Stat 643 Spring 2010 Assignment 04 31. / (T , F) (X , B) II II φ II ψ II I$ (R, B) T i. If φ(x) is an indicator function, that is, φ(x) = IA (x) for some A ∈ B. Define ψ(y) = IT (A) (y), which is (F, B)measurable. Then (ψ ◦ T )(x) = IT (A) (T (x)) = I{T (x) ∈ T (A)} = I{x ∈ A} = IA (x) = φ(x). Therefore, the result holds when φ is an indicator function. Pn n ii. If φ(x) = i=1 ai IAi (x), suchP that φi ◦T (x) = IAi (x), 1 ≤ Pn then by i., there exist {ψi }i=1 who are (F, B)-measurable Pn n i ≤ n. Then let ψ = i=1 ai ψi . ψ is (F, B)-measurable and ψ ◦ T = i=1 ai ψi ◦ T = i=1 ai IAi = φ. iii. For a general measurable function φ, there exist a sequence of simple functions {φi }i≥1 which converges to φ (One can think about φ = φ+ − φ− ). Then correspondingly, there exist a sequence of (F, B)-measurable function {ψi }i≥1 such that φi = ψi ◦ T . Let ψ = limn→∞ ψn (It is well-defined!). Then ψ ◦ T = φ and ψ is still (F, B)-measurable. In conclusion, for and measurable function φ, there exists a (F, B)-measurable function ψ such that φ = ψ ◦ T . 32. Let λ = m + ν, where m is the Lebesgue measure and ν is the counting measure on N = {0, 1, 2, 3, 4, . . .}. Adopt the notation P X = P ◦ X −1 for any random variable X. First of all, P X1 = pP W1 + (1 − p)P Z1 . One could think of X1 = R1 W1 + (1 − R1 )Z1 , where R1 is independent of (W1 , Z1 ) and follows Bernoulli(p). Secondly, we claim that 1 e−µ µx dP X1 (x) = p √ exp{−(x − µ)2 /2}I{x ∈ / N} + (1 − p) I{x ∈ N}, dλ x! 2π a.e.(λ). This is because, for any A ∈ B(R), P X1 (A) = = = pP W1 (A) + (1 − p)P Z1 (A) Z Z e−µ µx 1 dν p √ exp{−(x − µ)2 /2}dm + (1 − p) x! 2π A A Z Z 1 e−µ µx p √ exp{−(x − µ)2 /2}I{x ∈ / N}d(m + ν) + (1 − p) I{x ∈ N}d(m + ν). x! 2π A A Now denote δi = I{Xi ∈ N}. Then ( ) δi n −µ x 1−δi Y 1 e µ p √ exp{−(x − µ)2 /2} (Xi ) = (1 − p) dλ x! 2π i=1 ( n ) n X X Pn 2 = exp µ δi Xi + [µ − µ /2 + log{p/(1 − p)}] δi µ i=1 (1−δi )Xi n Y dP Xi i=1 · exp{− i=1 n X i=1 δi Xi2 /2} i=1 n Y (Xi !)δi −1 · (1 − p)n i=1 By Factorization theorem, one sufficient statistic, which satisfies both (a) and (b), is ( n X i=1 δi , n X δi Xi , i=1 n X i=1 1 (1 − δi )Xi ). 33. The Pθ in this context really means PθX , that is, PθX ({x}) = (1 − p)px−θ I{x − θ ∈ N}, where N shares the same definition as in Problem 32. (a) Suppose Pθ can be dominated by some σ-finite measure µ, then Pθ must be dominated by a probability measure λ (Lemma 52). As PθX ({θ}) = 1−p > 0 for all θ, this means λ({θ}) > 0 for all θ. This is not possible for a probability measure λ. Specifically, let D0 = {x ∈ R | λ({x}) > 1}, Dn = {x ∈ R | 1/(n + 1) < λ({x}) ≤ 1/n}, n ≥ 1. Then R = ∪n≥0 Dn . This means, there exists uncountable set Dn0 . From the construction of Dn0 we know that λ(Dn0 ) = ∞ as it contains infinitely many points. (b) Consider PP (X1 = x1 , . . . , Xn = xn |X(1) = t1 , Sn = t2 ), where X(1) is the minimum order statistic of X1 , . . . , Xn n and Sn = i=1 Xi . We would assume that t1 > θ, since otherwise Pn the conditioning set is empty. Secondly, this probability would be 0, if (x1 , . . . , xn ) ∈ / E(t1 , t2 ) = {(k1 , . . . , kn )| i=1 ki = t2 , min1≤i≤n ki = t1 , ki − θ ∈ N}. Thus P (X1 = x1 , . . . , Xn = xn , X(1) = t1 , Sn = t2 ) = (1 − p)n pt2 −nθ I{(x1 , . . . , xn ) ∈ E(t1 , t2 )}. Furthermore, P (X(1) = P P t1 , Sn = t2 ) = (k1 ,...,kn )∈E(t1 ,t2 ) P (X1 = k1 , . . . , Xn = kn ) = (k1 ,...,kn )∈E(t1 ,t2 ) (1 − p)n pt2 −nθ . Consequently, P (X1 = x1 , . . . , Xn = xn |X(1) = Ptn1 , Sn = t2 ) = 1/|E(t1 , t2 )|, where |E(t1 , t2 )| is the size of E(t1 , t2 ). Notice further that E(t , t ) = {(k , . . . , k ) | 1 2 1 n i=1 (ki −t1 ) = t2 −nt1 , ki −t1 ∈ N, min1≤i≤n (ki −t1 ) = 0} = {(r1 +t1 , . . . , rn +tn ) | Pn r = t − nt , r ∈ N, min ri = 0}, which implies that it is independent of θ and only depends on t1 , t2 . i 2 1 i 1≤i≤n i=1 Thus (X(1) , Sn ) is sufficient for {Pθ }θ∈Θ . (c) When θ is known, Θ = (0, 1) and PθX is dominated by counting measure ν, with R-N derivative dPθX (x) = (1 − p)px−θ I{x − θ ∈ N}, a.e.(ν). dν By factorization theorem, P i Xi would be a sufficient statistic. 34. First we claim that σ(T ) = {C ∈ B k | π(C) = C, ∀π}. This is because permutation on x ∈ Rk would not change the order of it, that is, T (π(x)) = T (x), ∀π. This suggests that the collection of σ(T )-measurable functions is {f : Rk → R | f is a measurable function and f ◦ π = f, ∀π} P 0 Now consider Y (x) = σ(T )-measurable as Y ◦ π Iπ(B) (x)/k!, where x = (x1 , . . . , xk ) .P First of all, Y (X) is P π(x) = Y (x), ∀π. Secondly, for any C ∈ σ(T ), E(I Y (X)) = P (π(B) ∩ C)/k! = B π π P (π(B) ∩ π(C)/k! = P P P (π(B ∩ C))/k! = P (B ∩ C)/k! = P (B ∩ C) = E(I I ). Therefore, P (B|σ(T )) = Y (X), a.s.(P ). Note that B C π π P Y (x) = π IB (π(x))/k! as well. 2