Prefatory Comments

The book has been well recieved in reviews. I would much appreciate feedback
on its use.

Ian Jordaan
jordaan@nf.sympatico.ca

@seismicisolation
@seismicisolation
Corrigenda
Chapter 3
Exercise 3.22:...exponential distribution
µ ¶
1 x
fX (x) = exp − for x ≥ 0
β β

Chapter 4
Figure 4.13: values on the y-axis below the x-axis should have a minus sign
added; Utiles of −20, −40, etc.
Figure 4.25(b): end of upper branch, should read −0.5
Exercise 4.7: units of intensity are Wm−2 s−1

Chapter 5
p 224, Example 5.1, second line from bottom:
...; for B τ 1 is ‘heads’ and...
p 230, Figure 5.3: dimension should read

nx =

(curly )

Chapter 6
Example 6.4, after Equation 6.35 add full stop, omit "and where. . . Lagrangian
L"

Chapter 8
p. 383 Line after (8.1), “Series (8.1)”, not “Equation (8.1)”
p. 383 Line before (8.2), “...using the usual notation
Ei = 1 (success) and Ei = 0 or Ẽi = 1 (failure):”
p. 416 bottom line "theparameter" → "the parameter"
p. 432 Note: varkapa (κ 2 ) used for particular values of of chisquare (χ2 )
p 435 under “Methodology”, line 2 “An hypothesis is an assertion
..."; line 6, ...“on a road construction site...”

@seismicisolation
@seismicisolation
 Figure 1:

Chapter 9
Figure 9.3: Ordinate (y-axis) should be fZ (z) (probability density), not FZ (z)
Corrections to the ordinate scales (y-axes) of Figures:
Figure 9.3(c), multiply scale by 0.1
Figure 9.4, divide scale by 0.01
Figure 9.6, divide scale by 5
Figure 9.7 divide scale by 5
Equations (9.30) and (9.31) should be written as
" µ ¶ #
2
4h h
fH (h) = 2 exp −2 .
hs hs
and " µ ¶2 #
h
FH (h) = 1 − exp −2 .
hs
respectively.

Exercise
9.9 An approximation that is often useful for very small p (¿ 1/n) is
n
(1 − p) ' 1 − np. ((9.172))

@seismicisolation
@seismicisolation
Figure 2: Figure 9.4, divide scale by 0.01

Figure 3: Figure 9.6, divide scale by 5

@seismicisolation
@seismicisolation
 Figure 4: Figure 9.7 divide scale by 5

(a) Apply this in Exercise 9.8.

(b) Equation (9.49) states that
µ ¶n
n 1 1
FZ (xn ) = FX (xn ) = 1 − → ,
n e

in other words the n-year load on a structure has a 63% chance of

being exceeded in n years for large n. A designer wishes therefore to
choose a more conservative value. He then considers a value z̆ with
a small probability of exceedance on the n-year load pne such that

pne = 1 − FZ (z̆) ((9.173))

rather than the mode. Then if X is the parent,

[FX (z̆)]n = 1 − pne , or ((9.174))

pne
FX (z̆) = (1 − pne )1/n '1− , ((9.175))
n
using a variation of the rule (9.172). Obtain pne in terms of an , bn
and α for a Weibull parent distribution based on Equation (9.119),
with X replacing W :
· µ ¶α ¸
x − an
FX (x) = 1 − exp − . ((9.176))
bn

@seismicisolation
@seismicisolation
 An engineer in offshore practice would realize that a load factor is
usually applied to the 100-year load, giving a result that approaches
the situation described.

Chapter 10
Figure 10.7(a): (Clarification) This is drawn correponding to the proof for Equa-
tion (10.6) rather thqn (10.5):
dpf = fT (x) [1 − FS (x)] dx.
Figure 10.8: σ and µ should have upper case subscripts, thus: σM and µM

Chapter 11
p 603 7 lines from bottom, “girth” should be “girths”

Appendix
p 633 Lognormal distribution add subscripts “Y ” to µ and σ where appropriate:

Form
¡ ¢
If a random quantity X = ln Y¡ is normally
¢ distributed µ, σ 2 , then Y is said
to be lognormally distributed µY , σ 2Y .

pdf " #
1 −1 (ln y − µ)2
fY (y) = √ y exp − , for y > 0 (1)
σ 2π 2σ 2

cdf

Mean and Variance

¡ ¢
Let a = exp µ, b = exp σ 2

µ ¶
√ σ2
µY = a b = exp µ + (2)
2
¡ ¢£ ¡ ¢ ¤
σ 2Y = a2 b (b − 1) = exp 2µ + σ 2 exp σ 2 − 1 (3)
Mode is at ¡ ¢
exp µ − σ 2 (4)
Median is at
exp µ (5)

@seismicisolation
@seismicisolation
Solutions for Chapter 1
Exercise 1.1
1. Draw a decision tree, involving choice and chance forks, for each of

(a) a personal decision (for example, for which term should one renew a
mortgage; one three or five years?).
(b) an engineering problem.

Solution
Everyday decisions
Every decision in our lives is by choice and chance, preference and probability,
in some combination, sometimes one dominating, sometimes the other, usually
both being important. It is helpful for beginners to practice drawing decision
trees so as to idealize problems, much as engineers studying mechanics learn to
draw free body diagrams.
To start with the question of the renewing a mortgage: if the interest rate
at present is low, and we feel that it might rise, then we would consider a longer
term mortgage. Conversely, if we foresee rates declining, a short term mortgage
might be a good idea, so that one can remortgage later at a lower rate. See Tree
1.

Tree 1

We might be considering which party to go to on a Friday night; we are

lucky to have two invitations (party A and Party B in Tree 2).

@seismicisolation
@seismicisolation
 Tree 2

We consider here a winelover who detests beer. The decision-maker has to

gauge whether there is a greater chance of wine, as against beer at party A or B.
Of course, there are many other choices at parties; presence of women or men,
for example, will attract depending on the decision-maker’s preferences.

Engineering decisions

Tree 3

@seismicisolation
@seismicisolation
 Should one design for a very high load or demand, or is a lower demand level
adequate? The decisions hinges on the judgement of probability of the actual
occurrence of high or low loads (demand) and the consequences of failure.
The book contains many examples of engineering decision trees.

@seismicisolation
@seismicisolation
Solutions for Chapter 2
Exercise 2.1
An engineer states that the odds of a beam surviving a load of 3 MN for 25
years is “5 to 1 on”. What is the engineer’s probability for this event? Draw a
decision tree as in Figure 2.1(a) to illustrate the assignment of probability.

Solution
Let E be the event. Then
5 5
Pr(E) = =
5+1 6
r
since odds = r = 5 to 1, and Pr(E) = .
r+1

6
1
0
Status quo
Exercise 2.2
The odds of a prototype microchip surviving a current of 3 µA for 2 hours are
estimated as 4 to 3 against. What is the estimated probability of this event?
Draw a decision tree as in Figure 2.1(a) to illustrate the estimator’s assignment
of probability.

@seismicisolation
@seismicisolation
Solution
Let E be the event. Odds of "4 to 3 against" are the same as "3 to 4 on." Then
r = 3/4, and
3 3
Pr(E) = =
3+4 7

7
3
0
Status quo
Exercise 2.3
The event E = (the score on a roll of a fair die is a composite number; that is,
a non-prime number greater than 1). Calculate the odds on E, expressed as a
ratio reduced to its lowest terms.

Solution
E = {4, 6} , and
2 1
Pr(E) = =
6 3
2 1
Pr(E) = =
6 3 ¡1¢
Odds on E are = ¡ 32 ¢ , or 1:2 on, or 2:1 against.
3

@seismicisolation
@seismicisolation
Exercise 2.4
Odds vs. bookie’s odds. In a five horse race, a bookmaker sets a stake of $240.
The bookie quotes odds of r1 = 5 to 3 on, r2 = 2 to 1 against, r3 = 5 to 1
against, r4 = 19 to 1 against and r5 = 19 to 1 against. Calculate “probabilities”
qi (i = 1, 2, . . . , 5) based on these odds. You may place a bet of $(240 · qi ) on
any of the five horses and if event Ei (= horse i wins) occurs, then you win the
bookie’s stake of $240.

(a) From your point of view, do these odds constitute fair bets? Are the qi
acceptable as probabilities? Justify your answer by determining whether
or not the probabilities associated with these odds are coherent.
(b) What is the bookie’s guaranteed profit if you place one bet on each of all
five horses?

Solution
(a) µ ¶
5 5/3 5
r1 = ⇒ q1 = =
3 8/3 8
µ ¶
1 1/2 1
r2 = ⇒ q2 = =
2 3/2 3
µ ¶
1 1/5 1
r3 = ⇒ q3 = =
5 6/5 6
µ ¶
1 1/19 1
r4 = r5 = ⇒ q4 = q5 = =
19 20/19 20

The sum of these five probabilities is 49/40 which is not equal to 1. The
probabilities are therefore incoherent.

(b) The bookies net gain if horse h wins is

³X ´
gh = pi si − sh
h³X ´ i
= s pi − 1

(because the bookie’s stake is the same for all five horses).

(49 − 40)
= 240 × = 54,
40
(using pi = qi ).
The guaranteed profit is $54.

@seismicisolation
@seismicisolation
Exercise 2.5
Fair vs. unfair odds. Six teams are competing in a design competition. An
engineer offers a stake of $100 to be won by a colleague who guesses the winner.
He offers odds of 19 to 1 against team A, 9 to 1 against team B, 9 to 1 against C,
3 to 1 on team D, 4 to 1 against E, and 7 to 3 against F. Calculate “probabilities”
based on these odds. Are they coherent? If one person bets on each team,
what is the engineer’s guaranteed profit? Assume that the odds reflect well the
relative likelihood of each team winning, so that for example teams B and C
are considered equally likely to win. Rescale the probabilities and odds so that
the bets become fair bets and the probabilities become coherent.

Solution
1 1
pA = = = 0.05
19 + 1 20
1 1
pB = = = 0.10
9+1 10
1
pC = pB = = 0.10
10
3 3
pD = = = 0.75
3+1 4
1 1
pE = = = 0.20
4+1 5
3 3
pF = = = 0.30
3+7 10
The sum of the probabilities is 1.50, so they are not coherent. For a stake of
$100, the engineer receives bets of $100 × pi for each outcome i. The engineer’s
total receipts are thus

$5 + $10 + $10 + $75 + $20 + $30 = $150.

Only one of these six bets will receive the stake of $100 from the engineer.
The engineer’s guaranteed profit
P is therefore $150 − $100 = $50.
µ For
¶ coherence we require pi = 1. Adjusting each probability by a factor
1 2
= will achieve coherence.
1.5 3
µ ¶ µ ¶ ¡1¢
2 1 1 pA 1
pA = × = ⇒ rA = = ¡ 30 ¢
29 = 29
3 20 30 (1 − pA ) 30
or 29:1 against.

@seismicisolation
@seismicisolation
 µ ¶ µ ¶ ¡1¢
2 1 1 pB 1
pC = pB = × = ⇒ rC = rB = = ¡ 15
14
¢=
3 10 15 (1 − pB ) 15
14

or 14:1 against.
µ ¶ µ ¶ ¡1¢
2 3 1 pD
pD = × = ⇒ rD = = ¡ 21 ¢ = 1
3 4 2 (1 − pD ) 2

or even odds (1:1).

µ ¶ µ ¶ ¡2¢
2 1 2 pE 2
pE = × = ⇒ rE = = ¡ 15 ¢
13 = 13
3 5 15 (1 − pE ) 15
or 13:2 against.
µ ¶ µ ¶ ¡1¢
2 3 1 pF 1
pF = × = ⇒ rF = = ¡ 54 ¢ =
3 10 5 (1 − pF ) 5
4
or 4:1 against.

Exercise 2.6
Derive the condition of coherence for a bookmaker by using equation (2.7) with-
out making the assumption that s1 = s2 = · · · = sn = s. (de Finetti, 1937).

Solution
Equation (2.6) is
Xn
mh = pi si − sh ; h = 1, 2, . . . , n.
i=1

This can be written as

mh = [q]sh ,
where  
p1 − 1 p2 p3 ··· pn
 p1 p2 − 1 p3 ··· pn 
 
 p1 p2 p3 − 1 ··· pn 
[q] =  
 .. .. .. .. 
 . . . . 
p1 p2 p3 ··· pn − 1
How can we prevent the bookmaker from solving for s to find arbitrary m? The
answer is: arrange your values p1 , p2 , . . . , pn so that he cannot invert [q]. This
operation, the inversion of [q], is possible provided that the determinant |q| 6= 0.

@seismicisolation
@seismicisolation
Since you control [q] via your probabilities, you will arrange the latter so that
|q| = 0. It can be shown that
Xn
|q| = (−1)n−1 p1
i=1

by carrying out the following operation on the determinant of the matrix :

1. Take (−1) times the first row and add this to the 2nd, 3rd, . . ., nth rows.
2. Add the 2nd, 3rd, . . ., nth columns to the first column.

These operations do not change the value of the determinant, and we have
¯ P ¯
¯ pi − 1 p2 · · · pn ¯¯
¯
¯ 0 −1 · · · 0 ¯¯
¯
¯ 0 0 · · · 0 ¯¯
|q| = ¯
¯ .. .. .. ¯
¯ . . . ¯¯
¯
¯ 0 0 · · · −1 ¯

The determinant is equal to the sum of the products of the elements of the first
row and their cofactors; The consequence of equating |q| = 0 is simply expressed
as Xn
pi = 1.
i=1

Exercise 2.7 - 2.11

Solutions not needed.

Exercise 2.12
Expand equation (2.17) for n = 3. Write the corresponding probabilistic equa-
tion using the principle in Equations (2.11) and (2.12).

Solution
From equation (2.17)

Y3
E1 ∨ E2 ∨ E3 = 1− (1 − Ei )
i=1
= 1 − (1 − E1 ) (1 − E2 ) (1 − E3 )
= E1 + E2 + E3 − E1 E2 − E1 E3 − E2 E3 + E1 E2 E3

Therefore using equation (2.10)

@seismicisolation
@seismicisolation
 Pr(E1 ∨ E2 ∨ E3 ) = Pr(E1 ) + Pr(E2 ) + Pr(E3 ) − Pr(E1 E2 )
−Pr(E1 E3 ) − Pr(E2 E3 ) + Pr(E1 E2 E3 )

Can check with Venn Diagram:

E1 E2

E3

Exercise 2.13
For the sample space S which consists of the first ten natural numbers {1, 2, . . . , 10},
events A, B, C are defined as: A = {even numbers in S}, B = {prime numbers
in S}, and C = {3, 5}.

(a) Represent these sets in a Venn diagram.

(b) Express the set D = {even prime numbers in S} as a set function of the
above sets. How many elements are in this set?
(c) Simplify the expression B ∩ C.
(d) Simplify the expression B ∪ C.
(e) Simplify the expression A ∪ C.
(f) What elements are in the set à ∩ B̃ ∩ C̃? What are the elements in the
set à ∩ B̃ ∩ C̃ ?

@seismicisolation
@seismicisolation
Solution
(a)

A 10 7 B
8
2
6 5
4 3
C
1 9
(b) D = {2} = A ∩ B. n(D) = 1.
(c) The set C is entirely inside (is a subset of) the set B. Therefore B∩C = C.
(d) For the same reason as in (c), B ∪ C = B.
(e) Sets A and C are mutually exclusive (no overlap at all). Therefore their
intersection is empty: A ∩ C = ∅.
(f) These are numbers which are neither even nor prime. Then ˜A∩˜B∩˜C =
{1, 9} .

[Note: 1 is not prime because it does not have exactly two factors.]

Exercise 2.14
There are three possible causes for failure of a computer system. One or more
of the causes may be associated with a failure. Let Ai = {cause i is associated
with a failure}, for i = 1, 2, 3 and suppose that Pr (A1 ) = 0.30, Pr[A2 ] =
0.39, Pr (A3 ) = 0.32, Pr (A1 ∧ A2 ) = 0.16, Pr (A2 ∧ A3 ) = 0.12, Pr (A3 ∧ A1 ) =
0.12, and Pr (A1 ∧ A2 ∧ A3 ) = 0.02.Draw a Venn diagram that represents this
situation. Calculate the probability of each of these events:

(a) A1 ∨ A2
(b) Ã1 ∧ Ã2

@seismicisolation
@seismicisolation
 (c) A1 ∨ A2 ∨ A3
(d) Ã1 ∧ Ã2 ∧ Ã3
(e) Ã1 ∧ Ã2 ∧ A3

Solution

A1 A2
a b c
0.02
d e

f
A3

Pr[A1 ∧ A2 ] = 0.16 = b + 0.02 ⇒ b = 0.14

Pr[A2 ∧ A3 ] = 0.12 = b + 0.02 ⇒ e = 0.10
Pr[A3 ∧ A1 ] = 0.12 = b + 0.02 ⇒ d = 0.10

@seismicisolation
@seismicisolation
 A1 A2
a 0.14
c
0.02
0.10 0.10

f
A3
Pr[A1 ] = 0.30 = a + 0.10 + 0.02 + 0.14 ⇒ a = 0.04
Pr[A2 ] = 0.39 = c + 0.10 + 0.02 + 0.14 ⇒ c = 0.13
Pr[A3 ] = 0.32 = f + 0.10 + 0.02 + 0.10 ⇒ f = 0.10

The region outside all three events has a probability of

1 − (0.02 + a + b + c + d + e + f ) = 0.0.37

A1 A2
0.04 0.13
0.14
0.02
0.10 0.10

0.10
0.37 A3
10

@seismicisolation
@seismicisolation
 Values (a) to (e):

Pr (A1 ∨ A2 ) = 0.53
³ ´
Pr Ã1 ∧ Ã2 = 0.47
Pr (A1 ∨ A2 ∨ A3 ) = 0.63
³ ´
Pr Ã1 ∧ Ã2 ∧ Ã3 = 0.37
³ ´
Pr Ã1 ∧ Ã2 ∧ A3 = 0.10

Exercise 2.15
The quality of mass produced ceramic components for an electricity supply
system is checked by an inspection system. The products are unacceptable if
they are fractured, oversize or warped. From past experience, the probability
that a component is warped is 0.0015, the probability that it is fractured is
0.0020, the probability that it is oversize is 0.0014, the probability that it is
either warped or fractured or both is 0.0028, the probability that it is either
fractured or oversize or both is 0.0029, the probability that it is either oversize
or warped or both is 0.0025 and finally the probability that it has at least one
of these imperfections is 0.0034.

