This action might not be possible to undo. Are you sure you want to continue?

:

Introduction to Probability Models: Eighth Edition

by Sheldon M. Ross.

John L. Weatherwax

∗

October 26, 2008

Introduction

Chapter 1: Introduction to Probability Theory

Chapter 1: Exercises

Exercise 8 (Bonferroni’s inequality)

From the inclusion/exclusion identity for two sets we have

P(E ∪ F) = P(E) +P(F) −P(EF) .

Since P(E ∪ F) ≤ 1, the above becomes

P(E) +P(F) −P(EF) ≤ 1 .

or

P(EF) ≥ P(E) +P(F) −1 ,

which is known as Bonferroni’s inequality. From the numbers given we ﬁnd that

P(EF) ≥ 0.9 + 0.8 −1 = 0.7 .

∗

wax@alum.mit.edu

1

1 2 3 4 5 6

1 2 3 4 5 6 7

2 3 4 5 6 7 8

3 4 5 6 7 8 9

4 5 6 7 8 9 10

5 6 7 8 9 10 11

6 7 8 9 10 11 12

Table 1: The possible values for the sum of the values when two die are rolled.

Exercise 10 (Boole’s inequality)

We begin by decomposing the countable union of sets A

i

A

1

∪ A

2

∪ A

3

. . .

into a countable union of disjoint sets C

j

. Deﬁne these disjoint sets as

C

1

= A

1

C

2

= A

2

\A

1

C

3

= A

3

\(A

1

∪ A

2

)

C

4

= A

4

\(A

1

∪ A

2

∪ A

3

)

.

.

.

C

j

= A

j

\(A

1

∪ A

2

∪ A

3

∪ · · · ∪ A

j−1

)

Then by construction

A

1

∪ A

2

∪ A

3

· · · = C

1

∪ C

2

∪ C

3

· · · ,

and the C

j

’s are disjoint, so that we have

Pr(A

1

∪ A

2

∪ A

3

∪ · · · ) = Pr(C

1

∪ C

2

∪ C

3

∪ · · · ) =

j

Pr(C

j

) .

Since Pr(C

j

) ≤ Pr(A

j

), for each j, this sum is bounded above by

j

Pr(A

j

) ,

Problem 11 (the probability the sum of the die is i)

We can explicitly enumerate these probabilities by counting the number of times each oc-

currence happens, in Table 1 we have placed the sum of the two die in the center of each

square. Then by counting the number of squares where are sum equals each number from

two to twelve, we have

P

2

=

1

36

, P

7

=

6

36

=

1

6

P

3

=

2

36

=

1

18

, P

8

=

5

36

P

4

=

3

36

=

1

12

, P

9

=

4

36

=

1

9

P

5

=

4

36

=

1

9

, P

10

=

3

36

=

1

12

P

6

=

5

36

, P

11

=

2

36

=

1

18

, P

12

=

1

36

.

Problem 13 (winning at craps)

From Problem 11 we have computed the individual probabilities for various sum of two

random die. Following the hint, let E

i

be the event that the initial die sum to i and that

the player wins. We can compute some of these probabilities immediately P(E

2

) = P(E

3

) =

P(E

12

) = 0, and P(E

7

) = P(E

11

) = 1. We now need to compute P(E

i

) for i = 4, 5, 6, 8, 9, 10.

Again following the hint deﬁne E

i,n

to be the event that the player initial sum is i and wins

on the n-th subsequent roll. Then

P(E

i

) =

∞

n=1

P(E

i,n

) ,

since if we win, it must be either on the ﬁrst, or second, or third, etc roll after the initial

roll. We now need to calculate the P(E

i,n

) probabilities for each n. As an example of this

calculation ﬁrst lets compute P(E

4,n

) which means that we initially roll a sum of four and

the player wins on the n-th subsequent roll. We will win if we roll a sum of a four or loose

if we roll a sum of a seven, while if roll anything else we continue, so to win when n = 1 we

see that

P(E

4,1

) =

1 + 1 + 1

36

=

1

12

,

since to get a sum of four we can roll pairs consisting of (1, 3), (2, 2), and (3, 1).

To compute P(E

4,2

) the rules of craps state that we will win if a sum of four comes up (with

probability

1

12

) and loose if a sum of a seven comes up (with probability

6

36

=

1

6

) and continue

playing if anything else is rolled. This last event (continued play) happens with probability

1 −

1

12

−

1

6

=

3

4

.

Thus P(E

4,2

) =

_

3

4

_

1

12

=

1

16

. Here the ﬁrst

3

4

is the probability we don’t roll a four or a

seven on the n = 1 roll and the second

1

12

comes from rolling a sum of a four on the second

roll (where n = 2). In the same way we have for P(E

4,3

) the following

P(E

4,3

) =

_

3

4

_

2

1

12

.

Here the ﬁrst two factors of

3

4

are from the two rolls that “keep us in the game”, and the

factor of

1

12

, is the roll that allows us to win. Continuing in this in this manner we see that

P(E

4,4

) =

_

3

4

_

3

1

12

,

and in general we ﬁnd that

P(E

4,n

) =

_

3

4

_

n−1

1

12

for n ≥ 1 .

To compute P(E

i,n

) for other i, the derivations just performed, only change in the probabil-

ities required to roll the initial sum. We thus ﬁnd that for other initial rolls (heavily using

the results of Problem 24) that

P(E

5,n

) =

1

9

_

1 −

1

9

−

1

6

_

n−1

=

1

9

_

13

18

_

n−1

P(E

6,n

) =

5

36

_

1 −

5

36

−

1

6

_

n−1

=

5

36

_

25

36

_

n−1

P(E

8,n

) =

5

36

_

1 −

5

36

−

1

6

_

n−1

=

5

36

_

25

36

_

n−1

P(E

9,n

) =

1

9

_

1 −

1

9

−

1

6

_

n−1

=

1

9

_

13

18

_

n−1

P(E

10,n

) =

1

12

_

1 −

1

12

−

1

6

_

n−1

=

1

12

_

3

4

_

n−1

.

To compute P(E

4

) we need to sum the results above. We have that

P(E

4

) =

1

12

n≥1

_

3

4

_

n−1

=

1

12

n≥0

_

3

4

_

n

=

1

12

1

_

1 −

3

4

_ =

1

3

.

Note that this also gives the probability for P(E

10

). For P(E

5

) we ﬁnd P(E

5

) =

2

5

, which

also equals P(E

9

). For P(E

6

) we ﬁnd that P(E

6

) =

5

11

, which also equals P(E

8

). Then our

probability of winning craps is given by summing all of the above probabilities weighted by

the associated priors of rolling the given initial roll. We ﬁnd by deﬁning I

i

to be the event

that the initial roll is i and W the event that we win at craps that

P(W) = 0 P(I

2

) + 0 P(I

3

) +

1

3

P(I

4

) +

4

9

P(I

5

) +

5

9

P(I

6

)

+ 1 P(I

7

) +

5

9

P(I

8

) +

4

9

P(I

9

) +

1

3

P(I

10

) + 1 P(I

11

) + 0 P(I

12

) .

Using the results of Exercise 25 to evaluate P(I

i

) for each i we ﬁnd that the above summation

gives

P(W) =

244

495

= 0.49292 .

These calculations are performed in the Matlab ﬁle chap 1 prob 13.m.

Exercise 15 (some set identities)

We want to prove that E = (E ∩ F) ∪ (E ∩ F

c

). We will do this using the standard proof

where we show that each set in the above is a subset of the other. We begin with x ∈ E.

Then if x ∈ F, x will certainly be in E ∩ F, while if x / ∈ F then x will be in E ∩ F

c

. Thus

in either case (x ∈ F or x / ∈ F) x will be in the set (E ∩ F) ∪ (E ∩ F

c

).

If x ∈ (E ∩ F) ∪ (E ∩ F

c

) then x is in either E ∩ F, E ∩ F

c

, or both by the deﬁnition of

the union operation. Now x cannot be in both sets or else it would simultaneously be in F

and F

c

, so x must be in one of the two sets only. Being in either set means that x ∈ E and

we have that the set (E ∩ F) ∪ (E ∩ F

c

) is a subset of E. Since each side is a subset of the

other we have shown set equality.

To prove that E∪F = E∪(E

c

∩F), we will begin by letting x ∈ E∪F, thus x is an element

of E or an element of F or of both. If x is in E at all then it is in the set E ∪ (E

c

∩ F). If

x / ∈ E then it must be in F to be in E ∪ F and it will therefore be in E

c

∩ F. Again both

sides are subsets of the other and we have shown set equality.

Exercise 23 (conditioning on a chain of events)

This result follows for the two set case P{A∩B} = P{A|B}P{B} by grouping the sequence

of E

i

’s in the appropriate manner. For example by grouping the intersection as

E

1

∩ E

2

∩ · · · ∩ E

n−1

∩ E

n

= (E

1

∩ E

2

∩ · · · ∩ E

n−1

) ∩ E

n

we can apply the two set result to obtain

P{E

1

∩ E

2

∩ · · · ∩ E

n−1

∩ E

n

} = P{E

n

|E

1

∩ E

2

∩ · · · ∩ E

n−1

} P{E

1

∩ E

2

∩ · · · ∩ E

n−1

} .

Continuing now to peal E

n−1

from the set E

1

∩E

2

∩· · · ∩E

n−1

we have the second probability

above equal to

P{E

1

∩ E

2

∩ · · · ∩ E

n−2

∩ E

n−1

} = P{E

n−1

|E

1

∩ E

2

∩ · · · ∩ E

n−2

}P{E

1

∩ E

2

∩ · · · ∩ E

n−2

} .

Continuing to peal oﬀ terms from the back we eventually obtain the requested expression

i.e.

P{E

1

∩ E

2

∩ · · · ∩ E

n−1

∩ E

n

} = P{E

n

|E

1

∩ E

2

∩ · · · ∩ E

n−1

}

× P{E

n−1

|E

1

∩ E

2

∩ · · · ∩ E

n−2

}

× P{E

n−2

|E

1

∩ E

2

∩ · · · ∩ E

n−3

}

.

.

.

× P{E

3

|E

1

∩ E

2

}

× P{E

2

|E

1

}

× P{E

1

} .

Exercise 30 (target shooting with Bill and George)

warning! not ﬁnished...

Let H be the event that the duck is “hit”, by either Bill or George’s shot. Let B and G be the

events that Bill (respectively George) hit the target. Then the outcome of the experiment

where both George and Bill ﬁre at the target (assuming that their shots work independently

is)

P(B

c

, G

c

) = (1 −p

1

)(1 −p

2

)

P(B

c

, G) = (1 −p

1

)p

2

P(B, G

c

) = p

1

(1 −p

2

)

P(B, G) = p

1

p

2

.

Part (a): We desire to compute P(B, G|H) which equals

P(B, G|H) =

P(B, G, H)

P(H)

=

P(B, G)

P(H)

Now P(H) = (1 −p

1

)p

2

+p

1

(1 −p

2

) +p

1

p

2

so the above probability becomes

p

1

p

2

(1 −p

1

)p

2

+p

1

(1 −p

2

) +p

1

p

2

=

p

1

p

2

p

1

+p

2

−p

1

p

2

.

Part (b): We desire to compute P(B|H) which equals

P(B|H) = P(B, G|H) +P(B, G

c

|H) .

Since the ﬁrst termP(B, G|H) has already been computed we only need to compute P(B, G

c

|H).

As before we ﬁnd it to be

P(B, G

c

|H) =

p

1

(1 −p

2

)

(1 −p

1

)p

2

+p

1

(1 −p

2

) +p

1

p

2

.

So the total result becomes

P(B|H) =

p

1

p

2

+p

1

(1 −p

2

)

(1 −p

1

)p

2

+p

1

(1 −p

2

) +p

1

p

2

=

p

1

p

1

+p

2

−p

1

p

2

.

Exercise 33 (independence in class)

Let S be a random variable denoting the sex of the randomly selected person. The S can

take on the values m for male and f for female. Let C be a random variable representing

denoting the class of the chosen student. The C can take on the values f for freshman and

s for sophomore. We want to select the number of sophomore girls such that the random

variables S and C are independent. Let n denote the number of sophomore girls. Then

counting up the number of students that satisfy each requirement we have

P(S = m) =

10

16 +n

P(S = f) =

6 +n

16 +n

P(C = f) =

10

16 +n

P(C = s) =

6 +n

16 +n

.

The joint density can also be computed and are given by

P(S = m, C = f) =

4

16 +n

P(S = m, C = s) =

6

16 +n

P(S = f, C = f) =

6

16 +n

P(S = f, C = s) =

n

16 +n

.

Then to be independent we must have P(C, S) = P(S)P(C) for all possible C and S values.

Considering the point case where (S = m, C = f) we have that n must satisfy

P(S = m, C = f) = P(S = m)P(C = f)

4

16 +n

=

_

10

16 +n

__

10

16 +n

_

which when we solve for n gives n = 9. Now one should check that this value of n works for

all other equalities that must be true, for example one needs to check that when n = 9 the

following are true

P(S = m, C = s) = P(S = m)P(C = s)

P(S = f, C = f) = P(S = f)P(C = f)

P(S = f, C = s) = P(S = f)P(C = s) .

As these can be shown to be true, n = 9 is the correct answer.

Exercise 36 (boxes with marbles)

Let B be the event that the drawn ball is black and let X

1

(X

2

) be the event that we select

the ﬁrst (second) box. Then to calculate P(B) we will condition on the box drawn from as

P(B) = P(B|X

1

)P(X

1

) +P(B|X

2

)P(X

2

) .

Now P(B|X

1

) = 1/2, P(B|X

2

) = 2/3, P(X

1

) = P(X

2

) = 1/2 so

P(B) =

1

2

_

1

2

_

+

1

2

_

2

3

_

=

7

12

.

Exercise 37 (observing a white marble)

If we see that the ball is white (i.e. it is not black i.e event B

c

has happened) we now want

to compute that it was drawn from the ﬁrst box i.e.

P(X

1

|B

c

) =

P(B

c

|X

1

)P(X

1

)

P(B

c

|X

1

)P(X

1

) +P(B

c

|X

2

)P(X

2

)

=

3

5

.

Problem 40 (gambling with a fair coin)

Let F denote the event that the gambler is observing results from a fair coin. Also let O

1

,

O

2

, and O

3

denote the three observations made during our experiment. We will assume that

before any observations are made the probability that we have selected the fair coin is 1/2.

Part (a): We desire to compute P(F|O

1

) or the probability we are looking at a fair coin

given the ﬁrst observation. This can be computed using Bayes’ theorem. We have

P(F|O

1

) =

P(O

1

|F)P(F)

P(O

1

|F)P(F) +P(O

1

|F

c

)P(F

c

)

=

1

2

_

1

2

_

1

2

_

1

2

_

+ 1

_

1

2

_ =

1

3

.

Part (b): With the second observation and using the “posteriori’s become priors” during a

recursive update we now have

P(F|O

2

, O

1

) =

P(O

2

|F, O

1

)P(F|O

1

)

P(O

2

|F, O

1

)P(F|O

1

) +P(O

2

|F

c

, O

1

)P(F

c

|O

1

)

=

1

2

_

1

3

_

1

2

_

1

3

_

+ 1

_

2

3

_ =

1

5

.

Part (c): In this case because the two-headed coin cannot land tails we can immediately

conclude that we have selected the fair coin. This result can also be obtained using Bayes’

theorem as we have in the other two parts of this problem. Speciﬁcally we have

P(F|O

3

, O

2

, O

1

) =

P(O

3

|F, O

2

, O

1

)P(F|O

2

, O

1

)

P(O

3

|F, O

2

, O

1

)P(F|O

2

, O

1

) +P(O

3

|F

c

, O

2

, O

1

)P(F

c

|O

2

, O

1

)

=

1

2

_

1

5

_

1

2

_

1

5

_

+ 0

= 1 .

Verifying what we know must be true.

Problem 46 (a prisoners’ dilemma)

I will argue that the jailers reasoning is sound. Before asking his question the probability of

event A (A is executed) is P(A) = 1/3. If prisoner A is told that B (or C) is to be set free

then we need to compute P(A|B

c

). Where A, B, and C are the events that prisoner A, B,

or C is to be executed respectively. Now from Bayes’ rule

P(A|B

c

) =

P(B

c

|A)P(A)

P(B

c

)

.

We have that P(B

c

) is given by

P(B

c

) = P(B

c

|A)P(A) +P(B

c

|B)P(B) +P(B

c

|C)P(C) =

1

3

+ 0 +

1

3

=

2

3

.

So the above probability then becomes

P(A|B

c

) =

1(1/3)

2/3

=

1

2

>

1

3

.

Thus the probability that prisoner A will be executed has increased as claimed by the jailer.

Chapter 4: Markov Chains

Chapter 4: Exercises

Exercise 6 (an analytic calculation of P

(n)

)

Given the transition probability matrix P =

_

p 1 −p

1 −p p

_

, by matrix multiplication we

see that P

(2)

is given as

P

(2)

=

_

p 1 −p

1 −p p

_ _

p 1 −p

1 −p p

_

=

_

p

2

+ (1 −p)

2

2p(1 −p)

2p(1 −p) p

2

+ (1 −p)

2

_

.

We desire to prove that P

(n)

is given as

P

(n)

=

_

1

2

+

1

2

(2p −1)

n 1

2

−

1

2

(2p −1)

n

1

2

−

1

2

(2p −1)

n 1

2

+

1

2

(2p −1)

n

_

. (1)

We will do this by mathematical induction. We begin by verifying that the above formula

is valid for n = 1. Evaluating the above expression for n = 1 we ﬁnd

P

(1)

=

_

1

2

+

1

2

(2p −1)

1

2

−

1

2

(2p −1)

1

2

−

1

2

(2p −1)

1

2

+

1

2

(2p −1)

_

=

_

p 1 −p

1 −p p

_

,

as required. Next we assume the relationship in Equation 1 is true for all n ≤ k and we

desire to show that it is true for n = k +1. Since P

(k+1)

= P

(1)

P

(k)

by matrix multiplication

we have that

P

(k+1)

= P

(1)

P

(k)

=

_

p 1 −p

1 −p p

_ _

1

2

+

1

2

(2p −1)

k 1

2

−

1

2

(2p −1)

k

1

2

−

1

2

(2p −1)

k 1

2

+

1

2

(2p −1)

k

_

=

_

p

2

+

p

2

(2p −1)

k

+

1−p

2

−

1−p

2

(2p −1)

k p

2

−

p

2

(2p −1)

k

+

1−p

2

+

1−p

2

(2p −1)

k

1−p

2

+

1−p

2

(2p −1)

k

+

p

2

−

p

2

(2p −1)

k 1−p

2

−

1−p

2

(2p −1)

k

+

p

2

+

p

2

(2p −1)

k

_

=

_

1

2

+

1

2

(p −(1 −p))(2p −1)

k 1

2

+

_

−

p

2

+

1−p

2

_

(2p −1)

k

1

2

+

1

2

(1 −p −p)(2p −1)

k 1

2

+

1

2

(−(1 −p) +p)(2p −1)

k

_

=

_

1

2

+

1

2

(2p −1)

k+1 1

2

−

1

2

(2p −1)

k+1

1

2

−

1

2

(2p −1)

k+1 1

2

+

1

2

(2p −1)

k+1

_

,

which is the desired result for n = k + 1.

Chapter 5: The Exponential Distribution

and the Poisson Process

Chapter 5: Exercises

Exercise 1 (exponential repair times)

We are told that T is distributed with a exponential probability distribution function with

mean 1/2. This means that the distribution function of T, f

T

(t), is given by

f

T

(t) =

_

2e

−2t

t ≥ 0

0 t < 0

Part (a): Now for the repair time to take longer than 1/2 an hour will happen with

probability that is the complement of the probability that it will take less than 1/2 of an

hour. We have

1 −P{T < 1/2} = 1 −

_

1/2

0

2e

−2t

dt

= 1 −

2e

−2t

−2

¸

¸

¸

¸

1/2

0

= 1 +

_

e

−1

−1

_

= e

−1

.

Part (b): Since the exponential distribution has no memory, the fact that the repair is still

going after 12 hours is irrelevant. Thus we only need to compute the probability that the

repair will last at least 1/2 more. This probability is the same as that calculated in Part (a)

of this problem and is equal to e

−1

.

Exercise 2 (the expected bank waiting time)

By the memoryless property of the exponential distribution now the fact that one person is

being served makes no diﬀerence. We will have to wait until an amount of time given by

T =

6

i=1

X

i

,

where X

i

are independent exponential random variables with rate µ. Taking the expectation

of the variable T we have

E[T] =

6

i=1

E[X

i

] =

6

i=1

1

µ

=

6

µ

.

We sum to six to allow the teller to service the ﬁve original customers and then you.

Exercise 3 (expectations with exponential random variables)

I will argue that E[X

2

|X > 1] = E[(X+1)

2

]. By the memoryless property of the exponential

random variable the fact that we are conditioning on the event that X > 1 makes no

diﬀerence relative to the event X > 0 (i.e. no restriction on the random variable X).

Removing the conditional expectation is equivalent to “starting” the process at x = 0. This

can be performed as long as we “shift” the expectation’s argument accordingly i.e. from X

2

to (X + 1)

2

. The other two expressions violate the nonlinearity of the function X

2

. We can

prove that this result is correct by explicitly evaluating the original expectation. We ﬁnd

E[X

2

|X > 1] =

_

∞

0

ξ

2

p

X

(ξ|X > 1)dξ .

Now this conditional probability density is given by

p

X

(ξ|X > 1) =

p(X = ξ, X > 1)

p(X > 1)

=

p(X = ξ, X > 1)

1 −p(X < 1)

=

p(X = ξ, X > 1)

e

−λ

=

λe

−λξ

H(ξ −1)

e

−λ

= λe

−λ(ξ−1)

H(ξ −1) .

Here H(·) is the Heaviside function deﬁned as

H(x) =

_

0 x < 0

1 x > 0 ,

and this function enforces the constraint that X > 1. With this deﬁnition we then have that

E[X

2

|X > 1] =

_

∞

0

ξ

2

λe

−λ(ξ−1)

H(ξ −1)dξ

=

_

∞

1

ξ

2

λe

−λ(ξ−1)

dξ ,

Letting u = ξ −1, so that du = dξ the above becomes

_

∞

0

(u + 1)

2

λe

−λu

du = E[(X + 1)

2

] ,

where X is an exponential random variable with rate parameter λ. This is the expression

we argued at the beginning of this problem should hold true.

Exercise 4 (the post oﬃce)

Part (a): In this case, it is not possible for A to still be in the post oﬃce because in

ten minutes time A and B will both ﬁnish their service times and exit the service station

together. Thus there is no way for C to get served before A ﬁnishes.

Part (b): A will still be in the post oﬃce if both B and C are served before A. If we let

A, B, and C be the amount of time that each respective person spends with their clerks,

then the event that A is the last person in the post oﬃce is equivalent to the constraint that

A ≥ B+C. Here we have assumed (by using ≥) that an equality constraint is acceptable for

determining if A leaves last. For notational purposes we will deﬁne the event that A leaves

last as E. To compute the probabilities that this event happens we can condition on the

possible sums of B and C (the times B and C spend at their clerks). We have

P(E) = P(E|B +C = 2)P(B +C = 2) +P(E|B +C = 3)P(B +C = 3)

+ P(E|B +C = 4)P(B +C = 4) +P(E|B +C = 5)P(B +C = 5)

+ P(E|B +C = 6)P(B +C = 6) .

Now P(E|B + C = 4) = P(E|B + C = 5) = P(E|B + C = 6) = 0, since A will certainly

ﬁnish in a time less than four units. Also

P(E|B +C = 2) =

2

3

,

since to have A ≥ B +C, A can ﬁnishes in two or three units. While ﬁnally

P(E|B +C = 3) =

1

3

,

since to have A ≥ B+C, A can ﬁnish in three units. Now we have that our priors are given

by P(B+C = 2) =

1

9

and P(B+C = 3) =

2

9

, which gives for P(E) using the above formula

P(E) =

4

27

.

If we want to work this problem assuming strict inequality in the time relationships i.e. that

A will leave last only if A > B + C, we ﬁnd that our conditional probabilities must be

adjusted. For example

P(E|B +C = 2) =

1

3

,

since A must now ﬁnish in three units of time. Also P(E|B +C = 3) = 0. These then give

in the same way that

P(E) =

1

3

·

1

9

=

1

27

.

Part (c): For this part we assume that the service time of each clerk is exponentially dis-

tributed. Now because the random variables are continuous there is no need to diﬀerentiate

between greater than and greater than or equal signs in the inequality denoting the event

that A ﬁnishes last. Thus we can take as an expression of the event that A will be the last

person served if A > B +C. This means that the time to service A takes more time than to

service both B and C. To evaluate the probability that this event happens we will condition

on the random variable which is the sum of B and C, i.e.

P{A > B +C} =

_

p{A > B +C|B +C = t}P{B +C = t}dt .

Now since both B and C are exponential random variables with the same rate the sum of

them is a random variable distributed with a gamma distribution (see the section on further

properties of the exponential distribution). We thus have that the distribution of the sum

of B and C is given by

f

B+C

(t) =

1!e

−µt

(µt)

1!

= µ

2

te

−µt

.

So that our integral is given by

P{A > B +C} =

_

P

A

{A > t}µ

2

te

−µt

dt .

Now P

A

{A > t} = 1 −(1 −e

−µt

) = e

−µt

, so that this integral becomes

P{A > B +C} =

_

∞

0

µ

2

te

−2µt

dt

= µ

2

te

−2µt

(−2µ)

¸

¸

¸

¸

∞

0

−

µ

2

(−2µ)

_

∞

0

e

−2µt

dt

=

µ

2

_

e

−2µt

(−2µ)

¸

¸

¸

¸

∞

0

=

1

4

.

Exercise 5 (old radios)

Because of the memoryless property of the exponential distribution the fact that the radio

is ten years old makes no diﬀerence. Thus the probability that the radio is working after an

additional ten years is given by

P{X > 10} =

_

∞

10

λe

−λt

dt = 1 −F(10) = 1 −

1

10

e

−

1

10

(10)

= 1 −

1

10

e

−1

= 0.96321 .

Exercise 6 (the probability that Smith is last)

Now one of Mr. Jones or Mr. Brown will ﬁnish their service ﬁrst. Once this person’s

service station is open, Smith will begin his processing there. By the memoryless property

of exponential the fact that Smiths “competitor” (the remaining Mr. Jones or Mr. Brown

who did not ﬁnish ﬁrst) is in the middle of their service has no eﬀect on whether Smith or

his comp editor ﬁnishes ﬁrst. Let E be the event that Smith is ﬁnishes last. Then we have

that P(E

c

) = 1 − P(E), where E

c

is the desired event (i.e. that Smith is last). Now let

JB be the event that Mr. Jones ﬁnishes before Mr. Brown (then Mr. Smith will take Mr.

Jones’ server position), and let JB be the event that Mr. Brown ﬁnishes before Mr. Jones

(so that Mr. Smith would take Mr. Brown’s server positions). We have that by conditioning

on the events JB and BJ that

P(E) = P(E|JB)P(JB) +P(E|BJ)P(BJ) .

Now

P(JB) =

λ

J

λ

J

+λ

B

and P(BJ) =

λ

B

λ

J

+λ

B

,

where λ

J

and λ

B

are the rates of the servers who are initially servicing Mr. Jones and Mr.

Brown when Mr. Smith enters the post oﬃce. By the discussion in the text on the section of

the text entitled “Further properties of the exponential distribution”, we have that P(E|JB)

is the probability that Mr. Smith is last given that he is being serviced by Mr. Jones’ old

server. This will happen if Mr. Brown ﬁnishes ﬁrst and thus has a probability of

λ

B

λ

B

+λ

J

.

In the same way we have

P(E|BJ) =

λ

J

λ

B

+λ

J

.

So that P(E) is given by

P(E) =

λ

B

λ

B

+λ

J

·

λ

J

λ

B

+λ

J

+

λ

J

λ

B

+λ

J

·

λ

B

λ

B

+λ

J

=

2λ

B

λ

J

(λ

B

+λ

J

)

2

.

