STAT 333 Assignment 1 SOLUTIONS

1. Consider 10 independent coin flips having probability p of landing heads. a. Find the expected number of changeovers. Let Xi = 1 if there is a changeover between trial i and trial i+1, 0 otherwise, i = 1, 2, , 9. Then X, the number of changeovers is X1 + X2 + + X9 E[Xi] = P(Changeover) = P(HT or TH) = p(1 p) + (1 p)p = 2(p p2), i = 1, 2, , 9. So E[X] = E[X1 + X2 ++ X9] = E[X1] + E[X2] + + E[X9] = 9[2(p p2)] = 18(p p2) b. Find the variance of the number of changeovers. Var(Xi) = P(Changeover)(1 P(Changeover)) = 2(p p2) (1 2(p p2)) = 2( 2p4 + 4p3 3p2 + p), i = 1, 2, , 9. Cov(Xi, Xj) will be 0 unless i and j are adjacent. We want i < j so let j = i + 1. E[XiXi+1] = P(2 changeovers) = P(HTH or THT) = p(1 p)p + (1 p)p(1 p) = p p2 Cov(Xi, Xi+1) = E[XiXi+1] E[Xi]E[Xi+1] = (p p2) (2(p p2))2 = 4p4 + 8p3 5p2 + p, i = 1, 2, , 8. So Var(X) = Var(Xi) + 2 Cov(Xi, Xi+1) = 9[2( 2p4 + 4p3 3p2 + p)] + 2*8[ 4p4 + 8p3 5p2 + p] = 100p4 + 200p3 134p2 + 34p c. Describe how the mean and variance of the number of changeovers behave for different values of p. Provide a brief logical explanation. Using Excel to plot E[X] and Var(X) for different values of p yields the following result:
5 4 3 2 1 0
0.01 0.05 0.09 0.13 0.17 0.21 0.25 0.29 0.33 0.37 0.41 0.45 0.49 0.53 0.57 0.61 0.65 0.69 0.73 0.77 0.81 0.85 0.89 0.93
E[X] Var(X)

Explanation: the expected value makes sense because with a higher likelihood of getting either all T or all H, we expect fewer changeovers. For the variance, as p goes to 0 or 1, the number of changeovers approaches 0 with certainty, and hence the variance approaches 0 too. In the middle, things are a little weird: variance is highest when p is around 0.78 and 0.22 and dips lower when p is closer to 0.5. It makes sense if you remember that variance is the average squared distance from the mean. When p is close to 0.5, the mean is 4.5, and the farthest away the actual number of changeovers could possibly be is 0 or 9 (a distance of 4.5). When p is slightly higher or lower, the mean drifts down, so the possibility of being really far away (e.g. 9 when the mean is 3, a distance of 6) drags the variance up.

0.97

2. Consider a Negative Binomial random variable Y ~ NB(r, p). Prove that Y is a proper rv iff p > 0 by the following methods: a. Express Y as the sum of r independent Geometric random variables, and apply a known result about Geometrics. Y = X1 + X2 + + Xr, where each Xi ~ Geo(p) P(Y < ) = P(X1 + X2 + + Xr < ) = P(X1 < , X2 < , Xr < ) for the sum to be finite, each Xi must be finite = P(X1 < )P(X2 < )P(Xr < ) by independence =1*1**1 since Geometrics are proper iff p > 0 =1 So the Negative Binomial is proper. b. Show that P(Y = ) = 0. For Y = , we would need somewhere before the rth S, an infinite sequence of Fs. P(Y = ) = P({any combination with up to r-1 Ss},F, F, F, F, .) = P({any such combination})P(F, F, F, ) by independence = c*(1-p)(1-p)(1-p) = 0 iff p > 0. Alternative Solution: P(Y = ) =1P(Y < ) = 1y=1 P(Y = y). Show that the sum is 1. c. Show (from first principles!) that E[Y] is r/p. Why does this imply Y is proper? First lets find the mean of a Geometric. E[X] = p + 2p(1-p) + 3p(1-p)2 + 4p(1-p)3 + = p{1 + (1-p) + (1-p)2 + (1-p)3 + + (1-p) + (1-p)2 + (1-p)3 + + (1-p)2 + (1-p)3 + } = p {1/p + (1-p)/p + (1-p)2/p + = p/p *1/(1-(1-p)) = 1/p (There are other tricks for finding the mean of a Geometric using a derivative trick, or multiplying the entire series through by (1-p) and subtracting) And since Y = X1 + X2 + + Xr, E[Y] = E[X1] + E[X2] + + E[Xr] = r E[Xi] = r/p. Alternative Solution: Work directly from the pmf of Y to show y=1 yP(Y = y) = r/p Since this is finite iff p > 0, this means Y must be proper, since if Y were improper, the mean would be infinite.

3. Suppose we have a series of independent trials, where the outcome is either S or F, but the probability of S on trial n is pn, not necessarily constant. Let X = number of trials until the first S, including that trial. a. Give an expression for the probability mass function, P(X = k). P(X = k) = P({F, F, , F, S}) (k 1 Fs followed by S) = P(F on trial 1)P(F on 2)P(F on k 1)P(S on k) since trials are independent = (1 p1)(1 p2)(1 pk1)pk, =

(1 p ) pk
n1 n

k 1

for k = 1, 2, 3,

b. Give an expression for P(X = ). P(X = ) = P({F, F, F, F, .}) infinite series of Fs = P(F on trial 1)P(F on 2)P(F on 3)P(F on 4). since trials are independent = (1 p1)(1 p2)(1 p3)(1 p4). =

(1 p )
n1 n

c. Prove that if pn = an, where 0 < a < , then

(1 p ) 1 a a2 ak
n1 n

Base case(s) [only one actually needed] k = 1: 1 a 1 a k = 2: (1 a)(1 a2) = 1 a (1 a)a2 > 1 a a2 2 k = 3: (1 a)(1 a )(1 a3) > (1 a a2)(1 a3) = 1 a a2 (1 a a2)a3 > 1 a a2 a3 Inductive Step
m

since (1 a) < 1 applying result from n = 2 since (1 a a2) < 1 if a <

Assume true for k = m. That is, assume Now for k = m + 1.