(a) What is the probability that a component is free of all imperfections?

(b) What is the probability that a component is oversize but is free of other
imperfections?
(c) What is the probability that a component has all of the specified imper-
fections?

Solution
Let
A = "specimen is warped"
B = "specimen is fractured"
C = "specimen is oversize"

(a)

Pr(no imperfections) = Pr(Ã ∧ B̃ ∧ C̃)

= 1 − Pr(A ∨ B ∨ C) (de Morgan’s Law)
= 1 − 0.0034 = 0.9966

11

@seismicisolation
@seismicisolation
 (b)
Pr(oversize only) = Pr(C ∧ Ã ∧ B̃)
= Pr(A ∨ B ∨ C) − Pr(A ∨ B)
= 0.0034 − 0.0028
= 0.0006

(c) This is the intersection of A, B and C. The probabilities for the various
events have been computed in the Venn Diagram below.

Pr(A ∧ B ∧ C) = 0.0001

A B
5 6 9
1
3 4 Divide by
10,000 for
probabilities
6
9966 C
Exercise 2.16
In a component failure, there are three possible causes Ei , i = 1, 2, 3. At failure,
a reliability analysis indicates that
E1 + 2E1 E2 + E3 = 1.

Solution
What can be said regarding the values Pr(Ei )?
By equation (2.10)

Pr(E1 ) + 2Pr(E1 E2 ) + Pr(E3 ) = 1

12

@seismicisolation
@seismicisolation
Exercise 2.17
The six faces of a cube of side 10 cm have been painted purple. The cube is
then cut into 1000 smaller cubes, each of side 1 cm, and these smaller cubes are
placed in an urn. One of these cubes is selected at random from the urn. What
is the probability that this cube will have two painted faces?

Solution
The edges of the large cube will have two sides painted, except the corners which
have three sides painted. The large cube has 12 edges, each 10 cm long. Of this
10 cm, 8 cm corresponds to edges that have two sides painted.
12 × 8
Required probability = = 0.096
1000

Exercise 2.18
Consider Raiffa’s urns of Figure 2.10. An urn is chosen at random from a room
containing 800 of type θ1 and 200 of type θ2 . Two balls are sampled without
replacement. What are the probabilities that the urn chosen is a θ1 or θ2 urn if
the sampled balls are
(a) both red;
(b) both pink; and
(c) red and pink?

Solution
(4/10)x(3/9)
RR
RP (4/10)x(6/9)

PR
1 (6/10)x(4/9)
0.8 (6/10)x(5/9)
(9/10)x(8/9)
2 RR
0.2 RP (9/10)x(1/9)

PR
(1/10)x(9/9)
PP (1/10)x(0/10)
Extensive Form

13

@seismicisolation
@seismicisolation
 (a)
4 3 8
10 × 9 × 10 2
Pr(θ1 |RR) = 4 =
10 × 9 + 10 × 89 × 10
3 9 2 5
3
Pr(θ2 |RR) =
5

(b)
6 5 8
10 × 9 × 10
Pr(θ1 |P P ) = 6 5 8 =1
10 × 9 × 10 +0
Pr(θ2 |P P ) = 0

(c)
4 6 8
10 × 9 × 10 32
Pr(θ1 |RP or P R) = 4 6 8 9 1 2 =
10 × 9 × 10 + 10 × 9 × 10
35
3
Pr(θ2 |RP or P R) =
35

Exercise 2.19
Everyday life 1.

(a) What kind of bets for the future are manifested in “puts and calls” in
security trading? Other bets, insurance or on sporting events might be
pondered.
(b) A Canadian automobile insurance company threatens to double the insur-
ance rate for a client if she puts in a claim after a minor accident. But they
will leave the rate the same if the claim is not entered even though they
know about the accident. Bayes’ theorem applies equally well as regards
the probability of a future accident whether the claim is entered or not
since the company knows of the accident. The client’s future risk is the
same in either case. Could it be that the insurance company is attempting
to force the client to pay for her repairs herself? Comment on this quite
common practice.

Solution
(a) A put is a contract that allows the holder to sell a stock at a specific
price within a specified period of time. A call gives the holder the right
to purchase a stock under similar conditions of specified price and period
of time. These are options. A put can, for example, be used to make a
profit from a decline in stock prices. The investor holding the put is able
to sell the stock at the higher price specified in the put. If the stock price

14

@seismicisolation
@seismicisolation
 rises, the investor can choose not to exercise the put option. In this case,
the investor loses the fee paid to obtain the option. Calls and puts are
bets as to future movements in the stock market.
(b) This practice does seem to be aimed at making the insured pay rather
than the insurer.

Exercise 2.20
Everyday life 2. An elderly public figure is accused by several women of his age of
having committed serious improprieties in his youth involving them individually.
The women do not know each other and appear reluctant to give evidence. They
are ordinary people with no apparent motive falsely to accuse the public figure.
Comment on the probability of guilt.

Solution
Probability and the law. The key point with regard to guilt is that it is either
true or it is not that the accused is guilty, but only very rarely will the judge or
jury have sufficient evidence to be certain of guilt. The example is one such
case. It is logically inadmissible, except in rare instances, for any pronounce-
ment of the kind “this person is guilty as charged” to be made. There can only
be a certain probability of guilt—not acknowledged in legal practice. It is gen-
erally permissible to make only a probabilistic statement based on the evidence
at hand. Much of the logic applied in the legal profession keeps probabilities in
the background (“beyond reasonable doubt”, for example) or ignores them com-
pletely. Examples abound of rather curious practices: for example, The Times
of London in September 24, 1982, reported the case of the “palace intruder” who
was acquitted by a jury of stealing a bottle of wine from Buckingham Palace
after he had admitted to taking and drinking it. In discussing probability and
the law, de Finetti (1974, Vol. 1, p. 158) goes so far as to suggest that “the
world will probably remain for quite some time at the mercy of a mentality so
distorted and arrogant that it neither retracts nor hesitates even when faced
with the most grotesque absurdities”.
In the example cited in the Exercise, the probability of guilt would approach
unity, on any reasonable assessment. The probability of guilt is very high since
the evidence comprises an independent series of allegations. The probability of
false accusations in the individual cases is very low. See equation below,
Yet the accused person was acquitted!
(Note: Based on a real case in Canada)

Pr(guilt | evidence)
Pr(evidence | guilt)Pr(guilt)
=
Pr(evidence | guilt)Pr(guilt) + Pr(evidence | no guilt)Pr(not guilt)
(Note: Pr(evidence|guilt) and Pr(evidence|no guilt) are relative values)

15

@seismicisolation
@seismicisolation
Exercise 2.21
Everyday life 3. Three door problem. This problem was a common topic of
discussion some years ago. The host in a television show has a set with three
doors. Behind one is a prize, say a certificate entitling the contestant to a brand
new Maserati. The contestant may choose one door at random, giving a one-
in-three chance of winning the prize. He selects a door. At this point the host
intervenes and offers to open one of the two remaining doors. She does so, and
offers the contestant the option of changing his selection to the remaining closed
door. Should he do so? It may be assumed that the host will not open a door
that contains the prize.

Solution
Three door problem. Yes the contestant should change to the remaining
closed door.
Initially, three doors are considered. Let:
D1 = "gift is behind Door 1"
D2 = "gift is behind Door 2"
D3 = "gift is behind Door 3"
1
Pr(D1 ) = Pr(D2 ) = Pr(D3 ) =
3
Assume that the contestant chooses Door 1. Then new information I is
obtained: Door 2 opened by host—no prize. Therefore

Pr (I | G1 ) = 0.5
Pr (I | G2 ) = 0, and
Pr (I | G3 ) = 1.

By Bayes’ Theorem:

Pr(I | G1 )Pr(G1 )
Pr (G1 | I) =
Pr(I | G1 )Pr(G1 ) + Pr(I | G2 )Pr(G2 ) + Pr(I | G3 )Pr(G3 )

Then, substituting,

¡ ¢
(0.5) 13 1
Pr (G1 | I) = ¡1¢ ¡ ¢ ¡ ¢=
(0.5) 3 + (0) 13 + (1) 13 3
Pr (G2 | I) = 0
2
Pr (G3 | I) =
3
A case for changing doors is made!

16

@seismicisolation
@seismicisolation
Exercise 2.22
Warning light 1. Warning lights and devices are used in many applications,
for example fire alarms, lights for oil level in an automobile and devices for
methane gas detection in coal mines. In automobile dashboards, warning lights
are used rather than the more expensive gauges. Consider for example, a device
to indicate event T = “oil level below a specified level”. A warning light comes
on–event L–to indicate ³ that´ T might have occurred. Tests can be conducted to
obtain Pr(L|T ) and Pr L̃|T̃ . Let these values be denoted p and q respectively,
³ ´ ³ ´ ³ ´
and let Pr(T ) = t. Find Pr(T |L) , Pr T̃ |L , Pr T |L̃ and Pr T̃ | L̃ in terms
of p, q and t.

Solution

Pr(L | T )Pr(T )
Pr (T | L) =
Pr(L | T )Pr(T ) + Pr(L | T̃ )Pr(T̃ )
pt pt
= = .
pt + (1 − q)(1 − t) pt + q̃ t̃
Similarly:

³ ´ q̃ t̃
Pr T̃ | L̃ = ,
pt + q̃ t̃
³ ´ q t̃
Pr T̃ | L̃ = , and
q t̃ + p̃t
³ ´ p̃t̃
Pr T̃ | L̃ = .
q t̃ + p̃t

Exercise 2.23
Warning light 2. The devices in problem 2.22 are designed to indicate the
occurrence of a rare event. This can pose difficulties in the design of the system.
Consider the following problem. A warning device on an aircraft includes a light
that is “on” when the landing gear is correctly positioned. It accurately reflects
the landing gear correctly in position 99.9% of the time. The remaining 0.1%
corresponds to failure of the lighting system in the device, incorrectly indicating
failure of the landing gear to position itself. Given that failure of the landing gear
occurs on average once in 106 landings, what is the probability that the landing
gear is incorrectly positioned if the light indicates “failure”? Assume that the
warning device works perfectly when the landing gear has failed. Comment.

17

@seismicisolation
@seismicisolation
 This problem might be analyzed with a real-world application in mind; see
Aviation Weekly and Space Technology, July 30 and August 6, 1973. A Lockheed
L-1011 aircraft crashed with many fatalities after there was doubt whether the
front landing gear was in the down position. The bulbs in the system indicating
the position of the landing gear were found after the crash to be burned out, and
the probable cause of the accident was preoccupation of the crew regarding the
landing gear position, resulting in distraction from the instruments and altitude
of the aircraft.

Solution
Let A = "Landing gear correctly positioned,"
à = "Landing gear incorrectly positioned,"
L = "Light on," and
L̃ = "Light off."
L
0.999

A
-6

1-10 0.001

A L
10
-6
0.0

1.0

³ ´ Pr(L̃ | Ã)Pr(Ã)
Pr à | L̃ =
Pr(L̃ | Ã) Pr(Ã) + Pr(L̃ | A)Pr(A)
¡ ¢
(1.00) 10−6
=
(1.00) (10−6 ) + (0.001) (1 − 10−6 )
1 ∼
= = 10−3
1001.001
Not a good system. When the indicator light is "off," there remains approx-
imately 0.999 probability that the landing gear is correctly positioned.

18

@seismicisolation
@seismicisolation
 The problem hinges on the relative values 10−6 and 10−3 ; the warning system
is likely to fail 1000 times more often than the landing gear.

Exercise 2.24
An electronic company “A” manufactures microprocessors. An important com-
petitor, “B”, has introduced a new product which is selling well. Company
A has to decide whether to introduce a competing microprocessor. There are
concerns regarding the response of company B, which could expand production,
keep it at the same level, or reduce. From past experience, the chances of the
three actions are respectively 60%, 30%, and 10%. One source of information to
company A is to study B’s advertising campaign. >From past experience, they
have assigned probabilities to the rival reducing, maintaining, and increasing
their advertising given that it is their intention to cut, keep at some level, or
expand production as in the table below.
Advertising is
Decreased Maintained Increased
Company B Cut back 0.60 0.30 0.10
is going to Keep same level 0.30 0.50 0.20
Expand 0.10 0.30 0.60
It is then observed that advertising is in fact increased. What are the revised
probabilities for the rival’s course of action?

Solution
Let E = "B expands production,"
M = "B maintains present level,"
R = "B reduces production," and
I = "B’s advertising is increased."

19

@seismicisolation
@seismicisolation
 I
0.6

E
I
M 0.2

R
I
0.1

(0.6) (0.6)
Pr (E | I) =
(0.6) (0.6) + (0.3) (0.2) + (0.1) (0.1)
36 36
= = ,
36 + 6 + 1 43
6
Pr (M | I) = , and
43
1
Pr (R | I) =
43

Exercise 2.25
In an industrial plant, a flow meter is used to measure the direction of flow in a
pipe, with an electrical signal being sent back to the control room. Under ideal
conditions, a positive signal should arrive when the flow is in one direction–
positive flow–with a negative signal level when the flow is in the opposite, or
negative, direction. In reality, a positive signal level can be observed even if the
flow is negative and vice versa, as a result of turbulence and system noise. The
possible received signal levels X are −2, −1, 0, +1, +2. The probabilities of the
value of X conditional on the direction of the flow are as follows.
x Pr (X = x| positive flow) Pr (X = x| negative flow)
+2 0.40 0.00
+1 0.30 0.15
0 0.20 0.15
−1 0.05 0.30
−2 0.05 0.40

20

@seismicisolation
@seismicisolation
 It is also known that positive flow is 1.5 times as likely as negative flow.
The operator in the control room must make a decision based on the received
signal level. She assumes that the more likely flow direction indicates the
actual flow direction. In other words, upon receiving signal level X = x, if
Pr (positive flow | signal level X = x) ≥
Pr (negative flow | signal level X = x), the operator takes the flow direction
to be positive. Otherwise the flow direction is taken to be negative.

(a) Find the signal levels which cause the operator to assume that the flow is
positive.
(b) What is the probability that the operator’s assumption of flow direction
is incorrect?

Solution
x= (3/5)(4/10)=12/50=6/25
+2

0.4 +1 (3/5)(3/10)=9/50
0.3
0 (3/5)(2/10)=6/50=3/25
0.2
w
Flo
ve
siti 3/5
0.05 -1 (3/5)(5/100)=3/100
Po 0.05
(3/5)(5/100)=3/100
-2
Ne
gat
ive +2 0
Flo (2/5)(15/100)=6/100=3/50
w 0 +1
0.15
2/5
0 (2/5)(15/100)=6/100=3/50
0.15

0.3 -1 (2/5)(3/10)=6/50=3/25
0.4
-2 (2/5)(4/10)=8/50=4/25

(a) Let E = positive flow

x Pr(E | X = x) Pr(Ẽ | X = x) Assumption

+2 1.0 0.0 Positive
+1 3/4 1/4 Positive
0 2/3 1/3 Positive
−1 1/5 4/5 Negative
−2 3/19 16/19 Negative

(b) From tree the probability is

3 3 3 3 9
= + + + = .
100 100 50 50 50

21

@seismicisolation
@seismicisolation
Exercise 2.26
In the construction of a gas pipeline, the number of flaws due to welding at joints
in the metal is a question of major importance since the flaws are associated with
possible subsequent failure of the pipeline. The engineer in charge first studies
previous pipelines of similar construction to estimate the number of major flaws,
S, per section. On this basis, the engineer decides that S = 0, 1, or 2 major
flaws are possible, with probabilities of 0.92, 0.05, and 0.03 respectively. An
expensive x-ray testing procedure can be used to detect the presence and depth
of the flaws; the results are embodied in the values I = 0, 1 or 2, representing the
number of major flaws indicated by the x-ray machine. Non-destructive testing
methods of this kind are not perfect, and previous use of similar equipment
yields the following conditional probabilities Pr(I = i|S = s) .
I
S 0 1 2
0 0.80 0.15 0.05
1 0.10 0.70 0.20
2 0.10 0.10 0.80
Calculate the matrix of values Pr(S = s|I = i).