The probability that we want is P(E

c

) and is given by

P(E

c

) = 1 −P(E)

=

(λ

B

+λ

J

)

2

(λ

B

+λ

J

)

2

−

2λ

B

λ

J

(λ

B

+λ

J

)

2

=

λ

2

B

(λ

B

+λ

J

)

2

+

λ

2

J

(λ

B

+λ

J

)

2

,

as we where to show.

Exercise 7 (the probability X

1

< X

2

given the minimum value of X

1

and X

2

)

For this exercise we will be considering the expression P{X

1

< X

2

| min(X

1

, X

2

) = t}. Using

the deﬁnition of conditional probability we have that this is equal to

P{X

1

< X

2

| min(X

1

, X

2

) = t} =

P{X

1

< X

2

, min(X

1

, X

2

) = t}

P{min(X

1

, X

2

) = t}

=

P{X

1

< X

2

, X

1

= t}

P{min(X

1

, X

2

) = t}

=

P{t < X

2

, X

1

= t}

P{min(X

1

, X

2

) = t}

=

P{X

2

> t}P{X

1

= t}

P{min(X

1

, X

2

) = t}

.

Where in the last step we used the independence of X

1

and X

2

. We now need to be able to

relate the expressions above in terms of the failure rates of the random variables X

1

and X

2

.

Given the failure rate of a random variable as a function of t say r(t) from the discussion in

the book the distribution function F(·) in this case is given by

F(t) = e

−

R

t

0

r(τ)dτ

.

Taking the derivative of this expression we ﬁnd that the density function f(·) for our random

variable with failure rate r(t) is given by

f(t) = r(t)e

−

R

t

0

r(τ)dτ

.

With these two expressions we can evaluate part of the above fraction. We ﬁnd that

P{X

2

> t}P{X

1

= t}

P{min(X

1

, X

2

) = t}

=

(e

−

R

t

0

r

2

(τ)dτ

)r

1

(t)e

−

R

t

0

r

1

(τ)dτ

P{min(X

1

, X

2

) = t}

.

Lets now work on evaluating the denominator of the above expression, which we do in the

same way as in the book. We ﬁnd that

P{min(X

1

, X

2

) > t} = P{X

i

> t , ∀i}

=

2

i=1

P{X

i

> t}

= e

−

R

t

0

r

1

(τ)dτ

e

−

R

t

0

r

2

(τ)dτ

= e

−

R

t

0

(r

1

(τ)+r

2

(τ))dτ

.

From which we see that min(X

1

, X

2

) is a random variable with a failure rate given by

r

1

(t)+r

2

(t). This then means that our density function for this random variable min(X

1

, X

2

)

is given by

P{min(X

1

, X

2

) = t} = (r

1

(t) +r

2

(t))e

−

R

t

0

(r

1

(τ)+r

2

(τ))dτ

.

Using this result in our desired expression we ﬁnally conclude that

P{X

1

< X

2

| min(X

1

, X

2

) = t} =

r

1

(t)e

−

R

t

0

(r

1

(τ)+r

2

(τ))dτ

(r

1

(t) +r

2

(t))e

−

R

t

0

(r

1

(τ)+r

2

(τ))dτ

=

r

1

(t)

r

1

(t) +r

2

(t)

,

as we were to show.

Exercise 8 (the expectation of 1/r(x))

We will use an equivalent expression for the expectation of X which is derived in many books

on probability, for example see ??, which is that

E[X] =

_

P{X > x}dx.

From this expression and with the following manipulations to introduce the failure rate

function r(X) we ﬁnd that

E[X] =

_

P{X > x}dx

=

_

1 −F(x)

f(x)

f(x)dx

=

_

1

r(x)

f(x)dx

= E[

1

r(X)

] .

Here we have used f(·) for the probability density function, F(·) for the cumulative density

function and r(·) for the failure rate function, i.e. r(x) ≡

f(x)

1−F(x)

.

Exercise 9 (the probability we fail ﬁrst after working t time)

We are told that machine one is currently working and at time t, machine two will be put

into service. At that point if the ﬁrst machine is still working they will be working in tandem.

We want to calculate the probability that machine one is the ﬁrst machine to fail. Let E

be the event that machine one fails ﬁrst and F the event that machine one fails in the time

(0, t). Now conditioning on F and F

c

we have that

P(E) = P(E|F)P(F) +P(E|F

c

)P(F

c

) .

Now introducing X

1

and X

2

be the random variables denoting the time when machine

one/two fails.

P(F) = P{X

1

< t} = 1 −e

−λ

1

t

P(F

c

) = P{X

1

> t} = e

−λ

1

t

.

Now by the memoryless property of the exponential distribution when the second machine

is brought on-line with the ﬁrst machine (assuming that machine one has not failed in the

time (0, t)) the probability that machine one fails before machine two is given by

λ

1

λ

1

+λ

2

,

which is the probability P(E|F

c

). In addition we have that P(E|F) = 1, so that P(E)

becomes

P(E) = 1 · (1 −e

−λ

1

t

) +

λ

1

λ

1

+λ

2

· e

−λ

1

t

= 1 −

_

λ

2

λ

1

+λ

2

_

e

−λ

1

t

Exercise 10 (some expectations)

Part (a): We want to evaluate E[MX|M = X]. Note that if we try to evaluate this as

E[MX|M = X] = E[X

2

] ,

this result cannot be correct since we have eﬀectively lost the information that our minimum

of X and Y is X since the last expression involves only X. Instead we ﬁnd that

E[MX|M = X] = E[M

2

] =

_

∞

0

m

2

e

−(λ+µ)m

dm =

2

(λ +µ)

2

,

when we perform the integration.

Part (b): To evaluate E[MX|M = Y ], we will stress the information that Y = M is

less that X by writing X = M + X

′

. Now by the memoryless property of the exponential

distribution X

′

is another exponential random variable with failure rate λ. With this we

ﬁnd that

E[MX|M = Y ] = E[M(M +X

′

)]

= E[M

2

] +E[MX

′

]

= E[M

2

] +E[M]E[X

′

]

=

2

(λ +µ)

2

+

1

λ +µ

·

1

λ

.

Part (c): We want to evaluate Cov(X, M), which we will do by simplifying the deﬁnition

of the covariance to ﬁnd

Cov(X, M) = E[XM] −E[X]E[M] .

To compute E[XM] we can condition on whether the minimum is X or Y . We ﬁnd that

E[XM] = E[XM|M = X]P{M = X} +E[XM|M = Y ]P{M = Y }

=

2

(λ +µ)

2

_

λ

λ +µ

_

+

_

2

(λ +µ)

2

+

1

λ(λ +µ)

_

µ

λ +µ

=

2λ +µ

λ(λ +µ)

2

,

so that with this result Cov(X, M) is then given by

Cov(X, M) =

2λ +µ

λ(λ +µ)

2

−

1

λ +µ

_

1

λ

_

=

1

(λ +µ)

2

.

Exercise 12 (probabilities with three exponential random variables )

Part (a): We can evaluate P{X

1

< X

2

< X

3

} using the deﬁnition of this expression. We

ﬁnd.

P{X

1

< X

2

< X

3

} =

_

∞

x

1

=0

_

∞

x

2

=x

1

_

∞

x

3

=x

2

p(x

1

, x

2

, x

3

)dx

3

dx

2

dx

1

=

_

∞

x

1

=0

_

∞

x

2

=x

1

_

∞

x

3

=x

2

p

X

1

(x

1

)p

X

2

(x

2

)p

X

3

(x

3

)dx

3

dx

2

dx

1

=

_

∞

x

1

=0

p

X

1

(x

1

)

_

∞

x

2

=x

1

p

X

2

(x

2

)

_

∞

x

3

=x

2

p

X

3

(x

3

)dx

3

dx

2

dx

1

,

by the independence of the random variables X

i

. Because X

i

are exponential with failure

rate λ

i

the above becomes

P{X

1

< X

2

< X

3

} =

_

∞

x

1

=0

λ

1

e

−λ

1

x

1

_

∞

x

2

=x

1

λ

2

e

−λ

2

x

2

_

∞

x

3

=x

2

λ

3

e

−λ

3

x

3

dx

3

dx

2

dx

1

.

To compute the above probability We can perform the above integrations one at a time from

the outside inward. The ﬁrst integral is with respect to x

3

, the second with respect to x

2

,

and ﬁnally the third integral is with respect to x

1

. When this is done we obtain

P{X

1

< X

2

< X

3

} =

λ

1

λ

2

(λ

2

+λ

3

)(λ

1

+λ

2

+λ

3

)

.

See the Mathematica ﬁle chap 5 prob 12.nb for this algebra.

Part (b): We want to evaluate P{X

1

< X

2

| max(X

1

, X

2

, X

3

) = X

3

}. Using the deﬁnition

of the conditional expectation we ﬁnd that

P{X

1

< X

2

| max(X

1

, X

2

, X

3

) = X

3

} =

P{X

1

< X

2

, max(X

1

, X

2

, X

3

) = X

3

}

P{max(X

1

, X

2

, X

3

) = X

3

}

=

P{X

1

< X

2

< X

3

}

P{max(X

1

, X

2

, X

3

) = X

3

}

,

since the two events X

1

< X

2

and max(X

1

, X

2

, X

3

) = X

3

imply that X

1

< X

2

< X

3

.

The probability of this event was calculated in Part (a) of this problem. Now the event

max(X

1

, X

2

, X

3

) = X

3

is equivalent to the union of the disjoint events

X

1

< X

2

< X

3

and X

2

< X

1

< X

3

.

Each of these events can be computed using the results from Part (a). Speciﬁcally

P{X

1

< X

2

< X

3

} =

λ

1

λ

2

(λ

2

+λ

3

)(λ

1

+λ

2

+λ

3

)

P{X

2

< X

1

< X

3

} =

λ

1

λ

2

(λ

1

+λ

3

)(λ

1

+λ

2

+λ

3

)

,

so that

P{max(X

1

, X

2

, X

3

) = X

3

} = P{X

1

< X

2

< X

3

} +P{X

2

< X

1

< X

3

} .

Thus using these results we can calculate our desired probability

P{X

1

< X

2

| max(X

1

, X

2

, X

3

) = X

3

} =

λ

1

λ

2

(λ

2

+λ

3

)(λ

1

+λ

2

+λ

3

)

λ

1

λ

2

(λ

2

+λ

3

)(λ

1

+λ

2

+λ

3

)

+

λ

1

λ

2

(λ

1

+λ

3

)(λ

1

+λ

2

+λ

3

)

=

λ

1

+λ

3

λ

1

+λ

2

+ 2λ

3

.

Part (c): We want to evaluate E[max(X

i

)|X

1

< X

2

< X

3

]. From the deﬁnition of condi-

tional expectation we have that (if we deﬁne the random variable M to be M = max(X

i

))

E[max(X

i

)|X

1

< X

2

< X

3

] =

_

∞

0

mp

M

(m|X

1

< X

2

< X

3

)dm.

So that we see that to evaluate this expectation we need to be able to compute the conditional

density p

M

(m|X

1

< X

2

< X

3

). Using the deﬁnition of conditional density we have that

p

M

(m|X

1

< X

2

< X

3

) =

P{M = m, X

1

< X

2

< X

3

}

P{X

1

< X

2

< X

3

}

.

From Part (a) of this problem we know how to compute P{X

1

< X

2

< X

3

}, so our problem

becomes how to calculate the expression P{M = m, X

1

< X

2

< X

3

}. We ﬁnd that

P{M = m, X

1

< X

2

< X

3

} = P{max(X

1

, X

2

, X

3

) = m, X

1

< X

2

< X

3

}

= P{X

3

= m, X

1

< X

2

< X

3

}

= P{X

3

= m, X

1

< X

2

< m}

= P{X

3

= m}P{X

1

< X

2

< m} ,

using the independence of the variables X

i

. Since X

3

is an exponential random variable

we know that P{X

3

= m} = λ

3

e

−λ

3

m

. We can also compute the second probability in the

product above as

P{X

1

< X

2

< m} =

_ _

Ω:{X

1

<X

2

<m}

p

X

1

(x

1

)p

X

2

(x

2

)dx

2

dx

1

=

_

m

x

2

=0

_

x

2

x

1

=0

λ

1

e

−λ

1

x

1

λ

2

e

−λ

2

x

2

dx

1

dx

2

.

We can perform the above integrations one at a time from the outside inward. The ﬁrst

integral is with respect to x

1

and the second is with respect to x

2

. When this is done we

obtain

P{X

1

< X

2

< m} =

λ

1

λ

1

+λ

2

−e

−λ

2

m

+

λ

2

λ

1

+λ

2

e

−(λ

1

+λ

2

)m

.

See the Mathematica ﬁle chap 5 prob 12.nb for this algebra. Now multiplying this ex-

pression by P{X

3

= m} = λ

3

e

−λ

3

m

we have (doing the steps very slowly) that

P{M = m, X

1

< X

2

< X

3

} =

λ

1

λ

1

+λ

2

· λ

3

e

−λ

3

m

−λ

3

e

−(λ

2

+λ

3

)m

+

λ

2

λ

3

λ

1

+λ

2

e

−(λ

1

+λ

2

+λ

3

)m

=

λ

1

λ

1

+λ

2

· λ

3

e

−λ

3

m

−

λ

3

λ

2

+λ

3

(λ

2

+λ

3

)e

−(λ

2

+λ

3

)m

+

λ

2

λ

3

λ

1

+λ

2

+λ

3

(λ

1

+λ

2

+λ

3

)e

−(λ

1

+λ

2

+λ

3

)m

.

The manipulations performed above were to enable us to recognize the above expression as

a linear combination of exponential random variables. We explicitly demonstrate this fact

by introducing normalized exponential random variables in the second line above. With this

representation, taking the expectation of this expression is simple to do. At this point in the

calculation we do not yet need divide by P{X

1

< X

2

< X

3

} since it is just a scalar multiplier

of the above and can be done after the integration. Performing the integration over m in the

above expression and remembering the expression for the mean of a exponentially distributed

random variable we ﬁnd that

E[max(X

i

)|X

1

< X

2

< X

3

] ∝

λ

1

λ

1

+λ

2

·

1

λ

3

−

λ

3

λ

2

+λ

3

·

1

λ

2

+λ

3

+

λ

2

λ

3

λ

1

+λ

2

+λ

3

·

1

λ

1

+λ

2

+λ

3

.

So that when we divide by P{X

1

< X

2

< X

3

} we ﬁnd that we obtain for E[max(X

i

)|X

1

<

X

2

< X

3

] the following

(λ

2

+λ

3

)(λ

1

+λ

2

+λ

3

)

λ

1

λ

2

×

_

λ

1

λ

3

(λ

1

+λ

2

)

−

λ

3

(λ

2

+λ

3

)

2

+

λ

3

(λ

1

+λ

2

+λ

3

)

2

_

.

Which one might be able to simplify further but we stop here.

Part (d): To evaluate E[max(X

1

, X

2

, X

3

)] we will use the mins for maxs identity which is

max(X

1

, X

2

, X

3

) =

3

i=1

X

i

−

i<j

min(X

i

, X

j

) +

i<j<k

min(X

i

, X

j

, X

k

)

= X

1

+X

2

+X

3

−min(X

1

, X

2

) −min(X

1

, X

3

) −min(X

2

, X

3

)

+ min(X

1

, X

2

, X

3

) ,

Sine the X

i

’s are independent exponential distributed random variables the random variables

which are the minimization’s of such random variables are themselves exponential random

variables with failure rates given by the sum of the failure rates of their various components.

Speciﬁcally the random variable min(X

1

, X

2

) is an exponential random variable with failure

rate give by λ

1

+λ

2

. Taking the expectation of this expression we ﬁnd that

E[max(X

1

, X

2

, X

3

)] =

1

λ

1

+

1

λ

2

+

1

λ

3

−

1

λ

1

+λ

2

−

1

λ

1

+λ

3

−

1

λ

2

+λ

3

+

1

λ

1

+λ

2

+λ

3

.

Note that by using the various pieces computed in this problem one could compute E[max(X

i

)]

by conditioning on events like in Part (c). For example

E[max(X

i

)] = E[max(X

i

)|X

1

< X

2

< X

3

]P{X

1

< X

2

< X

3

]}

+ E[max(X

i

)|X

1

< X

3

< X

2

]P{X

1

< X

3

< X

2

]}

+ E[max(X

i

)|X

2

< X

3

< X

1

]P{X

2

< X

3

< X

1

]}

+ · · ·

The two approaches should give the same answer but the second seems more tedious.

Exercise 14 (are we less than c?)

Part (a): We have from the deﬁnition of conditional expectation that

E[X|X < c] =

_

xp

X

(x|X < c)dx.

Now the density function in the above expression is by deﬁnition given by

p

X

(x|X < c) =

P{X = x, X < c}

P{X < c}

.

Now in the above expression the numerator is given by

P{X = x, X < c} = H(−(x −c))p

X

(x) = H(c −x)p

X

(x) .

here H(·) is the Heaviside function and enforces the requirement that X < c. The denomina-

tor is then given by the cumulative distribution function for an exponential random variable

i.e.

P{X < c} = F

X

(c) = 1 −λe

−λc

.

Using both of these expressions the above becomes

_

∞

0

x

_

H(c −x)p

X

(x)

1 −λe

−λc

_

dx =

1

1 −λe

−λc

_

c

0

xλe

−λx

dx

=

1

λ(1 −λe

−λc

)

(1 −e

−λc

−cλe

−λc

) ,

for the requested expression.

Part (b): We now want to ﬁnd the E[X|X < c] by conditioning on E[X]. To begin we

have

E[X] = E[X|X < c]P{X < c} +E[X|X > c]P{X > c} .

Now from the properties of an exponential distribution we know several of these pieces. We

know that

E[X] =

1

λ

P{X < c} = 1 −λe

−λc

P{X > c} = λe

−λc

.

Thus using all of this information in the ﬁrst expression derived in this part of this problem

we ﬁnd

1

λ

= E[X|X < c](1 −λe

−λc

) +E[X|X > c](λe

−λc

) .

Now by the memoryless property of the exponential distribution we have the key observation

that

E[X|X > c] = E[X +c] = E[X] +c =

1

λ

+c ,

so the formula above becomes

1

λ

= E[X|X < c](1 −λe

−λc

) +

_

1

λ

+c

_

(λe

−λc

) ,

which when we solve for E[X|X < c] we ﬁnd that

E[X|X < c] =

1 −(1 +cλ)e

−λc

λ(1 −λe

−λc

)

,

the same expression as in Part (a).

Exercise 18 (mins and maxes of same rate exponentials)

Part (a): Since X

(1)

is deﬁned as X

(1)

= min(X

1

, X

2

) it is an exponential random variable

with rate 2µ so we ﬁnd that

E[X

(1)

] =

1

2µ

.

Part (b): Since X

(1)

is deﬁned as X

(1)

= min(X

1

, X

2

) it is an exponential random variable

with rate 2µ so we ﬁnd that

Var(X

(1)

) =

1

(2µ)

2

=

1

4µ

2

.

Part (c): To compute moments of X

(2)

= max(X

1

, X

2

) recall from the discussion on order

statistics that the density function for the ith order (of n independent random variables with

density/distribution function f/F) is given by

f

X

(i)

(x) =

n!

(n −i)!(i −1)!

f(x)F(x)

i−1

(1 −F(x))

n−i

,

so the distribution of the maximum X

(n)

is given by

f

X

(n)

(x) =

n!

0!(n −1)!

f(x)F(x)

n−1

= nf(x)F(x)

n−1

.

When there are only two independent exponential random variables this becomes

f

X

(2)

(x) = 2µe

−µx

−2µe

−2µx

.

Notice that this is a linear combination of two exponential densities i.e. one with rate µ and

one with rate 2µ. This observation helps in computing various quantities. For example for

the expectation we ﬁnd that

E[X

(2)

] = 2

_

1

µ

_

−

1

2µ

=

3

2µ

.

We note that this problem could also be solved by using the “mins” for “maxes” relationship

which for two variables is given by

max(X

1

, X

2

) = X

1

+X

2

−min(X

1

, X

2

) , (2)

where we know that since X

i

is distributed as an exponential random variable with rate

µ the random variable min(X

1

, X

2

) is an exponential random variable with rate 2µ. Thus

taking the expectation of the above (and using the known value of the expectation for an

exponential random variable) we ﬁnd that

E[max(X

1

, X

2

)] =

1

µ

+

1

µ

−

1

2µ

=

3

2µ

,

as before. The variance calculation using this approach would be more complicated because

it would involve products of terms containing X

1

and min(X

1

, X

2

) which are not independent

and would have to be computed in some way. Finally a third way to compute this expectation

is to recall that by the memoryless property of the exponential distribution that the random

variable X

(2)

is related to the random variable X

(1)

by an “oﬀset” random variable (say A)

in such a way that X

(2)

= X

(1)

+A, where A is an exponential random variable with rate µ

and X

(1)

and A are independent. Thus we then ﬁnd using this method that

E[X

(2)

] = E[X

(1)

] +E[A] =

1

2µ

+

1

µ

=

3

2µ

,

the same as earlier.

Part (d): To compute the variance of X

(2)

we can use

Var(X

(2)

) = E[X

2

(2)

] −E[X

(2)

]

2

,

and we can calculate E[X

2

(2)

] from the density given in Part (c). Remembering that the

second moment of an exponential random variable (say X) with failure rate λ is given by

E[X

2

] =

2

λ

2

,

given the density for X

(2)

derived above we ﬁnd that

E[X

2

(2)

] = 2

_

2

µ

2

_

−

2

(2µ)

2

=

7

2µ

2

.

Using this result we see that Var(X

(2)

) is given by

Var(X

(2)

) =

7

2µ

2

−

9

4

1

µ

2

=

5

4µ

2

.

Another way to compute this variance is to recall that by the memoryless property of the

exponential distribution that the random variable X

(2)

is related to the random variable X

(1)

by an “oﬀset” random variable (say A) in such a way that X

(2)

= X

(1)

+ A, where A is an

exponential random variable with rate µ and X

(1)

and A are independent. Thus we then

ﬁnd using this method that

Var(X

(2)

) = Var(X

(1)

) + Var(A) =

1

(2µ)

2

+

1

µ

2

=

5

4µ

,

the same as before.

Exercise 19 (mins and maxes of diﬀerent rate exponentials)

Part (a): Using much of the discussion from Exercise 18 we have that X

(1)

is an exponential

random variable with rate µ

1

+µ

2

, so that

E[X

(1)

] =

1

µ

1

+µ

2

.

Part (b): As in Part (a) the variance of X

(1)

is simple to calculate and we ﬁnd that

Var(X

(1)

) =

2

(µ

1

+µ

2

)

2

.

Part (c): To compute E[X

(2)

] we condition on whether X

1

< X

2

or not. We have

E[X

(2)

] = E[X

(2)

|X

1

< X

2

]P{X

1

< X

2

} +E[X

(2)

|X

1

> X

2

]P{X

1

> X

2

}

= E[X

(2)

|X

1

< X

2

]

_

λ

1

λ

1

+λ

2

_

+E[X

(2)

|X

1

> X

2

]

_

λ

2

λ

1

+λ

2

_

.

We will now evaluate E[X

(2)

|X

1

< X

2

]. We ﬁrst deﬁne the random variable M = max(X

1

, X

2

)

and use the deﬁnition of conditional expectation. We have

E[X

(2)

|X

1

< X

2

] =

_

∞

0

mp

M

(m|X

1

< X

2

)dm.

The conditional probability distribution p

M

(m|X

1

< X

2

) is given by some simple manipula-

tions. We ﬁnd

p

M

(m|X

1

< X

2

) =

P{M = m, X

1

< X

2

}

P{X

1

< X

2

}

=

P{max(X

1

, X

2

) = m, X

1

< X

2

}

P{X

1

< X

2

}

=

P{X

2

= m, X

1

< X

2

}

P{X

1

< X

2

}

=

P{X

2

= m, X

1

< m}

P{X

1

< X

2

}

=

P{X

2

= m}P{X

1

< m}

P{X

1

< X

2

}

,

where we have used the independence of X

1

and X

2

in the last step. From earlier we know

that the denominator of the above is given by P{X

1

< X

2

} =

λ

1

λ

1

+λ

2

, and the numerator is

given by

P{X

2

= m}P{X

1

< m} = λ

2

e

−λ

2

m

(1 −e

−λ

1

m

) = λ

2

e

−λ

2

m

−λ

2

e

−(λ

1

+λ

2

)m

.

Combining all of this information we ﬁnd that our probability density is given by

p

M

(m|X

1

< X

2

) =

_

λ

1

+λ

2

λ

1

_

(λ

2

e

−λ

2

m

−λ

2

e

−(λ

1

+λ

2

)m

) .

Multiplying by m and integrating we can use the fact that p

M

(m|X

1

< X

2

) is a linear

combination of exponential densities to ﬁnd that

E[X

(2)

|X

1

< X

2

] =

λ

1

+λ

2

λ

1

λ

2

−

λ

2

λ

1

_

1

λ

1

+λ

2

_

.

By the same arguments (eﬀectively exchanging 1 with 2) we ﬁnd that

E[X

(2)

|X

1

> X

2

] =

λ

1

+λ

2

λ

1

λ

2

−

λ

1

λ

2

_

1

λ

1

+λ

2

_

.

So combining these two results we can ﬁnally obtaining E[X

(2)

] as

E[X

(2)

] =

1

λ

1

+

1

λ

2

−

1

λ

1

+λ

2

.

Note that this could also have been obtained (with a lot less work) by using the mins for

max identity given by Equation 2.

Part (d): We want to evaluate Var(X

(2)

), which we will do by evaluating E[X

2

(2)

] −E[X

(2)

]

2

using the results from Part (c) above. As before we can evaluate E[X

2

(2)

] by conditioning on

whether X

1

< X

2

or not. From the expression for p

M

(m|X

1

< X

2

) derived in Part (c) we

ﬁnd that

E[X

2

(2)

|X

1

< X

2

] =

λ

1

+λ

2

λ

1

_

2

λ

2

2

−

λ

2

λ

1

+λ

2

·

2

(λ

1

+λ

2

)

2

_

,

and in the same way we then have

E[X

2

(2)

|X

1

> X

2

] =

λ

1

+λ

2

λ

2

_

2

λ

2

1

−

λ

1

λ

1

+λ

2

·

2

(λ

1

+λ

2

)

2

_

.

So that using the results above, the expectation of X

2

(2)

is given by

E[X

2

(2)

] =

2

λ

2

1

+

2

λ

2

2

−

2

(λ

1

+λ

2

)

2

.

Giving ﬁnally the desired result for Var(X

(2)

) of

Var(X

(2)

) =

2

λ

2

1

+

2

λ

2

2

−

2

(λ

1

+λ

2

)

2

−

_

1

λ

1

+

1

λ

2

−

1

λ

1

+λ

2

_

2

=

1

λ

2

1

+

1

λ

2

2

−

3

(λ

1

+λ

2

)

2

,

after some algebra.

Exercise 21 (waiting time at a two station queue)

For simplicity let X

i

for i = 1, 2 be an exponential random variables with parameter µ

1

,

and Y

i

for i = 1, 2 be exponential random variables with parameter µ

2

. When we enter the

system we will have to ﬁrst wait for the ﬁrst server to ﬁnish. By the memoryless property

of exponential random variables we have that the amount of time we have to wait for this

to happen will be X

1

. At this point we will be at server ones station while the person

ahead of us will be at sever number two’s station. The total time for us to wait until we

move to server number two’s station is given by the maximum amount of time spent at

either server number one or server number two and again using the memory less property of

the exponential distribution is represented as the random variable max(X

2

, Y

1

). Once this

amount of time has completed we will be a the second server where we will have to wait an

additional Y

2

amount of time. Thus if T denotes the random variable representing the total

amount of time we will have to wait we have that

T = X

1

+ max(X

2

, Y

1

) +Y

2

.

