Final – Math 5080 – Spring 2004 Name: Instructions. READ CAREFULLY.: (i) The work you turn in must be your own. You may not discuss the final with anyone, either in the class or outside the class. [You may of course consult with me for clarification of any of the problems.] Failure to follow this policy will be considered cheating and will result in a course grade of E. (ii) You may consult the textbook and your notes. In particular, feel free to use any of the information in the tables of distributions in Appendix B, for example, the moment generating functions for specific distributions. You can use a general mathematical reference, for example a calculus text or a table of integrals. You may not use any statistical textbook or written source material concerning the specific subject matter of the course. Failure to follow this policy will be considered cheating and will result in a course grade of E. (iii) Your final must be clearly written and legible. I will not grade problems which are sloppily presented and such problems will receive a grade of 0. If you are unable to write legibly and clearly, use of a word processor. You have at least 7 days to complete the final; budget time for writing up your solutions. (iv) Think about your exposition. Someone (me) has to read what you have written. Your answer is only correct if I can understand what you have done. Style matters. (v) Finals are due at 6 PM on Wednesday May 5, 2004. Late finals will not be accepted, except for reasons of death or serious illness. It is highly recommended that you turn in your exam to me personally. I will be in my office (LCB 209) at 6 PM May 5 2004 to accept exams then, and will check my box in the math department at that time. I cannot guarantee that exams left in my box will be received. 1 Sign here to indicate you have read and understand these instructions. 2 Problem 1. For each n, let Xn,1 , Xn,2 , . . . , Xn,n be independent Bernoulli random variables. That is, Xn,i 1 with probability pn,i , = 0 with probability 1 − p . n,i Note that the pn,i are not assumed to be identical. Suppose that lim n→∞ n X pn,i = µ , and lim n→∞ i=1 Find a random variable Y so that Sn = Pn i=1 n X p2n,i = 0 . Xn,i converges in distribution to Y . Prove your answer. Solution. Let Mn,i (t) be the mgf for Xn,i : Mn,i (t) = E etXn,i = et pn,i + (1 − pn,i ) = 1 + pn,i et − 1 . If Mn (t) is the mgf for Sn , we have Mn (t) = n Y Mn,i (t) i=1 n Y = 1 + pn,i (et − 1) log Mn (t) = i=1 n X log 1 + pn,i (et − 1) . i=1 Now log(1 + x) = x + ε(x), where |ε(x)| ≤ x2 . Thus n X Mn (t) = pn,i (et − 1) + ε pn,i (et − 1) i=1 = (et − 1) n X pn,i + i=1 n X i=1 3 (1) i=1 ε pn,i (et − 1) . Now, n n X X t ε pn,i (et − 1) ε pn,i (e − 1) ≤ i=1 ≤ i=1 n X 2 p2n,i et − 1 i=1 n 2 X = e −1 p2n,i . t i=1 We conclude from the hypotheses (1) that log Mn (t) → µ(et − 1) . Thus, we conclude that Sn converges in distribution to a Poisson random variable with mean µ. Problem 2. Let X1 , X2 , . . . , Xn be a sequence of i.i.d. random variables with density 2θ−2 x if 0 ≤ x ≤ θ f (x; θ) = 0 otherwise. (i) Find the MLE for θ. (ii) Compute the bias of the MLE for θ. Solution. The likelihood function is n Y 2 L(θ; x) = 1 {0 ≤ xi ≤ θ} θ2 i=1 2n = 2n 1 x(n) ≤ θ . θ By graphing this function, we see it is largest at θ = x(n) , and so the MLE is θ̂ = X(n) . We need to find the distribution of X(n) : For 0 ≤ x ≤ θ P X(n) ≤ x = F (x)n x2n = 2n , θ 4 and so fX(n) (x) = 2nx2n−1 θ2n for x ∈ [0, θ]. Then Zθ E X(n) = θ 2nx2n−1 2nx2n+1 2n θ x dx = = 2n 2n θ (2n + 1)θ 2n + 1 0 0 Thus β =θ− 2n θ θ= . 2n + 1 2n + 1 Problem 3. Let X and Y be two independent random variables with respective density functions f (x) = 1 2 if 0 ≤ x ≤ 2 0 otherwise. e−x if 0 ≤ x < ∞ g(x) = 0 otherwise. Compute the density of X + Y . Solution. Let S = X + Y . Using the convolution formula, for s ≥ 0, and letting a ∧ b = 5 min{a, b}, Z∞ fX (x)fY (s − x)dx fS (s) = −∞ Z∞ = 1 1 {0 ≤ x ≤ 2} e−(s−x) 1 {0 ≤ s − x} dx 2 −∞ Z2∧s 1 −(s−x) e dx 2 0 2∧s 1 −(s−x) = e 2 = 0 1 1 = e−(s−2∧s) − e−s . 