Solution
I=1
2

3
S=1
I=1
S=2 2

3
S=3 I=1
2

3
Typically

(0.8) (0.92)
Pr(S = 0 | I = 0) = = 0.989
(0.8) (0.92) + (0.1) (0.05) + (0.1) (0.03)

Values of Pr(S = s | I = i) are given in the following table.

22

@seismicisolation
@seismicisolation
 I
0 1 2
S 0 0.989 0.784 0.575
1 0.007 0.199 0.125
2 0.004 0.017 0.300

Exercise 2.27
A message is sent from a transmitter to a receiver via three relays. Bits (0 or 1)
are transmitted. The relays may commit a reversal error in which case a “1” is
sent when a “0” is received and vice versa. The probability of a reversal error
in each relay is 0.03. Stochastic independence from one relay to another can be
assumed.
(a) Calculate the probability that the received bit is identical to the transmit-
ted bit.
(b) From the past, it is known that 60% of all bits transmitted are “1”. Find
the probability that the received bit is a “1” and the probability that a
“1” was transmitted given that a “1” has been received.

Solution
(a) This will happen if and only if the number of reversals is 0 or 2. For X
reversals:

Pr(X = 0) = 0.973 = 0.912673

2
Pr(X = 2) = 3 (0.03) (0.97) = 0.002619
since there are 3 ways, and

Pr(received bit is the same as the one transmitted)

= 0.912673 + 0.002619 = 0.915292 (1)
say 92%
(b) Let R1 = "received bit is a 1", T1 = "transmitted bit is a 1", and T0 =
"transmitted bit is a 0". Then

Pr(R1 ) = Pr(R1 | T1 )Pr(T1 ) + Pr(R1 | T0 )Pr(T0 )

= (0.915)(0.60) + (1 − 0.915)(0.40) = 0.583
and

Pr(R1 | T1 )Pr(T1 ) (0.915)(0.60)

Pr(T1 | R1 ) = = = 0.942
Pr(R1 ) 0.583

23

@seismicisolation
@seismicisolation
Exercise 2.28
In the two systems below, the components have a reliability of 98%. This means
that the probability of failure is 0.02. Find the system reliability in each case.
Stochastic independence should be assumed. The systems are operational if
there is a working path from A to B.

Solution
Series System
Both components S1 and S2 must be operational.
Required probability = 0.98 × 0.98 = 0.9604
Parallel System
Either S1 or S2 or both have to be operational.
Required probability = 0.98 + 0.98 − (0.98)2 = 0.9996
[Probability of failure = (0.02)2 = 0.0004 = 1 − 0.9996]

Exercise 2.29
Describe an example of dependence that is conditional on our state of knowledge.

Solution
(Bayes networks are a very elegant and contemporary method of idealizing con-
ditional probability.)
An offshore rig has failed. Two main causes are considered possible: earth-
quake or hurricane. These are generally considered as being stochastically inde-
pendent. Evidence from a nearby seismic station indicates that an earthquake
was very unlikely. The probability of hurricane increases as a result of this
information. An example of stochastic dependence.

Exercise 2.30
A batch of 10 cables contains 3 defective cables. A random sample of 4 cables
is taken from this batch and tested. Find

(a) the number of ways there can be exactly two defective cables in the sample,
and
(b) the number of ways there can be at least two defective cables in the sample.

24

@seismicisolation
@seismicisolation
Solution

7 3
Good Def.

(a) There must be two defective cables and two good cables from the four.
The number of ways is
3! 7!
3 C2 × 7 C2 = × = 3 × 21 = 63.
2!1! 2!5!

(b) We have to work out the probability of 2 or 3 defectives. For 3 defectives,

the number of ways is

3 C3 × 7 C1 = 1 × 7 = 7.

For 2 or 3 defectives, using (a), the number of ways = 63 + 7 = 70.

Exercise 2.31
A water diviner claims that he can locate hidden sources of water using a carved
whalebone. A test is conducted by asking him to separate ten covered cans
into two groups of five, one consisting of cans containing water. It is arranged
beforehand that five cans contain water. What is the probability that the diviner
correctly puts at least three cans into the “water group”, entirely by chance?

25

@seismicisolation
@seismicisolation
Solution
Required probability is

5 C3 ·5 C2 5 C4 ·5 C1 5 C5 ·5 C0
+ +
10 C5 10 C5 10 C5
5! 5! 5! 5!
3!2! · 2!3! + 4!1! · 1!4! +1
= 10!
5!5!
100 + 25 + 1 1
= = .
252 2

Exercise 2.32
At the “Group of Eight” meetings in 2002, what is the probability that George
Bush and Tony Blair stand next to each other at the photo-op, purely by chance?
How many ways can the eight persons be seated around a circular table?

Solution
(a) Number of different line-ups = 8! = 40, 320

Now treat Bush and Blair as a single entity, tied together. There are now
7! = 5, 040 ways.
2 × 5040 1
Required probability = = (the factor 2 appears because we can
40320 4
have two permutations Bush-Blair and Blair-Bush).

(b) (n − 1)! = 7! = 5040

Exercise 2.33
An urn contains seven red balls and three green balls. A second urn contains
four red balls and six green balls. A ball is chosen randomly from the first urn
and placed in the second urn. Then a ball is chosen randomly from the second
urn and placed in the first urn.

(a) What is the probability that a red ball is selected both from the first urn
and then from the second urn?
(b) What is the probability that a red ball is selected from the second urn?
(c) At the conclusion of the selection process, what is the probability that
the numbers of red and green balls in each urn are identical to the initial
numbers?
Now consider the generalization to the case where there are n(> 1) balls
initially in each urn, with a (such that 0 < a < n) red balls in urn 1 and
b (with 0 < b < n) red balls in urn 2 initially.

26

@seismicisolation
@seismicisolation
 (d) Show that the probability that a red ball is selected from the second urn
during this process is, in general, different from the probability of drawing
a red ball directly from the second urn, without transferring a ball from
the first urn.
(e) Find a necessary and sufficient condition on the values of a and b for the
two unconditional probabilities in part (d) above to be equal.

Solution
Let R1 = "red ball selected from the first urn", R2 = "red ball selected from the
second urn", G1 = "green ball selected from the first urn" (= R̃1 ), and G2 =
"green ball selected from the second urn" (= R̃2 ).

(a) Then
7 3
Pr[R1 ] = , Pr[G1 ] = ,
10 10
5
Pr[R2 | R1 ] =
11
4
Pr[R2 | G1 ] =
11
and
µ ¶ µ ¶
5 7 35 7
Pr[R1 R2 ] = Pr[R2 | R1 ]×Pr[R1 ] = × = = = 0.3182.
11 10 110 22

(b)

Pr[R2 ] = Pr[R1 R2 ] + Pr[R̃1 R2 ] (total probability law)

= Pr[R2 | R1 ] × Pr[R1 ] + Pr[R2 | R̃1 ] × Pr[R̃1 ]
µ ¶ µ ¶ µ ¶ µ ¶ µ ¶ µ ¶
7 4 3 7 6 47
= + × = + = = 0.4273
22 11 10 22 55 110

(c) The desired event E = (numbers before and after are identical) can be
expressed as:
E = R1 R2 + G1 G2
But
¶ µ ¶ µ
5 7 7
Pr[R1 R2 ] = Pr[R2 | R1 ] × Pr[R1 ] = × = , and
11 10 22
µ ¶ µ ¶
7 3 21
Pr[G1 G2 ] = Pr[G2 | G1 ] × Pr[G1 ] = × = ,
11 10 110
and therefore
µ ¶ µ ¶
7 21 28
Pr[E] = + = = 0.5091
22 110 55

27

@seismicisolation
@seismicisolation
 A tree diagram may assist in the solution to this question:

R1 R 2
R2
5/11
R1 G2

7/10 6/11
R1 G 2
(7/22)+(21/110)
G1 R 2 = 28/55
G1 R2

3/10 4/11
G2
7/11
G1 G 2

(d) Urn 1 Urn 2

RED a b
NOT RED n−a n−b
Total n n
Let A = red ball drawn from urn 1, and B = red ball drawn from urn 2;
then
a
Pr[A] =
n
(n − a)
Pr[Ã] =
n
(b + 1) b
Pr[B | A] = ; Pr[B | Ã] =
(n + 1) (n + 1)
a (b + 1) (n − a) b
Pr[AB] = ; Pr[ÃB] = , and
[n (n + 1)] [n (n + 1)]
(ab + a + nb − ab) (a + nb)
Pr[B] = =
[n (n + 1)] [n (n + 1)]

Let C = red ball drawn from original contents of urn 2 (that is, when no
balls have been added from urn 1), then:
µ ¶
b
Pr[C] = 6= Pr(B)
n

in general.
(e) Since
Pr[B] = Pr[C],

28

@seismicisolation
@seismicisolation
 bn(n + 1) = n(a + nb), and
bn + b = a + nb.

Therefore b = a.

Exercise 2.34
Find a proof of Equation (2.49) based on the expansion of Newton’s formula for
(p + q)m+n .

Solution
¡m+n¢
The coefficient of pt q m+n−t is t , by Equation (2.46). We can also write

(p + q)m+n = (p + q)m (p + q)n

"m µ ¶ #" n µ ¶ #
X m X n
r m−r s n−s
= p q p q
r=0
r s=0
s
Xm X n µ ¶µ ¶
m n r+s m+n−r−s
= p q .
r=0 s=0
r s

The coefficient of pt q m+n−t can be obtained from this expression by considering

all terms for which (r + s) is equal to the designated value of t. The result
embodied in equation (2.49) is again found.]

Exercise 2.35
Two age-old problems.

(a) Find the number of distinct ways of arranging the letters in the word
“STATISTICS”.
(b) A hand of five cards is dealt from a standard deck of 52 cards. What are
the odds against the event that all five cards are of the same suit?

Solution
(a) "STATISTICS" contains 1A, 1C, 2I’s, 3S’s and 3T’s. The number of
distinct arrangements is
10!
= 50, 400.
1!1!2!3!3!

29

@seismicisolation
@seismicisolation
 (b) Let the event be ”E”. For 1 suit,

13 C5
Pr(E) =
52 C4

For 4 suits,
4 ·13 C5 33
Pr(E) = =
C
52 5 16660
Odds against E are
16660 − 33
= 16627 : 33 ∼
= 504 : 1.
33

Exercise 2.36
At a club for men on a wet and windy night all are wearing hats which they check
upon arrival. The person checking the hats becomes confused and returns them
at random. What is the probability that no man receives his hat? Consider n
men and the change in value of probability as n increases.

Solution
The hat problem also known as the matching problem. There are n! permuta-
tions of 1, 2, ..., n. If the original order is 1, 2, ..., n, then the permutations include
cases where the original and permuted numbers match. For example, if n = 7 :
1 2 3 4 5 6 7 Original
7 1 3 6 5 2 4 Permuted
In this case 3 and 5 match. These are termed "fixed points." A complete per-
mutation is one that has no fixed points. This corresponds to the case where
no man gets his hat back. We assume all permutations equally likely. We first
work out the probability that at least one fixed point (match) occurs. We start
with equation (2.17) which deals with general events Ei . The event E that at
least one Ei occurs is:

E = E1 ∨ E2 ∨ E3 ∨ · · · ∨ En = 1 − Πni=1 (1 − Ei ).
Upon expanding the right hand side, we obtain terms of the following kind:
X
Ei − Ei Ej + Ei Ej Ek − · · · ;
the number of terms in each numeration is
µ ¶ µ ¶ µ ¶
n n n
, ,...,
1 2 n
(see for example the binomial expansion (2.46)).
Let us now number the men 1, 2, ..., n, and their hats similarly. The prob-
ability of each permutation of hats is 1/n!. Let there be a fixed point (match)

30

@seismicisolation
@seismicisolation
in the ith place. There are (n − 1)! permutations of the remaining hats so that
the probability of the fixed point in the ith place is

(n − 1)! 1
=
n! n
Similarly for each combination of fixed points in places i and j (i 6= j), the
probability is

(n − 2)! 1
= ,
n! n (n − 1)
and so on. ¡ ¢ ¡ ¢
The numbers of combinations was shown above to be n1 , n2 , ..., so that the
final probability of at least one fixed point is
µ ¶ µ ¶ µ ¶
1 n 1 n 1 n
· − + − ···
n 1 n (n − 1) 2 n (n − 1) (n − 2) 3
This is equal to:
1 1 1
1− + − ··· ± ,...
2! 3! n!
for n = 1, 2, . . . . The last term is positive if n is odd. Now the probability of
no hats matching is the complement of the expression just deduced,
1 1 1 1
− + − .... ±
2! 3! 4! n!
Here n = 2, 3, ... (n = 1 could be included if a 0 is added at the beginning). The
last term is positive if n is even.
The following is found by computation:
n Probability that no
Number of Men man gets his hat
2 0.5
3 0.333
4 0.375
5 0.366
6 0.368
The series tends to e−1 = 0.367879 · · · . See Feller (1968) and Blom et al.
(1994).

Exercise 2.37
General model of drawing from urn. In an urn with balls of two colours, g1 are
of colour 1, and g2 of colour 2. A ball is drawn, and then replaced. Then h1 of
the colour drawn and h2 of the other colour are added to the urn and mixed.
A ball is now drawn from the urn, which contains g1 + g2 + h1 + h2 balls. The
values h1 and h2 are arbitrary and could be negative. See Feller (1968). We

31

@seismicisolation
@seismicisolation
consider now the case h2 = 0 and h1 > 0, which constitutes a general Pólya
urn. Show that the probability of drawing a ball of colour 2 at any drawing is
g2 / (g1 + g2 ) .

Solution
(by induction)
The statement is true for the first drawing. Assume that it is true for the
nth . Then for the (n + 1)st the probability is

g1 g2 g2 g2 + h1
· + ·
(g1 + g2 ) g1 + g2 + h1 (g1 + g2 ) g1 + g2 + h1
g1 g2 + g22 + g2 h1 g2 (g1 + g2 + h1 )
= =
(g1 + g2 ) (g1 + g2 + h1 ) (g1 + g2 ) (g1 + g2 + h1 )
g2
= .
g1 + g2

Exercise 2.38
Eight passengers share a compartment in a train which stops at eleven stations.
What is the probability that no two passengers leave at the same station?

Solution
Each passenger has 11 choices or 118 in all. For no two passengers to leave at
the same station, there are 11 choices for the first passenger, 10 for the second
passenger,. . . , and so on. The required probability is

(11)7
;
118
(11) (10) (9) (8) (7) (6) (5) (4) 6, 652, 800
= =
118 214, 358, 881
= 0.03103 · · · .

See Equation (2.53) et seq.

32

@seismicisolation
@seismicisolation
Solutions for Chapter 3
Exercise 3.1
A random quantity X is modelled using the probability mass function

pX (x) = k (4 − x) , for x = 0, 1, 2, 3
= 0, otherwise.

(a) Find the constant k and the pmf. Sketch the distribution.
(b) Find and sketch the cdf.
(c) Find Pr(X < 2) .

Solution
(a) x 0 1 2 3
px (x) 4k 3k 2k k
X
pX (x) = 10k = 1, for coherence;
1
=⇒ k =
10

0.4

p(x)
X

0 1 2 3 x
(b) x 0 1 2 3
Fx (x) 0.4 0.7 0.9 1.0

@seismicisolation
@seismicisolation
 F (x)
X

x
(c) 0.7

Exercise 3.2
For the following probability densities, determine the constant k, sketch the
density, and find and sketch the cdf.
¡ ¢
fX (x) = k 1 − x2 for 0 < x < 1, = 0, elsewhere.
fX (x) = exp (−k |x|) for all real x.
π
fX (x) = k sin (kx) for 0 < x < , = 0, elsewhere.
4
For the first pdf, find
¡ the ¢Pr(X > 0.5) , for the second find Pr(X > 1) and for
the third, find Pr X > π6 .

Solution
First
fX (x) = k(1 − x2 ), for 0 < x < 1

Coherence:
Z ∞ Z 1
fx (x)dx = k(1 − x2 )dx = 1
−∞ 0

2k 3
= 1, =⇒ k =
3 2
3¡ ¢
fX (x) = 1 − x2
2
cdf

@seismicisolation
@seismicisolation
 Z x
¡ ¢
3¡ 2
¢ x 3 − x2
FX (x) = = 1 − t dt = , 1 > x > 0;
0 2 2
= 1, for x > 1;
= 0, for x ≤ 0.

f (x)
X

F (x)
X

Pr(X > 0.5) = 1 − FX (0.5) = 0.3125

Second
fx (x) = e−k |x|

Coherence:
Z ∞ Z ∞
−k |x|
e dx = 1 = 2 e−k |x| dx, by symmetry.
−∞ −∞

1
2· = 1, so that k = 2.
k
cdf

@seismicisolation
@seismicisolation
 For x < 0, Z ∞
1 2x
FX (x) = e−2|x| dx = e ;
−∞ 2
1
Fx (0) = ,
2
by symmetry.
For x > 0, Z x
1
FX (x) = FX (0) + e−2t dt = 1 − e−2x
0 2
Hence ½ 1 2x
FX (x) = 2 e , for x ≤ 0
1 − 12 e−2x , for x > 0

f (x)
X

1.0

0 x

F (x)
X

1.0

0.5

0 x .