To evaluate this we need to compute the distribution function of a random variable A

deﬁned by A = max(X

2

, Y

1

) when X

2

and Y

2

are distributed as above. To do this consider

the distribution function of this random variable where we drop the subscripts on X and Y

we have

F(a) = P{max(X, Y ) ≤ a} =

_ _

Ωa

f(x, y)dxdy

where Ω

a

is the set of points in the XY plane where max(X, Y ) ≤ a. If we imagine a the

random variable X along the x axis of a Cartesian grid and the random variable Y along

the y axis of a Cartesian grid then the set of points where max(X, Y ) ≤ a is 0 ≤ X ≤ a and

0 ≤ Y ≤ a. Thus we can evaluate F(a) (using the independence of X and Y ) as

F(a) =

_

a

0

_

a

0

f(x, y)dxdy

=

_

a

0

_

a

0

µ

1

e

−µ

1

x

µ

2

e

−µ

2

y

dxdy

= µ

1

µ

2

_

a

0

e

−µ

1

x

dx

_

a

0

e

−µ

2

y

dy

= (1 −e

−µ

1

a

)(1 −e

−µ

2

a

)

= 1 −e

−µ

1

a

−e

−µ

2

a

+e

−(µ

1

+µ

2

)a

Thus our density function is then given by

f(a) = µ

1

e

−µ

1

a

+µ

2

e

−µ

2

a

−(µ

1

+µ

2

)e

−(µ

1

+µ

2

)a

.

Showing easily that the expectation of the random variable A is given by

E[A] =

1

µ

1

+

1

µ

2

−

1

µ

1

+µ

2

.

Thus we can now evaluate the expected length of time in the system T as

E[T] = E[X

1

] +E[max(X

2

, Y

1

)] +E[Y

2

]

=

_

1

µ

1

_

+

_

1

µ

1

+

1

µ

2

−

1

µ

1

+µ

2

_

+

_

1

µ

2

_

=

2

µ

1

+

2

µ

2

−

1

µ

1

+µ

2

.

Exercise 22 (more waiting at a two station queue)

In the same way as in the previous problem we will let X

i

be an exponential random variable

with parameter µ

1

and Y

i

be an exponential random variable with parameter µ

2

. Then to

have server one become available we must wait an amount max(X

1

, Y

1

). After this amount of

time you move to server number one. To move to server number two you must wait another

max(X

2

, Y

2

) amount of time. Finally you must wait for server number two to ﬁnish (an

additional amount of time Y

3

) serving you. Thus the total time T spent in the system is

given by

T = max(X

1

, Y

1

) + max(X

2

, Y

2

) +Y

3

.

So the expected time in the station is given by (using the results from the previous problem)

E[T] =

_

1

µ

1

+

1

µ

2

−

1

µ

1

+µ

2

_

+

_

1

µ

1

+

1

µ

2

−

1

µ

1

+µ

2

_

+

_

1

µ

2

_

=

2

µ

1

+

3

µ

2

−

2

µ

1

+µ

2

.

Exercise 23 (ﬂashlight batteries)

By the memory less property of the exponential distribution, when two batteries are being

used simultaneously the lifetime of either battery is given by an exponential distribution

that does not depend on how long the battery has been operational up to that point. Thus

if X and Y are exponential random variables denoting the life expectancy of two batteries

with the same parameter µ the probability that X will fail before Y is given by

P{X < Y } =

_ _

Ω

X<Y

f

X,Y

(x, y)dxdy

=

_

∞

x=0

_

x

y=0

f

X,Y

(x, y)dydx

=

_

∞

x=0

_

x

y=0

f

X,Y

(x, y)dydx

=

_

∞

x=0

_

x

y=0

µe

−µx

µe

−myy

dydx

= µ

_

∞

x=0

e

−µx

(1 −e

−µx

)dx

=

_

∞

x=0

µe

−µx

dx −

1

2

_

∞

x=0

(2µ)e

−2µx

dx

= 1 −

1

2

=

1

2

.

Thus when two batteries are in place due to the memoryless property of the exponential

distribution it is equally likely that either one of them will expire ﬁrst. With this information

we can calculate the various probabilities.

Part (a): P{X = n} is the probability that the last non failed battery is number n. Since

when the n the battery is placed in the ﬂashlight, with probability 1/2 it will last longer

than the other battery. Thus P{X = n} = 1/2.

Part (b): P{X = 1} is the probability that when the last working battery is observed it

happens to be the ﬁrst battery on the battery list. This will happen if the ﬁrst battery

lasts longer than the second battery, the third battery, the fourth batter, etc. up to the nth

battery. Since the event that the ﬁrst battery lasts longer than these n − 1 other batteries

this will happen with probability (1/2)

n−1

, thus

P{X = 1} =

_

1

2

_

n−1

.

Part (c): To evaluate P{X = i} means that when the ith battery is placed in the ﬂashlight

it proceeds to last longer than the remaining I, i + 1, i + 2, · · · , n other batteries. Here I is

the index of the other battery in the ﬂashlight when the ith battery is placed in the ﬂashlight

(1 ≤ I ≤ i −1). Thus the ith battery has n −(i +1) +1 + 1 = n −i +1 comparisons which

it must win. Thus

P{X = i} =

_

1

2

_

n−i+1

2 ≤ i ≤ n

The reason this expression does not work for i = 1 is that in that case there is no battery

“before” this one to compare to. Thus the index in the above expression is one two large for

the case i = 1.

Part (d): The random variable T (representing the entire lifetime of our ﬂashlight) can

be expressed in terms of sums of random variables representing the serial lifetimes of two

individual batteries. For example, the ﬁrst two batteries will produced a functional ﬂashlight

until the ﬁrst one burns out. If we let T

1

and T

2

represent the exponential random variables

representing the lifetime of the ﬁrst and second battery then the ﬂashlight will function for

a random amount of time F

1

= min(T

1

, T

2

). When one of these two batteries burns out

it is immediately replaces. By the memoryless property of the exponential distribution the

battery that is not replaced has a distribution as if it just started at this replacement time.

The next increment of time before the fourth battery is used is given by F

2

= min(T

∗

, T

3

),

where T

∗

is the battery not replaced earlier. Note that each of these minimizations is itself

a random variable with a rate given by 2µ. This process continues n −1 times. Thus if T is

the lifetime of the total ﬂashlight system we see that it is given by

T =

n−1

i=1

F

i

.

Since for each F

i

we have E[F

i

] =

1

2µ

, the expectation of T is given by

E[T] =

n−1

i=1

E[F

i

] =

n −1

2µ

.

Part (e): Since T is the sum of n − 1 exponential random variables with rates 2µ the

distribution of T is given by a gamma distribution with parameters n−1 and 2µ, speciﬁcally

f

T

(t) = 2µe

−2µt

(2µt)

n−2

(n −2)!

.

Exercise 24 (Laplace distributions)

Part (a): We want to show that the given f is a density function. We will integrate and

see if we obtain unity. We have

_

∞

−∞

f(x)dx =

_

0

−∞

λ

2

e

λx

dx +

_

∞

0

λ

2

e

−λx

dx

=

λ

2

_

e

λx

λ

¸

¸

¸

¸

0

−∞

+

λ

2

_

e

−λx

(−λ)

¸

¸

¸

¸

∞

0

=

1

2

(1) −

1

2

(−1) = 1 .

Part (b): The distribution function F(·) is given by

F(x) =

_

x

−∞

1

2

λe

λξ

dξ =

1

2

e

λξ

¸

¸

x

−∞

=

1

2

e

λx

for x < 0 ,

and

F(x) =

1

2

+

_

x

0

1

2

λe

−λξ

dξ =

1

2

+

λ

2

e

λξ

(−λ)

¸

¸

¸

¸

x

0

= 1 −

1

2

e

−λx

for x > 0 ,

Part (c): We are told that X and Y are exponential random variables then we want to

prove that X −Y is given by a Laplace (double exponential) distribution. To do this deﬁne

the random variable Z ≡ X−Y , which has a distribution function given by the convolution

of the random variables X and −Y . This means that

f

Z

(z) =

_

∞

−∞

f

X

(ξ)f

−Y

(z −ξ)dξ .

Now the density f

−Y

(y) is given by

f

−Y

(y) =

_

0 y > 0

λe

λy

y < 0

So to evaluate our convolution given above we begin by letting z < 0 and then since f

−Y

(z −

ξ) = f

−Y

(−(ξ−z)), plotted as a function of ξ this is the function f

−Y

(−ξ) shifted to the right

by z. Plotting the function f

−Y

(−(ξ − z)), for z = −1 and z = +1 we have the ﬁgure ??.

With this understanding if z < 0 our function f

Z

(z) is given by

f

Z

(z) =

_

∞

0

λe

−λξ

· λe

λ(z−ξ)

dξ =

λ

2

e

λz

for z < 0 ,

after some algebra. If z > 0 then f

Z

(z) is given by

f

Z

(z) =

_

∞

z

λe

−λξ

· λe

λ(z−ξ)

dξ =

λ

2

e

−λz

for z < 0 ,

Combining these results we ﬁnd that

f

Z

(z) =

_

λ

2

e

λz

z < 0

λ

2

e

−λz

z > 0

showing that Z is a Laplace or double exponential random variable.

Part (d): Now to show that IX is a Laplace random variable we remember that I = ±1

both with probability 1/2. Deﬁning Y ≡ IX the probability distribution of Y can be found

by conditioning on the value of I. We have

p

Y

(y) = P{IX = y|I = −1}P{I = −1} +P{IX = y|I = +1}P{I = +1}

= P{IX = y|I = −1}

1

2

+P{IX = y|I = +1}

1

2

= P{X = −y|I = −1}

1

2

+P{X = y|I = +1}

1

2

= P{X = −y}

1

2

+P{X = y}

1

2

,

by the independence of X and I. Since

P{X = −y} =

_

0 y > 0

λe

λy

y < 0

we have that p

Y

(y) is given by

p

Y

(y) =

_

1

2

λe

λy

y < 0

1

2

λe

−λy

y > 0

or a Laplace random variable.

Part (e): We deﬁne the random variable W as

W =

_

X I = 1

−Y I = −1

then this probability distribution can be calculated as in a Part (d) of this problem. For we

have

p

W

(w) =

1

2

p

X

(w) +

1

2

p

−Y

(w) =

1

2

H(w)λe

−λw

+

1

2

H(−w)λe

λw

,

with H(·) the Heaviside function. From which we see that the random variable W is a

Laplace random variable.

Exercise 29 (replacement kidneys)

Part (a): From the information given the time till a new kidney’s arrival, T, is distributed

as an exponential random variable with rate λ. Thus as long as A is still living when it

ﬁnally arrives she will receive it. If we denote L

A

as the random variable representing the

lifetime of A then A will be alive if L

A

> T. This event happens with probability

λ

µ

A

+λ

,

with µ

A

the exponential rate of the lifetime of A. This result as derived in the section of the

book entitled “Further Properties of the Exponential Distribution”.

Part (b): For B to receive a new kidney we must have had A expire, and B still alive when

the kidney arrives. Mathematically if we denote L

B

as the random variable the lifetime of

B this is the event that

L

B

> L

A

and L

B

> T .

By the independence of these random variables the probability of this event is

P{L

B

> L

A

}P{L

B

> T} =

_

µ

A

µ

A

+µ

B

__

λ

µ

B

+λ

_

.

Problem 30 (pet lifetimes)

Let L be the random variable denoting the additional lifetime of the surviving pet. We can

compute the expectation of L by conditioning on which pet dies ﬁrst. If we let L

D

and L

C

be random variables denoting the lifetimes of the dog and cat respectively, then the event

that the dog is the ﬁrst do die is mathematically stated as L

D

< L

C

. With these deﬁnitions

we can compute E[L] as

E[L] = E[L|L

D

< L

C

]P{L

D

< L

C

} + E[L|L

D

> L

C

]P{L

D

> L

C

}

= E[L|L

D

< L

C

]

_

λ

d

λ

d

+λ

c

_

+E[L|L

D

> L

C

]

_

λ

c

λ

d

+ λ

c

_

= E[L

C

]

_

λ

d

λ

d

+λ

c

_

+E[L

D

]

_

λ

c

λ

d

+λ

c

_

=

1

λ

c

_

λ

d

λ

d

+λ

c

_

+

1

λ

d

_

λ

c

λ

d

+λ

c

_

=

_

1

λ

d

+λ

c

__

λ

d

λ

c

+

λ

c

λ

d

_

.

In the above we have been able to make the substitution that E[L|L

D

< L

C

] = E[L

C

],

because of the memoryless property of the exponential random variables involved.

Problem 31 (doctors appointments)

If the 1:00 appointment is over before the 1:30 appointment then the amount of time the

1:30 patient spends in the oﬃce is given by an exponential random variable with mean 30

minutes. If the 1:00 appointment runs over (into the 1:30 patients time slot) the 1:30 patient

will spend more time in the oﬃce due to the fact that his appointment cannot start on

time. Thus letting T

1

and T

2

be exponential random variables denoting the length of time

each patient requires with the doctor, the expected time the 1:30 appointment spends at the

doctors oﬃce (L) is given by conditioning on the length of time the 1:00 patient requires

with the doctor. We ﬁnd

E[L] = E[L|T

1

< 30]P{T

1

< 30} +E[L|T

1

> 30]P{T

1

> 30} .

We can compute each of these expressions. We ﬁnd E[L|T

1

< 30] = E[T

2

] = 30 minutes,

P{T

1

< 30} = 1 −e

−(1/30)(30)

= 1 −e

−1

,

and

P{T

1

> 30} = e

−(1/30)(30)

= e

−1

.

To evaluate E[L|T

1

> 30] we can write it as follows.

E[L|T

1

> 30] = E[(T

1

−30) +T

2

|T

1

> 30]

= E[T

1

−30|T

1

> 30] +E[T

2

|T

1

> 30] .

By the memoryless property of the exponential random variable we have that

E[T

1

−30|T

1

> 30] = E[T

1

] = 30 .

Thus we ﬁnd

E[L|T

1

> 30] = 30 +E[T

2

|T

1

> 30] = 30 + 30 = 60 ,

and ﬁnally obtain

E[L] = 30(1 −e

−1

) + 60(e

−1

) = 41.03 ,

minutes for the expected amount of time the 1:30 patient spends at the doctors oﬃce.

Problem 34 (X given X +Y )

Part (a): Let begin by computing f

X|X+Y =c

(x|c). We have that

f

X|X+Y =c

(x|c) = P{X = x|X +Y = c}

=

P{X = x, X +Y = c}

P{X +Y = c}

=

P{X = x, Y = c −x}

P{X +Y = c}

=

P{X = x}P{Y = c −x}

P{X +Y = c}

,

By the independence of X and Y . Now deﬁning the random variable Z as Z = X + Y , we

see that Z has a distribution function given by the appropriate convolution

f

Z

(z) =

_

∞

−∞

f

X

(ξ)f

Y

(z −ξ)dξ ,

so that if z < 0 we have the above equal to

f

Z

(z) =

_

z

0

λe

−λξ

µe

−µ(z−ξ)

=

λµ

λ −µ

(e

−µz

−e

−λz

) ,

when we do the integration. Thus with this result we see that the expression for f

X|X+Y =c

(x|c)

is given by

f

X|X+Y =c

(x|c) =

(λe

−λx

)(µe

−µ(c−x)

)

λµ

λ−µ

(e

−µc

−e

−λc

)

=

(λ −µ)e

−(λ−µ)x

1 −e

−(λ−µ)c

for 0 < x < c .

The book gives a slightly diﬀerent result (which I believe is a typo). I believe the result

above is correct. One can integrate the above over the range (0, c) to verify that this function

integrates to one.

Part (b): From Part (a) of this problem we can integrate this expression to compute

E[X|X +Y = c]. We ﬁnd that

E[X|X +Y = c] =

(λ −µ)

(1 −e

−(λ−µ)c

)

_

c

0

xe

−(λ−µ)x

dx

=

1

λ −µ

−

ce

−(λ−µ)c

1 −e

−(λ−µ)c

.

Part (c): To ﬁnd E[Y |X + Y = c] we can use an idea like that in Exercise 14. Consider

the expression E[X +Y |X +Y = c], which we know must equal c since we are conditioning

on the event X +Y = c. By linearity of the expectation it must also equals

E[X|X +Y = c] +E[Y |X +Y = c] .

So solving for the desired E[Y |X +Y = c] we see that

E[Y |X +Y = c] = c −E[X|X +Y = c]

= c −

1

λ −µ

+

ce

−(λ−µ)c

1 −e

−(λ−µ)c

= c

_

1 −e

−(λ−µ)c

1 −e

−(λ−µ)c

_

−

1

λ −µ

+

ce

−(λ−µ)c

1 −e

−(λ−µ)c

=

c

1 −e

−(λ−µ)c

−

1

λ −µ

.

Problem 35 (equivalent deﬁnitions of a Poisson process)

We are asked to prove the equivalence of two deﬁnitions for a Poisson process. The ﬁrst

deﬁnition (Deﬁnition 5.1 in the book) is the following

The counting process {N(t), t ≥ 0} is said to be a Poisson process if:

i N(0) = 0

ii {N(t), t ≥ 0} has independent increments

iii The number of events in any interval of length t has a Poisson distribution with mean

λt. That is for s, t ≥ 0, we have

P{N(t +s) −N(s) = n} = e

−λt

(λt)

n

n!

n ≥ 0

We want to show that this deﬁnition is equivalent to the following (which is Deﬁnition 5.3

in the book)

i N(0) = 0

ii {N(t), t ≥ 0} has stationary, independent increments

iii P{N(t) ≥ 2} = o(t)

iv P{N(t) = 1} = λt +o(t)

We begin by noting that both deﬁnitions require N(0) = 0. From (ii) in Deﬁnition 5.1

we have the required independent increments needed in Deﬁnition 5.3 (ii). From (iii) in

Deﬁnition 5.1 we have that the distributions of X(t

2

+s) −X(t

1

+s) is given by a Poisson

distribution with mean λ(t

2

− t

1

) and the distribution of random variable X(t

2

) − X(t

1

) is

also given by a Poisson distribution with mean λ(t

2

− t

1

) showing that the process {N(t)}

also has stationary increments and thus satisﬁes the totality of Deﬁnition 5.3 (ii).

From (iii) in Deﬁnition 5.1 we have with s = 0 (and the fact that N(0) = 0) that

P{N(t) = n} =

e

−λt

(λt)

n

n!

.

So that

P{N(t) ≥ 2} =

∞

n=2

e

−λt

(λt)

n

n!

= e

−λt

_

∞

n=0

(λt)

n

n!

−1 −λt

_

= e

−λt

_

e

λt

−1 −λt

¸

= 1 −e

−λt

−λte

−λt

,

which (we claim) is a function that is o(t). To show that this is true consider the limit as t

goes to zero. Thus we want to evaluate

lim

t→0

1 −e

−λt

−λte

−λt

t

.

Since this is an indeterminate limit of type 0/0 we must use L’Hospital’s rule which gives

that the above limit is equal to the limit of the derivative of the top and bottom of the above

or

lim

t→0

λe

−λt

−λe

−λt

+λ

2

te

−λt

1

= λ −λ = 0 .

Proving that this expression is o(t) (since this limit equaling zero is the deﬁnition) and

proving that P{N(t) ≥ 2} = o(t). The ﬁnal condition required for Deﬁnition 5.3 is (iv). We

have from Deﬁnition 5.1 (iii) that

P{N(t) = 1} =

e

−λt

(λt)

1!

= λte

−λt

To show that this expression has the correct limiting behavior as t → 0, we ﬁrst prove that

e

−λt

= 1 −λt +o(t) as t → 0 ,

Which we do by evaluating the limit

lim

t→0

e

−λt

−1 +λt

t

= lim

t→0

−λe

−λt

+λ

1

= −λ +λ = 0 .

Where we have used L’Hospital’s rule again. With this result we see that

P{N(t) = 1} = λt(1 −λt +o(t))

= λt −λ

2

t

2

+o(t

2

)

= λt +o(t) ,

showing the truth of condition (iv) in Deﬁnition 5.3.

Exercise 37 (helping Reb cross the highway)

At the point where Reb wants to cross the highway the number of cars that cross is a Poisson

process with rate λ = 3, the probability that k appear is given by

P{N = k} =

e

−λt

(λt)

k

k!

.

Thus Reb will have no problem if no cars come during her crossing. If her crossing takes s

second this will happen with probability

P{N = 0} = e

−λs

= e

−3s

.

Note that this is the density function for a Poisson random variable (or the cumulative

distribution function of a Poisson random variable with n = 0). This expression is tabulated

for s = 2, 5, 10, 20 seconds in chap 5 prob 37.m.

Exercise 38 (helping a nimble Reb cross the highway)

Following the results from Exercise 29, Reb will cross unhurt, with probability

P{N = 0} +P{N = 1} = e

−λs

+e

−λs

(λs) = e

−3s

+ 3se

−3s

.

Not that this is the cumulative distribution function for a Poisson random variable. This

expression is tabulated for s = 5, 10, 20, 30 seconds in chap 5 prob 38.m.

Exercise 40 (the sum of two Poisson processes)

We will ﬁrst prove that the sum of two Poisson random variables is a Poisson random variable.

Let X and Y be Poisson random variables with parameters λ

1

and λ

2

respectively. We can

evaluate the distribution of X +Y by computing the characteristic function of X +Y . Since

X and Y are independent Poisson random variables the characteristic functions of X +Y is

given by

φ

X+Y

(u) = φ

X

(u)φ

Y

(u)

= e

λ

1

(e

iu

−1)

e

λ

2

(e

iu

−1)

= e

(λ

1

+λ

2

)(e

iu

−1)

.

From the direct connection between characteristic functions to and probability density func-

tions we see that the random variable X + Y is a Poisson random variable with parameter

λ

1

+λ

2

, the sum of the Poisson parameters of the random variables X and Y .

Now for the problem at hand, since N

1

(t) and N

2

(t) are both Poisson random variables with

parameters λ

1

t and λ

2

t respectively, then from the above discussion the random variable

N(t) deﬁned by N

1

(t) + N

2

(t) is a Poisson random variable with parameter λ

1

t + λ

2

t and

thus has a probability of the event N(t) = j given by

P{N(t) = j} =

e

−(λ

1

t+λ

2

t)

(λ

1

t +λ

2

t)

j

j!

=

e

−(λ

1

+λ

2

)t

((λ

1

+λ

2

)t)

j

j!

,

showing that N(t) is a Poisson process with rate λ

1

+λ

2

.

Exercise 41 (the probability that N

1

hits ﬁrst)

For this problem we are asked to evaluate

P{N

1

(t) = 1, N

2

(t) = 0|N

1

(t) +N

2

(t) = 1} ,

which we can do by using the deﬁnition of conditional probabilities as

P{N

1

(t) = 1, N

2

(t) = 0|N

1

(t) +N

2

(t) = 1} =

P{N

1

(t) = 1, N

2

(t) = 0}

P{N

1

(t) +N

2

(t) = 1}

=

P{N

1

(t) = 1}P{N

2

(t) = 0}

P{N

1

(t) +N

2

(t) = 1}

.

In the above we have used the independence of the process N

1

(·) and N

2

(·). The above then

equals

e

−λ

1

t

(λ

1

t)

1

1!

· e

−λ

2

t

e

−(λ

1

+λ

2

)t

((λ

1

+λ

2

)t)

1

1!

=

λ

1

λ

1

+λ

2

.

Exercise 42 (arrival times for a Poisson process)

Part (a): For a Poisson process the inter-arrival times T

i

are distributed as exponential

random variables with parameter λ. Thus the expectation of T

i

is well known since

E[T

i

] =

1

λ

.

With this results we see that the expected value of the fourth arrival time E[S

4

] is given by

E[S

4

] = E[

4

i=1

T

i

] =

4

i=1

E[T

i

] =

4

i=1

1

λ

=

4

λ

.

From this we see that in general we have that

E[S

n

] =

n

λ

.

Part (b): To calculate E[S

4

|N(1) = 2] we will use the memoryless property of the exponen-

tial distribution. Now since we are told that at t = 1 we have seen two of the four events (not

more) from the time t = 1 onward the evaluation of the above expression we only need to

compute the time for two more events to elapse. This last argument relies on the memoryless

property of the exponential distribution, i.e. no matter how much time has elapsed between

the last event seen, the time needed to elapse to the next event is given by an exponential

distribution that is unchanged. Mathematically we have

E[S

4

|N(1) = 2] = 1 +E[S

2

] = 1 +

2

λ

,

using the results from Part (a) of this problem.

Part (c): To calculate E[N(4) − N(2)|N(1) = 3] we will speciﬁcally use the stationary

increments property of the Poisson process. Subtracting one from each N(·) in the above

expectation we ﬁnd that

E[N(4) −N(2)|N(1) = 3] = E[N(3) −N(1)|N(1) = 3]

= E[N(3) −3]

= E[N(3)] −3

= 3λ −3 .

Using the result that E[N(t)] = λt which was shown in the book.

Exercise 50 (waiting for the train)

Part (a): We will solve this problem by conditioning on the time of arrival of the train t.

Speciﬁcally if X is the random variable denoting the number of passengers who get on the

next train the we have

E[X] = E[E[X|T]] ,

where T is the random variable denoting the arrival time of the next train after the previous

train has left the station (we are told that this random variable is uniform). We begin by

computing E[X|T]. Now assuming that the number of people that arrive to wait for the

next train is a Poisson process, the probability that there are n people waiting at the train

stop at time T = t is given by the standard expression for a Poisson process i.e.

P{X = n|T = t} =

e

−λt

(λt)

n

n!

,

where from the problem statement we have λ = 7 (per hour). This expression has an

expected value is given by

E[X|T = t] = λt .

The expectation of X can now be computed by conditioning on the random variable T. We

have

E[X] = E[E[X|T]] = E[λT] = λE[T] =

λ

2

.

Since we know T to be a uniform random variable over (0, 1).

Part (b): To compute the variance of X we will use the conditional variance formula given

by

Var(X) = E[Var(X|T)] + Var(E[X|T]) .

We will compute each term on the right hand side. We begin with the second term:

Var(E[X|T]). Since from the above the random variable E[X|T] is given by λT, the variance

of this expression is related to the variance of T and is given by

Var(E[X|T]) = Var(λT) = λ

2

Var(T) =

λ

2

12

.

Now for the ﬁrst term on the right hand side of the conditional variance expansion (and

the properties of the Poisson distribution), in exactly the same way as we computed the

expectation E[X|T = t], the conditional variance is given by

Var(X|T = t) = λt ,

so that the expectation of this expression then gives

E[Var(X|T)] = λE[T] =

λ

2

.

When we combine these two sub-results we ﬁnally conclude that

Var(X) =

λ

2

+

λ

2

12

.

Exercise 57 (Poisson events)

Part (a): Since our events are generated by a Poisson process we know that

P{N(t) = k} =

e

−λt

(λt)

k

k!

,

so in one hour (t = 1) we have that

P{N(t) = 0} =

e

−λ

(λ)

0

0!

= e

−λ

= e

−2

= 0.1353 ,

Part (b): A Poisson process has the time between events given by exponential random

variables with parameter λ. Thus the fourth event will occur at the time S

4

=

4

i=1

X

i

,

where each X

i

is an exponential random variable with parameter λ. Thus

E[S

4

] =

4

i=1

E[X

i

] =

4

i=1

1

λ

=

4

λ

= 2 .

Since λ is measured in units of reciprocal hours, this corresponds to 2 P.M.

Part (c): The probability that two or more events occur between 6 P.M. and 8 P.M. (a two

hour span) is the complement of the probability that less than two events occur in this two

hour span. This latter probability is given by

P{N(2) = 0} +P{N(2) = 1} =

e

−2λ

(2λ)

0

0!

+

e

−2λ

(2λ)

1

1!

= e

−2λ

+e

−2λ

2λ = 5e

−4

.

Thus the probability we seek is given by 1 −5e

−4

= 0.9084.

Exercise 58 (a Geiger counter)

To solve this problem we will use the following fact. If {N(t), t ≥} is a Poisson process with

rate λ then the process that counts each of the events from the process N(t) with probability

p is another Poisson process with rate pλ.

