Partial solutions to assignment 7 Math 5080-1, Spring 2011 p. 328, #33. (a) Here τ (µ) = µ; therefore, τ 0 (µ) = 1. Also, because I{X1 = 0, 1, . . .} = 1 with probability one, it follows that f (X1 , µ) = e−µ µX1 . X1 ! Therefore, ln f (X1 , µ) = −µ + X1 ln µ − ln (X1 !) , and hence X1 d ln f (X1 , µ) = −1 + . dµ µ As is always the case, the expectation of the preceding is zero. Therefore " 2 # d d E ln f (X1 , µ) = Var ln f (X1 , µ) dµ dµ X1 1 X1 Var(X1 ) = . = Var −1 + = Var = 2 µ µ µ µ Therefore, the CRLB gives the following: If T is unbiased for µ, then Var(T ) ≥ 1 µ = . n/µ n (b) Now, τ (µ) = θ = e−µ . Therefore, the numerator in CRLB changes from one [in part (a)] to the square of τ 0 (µ) = −e−µ , which is e−2µ . Therefore, the CRLB gives the following: If T is unbiased for e−µ , then Var(T ) ≥ e−2µ µe−2µ = . n/µ n (c) Var(X̄) = µ/n, which is the CRLB. Therefore, a UMVUE for µ is X̄. (d) By the invariance property, θ̂ = e−µ̂ , where µ̂ is the MLE for µ. Therefore, we compute µ̂ next: The likelihood function is Pn e−nµ µ L(µ) = Qn i=1 Xi j=1 (Xj !) e−nµ µnX̄ = Qn ; j=1 (Xj !) and therefore, ln L(µ) = −nµ + nX̄ ln µ − ln n Y (Xj !) . j=1 1 We differentiate to find that d nX̄ L(µ) = −n + . dµ µ Set this equal to zero to find that µ̂ = X̄ is the MLE for µ. Therefore, θ̂ = e−X̄ is the MLE for θ = e−µ . Pn (e) Let S := i=1 Xi and note that θ̂ = e−S/n . Because S ∼ Poisson(nµ), we can deduce that E θ̂ = E e−S/n = MS (−1/n), where MS (t) = E[etS ] is the MGF for S. The table in your text tells you that MS (t) = enµ[e and hence E(θ̂) = enµ[e t −1] −1/n , −1] . Therefore, Bias(θ̂) = enµ[e −1/n −1] h i −1/n 1 ] −1+ n −1 . − e−µ = e−µ enµ[e Therefore, θ̂ is biased. (f ) By Taylor’s expansion, e−1/n = 1 − 1 1 + ± ··· n 2n2 Therefore, e−1/n − 1 − 1 1 ≈ 2 n 2n as n → ∞. Plug to find that h i Bias(θ̂) ≈ e−µ eµ/(2n) − 1 . Another round of Taylor expansions yields eµ/(2n) ≈ 1 + µ . 2n Therefore, Bias(θ̂) ≈ µe−µ →0 2n as n → ∞. This shows that θ̂ is asymptotically unbiased. 2 (g) Let S := Pn i=1 Xi and α := (n − 1)/n. Because θ̃ = αS = eS ln α , it follows that E(θ̃) = MS (ln α) = enµ[e ln α −1] = enµ[α−1] = e−µ . Therefore, θ̃ is unbiased for θ = e−µ . (h) First of all, 2 ln α −1] . E [θ̃]2 = E α2S = E eS2 ln α = MS (2 ln α) = enµ[e But e 2 ln α 2 =α = n−1 n 2 = (n − 1)2 n2 − 2n + 1 2 1 = = 1 − + 2. n2 n2 n n Therefore, E [θ̃]2 = e−2µ+(µ/n) . This and the computation of the expectation in (g) together tell us that h i µe−2µ Var(θ̃) = e−2µ+(µ/n) − e2µ = e−2µ eµ/n − 1 ≈ n as n → ∞, after another round of Taylor expansion. This and the answer to (b) together show that θ̃ achieves asymptotically [as n → ∞] the minimum possible variance among all unbiased estimators. Therefore, in a sense, θ̃ is “asymptotically UMVUE” for θ. 3