2 2 In other words, 1 (1 − e−s ) if 0 ≤ s ≤ 2 , 2 fS (s) = 12 e−(s−2) − e−s if 2 ≤ 2 < ∞ , 0 if s < 0 . Check that it integrates to one: Z∞ 0 Z∞ −(s−2) e − e−s 1 − e−s ds + ds fS (s)ds = 2 2 2 0 ∞ −s 2 −(s−2) s+e −e + e−s = + 2 0 2 2 1 + e−2 1 − e−2 = + 2 2 = 1. Z2 6 Problem 4. Let X1 and X2 be independent random variables with the density function e−x if 0 ≤ x < ∞ f (x) = 0 otherwise. Compute the joint density function of Y1 = X1 + X2 and Y2 = 2X1 − X2 . Solution. Let g(x1 , x2 ) = (x1 + x2 , 2x1 − x2 ). We have " # " #" # y1 1 1 x1 = , y2 2 −1 x2 and since " 1 1 #−1 2 −1 We have g −1 (y1 , y2 ) = ( y31 + y2 2y1 , 3 3 " # " # 1 1 1 −1 −1 = 32 3 1 = −3 −2 1 −3 3 − y2 ), 3 and " Dg −1 (y1 , y2 ) = 1 3 2 3 1 3 # − 31 1 2 1 |Jg−1 (y1 , y2 )| = − = . 9 9 3 Thus fY1 ,Y2 (y1 , y2 ) = f (g1−1 (y1 , y2 ))f (g2−1 (y1 , y2 ))|Jg−1 (y1 , y2 )| y1 y2 = e− 3 − 3 1 {y1 + y2 ≥ 0} e− = 2y1 y + 32 3 1 {2y1 − y2 ≥ 0} 1 3 e−y1 1 {−y1 ≤ y2 ≤ 2y1 } . 3 Note that if you don’t indicate the region where the density is positive, your answer is wrong. In particular, it is not the case that fY1 ,Y2 (y1 , y2 ) = 31 e−y1 , as that does not integrate to one (it does not even integrate to a finite number!) 7 Problem 5. Let X1 , X2 , . . . , X6 be independent random variables. We assume that X1 and X2 are Normal(µ = 0, σ 2 = 2), while X3 , X4 , X5 , X6 are Normal(µ = 0, σ 2 = 4). Determine c such that P X1 +X2 is 2 2 χ4 . Thus Solution. Notice that 3, . . . , 6 so P6 2 i=3 Xi 4 is ! X1 + X2 p X32 + . . . + X62 ≤c = 0.9 . a Normal(0, 1) random variable. Also Xi2 /4 is χ21 for i = X1 +X2 q P6 2 2 i=3 Xi 4 X1 + X2 = 2 qP 6 2 /4 i=3 Xi has a t4 distribution. The 90the percentile of the t4 distribution is 1.533. Thus since ! ! X1 + X2 X1 + X2 P p 2 ≤ c = P 2p 2 ≤ 2c = P (t4 ≤ 2c) , X3 + . . . + X62 X3 + . . . + X62 Setting 2c = 1.533 makes the probability equal to 0.9 Thus c = 0.7665. Problem 6. Let X1 , X2 , . . . , X120 be independent and identically distributed random variables with density function 3x2 f (x) = 0 if 0 ≤ x ≤ 1 otherwise. Get an approximate value of c so that P (X1 + · · · + X120 ≤ c) = 0.99 . Solution. We have Z1 E(X1 ) = 0 1 3 4 3 x3x dx = x = , 4 0 4 2 8 and E X12 Z1 = 0 1 3 3 5 x 3x dx = x = . 5 0 5 2 2 Thus Var (X1 ) = Let Sn = P120 i=1 3 9 3 − = 5 16 80 Xi . Then 3 E(Sn ) = 120 = 90 , 4 r SD (Sn ) = 120 3 3 =√ 80 2 Thus Sn − 90 P (X1 + · · · + X120 ≤ c) = P √3 2 ≤ c − 90 ! √3 2 ≈ Φ(z) , where z = c−90 3 √ 2 . The 90th percentile of the standard normal distribution is 3.26. Thus solving 2.326 = c − 90 √3 2 for c gives c = 94.9342. Problem 7. Let X1 , X2 , . . . , Xn be i.i.d. Geometric(p) random variables: p(1 − p)x−1 if x = 1, 2, . . . , P (Xi = x; p) = 0 otherwise. (i) Find the MLE of p. (ii) Find the UMVU estimator of 1/p. [ Justify your answer. ] 9 Solution. We have f (x; p) = n Y P p(1 − p)xi −1 = pn (1 − p) xi −n i=1 X = exp n log p + (1 − p) xi − n log(1 − p) X p . = exp (1 − p) xi + n log 1−p P P Thus this is a regular exponential class, with sufficient statistic xi . It follows that xi is a complete minimal sufficient statistic. We have X `(p) = log(1 − p) xi + n log P ∂` xi n =− + . ∂p 1 − p p(1 − p) Setting ∂` ∂p p 1−p = 0 and solving for p gives X 0 = −p xi + n n p= P xi 1 p= . x̄ It can be checked that this is indeed a global maximum. Thus p̂ = 1/X̄. By the invariance d = X̄. of the MLE, we have 1/p It is clear that E X̄ = p1 , so the MLE is unbiased. Since X̄ is a function of the complete minimum sufficient statistic, it is UMVU (by Lehmann-Scheffe Theorem). Problem 8. Let (X1 , Y1 ), (X2 , Y2 ), . . . , (Xn , Yn ) be i.i.d. random vectors with the following density: f (x, y; ρ) = 1 πρ2 0 if (x, y) is in the disc centered at 0 with radius ρ , otherwise. Find the MLE of ρ. 10 Solution. The likelihood is q n Y 1 2 2 Xi + Yi ≤ ρ L(ρ) = 1 2 πρ i=1 q 1 2 2 = n 2n 1 max Xi + Yi ≤ ρ i π ρ Thus the MLE is q ρ̂ = max Xi2 + Yi2 . i Note that max i q Xi2 + Yi2 6= q 2 2 X(n) + Y(n) . [Try the two points (2, 1) and (1, 2): In this case x(2) = y(2) = 2, so p √ while max x2i + yi2 = 5.] 11 q 2 x2(2) + y(2) = √ 8,