Now for this speciﬁc problem the arrival of the particles at a rate of three a minute can be

modeled as a Poisson process with a rate λ = 3 per minute. The fact that only 2/3 of the

particles that arrive are actually measured means that this combined event can be modeled

using a Poisson process with a rate

2

3

λ = 2 particles per minute. Thus the process X(t) is a

Poisson process with rate 2 particles per minute. With this information the various parts of

the problem can be solved.

Part (a): In this case we have

P{X(t) = 0} =

e

−2t

(2t)

0

0!

= e

−2t

.

Part (b): In this case we have that

E[X(t)] = 2t .

Exercise 60 (bank arrivals)

Each part of this problem can be solved with the information that given n events have been

observed by time t, the location of any speciﬁc event is uniformly distributed over the time

interval (0, t).

Part (a): We desire that both events land in the ﬁrst 20/60 = 1/3 of the total one hour

time. This will happen with probability

_

1

3

_

2

=

1

9

.

Part (b): In this case we desire that at least one of the events fall in the ﬁrst third of the

total time. This can happen in several ways. The ﬁrst (or second) event falls in this interval

while the other event does not or both events happen in the ﬁrst third of our total interval.

This would give for a probability (reasoning like above) of

1

3

_

2

3

_

+

1

3

_

2

3

_

+

1

3

_

1

3

_

=

5

9

.

We note that the situation we desire the probability for is also the complement of the situation

where both events land in the last 2/3 of time. Using this idea the above probability could

be calculated as

1 −

_

2

3

_

2

=

5

9

.

Problem 81 (the distribution of the event times in a nonhomogenous process)

Part (a): Following the same strategy that the book used to compute the distribution

function for a homogeneous Poisson process, we will begin by assuming an ordered sequence

of arrival times

0 < t

1

< t

2

< · · · < t

n

< t

n+1

= t

and let h

i

be small increments such that t

i

+h

i

< t

i+1

, for i = 1, 2, · · · n. Then the probability

that a random sample of n arrival times S

i

happen at the times t

i

and conditioned on the

fact that we have n arrivals by time t can be computed by considering

P{t

i

≤ S

i

≤ t

i

+h

i

, for i = 1, 2, · · · , n|N(t) = n} .

Which if we deﬁne the event A to be the event that we have exactly one event in [t

i

, t

i

+h

i

]

for i = 1, 2, · · · , n and no events in the other regions then (by deﬁnition) the above equals

the following expression

P{A}

P{N(t) = n}

.

The probability that we have one event in [t

i

, t

i

+h

i

] is given by the fourth property in the

deﬁnition of a nonhomogenous Poisson and is given by

P{N(t

i

+h

i

) −N(t

i

) = 1} = λ(t

i

)h

i

+o(h

i

) (3)

To calculate the probability that we have no events in a given interval, we will derive this

from the four properties in the deﬁnition a nonhomogenous Poisson process. Speciﬁcally,

since the total probability must sum to one we have the constraint on increment variables

over the range of time [t

l

, t

r

],

P{N(t

r

) −N(t

l

) = 0} +P{N(t

r

) −N(t

l

) = 1} +P{N(t

r

) −N(t

l

) ≥ 2} = 1 .

or using properties (ii) and (iii) in the deﬁnition of a nonhomogenous Poisson process the

above becomes (solving for P{N(t

r

) −N(t

l

) = 0}) the following

P{N(t

r

) −N(t

l

) = 0} = 1 −λ(t

l

)(t

r

−t

l

) +o(t

r

−t

l

) . (4)

This result will be used in what follows. To evaluate P{A} we recognized that in the intervals

(0, t

1

), (t

1

+h

1

, t

2

), (t

2

+h

2

, t

3

), · · · , (t

n

+h

n

, t

n+1

) ,

no events occurs, while in the intervals

(t

1

, t

1

+h

1

), (t

2

, t

2

+h

2

), (t

3

, t

3

+h

3

), · · · , (t

n

, t

n

+h

n

) ,

one event occurs. By the independent increments property of nonhomogenous process the

event A can be computed as the product of the probabilities of each of the above intervals

event. The contributed probability P(A

1

) in the evaluation of P{A} from the intervals where

the count increase by one is given by

P(A

1

) =

n

i=1

{λ(t

i

)h

i

+o(h

i

)} =

n

i=1

λ(t

i

)h

i

+o(h) ,

where we have used Eq. 3 and the term o(h) represents terms higher than ﬁrst order in any

of the h

i

’s. By analogy with this result the contributed probability in the evaluation of P{A}

from the intervals where the count does not increase P(A

0

) is given by

P(A

0

) = (1 −λ(0)(t

1

) +o(t

1

))

n

i=1

{1 −λ(t

i

+h

i

)(t

i+1

−t

i

−h

i

) +o(t

i+1

−t

i

−h

i

)} .

This expression will take some manipulations to produce a desired expression. We begin our

sequence of manipulations by following the derivation in the book and recognizing that we

will eventually be taking the limits as h

i

→ 0. Since this expression has a ﬁnite limit we can

take the limit of the above expression as is and simplify some of the notation. Taking the

limit h

i

→ 0 and deﬁning t

0

= 0 the above expression becomes

P(A

0

) =

n

i=0

{1 −λ(t

i

)(t

i+1

−t

i

) +o(t

i+1

−t

i

)} .

We can simplify this product further by observing that the individual linear expressions we

multiply can be written as an exponential which will facilitate our evaluation of this product.

Speciﬁcally, it can be shown (using Taylor series) that

e

−λ(t

i

)(t

i+1

−t

i

)

= 1 −λ(t

i

)(t

i+1

−t

i

) +o(t

i+1

−t

i

) .

With this substitution the product above becomes a sum in the exponential and we have

P(A

0

) =

n

i=0

e

−λ(t

i

)(t

i+1

−t

i

)

= exp

_

−

n

i=0

λ(t

i

)(t

i+1

−t

i

)

_

.

Recognizing the above summation as an approximation to the integral of λ(·), we see that

the above is approximately equal to the following

P(A

0

) = exp

_

−

n

i=0

λ(t

i

)(t

i+1

−t

i

)

_

≈ exp

_

−

_

t

0

λ(τ)dτ

_

= e

−m(t)

.

With these expressions for P(A

1

) and P(A

0

), we can now evaluate our target expression

P{A}

P{N(t) = n}

=

P(A

1

)P(A

0

)

P{N(t) = n}

=

_

n!

e

−m(t)

m(t)

n

_

P(A

1

)P(A

0

)

=

_

n!

e

−m(t)

m(t)

n

_

__

n

i=1

λ(t

i

)h

i

+o(h)

_

e

−m(t)

_

= n!

_

n

i=1

λ(t

i

)

m(t)

h

i

+o(h)

_

.

It is this ﬁnal result we were after. After dividing by

n

i=1

h

i

and taking the limit where

h

i

→ 0, we can conclude that the probability of drawing a speciﬁc sample of n event times

(i.e. obtaining a draw of the random variables S

i

) for a nonhomogenous Poisson process with

rate λ(t) given that we have seen n events by time t is given by

f

S

1

,S

2

,··· ,Sn

(t

1

, t

2

, · · · , t

n

|N(t) = n) = n!

_

n

i=1

λ(t

i

)

m(t)

_

0 < t

1

< t

2

< · · · < t

n

< t (5)

We recognized that this expression is the same distribution as would be obtained for the

order statistics corresponding to n independent random variables uniformly distributed with

probability density function f(·) and a cumulative distribution function F(·) given by

f(x) =

λ(x)

m(t)

and F

′

(x) = f(x) .

By the deﬁnition of the function m(·) we have that λ(x) = m

′

(x), so that an equation for

our cumulative distribution function F is given by

F

′

(x) =

m

′

(x)

m(t)

.

This can be integrated to give

F(x) =

m(x)

m(t)

,

which can only hold if x ≤ t, while if x > t, F(x) = 1. This is the desired result.

Problem 82 (selecting events from an nonhomogenous Poisson process)

Part (a): We claim that the process {N

c

(t), t ≥ 0} is a nonhomogenous process with an

intensity function given by p(t)λ(t).

Part (b): The statement made in Part (a) can be proven by showing that N

c

(t) satisﬁes the

four conditions in the deﬁnition of a nonhomogenous Poisson process with intensity function

p(t)λ(t). We begin, by deﬁning N(t) to be the nonhomogenous Poisson process with intensity

function λ(t). We then have

• Because N(t) is a nonhomogenous Poisson process we have N(0) = 0. Because no

events have occurred at time t = 0 we then must necessarily have N

c

(0) = 0.

• As N(t) has the property of independent increments N

c

(t) will inherit this property

also. This can be reasoned by recognizing that if the increment variables for our original

nonhomogenous Poisson process N(t) are independent then when we select a subset of

these events by considering the process N

c

(t) and involving the probability function

p(·) we cannot introduce dependencies.

• We will compute the expression P{N

c

(t + h) − N

c

(t) = 1}. Note that the desired

event N

c

(t + h) − N

c

(t) = 1 only holds true if our original process has an event

N(t + h) − N(t) = 1. Thus since we desire this probability we begin by deﬁning the

event E

t

to be the event that (as h goes to zero) we count the single event that happens

in the interval (t, t + h) from our original nonhomogenous Poisson process N(t). We

then have conditioning on E

t

that

P{N

c

(t +h) −N

c

(t) = 1} = P{N(t +h) −N(t) = 1|E

t

}P(E

t

)

= P{N(t +h) −N(t) = 1}p(t)

= (λ(t)h +o(h))p(t)

= λ(t)p(t)h +o(h) .

Where we have used independence of E

t

and N(t).

• To compute P{N

c

(t + h) − N

c

(t) > 1} deﬁne E

G

to be the event that in the interval

(t, t +h) we count at least one event. This is the complement of the event we count no

events in this interval. Then we have

P{N

c

(t +h) −N

c

(t) > 1} = P{N(t +h) −N(t) > 1|E

G

}P(E

G

)

= P{N(t +h) −N(t) > 1}P(E

G

)

= o(h)P(E

G

) = o(h) .

Thus we have shown that N

c

(t) satisﬁes the four requirements in the deﬁnition of a nonho-

mogenous Poisson process with intensity function p(t)λ(t).

Problem 83 (time rescaling results in a nonhomogenous Poisson processes)

We will show that N(t) is an nonhomogenous Poisson process by showing that all four of

the required properties hold. We have

• N(0) = N

0

(m(0)) = N

0

_

_

0

0

λ(s)ds

_

= N

0

(0) = 0, since N

0

(t) is a Poisson process.

• To show that N(t) has independent increments consider two non overlapping intervals

of time (t

1

, t

2

) and (t

3

, t

4

). Then the since λ(·) is a non-negative function and the t’s

are ordered as

t

1

< t

2

≤ t

3

< t

4

we have that m(t) =

_

t

0

λ(s)ds is an increasing function of t. Thus our ordering of time

above introduced an ordering of m

i

≡ m(t

i

) and we see

m(t

1

) < m(t

2

) ≤ m(t

3

) < m(t

4

) ,

so the independent increments property of N

0

(t) on the “transformed” intervals (m(t

1

), m(t

2

))

and (m(t

3

), m(t

4

)) imply that same property for N(t).

• Consider

P{N(t +h) −N(t) = 1} = P{N

0

(m(t +h)) −N

0

(m(t)) = 1}

= P{N

0

(m(t) +m

′

(t)h +o(h)) −N

0

(m(t)) = 1}

= 1 · (m

′

(t)h +o(h))

= m

′

(t)h +o(h) .

since m(t) =

_

t

0

λ(s)ds we have that m

′

(t) = λ(t) and the above becomes

P{N(t +h) −N(t) = 1} = λ(t)h +o(h) .

• As in the above we ﬁnd

P{N(t +h) −N(t) > 1} = P{N

0

(m(t) +m

′

(t)h +o(h)) −N

0

(m(t)) > 1}

= o(m

′

(t)h +o(h)) = o(h) .

Thus we have shown that N(t) satisﬁes all four requirements for a nonhomogenous Poisson

process with rate function λ(t) as we were asked to show.

Problem 72 (the departure of cable car riders)

Part (a): By deﬁnition the last rider will depart after all n − 1 other riders have. Lets

assume that we start our cable car with all n passengers at the time t = 0 and deﬁne the

random variables X

i

to be the amount of time between successive stops. This means that we

will stop to let oﬀ the ﬁrst passenger at a time of X

1

, we will stop for the second passenger

to get oﬀ at a time X

1

+X

2

, for the third passenger to get oﬀ at a time of X

1

+X

2

+X

3

, etc.

In all of these cases X

i

is an exponentially distributed random variable with rate λ. Thus

the last or nth passenger gets oﬀ the cable car at a time T given by T =

n

i=1

X

i

. A sum of

n exponentially distributed random variables each with a rate of λ is distributed as Gamma

random variable with parameters (n, λ). That is the distribution function for T is given by

f

T

(t) = λe

−λt

(λt)

n−1

(n −1)!

.

Part (b): To solve this problem introduce the random variables Y

i

to be the time it takes

passenger i to arrive home after he/she is dropped oﬀ. Then for the next to last passenger

(the n − 1st rider) to be home before the last passenger (the nth rider) gets oﬀ requires

Y

n−1

< X

n

. Since Y

n−1

is an exponential random variable with rate µ the probability this is

true is given by

µ

µ+λ

. Given that the n−1 passenger makes it home before the nth passenger

gets oﬀ lets now consider the n−2nd passenger. Using the same logic for the n−2 passenger

to get home before the nth one gets oﬀ requires that

Y

n−2

< X

n

+X

n−1

.

In general, the requirement that the n − kth passenger get home before the nth rider has

departed is

Y

n−k

<

k−1

i=0

X

n−i

for k = 1, 2 · · · n −1 .

Since each Y

i

is an exponential random variable and each sum is a Gamma random variable,

to evaluate the probability of each of these events we need the following lemma.

Lemma: If Y is an exponential random variable with parameter µ and X is a Gamma

random variable with parameters (λ, n), independent of Y , we have

P{Y < X} =

_

∞

0

P{Y < X|X = x}P{X = x}dx

=

_

∞

0

P{Y < x}λe

−λx

(λx)

n−1

(n −1)!

dx

=

_

∞

0

(1 −e

−µx

)λe

−λx

(λx)

n−1

(n −1)!

dx

= 1 −

λ

n

(λ +µ)

n

,

when we perform the required integration.

Combining this result with the discussion above we see that the probability that all the other

riders are home (denoted here as P

h

) at the time the nth gets oﬀ is given by

P

h

=

n−1

k=1

P

_

Y

n−k

<

k−1

i=0

X

n−i

_

= P{Y

n−1

< X

n

}P{Y

n−2

< X

n

+X

n−1

} · · · P{Y

1

< X

n

+X

n−1

+· · · X

2

}

=

_

µ

µ +λ

__

1 −

λ

2

(µ +λ)

2

__

1 −

λ

3

(µ +λ)

3

_

· · ·

_

1 −

λ

n−1

(µ +λ)

n−1

_

=

n−1

k=1

_

1 −

λ

k

(λ +µ)

k

_

.

Problem 73 (shocks to the system)

Part (a): We are told that it takes n shocks to cause our system to fail. Because these

events are from a Poisson process the times between individual events X

i

are exponentially

distributed with a rate λ. Thus the time T can be written T =

n

i=1

X

i

, and we have

P{T = t|N = n} = P{

n

i=1

X

i

= t|N = n} .

Since the distribution of the sum of n exponential random variables is distributed as a

Gamma random variable we have that the above becomes

P{T = t|N = n} = λe

−λt

(λt)

n−1

(n −1)!

.

Part (b): We want to compute P{N = n|T = t} which we can compute using an application

of Bayes’ rule. We have

P{N = n|T = t} =

P{T = t|N = n}P{N = n}

P{T = t}

.

The ﬁrst factor in the numerator was calculated in Part (a). The second factor in the

numerator P{N = n} represents the probability it takes n trails to get one “success” where a

success represents the failure of the system. This is given by a negative-binomial distribution

and in this case is given by p(1 − p)

n−1

. The denominator P{T = t} can be calculated by

conditioning on N as

P{T = t} =

∞

n=1

P{T = t|N = n}P{N = n}

= λpe

−λt

∞

n=1

(λt(1 −p))

n−1

(n −1)!

= λpe

−λt

e

λt(1−p)

= λpe

−λpt

.

Note that this is an exponential distribution with rate λp, which could have been determined

without any calculations as follows. Recall that a Poisson process with rate λ that has its

events “ﬁltered” with a Bernoulli process with parameter p is another Poisson process with

rate λp. Since this second Poisson process represents the process of failures the ﬁrst failure

time will be distributed as an exponential random variable with rate λp exactly as found

above. Using these parts we then have

P{N = n|T = t} =

_

λe

−λt

(λt)

n−1

(n−1)!

_

p(1 −p)

n−1

λpe

−λtp

= e

−λ(1−p)t

(λ(1 −p)t)

n−1

(n −1)!

for n = 1, 2, · · · .

This looks exactly like a Poisson distribution with a mean λ(1 −p)t but starts at the value

n = 1. Deﬁning the random variable N

′

as N

′

≡ N −1, then from the above expression the

distribution of N

′

is given by

P{N

′

= n

′

|T = t} = e

−λ(1−p)t

(λ(1 −p)t)

n

′

(n

′

)!

for n

′

= 0, 1, 2, · · · .

So N

′

is a Poisson random variable with a mean λ(1 − p)t and we have that N is given as

1 +N

′

as we were to show.

Part (c): The desired probability P{N = n|T = t} for n = 1, 2, · · · could have been

determined without calculation using the following argument. The probability P{N = n|T =

t} means that the system failure happened on shock n at the time t. This failure must have

been caused by the last shock and requires that the previous n − 1 shocks did not result

in failure. Since non-failure causing shocks occur according to a Poisson process with rate

λ(1 −p), the number of such non-failure causing shocks, n−1, is given by a Poisson random

variable with mean λ(1 −p)t, exactly as calculated in Part (b).

Problem 93 (min/max identities)

Part (a): To prove the given identity

max(X

1

, X

2

) = X

1

+X

2

−min(X

1

, X

2

) ,

we can appeal to the law of the excluded middle. Which is a fancy way of saying we can

simply consider the two possible cases for the relationship between X

1

and X

2

. For example,

if X

1

≤ X

2

, then the left hand side of the above expression is max(X

1

, X

2

) = X

2

while the

right hand side is X

1

+ X

2

− min(X

1

, X

2

) = X

1

+ X

2

− X

1

= X

2

which are equal. If in

fact the other relationship holds between X

1

and X

2

, that is, X

1

≥ X

2

, a similar calculation

shows that the above identity to be true in this case also.

Part (b): I had a diﬃcult time proving this identity directly but can oﬀer the follow-

ing arguments in support of the given expression when X

i

≥ 0. Consider the expression

max(X

1

, X

2

, · · · , X

n

) in comparison to the expression

n

i=1

X

i

. Since one of the X

i

, say

X

i

∗, is the largest the summation expression will “over count” the maximum by

i=i

∗

X

i

.

That is by all X

i

that are less than X

i

∗ and we have

max(X

1

, X

2

, · · · , X

n

) ≤

n

i=1

X

i

.

We can “adjust” our sum on the right hand side to make it more like the left hand side

by subtracting oﬀ all terms smaller than X

i∗

. Terms that are smaller than X

i

∗ can be

represented as min(X

i

, X

j

) with i < j. We need to include the requirement that i < j since

otherwise considering all possible terms like min(X

i

, X

j

) we would double count all possible

minimums i.e. min(X

i

, X

j

) = min(X

j

, X

i

). Thus we now might consider the following

n

i=1

X

i

−

i<j

min(X

i

, X

j

) .

Note that in the above the second expression now subtracts too many terms since smaller

terms will certainly appear more than once. For example if X

1

= 1, X

2

= 2, and X

3

= 3

then max(X

1

, X

2

, X

3

) = 3, while the above gives

1 + 2 + 3 −(1 + 1 + 2) = 3 −1 = 3 .

Thus the value for X

1

was subtracted oﬀ twice. This logic allows one to conclude that

max(X

1

, X

2

, · · · , X

n

) ≥

n

i=1

X

i

−

i<j

min(X

i

, X

j

) .

Our approximation to the maximum expression can be improved on by adding terms like

min(X

i

, X

j

, X

k

) to give

max(X

1

, X

2

, · · · , X

n

) ≤

n

i=1

X

i

−

i<j

min(X

i

, X

j

) +

i<j<k

min(X

i

, X

j

, X

k

) .

This process is repeated until we add/subtract min(X

1

, X

2

, · · · , X

n

). Whether we add or

subtract depends on whether the number of variables we start with n is even or odd.

If we deﬁne X

i

for i = 1, 2, · · · n to be an indicator random variables denoting whether or

not the event A

i

occurred we see that

max(X

i

, X

j

, · · · , X

k

) = P(A

i

∪ A

j

· · · ∪ A

k

) and

min(X

i

, X

j

, · · · , X

k

) = P(A

i

A

j

· · · A

k

) .

With these expressions and using the max/min identity proven in Part (a) above we obtain

the well known identity that

P (∪

n

i=1

A

i

) =

i

P(A

i

) −

i<j

P(A

i

A

j

) +

i<j<k

P(A

i

A

j

A

k

) +· · · + (−1)

n−1

P(A

1

A

2

· · · A

n

) .

Part (c): There seemed to be two ways to do this part of the problem. To explicitly

use the results from above, let X

i

be the random variable denoting the time when the

ﬁrst event has occurred in the ith Poisson process. In a similar way, let X be a random

variable denoting the time at which and event has occurred in all n Poisson processes. Then

X = max(X

1

, X

2

, · · · , X

n

). We desire to calculate E[X]. From the identity in Part (b) we

have that X can be expressed in term of minimization’s as

X =

i

X

i

−

i<j

min(X

i

, X

j

) +

i<j<k

min(X

i

, X

j

, X

k

) +· · · +(−1)

n−1

min(X

1

, X

2

, · · · , X

n

) .

Taking the expectation of the above expression we ﬁnd that

E[X] =

i

E[X

i

] −

i<j

E[min(X

i

, X

j

)]

+

i<j<k

E[min(X

i

, X

j

, X

k

)] +· · · + (−1)

n−1

E[min(X

1

, X

2

, · · · , X

n

)] .

Using the independence of the X

i

, the fact that they are exponential with a rate λ

i

, and

earlier results from the book about the distribution of minimum of exponential random

variables we ﬁnd

E[X] =

n

i=1

1

λ

i

−

i<j

1

λ

i

+λ

j

+

i<j<k

1

λ

i

+λ

j

+λ

k

+· · · +

(−1)

n−1

λ

1

+λ

2

+· · · +λ

n

. (6)

As another way to solve this problem that results in a quite diﬀerent formula consider

the following where we compute the distribution function for X directly. Using the above

deﬁnitions we ﬁnd

P{X < t} = P{max

i

(X

i

) < t} = P{X

i

< t for all i}

=

n

i=1

(1 −e

−λ

i

t

) .

Now we can compute E[X] if we know P{X > t} as E[X] =

_

∞

0

P{X > t}dt. From this we

have

E[X] =

_

∞

0

_

1 −

n

i=1

(1 −e

−λ

i

t

)

_

dt . (7)

From the above functional form it appears that one should be able to expand the product

in the integrand, perform the integration, and show that it agrees with the Equation 6.

Chapter 5: Additional Exercises

The following are some problems that appeared in diﬀerent editions of this book in the

chapter on exponential distributions and Poisson processes.

Expectations of products of pairs of order statistics

The Problem: Let X

1

, X

2

, · · · , X

n

be independent exponential random variables, each

having rate λ. Also let X

(i)

be the ith smallest of these values i = 1, · · · , n. Find

(a) E[X

(1)

X

(2)

]

(b) E[X

(i)

X

(i+1)

] for i = 1, 2, · · · , n −1.

(c) E[X

(i)

X

(j)

] for i < j.

The Solution: To evaluate the above expectations we will simply use the joint density

of the order statistic X

(i)

and X

(j)

, when i < j and integrate over the appropriate region.

Rather than deriving this density we simply state the result, which can be discussed in [1].

We have

f

X

(i)

,X

(j)

(x

i

, x

j

) =

n!

(i −1)!(j −i −1)!(n −j)!

[F(x

i

)]

i−1

× [F(x

j

) −F(x

i

)]

j−i−1

[1 −F(x

j

)]

n−j

f(x

i

)f(x

j

) for x

i

< x

j

.

When X is an exponential random variable with rate λ we have that our density and distri-

bution functions f and F are given by the following special forms

f

X

(x) =

_

λe

−λx

x ≥ 0

0 x < 0

F

X

(x) = 1 −e

−λx

1 −F

X

(x) = e

−λx

.

So that our density above becomes in this speciﬁc case the following

f

X

(i)

,X

(j)

(x

i

, x

j

) =

λ

2

n!

(i −1)!(j −i −1)!(n −j)!

[1 −e

−λx

i

]

i−1

× [e

−λx

i

−e

−λx

j

]

j−i−1

e

−λ[(n−j+1)x

j

+x

i

]

for x

i

< x

j

.

Part (a): When i = 1 and j = 2 our density function above becomes

f

X

(1)

,X

(2)

(x

1

, x

2

) = λ

2

n(n −1)e

−λ((n−1)x

2

+x

1

)

for x

1

< x

2

.

So that our expectation is then given by

E[X

(1)

X

(2)

] =

_

∞

x

1

=0

_

∞

x

2

=x

1

x

1

x

2

f

X

(1)

,X

(2)

(x

1

, x

2

)dx

2

dx

1

= λ

2

n(n −1)

_

∞

x

1

=0

_

∞

x

2

=x

1

x

1

x

2

e

−λ((n−1)x

2

+x

1

)

dx

2

dx

1

= λ

2

n(n −1)

_

∞

x

1

=0

x

1

e

−λx

1

_

∞

x

2

=x

1

x

2

e

−λ(n−1)x

2

dx

2

dx

1

=

3n −2

(n −1)

_

1

nλ

_

2

,

See the Mathematica ﬁle chap 5 eprob pp.nb for the algebra for this problem. In addition,

for a Monte-Carlo veriﬁcation of the above see the Matlab ﬁle exp pp order stats pt a.m

and Figure 1.

n index

la

m

b

d

a

MC est. of the expectation: nSamples=50000

10 15 20 25

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

0.1

50

100

150

200

250

300

10

15

20

25

0

0.02

0.04

0.06

0.08

0.1

0

50

100

150

200

250

300

350

n index

The exact expectation

lambda

n index

la

m

b

d

a

Diff between the exact and approx: nSamples=50000

10 15 20 25

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

0.1

0.5

1

1.5

2

2.5

3

3.5

Figure 1: A comparison between a Monte-Carlo simulation of E[x

(1)

x

(2)

] and the derived

analytical expression. Left: the Monte-Carlo simulation plotted as a function of λ and n.

Middle: The analytical solution derived above. Right: The error between the two.

Part (b): When j = i + 1 our joint density function becomes

f

X

(i)

,X

(i+1)

(x

i

, x

i+1

) =

λ

2

n!

(i −1)!(n −i −1)!

(1 −e

−λx

i

)

i−1

e

−λ[(n−i)x

i+1

+x

i

]

for x

i

< x

i+1

.

So that our expectation is then given by

E[X

(i)

X

(i+1)

] =

_

∞

x

i

=0

_

∞

x

i+1

=x

i

x

i

x

i+1

f

X

(i)

,X

(i+1)

(x

i

, x

i+1

)dx

i+1

dx

i

=

λ

2

n!

(i −1)!(n −i −1)!

×

_

∞

x

i

=0

_

∞

x

i+1

=x

i

x

i

x

i+1

(1 −e

−λx

i

)

i−1

e

−λ((n−i)x

i+1

+x

i

)

dx

i+1

dx

i

=

λ

2

n!

(i −1)!(n −i −1)!