1 −2
Pr(X > 1) = 1−Pr(X ≤ 1) = 1−FX (1) = e = 0.0677, to 4 decimal places.
2
Third
π
fX (x) = k sin(kx), for 0 < x <
4

@seismicisolation
@seismicisolation
 Coherence:
Z π
4
k sin(kx)dx = 1;
0
π
[− cos(kx)]04 = 1
kπ kπ
1 − cos( ) = 1 =⇒ cos( ) = 0
4 4

kπ π
= + 2nπ, where n is an integer;
4 2
k = 2 + 4n Equation A
π
For the pdf to be non-negative inside the interval 0 < x < 4, (kx) should be
in the first quadrant, so that
kπ
0≤ ≤π
4
kπ
If k were negative, −π ≤ ≤ 0. Hence
4
|k| ≤ 4 Equation B

Combining Equations A and B:

k = ±2;
choose +2 (solutions are identical). Then
π
fX (x) = 2 sin(2x), for 0 < x <
4
cdf
Z x
FX (x) = 2 sin(2u)du = 1 − cos(2x)
0

 0; for x ≤ 0
π
FX (x) = 1 − cos(2x); for 0 < x < 4

1; for x ≥ π4

@seismicisolation
@seismicisolation
 2.0

f (x)
X

0
x
4

1.0

F (x)
X

0
x 4

π π π 1
Pr(x > ) = 1 − Pr(x ≤ ) = cos =
6 6 3 2

Exercise 3.3
A random quantity is modelled with a cdf

 0, for t < 1
FT (t) = t2 − 2t − 1, for 1 ≤ t ≤ 2

1, for t > 2

Determine the pdf fT (t) , and Pr(T ≤ 1.5) .

Solution
dFT (t)
fT (t) = ;
dt

 0, for t < 1
fT (t) = 2(t − 1), for 1 ≤ t ≤ 2

0, for t > 2

@seismicisolation
@seismicisolation
 Pr(T ≤ 1.5) = FT (1.5) = 0.25

Exercise 3.4
In planning of the use of a route for emergency vehicles in a city, there are ten
signalized intersections. At each one, the traffic light is green for 40 seconds,
yellow (amber) for 5 and red for 30 seconds. The emergency vehicle will stop at
an intersection only if it is red. The signalling system is not coordinated. What
probability distribution models the number of stops? Determine the probability
that the emergency vehicle will have to stop at

(a) the first intersection;

(b) none of the intersections;
(c) the first four intersections;
(d) the second, fourth,. . . , (even numbered) intersections but not at the oth-
ers;
(e) more than five intersections?
(f) It is desired to reduce the probability in (e) to 0.02 seconds. What should
the “red” time be so that this is obtained, keeping the amber and total
cycle times constant?

Solution
Let E = "vehicle stops"
30 30 2
Pr(E) = = = = 0.4
30 + 5 + 40 75 5
Binomial Distribution, where X = number of stops.
µ ¶
10
b(x, 10, 0.4) = (0.4)x (0.6)10−x
x

(a) 0.40 (Bernoulli Case)

10
(b) (0.6) = 0.0060466 · · · ∼
= 0.006
4
(c) (0.4) = 0.0256

(d) (0.4)5 (0.64 ) = 0.000796... ∼

= 0.0008

@seismicisolation
@seismicisolation
 (e)
X10 µ ¶
10
Pr(x > 5) = (0.4)x (0.6)10−x = 1 − FX (5) = 0.1663
x=6
x

(f)

X 5 µ ¶
10 x 10−x
1− p q = 0.02
x=0
x
X5 µ ¶
10 x
p (1 − p)10−x = 0.98
x=0
x
p = 0.233

"Red" time = (0.233)(75) = 17.5 seconds

Exercise 3.5
A lot of 1,000 transistors contains 900 that pass an acceptance test. The tran-
sistors are used in groups of five to construct a component. It is decided to
mix the failed transistors with those that have passed the test and to select five
randomly from the lot. The number of failed transistors in the group of five is
of interest, denoted X.

(a) What is the distribution of X? Can the binomial distribution be used as

an approximation?
(b) Find the probability that not all the transistors have passed the test, using
an exact method and a binomial approximation.
(c) What is the probability that more than two transistors in the component
have failed the test?
(d) What are the mean and the standard deviation? Use the approximate and
exact distributions.

Solution
(a) This is a series of trials with a complementary pair of outcomes. The
probability of success (test failed) is constant (see for example Exercise
1
2.37), and is equal to 10 , without knowledge of the results of trials. Given
the results of previous trials, the probability in the next trial is depen-
100 99
dent on these results. (Example: probability changes from 1000 to 999 if a
"failed" transistor is drawn on the first trial.) The distribution is hyper-
geometric, but the binomial distribution provides a good approximation,
with p = 0.1.

@seismicisolation
@seismicisolation
 (b) Binomial:

Pr(not all passed) = Pr(X > 0) = 1−Pr(X = 0) = 1−(0.9)5 = 0.4095 . . .

Exact:
900 C5 13517464688
Pr(X > 0) = 1 − Pr(X = 0) = 1 − =1−
1000 C5 22917475695
9400011007
= = 0.4102 . . .
22917475695

(c) Required probability (approx. method)

· µ ¶ ¸
5 4 5 2 3
1 − FX (2) = 1 − (0.9) + (5) (0.1) (0.9) + (0.1) (0.9)
3
= 1 − (0.59049 + 0.32805 + 0.0729) = 1 − 0.99144 = 0.00856

(Less than 1 in 100 chance of two or more failed transistros. If assembly

is robost, it might be viable to include these.
(d) Mean + SD
1 1
hXi = np = 5 × = = 0.5
10 2
Approx. SD
1 9 9
σ X = npq = 5 × × = = 0.45
10 10 20
Exact SD
(N − n) 1 9 990 33
σ X = npq =5× × × = = 0.4459 . . .
(N − 1) 10 10 999 74

Exercise 3.6
Summarize the four forms of the hypergeometric distribution as described and
emphasized by de Finetti.

Solution
The first,
¡g ¢¡ b−g ¢
x
Pr(X = x|b, n, g) = ¡ bn−x
¢ ,
n
represents the number of ways, or combinations, of choosing x red balls from g
and (n − x) pink balls from (b − g), divided by the number of ways of choosing
n from b. All of these choices are without regard to order: see Equation (2.44).

@seismicisolation
@seismicisolation
 The second, ¡n¢¡b−n¢
x g−x
Pr(X = x|b, n, g) = ¡b¢ ,
g

enumerates the number of paths to node (n, x) times the number of paths from
this node to the endpoint (b, g) divided by the number of paths from the start
to the endpoint (see Lotto Example 3.6).
The third form
· ¸
(b − g) (b − g − 1) (b − g − 2)
µ ¶ [g (g − 1) (g − 2) · · · (g − x + 1)]
n · · · (b − g − n + x + 1)
x b (b − 1) (b − 2) · · · (b − n + 1)
¡ ¢
gives the number of ways, nx , times the probability of drawing the red balls
first, followed by the pinks,
¡ ¢ taking into account the change in composition of
the urn. This represents nx times the probability of a single path
The fourth form is
 ³ ´³ ´ 
h³ ´³ ´ ³ ´i 1 − b−g 1 2
1 − b−g
1 − g1 1 − g2 · · · 1 − x−1  ³ ´ 
µ ¶ g
· · · 1 − n−x−1
n x n−x b−g
q q ¡ ¢¡ ¢ ¡ ¢
x 1 2 1 2
1 − b 1 − b ··· 1 − b n−1

where q1 = g/b, q2 = q̃1 = (b − g)/b. It gives the relation to the binomial

distribution; for large g and b with relatively small x and n. Then the ratio on
the right hand side tends to unity, and q1 and q2 can be interpreted as p and q
in the binomial distribution of Equation (3.3).

Exercise 3.7
In a telephone system, long distance telephone calls from a certain origin arrive
at a rate of 12 per five minutes. This rate can be considered as constant for the
times under consideration.

(a) What models should be used to describe the time T between successive
calls and for the number of calls X in a minute?
(b) If a time τ elapses after a call, during which time no call is received, what
is the distribution to the next call?
(c) What is the probability that two calls arrive less than 30 seconds apart?
What is the probability that the next call will arrive after a minute elapses?
(d) Find the probability that two successive calls are separated by an inter-
val that is within two standard deviations of the mean? Sketch the pdf
showing the calculation.

10

@seismicisolation
@seismicisolation
Solution
12
(a) T : use exponential distribution with λ = per minute.The pdf is
5
12 −12 t
fT (t) = e 5 , for t ≥ 0,
5
and t is in minutes.
12
(a) X : use Poission with ν = 5 · 1 = 2.4. The pmf is

e−2.4 (2.4)x
pX (x) = f orx = 0, 1, 2, . . .
x!

(b) Let t0 = t − τ ; and let the required distribution be fT 00 (t0 ). Then

Pr(t < T < t + dt) λe−λt

Pr(t < T < t + dt | t > τ ) = = −λt = λe−λ(t−τ ) , or
Pr(t > τ ) e
0
fT 0 (t0 ) = λe−λt .

This is the same as the original distribution but translated horizontally

by an amount τ .
(c) µ ¶
1 1
Pr(T < 30 sec .) = Pr(T < min .) = FT
2 2
Now

FT (t) = 1 − e−λt , and

µ ¶
1
FT = 0.699 · · ·
2
to 3 decimal places. Also
−12
Pr(T > 1) = 1 − FT (1) = e 5 = 0.0907 · · ·

to 3 significant figures.
(d) We require Pr(µ − 2σ < T < µ + 2σ). For exponential distribution
1 1
λe−λt , µ = ,σ = .
λ λ
Required probability =
−λ · 3
3
Pr(µ−2σ < T < µ+2σ) = Pr(T < ) = 1−e λ = 1−e−3 = 0.950 · · ·
λ
to 3 significant figures (note: this is true for any λ).

11

@seismicisolation
@seismicisolation
Exercise 3.8
The number of flaws X per square metre of the surface of a plate is a random
quantity. It is modelled using a Poisson distribution with parameter λ = 1.67
per square metre.
(a) What are the mean and standard deviation of X?
(b) What is the probability that a square metre of plate has no flaws?
If a square metre of plate has no flaws on its surface, it fails at a pressure
of 5.6 kPa. If there are one or more flaws, it fails at 2.8 kPa.
(c) Find the mean pressure at which a square metre of plate fails.
(d) Show that the standard deviation in kPa of the pressure for the square
metre of plate is
p
2.8 exp [(−1.67)] exp (1.67 − 1).

Solution
(a) Mean = λ = 1.67 kP a; Variance = σ 2 = λ = 1.67; σ = 1.292 kP a (to 3
decimal places)
(b)
λx e−λ
Pr(X = x) = , for x = 0, 1, 2, . . .
x!
Pr(X = 0) = e−λ = 0.188
(to 3 decimal places)
(c)
hY i = (5.6)(0.188) + (2.8)(0.812) = 3.327kP a
(to 3 decimal places). (where Y = pressure). Generally,
³y ´ ¡ ¢ ³ y1 ´ ¡ λ ¢
1
hY i = y1 e−λ + 1 − e−λ = e +1
2 2
(d)
­ 2® ³ y ´2 ¡ ¢ ³ y1 ´2 ¡ −2λ ¢
2 1
σ2Y = Y − hY i = y12 e−λ + 1 − e−λ − e + 2e−λ + 1
µ ¶ 2 2
1 1
= y12 e−λ − e−2λ , or
4 4
³ y ´2 ¡ ¢
1
σ2Y = e−2λ eλ − 1 , and
2 q
³y ´
1
σY = e−λ (eλ − 1)
2

σ 2Y = (3.327 − 2.8)2 (0.812) + (5.6 − 3.327)2 (0.188) = 1.197;

σY = 1.095kP a,

12

@seismicisolation
@seismicisolation
 (to 3 decimal places).

Exercise 3.9
Given that an earthquake has occurred, the magnitude R on a Richter scale is
modelled as the exponential distribution

fR (r) = 2.35 exp (−2.35r) for r > 0,

= 0, otherwise.

Find the cdf and the probability that the next earthquake will have a magnitude
greater than 6.0.

Solution
Z r
FR (r) = 2.35e−2.35t dt = 1 − e−2.35r , for r > 0
0
= 0, for r ≤ 0.

Pr(R > 6) = 1 − FR (6.0) = e−2.35r = e−14.1 = 7.52 × 10−7

(odds of less than one in a million).

Exercise 3.10
The pressure X acting on a ship’s hull during a collision is modelled by an
exponential distribution with a mean of 2 MPa. The number of impacts per
year is modelled by a Poisson distribution with a mean equal to 10. Consider
now only those impacts with X > x MPa.

(a) What model would you use for the number of these impacts (X > x) per
year? Find the probability that in a year the number of such impacts is
equal to zero for the case x = 6 MPa.
(b) Consider now the maximum pressure in a year, written as Z. Using the
result just obtained, derive the cdf FZ (z) .
(c) If the rate of impacts varies with time, with the same expected value,
would the results derived still apply?

13

@seismicisolation
@seismicisolation
Solution
−x
1
fX (x) = e 2
2
−x
Pr(X > x) = 1 − FX (x) = e 2
−x −x
(a) Model as a new Poisson process with rate = (10)(e 2 ) = 10e 2 = v in
e−v v n
. For x = 6, v = 0.498. Let N = number of impacts. Then
n!
e−0.498 (0.498)n
pN (n) = , for n = 0, 1, 2, . . .
n!
If n = 0,
pN (n) = e−0.498 = 0.6077
(to 4 decimal places).
(b) Let Z = x in previous result. Then Pr(number of impacts that exceed
Z = x) = e−v , where
−x
v = 10 [1 − FX (x)] = 10e 2
Therefore
−x
−10 exp( )
FZ (x) = e 2
or · ¸
−z
FZ (z) = exp −10 exp( )
2
which is the double-exponential distribution (see chapter 9).
(c) Yes. But tIme between impacts would no longer be exponential.

Exercise 3.11
The concentration X of pollutant in a lake has been modelled using a normal
distribution based on past measurements, with a mean of 2.21 and a standard
deviation of 0.46, in parts per million (ppm). Find the following.
(a) Pr(X < 1.5) .
(b) Pr(1.5 < X < 2.5) .
(c) Pr(X > 2.5) .
(d) The probability that in six samples, (a) occurs 3 times, (b) twice and (c)
once.

14

@seismicisolation
@seismicisolation
Solution
(a) Pr(X < 1.5) = Pr(X ≤ 1.5) for continuous RQ’s.
1.5 − 2.21
Pr(X ≤ 1.5) = Pr(Z ≤ ) = Pr(Z < −1.54) = 0.0618
0.46

(b)
2.5 − 2.21
Pr(1.5 < X < 2.5) = Pr(−1.54 < Z < )
0.46
= Pr(−1.54 < Z < 0.63) = 0.6739

(c)
Pr(X > 2.5) = Pr(Z > 0.63) = 0.2643
Check: 0.0618 + 0.6739 + 0.2643 = 1.0000.
(d) Assume independence. Required probability is

6! 3 2 1
(0.0618) (0.6739) (0.2643) = 0.001700
3!2!1!
(Multinomial distribution).

Exercise 3.12
Two rebars used in the construction of a gravity-based concrete offshore struc-
ture are connected by a coupling. The strength of each rebar under the cyclic
loading expected is normally distributed with a mean of 300 MPa and a standard
of 30 MPa.

(a) Let X denote the strength of a rebar. What is the value of strength
X = xf corresponding to FX (x) = 0.025?
(b) It is judged that the strengths of the two rebars are stochastically inde-
pendent. What is the probability that exactly one of the two rebars will
fail at a strength less than xf ?
(c) An assembly of two rebars is tested. Denote the strength of the assembly
as Y. What is the value of FY (xf ) given that FX (xf ) is 0.025?

Solution
(a) χf = µ − z0.025 σ = 300 − 1.96 × 30 = 241.2
(b) (2) (0.025) (0.975) (binomial) = 0.04875 or = 0.025 + 0.025 − 2(0.025)2 .

15

@seismicisolation
@seismicisolation
 The assembly will fail when the weaker of the two rebars fails. Let the
strength of thse be X1 and X2 . Then
Y = min(X1 , X2 ),
and
Pr(Y > y) = Pr [(X1 > y) (X2 > y)] = Pr [(X1 > y) ∧ (X2 > y)]
h i
2
1 − FY (y) = 1 − FX (y)

1 − FY (χf ) = (0.975)2 = 0.950625

FY (χf ) = 0.049375 ∼= 5%
h i
or 2 × 0.025 − (0.025)2

Exercise 3.13
Find the mean, variance and standard deviation for problem 3.1 and the first
of 3.2.

