Yale STAT 619 - Bayes estimation, minimaxity and Admissibility

Unformatted text preview:

Week 3Spring 2009Lecture 5. Bayes estimation, minimaxity and Admissibility (cont.).AdmissibilityConditions on priors and admissibility: conditions on the prior measurewhich guarantees that the corresponding generalized Bayes procedure is admis-sible.De…neJx(h) =Zh () ' (x  ) d.Let S1= fx 2 Rp: kxk  1g.AssumptionGrowth Condition:ZSc1g ()kk2log2(kk)< 1Asymptotic ‡atness condition:ZSc1Jx(rggJx(rg)Jx(g)2g)dx < 1It can be shown thatRSc1krgk2=gdx < 1 implies the ‡atness condition.Theorem. Let G be a prior satisfying two conditions above. Then Gisadmissible.For the normal mean estimation problem with squared error loss,Blyth’s Method. Let  be an estimator. Let fGjg be a sequence of …-nite prior measures such that:(i) r(Gj; )  r(Gj; Gj) ! 0 as j ! 1; (ii)infjfGj(S1)g > 0. Then  is an admissible estimator.hint : Let 00=0+ =2. If R; 0 R (; ) for all  and with strictinequality for some , then R; 00< R (; ) for all . For all j;r(Gj; Gj)  r(Gj; 00) ZS1R; 00Gj(d) +ZSc1R (; ) Gj(d)= r(Gj; ) +ZS1R; 00 R (; ) Gj(d)  r(Gj; )  "where " =RSR; 00 R (; ) Gj(d). Contradiction!Example. Let X  N (; 1). Let gj(x) =1p2expx22j, and gj=dGj=d where  is the Lebe sgue measu re. It is easy to showr(Gj; X) =pj, r(Gj; Gj) =pjjj + 11thenr(Gj; X) =pjj + 1! 0.Proposition.r(G; ) r(G; G) =ZkG k2g (x) dxProof :r(G; ) r(G; G) = EXEk  G+ G k2 k  Gk2jX= EXEkG k2jX=ZkG k2g (x) dxProof of the theorem: Please read page 374 of Stein (1961).De…ne gj= h2jg wherehj=8<:1 kk  11 log(kk)log j1  kk  j0 kk > j, j = 2; 3; : : : .It is easy to seeGj=Rh2jg () ' (x ) dRh2jg () ' (x ) d! Ga.s.andg j(x) =Zh2jg () ' (x ) d  g (x) .WriteG(x) = x +rg g = x +Jx(rg)Jx(g);Gj(x) = x +rg g = x +Jxh2jrg + grh2jJxh2jgwhere the second equality for each equation follows from integration by parts.Hencer(Gj; G)  r(Gj; Gj)=ZSc1G Gj2g j(x) dx +ZSc1G Gj2g j(x) dx ( apply DCT, since g j(x)  g (x) …nite.) 2ZSc1Jxgrh2jJx(gj)2g j(x) dx + 2ZSc1Jx(rg)Jx(g)Jxh2jrgJx(gj)2g j(x) dx + o (1)= 2Aj+ 2Bj+ o (1)2Show Aj! 0 by DCT:Aj=ZSc1Jxgrh2jJx(gj)2g j(x) dx= 4ZSc1Jx(ghjrhj)Jx(gj)2g j(x) dx = 4ZSc1Jxg1=2j g1=2jrhjJx(gj)2g j(x) dx 4ZSc1Jxg krhjk2dx (Cauchy-Schwartz inequality and gj g) 4ZSc1krhj()k2g () dandkrhj()k2=1kk2log2(j)I[1;j](kk) 1kk2log2(kk_ 2)I[1;j](kk) .Show Bj! 0 by DCT again:Jx(rg)Jx(g)Jxh2jrgJx(gj)2g j(x) (Note that h2jrg is gjrgg)=JxgjJx(rg)Jx(g ) h2jrg2Jx(gj)=JxhgjJx(rg)Jx(g )rggi2Jx(gj) Jx gjJx(rg)Jx(g)rgg2!