×

_

∞

x

i

=0

x

i

(1 −e

−λx

i

)

i−1

e

−λx

i

_

∞

x

i+1

=x

i

x

i+1

e

−λ(n−i)x

i+1

dx

i+1

dx

i

.

At this point even performing algebra by a computer becomes diﬃcult. See the Mathematica

ﬁle chap 5 eprob pp.nb for some of the algebra. A literature search reveals that this is a

rather diﬃcult problem to solve exactly, so instead I’ll use a Monte-Carlo calculation to

compute these expressions. See the Matlab ﬁle exp pp order stats pt b.m and Figure 2

for the computational experiment done for this part of the problem. Since we expect the

expression to scale as 1/λ

2

, we have assumed λ = 1 for the plots presented.

Part (c): As in Part (b) an exact analytic solution is quite diﬃcult so in the Matlab ﬁle

exp pp order stats pt c.m and Figure 3, I have presented the results from the Monte-Carlo

calculation done to evaluate the requested expression. Again since we expect the results to

scale as 1/λ

2

, we have assumed that λ = 1.

0 2 4 6 8 10 12 14

0

1

2

3

4

5

6

7

8

9

i index

E

[

x

i

x

i+

1

]

The appoximate solution

Figure 2: The Monte-Carlo simulation of E[x

(i)

x

(i+1)

] as a function of i for i = 1, 2, · · · , n−1.

Figure 3: The Monte-Carlo simulation of E[x

(i)

x

(j)

] as a function of i and j for i < j.

A Identity with Reciprocal Sums

The Problem: Argue that if λ

i

, i = 1, 2, · · · , n are distinct positive numbers then

n

i=1

1

λ

i

j=i

λ

j

λ

j

−λ

i

=

n

i=1

1

λ

i

.

Hint: Relate this problem to Section 5.2.4.

The Solution: From Section 5.2.4 we know that the random variable X deﬁned as the

sum of n exponential random variables X =

n

i=1

X

i

is called a hypo-exponential random

variable and has a probability density function given by

f

X

1

+X

2

+···+Xn

(x) =

n

i=1

j=i

λ

j

λ

j

−λ

i

λ

i

e

−λ

i

t

.

Since X is the sum of n exponential random variables when we multiply both sides of the

above by x and integrate the left hand side becomes the expectation of random variable

X =

n

i=1

X

i

, and equals

n

i=1

1

λ

i

.

While the right hand side results in a another expectation of sums of exponential random

variables. Taking these expectations we have an expression given by

n

i=1

1

λ

i

j=i

λ

j

λ

j

−λ

i

.

When we equate these two results we have the requested theorem.

We can derive the fact that P{N(t) = 1} = λt +o(t)

The Problem: Show that assumption (iv) of Deﬁnition 5.3 follows from assumptions (ii)

and (iii).

Hint: Derive a functional equation for g(t) = P{N(t) = 0}.

The Solution: Following the hint for this problem we will try to derive a functional rela-

tionship for P{N(t) = 0}, by considering P{N(t +s) = 0}. Now if N(t +s) = 0, this event

is equivalent to the event that N(t) = 0 and N(t +s) −N(t) = 0. so we have that

P{N(t +s) = 0} = P{N(t) = 0, N(t +s) −N(t) = 0}

= P{N(t) −N(0) = 0, N(t +s) −N(t) = 0}

= P{N(t) −N(0) = 0}P{N(t +s) −N(t) = 0}

= P{N(t) = 0}P{N(s) = 0} .

When we used the property of stationary independent increments. Thus deﬁning f(t) ≡

P{N(t) = 0}, from the above we see that f satisﬁes

f(t +s) = f(t)f(s) .

By the discussion in the book the unique continuous solution to this equation is f(t) = e

−λt

,

for some λ. Thus we have that P{N(t) = 0} = e

−λt

. Using (iii) from Deﬁnition 5.3 and the

fact that probabilities must be normalized (sum to one) we have that

P{N(t) = 0} +P{N(t) = 1} +P{N(t) ≥ 2} = 1 .

which gives us (solving for P{N(t) = 1}) the following

P{N(t) = 1} = 1 −P{N(t) = 0} −P{N(t) ≥ 2}

= 1 −e

−λt

−o(t)

= 1 −(1 −λt +o(t)) −o(t)

= λt +o(t) ,

as we were to show.

Cars v.s. Vans on the Highway

The Problem: Cars pass a point on the highway at a Poisson rate of one per minute. If 5

percent of the cars on the road are vans, then

• what is the probability that at least one van passes during the hour?

• given that ten vans have passed by in an hour, what is the expected number of cars to

have passes by in that time?

• if 50 cars have passed by in an hour, what is the probability that ﬁve of them are vans?

The Solution:

Part (a): The number of vans that have passed by in time N

v

(t) is given by a Poisson

process with rate pλ, where p is the fraction of the cars that are vans and λ is the Poisson

rate of the cars (we are told these values are λ = 1 (per minute) and p = 0.05). Because

of this fact the probability that at least one van passes by in one hour t = 60 (minutes) is

given by

P{N

v

(60) ≥ 1} = 1 −P{N

v

(60) = 0}

= 1 −e

−pλ60

= 1 −e

−0.05·1·60

= 0.9502 .

Part (b): We can think of the question posed: how many cars need to pass by to given ten

vans as equivalent to the statement of the expected number of trials will we need to perform

to guarantee ten successes, where the probability of success is given by p = 0.05. This is the

deﬁnition of a negative binomial random variable. If we deﬁne the random variable X to be

the number of trials needed to get r success where each success occurs with probability p we

have that

E[X] =

r

p

.

For this part of the problem at hand (r, p) = (10, 0.05), so the expected number of cars that

must have passed is given by

10

0.05

= 200 .

Part (c): If 50 cars have passed by and we have a 5 percent chance that each car is a

van, the probability that n from 50 are vans is a binomial random variable with parameters

(n, p) = (50, 0.05), so the answer to the question posed is given by

_

50

5

_

(0.05)

5

(0.95)

45

= 0.06584 .

References

[1] S. Ross. A First Course in Probability. Macmillan, 3rd edition, 1988.

1 2 3 4 5 6

1 2 3 4 5 6 7

2 3 4 5 6 7 8

3 4 5 6 7 8 9

4 5 6 5 6 7 6 7 8 7 8 9 8 9 10 9 10 11 10 11 12

Table 1: The possible values for the sum of the values when two die are rolled. Exercise 10 (Boole’s inequality) We begin by decomposing the countable union of sets Ai A1 ∪ A2 ∪ A3 . . . into a countable union of disjoint sets Cj . Deﬁne these disjoint sets as C1 C2 C3 C4 Cj Then by construction A1 ∪ A2 ∪ A3 · · · = C1 ∪ C2 ∪ C3 · · · , and the Cj ’s are disjoint, so that we have Pr(A1 ∪ A2 ∪ A3 ∪ · · · ) = Pr(C1 ∪ C2 ∪ C3 ∪ · · · ) =

j

= = = = . . . =

A1 A2 \A1 A3 \(A1 ∪ A2 ) A4 \(A1 ∪ A2 ∪ A3 ) Aj \(A1 ∪ A2 ∪ A3 ∪ · · · ∪ Aj−1 )

Pr(Cj ) .

**Since Pr(Cj ) ≤ Pr(Aj ), for each j, this sum is bounded above by Pr(Aj ) ,
**

j

Problem 11 (the probability the sum of the die is i) We can explicitly enumerate these probabilities by counting the number of times each occurrence happens, in Table 1 we have placed the sum of the two die in the center of each square. Then by counting the number of squares where are sum equals each number from

two to twelve, we have P2 = P3 = P4 = P5 = P6 = 1 , 36 2 = 36 3 = 36 4 = 36 5 , 36 P7 = 1 , 18 1 , 12 1 , 9 P11 6 1 = 36 6 5 P8 = 36 4 1 P9 = = 36 9 3 1 P10 = = 36 12 2 1 1 = = , P12 = . 36 18 36

Problem 13 (winning at craps) From Problem 11 we have computed the individual probabilities for various sum of two random die. Following the hint, let Ei be the event that the initial die sum to i and that the player wins. We can compute some of these probabilities immediately P (E2 ) = P (E3 ) = P (E12 ) = 0, and P (E7 ) = P (E11 ) = 1. We now need to compute P (Ei ) for i = 4, 5, 6, 8, 9, 10. Again following the hint deﬁne Ei,n to be the event that the player initial sum is i and wins on the n-th subsequent roll. Then

∞

P (Ei ) =

n=1

P (Ei,n ) ,

since if we win, it must be either on the ﬁrst, or second, or third, etc roll after the initial roll. We now need to calculate the P (Ei,n ) probabilities for each n. As an example of this calculation ﬁrst lets compute P (E4,n ) which means that we initially roll a sum of four and the player wins on the n-th subsequent roll. We will win if we roll a sum of a four or loose if we roll a sum of a seven, while if roll anything else we continue, so to win when n = 1 we see that 1 1+1+1 = , P (E4,1) = 36 12 since to get a sum of four we can roll pairs consisting of (1, 3), (2, 2), and (3, 1). To compute P (E4,2 ) the rules of craps state that we will win if a sum of four comes up (with 1 6 probability 12 ) and loose if a sum of a seven comes up (with probability 36 = 1 ) and continue 6 playing if anything else is rolled. This last event (continued play) happens with probability 1− 1 3 1 − = . 12 6 4

1 1 Thus P (E4,2 ) = 3 12 = 16 . Here the ﬁrst 3 is the probability we don’t roll a four or a 4 4 1 seven on the n = 1 roll and the second 12 comes from rolling a sum of a four on the second roll (where n = 2). In the same way we have for P (E4,3 ) the following

P (E4,3 ) =

3 4

2

1 . 12

Here the ﬁrst two factors of 3 are from the two rolls that “keep us in the game”, and the 4 1 factor of 12 , is the roll that allows us to win. Continuing in this in this manner we see that P (E4,4 ) = and in general we ﬁnd that P (E4,n ) = 3 4

n−1

3 4

3

1 , 12

1 12

for n ≥ 1 .

To compute P (Ei,n ) for other i, the derivations just performed, only change in the probabilities required to roll the initial sum. We thus ﬁnd that for other initial rolls (heavily using the results of Problem 24) that 1 P (E5,n ) = 9 1 1 1− − 9 6

n−1

1 = 9

13 18

n−1

5 1 5 1− − P (E6,n ) = 36 36 6 P (E8,n ) = 5 1 5 1− − 36 36 6 1 1 1− − 9 6

n−1

5 = 36 = 5 36 13 18

25 36 25 36

n−1

n−1

n−1

n−1

1 P (E9,n ) = 9

n−1

1 = 9

1 1 1 P (E10,n ) = 1− − 12 12 6

n−1

1 = 12

3 4

n−1

.

**To compute P (E4 ) we need to sum the results above. We have that P (E4 ) = = 1 12 3 4
**

3 4 n−1

= = 1 . 3

n≥1

1 12

n≥0

3 4

n

1 1 12 1 −

Note that this also gives the probability for P (E10 ). For P (E5 ) we ﬁnd P (E5 ) = 2 , which 5 5 also equals P (E9 ). For P (E6 ) we ﬁnd that P (E6 ) = 11 , which also equals P (E8 ). Then our probability of winning craps is given by summing all of the above probabilities weighted by the associated priors of rolling the given initial roll. We ﬁnd by deﬁning Ii to be the event that the initial roll is i and W the event that we win at craps that P (W ) = 0 P (I2 ) + 0 P (I3) + 4 5 1 P (I4 ) + P (I5) + P (I6 ) 3 9 9 5 4 1 + 1 P (I7 ) + P (I8 ) + P (I9 ) + P (I10 ) + 1 P (I11) + 0 P (I12 ) . 9 9 3

Using the results of Exercise 25 to evaluate P (Ii) for each i we ﬁnd that the above summation gives 244 = 0.49292 . P (W ) = 495 These calculations are performed in the Matlab ﬁle chap 1 prob 13.m.

Exercise 15 (some set identities) We want to prove that E = (E ∩ F ) ∪ (E ∩ F c ). We will do this using the standard proof where we show that each set in the above is a subset of the other. We begin with x ∈ E. Then if x ∈ F , x will certainly be in E ∩ F , while if x ∈ F then x will be in E ∩ F c . Thus / in either case (x ∈ F or x ∈ F ) x will be in the set (E ∩ F ) ∪ (E ∩ F c ). / If x ∈ (E ∩ F ) ∪ (E ∩ F c ) then x is in either E ∩ F , E ∩ F c , or both by the deﬁnition of the union operation. Now x cannot be in both sets or else it would simultaneously be in F and F c , so x must be in one of the two sets only. Being in either set means that x ∈ E and we have that the set (E ∩ F ) ∪ (E ∩ F c ) is a subset of E. Since each side is a subset of the other we have shown set equality. To prove that E ∪ F = E ∪ (E c ∩ F ), we will begin by letting x ∈ E ∪ F , thus x is an element of E or an element of F or of both. If x is in E at all then it is in the set E ∪ (E c ∩ F ). If x ∈ E then it must be in F to be in E ∪ F and it will therefore be in E c ∩ F . Again both / sides are subsets of the other and we have shown set equality.

Exercise 23 (conditioning on a chain of events) This result follows for the two set case P {A ∩ B} = P {A|B}P {B} by grouping the sequence of Ei ’s in the appropriate manner. For example by grouping the intersection as E1 ∩ E2 ∩ · · · ∩ En−1 ∩ En = (E1 ∩ E2 ∩ · · · ∩ En−1 ) ∩ En we can apply the two set result to obtain P {E1 ∩ E2 ∩ · · · ∩ En−1 ∩ En } = P {En |E1 ∩ E2 ∩ · · · ∩ En−1 } P {E1 ∩ E2 ∩ · · · ∩ En−1 } . Continuing now to peal En−1 from the set E1 ∩E2 ∩· · ·∩En−1 we have the second probability above equal to P {E1 ∩ E2 ∩ · · · ∩ En−2 ∩ En−1 } = P {En−1 |E1 ∩ E2 ∩ · · · ∩ En−2 }P {E1 ∩ E2 ∩ · · · ∩ En−2 } . Continuing to peal oﬀ terms from the back we eventually obtain the requested expression i.e. P {E1 ∩ E2 ∩ · · · ∩ En−1 ∩ En } = × × . . . × × × P {En |E1 ∩ E2 ∩ · · · ∩ En−1 } P {En−1 |E1 ∩ E2 ∩ · · · ∩ En−2 } P {En−2 |E1 ∩ E2 ∩ · · · ∩ En−3 } P {E3 |E1 ∩ E2 } P {E2 |E1 } P {E1 } .

Gc ) P (B c . Gc |H). Gc |H) = So the total result becomes P (B|H) = p1 p1 p2 + p1 (1 − p2 ) = . G) = P (H) P (H) Now P (H) = (1 − p1 )p2 + p1 (1 − p2 ) + p1 p2 so the above probability becomes p1 p2 p1 p2 = . (1 − p1 )p2 + p1 (1 − p2 ) + p1 p2 p1 + p2 − p1 p2 Part (b): We desire to compute P (B|H) which equals P (B|H) = P (B.Exercise 30 (target shooting with Bill and George) warning! not ﬁnished. (1 − p1 )p2 + p1 (1 − p2 ) + p1 p2 p1 + p2 − p1 p2 p1 (1 − p2 ) . G|H) has already been computed we only need to compute P (B. The S can take on the values m for male and f for female. (1 − p1 )p2 + p1 (1 − p2 ) + p1 p2 Exercise 33 (independence in class) Let S be a random variable denoting the sex of the randomly selected person. Since the ﬁrst term P (B. G|H) + P (B. G) = = = = (1 − p1 )(1 − p2 ) (1 − p1 )p2 p1 (1 − p2 ) p1 p2 . The C can take on the values f for freshman and s for sophomore.. H) P (B. G. Let C be a random variable representing denoting the class of the chosen student. G|H) which equals P (B. G|H) = P (B. by either Bill or George’s shot. Then the outcome of the experiment where both George and Bill ﬁre at the target (assuming that their shots work independently is) P (B c . Gc ) P (B.. G) P (B. Let H be the event that the duck is “hit”. Part (a): We desire to compute P (B. Gc |H) . We want to select the number of sophomore girls such that the random . Let B and G be the events that Bill (respectively George) hit the target. As before we ﬁnd it to be P (B.

Now one should check that this value of n works for all other equalities that must be true. for example one needs to check that when n = 9 the following are true P (S = m. C = f ) = 16 + n n P (S = f. C = s) = . C = f ) = P (S = f )P (C = f ) P (S = f. Then counting up the number of students that satisfy each requirement we have P (S = m) = 10 16 + n 6+n P (S = f ) = 16 + n 10 P (C = f ) = 16 + n 6+n P (C = s) = . C = f ) = 4 16 + n 6 P (S = m. C = f ) we have that n must satisfy P (S = m. 12 . 16 + n The joint density can also be computed and are given by P (S = m. Considering the point case where (S = m. P (X1 ) = P (X2 ) = 1/2 so P (B) = 1 2 1 2 + 1 2 2 3 = 7 . S) = P (S)P (C) for all possible C and S values. Then to calculate P (B) we will condition on the box drawn from as P (B) = P (B|X1 )P (X1) + P (B|X2)P (X2 ) . n = 9 is the correct answer. 16 + n Then to be independent we must have P (C. Let n denote the number of sophomore girls. Exercise 36 (boxes with marbles) Let B be the event that the drawn ball is black and let X1 (X2 ) be the event that we select the ﬁrst (second) box. C = f ) = P (S = m)P (C = f ) 4 10 10 = 16 + n 16 + n 16 + n which when we solve for n gives n = 9. C = s) = P (S = f )P (C = s) . C = s) = 16 + n 6 P (S = f. C = s) = P (S = m)P (C = s) P (S = f. Now P (B|X1 ) = 1/2.variables S and C are independent. P (B|X2) = 2/3. As these can be shown to be true.

P (B c |X1 )P (X1 ) + P (B c |X2 )P (X2 ) 5 Problem 40 (gambling with a fair coin) Let F denote the event that the gambler is observing results from a fair coin. O1 )P (F |O1) P (O2|F. O1 ) P (O3 |F. O1)P (F |O2. Verifying what we know must be true. Part (a): We desire to compute P (F |O1) or the probability we are looking at a fair coin given the ﬁrst observation. We will assume that before any observations are made the probability that we have selected the fair coin is 1/2.e event B c has happened) we now want to compute that it was drawn from the ﬁrst box i.e.Exercise 37 (observing a white marble) If we see that the ball is white (i. and O3 denote the three observations made during our experiment. P (X1|B c ) = P (B c |X1 )P (X1) 3 = . O2. This can be computed using Bayes’ theorem. O2 .e. it is not black i. O1 )P (F c|O2 . Speciﬁcally we have P (F |O3. O1 ) + P (O3|F c . We have P (F |O1) = P (O1 |F )P (F ) P (O1|F )P (F ) + P (O1|F c )P (F c ) 1 1 1 = 1 12 2 1 = . This result can also be obtained using Bayes’ theorem as we have in the other two parts of this problem. O1 ) = P (O2|F. O1) = = P (O3 |F. Also let O1 . 5 +1 3 2 3 Part (c): In this case because the two-headed coin cannot land tails we can immediately conclude that we have selected the fair coin. . O2 . O2. O1)P (F |O2. 3 +1 2 2 2 Part (b): With the second observation and using the “posteriori’s become priors” during a recursive update we now have P (F |O2. O1)P (F |O1) + P (O2|F c . O1)P (F c |O1 ) 1 1 1 = 1 12 3 2 = . O1 ) 1 1 2 5 1 1 + 2 5 0 = 1. O2 .

If prisoner A is told that B (or C) is to be set free then we need to compute P (A|B c ). and C are the events that prisoner A. or C is to be executed respectively. Before asking his question the probability of event A (A is executed) is P (A) = 1/3. . 3 3 3 P (B c |A)P (A) . Where A.Problem 46 (a prisoners’ dilemma) I will argue that the jailers reasoning is sound. 2/3 2 3 1 2 1 +0+ = . Now from Bayes’ rule P (A|B c ) = We have that P (B c ) is given by P (B c ) = P (B c |A)P (A) + P (B c |B)P (B) + P (B c |C)P (C) = So the above probability then becomes P (A|B c ) = 1(1/3) 1 1 = > . B. P (B c ) Thus the probability that prisoner A will be executed has increased as claimed by the jailer. B.

Since P (k+1) = P (1) P (k) by matrix multiplication we have that P (k+1) = P (1) P (k) = = = = p 1−p 1−p p 1 2 1 2 1 + 2 (2p − 1)k − 1 (2p − 1)k 2 1 2 1 2 1 − 2 (2p − 1)k 1 + 2 (2p − 1)k p p p + 2 (2p − 1)k + 1−p − 1−p (2p − 1)k 2 − p (2p − 1)k + 1−p + 1−p (2p 2 2 2 2 2 2 1−p p p + 1−p (2p − 1)k + p − 2 (2p − 1)k 1−p − 1−p (2p − 1)k + 2 + p (2p 2 2 2 2 2 2 p 1−p 1 1 1 k k + 2 (p − (1 − p))(2p − 1) + − 2 + 2 (2p − 1) 2 2 1 1 1 + 1 (1 − p − p)(2p − 1)k 2 + 2 (−(1 − p) + p)(2p − 1)k 2 2 1 + 1 (2p − 1)k+1 1 − 1 (2p − 1)k+1 2 2 2 2 . Evaluating the above expression for n = 1 we ﬁnd P (1) = 1 2 1 2 1 + 2 (2p − 1) − 1 (2p − 1) 2 1 2 1 2 − 1 (2p − 1) 2 + 1 (2p − 1) 2 = p 1−p 1−p p . (1) We will do this by mathematical induction. We begin by verifying that the above formula is valid for n = 1. We desire to prove that P (n) is given as P (n) = 1 2 1 2 1 + 2 (2p − 1)n 1 − 2 (2p − 1)n 1 2 1 2 − 1 (2p − 1)n 2 1 + 2 (2p − 1)n . as required. 1 1 1 − 2 (2p − 1)k+1 2 + 1 (2p − 1)k+1 2 2 − 1)k − 1)k which is the desired result for n = k + 1. .Chapter 4: Markov Chains Chapter 4: Exercises Exercise 6 (an analytic calculation of P (n) ) p 1−p . by matrix multiplication we 1−p p p2 + (1 − p)2 2p(1 − p) 2 2p(1 − p) p + (1 − p)2 Given the transition probability matrix P = see that P (2) is given as P (2) = p 1−p 1−p p p 1−p 1−p p = . Next we assume the relationship in Equation 1 is true for all n ≤ k and we desire to show that it is true for n = k + 1.

fT (t). We have 1/2 1 − P {T < 1/2} = 1 − 0 2e−2t dt 1/2 2e−2t −2 0 = 1 + e−1 − 1 = e−1 . µ µ We sum to six to allow the teller to service the ﬁve original customers and then you. the fact that the repair is still going after 12 hours is irrelevant. Taking the expectation of the variable T we have 6 6 E[T ] = i=1 E[Xi ] = i=1 1 6 = . = 1− Part (b): Since the exponential distribution has no memory. Thus we only need to compute the probability that the repair will last at least 1/2 more. We will have to wait until an amount of time given by 6 T = i=1 Xi . where Xi are independent exponential random variables with rate µ. This probability is the same as that calculated in Part (a) of this problem and is equal to e−1 . This means that the distribution function of T . is given by fT (t) = 2e−2t t ≥ 0 0 t<0 Part (a): Now for the repair time to take longer than 1/2 an hour will happen with probability that is the complement of the probability that it will take less than 1/2 of an hour. . Exercise 2 (the expected bank waiting time) By the memoryless property of the exponential distribution now the fact that one person is being served makes no diﬀerence.Chapter 5: The Exponential Distribution and the Poisson Process Chapter 5: Exercises Exercise 1 (exponential repair times) We are told that T is distributed with a exponential probability distribution function with mean 1/2.

from X 2 to (X + 1)2 . X > 1) p(X > 1) p(X = ξ. 0 where X is an exponential random variable with rate parameter λ. ∞ and this function enforces the constraint that X > 1. no restriction on the random variable X). This can be performed as long as we “shift” the expectation’s argument accordingly i.Exercise 3 (expectations with exponential random variables) I will argue that E[X 2 |X > 1] = E[(X +1)2 ]. Thus there is no way for C to get served before A ﬁnishes. This is the expression we argued at the beginning of this problem should hold true. We can prove that this result is correct by explicitly evaluating the original expectation. Exercise 4 (the post oﬃce) Part (a): In this case. e−λ Here H(·) is the Heaviside function deﬁned as H(x) = 0 x<0 1 x > 0. Removing the conditional expectation is equivalent to “starting” the process at x = 0. X > 1) = 1 − p(X < 1) p(X = ξ. . X > 1) = e−λ −λξ λe H(ξ − 1) = = λe−λ(ξ−1) H(ξ − 1) . so that du = dξ the above becomes ∞ (u + 1)2 λe−λu du = E[(X + 1)2 ] . it is not possible for A to still be in the post oﬃce because in ten minutes time A and B will both ﬁnish their service times and exit the service station together.e.e. The other two expressions violate the nonlinearity of the function X 2 . Now this conditional probability density is given by pX (ξ|X > 1) = p(X = ξ. 1 = Letting u = ξ − 1. By the memoryless property of the exponential random variable the fact that we are conditioning on the event that X > 1 makes no diﬀerence relative to the event X > 0 (i. With this deﬁnition we then have that E[X |X > 1] = 0 ∞ 2 ξ 2 λe−λ(ξ−1) H(ξ − 1)dξ ξ 2 λe−λ(ξ−1) dξ . We ﬁnd ∞ E[X 2 |X > 1] = 0 ξ 2 pX (ξ|X > 1)dξ .

3 since A must now ﬁnish in three units of time. i.e. While ﬁnally P (E|B + C = 3) = since to have A ≥ B + C. Now because the random variables are continuous there is no need to diﬀerentiate between greater than and greater than or equal signs in the inequality denoting the event that A ﬁnishes last. If we let A. since A will certainly ﬁnish in a time less than four units. 3 since to have A ≥ B + C. which gives for P (E) using the above formula 9 P (E) = 4 . we ﬁnd that our conditional probabilities must be adjusted. and C be the amount of time that each respective person spends with their clerks. Also P (E|B + C = 2) = 2 . A can ﬁnish in three units. Now we have that our priors are given 2 by P (B + C = 2) = 1 and P (B + C = 3) = 9 . 3 9 27 Part (c): For this part we assume that the service time of each clerk is exponentially distributed. P {A > B + C} = p{A > B + C|B + C = t}P {B + C = t}dt . B. To evaluate the probability that this event happens we will condition on the random variable which is the sum of B and C. that A will leave last only if A > B + C. Here we have assumed (by using ≥) that an equality constraint is acceptable for determining if A leaves last. These then give in the same way that 1 1 1 P (E) = · = . then the event that A is the last person in the post oﬃce is equivalent to the constraint that A ≥ B + C. To compute the probabilities that this event happens we can condition on the possible sums of B and C (the times B and C spend at their clerks). For example 1 P (E|B + C = 2) = . This means that the time to service A takes more time than to service both B and C. Also P (E|B + C = 3) = 0. For notational purposes we will deﬁne the event that A leaves last as E. Now P (E|B + C = 4) = P (E|B + C = 5) = P (E|B + C = 6) = 0.Part (b): A will still be in the post oﬃce if both B and C are served before A. Now since both B and C are exponential random variables with the same rate the sum of them is a random variable distributed with a gamma distribution (see the section on further .e. We have P (E) = P (E|B + C = 2)P (B + C = 2) + P (E|B + C = 3)P (B + C = 3) + P (E|B + C = 4)P (B + C = 4) + P (E|B + C = 5)P (B + C = 5) + P (E|B + C = 6)P (B + C = 6) . 27 If we want to work this problem assuming strict inequality in the time relationships i. 3 1 . Thus we can take as an expression of the event that A will be the last person served if A > B + C. A can ﬁnishes in two or three units.

