Probability Qualifying Exam Solution, Spring 2009 Shiu-Tang Li December 27, 2012 1 (1) E[SN ] = E[E[SN |N ]] ∞ X = E[ ]1{N =i} E[Si ]] i=0 ∞ X = E[ ]1{N =i} iµ] i=0 ∞ X = µE[ ]1{N =i} i] = µE[N ] i=0 (2) 2 V ar[SN ] = E[E[SN |N ]] − (ESN )2 ∞ X = E[ 1{N =i} E[Si2 ]] − (µE[N ])2 i=0 ∞ X = E[ 1{N =i} iσ 2 + i(i − 1)µ2 ] − µ2 (E[N ])2 i=0 2 = σ E[N ] + µ2 E[N 2 ] − E[N ] − µ2 (E[N ])2 = (σ 2 − µ2 )E[N ] + µ2 V ar[N ] 1 2 (a) n X 1 n k 1 E[ ]= p (1 − p)n−k · X +1 k+1 k k=0 n X 1 (n + 1)! = pk+1 (1 − p)n−k (n + 1)p k=0 (k + 1)!(n − k)! n+1 X 1 (n + 1)! = pk+1 (1 − p)n−k − pn+1 (n + 1)p k=0 (k + 1)!(n − k)! = 1 (1 − pn+1 ) (n + 1)p (b) ∞ ∞ X λk e−λ X λk+1 e−λ 1 1 E[ ]= · = X +1 k! k + 1 k=0 (k + 1)! k=0 ∞ 1 X λk+1 e−λ = λ k=0 (k + 1)! = ∞ 1 X λk e−λ − e−λ λ k=0 k! = 1 (1 − e−λ ) λ 3 (a)Given any A > 0, Z ∞ P (|Xn | > A) = 2 A n dx π(1 + n2 x2 ) 2 = tan−1 (nx)|∞ A π 2 π = ( − tan−1 (nA)) → 0 as n → ∞ π 2 Thus Xn → 0 in pr. 2 (b)E[|Xn − 0|] = E[|Xn |] = Xn 9 0 in L1 . 2 π R∞ 0 nx 1+n2 x2 = 1 nπ ln(1 + n2 x2 )|∞ 0 = ∞. Thus (c)Given A > 0, ∞ X ∞ 2X π ( − tan−1 (nA)) π n=1 2 Z ∞ 2 π π ≥ − tan−1 (x) dx − π 0 2 2 Z π 2 2 π = tan(x) dx − π 0 2 2 π π/2 = ln(sec x)|0 − =∞ π 2 P (|Xn − 0| > A) = n=1 By Borel-Cantelli lemma, P (|Xn − 0| > A i.o.) = 1. This implies the failure of a.s. convergence. 4 Hard. 5 S S S n−1 n−1 n n (a)E[Mn |Fn−1 ] = λ0 n−1 E[λX E[λX . 0 |Fn−1 ] = λ0 0 ] = λ0 1 (b)Since E[λX 0 ] = 1, we let λ0 = a(cos θ + i sin θ) and we have pa cos θ + (1 − p)a−1 cos θ = 1 (1) and pa sin θ − (1 − p)a−1 sin θ = 0 (2) q q or sin θ = 0. If a = 1−p , then by (1) we by (2), we have either a = 1−p p p have cos θ = √ 1 ≥ 1, and this implies cos θ = 1 and p = 21 ⇒a = 1. 2 p(1−p) If we start at the assumption that sin θ = 0, by (1) we have cos θ = 1 to make pa cos θ + (1 − p)a−1 cos θ positive. Hence pa2 − a + (1 − p) = 0 and we have a = 1 or a = 1−p . Therefore, λ0 equals 1 or 1−p , and for any n we have p p E[|Mn |] = E[|X1 |]n = p|λ| + (1 − p)|λ−1 | = pλ + (1 − p)λ−1 = 1 3 which shows {Mn } is L1 bounded; martingale convergence theorem tells us the limit exists and is finite a.s. (c)Let X1 takes value on 1 and 3 with probability 12 each, and λ0 = aeiθ . We have 1 1 a cos θ + a3 (4 cos3 θ − 3 cos θ) = 1 (3) 2 2 and 1 1 a sin θ + a3 (3 sin θ − 4 sin3 θ) = 0 (4) 2 2 By (4) we have sin2 θ = 34 + 4a12 = 1 − cosθ . Substitute 41 − 4a12 for cosθ in (3) we have cos θ = −1 . Therefore we have the equation ( −1 )2 = 14 − 4a12 , √ a3 √ a3 which solves to a = 2 since a > 0. Therefore, cos θ = −4 2 , and we choose √ sin θ = 414 . √ √ 1 1 1 We find that E[λX ] = 1 and P (|λX | = 2) = P (|λX | = 2 2) = 12 . 0 0 0 √ Thus |Mn | ≥ ( 2)n a.s., showing that the limit does not exist for a.s. ω. 6 Given i ∈ N, we may choose Ai s.t. P (|X| > Ai ) < 1/i and FX (x) is continuous at both Ai and −Ai , and Ai → ∞ as i → ∞. Therefore, Z Z xdµ = lim {x>0} xdµ by monotone convergence theorem Z Z Z ≤ lim sup | xdµ − xdµn | + lim sup | xdµn | i→∞ i→∞ {0<x≤Ai } {0<x≤Ai } {0<x≤Ai } Z Z ≤ lim sup | xdµ − xdµn | + (sup E[Xn2 ])1/2 i→∞ {0<x≤Ai } i→∞ ≤1+ {0<x≤Ai } (sup E[Xn2 ])1/2 , n≥1 {0<x≤Ai } n≥1 R R if we select n = n(i) above to satisfy | {0<x≤Ai } xdµ − {0<x≤Ai } xdµn | ≤ 1 for each i. This