Homework solutions Math526 Spring 2004 Text book: Bickel-Doksum, 2nd edition Assignment # 2 Section 4.4. 1. (a). Pivot: n X i=1 Solving 2 (Xi − X̄)2 /σ 2 = (n − 1)s2 /σ 2 ∼ Xn−1 2 Xn−1 α 2 α 2 ≤ (n − 1)s2 /σ 2 ≤ Xn−1 1− 2 We have (1 − α)-confidence intertval for log σ 2 : (n − 1)s2 (n − 1)s2 , log log α 2 2 Xn−1 Xn−1 1 − α2 2 (b). The level (1 − α) UCB is n X . 16.52 2 2 (Xi − X̄)2 Xn−1 (α) = 16.52/Xn−1 (0.01) = = 2.49 6.63 i=1 3. Let σo2 be assumed value of σ 2 . We have ~ = X̄ − µ(X) z(1 − α) √ σo n Then n √n(X̄ − µ) o σ σo o ~ =P P {µ ≥ µ(X)} ≤ z(1 − α) = Tn−1 z(1 − α) s s s where Tn−1 is the distribution function of the student distribution with parameter n − 1. 4. (a).Since θ < 0.1, 1 − α = P {θ ∈ [θ, θ̄]} = P {θ ∈ [θ, θ̄], θ < 1} = P {θ ∈ [min{θ, 0.1}, min{θ̄, 0.1}]} So the confidence interval is [min{θ, 0.1}, min{θ̄, 0.1}]. (b). In the case of part (a), the length of the interval is min{θ̄, 0.1} − min{θ, 0.1} = 1 ( 0, θ ≥ 0.1 0.1 − θ θ < 0.1 < θ̄ θ̄ − θ θ, θ̄ < 0.1 Clearly, in the first and third cases, the length is automatically bounded by 0.2. We only need to deal with the case θ < 0.1 < θ̄ In this case we need 0.1 − Or S + (kα2 /2) − kα S + (kα2 /2) − kα p [S(n − S)/n] + kα2 /4 ≤ 0.2 n + kα2 p [S(n − S)/n] + kα2 /4 ≥ −0.1 n + kα2 Note that S ≥ 0 and that S(n − S)/n ≤ n/4. The above will be satisfied if p (kα2 /2) − kα n/4 + kα2 /4 ≥ −0.1 n + kα2 which is 0.2(n + kα2 ) − kα So we need p n + kα2 + kα2 ≥ 0 √ p 0.2 1 + kα n + kα2 ≥ 0.4 Or h 1 + √0.2 2 i − 1 kα2 0.4 Note that kα = z(0.975) = 1.96. So n ≥ 46.45. We take n = 47. This is much smaller than the one for (4.4.3), where n = 9, 601 (p.238). n≥ 17. (a) Write Ui = F (Xi ). Then U1 , · · · , Un are i.i.d. with the uniform distribution on [0, 1]. Notice that n n i=1 i=1 X 1X F̂ (x) = 1{Xi ≤x} = 1{Ui ≤F (x)} = Û F (x) n Hence, √ n|Û F (x) − F (x)| q An (F ) = sup ; a ≤ F (x) ≤ b F (x) 1 − F (x) √ n|Û (u) − u| u=F (x) q = sup ; a ≤ u ≤ b = An (U ) u 1−u Let uα be decided by P {An (F ) ≤ uα } = 1 − α. Solving An (F ) ≤ uα we have nF̂ (y) + u2α nF̂ (y) ≤ F (y) ≤ n + u2α n + u2α 2 if a ≤ f (y) ≤ b So, if you know that a ≤ f (x) ≤ b, then the (1 − α)-level confidence interval is h nF̂ (x) , 2 n + uα nF̂ (x) + u2α i ∩ [a, b] n + u2α (b). Bn (F ) = sup Let u = F (x). Then F̂ −1 √ n|Û F (x) − F (x)| −1 −1 q ; F F̂ (a) ≤ F (x) ≤ F F̂ (b) Û F (x) 1 − Û F (x) n o n (a) = inf x; F̂ (x) ≥ a = inf x; n o 1X 1{Ui ≤F (x)} ≥ a n i=1 n = inf x; n o 1X 1{Xi ≤x} ≥ a n i=1 Notice that F is continuous and strictly increasing. n F F̂ −1 (a) = inf F (x); n = inf y; n n o 1X 1{Ui ≤F (x)} ≥ a n i=1 o n o 1X 1{Ui ≤y} ≥ a = inf y; Û (y) ≥ a = Û −1 (a) n i=1 Similarly, F F̂ −1 (b) = Û −1 (b). Hence, under the substitution u = F (x), Bn (F ) = sup √ n|Û (u) − u| −1 −1 q ; Û (a) ≤ u ≤ Û (b) = Bn (U ) Û (u) 1 − Û (u) 3