1! So that our integral is given by P {A > B + C} = PA {A > t}µ2 te−µt dt . Jones ﬁnishes before Mr. Then we have that P (E c ) = 1 − P (E). that Smith is last). Now P (JB) = λJ λJ + λB λB . λJ + λB and P (BJ) = . We thus have that the distribution of the sum of B and C is given by 1!e−µt (µt) fB+C (t) = = µ2 te−µt . Now PA {A > t} = 1 − (1 − e−µt ) = e−µt . 10 10 Exercise 6 (the probability that Smith is last) Now one of Mr. Jones (so that Mr. Jones or Mr. and let JB be the event that Mr. Jones or Mr. Smith will take Mr.96321 . Smith would take Mr. Now let JB be the event that Mr. so that this integral becomes ∞ P {A > B + C} = 0 µ2 te−2µt dt ∞ ∞ µ2 te−2µt − (−2µ) 0 (−2µ) ∞ µ e−2µt 1 = = . Brown who did not ﬁnish ﬁrst) is in the middle of their service has no eﬀect on whether Smith or his comp editor ﬁnishes ﬁrst. Brown will ﬁnish their service ﬁrst. We have that by conditioning on the events JB and BJ that P (E) = P (E|JB)P (JB) + P (E|BJ)P (BJ) . By the memoryless property of exponential the fact that Smiths “competitor” (the remaining Mr. Let E be the event that Smith is ﬁnishes last. where E c is the desired event (i. Smith will begin his processing there. Brown’s server positions). Jones’ server position).properties of the exponential distribution). Brown ﬁnishes before Mr. Once this person’s service station is open. Thus the probability that the radio is working after an additional ten years is given by ∞ P {X > 10} = 10 λe−λt dt = 1 − F (10) = 1 − 1 − 1 (10) 1 e 10 = 1 − e−1 = 0. Brown (then Mr. 2 (−2µ) 0 4 = µ2 e−2µt dt 0 Exercise 5 (old radios) Because of the memoryless property of the exponential distribution the fact that the radio is ten years old makes no diﬀerence.e.

X1 = t} = P {min(X1 . X2 ) = t} P {X2 > t}P {X1 = t} . = (λB + λJ )2 (λB + λJ )2 as we where to show. Brown ﬁnishes ﬁrst and thus has a probability of λB . Jones’ old server. Smith is last given that he is being serviced by Mr. X2 ) = t} P {X1 < X2 . Exercise 7 (the probability X1 < X2 given the minimum value of X1 and X2 ) For this exercise we will be considering the expression P {X1 < X2 | min(X1 . X2 ) = t} = P {X1 < X2 . This will happen if Mr. X2 ) = t} P {min(X1 . X2 ) = t}. By the discussion in the text on the section of the text entitled “Further properties of the exponential distribution”. X2 ) = t} Where in the last step we used the independence of X1 and X2 . X1 = t} = P {min(X1 . Given the failure rate of a random variable as a function of t say r(t) from the discussion in the book the distribution function F (·) in this case is given by F (t) = e− Rt 0 r(τ )dτ . . we have that P (E|JB) is the probability that Mr. Using the deﬁnition of conditional probability we have that this is equal to P {X1 < X2 | min(X1 . Jones and Mr. λB + λJ The probability that we want is P (E c ) and is given by P (E c ) = 1 − P (E) 2λB λJ (λB + λJ )2 − = 2 (λB + λJ ) (λB + λJ )2 λ2 λ2 J B + . λB + λJ λB + λJ λB + λJ λB + λJ (λB + λJ )2 λJ . min(X1 .where λJ and λB are the rates of the servers who are initially servicing Mr. λB + λJ In the same way we have P (E|BJ) = So that P (E) is given by P (E) = λJ λJ λB 2λB λJ λB · + · = . X2 ) = t} P {t < X2 . Brown when Mr. We now need to be able to relate the expressions above in terms of the failure rates of the random variables X1 and X2 . = P {min(X1 . Smith enters the post oﬃce.

for example see ??. which we do in the same way as in the book. X2 ) > t} = P {Xi > t . Exercise 8 (the expectation of 1/r(x)) We will use an equivalent expression for the expectation of X which is derived in many books on probability. X2 ) = t} Lets now work on evaluating the denominator of the above expression. We ﬁnd that P {X2 > t}P {X1 = t} (e− 0 r2 (τ )dτ )r1 (t)e− 0 r1 (τ )dτ = . This then means that our density function for this random variable min(X1 . X2 ) = t} = r1 (t)e− Rt 0 (r1 (τ )+r2 (τ ))dτ Rt 0 (r1 (t) + r2 (t))e− r1 (t) = . which is that E[X] = P {X > x}dx . P {min(X1 .Taking the derivative of this expression we ﬁnd that the density function f (·) for our random variable with failure rate r(t) is given by f (t) = r(t)e− Rt 0 r(τ )dτ . ∀i} 2 Rt Rt = = = e P {Xi > t} i=1 Rt Rt e− 0 r1 (τ )dτ e− 0 r2 (τ )dτ Rt − 0 (r1 (τ )+r2 (τ ))dτ . We ﬁnd that P {min(X1 . From which we see that min(X1 . X2 ) is a random variable with a failure rate given by r1 (t)+r2 (t). Using this result in our desired expression we ﬁnally conclude that P {X1 < X2 | min(X1 . X2 ) = t} = (r1 (t) + r2 (t))e− 0 (r1 (τ )+r2 (τ ))dτ . X2 ) = t} P {min(X1 . X2 ) is given by Rt P {min(X1 . r1 (t) + r2 (t) (r1 (τ )+r2 (τ ))dτ as we were to show. With these two expressions we can evaluate part of the above fraction. .

t).e. r(x) ≡ 1−F (x) . We want to calculate the probability that machine one is the ﬁrst machine to fail. machine two will be put into service. In addition we have that P (E|F ) = 1. λ1 + λ2 which is the probability P (E|F c ). Let E be the event that machine one fails ﬁrst and F the event that machine one fails in the time (0. t)) the probability that machine one fails before machine two is given by λ1 . P (F ) = P {X1 < t} = 1 − e−λ1 t P (F c ) = P {X1 > t} = e−λ1 t .From this expression and with the following manipulations to introduce the failure rate function r(X) we ﬁnd that E[X] = = P {X > x}dx 1 − F (x) f (x)dx f (x) 1 = f (x)dx r(x) 1 ]. = E[ r(X) Here we have used f (·) for the probability density function. Now conditioning on F and F c we have that P (E) = P (E|F )P (F ) + P (E|F c )P (F c ) . F (·) for the cumulative density f (x) function and r(·) for the failure rate function. Exercise 9 (the probability we fail ﬁrst after working t time) We are told that machine one is currently working and at time t. so that P (E) becomes P (E) = 1 · (1 − e−λ1 t ) + = 1− λ2 λ1 + λ2 λ1 · e−λ1 t λ1 + λ2 e−λ1 t . Now introducing X1 and X2 be the random variables denoting the time when machine one/two fails. At that point if the ﬁrst machine is still working they will be working in tandem. Now by the memoryless property of the exponential distribution when the second machine is brought on-line with the ﬁrst machine (assuming that machine one has not failed in the time (0. i.

(λ + µ)2 when we perform the integration. M) is then given by Cov(X. λ(λ + µ)2 so that with this result Cov(X. Part (b): To evaluate E[MX|M = Y ]. M) = 1 2λ + µ − 2 λ(λ + µ) λ+µ 1 . With this we ﬁnd that E[MX|M = Y ] = E[M(M + X ′ )] = E[M 2 ] + E[MX ′ ] = E[M 2 ] + E[M]E[X ′ ] 2 1 1 = + · . To compute E[XM] we can condition on whether the minimum is X or Y . 2 (λ + µ) λ+µ λ Part (c): We want to evaluate Cov(X. Note that if we try to evaluate this as E[MX|M = X] = E[X 2 ] . M). M) = E[XM] − E[X]E[M] . which we will do by simplifying the deﬁnition of the covariance to ﬁnd Cov(X. Instead we ﬁnd that ∞ E[MX|M = X] = E[M ] = 0 2 m2 e−(λ+µ)m dm = 2 . Now by the memoryless property of the exponential distribution X ′ is another exponential random variable with failure rate λ. We ﬁnd that E[XM] = E[XM|M = X]P {M = X} + E[XM|M = Y ]P {M = Y } 2 2 µ λ 1 = + + 2 2 (λ + µ) λ + µ (λ + µ) λ(λ + µ) λ + µ 2λ + µ = . this result cannot be correct since we have eﬀectively lost the information that our minimum of X and Y is X since the last expression involves only X.Exercise 10 (some expectations) Part (a): We want to evaluate E[MX|M = X]. we will stress the information that Y = M is less that X by writing X = M + X ′ . = (λ + µ)2 1 λ .

X3 ) = X3 }. and ﬁnally the third integral is with respect to x1 . . Because Xi are exponential with failure rate λi the above becomes ∞ ∞ ∞ P {X1 < X2 < X3 } = x1 =0 λ1 e−λ1 x1 x2 =x1 λ2 e−λ2 x2 x3 =x2 λ3 e−λ3 x3 dx3 dx2 dx1 . X2 . Part (b): We want to evaluate P {X1 < X2 | max(X1 . Now the event max(X1 . The ﬁrst integral is with respect to x3 . X2 . X3 ) = X3 } P {X1 < X2 < X3 } .Exercise 12 (probabilities with three exponential random variables ) Part (a): We can evaluate P {X1 < X2 < X3 } using the deﬁnition of this expression. P {X2 < X1 < X3 } = (λ1 + λ3 )(λ1 + λ2 + λ3 ) P {X1 < X2 < X3 } = so that P {max(X1 . The probability of this event was calculated in Part (a) of this problem. When this is done we obtain P {X1 < X2 < X3 } = λ1 λ2 .nb for this algebra. Each of these events can be computed using the results from Part (a). We ﬁnd. X3 ) = X3 imply that X1 < X2 < X3 . x3 )dx3 dx2 dx1 pX1 (x1 )pX2 (x2 )pX3 (x3 )dx3 dx2 dx1 x1 =0 ∞ x3 =x2 ∞ ∞ = = x1 =0 pX1 (x1 ) x2 =x1 pX2 (x2 ) x3 =x2 pX3 (x3 )dx3 dx2 dx1 . X3 ) = X3 } = P {X1 < X2 < X3 } + P {X2 < X1 < X3 } . = P {max(X1 . To compute the above probability We can perform the above integrations one at a time from the outside inward. X2 . X2 . X3 ) = X3 } since the two events X1 < X2 and max(X1 . X2 . Using the deﬁnition of the conditional expectation we ﬁnd that P {X1 < X2 | max(X1 . (λ2 + λ3 )(λ1 + λ2 + λ3 ) See the Mathematica ﬁle chap 5 prob 12. X2 . ∞ ∞ x2 =x1 ∞ x2 =x1 ∞ P {X1 < X2 < X3 } = x1 =0 ∞ x3 =x2 ∞ p(x1 . by the independence of the random variables Xi . X3 ) = X3 is equivalent to the union of the disjoint events X1 < X2 < X3 and X2 < X1 < X3 . x2 . the second with respect to x2 . X3 ) = X3 } = P {X1 < X2 . Speciﬁcally λ1 λ2 (λ2 + λ3 )(λ1 + λ2 + λ3 ) λ1 λ2 . X2 . X3 ) = X3 } P {max(X1 . X2 . max(X1 .

So that we see that to evaluate this expectation we need to be able to compute the conditional density pM (m|X1 < X2 < X3 ). λ1 + λ2 λ1 + λ2 See the Mathematica ﬁle chap 5 prob 12. X2 . Since X3 is an exponential random variable we know that P {X3 = m} = λ3 e−λ3 m . When this is done we obtain λ1 λ2 P {X1 < X2 < m} = − e−λ2 m + e−(λ1 +λ2 )m . X1 < X2 < X3 } P {X3 = m. + λ1 + λ2 + λ3 . We ﬁnd that P {M = m. We can also compute the second probability in the product above as P {X1 < X2 < m} = = pX1 (x1 )pX2 (x2 )dx2 dx1 Ω:{X1 <X2 <m} m x2 λ1 e−λ1 x1 λ2 e−λ2 x2 dx1 dx2 . so our problem becomes how to calculate the expression P {M = m. X3 ) = X3 } = = λ1 λ2 (λ2 +λ3 )(λ1 +λ2 +λ3 ) λ1 λ2 λ1 2 + (λ1 +λ3 )(λλ+λ2 +λ3 ) (λ2 +λ3 )(λ1 +λ2 +λ3 ) 1 λ1 + λ3 . The ﬁrst integral is with respect to x1 and the second is with respect to x2 . X1 < X2 < X3 } = λ1 λ2 λ3 −(λ1 +λ2 +λ3 )m · λ3 e−λ3 m − λ3 e−(λ2 +λ3 )m + e λ1 + λ2 λ1 + λ2 λ3 λ1 · λ3 e−λ3 m − (λ2 + λ3 )e−(λ2 +λ3 )m = λ1 + λ2 λ2 + λ3 λ2 λ3 (λ1 + λ2 + λ3 )e−(λ1 +λ2 +λ3 )m . X1 < X2 < X3 } P {X3 = m. x2 =0 x1 =0 We can perform the above integrations one at a time from the outside inward. X1 < X2 < X3 }. X3 ) = m. X1 < X2 < X3 } . λ1 + λ2 + 2λ3 Part (c): We want to evaluate E[max(Xi )|X1 < X2 < X3 ]. P {X1 < X2 < X3 } From Part (a) of this problem we know how to compute P {X1 < X2 < X3 }. X2 . From the deﬁnition of conditional expectation we have that (if we deﬁne the random variable M to be M = max(Xi )) ∞ E[max(Xi )|X1 < X2 < X3 ] = 0 mpM (m|X1 < X2 < X3 )dm . X1 < X2 < m} P {X3 = m}P {X1 < X2 < m} . using the independence of the variables Xi . Now multiplying this expression by P {X3 = m} = λ3 e−λ3 m we have (doing the steps very slowly) that P {M = m. X1 < X2 < X3 } = = = = P {max(X1 . Using the deﬁnition of conditional density we have that pM (m|X1 < X2 < X3 ) = P {M = m.Thus using these results we can calculate our desired probability P {X1 < X2 | max(X1 .nb for this algebra.

X2 ) − min(X1 . Xj . Speciﬁcally the random variable min(X1 . Performing the integration over m in the above expression and remembering the expression for the mean of a exponentially distributed random variable we ﬁnd that E[max(Xi )|X1 < X2 < X3 ] ∝ λ1 1 λ3 1 · − · λ1 + λ2 λ3 λ2 + λ3 λ2 + λ3 1 λ2 λ3 · . For example E[max(Xi )] = + + + E[max(Xi )|X1 < X2 < X3 ]P {X1 < X2 < X3 ]} E[max(Xi )|X1 < X3 < X2 ]P {X1 < X3 < X2 ]} E[max(Xi )|X2 < X3 < X1 ]P {X2 < X3 < X1 ]} ··· The two approaches should give the same answer but the second seems more tedious. X3 ) = i=1 Xi − i<j min(Xi . X3 ) + min(X1 . Sine the Xi ’s are independent exponential distributed random variables the random variables which are the minimization’s of such random variables are themselves exponential random variables with failure rates given by the sum of the failure rates of their various components. We explicitly demonstrate this fact by introducing normalized exponential random variables in the second line above. taking the expectation of this expression is simple to do. + λ1 + λ2 + λ3 λ1 + λ2 + λ3 So that when we divide by P {X1 < X2 < X3 } we ﬁnd that we obtain for E[max(Xi )|X1 < X2 < X3 ] the following (λ2 + λ3 )(λ1 + λ2 + λ3 ) × λ1 λ2 λ3 λ3 λ1 − + 2 λ3 (λ1 + λ2 ) (λ2 + λ3 ) (λ1 + λ2 + λ3 )2 . X2 . X2 . X2 . Part (d): To evaluate E[max(X1 . Taking the expectation of this expression we ﬁnd that E[max(X1 . X3 ) − min(X2 . X3 )] = 1 1 1 1 1 1 + + − − − λ1 λ2 λ3 λ1 + λ2 λ1 + λ3 λ2 + λ3 1 . Which one might be able to simplify further but we stop here. Xk ) = X1 + X2 + X3 − min(X1 . With this representation. X2 ) is an exponential random variable with failure rate give by λ1 + λ2 . + λ1 + λ2 + λ3 Note that by using the various pieces computed in this problem one could compute E[max(Xi )] by conditioning on events like in Part (c). X2 . X3 ) . . X3 )] we will use the mins for maxs identity which is 3 max(X1 .The manipulations performed above were to enable us to recognize the above expression as a linear combination of exponential random variables. At this point in the calculation we do not yet need divide by P {X1 < X2 < X3 } since it is just a scalar multiplier of the above and can be done after the integration. Xj ) + i<j<k min(Xi .

P {X < c} = FX (c) = 1 − λe−λc . We know that 1 λ P {X < c} = 1 − λe−λc P {X > c} = λe−λc . E[X] = Thus using all of this information in the ﬁrst expression derived in this part of this problem we ﬁnd 1 = E[X|X < c](1 − λe−λc ) + E[X|X > c](λe−λc ) .e. Now the density function in the above expression is by deﬁnition given by pX (x|X < c) = P {X = x. The denominator is then given by the cumulative distribution function for an exponential random variable i. λ . Now from the properties of an exponential distribution we know several of these pieces. here H(·) is the Heaviside function and enforces the requirement that X < c. Using both of these expressions the above becomes ∞ x 0 H(c − x)pX (x) 1 − λe−λc dx = c 1 xλe−λx dx 1 − λe−λc 0 1 (1 − e−λc − cλe−λc ) . X < c} = H(−(x − c))pX (x) = H(c − x)pX (x) . To begin we have E[X] = E[X|X < c]P {X < c} + E[X|X > c]P {X > c} .Exercise 14 (are we less than c?) Part (a): We have from the deﬁnition of conditional expectation that E[X|X < c] = xpX (x|X < c)dx . = λ(1 − λe−λc ) for the requested expression. λ Now by the memoryless property of the exponential distribution we have the key observation that 1 E[X|X > c] = E[X + c] = E[X] + c = + c . X < c} . P {X < c} Now in the above expression the numerator is given by P {X = x. Part (b): We now want to ﬁnd the E[X|X < c] by conditioning on E[X].

Notice that this is a linear combination of two exponential densities i. (2µ)2 4µ Part (c): To compute moments of X(2) = max(X1 . 1 − (1 + cλ)e−λc . This observation helps in computing various quantities. λ(1 − λe−λc ) 1 + c (λe−λc ) . For example for the expectation we ﬁnd that E[X(2) ] = 2 1 µ − 1 3 = .so the formula above becomes 1 = E[X|X < c](1 − λe−λc ) + λ which when we solve for E[X|X < c] we ﬁnd that E[X|X < c] = the same expression as in Part (a). 2µ 2µ . E[X(1) ] = 2µ Part (b): Since X(1) is deﬁned as X(1) = min(X1 . (n − i)!(i − 1)! so the distribution of the maximum X(n) is given by fX(n) (x) = n! f (x)F (x)n−1 = nf (x)F (x)n−1 . X2 ) recall from the discussion on order statistics that the density function for the ith order (of n independent random variables with density/distribution function f /F ) is given by fX(i) (x) = n! f (x)F (x)i−1 (1 − F (x))n−i . X2 ) it is an exponential random variable with rate 2µ so we ﬁnd that 1 . one with rate µ and one with rate 2µ. X2 ) it is an exponential random variable with rate 2µ so we ﬁnd that Var(X(1) ) = 1 1 = 2.e. λ Exercise 18 (mins and maxes of same rate exponentials) Part (a): Since X(1) is deﬁned as X(1) = min(X1 . 0!(n − 1)! When there are only two independent exponential random variables this becomes fX(2) (x) = 2µe−µx − 2µe−2µx .

Thus we then ﬁnd using this method that Var(X(2) ) = Var(X(1) ) + Var(A) = the same as before. X2 )] = 1 1 1 3 + − = . Thus taking the expectation of the above (and using the known value of the expectation for an exponential random variable) we ﬁnd that E[max(X1 . λ2 derived above we ﬁnd that E[X 2 ] = 2 E[X(2) ] = 2 2 µ2 − 2 7 = 2. X2 ) which are not independent and would have to be computed in some way. Finally a third way to compute this expectation is to recall that by the memoryless property of the exponential distribution that the random variable X(2) is related to the random variable X(1) by an “oﬀset” random variable (say A) in such a way that X(2) = X(1) + A. 1 5 1 + 2 = . Thus we then ﬁnd using this method that E[X(2) ] = E[X(1) ] + E[A] = the same as earlier.We note that this problem could also be solved by using the “mins” for “maxes” relationship which for two variables is given by max(X1 . X2 ) . 2µ µ 2µ given the density for X(2) 2 . (2) where we know that since Xi is distributed as an exponential random variable with rate µ the random variable min(X1 . 2 (2µ) µ 4µ . X2 ) is an exponential random variable with rate 2µ. Remembering that the second moment of an exponential random variable (say X) with failure rate λ is given by 1 1 3 + = . where A is an exponential random variable with rate µ and X(1) and A are independent. Part (d): To compute the variance of X(2) we can use 2 Var(X(2) ) = E[X(2) ] − E[X(2) ]2 . 2 (2µ) 2µ Using this result we see that Var(X(2) ) is given by Var(X(2) ) = 9 1 5 7 − = 2. 2 and we can calculate E[X(2) ] from the density given in Part (c). 2 2 2µ 4µ 4µ Another way to compute this variance is to recall that by the memoryless property of the exponential distribution that the random variable X(2) is related to the random variable X(1) by an “oﬀset” random variable (say A) in such a way that X(2) = X(1) + A. The variance calculation using this approach would be more complicated because it would involve products of terms containing X1 and min(X1 . X2 ) = X1 + X2 − min(X1 . where A is an exponential random variable with rate µ and X(1) and A are independent. µ µ 2µ 2µ as before.

Combining all of this information we ﬁnd that our probability density is given by pM (m|X1 < X2 ) = λ1 + λ2 λ1 (λ2 e−λ2 m − λ2 e−(λ1 +λ2 )m ) .Exercise 19 (mins and maxes of diﬀerent rate exponentials) Part (a): Using much of the discussion from Exercise 18 we have that X(1) is an exponential random variable with rate µ1 + µ2 . We have E[X(2) ] = E[X(2) |X1 < X2 ]P {X1 < X2 } + E[X(2) |X1 > X2 ]P {X1 > X2 } λ2 λ1 + E[X(2) |X1 > X2 ] . X2 ) = m. The conditional probability distribution pM (m|X1 < X2 ) is given by some simple manipulations. so that E[X(1) ] = 1 . We ﬁrst deﬁne the random variable M = max(X1 . From earlier we know that the denominator of the above is given by P {X1 < X2 } = λ1λ1 2 . X2 ) and use the deﬁnition of conditional expectation. (µ1 + µ2 )2 Part (c): To compute E[X(2) ] we condition on whether X1 < X2 or not. and the numerator is +λ given by P {X2 = m}P {X1 < m} = λ2 e−λ2 m (1 − e−λ1 m ) = λ2 e−λ2 m − λ2 e−(λ1 +λ2 )m . . X1 < X2 } P {X1 < X2 } P {X2 = m. = E[X(2) |X1 < X2 ] λ1 + λ2 λ1 + λ2 We will now evaluate E[X(2) |X1 < X2 ]. P {X1 < X2 } where we have used the independence of X1 and X2 in the last step. X1 < X2 } P {X1 < X2 } P {max(X1 . X1 < X2 } P {X1 < X2 } P {X2 = m. X1 < m} P {X1 < X2 } P {X2 = m}P {X1 < m} . µ1 + µ2 Part (b): As in Part (a) the variance of X(1) is simple to calculate and we ﬁnd that Var(X(1) ) = 2 . We ﬁnd pM (m|X1 < X2 ) = = = = = P {M = m. We have ∞ E[X(2) |X1 < X2 ] = 0 mpM (m|X1 < X2 )dm .

2 Part (d): We want to evaluate Var(X(2) ). which we will do by evaluating E[X(2) ] − E[X(2) ]2 2 using the results from Part (c) above. When we enter the . So combining these two results we can ﬁnally obtaining E[X(2) ] as E[X(2) ] = 1 1 1 + − . · − E[X(2) |X1 < X2 ] = λ1 λ2 λ1 + λ2 (λ1 + λ2 )2 2 and in the same way we then have 2 E[X(2) |X1 > X2 ] = λ1 + λ2 λ2 2 2 λ1 · − 2 λ1 λ1 + λ2 (λ1 + λ2 )2 . 2 So that using the results above. 2 be an exponential random variables with parameter µ1 . the expectation of X(2) is given by 2 E[X(2) ] = 2 2 2 + 2− . and Yi for i = 1. 2 be exponential random variables with parameter µ2 . From the expression for pM (m|X1 < X2 ) derived in Part (c) we ﬁnd that 2 λ2 λ1 + λ2 2 2 . By the same arguments (eﬀectively exchanging 1 with 2) we ﬁnd that E[X(2) |X1 > X2 ] = λ1 + λ2 λ1 − λ1 λ2 λ2 1 λ1 + λ2 . 1 1 1 + − λ1 λ2 λ1 + λ2 2 Exercise 21 (waiting time at a two station queue) For simplicity let Xi for i = 1. λ1 λ2 λ1 + λ2 Note that this could also have been obtained (with a lot less work) by using the mins for max identity given by Equation 2. 2 λ1 λ2 (λ1 + λ2 )2 Giving ﬁnally the desired result for Var(X(2) ) of 2 2 2 + 2− − Var(X(2) ) = 2 λ1 λ2 (λ1 + λ2 )2 1 3 1 + 2− .Multiplying by m and integrating we can use the fact that pM (m|X1 < X2 ) is a linear combination of exponential densities to ﬁnd that E[X(2) |X1 < X2 ] = λ1 + λ2 λ2 − λ1 λ2 λ1 1 λ1 + λ2 . As before we can evaluate E[X(2) ] by conditioning on whether X1 < X2 or not. = 2 λ1 λ2 (λ1 + λ2 )2 after some algebra.