can be done because Xn ⇒ X implies E[f (Xn )] → E[f (X)] forRany bounded continuous function. Similarly we can prove the finiteness of {x≤0} xdµ. This proves X ∈ L1 . For the rest part, 4 Z Z x dµn (x) − x dµ(x) ≤ Z x dµn (x) + |x|>Ai Z Z x dµn (x) − + |x|≤Ai Z x dµ(x) |x|>Ai x dµ(x) |x|≤Ai 1 ≤ ( · sup E[Xn2 ])1/2 + E[X; |X| > Ai ] i n≥1 Z Z x dµn (x) − x dµ(x). + |x|≤Ai |x|≤Ai Now we let n → ∞ to obtain 1 lim sup |EXn − EX| ≤ ( · sup E[Xn2 ])1/2 + E[X; |X| > Ai ] i n≥1 n And finally we let i → ∞ to have RHS → 0 and get the desired result. 7 First we show that we still need an extra condition so that the weak convergence works. We compute −1/2 E[ n n X 4 1 X (X1 + · · · + Xn ) ] = 2 E[Xi4 ] + 2 E[Xi2 ]E[Xj2 ] n i=1 1≤i<j≤n ≤ 1 (n + n(n − 1)) = 1. n2 Use similar techniques as we did in the previous problem using the fact that {Xn } has bounded fourth moments, we claim that E[ n−1/2 (X1 + · · · + 2 σ). Xn ) ] → E[N (0, σ)2 ], if we assume that n−1/2 (X1 + · · · + Xn ) → N P(0, n 1 2 As a result, we have to impose an additional condition: limn→∞ 3n i=1 ai exists and is finite. This value is actually σ 2 . Now it’s time for applying Lindeberg-Feller theorem, which states as follows (See Durrett’s book): For each n, let Xn,m , 1 ≤ m ≤ n be independent random variables with E[Xn,m ] = 0. Suppose 5 P 2 (i) nm=1 E[Xn,m ] → σ 2 >P 0; 2 (ii)For all > 0, limn→∞ nm=1 E[Xn,m ; |Xn,m | > ] = 0. Then Sn = Xn,1 + · · · + Xn,n ⇒ N (0, σ 2 ) as n → ∞. −1/2 Let Xn,m Xm . The condition (i) is satisfied since we assume Pn= n2 1 limn→∞ 3n i=1 ai = σ 2 exists and is finite. For (ii), just note that P (|Xn,m | > Thus by ) = P (|Xm | > n1/2 ) = 0 when n > 12 for every 1 ≤ m ≤ n. P 1 −1/2 Lindeberg-Feller theorem, n (X1 + · · · + Xn ) → N (0, limn→∞ 3n ni=1 a2i ). 8 R∞ R∞ (a)The fact that g ≥ 0 a.e. in R2 is obvious. It is left to show −∞ −∞ g(x, y) dx dy = 1. To see this, Z ∞Z y Z ∞Z x 2f (x)f (y) dx dy = 2f (y)f (x) dy dx −∞ −∞ −∞ −∞ Z ∞Z ∞ = 2f (y)f (x) dx dy −∞ y Z ∞Z ∞ Z Z 1 ∞ y 2f (y)f (x) dx dy) 2f (x)f (y) dx dy + = ( 2 −∞ −∞ −∞ y Z ∞Z ∞ 1 = ·2 f (x)f (y) dx dy = 1 2 −∞ −∞ Rt R∞ (b) First we choose t ∈ R so that −∞ f (x) dx > 0 and t f (x) dx > 0. R∞ Rt We have P (X ≤ t, Y > t) = 0, but both P (X ≤ t) = −∞ −∞ 2f (x)f (y) dx dy = Rt R∞ 2 −∞ f (x) dx > 0 and P (Y > t) = 2 y f (y) dy > 0. This shows X and Y are not independent. 9 P (n−1/α max(X1 , · · · , Xn ) ≤ x) = P (max(X1 , · · · , Xn ) ≤ n1/α x) = P (1 − (n1/α x)−α )n x−α n −α = P (1 − ) → e−x n −α as n → ∞. We may define F (x) = e−x , and it is continuous and nondecreasing on R, with F (∞) = 1, F (−∞) = 0. Thus it is the distribution −α function of some random variable X. So P (X > x) = 1 − F (x) = 1 − e−x . 6 10 (a) ∞ λα α−1 −λx itx x e e dx Γ(α) 0 Z ∞ α ∞ X λ (itx)n α−1 −λx = x e dx Γ(α) n! 0 n=0 Z ∞ X (it)n λα Γ(n + α) ∞ λn+α = · · xn+α−1 e−λx dx n+α n! Γ(α) λ Γ(n + α) 0 n=0 ∞ X α it it = (1 + )α . = ( )n n λ λ n=0 Z f (t) = (b) fX1 +···+Xn (t) = fX1 (t)n = (1 + itλ )n . That is, it is Gamma(n, λ). (c) We have: Sn + Xn+1 > k, Sn < k. Therefore, the probability is Z k 0 Z = 0 Z ∞ λe−λx k−s k λn sn−1 e−λs ds (n − 1)! Z k −λk e sn−1 ds e−λ(k−s) λn (n − 1)! (λk)n −λk = e . n! = λn sn−1 e−λs dx ds (n − 1)! 0 7