Solution
Problem 3.1

hxi = (0.3) (0) + (0.3) (1) + (0.2) (2) + (0.1) (3) = 1.0

­ ® h i
σ 2x = x2 − µ2x = (0.3) (1)2 + (0.2) (2)2 + (0.1) (3)2 − 1 = 1
σx = 1
Problem 3.2
3
fX (x) = (1 − x2 ), for 0 < x < 1
2
Z 1 µ ¶ · ¸1
3 3 x2 x4 3
hXi = x (1 − x2 ) dx = − =
0 2 2 2 4 0 8
­ ®
σ 2X = X 2 − µ2x
Z 1 µ ¶
­ 2® 3 1
X = x2 (1 − x2 ) dx =
0 2 5

1 9 19
σ 2X = − = = 0.0594
5 64 320
σX = 0.2437

16

@seismicisolation
@seismicisolation
Exercise 3.14
The Cauchy pdf with parameter a,
k
fX (x) = , for − ∞ < x < ∞,
x2 + a2
is a bell-shaped curve resembling in a general way the normal distribution. But
it has much thicker tails.

(a) Find the value of k so that the pdf is well-defined.

(b) Find the cdf FX (x) .
1
(c) Find the inter-quartile range, the difference in x between FX (xL ) = 4
and FX (xU ) = 34 .
(d) Find the mean and standard deviation.

Solution
(a) f (x) ≥ 0 for all x if an only if k ≥ 0.
Z ∞ Z ∞
k
f (x; a) dx = 2 2
dx
−∞ −∞ x + a
µ ¶
k h −1 ³ x ´i∞ k π −π kπ
= tan = − =
a a −∞ a 2 2 a
Z ∞
a
f (x; a) dx = 1 =⇒ k =
−∞ π
Therefore, for f (x; a) to be a well defined probability density function, we
must have
a
k=
π
(b)
Z x Z x · µ ¶¸x
a 1 1 −1 t
F (x; a) = f (t; a)dt = dt = tan
−∞ π −∞ t2 + a2 π a −∞
1 ³x´ 1
F (x; a) = tan−1 +
π a 2

(c)
³x ´ µ ¶
1 1 L 1 1
F (xL ; a) = =⇒ tan−1 + =
4 π a 2 4
³ x ´ −π
L
=⇒ tan−1 =
aµ ¶4
xL −π
=⇒ = tan = −1
a 4

17

@seismicisolation
@seismicisolation
 Therefore xL = −a and, by symmetry, xU = +a, and

IQR = 2a

By symmetry, µ = 0; or evaluate the limit:

Z k
a x a £ ¡ ¢¤k
lim = lim ln x2 + a2 −k = 0
k→∞ π −k x2 2
+ a dx 2π k→∞

2 £ ¤
V [X] = E[X 2 ] − (E [X]) = E X 2 − 0
Z Z
a ∞ x2 a ∞ x2 + a2 − a2
= dx = dx
π −∞ x2 + a π −∞ x2 + a2
Z ∞µ ¶
a a2 a ∞
= 1− 2 2
dx = [x]∞ − 1 = ∞
π −∞ x +a π

Therefore
σ=∞

Exercise 3.15
The Pareto pdf is given by

ktk
fX (x) = , for x ≥ t
xk+1
= 0, for x < t.

(a) Find necessary conditions on k to ensure that hX n i is finite.

(b) Find the range of values for which the variance is finite. Find an expression
for this variance.

Solution
(a)
Z ∞ Z t Z ∞
n n ktk
E [X ] = x f (x) dx = 0dx + xn dx
−∞ −∞ t xk+1
If k = n, then
E [X n ] = ntn [ln x]∞
t =∞

Otherwise

· ¸∞  −0 + ktn
ktk xn−k k>n
E [X n ] = = k−n
n−k t  ∞ k≤n

18

@seismicisolation
@seismicisolation
 (b) £ ¤ 2
V [X] = E X 2 − (E [X])
£ ¤
E X 2 is finite provided k > 2; E [X] is finite provided k > 1. V [X] is
finite iff E[X 2 ] and E [X] are both finite. Therefore V [X] is finite provided
k > 2.
³ ´
2
µ ¶2 kt2 (k − 1)2 − k (k − 2)
kt kt
V [X] = − =
k−2 k−1 (k − 1)2 (k − 2)

kt2
V [X] =
(k − 1)2 (k − 2)

Exercise 3.16
An object (possibly an iceberg) is drifting on the ocean surface with a speed X
that is random and modelled by a gamma distribution:

xα−1 exp (−x/β)

fX (x) = .
β α Γ (α)

This distribution is based on a survey of all objects in a large area. Consider

now the event that the object collides with an offshore installation. The prob-
ability of this event is proportional to the speed X. Derive an updated velocity
distribution using Bayes’ theorem given the event and using the pdf noted as a
prior distribution. Show that this updated distribution has a gamma distribu-
tion with shape parameter changed from α to α + 1. Plot the prior and updated
distributions for α = 2.5, and with β = 0.11 and 0.38 ms−1 .

Solution
Let V = velocity, a random quantity, with pdf fV (v) . Then

Pr(v < V < v + dv) = fV (v) dv

(prior to collision). Let C = the event of a collision; then

Pr(C | v < V < v + dv)fV (v) dv

Pr(v < V < v + dv | C) = R ,
all V
(above)
by Bayes’ Theorem.
Thus

vfV (v) dv
fV |C (v)dv = R
f (v)dv
all V V
Substituting the equation for the gamma distribution:

19

@seismicisolation
@seismicisolation
 −v
v·v e βα−1
fV |C (v) = α R ∞ ........Equation A
β Γ(α) 0 · · ·
The numerator takes the form
−v −v
0
v e β =v
α α −1
e β

where α0 = α + 1
Therefore equation A above is a new gamma distribution
−v
0
v α −1 e β
fV |C (v) = 0
β α Γ (α0 )
The numerator is part of the gamma family so that denominator follows this
distribution, because it must integrate out to give a distribution with an area
of unity under the curve.
[If you are not convinced by the preceding arguments, consider:
Equation A above
−v
vα e β
−v
R ∞ vαe β
β α Γ (α) 0 dv
β α Γ (α)
where the β α Γ (α) terms cancel. Thus A :

−v
v e β α
fV |C (v) = −v
R∞
v e β dv
0
α

v
Put α0 = α + 1, = v 0 ,dv = βdv 0 , then
β
Z ∞ Z ∞ 0
0
0 α −1 −v 0 α0 −1 α −1 0 0
(βv ) e 0
βdv = β β v0 e−v dv 0 = β α Γ (α0 )
0 0

(latter by definition)

20

@seismicisolation
@seismicisolation
1 Exercise 3.17
An interesting example arises in arctic engineering. The area of an ice floe
(assuming it to be circular of area A and radius R, approximately) is given by
A = πR2 . If R is random, then so is A. What is the mean area?

21

@seismicisolation
@seismicisolation
1.1 Solution
A = πR2
­ ®
hAi = π R2 .

From equation (3.70), the mean area is

­ ®
π R2 = π(µ2R + σ 2R ).

Contrast with the usual A = πR2 .

2 Exercise 3.18
Geometric probability 1. A space probe is planned to land at a point P on the
surface of a planet. A local coordinate system is constructed in the tangent to
the surface at P. The probability of landing outside a circle of radius 5 km with
P as centre is considered to be zero, and the probability within is judged to be
uniformly distributed over the circular area with the density proportional to the
area. Find the probability that the space probe will land

(a) within 2 km of P,
(b) between 2 and 3 km from P,
(c) north of the aiming point,
(d) south of the aiming point and more than 2 km from P, and
(e) exactly due north of P.

Solution
(a) Required probability
π (2)2 4
= 2 = = 0.16
π (5) 25

(b) Required probability

¡ ¢
π 32 − 22 5 1
= 2 = = = 0.20
π (5) 25 5

(c) Required probability

1
=
2

22

@seismicisolation
@seismicisolation
 (d) Required probability
¡ ¢
1 π 52 − 22 21
= · 2 = = 0.42
2 π (5) 50

(e) Since the area associated with this event is zero,

required probability = 0.

Exercise 3.19
Geometric probability 2. Two signals are sent to a receiver, and arrive within
the time interval t. The instants at which the signals arrive are denoted X1 and
X2 . If the signals arrive within the small interval τ of each other, interference
and deterioration of the signals result. Find the probability of this event, on the
assumption that X1 and X2 are independent and equally likely at any point in
the interval [0, t] .

Solution
Draw the square {(0, t) , (0, t)} in the first quadrant with axes x1 and x2 . Shade
the following area
|x1 − x2 | ≤ τ

x x -x
1 2
2

t x 1 .
Required probability is

shaded area t2 − (t − τ )2 ³ τ ´2
= = = 1 − 1 −
total area t2 t

23

@seismicisolation
@seismicisolation
Exercise 3.20
Geometric probability 3.

(a) Two ships berth at the same quay in a harbour. The arrival times for
the ships are independent and uniformly distributed (equally probable)
during any 24 hour period. If, during the time wi (needed for mooring,
loading and unloading) between arrival and departure of either ship, the
other ship arrives, then the other ship has to wait to use the berth. For
the first ship this waiting time is w1 = 1 hour. For the second ship the
waiting time is w2 = 2 hours. Estimate the probability that one of the
ships will have to wait to use the berth during the next 24 hour period.
(b) A ferry carries freight continually from one bank of a river to another.
This takes one hour. A ship moving along the river at right angles to the
direction of travel of the ferry can be seen from the ferry if it passes the
trajectory of the ferry less than 20 minutes away from its position. All
times and places of intersection are judged to be equally likely. What is
the probability that the ship will be seen from the ferry?

Solution
(a) Let X1 = arrival time for first ship; X2 = arrival time for second ship.

x 2

_1
2
24 hrs
2_
2
1

2 t x 1
24 hrs
½
x2 − x1 ≤ 1,
x2 ≥ x1
x2 = x1 + 1
½
x1 − x2 ≤ 2
x1 ≥ x2
x2 = x1 − 2

24

@seismicisolation
@seismicisolation
  µ ¶ 
 √ 1 1 

 23 2 + √ √ = 23.5 

Central area = µ 2¶ 2 69.5
 √ 2 2 

 22 2 + √ √ = 46 

2 2
Required probability
69.5
= 2 = 0.12069572̇
(24)

(b)

River Banks

Ferry
x

20 min 20 min 20 min

1 1
First or last of ferry’s passage:
3 3
Z 1 µ ¶
3 1 1
Pr(ship seen) = x+ dx =
0 3 6

1
Middle :
3
1 2 2
Pr(ship seen) = · =
3 3 9
Required probability:
1 2 5
=2· + =
6 9 9

25

@seismicisolation
@seismicisolation
Exercise 3.21
Geometric probability 4. On the surface of the earth the area bounded by two de-
grees of longitude and by the line of latitude θ degrees to θ+1 is to a good approx-
imation, a rectangle of sides 111.12 km (north-south) and 111.12cos (θ + 0.5) km
(east-west). A ship of beam (width) b is crossing the degree-rectangle on a path
of total length d. A stationary iceberg (or other object) of width w is placed at
a random position in the degree-rectangle.

(a) Find the probability that the ship will collide with the object during the
crossing of the degree-rectangle and evaluate the probability given θ =
41◦ , b = 28 m, w = 110 m and d = 83.2 km (these are plausible values for
the RMS Titanic in 1912), or using some other values of interest.
(b) Using the values obtained in (a) find the probability that there are no
collisions after 100 crossings, and the number of crossings above which
the probability of a collision just exceeds 0.50.

Solution
(a)

(b + w) · d
(if lengths all in metres)
1111202 · cos(θ + 0.5)
= 0.0012415 (correct to 5 significant figures)

(b)
88%
558

Exercise 3.22
Find the skewness of the exponential distribution
µ ¶
1 x
fX (x) = exp − for x ≥ 0
β β
= 0 otherwise.

Solution
Z −x
∞
3 1 β
m(3)
x = (x − β) e dx;
0 β
x
P ut y =
β

26

@seismicisolation
@seismicisolation
 Z ∞
m(3)
x = β 3
(y − 1)3 e−y dy
Z0
3
¡ ¢
= β [ y y 2 − 2y + 1 e−y dy − 1]
Z Z
3
¡ 3 −y ¢
= β [ y e dy − 2 y e dy + ye−y dy − 1]
2 −y

¡ ¢
= β 3 [ y 3 e−y dy − 4
Substitute u = y 3 and integrate by parts:
3 3
m(3)
x = β [6 − 4] = 2β

Skewness is
m3x 2β 3
γ1 = 3
= 3 =2
σx β

Exercise 3.23
The joint probability distribution pX,Y (x, y) of the random quantities X and
Y is as follows.
x
0 1 2 3
1. 1 2/27 0 0 1/27
y 2 6/27 6/27 6/27 0
3 0 6/27 0 0
(a) Check for coherence and determine the marginal distributions pX (x) and
pY (y) .
(b) Determine pX|Y (x, y) , pY |X (x, y) and pY |X=0 (y) .
(c) Find the means µX , µY , the standard deviations σ X , σ Y as well as hXY i
and the covariance σ X,Y .
(d) Determine the covariance σ X,Y .
(e) Are X and Y independent?

Solution
PP
(a) pX,Y (x, y) = 1
0 1 2 3 pY (y)
1 2/27 0 0 1/27 3/27 = 1/9
y 2 6/27 6/27 6/27 0 18/27 = 2/3
3 0 6/27 0 0 6/27 = 2/9
pX (x) 8/27 12/27 6/27 1/27 27/27

27

@seismicisolation
@seismicisolation
 (b) pX|Y

x
0 1 2 3
1 2/3 0 0 1/3
y 2 1/3 1/3 1/3 0
3 0 1 0 0
pY |X
x
0 1 2 3
1 1/4 0 0 1
y 2 3/4 1/2 1 0
3 0 1/2 0 0
pY |X=0 is given by the first column in the preceding table.

(c)
X µ ¶ µ ¶ µ ¶ µ ¶
8 12 6 1
µX = hXi = xpX (x) = 0· + 1· + 2· + 3· =1
27 27 27 27
19 1
µY = =2
9 9

2 8 2 6 2 1 18 2
σ 2X = (1) + (1) + (2) = =
27 27 27 27 3
σX ∼
= 0.817
µ ¶2 µ ¶2 µ ¶ µ ¶2
10 3 1 18 8 6 234
σ2Y = + + =
9 27 9 27 9 27 729
σY ∼
= 0.567
XX 19
hX Y i = xypX,Y (x, y) =
9
σ X, Y = hX Y i − µX µY = 0

(e) No. Since σ X,Y = 0,either X and Y are stochastically independent or

the dependence is non-linear. For definiteness, need pX,Y (x, y) = pX (x) ·
pY (y) for all x, y.

Checking the entries above it is found that this is not true.

Exercise 3.24
Given the pdf

fX,Y (x, y) = 2 exp (−x − y) , for 0 ≤ x ≤ y and 0 ≤ y < ∞

= 0 otherwise,

find the marginal distributions and check for independence.

28

@seismicisolation
@seismicisolation
Solution
It appears as if the density is separable into fX · fY since:
¡ −x ¢ −y
fX,Y = 2e (e )
| {z }
→ not a normalizable pdf.
But the first expression is not a normalizale pdf, as indicate.
In more detail:

Z ∞
fX (x) = 2 ex−y dy = 2e−2x ;
Zxy
£ ¤
fY (y) = 2 e−x−y dx = 2e−y 1 − e−y ;
0

fX, Y 6= fX · fY ,
so X, Y are not independent.

Exercise 3.25
Prove that the correlation coefficient ρXY is such that

−1 ≤ ρX,Y ≤ 1,

always. Comment on the meaning of ρXY = 0.

Solution
Let

U = aX + bY
σ 2U = a2 σ 2X + b2 σ 2Y + 2abσX,Y Equation 3.102
> 0 since σ 2U > 0

The discriminant of this quadratic expression is positive; hence:

σ2X σ 2Y > σ 2X,Y

If
aX + bY = c
(c a constant) giving a linear relationship,

σ2X σ 2Y = σ 2X,Y

29

@seismicisolation
@seismicisolation
In general
σ2X σ 2Y ≥ σ 2X,Y
and
σ X,Y
−1 ≤ ≤1
σX σY
Note that

pX,Y = 0 −→ σ X,Y = 0
implying a total absence of a linear relationship.

Exercise 3.26
Raiffa decision problem. Consider an additional choice in the examples ?? and
?? of section 3.5.2. You are offered the option of sampling two balls for a fee
of $12 rather than the single ball at $8 in example ??. The contracts offered
remain the same as before. Note that the updated probabilities are the subject
of exercise 2.18. Is the new option an attractive one based on EMV compared
to the others?

Solution
Raiffa: Problem 2.18 has probabilites.
Note that:

24
Pr(RR) = ;
90
42
Pr(RP or P R) =
90
24
Pr(P P ) =
90

30

@seismicisolation
@seismicisolation
 The option presented (e2 ) is optimal (based on EMV ).

3 Exercise 3.27
A long distance telephone company is planning to increase the capacity of its
networking system. A team set up to study the further demand is unable to
reach a firm decision and concludes that the annual growth rate will be 1, 2,
or 3 percent with equal probability. Three capacities will be adequate for these
three growth rates for the design period. The respective capacities are 5, 10,
and 15, in arbitrary units. Economics of scale given by the equation

Cost = k(capacity)0.6

can be applied to the cost of the three capacity units. Draw a decision tree
for this problem and indicate what decision you would recommend on the basis
of maximum expected monetary value. Complete the calculation in terms of
present monetary units, i.e. ignore interest on money and the effects of inflation.