The total time for us to wait until we move to server number two’s station is given by the maximum amount of time spent at either server number one or server number two and again using the memory less property of the exponential distribution is represented as the random variable max(X2 . Y1 ) + Y2 . At this point we will be at server ones station while the person ahead of us will be at sever number two’s station. y)dxdy µ1 e−µ1 x µ2 e−µ2 y dxdy 0 0 a a = = µ1 µ2 e −µ1 x dx e−µ2 y dy 0 −µ2 a = (1 − e )(1 − e ) −µ1 a −µ2 a = 1−e −e + e−(µ1 +µ2 )a Thus our density function is then given by f (a) = µ1 e−µ1 a + µ2 e−µ2 a − (µ1 + µ2 )e−(µ1 +µ2 )a . Showing easily that the expectation of the random variable A is given by E[A] = 1 1 1 + − . If we imagine a the random variable X along the x axis of a Cartesian grid and the random variable Y along the y axis of a Cartesian grid then the set of points where max(X. Y1) when X2 and Y2 are distributed as above. µ1 µ2 µ1 + µ2 1 µ2 . By the memoryless property of exponential random variables we have that the amount of time we have to wait for this to happen will be X1 . To do this consider the distribution function of this random variable where we drop the subscripts on X and Y we have F (a) = P {max(X. µ1 µ2 µ1 + µ2 0 −µ1 a Thus we can now evaluate the expected length of time in the system T as E[T ] = E[X1 ] + E[max(X2 . To evaluate this we need to compute the distribution function of a random variable A deﬁned by A = max(X2 .system we will have to ﬁrst wait for the ﬁrst server to ﬁnish. Y ) ≤ a. Thus if T denotes the random variable representing the total amount of time we will have to wait we have that T = X1 + max(X2 . Thus we can evaluate F (a) (using the independence of X and Y ) as a a F (a) = 0 a 0 a f (x. Y1 ). Y ) ≤ a is 0 ≤ X ≤ a and 0 ≤ Y ≤ a. Once this amount of time has completed we will be a the second server where we will have to wait an additional Y2 amount of time. Y ) ≤ a} = f (x. y)dxdy Ωa where Ωa is the set of points in the XY plane where max(X. Y1)] + E[Y2 ] 1 1 1 1 + + + − = µ1 µ1 µ2 µ1 + µ2 2 2 1 = + − .

y)dydx µe−µx µe−myy dydx x=0 y=0 ∞ −µx = = µ = x=0 e (1 − e−µx )dx 1 2 ∞ x=0 ∞ µe−µx dx − (2µ)e−2µx dx x=0 1 1 = 1− = . Thus the total time T spent in the system is given by T = max(X1 . . y)dydx x=0 ∞ y=0 x = = x=0 ∞ y=0 x fX. Y1 ). Thus if X and Y are exponential random variables denoting the life expectancy of two batteries with the same parameter µ the probability that X will fail before Y is given by P {X < Y } = ΩX<Y ∞ x fX. Y2 ) + Y3 . After this amount of time you move to server number one.Exercise 22 (more waiting at a two station queue) In the same way as in the previous problem we will let Xi be an exponential random variable with parameter µ1 and Yi be an exponential random variable with parameter µ2 .Y (x. µ1 µ2 µ1 + µ2 + 1 1 1 + − µ1 µ2 µ1 + µ2 + 1 µ2 Exercise 23 (ﬂashlight batteries) By the memory less property of the exponential distribution. Y1 ) + max(X2 . To move to server number two you must wait another max(X2 .Y (x. 2 2 Thus when two batteries are in place due to the memoryless property of the exponential distribution it is equally likely that either one of them will expire ﬁrst. Y2 ) amount of time. With this information we can calculate the various probabilities. y)dxdy fX. So the expected time in the station is given by (using the results from the previous problem) E[T ] = 1 1 1 + − µ1 µ2 µ1 + µ2 2 3 2 = + − . when two batteries are being used simultaneously the lifetime of either battery is given by an exponential distribution that does not depend on how long the battery has been operational up to that point. Then to have server one become available we must wait an amount max(X1 .Y (x. Finally you must wait for server number two to ﬁnish (an additional amount of time Y3 ) serving you.

the fourth batter. · · · . 2µ the expectation of T is given by n−1 E[T ] = i=1 E[Fi ] = n−1 . If we let T1 and T2 represent the exponential random variables representing the lifetime of the ﬁrst and second battery then the ﬂashlight will function for a random amount of time F1 = min(T1 . the ﬁrst two batteries will produced a functional ﬂashlight until the ﬁrst one burns out. Thus P {X = n} = 1/2. Part (c): To evaluate P {X = i} means that when the ith battery is placed in the ﬂashlight it proceeds to last longer than the remaining I. etc. This process continues n − 1 times. By the memoryless property of the exponential distribution the battery that is not replaced has a distribution as if it just started at this replacement time. the third battery. Since the event that the ﬁrst battery lasts longer than these n − 1 other batteries this will happen with probability (1/2)n−1 . 2µ . Thus if T is the lifetime of the total ﬂashlight system we see that it is given by n−1 T = i=1 Fi . Here I is the index of the other battery in the ﬂashlight when the ith battery is placed in the ﬂashlight (1 ≤ I ≤ i − 1). Thus n−i+1 1 2≤i≤n P {X = i} = 2 The reason this expression does not work for i = 1 is that in that case there is no battery “before” this one to compare to. When one of these two batteries burns out it is immediately replaces. Part (b): P {X = 1} is the probability that when the last working battery is observed it happens to be the ﬁrst battery on the battery list. with probability 1/2 it will last longer than the other battery. where T∗ is the battery not replaced earlier. Since when the n the battery is placed in the ﬂashlight. thus P {X = 1} = 1 2 n−1 . T2 ). up to the nth battery. Thus the ith battery has n − (i + 1) + 1 + 1 = n − i + 1 comparisons which it must win. i + 1. i + 2. Part (d): The random variable T (representing the entire lifetime of our ﬂashlight) can be expressed in terms of sums of random variables representing the serial lifetimes of two individual batteries. Note that each of these minimizations is itself a random variable with a rate given by 2µ. This will happen if the ﬁrst battery lasts longer than the second battery. Thus the index in the above expression is one two large for the case i = 1. For example. n other batteries. Since for each Fi we have E[Fi ] = 1 . T3 ). The next increment of time before the fourth battery is used is given by F2 = min(T∗ .Part (a): P {X = n} is the probability that the last non failed battery is number n.

This means that ∞ fZ (z) = −∞ fX (ξ)f−Y (z − ξ)dξ . plotted as a function of ξ this is the function f−Y (−ξ) shifted to the right by z. . Part (c): We are told that X and Y are exponential random variables then we want to prove that X − Y is given by a Laplace (double exponential) distribution. Plotting the function f−Y (−(ξ − z)). speciﬁcally fT (t) = 2µe−2µt (2µt)n−2 . We will integrate and see if we obtain unity. Now the density f−Y (y) is given by f−Y (y) = 0 y>0 λeλy y < 0 So to evaluate our convolution given above we begin by letting z < 0 and then since f−Y (z − ξ) = f−Y (−(ξ −z)). 2 2 = Part (b): The distribution function F (·) is given by x F (x) = −∞ 1 1 λξ λe dξ = eλξ 2 2 x −∞ 1 = eλx 2 x 0 for x < 0 . and 1 F (x) = + 2 x 0 1 −λξ 1 λ eλξ λe dξ = + 2 2 2 (−λ) 1 = 1 − e−λx 2 for x > 0 . With this understanding if z < 0 our function fZ (z) is given by ∞ fZ (z) = 0 λe−λξ · λeλ(z−ξ) dξ = λ λz e 2 for z < 0 . To do this deﬁne the random variable Z ≡ X − Y . which has a distribution function given by the convolution of the random variables X and −Y .Part (e): Since T is the sum of n − 1 exponential random variables with rates 2µ the distribution of T is given by a gamma distribution with parameters n − 1 and 2µ. for z = −1 and z = +1 we have the ﬁgure ??. We have ∞ 0 f (x)dx = −∞ −∞ λ λx e dx + 2 λx 0 ∞ 0 λ −λx e dx 2 ∞ 0 λ e λ e−λx + 2 λ −∞ 2 (−λ) 1 1 = (1) − (−1) = 1 . (n − 2)! Exercise 24 (Laplace distributions) Part (a): We want to show that the given f is a density function.

2 2 2 2 with H(·) the Heaviside function. For we have 1 1 1 1 pW (w) = pX (w) + p−Y (w) = H(w)λe−λw + H(−w)λeλw . Part (e): We deﬁne the random variable W as W = X −Y I=1 I = −1 1 λeλy 2 1 λe−λy 2 0 y>0 λeλy y < 0 y<0 y>0 then this probability distribution can be calculated as in a Part (d) of this problem.after some algebra. Part (d): Now to show that IX is a Laplace random variable we remember that I = ±1 both with probability 1/2. From which we see that the random variable W is a Laplace random variable. If z > 0 then fZ (z) is given by ∞ fZ (z) = z λe−λξ · λeλ(z−ξ) dξ = λ −λz e 2 for z < 0 . Combining these results we ﬁnd that fZ (z) = λ λz e 2 λ −λz e 2 z<0 z>0 showing that Z is a Laplace or double exponential random variable. Since P {X = −y} = we have that pY (y) is given by pY (y) = or a Laplace random variable. . 2 2 by the independence of X and I. We have pY (y) = P {IX = y|I = −1}P {I = −1} + P {IX = y|I = +1}P {I = +1} 1 1 = P {IX = y|I = −1} + P {IX = y|I = +1} 2 2 1 1 = P {X = −y|I = −1} + P {X = y|I = +1} 2 2 1 1 = P {X = −y} + P {X = y} . Deﬁning Y ≡ IX the probability distribution of Y can be found by conditioning on the value of I.

µA + λ with µA the exponential rate of the lifetime of A. Problem 30 (pet lifetimes) Let L be the random variable denoting the additional lifetime of the surviving pet. T . By the independence of these random variables the probability of this event is P {LB > LA }P {LB > T } = µA µA + µB λ µB + λ . Thus as long as A is still living when it ﬁnally arrives she will receive it. Problem 31 (doctors appointments) If the 1:00 appointment is over before the 1:30 appointment then the amount of time the 1:30 patient spends in the oﬃce is given by an exponential random variable with mean 30 . With these deﬁnitions we can compute E[L] as E[L] = E[L|LD < LC ]P {LD < LC } + E[L|LD > LC ]P {LD > LC } λd λc + E[L|LD > LC ] = E[L|LD < LC ] λd + λc λd + λc λc λd + E[LD ] = E[LC ] λd + λc λd + λc 1 1 λd λc 1 λd λc = + = . is distributed as an exponential random variable with rate λ. and B still alive when the kidney arrives. This result as derived in the section of the book entitled “Further Properties of the Exponential Distribution”. Mathematically if we denote LB as the random variable the lifetime of B this is the event that LB > LA and LB > T . + λc λd + λc λd λd + λc λd + λc λc λd In the above we have been able to make the substitution that E[L|LD < LC ] = E[LC ]. If we let LD and LC be random variables denoting the lifetimes of the dog and cat respectively. Part (b): For B to receive a new kidney we must have had A expire. then the event that the dog is the ﬁrst do die is mathematically stated as LD < LC . We can compute the expectation of L by conditioning on which pet dies ﬁrst. This event happens with probability λ .Exercise 29 (replacement kidneys) Part (a): From the information given the time till a new kidney’s arrival. If we denote LA as the random variable representing the lifetime of A then A will be alive if LA > T . because of the memoryless property of the exponential random variables involved.

Y = c − x} = P {X + Y = c} P {X = x}P {Y = c − x} . X + Y = c} = P {X + Y = c} P {X = x. we see that Z has a distribution function given by the appropriate convolution ∞ fZ (z) = −∞ fX (ξ)fY (z − ξ)dξ . Problem 34 (X given X + Y ) Part (a): Let begin by computing fX|X+Y =c (x|c). We ﬁnd E[L] = E[L|T1 < 30]P {T1 < 30} + E[L|T1 > 30]P {T1 > 30} .minutes. Now deﬁning the random variable Z as Z = X + Y . We have that fX|X+Y =c (x|c) = P {X = x|X + Y = c} P {X = x. By the memoryless property of the exponential random variable we have that E[T1 − 30|T1 > 30] = E[T1 ] = 30 . We ﬁnd E[L|T1 < 30] = E[T2 ] = 30 minutes. minutes for the expected amount of time the 1:30 patient spends at the doctors oﬃce. We can compute each of these expressions. P {T1 < 30} = 1 − e−(1/30)(30) = 1 − e−1 .03 . . = P {X + Y = c} By the independence of X and Y . If the 1:00 appointment runs over (into the 1:30 patients time slot) the 1:30 patient will spend more time in the oﬃce due to the fact that his appointment cannot start on time. Thus we ﬁnd E[L|T1 > 30] = 30 + E[T2 |T1 > 30] = 30 + 30 = 60 . Thus letting T1 and T2 be exponential random variables denoting the length of time each patient requires with the doctor. the expected time the 1:30 appointment spends at the doctors oﬃce (L) is given by conditioning on the length of time the 1:00 patient requires with the doctor. To evaluate E[L|T1 > 30] we can write it as follows. and P {T1 > 30} = e−(1/30)(30) = e−1 . E[L|T1 > 30] = E[(T1 − 30) + T2 |T1 > 30] = E[T1 − 30|T1 > 30] + E[T2 |T1 > 30] . and ﬁnally obtain E[L] = 30(1 − e−1 ) + 60(e−1 ) = 41.

Part (b): From Part (a) of this problem we can integrate this expression to compute E[X|X + Y = c]. The book gives a slightly diﬀerent result (which I believe is a typo). Consider the expression E[X + Y |X + Y = c]. The ﬁrst deﬁnition (Deﬁnition 5. λ−µ = when we do the integration.1 in the book) is the following . By linearity of the expectation it must also equals E[X|X + Y = c] + E[Y |X + Y = c] .so that if z < 0 we have the above equal to z fZ (z) = 0 λe−λξ µe−µ(z−ξ) λµ (e−µz − e−λz ) . c) to verify that this function integrates to one. which we know must equal c since we are conditioning on the event X + Y = c. One can integrate the above over the range (0. We ﬁnd that E[X|X + Y = c] = c (λ − µ) xe−(λ−µ)x dx (1 − e−(λ−µ)c ) 0 ce−(λ−µ)c 1 − . So solving for the desired E[Y |X + Y = c] we see that E[Y |X + Y = c] = c − E[X|X + Y = c] ce−(λ−µ)c 1 + = c− λ − µ 1 − e−(λ−µ)c 1 − e−(λ−µ)c 1 ce−(λ−µ)c = c − + 1 − e−(λ−µ)c λ − µ 1 − e−(λ−µ)c 1 c − . = 1 − e−(λ−µ)c λ − µ Problem 35 (equivalent deﬁnitions of a Poisson process) We are asked to prove the equivalence of two deﬁnitions for a Poisson process. = λ − µ 1 − e−(λ−µ)c Part (c): To ﬁnd E[Y |X + Y = c] we can use an idea like that in Exercise 14. I believe the result above is correct. Thus with this result we see that the expression for fX|X+Y =c (x|c) is given by fX|X+Y =c (x|c) = = (λe−λx )(µe−µ(c−x) ) λµ (e−µc − e−λc ) λ−µ (λ − µ)e−(λ−µ)x 1 − e−(λ−µ)c for 0 < x < c .

That is for s. n! P {N(t) ≥ 2} = n=2 e−λt (λt)n n! ∞ = e = e −λt n=0 λt (λt)n − 1 − λt n! −λt e − 1 − λt = 1 − e−λt − λte−λt . t ≥ 0} has independent increments iii The number of events in any interval of length t has a Poisson distribution with mean λt. t ≥ 0} has stationary. From (ii) in Deﬁnition 5.The counting process {N(t). we have P {N(t + s) − N(s) = n} = e−λt (λt)n n! n≥0 We want to show that this deﬁnition is equivalent to the following (which is Deﬁnition 5.3 in the book) i N(0) = 0 ii {N(t). which (we claim) is a function that is o(t). t→0 t .3 (ii). t ≥ 0. independent increments iii P {N(t) ≥ 2} = o(t) iv P {N(t) = 1} = λt + o(t) We begin by noting that both deﬁnitions require N(0) = 0.1 we have with s = 0 (and the fact that N(0) = 0) that P {N(t) = n} = So that ∞ e−λt (λt)n . From (iii) in Deﬁnition 5. Thus we want to evaluate lim 1 − e−λt − λte−λt .1 we have the required independent increments needed in Deﬁnition 5. t ≥ 0} is said to be a Poisson process if: i N(0) = 0 ii {N(t).3 (ii).1 we have that the distributions of X(t2 + s) − X(t1 + s) is given by a Poisson distribution with mean λ(t2 − t1 ) and the distribution of random variable X(t2 ) − X(t1 ) is also given by a Poisson distribution with mean λ(t2 − t1 ) showing that the process {N(t)} also has stationary increments and thus satisﬁes the totality of Deﬁnition 5. To show that this is true consider the limit as t goes to zero. From (iii) in Deﬁnition 5.

Note that this is the density function for a Poisson random variable (or the cumulative distribution function of a Poisson random variable with n = 0).3 is (iv). If her crossing takes s second this will happen with probability P {N = 0} = e−λs = e−3s . 5. This expression is tabulated for s = 2. Which we do by evaluating the limit lim e−λt − 1 + λt −λe−λt + λ = lim = −λ + λ = 0 . Exercise 37 (helping Reb cross the highway) At the point where Reb wants to cross the highway the number of cars that cross is a Poisson process with rate λ = 3.3. . the probability that k appear is given by P {N = k} = e−λt (λt)k . showing the truth of condition (iv) in Deﬁnition 5.Since this is an indeterminate limit of type 0/0 we must use L’Hospital’s rule which gives that the above limit is equal to the limit of the derivative of the top and bottom of the above or λe−λt − λe−λt + λ2 te−λt lim = λ− λ = 0. The ﬁnal condition required for Deﬁnition 5.1 (iii) that P {N(t) = 1} = e−λt (λt) = λte−λt 1! To show that this expression has the correct limiting behavior as t → 0. With this result we see that P {N(t) = 1} = λt(1 − λt + o(t)) = λt − λ2 t2 + o(t2 ) = λt + o(t) . 10.m. t→0 1 Proving that this expression is o(t) (since this limit equaling zero is the deﬁnition) and proving that P {N(t) ≥ 2} = o(t). t→0 t→0 t 1 Where we have used L’Hospital’s rule again. We have from Deﬁnition 5. 20 seconds in chap 5 prob 37. we ﬁrst prove that e−λt = 1 − λt + o(t) as t → 0 . k! Thus Reb will have no problem if no cars come during her crossing.

which we can do by using the deﬁnition of conditional probabilities as P {N1 (t) = 1. Reb will cross unhurt. = P {N1 (t) + N2 (t) = 1} . 30 seconds in chap 5 prob 38. with probability P {N = 0} + P {N = 1} = e−λs + e−λs (λs) = e−3s + 3se−3s . the sum of the Poisson parameters of the random variables X and Y . since N1 (t) and N2 (t) are both Poisson random variables with parameters λ1 t and λ2 t respectively. Let X and Y be Poisson random variables with parameters λ1 and λ2 respectively. Exercise 41 (the probability that N1 hits ﬁrst) For this problem we are asked to evaluate P {N1 (t) = 1.Exercise 38 (helping a nimble Reb cross the highway) Following the results from Exercise 29. 10. N2 (t) = 0|N1 (t) + N2 (t) = 1} . 20. Now for the problem at hand. This expression is tabulated for s = 5. N2 (t) = 0} P {N1 (t) + N2 (t) = 1} P {N1 (t) = 1}P {N2(t) = 0} . j! j! iu iu showing that N(t) is a Poisson process with rate λ1 + λ2 . then from the above discussion the random variable N(t) deﬁned by N1 (t) + N2 (t) is a Poisson random variable with parameter λ1 t + λ2 t and thus has a probability of the event N(t) = j given by P {N(t) = j} = e−(λ1 +λ2 )t ((λ1 + λ2 )t)j e−(λ1 t+λ2 t) (λ1 t + λ2 t)j = . From the direct connection between characteristic functions to and probability density functions we see that the random variable X + Y is a Poisson random variable with parameter λ1 + λ2 . We can evaluate the distribution of X + Y by computing the characteristic function of X + Y . Not that this is the cumulative distribution function for a Poisson random variable.m. Exercise 40 (the sum of two Poisson processes) We will ﬁrst prove that the sum of two Poisson random variables is a Poisson random variable. N2 (t) = 0|N1 (t) + N2 (t) = 1} = P {N1 (t) = 1. Since X and Y are independent Poisson random variables the characteristic functions of X + Y is given by φX+Y (u) = φX (u)φY (u) = eλ1 (e −1) eλ2 (e −1) iu = e(λ1 +λ2 )(e −1) .

the time needed to elapse to the next event is given by an exponential distribution that is unchanged. Subtracting one from each N(·) in the above expectation we ﬁnd that E[N(4) − N(2)|N(1) = 3] = = = = E[N(3) − N(1)|N(1) = 3] E[N(3) − 3] E[N(3)] − 3 3λ − 3 . Thus the expectation of Ti is well known since E[Ti ] = 1 . i. .In the above we have used the independence of the process N1 (·) and N2 (·). Mathematically we have E[S4 |N(1) = 2] = 1 + E[S2 ] = 1 + using the results from Part (a) of this problem. λ Using the result that E[N(t)] = λt which was shown in the book. λ Part (b): To calculate E[S4 |N(1) = 2] we will use the memoryless property of the exponential distribution. λ With this results we see that the expected value of the fourth arrival time E[S4 ] is given by 4 4 4 E[S4 ] = E[ i=1 Ti ] = i=1 E[Ti ] = i=1 1 4 = .e. λ λ From this we see that in general we have that E[Sn ] = n . Now since we are told that at t = 1 we have seen two of the four events (not more) from the time t = 1 onward the evaluation of the above expression we only need to compute the time for two more events to elapse. no matter how much time has elapsed between the last event seen. Part (c): To calculate E[N(4) − N(2)|N(1) = 3] we will speciﬁcally use the stationary increments property of the Poisson process. −(λ1 +λ2 )t ((λ +λ )t)1 e 1 2 λ1 + λ2 1! Exercise 42 (arrival times for a Poisson process) Part (a): For a Poisson process the inter-arrival times Ti are distributed as exponential random variables with parameter λ. This last argument relies on the memoryless property of the exponential distribution. The above then equals e−λ1 t (λ1 t)1 · e−λ2 t λ1 1! = . 2 .

n! where from the problem statement we have λ = 7 (per hour). the probability that there are n people waiting at the train stop at time T = t is given by the standard expression for a Poisson process i. 2 Since we know T to be a uniform random variable over (0. 2 When we combine these two sub-results we ﬁnally conclude that Var(X) = λ λ2 + . We will compute each term on the right hand side. We begin by computing E[X|T ]. Part (b): To compute the variance of X we will use the conditional variance formula given by Var(X) = E[Var(X|T )] + Var(E[X|T ]) . 2 12 . We have λ E[X] = E[E[X|T ]] = E[λT ] = λE[T ] = . The expectation of X can now be computed by conditioning on the random variable T .e. the variance of this expression is related to the variance of T and is given by Var(E[X|T ]) = Var(λT ) = λ2 Var(T ) = λ2 . This expression has an expected value is given by E[X|T = t] = λt . the conditional variance is given by Var(X|T = t) = λt . in exactly the same way as we computed the expectation E[X|T = t]. We begin with the second term: Var(E[X|T ]). P {X = n|T = t} = e−λt (λt)n . 1). so that the expectation of this expression then gives E[Var(X|T )] = λE[T ] = λ . Now assuming that the number of people that arrive to wait for the next train is a Poisson process. 12 Now for the ﬁrst term on the right hand side of the conditional variance expansion (and the properties of the Poisson distribution). Since from the above the random variable E[X|T ] is given by λT .Exercise 50 (waiting for the train) Part (a): We will solve this problem by conditioning on the time of arrival of the train t. where T is the random variable denoting the arrival time of the next train after the previous train has left the station (we are told that this random variable is uniform). Speciﬁcally if X is the random variable denoting the number of passengers who get on the next train the we have E[X] = E[E[X|T ]] .

9084. Now for this speciﬁc problem the arrival of the particles at a rate of three a minute can be modeled as a Poisson process with a rate λ = 3 per minute. 0! . this corresponds to 2 P.M. Thus 4 4 e−λt (λt)k . With this information the various parts of the problem can be solved. Thus the process X(t) is a Poisson process with rate 2 particles per minute. i=1 where each Xi is an exponential random variable with parameter λ.M. This latter probability is given by P {N(2) = 0} + P {N(2) = 1} = e−2λ (2λ)0 e−2λ (2λ)1 + 0! 1! −2λ −2λ = e + e 2λ = 5e−4 . P {N(t) = 0} = 0! Part (b): A Poisson process has the time between events given by exponential random variables with parameter λ. and 8 P.1353 . Thus the probability we seek is given by 1 − 5e−4 = 0.Exercise 57 (Poisson events) Part (a): Since our events are generated by a Poisson process we know that P {N(t) = k} = so in one hour (t = 1) we have that e−λ (λ)0 = e−λ = e−2 = 0. λ λ Since λ is measured in units of reciprocal hours. t ≥} is a Poisson process with rate λ then the process that counts each of the events from the process N(t) with probability p is another Poisson process with rate pλ.M. (a two hour span) is the complement of the probability that less than two events occur in this two hour span. If {N(t). The fact that only 2/3 of the particles that arrive are actually measured means that this combined event can be modeled 2 using a Poisson process with a rate 3 λ = 2 particles per minute. Part (c): The probability that two or more events occur between 6 P. Part (a): In this case we have P {X(t) = 0} = e−2t (2t)0 = e−2t . k! E[S4 ] = i=1 E[Xi ] = i=1 4 1 = = 2. Exercise 58 (a Geiger counter) To solve this problem we will use the following fact. Thus the fourth event will occur at the time S4 = 4 Xi .

· · · . This would give for a probability (reasoning like above) of 1 3 2 3 + 1 3 2 3 + 1 3 1 3 = 5 . Part (a): We desire that both events land in the ﬁrst 20/60 = 1/3 of the total one hour time. Which if we deﬁne the event A to be the event that we have exactly one event in [ti . 9 We note that the situation we desire the probability for is also the complement of the situation where both events land in the last 2/3 of time. This can happen in several ways. 2. The ﬁrst (or second) event falls in this interval while the other event does not or both events happen in the ﬁrst third of our total interval. t). Then the probability that a random sample of n arrival times Si happen at the times ti and conditioned on the fact that we have n arrivals by time t can be computed by considering P {ti ≤ Si ≤ ti + hi . for i = 1. the location of any speciﬁc event is uniformly distributed over the time interval (0. · · · . 2. ti + hi ] for i = 1. n and no events in the other regions then (by deﬁnition) the above equals . This will happen with probability 1 3 2 = 1 . for i = 1. · · · n.Part (b): In this case we have that E[X(t)] = 2t . Using this idea the above probability could be calculated as 2 5 2 = . 2. 1− 3 9 Problem 81 (the distribution of the event times in a nonhomogenous process) Part (a): Following the same strategy that the book used to compute the distribution function for a homogeneous Poisson process. Exercise 60 (bank arrivals) Each part of this problem can be solved with the information that given n events have been observed by time t. 9 Part (b): In this case we desire that at least one of the events fall in the ﬁrst third of the total time. we will begin by assuming an ordered sequence of arrival times 0 < t1 < t2 < · · · < tn < tn+1 = t and let hi be small increments such that ti +hi < ti+1 . n|N(t) = n} .

