 , l* * n* N n l* l n* N n* l l n* N e l* l n* N 1 . * n* N It follows that n N and ̂ms * l* l (ANS) 2 E ˆms E 2 E X ˆ2ms n N 2 n*2 n* E X * 2 l* l* l n*2 n* l*2 n*2 n* l*2 n* n* 1n* N 2 n* N 1n* N l*2 , n n 1n* N * * n* N 1l*2 n* n* 1 n* N 1l*2 . Find MAP estimate ln p x X n* N ln l* l ln n* N l* l n* N 1ln( ) , ln p x X n N 1 l* l * 0 (ANS) Hence, ˆMAP n* N 1 ˆ MS l* l Prob. 2.4.3 r an n ~ N 0, n R A2 pr a R A exp n 2 2 n2 1. A m 2 k0 0 p a A exp 2 n 2 2 n / k 02 1 Since both a and n have Gaussian distributions and are independent, r is also Gaussian distributed with Er En Ea m0 and var r var n var a n2 k 02 k02 1. Hence, we can write p r R n k 2 R m 2 0 exp 0 . 2 2 2 2 k 1 k 0 1 2 n 0 k0 The posterior density is given by p r a R a p a ( A) pa r A R , p r R A m 2 R A2 k 0 0 exp exp 2 2 n 2 2 n / k 02 2 n n 2 , k 2 R m 2 k0 0 exp 0 2 2 2 2 k 1 n k 0 1 2 n 0 1 k02 1 exp R A2 A m0 2 k02 R m0 2 . n 2 2 n2 2 n2 / k 02 2 n2 k 02 1 Consider on the exponent term. (1) R A2 A m0 2 k 2 R m0 2 R2 2 RA A 2 k 02 A 2 0 2 n2 / k 02 2 n2 k 02 1 2 n2 2 n2 2 n2 2 n2 2 n2 k 02 Am0 k 02 m0 2 k 02 R m0 2 2 n2 2 n2 2 n2 k 02 1 A2 2 n2 1 k 22A R m k 2R 2 0 2 n k 2 m 2 k 2 R m0 2 - 0 0 0 , 2 2 2 2 2 2 k 1 n n n 0 2 2 0 0 2 2 2 1 k 02 R m0 k 02 R m0 k 02 A 2 2 2 1 k 2 n2 1 k 02 n 0 2 , 2 2 2 2 2 2 2 1 k 02 A R m0 k 0 R m0 k 0 R - k 0 m0 k 0 R m0 , 2 2 1 k 02 2 n2 1 k 02 2 n2 2 n2 2 n2 k 02 1 n R 2 k 02 R 2 m02 k 04 k 02 m0 2 k 02 R 2 2 Rm 0 k 02 m02 k 02 , 2 n2 k 02 1 2 2 1 k 02 R m0 k 02 R m0 k 02 A 2 2 1 k 02 2 n2 1 k 02 n R 2 m02 k 04 2 Rm 0 k 02 , 2 n2 k 02 1 2 2 2 2 2 2 1 k 02 A R m0 k 0 R m0 k 0 R m0 k 0 , 2 2 1 k 02 2 n2 1 k 02 2 n2 1 k 02 n 2 2 1 k 02 A R m0 k 0 . 2 2 1 k 02 n Hence, (1) becomes pa r A R k02 1 exp 1 k02 A R m0 k02 2 n 2 2 2 n 2. Extend to N independent observations We know that 1 k 02 (QED) pr R pr a R a p a ( A)dA N 2 . Ri A 2 k0 A m i 1 0 dA exp N 1 / 2 2 2 2 N 1 2 n 2 n / k 0 2 n Let us consider on the exponent term. Hence, we have (2) N Ri A2 i 1 2 n2 A m0 2 2 n2 / k 02 N 1 N 2 R 2 A Ri NA2 A 2 k 02 2 Am0 k 02 m02 k 02 , i 2 n2 i 1 i 1 NR N 1 2 A N k 02 2 A N i m0 k 02 Ri2 m02 k 02 , 2 n2 i 1 N i 1 2 2 Nl m0 k 02 Nl m0 k 02 N k 02 1 N 2 2 2 Ri m0 k 0 , A 2 2 N k 2 2 2 i 2 n2 2 N k 1 0 n n 0 2 Nl m0 k 02 N k 02 f (R ) , (3) A 2 2 2 n N k 0 where 2 Nl m0 k 02 1 N 2 2 2 f (R ) . Ri m0 k 0 2 2 2 n2 i 1 2 n N k 0 Hence, (2) becomes pr R n 2 N 1 2 N 1 / 2 n k 02 N k02 exp f R , The posterior density is given by pr a R a p a ( A) pa r A R , pr R N 2 Ri A 2 A m0 k0 exp i 1 N 1 / 2 2 N 1 2 n 2 n2 / k 02 n 2 n 2 N 1 2 N 1 / 2 n k 02 , N k02 exp f R N 2 Ri A N A m0 2 i 1 exp f R 2 2 2 n 2 2 n 2 n / k 0 Using (3), we have 2 2 Nl m0 k 02 N k 02 N k 0 f (R ) f R , pa r A R exp A N k 2 n 2 2 n2 0 k 02 pa r A R N k02 exp N k02 A Nl m0 k02 2 N k 2 n 2 2 n2 0 (ANS) Prob. 2.4.9 1. Does an efficient estimate of the standard deviation of a zero-mean Gaussian density exist? pr R 1 2 R2 2 e 2 Taking ln(.) both sides of the above equation, we obtain ln pr R ln( ) 0.5 ln 2 R2 . 2 2 Differentiate the above equation with respect to and let it equal to zero; i.e., ln p r R 1 (2) R 2 0 2 3 (1) Hence, we have 2 ˆ ML R2 , or ̂ ml R Replace ̂ ml R in (1). We reach 2 ln p r R 1 (2)ˆ ml 2 3 Obviously, the above equation cannot be written in the form of ln p r R k ( )ˆ ml . As a result, we conclude that the efficient estimate of a standard deviation does not exist. 2. Does an efficient estimate of the variance 2 of a zero-mean Gaussian density exist? We begin the problem by rewriting the conditional probability of R given 2 , and we obtain p R 2 r 2 1 2 2 R2 2 e 2 To avoid the confusion, we denote 2 by v. Thus, the above equation becomes pr v R v 1 2v e R2 2v Taking ln(.) of the above equation, we have 1 R2 ln pr v R v ln( v) ln( 2 ) 2 2v Next, we differentiate the above equation with respect to v, and we have ln p r v R v 1 R2 v 2v 2v 2 Let (2) equal to zero; i.e., 1 R2 0, 2vml 2vml 2 vˆml R 2 We replace vˆml R 2 in (2), and obtain ln p r v R v 1 vˆ ml , v 2v 2v 2 1 2v 2 vˆml v k v vˆml v Hence, the efficient estimate of variance exists. (2) Prob. 2.4.12 s1 x11 A1 x12 A2, (a) s 2 x21 A1 x22 A2, and r1 s1 n1 , (b) r2 s 2 n2 where ni is i.i.d. N(0, n ) pr s R S R s 2 R s 2 2 2 exp 1 1 2 2 n 2 2 n 1 Hence, R x A x A 2 R x A x A 2 1 11 1 12 2 2 21 1 22 2, pr a R A exp 2 2 n 2 2 n Taking ln(.) the above equation, we obtain 1 R1 x11 A1 x12 A2 2 R2 x21 A1 x22 A2, 2 2 ln pr a R A ln n 2 2 n2 (1) Differentiate (1) with respect to A1 we have R x A x A x R2 x21 A1 x22 A2 x21 ln pr a R A 1 11 1 12 2 11 A1 n2 Letting the above equation equal to zero, we achieve R1 x11 A1 x12 A2 x11 R2 x21 A1 x22 A2 x21 n2 0, 2 2 A1 x11 x21 A2 x11 x12 x21 x22 x11 R1 x21 R2 Again, differentiate (1) with respect to A2 we have R x A x A x R2 x21 A1 x22 A2 x22 ln pr a R A 1 11 1 12 2 12 A2 n2 (2) Letting the above