31

@seismicisolation
@seismicisolation
Solution
-2.63

-2*2.63
Accept either
alternative -3*2.63

-2.63

-2*2.63
5
5 -(2.63+3.98)

10 -3.98

-3.98
15
-(3.98+2.63)

-5.08
-5.08

-5.08
In the above we have taken k = 1, i.e.
Capacity Cost
5 (5)0.6 = 2.63
10 (10)0.6 = 3.98
0.6
15 (15) = 5.08
Choose α2 or α3 depending on strategy for adding capacity.

Exercise 3.28
Aggregates for a highway pavement are extracted from a gravel pit. Based on
experience with the material from this area, it is estimated that the probabilities
are Pr(G) =Pr(good-quality aggregate) = 0.70, and Pr(G̃) = Pr(poor-quality
aggregate) = 0.30. In order to improve this prior information, the engineer
considers a testing programme to be conducted on a sample of the aggregate.
The test method is not perfectly reliable–the probability that a perfectly good-
quality aggregate will pass the test is 80%, whereas the probability of a poor-
quality aggregate passing the test is 10%.

(a) Draw both Bayes calculation and decision trees for the problem. The de-
cision tree should include the options “do not implement test programme”
and “implement test programme”, as well as “accept the aggregate” and
“reject the aggregate”. Determine the optimal decision on the basis of
minimum-expected-cost, given the following costs (on a nominal scale).
Accept, good aggregate, 15 units; accept, poor aggregate, −8 units; re-
ject, good or poor aggregate, 5 units; cost of testing, 2 units.

32

@seismicisolation
@seismicisolation
 (b) If a sample is tested, and if it passes, what is the probability of good-
quality aggregate?
(c) If a second test is performed, and the sample again passes, what is the
probability of good-quality aggregate? One can use the result of part (b)
as a prior probability in this case.

Solution
g 15

~
g
-8
g 5

~
g 5

13
g

~
g
-10
g 3

~
g 3
g 13

~
g
g 3

~
g 3
(a)
(b) 0.95 (from above)
(c)
(0.95) (0.8) 0.76
= = 0.993
(0.95) (0.8) + (0.005) (0.1) 0.76 + 0.005

Exercise 3.29
In the design of a dam for flood control, the height of the dam depends on
two factors. The first is the level of water in the dam, denoted X, and the
second is the increase caused by flood waters coming into the reservoir, de-
noted Y . The level X takes the values {30, 40, 50 metres} with probabilities
{0.2, 0.5, 0.3} respectively while Y takes values {10, 20, 30 metres} with prob-
abilities {0.3, 0.4, 0.3} respectively. It is decided to use a decision analysis to
determine the dam height. Two values are under consideration, 65 and 75m,
with costs of 12 and 15 units respectively. The consequences of overtopping
are flooding downstream with costs of 15 units. (All costs can be treated as
additive). A hydrological study is under consideration in order to obtain a bet-
ter estimate of the probabilities of the flood levels noted. For the purposes of

33

@seismicisolation
@seismicisolation
the present analysis, the results of the study can be summarized by the values
of R = 1 (low), R = 2 (medium) and R = 3 (high). The method has the
reliabilities indicated in the following table.
Pr(R|Y )
R=1 R=2 R=3
Y = 10 0.70 0.20 0.10
Y = 20 0.10 0.80 0.10
Y = 30 0.05 0.05 0.90
Update the probabilities of Y, i.e. determine the values Pr(Y |R) . Complete
a decision analysis with tree, including the option: (complete study, do not
complete study) and the option (dam height= 65m, dam height = 75m). The
cost of the study is 2 units.

Solution
Need Pr(Y | R);(Bayes)

Pr(R = r | Y = y)Pr(Y = y)
Pr(Y = y | R = r) = P
Pr(R = r | Y = y)Pr(Y = y)
all y

R=1 R=2 R=3

Pr(R = 1) = 0.265
Y = 10 0.792 0.152 0.088
Pr(R = 2) = 0.396
Y = 20 0.151 0.810 0.118
Pr(R = 3) = 0.340
Y = 30 0.057 0.038 0.794
Overtopping occurs if X + Y > height
Y X
40 30
65)
50 20
50 30

Y X
75)
30 50

34

@seismicisolation
@seismicisolation
 0.36
17.4
0.64
0.09
75
16.35 0.91
e
0
0.0909
13.36
65
15.36 0.9091
0.0171
75
e 15.26
1
R =1
0.2734
65 0.7266
R =2 16.10 0.0114

17.17 75 0.9886
15.17
0.6706
R =3
65 0.3274
22.06
0.2382
20.57 75
18.57 0.7618

Exercise 3.30
Referring to problem 2.26, the engineer has to decide whether or not to close
down a section of the pipeline to conduct major repairs. The following values of
utility are assigned. Subtract 2 units as a “cost” if tests are undertaken. Draw a
decision tree and determine the optimal course of action based on the criterion
of maximum expected utility. the symbol R denotes “repairs undertaken”.
S Given R Given R̃
0 7 10
1 6 5
2 5 0

35

@seismicisolation
@seismicisolation
Solution
6.89 7
R 6
5
9.45 10
~
R 5
0
4.985 5
7.925 R 4
3
7.925 8
~ 3
R -2
e 0
I=0
4.767
6.835
I=1 R
e 1

6.835
~
R
I= 4.275
4.375 R

~
R
4.375

Exercise 3.31
The speeds of vehicles on a highway are modelled by a normal distribution with
a mean of 60 kmh−1 and a standard deviation of 15. Determine the pdf of
the relative speed of two vehicles moving in opposite direction. Assume now
that the drivers, having seen a police car in the area, obey a speed limit of 75
kmh−1 ; those exceeding the limit now drive at the limit, with the other speeds
remaining unchanged. Sketch and describe the new distribution of speeds (in
one direction).

Solution
Let V1 = speed of car in one direction, and V2 = speed of car in opposite
direction. Assume independence of V1 and V2 (reasonable).
(V1 + V2 ) is N (60 + 60, 225 + 225) or N (120, 450), i.e a s.d. of 21.21km/h

µ ¶
75 − 60
Pr(V > 75) = 1 − FV (75) = 1 − FZ = 1 − FZ (1.00) = 0.1587
15

The right tail is replaced by a Dirac Delta function

36

@seismicisolation
@seismicisolation
 Spike of
Probability

60 75

Exercise 3.32
Consider a sequence of four waves in a storm. The wave heights are {X1 , X2 , X3 , X4 } .
These quantities are judged to be exchangeable, so that all 24 permutations such
as {X1 , X3 , X2 , X4 } or {X4 , X2 , X3 , X1 } are considered equivalent. For quanti-
ties that are not exchangeable there are four values hXh i , six hXh Xi i for (h < i),
and four hXh Xi Xj i for (h < i < j) . What can be said of these in the case of
exchangeable random quantities?

Solution
There is only one value of each expected value.
hXh i is the same for all h
hXh , Xi i is the same for all h, i
hXh , Xi , Xj i is the same for all h, i, j

Exercise 3.33
Sketch random processes that illustrate the four combinations of: discrete or
continuous in time and discrete or continuous random quantities.

Solution
See Ochi, 1990

37

@seismicisolation
@seismicisolation
Exercise 3.34
In a Poisson process with constant intensity λ, show that the time T to the kth
event is modelled by the Erlang distribution

λ (λt)k−1 exp (−λt)

fT (t) = , for t > 0, k ≥ 1.
(k − 1)!

Verify that this is a special case of the gamma distribution.

If accidents occur at the rate of 0.2 per month on a highway, describe the
distribution of the time in months to the third accident.

Solution
(λt)k−1 e−λt
fT (t)dt = · λdt
(k − 1)!
The terms on the r.h.s represent Pr(k − 1 events in time up to t, Poisson
distribution) and Pr(kth accident during t to t+dt), respectively.
Time to third accident is represented by the pdf

(0.2) (0.2t)2 e−0.2t

fT (t) = = 0.004t2 e−0.2t
2

Exercise 3.35
Aircraft are serviced at a loading bay. An engineer is studying the waiting times
and usage of space and decides to model the situation by dividing time into a
discrete set of small time intervals during which one arrival is expected. During
one such time step, the crew at the loading bay can service {0, 1, 2} aircraft with
probabilities {3/8, 1/2, 1/8} . The following strategy is adopted. One aircraft is
allowed to wait to enter the bay if one is being serviced there. Consequently,
when the bay has zero or one aircraft being serviced, an additional aircraft is
then permitted to enter or to wait. In the case where one is being serviced,
and one is waiting, any arriving aircraft are routed elsewhere. The states under
consideration are {0, 1, 2} , corresponding to the number of aircraft at the bay
(waiting or being serviced) at the end of each time step. The end of a time step
is defined such that any completed aircraft have moved on and before a new one
is accepted.

(a) Derive the one-step transition matrix.

(b) Find the long-run stationary state, if there is one.
(c) If the bay starts out the day empty, so that an aircraft is admitted at the
beginning of the next step, write down the first three probability vectors
qi .

38

@seismicisolation
@seismicisolation
Solution
(a)    
p00 p01 p02 5/8 1/8 1/8
pi| j = pi j =  p10 p11 p12  =  3/8 1/2 1/2 
p20 p21 p22 0 3/8 3/8
¡ ¢
Think of pi| j !
(b)
[pi j ] qj∗ = qi∗
¡5 ¢ 
8 − 1¡ q1∗ + ¢18 q2∗ + 18 q3∗ = 0 

3 ∗ 1 ∗ 1 ∗ 
8 q1 + 2 − 1 q2 + ¡ 32 q3 =¢ 0∗
3 ∗
8 q2 + 8 − 1 q3 = 0 


q1∗ + q2∗ + q3∗ = 1

1 ∗ 15 ∗ 9
q1∗ = ;q = ;q =
4 2 32 3 32
(c)      
 1   5/8   28/64 = 7/16 
q0 = 0 ;q = 3/8 ;q = 27/64
  1   2  
0 0 9/64

Exercise 3.36
Find the autocovariance of

(a)
X (t) = A cos (ωt + Ξ) ,
where A and Ξ are independent random quantities, the latter uniformly
distributed in the interval [−π, π] .
(b)
Xn
X (t) = [Ai cos (ω i t) + Bi sin (ω i t)] ,
i=1
where all of the Ai and the Bi are independent random quantities with
hAi i = hBi i = 0 and σ 2Ai = σ 2Bi = σ 2i .

Solution
(a)

hX(t)i = hAi hcos(wt + Ξ)i

= 0

39

@seismicisolation
@seismicisolation
 (independence)
® ­
σ X,X (τ ) = A2 hcos(wt + Ξ) · cos [w (t + τ ) + Ξ]i
1 ­ 2®
= = A cos (wτ )
2

(b)
n
X
hX(t)i = [hAi i cos (wi t) + hBi i sin (wi t)]
i=1
= 0
 µ ¶ 
* P +

 Ah cos (wh t) + Bh sin (wh t) 

σ X,X (τ ) = µ h ¶

 P 
 · Ai cos [wi (t + τ )] + Bi sin [wi (t + τ )] 
i

Note: hAh Bi i = hAh i hBi i (independence). Similarly hAh Ai i = 0, if

h 6= i. Therefore:
X ­ ® ­ ®
σX,X (τ ) = { A2h cos (wh t) cos [wh (t + τ )] + Bh2 σ 2h sin(wh t) sin [wh (t + τ )]}
| {z } | {z }
σ 2h σ2h
X
= σ 2h {cos (wh t) cos [wh (t + τ )] + sin(wh t) sin [wh (t + τ )]}
h
X
= σ 2h cos (wh τ )

40

@seismicisolation
@seismicisolation
Solutions for Chapter 4
Exercise 4.1
Benefit-cost ratio and qualitative measures. An authority is considering dam
construction to supply clean water to the region. Two alternatives, 1 and 2
below, are being considered for water supply and flood control. The following
table shows costs and benefits associated with the two alternatives together with
other consequences. The costs and benefits are annual values.

Alternative Water Supply Flood Control

1: one large dam Costs $20.2 ×106 $8.0 ×106
Benefits $23.2 ×106 $8.3 ×106
2: two small dams Costs $28.0 ×106 $9.1 ×106
Benefits $24.5 ×106 $10.0 ×106

Alternative Other Consequences

1: one large Costs Flooding of 50 km2 of natural woodland and habitat
dam Relocation of people, reduction in wild life
Benefits Improved human health for local population
Recreational facilities for 100,000 persons per year
2: two small Costs Flooding of 23 km2 of natural woodland and habitat
dams Relocation of people, reduction in wild life
Benefits Improved human health for local population
Recreational facilities for 200,000 persons per year

(a) Which project has the better benefit-cost ratio? Use stated dollar amounts.
(b) Suggest ways to assist the authority to make a decision taking into account
the other qualitatively measured consequences listed above.
(c) Comment on the merits of benefit-cost ratios in decision-making in par-
ticular in the above.

Solution
Benefits: 31.5 × 106
(a) Alternative 1: Costs: 28.2 × 106
Benefit-cost ratio = 1.12
Benefits: 34.5 × 106
Alternative 2: Costs: 37.1 × 106
Benefit-cost ratio = 0.93

On this basis, Alternative 1 is the better choice. But the other factors
favour Alternative 2.

@seismicisolation
@seismicisolation
 (b) Following are main points

• Use of subjective judgement

• Reduction of all iterms to equivalent dollar value (difficult in some
cases)
• Ordinal scales - and judgement (cannot use probablisitic weighting)
• Develop interval/ratio scales and develop utility measure.

(c) Very arbitrary. No risk aversion. Hard to deal with qualitative factors.
Interest rates can be arbitrary and tailored to achieve desired answers.

Exercise 4.2
Time value of money. Commodity prices can be very volatile and the problem
is posed in arbitrary units. An oil company has to make a decision on whether
to invest in offshore production from a marginal oil field, the size of which is
estimated at 200 MB (million barrels). The development requires capital ex-
penditure of 150 million per year for five years at which point production would
commence at the rate of 10 MB per year for 20 years. The production period
would also incur operating and maintenance costs of 75 million per year. The
costs noted so far are in real (constant-worth) dollars, with constant purchasing
power with regard to year 1 of the project. In a preliminary analysis of the prob-
lem, an engineer considers two scenarios as described in the table below. These
represent two possible developments of the world economy in average terms, the
second corresponding to a period of conflict in the middle east (rather alarmist
but useful for contrast). The values are in actual (inflated) dollars.
Calculate the net present worth for the two scenarios. For real dollars, use
a real interest rate calculated from the figures in the table. Calculate the EMV
based on a probability of 0.80 for scenario as against 0.2 for scenario 2.

Scenario 1–stable
Oil Price Interest Rate Inflation Rate
15 per bbl 8% per year 4% per year
increasing at over 20 year over 20 year
5% per year period period

Scenario 2–volatile
Oil Price Interest Rate Inflation Rate
25 per bbl 15% per year 12% per year
increasing at over 20 year over 20 year
10% per year period period

@seismicisolation
@seismicisolation
Solution
Let ir = real interest rate, f = inflation rate, im = money interest rate (actual
$’s). Then
im = ir + f + ir f

1 + im
ir = −1
1+f
(∼
= im − f for small rates)

Scenario 1
1 + im 1 + 0.08
ir = −1= − 1 = 0.0385 = 3.85%
1+f 1.04
For capital investment O and M,

5 20
(1.0385) − 1 (1.0385) −1 1
NP W = −150 − 75 ·
(0.0385) (1.0385)5 (0.0385) (1.0385)20 (1.0385)20
= −671 − 485 = −1156

(in millions of monetary units). ¡ ¢n

For oil sales use actual $’s and interest corresponding to year n of 15 1.05
1.08
mon units
←
bbl
1
→ 1.0286 =⇒ 2.86%

(1.0286)20 − 1 1
N P W = 150 20 · 20 = 1286
0.0286 (1.0286) (1.0286)

Total N P W = 130
Scenario 2
ir = 1.15
1.12 − 1 = 0.0268 = 2.68%
For investment and O and M:

N P W = −150(4.60) − 75(14.36)(0.589) = −690 − 634 = −1324

For oil sales:

N P W = 250(12.95)(0.411) = 1331

Total N P W = 7

@seismicisolation
@seismicisolation
 }
130
0.8
EMV = 105.4
Invest
0.2 7

Status quo

Note re oil price (io is the increase in oil proce):

· ¸n 
(1 + io ) 1 

(price) ·  · 1 ¸n
1
· + f 1
¸n + ir
1 + io 
 1 + iom
= 
1 + im
where iom is a special factor.

Exercise 4.3
Consider the following payoff matrix for a fisher, who has to decide which species
to fish on in a season. The states of nature are uncertain. Solve the problem
using the maximin, maximax, Hurwicz (α = 0.5) and minimax regret criteria.
Comment on the strategies. What is meant by utility dominance? Give an
example.
State of Nature
θ1 θ2 θ3
Species A 1 1 1
Species B 1 3 0
Species C 0 4 0
Species D 1 2 0

@seismicisolation
@seismicisolation
Solution
Maximin and State of Nature
Maximax Ä1 Ä2 Ä3 Maximin Maximax
A 1 1 1 1 1
B 1 3 0 0 3
C 0 4 0 0 4
D 1 2 0 0 2
Hurwicz α = 0.5
Min. Max.
A 1 1 0.5 × (1) + 0.5 × (1) 1.0
B 0 3 0.5 × (0) + 0.5 × (3) 1.5
C 0 4 0.5 × (0) + 0.5 × (4) 2.0
D 0 2 0.5 × (0) + 0.5 × (2) 1.0
Regret State of Nature
Matrix Ä1 Ä2 Ä3 Regret| Ä1 Regret| Ä2 Regret| Ä3 Row Max.