(4) This result will be used in what follows. To evaluate P {A} we recognized that in the intervals (0. (tn + hn . tn+1 ) . no events occurs. Taking the limit hi → 0 and deﬁning t0 = 0 the above expression becomes n P (A0 ) = i=0 {1 − λ(ti )(ti+1 − ti ) + o(ti+1 − ti )} . one event occurs. (t1 + h1 . We begin our sequence of manipulations by following the derivation in the book and recognizing that we will eventually be taking the limits as hi → 0. (t2 + h2 . P {N(tr ) − N(tl ) = 0} + P {N(tr ) − N(tl ) = 1} + P {N(tr ) − N(tl ) ≥ 2} = 1 . 3 and the term o(h) represents terms higher than ﬁrst order in any of the hi ’s. tr ]. Speciﬁcally. t2 + h2 ). P {N(t) = n} The probability that we have one event in [ti . .the following expression P {A} . t1 + h1 ). (tn . or using properties (ii) and (iii) in the deﬁnition of a nonhomogenous Poisson process the above becomes (solving for P {N(tr ) − N(tl ) = 0}) the following P {N(tr ) − N(tl ) = 0} = 1 − λ(tl )(tr − tl ) + o(tr − tl ) . t1 ). By analogy with this result the contributed probability in the evaluation of P {A} from the intervals where the count does not increase P (A0 ) is given by n P (A0 ) = (1 − λ(0)(t1 ) + o(t1 )) i=1 {1 − λ(ti + hi )(ti+1 − ti − hi ) + o(ti+1 − ti − hi )} . we will derive this from the four properties in the deﬁnition a nonhomogenous Poisson process. since the total probability must sum to one we have the constraint on increment variables over the range of time [tl . t3 + h3 ). t3 ). (t3 . · · · . (t2 . where we have used Eq. The contributed probability P (A1 ) in the evaluation of P {A} from the intervals where the count increase by one is given by n n P (A1 ) = i=1 {λ(ti )hi + o(hi )} = i=1 λ(ti )hi + o(h) . This expression will take some manipulations to produce a desired expression. · · · . tn + hn ) . Since this expression has a ﬁnite limit we can take the limit of the above expression as is and simplify some of the notation. t2 ). By the independent increments property of nonhomogenous process the event A can be computed as the product of the probabilities of each of the above intervals event. ti + hi ] is given by the fourth property in the deﬁnition of a nonhomogenous Poisson and is given by P {N(ti + hi ) − N(ti ) = 1} = λ(ti )hi + o(hi ) (3) To calculate the probability that we have no events in a given interval. while in the intervals (t1 .

m(t) . t2 .We can simplify this product further by observing that the individual linear expressions we multiply can be written as an exponential which will facilitate our evaluation of this product. we can conclude that the probability of drawing a speciﬁc sample of n event times (i. m(t) which can only hold if x ≤ t. This is the desired result. · · · . while if x > t.Sn (t1 . F (x) = m′ (x) . Recognizing the above summation as an approximation to the integral of λ(·). obtaining a draw of the random variables Si ) for a nonhomogenous Poisson process with rate λ(t) given that we have seen n events by time t is given by n fS1 . It is this ﬁnal result we were after. After dividing by n hi and taking the limit where i=1 hi → 0. With these expressions for P (A1 ) and P (A0 ). Speciﬁcally. so that an equation for our cumulative distribution function F is given by F ′ (x) = This can be integrated to give m(x) . By the deﬁnition of the function m(·) we have that λ(x) = m′ (x). we can now evaluate our target expression P {A} P (A1 )P (A0) = P {N(t) = n} P {N(t) = n} n! = P (A1 )P (A0) −m(t) m(t)n e = n! −m(t) m(t)n e n n λ(ti )hi + o(h) e−m(t) i=1 = n! i=1 λ(ti ) hi + o(h) m(t) .··· . tn |N(t) = n) = n! i=1 λ(ti ) m(t) 0 < t1 < t2 < · · · < tn < t (5) We recognized that this expression is the same distribution as would be obtained for the order statistics corresponding to n independent random variables uniformly distributed with probability density function f (·) and a cumulative distribution function F (·) given by f (x) = λ(x) m(t) and F ′ (x) = f (x) . F (x) = 1.e. With this substitution the product above becomes a sum in the exponential and we have n n P (A0 ) = i=0 e−λ(ti )(ti+1 −ti ) = exp − i=0 λ(ti )(ti+1 − ti ) . it can be shown (using Taylor series) that e−λ(ti )(ti+1 −ti ) = 1 − λ(ti )(ti+1 − ti ) + o(ti+1 − ti ) .S2 . we see that the above is approximately equal to the following n t P (A0 ) = exp − i=0 λ(ti )(ti+1 − ti ) ≈ exp − 0 λ(τ )dτ = e−m(t) .

• We will compute the expression P {Nc (t + h) − Nc (t) = 1}. by deﬁning N(t) to be the nonhomogenous Poisson process with intensity function λ(t). Then we have P {Nc (t + h) − Nc (t) > 1} = P {N(t + h) − N(t) > 1|EG }P (EG ) = P {N(t + h) − N(t) > 1}P (EG ) = o(h)P (EG ) = o(h) . We then have • Because N(t) is a nonhomogenous Poisson process we have N(0) = 0. Because no events have occurred at time t = 0 we then must necessarily have Nc (0) = 0. Where we have used independence of Et and N(t). • To compute P {Nc (t + h) − Nc (t) > 1} deﬁne EG to be the event that in the interval (t. Part (b): The statement made in Part (a) can be proven by showing that Nc (t) satisﬁes the four conditions in the deﬁnition of a nonhomogenous Poisson process with intensity function p(t)λ(t). Thus since we desire this probability we begin by deﬁning the event Et to be the event that (as h goes to zero) we count the single event that happens in the interval (t. Thus we have shown that Nc (t) satisﬁes the four requirements in the deﬁnition of a nonhomogenous Poisson process with intensity function p(t)λ(t). This can be reasoned by recognizing that if the increment variables for our original nonhomogenous Poisson process N(t) are independent then when we select a subset of these events by considering the process Nc (t) and involving the probability function p(·) we cannot introduce dependencies. We begin. This is the complement of the event we count no events in this interval. . t + h) we count at least one event. We then have conditioning on Et that P {Nc (t + h) − Nc (t) = 1} = = = = P {N(t + h) − N(t) = 1|Et }P (Et ) P {N(t + h) − N(t) = 1}p(t) (λ(t)h + o(h))p(t) λ(t)p(t)h + o(h) . t ≥ 0} is a nonhomogenous process with an intensity function given by p(t)λ(t). • As N(t) has the property of independent increments Nc (t) will inherit this property also. Note that the desired event Nc (t + h) − Nc (t) = 1 only holds true if our original process has an event N(t + h) − N(t) = 1. t + h) from our original nonhomogenous Poisson process N(t).Problem 82 (selecting events from an nonhomogenous Poisson process) Part (a): We claim that the process {Nc (t).

we will stop for the second passenger . This means that we will stop to let oﬀ the ﬁrst passenger at a time of X1 . • To show that N(t) has independent increments consider two non overlapping intervals of time (t1 . Thus our ordering of time above introduced an ordering of mi ≡ m(ti ) and we see m(t1 ) < m(t2 ) ≤ m(t3 ) < m(t4 ) . m(t4 )) imply that same property for N(t). since N0 (t) is a Poisson process. m(t2 )) and (m(t3 ). Problem 72 (the departure of cable car riders) Part (a): By deﬁnition the last rider will depart after all n − 1 other riders have. We have • N(0) = N0 (m(0)) = N0 0 0 λ(s)ds = N0 (0) = 0. • As in the above we ﬁnd P {N(t + h) − N(t) > 1} = P {N0 (m(t) + m′ (t)h + o(h)) − N0 (m(t)) > 1} = o(m′ (t)h + o(h)) = o(h) . Then the since λ(·) is a non-negative function and the t’s are ordered as t1 < t2 ≤ t3 < t4 we have that m(t) = 0 λ(s)ds is an increasing function of t. • Consider P {N(t + h) − N(t) = 1} = = = = since m(t) = t 0 t P {N0 (m(t + h)) − N0 (m(t)) = 1} P {N0 (m(t) + m′ (t)h + o(h)) − N0 (m(t)) = 1} 1 · (m′ (t)h + o(h)) m′ (t)h + o(h) .Problem 83 (time rescaling results in a nonhomogenous Poisson processes) We will show that N(t) is an nonhomogenous Poisson process by showing that all four of the required properties hold. Thus we have shown that N(t) satisﬁes all four requirements for a nonhomogenous Poisson process with rate function λ(t) as we were asked to show. Lets assume that we start our cable car with all n passengers at the time t = 0 and deﬁne the random variables Xi to be the amount of time between successive stops. t2 ) and (t3 . t4 ). so the independent increments property of N0 (t) on the “transformed” intervals (m(t1 ). λ(s)ds we have that m′ (t) = λ(t) and the above becomes P {N(t + h) − N(t) = 1} = λ(t)h + o(h) .

Thus the last or nth passenger gets oﬀ the cable car at a time T given by T = n Xi . n). (n − 1)! Part (b): To solve this problem introduce the random variables Yi to be the time it takes passenger i to arrive home after he/she is dropped oﬀ. Combining this result with the discussion above we see that the probability that all the other . Since each Yi is an exponential random variable and each sum is a Gamma random variable.to get oﬀ at a time X1 + X2 . etc. Then for the next to last passenger (the n − 1st rider) to be home before the last passenger (the nth rider) gets oﬀ requires Yn−1 < Xn . Lemma: If Y is an exponential random variable with parameter µ and X is a Gamma random variable with parameters (λ. the requirement that the n − kth passenger get home before the nth rider has departed is k−1 Yn−k < i=0 Xn−i for k = 1. In all of these cases Xi is an exponentially distributed random variable with rate λ. λ). for the third passenger to get oﬀ at a time of X1 + X2 + X3 . Using the same logic for the n − 2 passenger to get home before the nth one gets oﬀ requires that Yn−2 < Xn + Xn−1 . That is the distribution function for T is given by fT (t) = λe−λt (λt)n−1 . to evaluate the probability of each of these events we need the following lemma. independent of Y . 2 · · · n − 1 . we have ∞ P {Y < X} = 0 ∞ P {Y < X|X = x}P {X = x}dx P {Y < x}λe−λx (λx)n−1 dx (n − 1)! 0 ∞ (λx)n−1 = (1 − e−µx )λe−λx dx (n − 1)! 0 λn . Given that the n − 1 passenger makes it home before the nth passenger gets oﬀ lets now consider the n − 2nd passenger. A sum of i=1 n exponentially distributed random variables each with a rate of λ is distributed as Gamma random variable with parameters (n. In general. Since Yn−1 is an exponential random variable with rate µ the probability this is µ true is given by µ+λ . = 1− (λ + µ)n = when we perform the required integration.

e Note that this is an exponential distribution with rate λp. We have P {N = n|T = t} = P {T = t|N = n}P {N = n} .riders are home (denoted here as Ph ) at the time the nth gets oﬀ is given by n−1 k−1 Ph = k=1 P Yn−k < i=0 Xn−i = P {Yn−1 < Xn }P {Yn−2 < Xn + Xn−1 } · · · P {Y1 < Xn + Xn−1 + · · · X2 } µ λ2 λ3 λn−1 = 1− 1− ··· 1− µ+λ (µ + λ)2 (µ + λ)3 (µ + λ)n−1 n−1 = k=1 1− λk (λ + µ)k . Because these events are from a Poisson process the times between individual events Xi are exponentially n distributed with a rate λ. Recall that a Poisson process with rate λ that has its . The second factor in the numerator P {N = n} represents the probability it takes n trails to get one “success” where a success represents the failure of the system. Since the distribution of the sum of n exponential random variables is distributed as a Gamma random variable we have that the above becomes P {T = t|N = n} = λe−λt (λt)n−1 . and we have n P {T = t|N = n} = P { i=1 Xi = t|N = n} . Problem 73 (shocks to the system) Part (a): We are told that it takes n shocks to cause our system to fail. The denominator P {T = t} can be calculated by conditioning on N as ∞ P {T = t} = n=1 P {T = t|N = n}P {N = n} ∞ −λt = λpe = λpe n=1 −λt λt(1−p) (λt(1 − p))n−1 (n − 1)! = λpe−λpt . This is given by a negative-binomial distribution and in this case is given by p(1 − p)n−1 . (n − 1)! Part (b): We want to compute P {N = n|T = t} which we can compute using an application of Bayes’ rule. Thus the time T can be written T = i=1 Xi . which could have been determined without any calculations as follows. P {T = t} The ﬁrst factor in the numerator was calculated in Part (a).

Since non-failure causing shocks occur according to a Poisson process with rate λ(1 − p). For example. then from the above expression the distribution of N ′ is given by P {N = n |T = t} = e ′ ′ −λ(1−p)t (λ(1 − p)t)n (n′ )! ′ for n′ = 0. .events “ﬁltered” with a Bernoulli process with parameter p is another Poisson process with rate λp. then the left hand side of the above expression is max(X1 . 2. This looks exactly like a Poisson distribution with a mean λ(1 − p)t but starts at the value n = 1. Using these parts we then have P {N = n|T = t} = λe−λt (λt)n−1 (n−1)! p(1 − p)n−1 λpe−λtp (λ(1 − p)t)n−1 = e−λ(1−p)t (n − 1)! for n = 1. If in fact the other relationship holds between X1 and X2 . is given by a Poisson random variable with mean λ(1 − p)t. 2. · · · could have been determined without calculation using the following argument. The probability P {N = n|T = t} means that the system failure happened on shock n at the time t. X2 ) = X1 + X2 − min(X1 . is the largest the summation expression will “over count” the maximum by i=i∗ Xi . Deﬁning the random variable N ′ as N ′ ≡ N − 1. Problem 93 (min/max identities) Part (a): To prove the given identity max(X1 . the number of such non-failure causing shocks. X1 ≥ X2 . X2 . exactly as calculated in Part (b). Since this second Poisson process represents the process of failures the ﬁrst failure time will be distributed as an exponential random variable with rate λp exactly as found above. X2 ) = X1 + X2 − X1 = X2 which are equal. So N ′ is a Poisson random variable with a mean λ(1 − p)t and we have that N is given as 1 + N ′ as we were to show. Part (c): The desired probability P {N = n|T = t} for n = 1. n − 1. This failure must have been caused by the last shock and requires that the previous n − 1 shocks did not result in failure. that is. · · · . if X1 ≤ X2 . say i=1 Xi∗ . Since one of the Xi . Part (b): I had a diﬃcult time proving this identity directly but can oﬀer the following arguments in support of the given expression when Xi ≥ 0. 2. Which is a fancy way of saying we can simply consider the two possible cases for the relationship between X1 and X2 . Consider the expression max(X1 . Xn ) in comparison to the expression n Xi . X2 ) . a similar calculation shows that the above identity to be true in this case also. · · · . 1. X2 ) = X2 while the right hand side is X1 + X2 − min(X1 . · · · . we can appeal to the law of the excluded middle.

Thus we now might consider the following n Xi − i=1 i<j min(Xi . · · · . This logic allows one to conclude that n max(X1 . Xi ). Xk ) = P (Ai Aj · · · Ak ) . Xj . Xj . · · · . Xn ) ≥ i=1 Xi − i<j min(Xi . Xj ) . X2 . Xj ) we would double count all possible minimums i.That is by all Xi that are less than Xi∗ and we have n max(X1 . Our approximation to the maximum expression can be improved on by adding terms like min(Xi . Xn ). Xj ) with i < j. This process is repeated until we add/subtract min(X1 . Xj ) = min(Xj . X2 . We need to include the requirement that i < j since otherwise considering all possible terms like min(Xi . Xk ) = P (Ai ∪ Aj · · · ∪ Ak ) and min(Xi . let X be a random . while the above gives 1 + 2 + 3 − (1 + 1 + 2) = 3 − 1 = 3 . For example if X1 = 1. X2 . Xn ) ≤ i=1 Xi . 2. In a similar way. · · · n to be an indicator random variables denoting whether or not the event Ai occurred we see that max(Xi . X2 = 2. X2 . Xj . Xk ) to give n max(X1 . X3 ) = 3. · · · . Xk ) . min(Xi . · · · . Xj ) . Whether we add or subtract depends on whether the number of variables we start with n is even or odd. X2 .e. If we deﬁne Xi for i = 1. Xn ) ≤ i=1 Xi − i<j min(Xi . Thus the value for X1 was subtracted oﬀ twice. With these expressions and using the max/min identity proven in Part (a) above we obtain the well known identity that n P (∪i=1 Ai ) = i P (Ai ) − i<j P (Ai Aj ) + i<j<k P (Ai Aj Ak ) + · · · + (−1)n−1 P (A1 A2 · · · An ) . · · · . · · · . Xj . let Xi be the random variable denoting the time when the ﬁrst event has occurred in the ith Poisson process. Terms that are smaller than Xi∗ can be represented as min(Xi . Note that in the above the second expression now subtracts too many terms since smaller terms will certainly appear more than once. Part (c): There seemed to be two ways to do this part of the problem. To explicitly use the results from above. We can “adjust” our sum on the right hand side to make it more like the left hand side by subtracting oﬀ all terms smaller than Xi∗ . and X3 = 3 then max(X1 . Xj ) + i<j<k min(Xi .

· · · . Xn )] . Then X = max(X1 .variable denoting the time at which and event has occurred in all n Poisson processes. Xj )] + i<j<k E[min(Xi . Xj . and earlier results from the book about the distribution of minimum of exponential random variables we ﬁnd n E[X] = i=1 1 − λi i<j 1 1 (−1)n−1 + +···+ . X2 . n. Using the independence of the Xi . Expectations of products of pairs of order statistics The Problem: Let X1 . From this we (7) E[X] = 0 1− i=1 (1 − e−λi t ) dt . · · · . X2 . From the above functional form it appears that one should be able to expand the product in the integrand. the fact that they are exponential with a rate λi . ∞ 0 Now we can compute E[X] if we know P {X > t} as E[X] = have ∞ n P {X > t}dt. We desire to calculate E[X]. Using the above deﬁnitions we ﬁnd P {X < t} = P {max(Xi ) < t} = P {Xi < t for all i} i n = i=1 (1 − e−λi t ) . Find . · · · . perform the integration. Also let X(i) be the ith smallest of these values i = 1. Xj ) + i<j<k min(Xi . Xn be independent exponential random variables. · · · . Taking the expectation of the above expression we ﬁnd that E[X] = i E[Xi ] − i<j E[min(Xi . Xn ). each having rate λ. From the identity in Part (b) we have that X can be expressed in term of minimization’s as X= i Xi − i<j min(Xi . Chapter 5: Additional Exercises The following are some problems that appeared in diﬀerent editions of this book in the chapter on exponential distributions and Poisson processes. Xn ) . · · · . Xj . λi + λj i<j<k λi + λj + λk λ1 + λ2 + · · · + λn (6) As another way to solve this problem that results in a quite diﬀerent formula consider the following where we compute the distribution function for X directly. and show that it agrees with the Equation 6. X2 . X2 . Xk )] + · · · + (−1)n−1 E[min(X1 . Xk ) + · · · + (−1)n−1 min(X1 .

n − 1.(a) E[X(1) X(2) ] (b) E[X(i) X(i+1) ] for i = 1. xj ) = λ2 n! [1 − e−λxi ]i−1 (i − 1)!(j − i − 1)!(n − j)! × [e−λxi − e−λxj ]j−i−1 e−λ[(n−j+1)xj +xi ] for xi < xj . xj ) = n! [F (xi )]i−1 (i − 1)!(j − i − 1)!(n − j)! × [F (xj ) − F (xi )]j−i−1 [1 − F (xj )]n−j f (xi )f (xj ) for xi < xj . for a Monte-Carlo veriﬁcation of the above see the Matlab ﬁle exp pp order stats pt a. Rather than deriving this density we simply state the result. In addition. . Part (a): When i = 1 and j = 2 our density function above becomes fX(1) . · · · .X(2) (x1 .X(2) (x1 .m and Figure 1. See the Mathematica ﬁle chap 5 eprob pp. x2 )dx2 dx1 ∞ x1 =0 ∞ ∞ = λ2 n(n − 1) x2 =x1 x1 x2 e−λ((n−1)x2 +x1 ) dx2 dx1 ∞ = λ2 n(n − 1) = 3n − 2 (n − 1) 1 nλ x1 =0 2 x1 e−λx1 x2 =x1 x2 e−λ(n−1)x2 dx2 dx1 . (c) E[X(i) X(j) ] for i < j. when i < j and integrate over the appropriate region. The Solution: To evaluate the above expectations we will simply use the joint density of the order statistic X(i) and X(j) .X(j) (xi . We have fX(i) . x2 ) = λ2 n(n − 1)e−λ((n−1)x2 +x1 ) So that our expectation is then given by ∞ ∞ x2 =x1 for x1 < x2 .nb for the algebra for this problem. which can be discussed in [1]. So that our density above becomes in this speciﬁc case the following fX(i) . 2.X(j) (xi . When X is an exponential random variable with rate λ we have that our density and distribution functions f and F are given by the following special forms fX (x) = λe−λx x ≥ 0 0 x<0 FX (x) = 1 − e−λx 1 − FX (x) = e−λx . E[X(1) X(2) ] = x1 =0 x1 x2 fX(1) .

5 lambda 0. See the Mathematica ﬁle chap 5 eprob pp.5 0.X(i+1) (xi . xi+1 ) = λ2 n! (1 − e−λxi )i−1 e−λ[(n−i)xi+1 +xi ] (i − 1)!(n − i − 1)! for xi < xi+1 .X(i+1) (xi .02 0.08 0.05 0.1 300 The exact expectation 0.MC est.02 0. Again since we expect the results to scale as 1/λ2 .01 10 15 n index 20 25 0.03 0. we have assumed that λ = 1.09 350 0.nb for some of the algebra.08 3 300 0.03 50 0.06 0. Left: the Monte-Carlo simulation plotted as a function of λ and n. Part (c): As in Part (b) an exact analytic solution is quite diﬃcult so in the Matlab ﬁle exp pp order stats pt c.m and Figure 2 for the computational experiment done for this part of the problem.04 1 0.04 20 0. we have assumed λ = 1 for the plots presented.07 200 250 200 lambda 0.5 0. A literature search reveals that this is a rather diﬃcult problem to solve exactly. So that our expectation is then given by ∞ ∞ xi+1 =xi 2 E[X(i) X(i+1) ] = xi =0 xi xi+1 fX(i) . Since we expect the expression to scale as 1/λ2 .1 Diff between the exact and approx: nSamples=50000 3. Part (b): When j = i + 1 our joint density function becomes fX(i) . I have presented the results from the Monte-Carlo calculation done to evaluate the requested expression.m and Figure 3.08 0. . At this point even performing algebra by a computer becomes diﬃcult.04 100 50 0 0.02 15 0 10 n index 25 0. so instead I’ll use a Monte-Carlo calculation to compute these expressions.1 0.5 150 100 0.09 250 0.06 150 0.07 2. of the expectation: nSamples=50000 0. See the Matlab ﬁle exp pp order stats pt b. Right: The error between the two. Middle: The analytical solution derived above.05 1.01 10 15 n index 20 25 lambda Figure 1: A comparison between a Monte-Carlo simulation of E[x(1) x(2) ] and the derived analytical expression.06 2 0. xi+1 )dxi+1 dxi = × λ n! (i − 1)!(n − i − 1)! ∞ xi =0 ∞ xi xi+1 (1 − e−λxi )i−1 e−λ((n−i)xi+1 +xi ) dxi+1 dxi λ n! (i − 1)!(n − i − 1)! ∞ ∞ xi+1 =xi 2 = × xi (1 − e−λxi )i−1 e−λxi xi =0 xi+1 =xi xi+1 e−λ(n−i)xi+1 dxi+1 dxi .

A Identity with Reciprocal Sums The Problem: Argue that if λi . λj − λi Since X is the sum of n exponential random variables when we multiply both sides of the above by x and integrate the left hand side becomes the expectation of random variable X = n Xi .2. λi i=1 While the right hand side results in a another expectation of sums of exponential random variables.4 we know that the random variable X deﬁned as the sum of n exponential random variables X = n Xi is called a hypo-exponential random i=1 variable and has a probability density function given by n fX1 +X2 +···+Xn (x) = i=1 j=i λj λi e−λi t . Taking these expectations we have an expression given by n i=1 1 λi j=i λj . · · · . 2. The Solution: From Section 5. Figure 3: The Monte-Carlo simulation of E[x(i) x(j) ] as a function of i and j for i < j. 2. .4. i = 1. and equals i=1 n 1 . λi Hint: Relate this problem to Section 5. n−1.2.The appoximate solution 9 8 7 6 E[xi xi+1] 5 4 3 2 1 0 0 2 4 6 i index 8 10 12 14 Figure 2: The Monte-Carlo simulation of E[x(i) x(i+1) ] as a function of i for i = 1. · · · . λj − λi When we equate these two results we have the requested theorem. n are distinct positive numbers then n i=1 1 λi j=i λj = λj − λi n i=1 1 .

from the above we see that f satisﬁes f (t + s) = f (t)f (s) . Thus deﬁning f (t) ≡ P {N(t) = 0}.s. what is the expected number of cars to have passes by in that time? . Hint: Derive a functional equation for g(t) = P {N(t) = 0}. for some λ.We can derive the fact that P {N(t) = 1} = λt + o(t) The Problem: Show that assumption (iv) of Deﬁnition 5.3 and the fact that probabilities must be normalized (sum to one) we have that P {N(t) = 0} + P {N(t) = 1} + P {N(t) ≥ 2} = 1 . which gives us (solving for P {N(t) = 1}) the following P {N(t) = 1} = = = = as we were to show. By the discussion in the book the unique continuous solution to this equation is f (t) = e−λt . N(t + s) − N(t) = 0} P {N(t) − N(0) = 0.3 follows from assumptions (ii) and (iii). 1 − P {N(t) = 0} − P {N(t) ≥ 2} 1 − e−λt − o(t) 1 − (1 − λt + o(t)) − o(t) λt + o(t) . by considering P {N(t + s) = 0}. Using (iii) from Deﬁnition 5. If 5 percent of the cars on the road are vans. so we have that P {N(t + s) = 0} = = = = P {N(t) = 0. Thus we have that P {N(t) = 0} = e−λt . N(t + s) − N(t) = 0} P {N(t) − N(0) = 0}P {N(t + s) − N(t) = 0} P {N(t) = 0}P {N(s) = 0} . Now if N(t + s) = 0. The Solution: Following the hint for this problem we will try to derive a functional relationship for P {N(t) = 0}. then • what is the probability that at least one van passes during the hour? • given that ten vans have passed by in an hour. When we used the property of stationary independent increments. this event is equivalent to the event that N(t) = 0 and N(t + s) − N(t) = 0. Cars v. Vans on the Highway The Problem: Cars pass a point on the highway at a Poisson rate of one per minute.

where the probability of success is given by p = 0. p) = (10.05).05). p For this part of the problem at hand (r.05). p) = (50. so the expected number of cars that must have passed is given by 10 = 200 . . If we deﬁne the random variable X to be the number of trials needed to get r success where each success occurs with probability p we have that r E[X] = .9502 . Because of this fact the probability that at least one van passes by in one hour t = 60 (minutes) is given by P {Nv (60) ≥ 1} = = = = 1 − P {Nv (60) = 0} 1 − e−pλ60 1 − e−0. This is the deﬁnition of a negative binomial random variable. so the answer to the question posed is given by 50 5 (0. the probability that n from 50 are vans is a binomial random variable with parameters (n.95)45 = 0.• if 50 cars have passed by in an hour.06584 .05·1·60 0. 0. 0. 0. where p is the fraction of the cars that are vans and λ is the Poisson rate of the cars (we are told these values are λ = 1 (per minute) and p = 0.05 Part (c): If 50 cars have passed by and we have a 5 percent chance that each car is a van.05)5 (0. what is the probability that ﬁve of them are vans? The Solution: Part (a): The number of vans that have passed by in time Nv (t) is given by a Poisson process with rate pλ. Part (b): We can think of the question posed: how many cars need to pass by to given ten vans as equivalent to the statement of the expected number of trials will we need to perform to guarantee ten successes.05.

. 3rd edition.References [1] S. Ross. A First Course in Probability. Macmillan. 1988.

Sign up to vote on this title

UsefulNot useful- Solution Manual for Introduction to Probability Models 7thby José Carlos
- Ross, Sheldon M - Introduction to Probability Models (Sixth Ed)by Zitein Dhuria
- Applied Probability Models with Optimization Applications By Sheldon M Ross solution manual 2007by Mayez Chikhany
- Ross Solution Manual_OddExersicesby tlanggy

- Solution Manual for Introduction to Probability Models 7th
- Ross, Sheldon M - Introduction to Probability Models (Sixth Ed)
- Applied Probability Models with Optimization Applications By Sheldon M Ross solution manual 2007
- Ross Solution Manual_OddExersices
- HW3 Selected Solutions
- Sheldon Ross
- Solution.manual.to.Introduction.to.Mathematical.statistics..6ed..Hogg..McKean.and.Craig
- Stochastic Processes - Ross
- Solutions Manual - Introduction to Probability (2e) by Dimitri P. Bertsekas & John N. Tsitsiklis
- Weatherwax Ross Solutions
- Stochastic Solutions Manual
- HW5 Selected Solutions
- Chapter 06
- Chapter 05
- Chapter 16
- Chapter 04
- Chapter 14
- Weather Wax Ross Solutions
- Ross S.M. Introduction to Probability Models(6ed., AP, s
- Stochastic Process_Sheldon Ross 2ed
- A First Course in Stochastic Processes - Karlin S., Taylor H.M
- solution manual
- Solution Chapter8
- Introduction to Stochastic Processes
- Markov Chain Random Walk
- Mathematical Modelling
- Bowers Solutions
- Chapter 03
- Complex Analysis (Solutions) - Stein
- Weather Wax Ross Solutions