(Cauchy-Schwartz inequality) Jx gJx(rg)Jx(g)rgg2!.Admissibility of 0= X for p= 1; 2Let g () = 1, thenZSc1g ()kk2log2(kk)= 2Z121r log2rdr < 1.Homework problem (you pick one part to work on). Let Xi Poisson (i)be independent, i = 1; 2; : : : ; p. Denote X = (X1; : : : ; Xp) and  = (1; : : : ; p).Under the loss L (; ) =Ppi=1(i i)2=i, show that (i) for p = 1, X is anadmissible estimator of  using Blyth’s method; (2) for p  2, X is not anadmissible estimator of .3Reference: (i) Clevenson and Zidek (1975), Simultaneous Estimation of theMeans of Independent Poisson Laws, JASA.(ii) Brown and Hwang (1982), A uni…ed admissibility proof, Statistical De-cision Theory and Related Topics, III. S. S. Gupta and J. O. Berger (eds.)4Lecture 6. Bayes estimation, minimaxity and Admissibility (cont.).Superharmonic priors and minimaxity.De…nition. Let h : Rp! R be twice di¤erentiable. We call h super-harmonic ifr2h (x)  0 for all x 2 Rp, where r2h (x) =pXj=1@2@x2ih (x)and call h harmonic if r2h (x) = 0 for all x 2 Rp. The operator r2is calledLaplace operator (other notations include  and r r).Let X  N (; I). Let G be any prior such that g (x) < 1 for all x. Thenthe generalized Bayes estimator GisE (jX = x) = x + r(log (g (x))) .By Ste in’s unbiased estimate of the risk, we haveR (; G)  p = E(2r rg g +rg g 2)= E(2r2g g rg g 2).Let h =pg . We seeR (; G)  p = 4Er2h h (or 4Er2pg pg ).sinceE(2r rg g +rg g 2)= 4E(r rh h (x)+rh h 2)= 4E(r2h h krh k2(h )2+rh h 2)Theorem. If r2pg  0 (or r2g ), then Gis minimax.SinceR (; G)  p = E(2r2g g rg g 2).andr2g (x) =Zr2g () ' (x ) d,then Gis min imax if g is super-harmonic.Calculation of the Harmonic Bayes procedure.5Letg () = 1= kkp2This prior density is called the “harmonic prior”because it is harmonic at everypoint except  = 0. Note that it is a a valid generalized prior density whenp  3 in the sense that g (x) < 1 for all x . And it is true that g (x) issuperharmonic.Writeg () = 1= kkp2_Z10!(p4)=2e!kk2=2d!=Z101v2v1  vp=2evkk22(1v)dvtheng (x) _Z10v(p4)=2exp kxk2v2!dv,and thusrg g =R10vp=21expkxk2v2dvR10vp=22expkxk2v2dvx =1kxk2Rkxk20yp=21exp (y=2) dyRkxk20yp=22exp (y=2) dyxWe haveG(x) = x + r(log (g (x))) = 1 p  2kxk21  ekxk2=2Tp+2(kxk2)1  ekxk2=2Tp(kxk2)!xwhereTpkxk2=(p=22k=0(s=2)k=k! p  4 even(p1)=21k=02kk!s2k+1= (2k + 1)! p  3


View Full Document

Yale STAT 619 - Bayes estimation, minimaxity and Admissibility

Download Bayes estimation, minimaxity and Admissibility
Our administrator received your request to download this document. We will send you the file to your email shortly.
Loading Unlocking...
Login

Join to view Bayes estimation, minimaxity and Admissibility and access 3M+ class-specific study document.

or
We will never post anything without your permission.
Don't have an account?
Sign Up

Join to view Bayes estimation, minimaxity and Admissibility 2 2 and access 3M+ class-specific study document.

or

By creating an account you agree to our Privacy Policy and Terms Of Use

Already a member?