Probability Qualifying Exam Solution, Spring 2010 Shiu-Tang Li September 6, 2012 1 Define f (x) := inf{k ≥ 0 : Φ(k) > x}. Note that since Φ is an increasing function and f is a decreasing function, {a > 0 : Φ(f (a)) > a} is a Borel set in [0, ∞). When a is chosen such that Φ(f (a)) > a, then {x : Φ(x) > a} = {x : x ≥ f (a)}. If Φ(f (a)) ≤ a, then {x : Φ(x) > a} = {x : x > f (a)}. ∞ Z P (Φ(X) > a) da E[Φ(X)] = Z0 Z ≤ P (X ≥ f (a)) da + P (X > f (a)) da [0,∞)∩{Φ(f (a))>a} [0,∞)∩{Φ(f (a))≤a} Z Z ≤ P (Y ≥ f (a)) da + [0,∞)∩{Φ(f (a))>a} Z ∞ ≤ P (Y > f (a)) da [0,∞)∩{Φ(f (a))≤a} P (Φ(Y ) > a) da = E[Φ(Y )] 0 2 (1)-1. Let {Xn }n∈N be a (sub)martingale such that E[|Xn |] is bounded over all n. Then, with probability one, the limit X = limn Xn exists and is finite. (1)-2. For a martingale {Xn , Fn }, TFAE: 1 (i) it is uniformly integrable. (ii) It converges to X 0 a.s. and in L1 , and Xn = E[X 0 |Fn ]. (iii) It converges to X 0 in L1 , and Xn = E[X 0 |Fn ]. (iv) There is some X ∈ L1 s.t. Xn = E[X|Fn ]. (Note that the X 0 in (ii) and (iii) is the same one but is different from the X in (iv).) (2) For each n, E[E[Z|Fn ]|Fn−1 ] = E[Z|Fn−1 ], by the definition of conditional expectation, and hence {E[Z|Fn ], Fn } is a martingale. We observe that W∞ X = E[Z|F∞ ] satisfies E[Z|Fn ] = E[E[Z|F∞ ]|Fn ], where F∞ := n=1 Fn , so by Doob’s martingale convergence theorem E[Z|Fn ] → Y a.s. and in L1 for some Y , and E[Y |Fn ] = E[Z|Fn ] by the statements in martingale convergence theorem. We claim that Y = E[Z|F∞ ]. R R S∞ F , we have Y dP = ZdP . The set {B : For any A ∈ n n=1 A A R R Y dP = ZdP } is a λ-system (The monotone class property follows B B S∞ from LDCT), which contains n=1 Fn , Rwhich is a Rπ-system that generates F∞ . Thus by π − λ theorem we have A Y dP = A ZdP for A ∈ F∞ , so Y = E[Z|Fn ]. 3 First note that a + enough, we have 1 λ ln n > 0 for all n large. Therefore, when n is large P (max(X1 , · · · , Xn ) ≤ a + λa 1 1 ln n) = (1 − eλ(a+ λ ln n) )n λ e−λa n = (1 − ) n λa = e−e as n → ∞ Let F (a) = 1 − e−e . It is the distribution function of some r.v. X and we conclude that max(X1 , · · · , Xn ) − λ1 ln n ⇒ X. 2 4 5 6 (a) P (N = k; N1 < N2 ) = P (N1 = k; N2 > k) = P (X1 = 3, · · · , Xk−1 = 3, Xk = 1) − P (X1 = 3, · · · , Xk−1 = 3, Xk = 1; Xn 6= 2 for all n > k) = P (X1 = 3, · · · , Xk−1 = 3, Xk = 1) = pk−1 3 p1 2 = pk−1 3 (p1 + p2 ) · (p1 + p1 p3 + p1 p3 + · · · ) = P (X1 = 3, · · · , Xk−1 = 3, Xk 6= 3) ∞ X P (X1 = 3, · · · Xn = 3, Xn+1 = 1) × n=0 = P (N = k)P (N1 < N2 ) (b) E[N ; N1 < N2 ] P (N1 < N2 ) E[N ]P (N1 < N2 ) = P (N1 < N2 ) ∞ X nP (N = n) = E[N1 |N1 < N2 ] = n=1 = ∞ X np3n−1 (p1 + p2 ) = n=1 1 1 − p3 (c)Let N1 be the numer of trials needed for 6 to appear, N2 be the numer of trials needed for 7 to appear, N3 be the numer of trials needed for any number other that 6 or 7 to appear. By (b), the conditional expected numer 1 of rolls is 11/36 = 36/11. 