Homework solutions Math525 Fall 2003 Text book: Bickel-Doksum, 2nd edition Assignment # 6 Section 3.4. 1. E X̄ = θ. So X̄ is unbiased. p(x, θ) = θ x (1 − θ)1−x , x = 0, 1. log p(x, θ) = x log θ + (1 − x) log(1 − θ) and x 1−x ∂ log p(x, θ) = − ∂θ θ 1−θ hX i2 h∂ 1 − X i2 log p(X, θ) = E − ∂θ θ 1−θ h X2 hX 2i X(1 − X) (1 − X) 1−X i 1 1 1 =E 2 −2 + + = = E = + 2 2 2 θ θ(1 − θ) (1 − θ) θ (1 − θ) θ 1−θ θ(1 − θ) I1 (θ) = E IX~ (θ) = nI1 (θ) = n The information inequality lower bound is 1 θ(1 − θ) θ(1 − θ) 1 = IX~ (θ) n On the other hand, V ar(X̄) = θ(1 − θ) n So X̄ is UMVU estimate of θ. 2. n o R(θ, δ) = El θ, δ(X) = E E l θ, δ(X) |T (X) ≥ El θ, E δ(X)|T (X) = El θ, δ ∗ (X) = R(θ, δ ∗ ) 12. (a). f (x, θ) = θxθ−1 , 0 ≤ x ≤ 1. p(~x, θ) = θ n n Y i=1 xθ−1 i and log p(~x, θ) = n log θ + (θ − 1) Solve n ∂ n X log p(~x, θ) = + log xi = 0 ∂θ θ i=1 we have n 1X 1 =− log xi θ n i=1 1 n X i=1 log xi So the MLE of 1/θ is n ~ =− δ(X) 1X log Xi n i=1 Since ~ = −E log X = − Eδ(X) Z 1 0 log x · θxθ−1 dx = 1 θ ~ is unbiased. δ(X) n h∂ i2 i2 i2 hn X h1 ~ I(θ) = E log p(X, θ) = E + log Xi = nE + log X ∂θ θ i=1 θ Z n 1 n2 n 1o 1o n 1o (log x)2 · θxθ−1 dx − 2 = n 2 − 2 = 2 = n E(log X)2 − 2 = n θ θ θ θ θ 0 Write ψ(θ) = θ1 . The information inequality lower bound is [ψ 0 (θ)]2 1 = I(θ) nθ 2 On the other hand, n 1X 1 1 [ψ 0 (θ)]2 ~ V ar δ(X) = V ar − log Xi = V ar(log X) = = n n nθ 2 I(θ) i=1 So the information inequality lower bound is achieved. (b). E X̄ = EX = So X̄ is unbiased. Write ψ(θ) = lower bound is θ 1+θ . Z 1 0 x · θxθ−1 dx = Then ψ 0 (θ) = θ 1+θ 1 (1+θ)2 . So the information inequality θ2 [ψ 0 (θ)]2 = I(θ) n(1 + θ)4 On the other hand, Z θ 2 o θ 2 o 1n 1 1n 1 2 2 V ar(X̄) = V ar(X) = EX − = x · θxθ−1 dx − n n 1+θ n 1+θ 0 n o 2 θ 1 θ 1 θ = = − n 2+θ 1+θ n (2 + θ)(1 + θ)2 By the simple fact that θ 1 θ2 > n (2 + θ)(1 + θ)2 n(1 + θ)4 2 the information inequality lower bound is not achieved by X̄. Section 4.1 2. (a). By Problem B.3.4, 2λ n X i=1 d 2 Xi = X2n So the power β(λ) = Pλ {X̄ ≥ µ0 x(1 − α)/2n} n n X o n o 2 = Pλ 2λ Xi ≥ λµ0 x(1 − α) = P X2n ≥ λµ0 x(1 − α) i=1 When H is true, or λ ≥ 1/µ0 , n o 2 Pλ {X̄ ≥ µ0 x(1 − α)/2n} ≤ P X2n ≥ x(1 − α) = α and the equality holds as λ = 1/µ0 . So sup Pλ {X̄ ≥ µ0 x(1 − α)/2n} = α λ≥1/µ0 Hence, this test has level α. (b) the power is n o 2 beta(λ) = P X2n ≥ λµ0 x(1 − α) (c). Let Z1 , · · · , Z2n be i.i.d. N (0, 1) random variables. By Theorem B.3.1, p. 491, d 2 2 = Z12 + · · · + Z2n X2n As n is sufficiently large, n o n X 2 − 2nEZ 2 λµ0 x(1 − α) − 2nEZ 2 o 2 p β(λ) = P X2n ≥ λµ0 x(1 − α) = P p2n ≥ 2nV ar(Z 2 ) 2nV ar(Z 2 ) λµ x(1 − α) − 2nEZ 2 0 p ≈ 1−Φ 2nV ar(Z 2 ) where the approximation follows from the central limit theorem. Notice that EZ 2 = 1 and V ar(Z 2 ) = EZ 4 − 1 = 3 − 1 = 2. Therefore, √ n o λµ x(1 − α) − 2n λµ0 x(1 − α) 0 2 √ √ =Φ n− β(λ) = P X2n ≥ λµ0 x(1 − α) ≈ 1 − Φ 2 n 2 n 3 Take µ = 1/µ0 gives that λµ0 x(1 − α) − 2n √ ≈ z(1 − α) 2 n √ Or, x(1 − α) ≈ 2 nz(1 − α) + 2n. Therefore, √ √n(µ − µ ) µ z 0 0 + (1 − α) β(λ) ≈ Φ n(1 − λµ0 ) + λµ0 z(1 − α) = Φ µ µ as λ = 1/µ. 8. (a). For each x, |F̂ (x) − F0 (x)| ≤ Dn . Hence, |F̂ (x) − F0 (x)| ≥ kα ⊂ {Dn ≥ kα } Consequently, P |F̂ (x) − F0 (x)| ≤ kα ≤ P {Dn ≥ kα } Taking supremum over x on the left hand side gives sup P |F̂ (x) − F0 (x)| ≤ kα ≤ P {Dn ≥ kα } x (b). Omitted (c). By the fact that F 6= F0 , there is a x such that F (x) 6= F0 (x). As kα < |F (x) − F0 (x)|, by the law of large numbers, P |F̂ (x) − F0 (x)| ≤ kα −→ P |F (x) − F0 (x)| ≤ kα = 1 So the conclusion follows from Part (a). Section 4.2 2. (a). n Y f (xi , θ) = f (1, θ)N1 f (2, θ)N2 f (3, θ)N3 = f (1, θ)N1 f (2, θ)N2 f (3, θ)n−N1 −N2 i=1 = θ 2N1 2θ(1 − θ) Hence, L(~x, θ0 , θ1 ) = N2 n Y i=1 (1 − θ)2(n−N1 −N2 ) = 2N2 (1 − θ)2n f (xi , θ1 ) n .Y f (xi , θ0 ) = i=1 θ 2N1 +N2 1−θ 1 − θ 2n 1 − θ θ 2N1 +N2 1 0 1 1 − θ0 1 − θ 1 θ0 So the conclusion follows from the fact that 1 − θ 0 θ1 >1 1 − θ 1 θ0 (b) Write 1 − θ 2n 1 − θ θ c 0 1 1 k= 1 − θ0 1 − θ 1 θ0 Then L(~x, θ0 , θ1 ) ≥ k if and only if 2N1 + N2 ≥ c. The rest follows from Theorem 4.2.1-(a). 4