equation equal to zero, we achieve 2 2 A1 x11 x12 x21 x22 A2 x12 x22 x12 R1 x22 R2 (3) Using the linear algebra to solve (2) and (3), we obtain that aˆ1ml x22 R1 x12 R2 , x11 x22 x12 x21 (4) aˆ 2 m l x21 R1 x11 R2 x11 x22 x12 x21 (5) and 2 Compute the variances of estimates x R x R E aˆ1ml E 22 1 12 2 , x11 x22 x12 x21 x ER1 x12 ER2 E aˆ1ml 22 x11 x22 x12 x21 var aˆ1m l E aˆ1m l E aˆ1m l 2 , x R E R x R E R 2 1 12 2 2 , E 22 1 x11 x22 x12 x21 2 2 x22 var( R1 ) x12 var( R2 ) x11 x22 x12 x21 2 2 2 x22 x12 2, 2 n x11 x22 x12 x21 because R1 and R2 are statistically independent. , (6) By using the same procedure, we obtain var aˆ 2 ml E aˆ 2 ml E aˆ 2 ml 2 Next, find the cross correlation 2 2 x21 x11 x11 x22 x12 x21 2 n2 (7) E aˆ1m l E aˆ1m l aˆ 2 m l E aˆ 2 m l x R E R1 x12 R2 E R2 x21 R1 E R1 x11 R2 E R2 , E 22 1 x x x x x x x x 11 22 12 21 11 22 12 21 x x E R1 E R1 2 x12 x11 E R2 E R2 2 , 22 21 x11 x22 x12 x21 2 x x x x 22 21 12 11 n2 x11 x22 x12 x21 2 Hence, we have 2 2 x22 x12 x11 x22 x12 x21 2 x22 x21 x12 x11 n2 x22 x21 x12 x11 2 2 x21 x11 3. Find fisher’s information by continue differentiation 2 ln pr a R A A12 2 2 x11 x21 n2 x11 x12 x21 x22 2 , ln pr a R A A1A2 n2 x11 x12 x21 x22 2 ln pr a R A A2 A1 n2 2 2 x22 x12 2 ln pr a R A A22 n2 Hence, J we have 2 2 1 x11 x21 n2 x11 x12 x21 x22 x11 x12 x21 x22 , 2 2 x22 x12 and J 1 2 2 x22 x12 x11 x22 x12 x21 2 x11 x12 x21 x22 n2 We observe that J 1 Σ . x11 x12 x21 x22 , 2 2 x11 x21 Hence, we can conclude that the estimates are efficient. Prob. 2.4.17 Eaˆ A A B( A) (1) ซึง่ หมายความว่า Eaˆ A B( A) A aˆ A B( A) pr a R AdR 0 (2) differentiate สมการข้ างบนเทียบกับ A จะได้ 1 dB A / dA pr a R AdR aˆ A B( A) pr a R A A dR 0 (3) 1 จาก calculus เรารู้วา่ ln u 1 u x u x (4) ดังนันแล้ ้ ว เราจะได้ วา่ pr a R A ln pr a R A A (5) ln pr a R A dR 1 dB A / dA A (6) A pr a R A เอาสมการ (5) .ใส่ไปใน (3) จะได้ aˆ A B( A) pr a R A เขียนสมการข้ างบนใหม่ได้ เป็ น ln pr a R A A pr a R A aˆ A B( A) pr a R A dR 1 dB A / dA ใช้ Schwartz’s inequality ที่เขียนว่า (7) 2 2 x t dt y t dt xt y t dt จะเท่ากันก็ตอ่ เมื่อ y (t ) kx(t ) 2 (8) เมื่อ k เป็ นค่าคงที่ ดังนันสมการ ้ (7) เขียนใหม่เป็ น ln p R A 2 ra pr a R AdR A 2 2 aˆ A B( A) pr a R AdR 1 dB A / dA (9) จะเท่ากันก็ตอ่ เมื่อ ln pr a R A pr a R A k ( A)aˆ A B( A) pr a R A A (10) สมการ (9) สมารถเขียนใหม่เป็ น ln p R A 2 ra E A AE aˆ A B( A) 2 A 1 dB A / dA2 หรื อ ln p R A 2 ra var aˆ A A 1 dB A / dA2 E A 1 (11)