7 it (a) The characteristic function φ(t)√ of X ∼ Poisson(λ) is eλ(e −1) . The √ p it/ λ −1) −iλ(t/ λ) ch.f. of (X − EX)/ V ar(X) is eλ(e e . We observe that 3 √ it/ λ lim λ(e λ→∞ √ it/ λ √ eit/λ − 1 − it/λ − 1 − it/ λ) = lim λ→∞ 1/λ2 −iteit/λ /λ2 + it/λ2 = lim λ→∞ −2/λ3 −iteit/λ + it = lim λ→∞ −2/λ (it)2 eit/λ /λ2 + it = lim λ→∞ 2/λ2 −t2 = 2 √ 2 /2 −1) −iλ(t/ λ) As a result, eλ(e e → e−t shown the desired weak convergence. while λ → ∞, and we have 8 X n 2 2 1 Xi Xj − µ )] E[ n ( 2 2 1≤i<j≤n 2 X X X n 2 n 4 X X ] − 2E[ X X ] µ + µ4 ) (E[ X X ]E[ = 2 k l i j i j 2 2 n (n − 1)2 1≤i<j≤n 1≤i<j≤n 1≤k<l≤n X X 4 = 2 (6E[ X X X X ] + 2E[ Xi2 Xj Xk ] i j k l 2 n (n − 1) 1≤i<j<k<l≤n 1≤i,j,k<l≤n,i6=j,j6=k,k6=i 2 X X n 2 n 2 2 +E[ Xi Xj ] − 2E[ Xi Xj ] µ + µ4 ) 2 2 1≤i<j≤n 1≤i<j≤n 2 2 n 2 2 n n n 4 n 4 2 2 2 2 4 = 2 (6 µ + 2 · 3 · µ (σ + ν ) + (σ + ν ) − 2 µ + µ4 ) 2 n (n − 1) 4 3 2 2 2 → 0 as n → ∞ We have proved the L2 convergence of it and hence convergence in pr. 4 9 Let B, B 0 ∈ B(R). Z 1{X1 ∈B 0 } dP = P (X1 ∈ B 0 , X2 + · · · + Xn ∈ B − B 0 ) {Sn ∈B} = P (Xk ∈ B 0 , Sn − Xk ∈ B − B 0 ) Z 1{Xk ∈B 0 } dP = {Sn ∈B} We thus have m m2 X Z {Sn ∈B} j=1 j j−1 j 1 dP = 2m {X1 ∈[ 2m , 2m )} m m2 X Z {Sn ∈B} j=1 j j−1 j 1 dP 2m {Xk ∈[ 2m , 2m )} and let m → ∞ using the fact that E[X1 ] < ∞, we have Z Z + X1 dP = Xk+ dP. {Sn ∈B} {Sn ∈B} Similarly, we have Z X1− Z dP = {Sn ∈B} {Sn ∈B} Xk− dP For any 1 ≤ k ≤ n. Therefore, R {Sn ∈B} Xk dP = R {Sn ∈B} X1 +···+Xn n dP , and thus E[Xk |Sn ] = Sn . n (b) Let Fk = σ(Sn , · · · , Sn−k+1 ). Sn−k+1 |σ(Sn , · · · , Sn−k+2 )] n−k+1 1 = E[Sn−k+1 |σ(Sn−k+2 , Xn−k+3 , · · · , Xn )] n−k+1 1 = E[Sn−k+1 |σ(Sn−k+2 )] Since Xi0 s are independent. n−k+1 1 Sn−k+2 = (n − k + 1) = Mk−1 . n−k+1 n−k+2 E[Mk |Fk−1 ] = E[ Remarks. If σ(X, Y ) is independent if Z, then E[X|Y, Z] = E[X|Y ]. To 5 see this, Z Z E[X|Y, Z] dP = X dP {Y ∈B1 ,Z∈B2 } {Y ∈B1 ,Z∈B2 } Z = P (Z ∈ B2 ) X dP {Y ∈B1 } Z E[X|Y ] dP = P (Z ∈ B2 ) {Y ∈B1 } Z E[X|Y ]1{Z∈B2 } dP = {Y ∈B1 } Z E[X|Y ] dP. = {Y ∈B1 ,Z∈B2 } 10 Let a, b > 0, 0 < x < 1. Let p = 1 ,q 1−x = x1 , we have p + q = 1. (1 − x) ln E|X|a + x ln E|X|b = ln(E[|X|a ])1−x E[|X|b ]x ≥ ln E[|X|a(1−x) |X|bx ] = ln E[|X|a(1−x)+bx ] 6