}
A 1 1 1 0 3 0 3
B 1 3 0 0 1 1 1
C 0 4 0 1 0 1 1 Minimax
D 1 2 0 0 2 1 2

Utility Dominance occurs if all the utilities for a particular strategy are
≥ those for another strategy, e.g., utiltiies for strategy B is ≥ those for D in
the above question. We also say strategy B dominates strategy D. In strict
dominance, > replaces ≥ (there is no example in this question)

Exercise 4.4
Imagine that you are a member of the Board of Directors of a small company,
with assets of about $50,000. The table below shows a choice to be made
by the board between three contracts (α1 , α2 , and α3 ). The entries show the
consequences with associated probabilities in brackets. Draw the decision tree.
Infer utility values for each outcome using π-basic-reference-lottery-tickets. Plot
the resulting values of utility against money. Iterate if necessary. Complete the
decision analysis using expected monetary value and expected utility criteria.
Consequences in $ (probability)
α1 −80, 000 (0.2) 50, 000 (0.8)
α2 20, 000 (0.6) 4, 000 (0.4)
α3 100, 000 (0.2) 0 (0.8)

@seismicisolation
@seismicisolation
Solution
EMV Solution
α1 : 0.2 (−80, 000) + 0.8 (50000) = $24, 000 Optimal - but risky
α2 : 0.6(20, 000) + 0.4(4000) = $13, 600
α3 : 0.2(100, 000) + 0.8(0) = $20, 000
Utilities—important to note that these are subjective, and can vary from
person to person.
For the determination of utility, consider the π − brlt

100,000
π

1- π

-80,000

Typical values might be as follows:

π Consequences ($)
1.00 100, 000
0.93 50, 000
0.83 20, 000
0.78 4, 000
0.75 0
0.00 −80, 000

Expected utilities:
Choice E(u)
α1 0.74
α2 0.81
α3 0.80

@seismicisolation
@seismicisolation
 Here, α2 is optimal.

Exercise 4.5
Consider the statement: “A decision-maker is risk-averse if she or he prefers the
expected consequences of a lottery to the lottery itself”. Take two examples:

(i) an attribute with increasing preferences, e.g., money.

(ii) an attribute with decreasing preferences, e.g. size of an oil spill.

Illustrate the statement above with regard to these two cases and sketch the
utility functions. Assume that the utilities range between 0 and 1. Use lotteries
with 50-50% probabilities.

Solution
See Figure 4.16 in text.

Exercise 4.6
An engineer is working with a client in formulating a response to a potential oil
spill in an offshore region. The volumes of spill under consideration are in the
range 0-50,000 m3 . Sketch a utility function expressing risk aversion. Prove that
the shape shown must express risk aversion by using the certainty equivalent of
h0, 50000i.Consider the function

u (v) = b [c − exp (av)]

where v = volume of spill, and a, b and c are constants. Discuss whether this
might be used as a utility function. Let u (0) = 1, and u (50000) = 0. Show how
the certainty equivalent noted can be used with this information in evaluating
the constants a, b and c.

@seismicisolation
@seismicisolation
Solution

0
0.5

v̂ ≈ h0, 50, 000i = ≈

0.5

50,000

@seismicisolation
@seismicisolation
 u(0)
0.5

0.5

u(50,000)
If we accept v̂ > v̄ in exchange for < 0, 50, 000 >, we express risk aversion,
since we accept the excess (v̂ − v̄) above the mean (v̄), to avoid the possibility
of a larger spill (of 50, 000 m3 ). Now

u (v̂) = 0.5 [u (0) + u (50, 000)] ;

this dictates the shape as shown above.

u(0) = b(c − 1) = 1
u(50, 000) = 0 = b(c − e50,000a )
0.5 = b(c − ev̂a ), given v̂

Now solve for "a," "b" and "c"

u(0) = b(c − 1) = 1
¡ ¢
u(50, 000) = 0 = b(c − e50,000a ) =⇒ −1 = b 1 − e50,000a
1 ¡ ¢
u(35, 000) = = b(c − e35,000a ) =⇒ −0.5 = b 1 − e35,000a
2

£ ¤
1 − e50,000a
2 = =⇒ 1 = 2e35,000a − e50,000a
[1 − e35,000a ]
a = 0.036

By trial and error, iteration. Hence

b = 0.198
c = 6.05

@seismicisolation
@seismicisolation
Exercise 4.7
Study the problem of noise, for example in reduction of noise pollution in siting
an airport. What would constitute a good attribute in attempting to formu-
late utility? Consider intensity (energy per unit area per unit time–Wm−2 ).
For utility, would decibels serve? (Note that doubling the decibel level “buys”
intensity on a log scale).

Solution
Noise intensity is a decreasing attribute.
A risk averse utility function would look like:

Now
µ ¶
i
β = 10 log10
io
in decibels.

10

@seismicisolation
@seismicisolation
 β

Conclude: decibel system, taken as u = −β, is risk prone!!

i (energy units) dB
1014 140 Jet aircraft (near)
130 Pain threshold
120 Thunderclap (near)
100 Heavy manufacturing
80 3rd floor high rise downtown
65 Urban shopping area
40 − 60 Residential area
40 Farm in a valley
20 Grand canyon
One "buys" more and more sound (relatively) with the same increase in dB.

Exercise 4.8
Contrast the exponential form of Example 4.3 and the logarithmic form

u (x) = ln [a (x + b)]

for expressing risk aversion. Obtain the local risk aversion measure γ (x) of
Equation (4.17). Use examples from your experience (cost, money, time saved
in construction or manufacturing, delays, quality of product,. . . ).

11

@seismicisolation
@seismicisolation
Solution
We saw that the local risk aversion (lra) measure γ(x) for (−e−cx ) was a con-
stant, c. For
U (x) = ln [a (x + b)] ,
note first that in
u(x) = ln(a) + ln(x + b)
| {z }
the constant indicated dissapears upon integration.
1
u0 =
x+b
−1
u00 = 2
(x + b)

and
1
γ(x) =
x+b
The lra measure γ is positive and decreasing in x for x > −b (or x + b > 0).
This might be more appropriate for money than constant risk aversion as in a
logarithmic function (Daniel Bernoulli).

Exercise 4.9
Find the local risk aversion measure γ (x) for the quadratic utility function

u (x) = −ax2 + bx + c

with a, b > 0, and x < b/2a, so that u is increasing. Comment.

Solution

u0 = −2ax + b
u00 = −2a

and
2a
γ(x) = .
−2ax + b
The measure γ > 0 under condiitons stated. But γ increases with x so we
have increasing risk aversion.

12

@seismicisolation
@seismicisolation
Exercise 4.10
Strategic equivalence of two utility functions is defined such that the same pref-
erences are implied for any lotteries with the same consequences. Show that the
relationship between the two utility functions uA and uB of the kind given in
Equation (4.6), i.e. uB = a + buA , b > 0 results in strategic equivalence. Use
Equation (4.15) to assist in the solution. Show also that strategic equivalence
results in the positive linear transformation noted.

Solution
Utility functions are UA and UB ;
x̂ = u−1 −1
A [hUA (X)i] = uB [hUB (X)i]

Then
uA (x̂) = hUA (X)i
uB (x̂) = hUB (X)i = a + b hUA (X)i = a + buA (x̂)
This is true for all x̂, i.e. for all x. So we have strategic equivalence.
If we have strategic equivalence , we can suppose that the lotteries are the
kind:

hx1 , π, x2 i ≈ x
be considered for all x.
uA (x) = πuA (x1 ) + (1 − π)uA (x2 ) Equation 1
uB (x) = πuB (x1 ) + (1 − π)uB (x2 ) Equation 2
Solve for π from 2 and substitute into 1; this gives the desired result.

Exercise 4.11
Cost and safety
Cost - dollars, x
Safety - lives lost (or −ve), y
Consider additive independence for example:

13

@seismicisolation
@seismicisolation
 One could accept this, considering that "cost" does not matter where lives
are at stake.
Then we can write:
u (x, y) = k1 u1 (x) + k2 u2 (y) .
If the sums of money become exorbitantly large, (millions....) we might have
to reconsider. For example: ¾
x = 100, 000
? Use money elsewhere ?
x1 = 1, 000, 000 or 10, 000, 000

Exercise 4.12
Raiffa problem.

14

@seismicisolation
@seismicisolation
 σ1: e0 , α1

σ2: e0 , α2

σ3: e1; α1 if R, α1 if P

σ4: e1; α1 if R, α2 if P

σ5: e1; α2 if R, α1 if P

σ6: e1; α1 if R, α2 if P

Strategy u (σ i |θ1 ) u (σi |θ2 )

σ1 40 −20
σ2 −5 100
σ3 −8 + 40 (4/10) + −8 + (−20) (9/10) +
40 (6/10) = 32 (−20) (1/10) = −28
σ4 −8 + 40 (4/10) + −8 + (−20) (9/10) +
(−5) (6/10) = 5 (100) (1/10) = −16
σ5 −8 + (−5) (4/10) + −8 + (100) (9/10) +
40 (6/10) = 14 (−5) (1/10) = 81.5
σ6 −8 + (−5) (4/10) + −8 + (100) (9/10) +
(−5) (4/10) = −13 (100) (9/10) = 92

Typically:

α2
R

θ1
P
α1
α2
R
θ2

P
α1

15

@seismicisolation
@seismicisolation
 By inspction, σ 1 will be optimal for Pr(Ä1 ) = 0.8

Pr(Ä2 ) Optimal strategy

0 − 0.49 σ2
0.49 − 0.65 σ5
0.65 − 1 σ1

Exercise 4.13

(0.2) (0.5)
Pr(M = 0.2 | Y = 0.6) =
(0.2) (0.5) + (0.3) (0.4) + (0.7) (0.1)
0.1 0.10
= = = 0.345
0.1 + 0.12 + 0.07 0.29
12
Pr(M = 0.4 | Y = 0.6) = = 0.414
29

16

@seismicisolation
@seismicisolation
 7
Pr(M = 0.6 | Y = 0.6) = = 0.241
29
Conclude: e0 , B̃

Exercise 4.14
e0 = "do not take survey" α1 = "build system"
e1 = "take survey" α2 = "do not build system"
Possible strategies:
σ1 : e0 ; α1
σ2 : e0 ; α1
σ3 : e1 ; α1 if Y = 0.2, α1 if Y = 0.4, α1 if Y = 0.6
σ4 : e1 ; α2 if Y = 0.2, α1 if Y = 0.4, α1 if Y = 0.6
σ5 : e1 ; α1 if Y = 0.2, α2 if Y = 0.4, α1 if Y = 0.6
σ6 : e1 ; α1 if Y = 0.2, α1 if Y = 0.4, α2 if Y = 0.6
σ7 : e1 ; α2 if Y = 0.2, α2 if Y = 0.4, α1 if Y = 0.6
σ8 : e1 ; α2 if Y = 0.2, α1 if Y = 0.4, α2 if Y = 0.6
σ9 : e1 ; α1 if Y = 0.2, α2 if Y = 0.4, α2 if Y = 0.6
σ 10 : e1 ; α2 if Y = 0.2, α2 if Y = 0.4, α2 if Y = 0.6
The utilities u (σ i | M = 0.2, 0.4, 0.6) are easily calculated, for example:
u (σ 9 | M = 0.4) = −2 + 60 Pr(Y = 0.2 | M = 0.4)
+60 Pr(Y = 0.4 | M = 0.4) + 60 Pr(Y = 0.6 | M = 0.4)
= −2 + 60(0.1) + 60(0.6) + 60(0.3) = 58
All of the values are summarised in the following table:
σ i u(σi | M = 0.2) u(σ i | M = 0.4) u(σ i | M = 0.6) u(σ i )
σ1 30 60 90 48
σ2 60 60 60 60
σ3 28 58 88 46
σ4 43 58 83.5 53.1
σ5 37 58 83.5 50.1
σ6 34 58 67 46.9
σ7 52 58 79 57.1
σ8 49 58 62.5 54.0
σ9 43 58 62.5 51.0
σ 10 58 58 58 58
σ2 is the optimal strategy, i.e. no study, no LRT!

17

@seismicisolation
@seismicisolation
 Maximin 60
σ i u(σi | M = 0.2) u(σ i | M = 0.4) u(σ i | M = 0.6)
σ1 30 60 90 30
σ2 60 60 60 60
σ3 28 58 88 28
σ4 43 58 83.5 43
σ5 37 58 83.5 37
σ6 34 58 67 34
σ7 52 58 79 52
σ8 49 58 62.5 49
σ9 43 58 62.5 43
σ 10 58 58 58 58
Choose σ 2

Maximax 90
σ i u(σi | M = 0.2) u(σ i | M = 0.4) u(σ i | M = 0.6)
σ1 30 60 90 90
σ2 60 60 60 60
σ3 28 58 88 88
σ4 43 58 83.5 83.5
σ5 37 58 83.5 83.5
σ6 34 58 67 67
σ7 52 58 79 79
σ8 49 58 62.5 62.5
σ9 43 58 62.5 62.5
σ 10 58 58 58 58
Choose σ 1

Hurwicz
σ i u(σi | M = 0.2) u(σ i | M = 0.4) u(σ i | M = 0.6)
σ1 30 60 90
σ2 60 60 60
σ3 28 58 88
σ4 43 58 83.5
σ5 37 58 83.5
σ6 34 58 67
σ7 52 58 79
σ8 49 58 62.5
σ9 43 58 62.5
σ 10 58 58 58
Say α = 0.4
α index of σ 1 = (0.4) (30) + (0.60) (90) = 66
α index of σ 2 = (0.4) (60) + (0.60) (60) = 60
α index of σ 3 = (0.4) (28) + (0.60) (88) = 64
α index of σ 4 = (0.4) (43) + (0.60) (83.5) = 67.30
α index of σ 5 = (0.4) (37) + (0.60) (83.5) = 64.90

18

@seismicisolation
@seismicisolation
 α index of σ 6 = (0.4) (34) + (0.60) (67) = 53.8
α index of σ 7 = (0.4) (52) + (0.60) (79) = 68.20 ←− Optimal decision
α index of σ 8 = (0.4) (49) + (0.60) (62.5) = 57.10
α index of σ 9 = (0.4) (43) + (0.60) (62.5) = 54.70
α index of σ 10 = (0.4) (58) + (0.60) (58) = 58

Minimax Regret
σ i u(σi | M = 0.2) u(σ i | M = 0.4) u(σ i | M = 0.6) Row maxima
σ1 30 0 0 30
σ2 0 0 30 30
σ3 32 2 2 32
σ4 17 2 6.5 17
σ5 23 2 6.5 23
σ6 26 2 23 26
σ7 8 2 11 11
σ8 11 2 27.5 27.5
σ9 17 2 27.5 27.5
σ 10 2 2 32 32
Strategy σ 7 is optimal in this case.

(ii) ******* Listed as page 7 of hand written solutions ***********

Note: σ4 is on the line joining σ 1 and σ 7 : hence, if one is indifferent between

strategies σ1 and σ 7 , one is indifferent between σ1 , σ 7 and σ 4
(and of course any other point on line σ 1 − σ 7 ). The equation of a line in a
plane ¯ (x1 , y1 ) and (x2 , y2 ) is
¯ joining points
¯ x y 1 ¯
¯ ¯
¯ x1 y1 1 ¯ = 0
¯ ¯
¯ x2 y2 1 ¯
¯ ¯ ¯ ¯ ¯ ¯
¯ y1 1 ¯ ¯ x1 1 ¯ ¯ x1 y1 ¯
or ¯ ¯ ¯ ¯ ¯ ¯
¯ y2 1 ¯ x − ¯ x2 1 ¯ y + ¯ x2 y2 ¯ = 0

19

@seismicisolation
@seismicisolation
 i.e. (y1 − y2 ) x − (x1 − x2 ) y + x1 y2 − x2 y1 = 0
or Ax + By = D
The Direction cosines of the normal are given by:
A B
, where ∆2 = A2 + B 2
∆ ∆
σ1 − σ7 Direction cosines of the normal are given by:
90 − 79 (30 − 52)
1 − i.e. 0.45, 0.89
2 2
(11 + 22 ) 2
1
(112 + 222 ) 2 | {z }
Vectorial
Length =1

=⇒ 0.33, 0.67
| {z } P
Probabilities (normalised so that pi = 1

σ7 − σ2
79 − 60 (52 − 60)
1 − , i.e. 0.92, 0.39 =⇒ 0.70, 0.30
(192 + 82 ) 2 192 + 82

Prior Probability of M = 0.2 Prior Probability of M = 0.2 Optimal Strategy

0 − 0.33 1 − 0.67 σ1
0.33 − 0.70 0.67 − 0.30 σ7
0.70 − 1 0.30 − 0 σ2
At Pr(= 0.2) = 0.33, one is indifferent between σ 1 , σ 4 and σ 7

Exercise 4.15
(a)
15
S
Accept
batch

-50
P
Reject
batch 0
15
e 1
S
Accept
batch

e 0
-50

Reject
batch 0

20

@seismicisolation
@seismicisolation
(b)
Pr(S) = 0.98 Pr(S̃) = 0.02
(0.15) (0.98)
Pr(S | P̃ ) = = 0.913
(0.15) (0.98) + (0.70) (0.02)
Pr(S̃ | P̃ ) = 0.087
hU i = (0.913) (15) + (0.087) (−50) = 9.35
This is > 0 (utility for rejecting batch)

∴ Accept.
(c)
(0.15) (0.5) 0.15
Pr(S | P̃ ) = = = 0.176
(0.15) (0.5) + (0.70) (0.5) 0.85
Pr(S̃ | P̃ ) = 0.824
hU i = (0.176) (15) + (0.824) (−50) = −38.6
∴ Reject batch.
(d)
σ 1 : P − accept, P̃ − accept
σ 2 : P − accept, P̃ − reject
σ 3 : P − reject, P̃ − accept
σ 4 : P − reject, P̃ − reject
³ ´
σi u (σ i | S) u σi | S̃
σ1 15 −50
σ2 12.75 −15
σ3 2.25 −35
σ4 0 0

(no figures on cost of testing - assumed to be included in utility).