(1 a ) 1 a a2 am
n n1

(1 an ) =
n1

m1

(1 a ) (1 am+1)
n n1

> (1 a a2 am)(1 am+1) applying the Induction Hypothesis 2 m 2 = 1 a a a (1 a a am)am+1 > 1 a a2 am am+1 since (1 a am) < 1 if a < So the inequality holds for k = m+1, and thus holds for all positive integers k by the principle of mathematical induction.

d. Using the general result in c, when we take the limit as k goes to , what happens to

(1 pn ) if
n1

pn diverges? What if
n 1

p
n 1

converges to something < 1?

We showed that If

(1 pn ) 1 pn . So as k approaches ,
n1 n 1

(1 pn ) 1 pn .
n1 n 1

p
n 1

diverges,

(1 p )
n1 n
n1

But recall these pns are probabilities, so the only thing we know is (1 pn ) 0. (It actually turns out that

(1 p ) = 0 in this case!)
n1 n n1

On the other hand, if

pn converges to a value < 1 (say b), then (1 pn ) 1 b > 0

(This is a strict inequality, rather than simply knowing the product is non-negative.) See the new handout for a summary of this rule. e. Back to our trials. Suppose pn = 3n. Determine using the results in c and d whether X is proper or improper by finding P(X = ). P(X = ) = =

(1 p )
(1 3
n1

n1

from b in this case from c and d, where a = 31 = 1/3

1 =1

3
n 1

1/ 3 = 0.5 11/ 3

The probability of X equaling is at least 0.5, which is clearly greater than 0. So X is an improper random variable.

4. Suppose X1, X2, X3, and Y are independent continuous uniform random variables over the unit interval (0,1). a. Find P(Xi < Y). quick method: Since Xi and Y are i.i.d. random variables, then, by symmetry, each is equally likely to be the larger of the two. Therefore P(X i < Y ) = P(Xi > Y) = 1/2. longer method (by conditioning on Y ): P(Xi < Y) = 01 P(Xi < Y | Y = y)dy since the pdf f(y) = 1 for U(0,1) 1 = 0 P(Xi < y | Y = y)dy substitution = 01 P(Xi < y)dy since Xi and Y are indep 1 = 0 y dy since Xi is U(0,1) and the cdf F(y) = y = 1/2 b. Find P(Xi < Y AND Xj < Y) for i j. quick method: Again by symmetry, since each of Xi, Xj, and Y is equally likely to be the largest of the three, it follows that P(Xi < Y AND Xj < Y) = 1/3. longer method (by conditioning on Y ): P(Xi < Y AND Xj < Y) = 01 P(Xi < Y AND Xj < Y | Y = y)dy since the pdf f(y) = 1 1 = 0 P(Xi < y AND Xj < y | Y = y)dy substitution = 01 P(Xi < y AND Xj < y)dy Xi and Xj indep of Y 1 = 0 P(Xi < y)P(Xj < y)dy Xi and Xj indep of each other = 01 y2 dy since Xi and Xj are U(0,1) = 1/3 Note: even though the four variables are independent, the events X i < Y and Xj < Y are NOT independent, since they both depend on Y. So we CANNOT say its * = c. Use the above to evaluate Var(Z), where Z is the number of Xis that are < Y We can express Z as the sum of three indicators, one for each of the X is being < Y Z = I1 + I 2 + I3 Var(Ii) = P(Xi < Y)(1 P(Xi < Y)) = 1/2(1 1/2) from a = 1/4 Cov(Ii, Ij) = P(Xi < Y AND Xj < Y) P(Xi < Y)P(Xj < Y) = 1/3 (1/2)2 from b and a = 1/12 So Var(Z) = Var(I1) + Var(I2) + Var(I3) + 2 Cov(Ii, Ij) = 3Var(Ii) + 6Cov(Ii, Ij) = 3(1/4) + 6(1/12) = 5/4 This can also be done using symmetry arguments. The position of Y when the variables are ordered smallest to largest is equally likely to be smallest (in which case Z = 0), second-smallest (Z = 1), second-largest (Z = 2) or largest (Z = 3). E[Z] = 1/4(0 + 1 + 2 + 3) = 6/4 E[Z2] = 1/4(02 + 12 + 22 + 32) = 14/4 So Var(Z) = 14/4 (6/4)2 = 5/4

5. A gambler wins each game with probability p. Find the expected total number of wins if a. The gambler plays n games. If he/she wins X of these games, then he/she will play an additional X games, and then stop. Let W = total number of wins. E[W|X=x] = x + px since we know they win x of the first n games, and they will on average win px of the additional x games So E[W|X] = (1+p)X E[W] = E[E[W|X]] by double averaging = E[(1+p)X] = (1+p)E[X] = (1+p)np since X ~ Bin(n, p) with mean np b. The gambler plays until he/she wins once. If it takes him/her Y games to get this win, he/she will play an additional Y games, and then stop. Again let W = total number of wins. E[W|Y=y] = 1 + py since we know they win 1 of the first y games, and they will on average win py of the additional x games So E[W|Y] = 1 +pY E[W] = E[E[W|Y]] = E[1+pY] = 1+pE[Y] = 1+p(1/p) =2 by double averaging

since Y ~ Geo(p) with mean 1/p