21

@seismicisolation
@seismicisolation
Exercise 4.16
³ ´
(1) (1) (1)
Let us take the case of a plane through three points x1 , x2 , x3 ,
³ ´ ³ ´
(2) (2) (2) (3) (3) (3)
x1 , x2 , x3 and x1 , x2 , x3 .
The equation of the plane in determinantal form is:

¯ ¯
¯ x1 x2 x3 1 ¯
¯ ¯
¯ (1)
x1
(1)
x2
(1)
x3 1 ¯
¯ ¯
¯ (2) (2) (2) ¯=0
¯ x1 x2 x3 1 ¯
¯ ¯
¯ (3)
x1
(3)
x2
(3)
x3 1 ¯

Expand in terms of the first row:

¯ ¯ ¯ ¯ ¯ ¯
¯ x(1) x(1) 1 ¯ ¯ x(1) x(1) 1 ¯ ¯ x(1) (1)
x2 1 ¯
¯ 2 3 ¯ ¯ 1 3 ¯ ¯ 1 ¯
¯ (2) (2) ¯ ¯ (2) (2) ¯ ¯ (2) (2) ¯
¯ x2 x3 1 ¯ x1 − ¯ x1 x3 1 ¯ x2 − ¯ x1 x2 1 ¯ x3
¯ (3) ¯ ¯ (3) ¯ ¯ (3) ¯
¯ x 1 ¯ ¯ x ¯ ¯ x ¯
(3) (3) (3)
2 x3 1 x3 1 1 x2 1
¯ ¯
¯ x(1) (1)
x2
(1)
x3 ¯
¯ 1 ¯
¯ (2) (2) (2) ¯
− ¯ x1 x2 x3 ¯ =0
¯ (3) ¯
¯ x (3)
x2
(3)
x3 ¯
1

Write this as ax1 + bx2 + cx3 = d, where a, b, c, d are defined by the deter-
minants above.
The direction cosines of the normal to the plane are (numerically)
A B C
, , , where ∆2 = A2 + B 2 + C 2 .
∆ ∆ ∆
One would have first to identify the efficient set - vertices of the feasible set.
Then the feasible planes (hyperplanes) would be analysed as above. General-
izations of this method can be developed.

Exercise 4.17
Abandoned Game Example
Three players are taking turns to win a game of chance.
The winner will receive all of a pot of S dollars. The losers receive nothing.
When it is Player A0 s turn, the probability that Player A wins is immediately
pA .
If Player A does not win immediately, then it becomes Player B’s turn.
When it is Player B 0 s turn, the probability that Player B wins is immediately
pB .
If Player B does not win immediately, then it becomes Player C’s turn.

22

@seismicisolation
@seismicisolation
 When it is Player A0 s turn, the probability that Player C wins is immediately
pC .
If Player C does not win immediately, then it becomes Player A’s turn again.
The game is still in progress, with Player A about to play, when the game
has to be abandoned. What is the fairest way to divide the pot?
The event "Player A wins" is the union of an infinite set of events.
If Player A does not win on the first turn after the instant when the game
is abandoned, an if the other two players also fail, then Player A gets another
chance in turn #4.
If all players fail in their first two attempts, then player A gets another
chance in turn #7, and so on indefinitely.
Let A, [B, C] represent the event "player A, [B, C] wins on the current
turn". Then the event "Player A wins" is:
A v ÃB̃ C̃A v ÃB̃ C̃ ÃB̃ C̃A v ÃB̃ C̃ ÃB̃ C̃ ÃB̃ C̃A v ....
The probability of this event is:

2
pA + (1 − pA ) (1 − pB ) (1 − pC ) pA + ((1 − pA ) (1 − pB ) (1 − pC )) pA
3
+ ((1 − pA ) (1 − pB ) (1 − pC )) pA + ....

∞
X k
= pA · ((1 − pA ) (1 − pB ) (1 − pC ))
k=0
pA
= (geometric series)
1 − (1 − pA ) (1 − pB ) (1 − pC )

The event "Player B wins" is:

ÃB v ÃB̃ C̃ ÃB v ÃB̃ C̃ ÃB̃ C̃ ÃB v ÃB̃ C̃ ÃB̃ C̃ ÃB̃ C̃ ÃB v ....
The probability of this event is:

(1 − pA ) pB + (1 − pA ) (1 − pB ) (1 − pC ) (1 − pA ) pB
+ (1 − pA ) · ((1 − pB ) (1 − pC ) (1 − pA ))2 pB
+ (1 − pA ) · ((1 − pB ) (1 − pC ) (1 − pA ))3 pB + ....

∞
X k
= (1 − pA ) · pB · ((1 − pA ) (1 − pB ) (1 − pC ))
k=0

(1 − pA ) pB
= (geometric series)
1 − (1 − pA ) (1 − pB ) (1 − pC )

The event "Player C wins" is:

ÃB̃C v ÃB̃ C̃ ÃB̃C v ÃB̃ C̃ ÃB̃ C̃ ÃB̃C v ÃB̃ C̃ ÃB̃ C̃ ÃB̃ C̃ ÃB̃C v ....

23

@seismicisolation
@seismicisolation
 The probability of this event is:

(1 − pA ) (1 − pB ) pC + (1 − pA ) (1 − pB ) pC
2
+ (1 − pA ) (1 − pB ) · ((1 − pC ) (1 − pA ) (1 − pB )) + ....

∞
X
= (1 − pA ) · (1 − pB ) · pC · ((1 − pA ) (1 − pB ) (1 − pC ))k
k=0

(1 − pA ) (1 − pB ) pC
=
1 − (1 − pA ) (1 − pB ) (1 − pC )

Therefore the fairest way to divide the pot is in the ratio

pA : (1 − pA ) · pB : (1 − pA ) · (1 − pB ) · pC

(a)
pA = pB = pC = 16
(which is the case if the winner is the first to throw "doubles" on a pair of
fair standard dice) and a pot of S = $273.

The fair ratio becomes

1 5 1 25 1
: · : · = 36 : 30 : 25
6 6 6 36 6
Therefore the fairest way to split the pot is not equal shares of $91, but
Player A receives $108
Player B receives $90
Player C receives $75

(b)
pA = 0.2, pB = 0.5, pC = 0.9
If the pot were to be divided into equal shares, then each player would receive
$32.
If the pot were to be divdided int he raio pA : pB : pC , then the shares would
be
Player A receives $12
Player B receives $30
Player C receives $54
However the fair ratio is:
1 4 1 4 1 9
: · : · · = 5 : 10 : 9
5 5 2 5 2 10

24

@seismicisolation
@seismicisolation
 Therefore the fairest way to split the pot is not shares of $12, $30, and $54,
but
Player A receives $20
Player B receives $40
Player C receives $36

25

@seismicisolation
@seismicisolation
Solutions for Chapter 5
Exercise 5.2
Solve the following optimization problem, in which v represents the volume of
a tank.
Max ( v = x1 x2 x3 )
subject to

x1 + x2 + 2x3 = b and
x1 , x2 , x3 ≥ 0.

Solution
L = x1 x2 x3 − λ (x1 + x2 + 2x3 − b)

∂L
= x2 x3 − λ = 0
∂x1
∂L
= x1 x3 − λ = 0
∂x2
∂L
= x1 x2 − 2λ = 0
∂x3
∂L
= x1 + x2 + 2x3 − b = 0
∂λ
Solving:

b b b b3
x1 = , x2 = , x3 = ; V =
3 3 6 54
or
V = x1 x2 (0.5b − 0.5x1 − 0.5x2 )

∂V
= 0.5bx2 − x1 x2 − 0.5x22 = 0
∂x1
∂V
= 0.5bx1 − x1 x2 − 0.5x21 = 0
∂x2
Solve to get same result as above.
Check for critical point:
∂V 2 b ∂V 2 ∂V 2
2 = −x2 = − 3 , ∂x22
= −x1 = − 3b , b
∂x1 ∂x2 = 0.5b − x1 − x2 = − 6
¯∂x1 ¯
¯ −b/3 −b/6 ¯
¯ ¯ hence f (x1 , x2 ) is concave and also the max.
¯ −b/6 −b/3 ¯ > 0

@seismicisolation
@seismicisolation
Exercise 5.3
A closed right cylindrical pressure vessel of volume v is to be constructed so as
to have minimum surface area a. Find the radius r and length that achieve
this. Discuss how one might investigate the character of the stationary point
found.

Solution
¡ ¢
Min a = 2πr2 + 2πrl , or
¡ ¢
Min r2 + rl

v
v = πr2 l = const −→ r2 l = const = c =
π
¡ ¢
L = r2 + rl − λ r2 l − c

∂L
¾ 1
∂r = 2r + l = 2rλl = 0 r= λ
∂L 2 2
∂l = r − r λ = 0
l= λ

∂L
= r2 l − c = 0
∂λ
Then
r r
3 2 1 3
c 3
v
λ = ; = =
c λ 2 2π
r
v
r = 3
2π
r
v
l = 23
2π
This actually corresponds to a saddle point in L. The character of this point
in area a above can be investigated by the method of constrained variation
(Beveridge and Schechter,1970) and found to be a minimum.

Exercise 5.4
Solve the following optimization problem, in which H represents the entropy.
There is a countable set of states xi , i = 0, 1, 2, . . . , of the random quantity X
with associated probabilities pi .
³ Xn ´
Max H = − pi ln pi
i=1

@seismicisolation
@seismicisolation
subject to coherence X∞
pi = 1,
i=0
and X∞
pi xi = hXi = µ.
i=0

Solution
Standard result (see Chapter 6)

e−λxi
pi = P −λxi
ie

Exercise 5.5
Sketch the feasible region and stationary points of Example 5.6. Investigate and
determine the minima of Equation (5.22).

For Solution
See Beveridge and Schechter (1970) pages 269,270, 275
Tables V11-2 and V11-3

Exercise 5.6
Solve by the simplex method and check the answer graphically.

Min z = −10x1 + 4x2

Subject to 2x1 + x2 ≤ 9
x1 − 2x2 ≤ 2
−3x1 + 2x2 ≤ 1
x1 , x2 ≥ 0.

Solution
Simplex Method

Min z = −10x1 + 4x2

Subject to 2x1 + x2 + x3 = 9
x1 − 2x2 + x4 = 2
−3x1 + 2x2 + x5 = 1

xi ≥ 0, i = 1, 2, ..., 5

@seismicisolation
@seismicisolation
 x1 x2 x3 x4 x5 Basis
9 2 1 1 x3 x3
2 1 −2 1 x4 x4
1 −3 2 1 x5
z − 0 −10 4
5 5 1 −2 x3
2 1 −2 1 x1
7 −4 3 1 x5
z + 20 −16 10
1 1 1/5 −2/3 x2
4 1 2/3 1/5 x1
11 4/5 7/5 1 x5
z + 36 16/5 18/5
x∗1 = 4, x∗2 = 1, (other variables = 0), and
0
zmax = 36, zmin = −36

@seismicisolation
@seismicisolation
Exercise 5.7
Solve by the simplex method.

Max z 0 = −3x1 − 6x2 + 3x3 − 3x4

Subject to x1 − x2 + x3 ≤ 5
x1 + x2 + 2x4 ≤ 6
x2 − 2x3 + x4 ≤ 3
−x1 + 2x2 + x3 ≤ 2
x1 , x4 ≥ 0
x2 , x3 : unrestricted in sign.

Solution
Replacing the unrestricted variables x2 , x3 by the difference of two non-negative
variables, i.e.,

x2 = x+ −
2 − x2
x3 = x3 − x−
+
3

and adding the slack variables x5 , x6 , x7 and x8 , yields the following simplex
tableau, in which "Max" has been converted to "Min".
x1 x+ 2 x−
2 x+
3 x−
3 x4 x5 x6 x7 x8 Basis
5 1 −1 1 1 −1 1 x5
6 1 1 −1 2 1 x6
3 1 −1 −2 2 1 1 x7
2 −1 2 −2 1 −1 1 x8
z−0 3 6 −6 −3 3 3
5 1 −1 1 1 −1 1 x2
11 2 1 −1 2 1 1 x6
8 1 −1 1 1 1 1 x7
12 1 3 −3 2 1 x8
z + 30 9 3 −3 3 6
13 2 −1 1 1 2 1 x2
19 3 3 2 1 1 x6
8 1 −1 1 1 1 1 x3
36 4 3 5 3 x8
z + 54 12 6 9 3 1

x∗1 = 0
x∗2 = x+ −
2 − x2 = 0 − 10 = −13
x3 = x3 − x−
∗ +
3 = 0 − 6 = −8
x∗4 = 0
zmax = 54

@seismicisolation
@seismicisolation
Exercise 5.8
Solve by the simplex method.

Max z 0 = −2x1 + 4x2 − 2x3

Subject to 3x1 − x2 + x3 ≤ 4
x1 − x2 + x3 ≤ 2
−2x1 + x2 − x3 ≤ 4
x1 , x2 , x3 ≥ 0.

Solution
Convert "Max" to "Min", add three slack variables:
x1 x2 x3 x4 x5 x6 Basis
4 3 −1 1 1 x4
2 1 −1 1 1 x5
4 −2 1 −1 1 x6
z−0 2 −4 2
8 1 1 1 x4
6 −1 1 1 x5
4 −2 1 −1 1 x2
z + 16 −6 −2 4
8 1 1 1 x1
14 1 1 2 x5
20 1 −1 2 x2
z + 64 −2 6 10

The negative coefficient in the objective function of the tableau indicates

that further improvement is possible, and that x3 should enter the basis. The
absence of x3 in the first two constraints and its negative coefficient in the third
imply that x3 is unconstrained. Hence the solution is unbounded.

Exercise 5.9
Solve by the simplex method.

Min z = −x1 − x2 − 2x3

Subject to 4 ≥ x1 ≥ 1
3x2 − 2x3 = 6
2 ≥ x3 ≥ −1
x2 ≥ 0.

@seismicisolation
@seismicisolation
Solution

Min z = −x1 − x2 − 2x+ −

3 + 2x3
Subject to x1 + x4 = 4
x1 − x5 + x6 = 1
¡ ¢
3x2 − 2 x+ −
3 − x3 + x7 = 6
x+ −
3 − x3 + x8 = 2
−x+ −
3 + x3 + x9 = 1
+ −
¡xi +≥ 0, i−=
¢ 1, . . . , 9; x3 , x3 ≥ 0. Since x3 is unrestricted, it has been replaced
by x3 − x3 ; x6 and x7 are artificials. The objective of phase 1 routine is to
reduce the artificial variables to zero, i.e.,
Min: w = x6 + x7 = 7 − x1 − 3x2 + 2x+ −
3 − 2x3 + x5

@seismicisolation
@seismicisolation
 x1 x2 x+
3 x−
3 x4 x5 x6 x7 x8 x9 Basis
4 1 1 x4
1 1 −1 1 x6
6 3 −2 2 1 x7
2 1 −1 1 x8
1 −1 1 1 x9
z−0 −1 −1 −2 2
w−7 −1 −3 2 −2 1
4 1 1 x4
1 1 −1 1 x6
1
2 1 −2/3 2/3 3 x2
2 1 −1 1 x8
1 −1 1 1 x9
1
z+2 −1 8/3 8/3 3
w−1 −1 1
3 1 1 −1 x4
1 1 −1 1 x1
2 1 −2/3 2/3 x2
2 1 −1 1 x8
1 −1 1 1 x9
z+3 −1 1
w−0 1
3 1 1 x4
1 1 −1 x1
10
3 1 x2
2 1 −1 x3
3 x9
z + 25/3 −1
3 1 1 x6
4 1 1 x1
10/3 1 2/3 x2
2 1 −1 1 x3
3 1 1 x9
z + 34/3 1 0 8/3
x∗1 = 4
x∗2 = 10
3
x∗3 = 2
0
zmax = 343 ; zmin =
−34
3

Exercise 5.10
Check the answers to problems 5.6 to 5.9 by using computer software.

@seismicisolation
@seismicisolation
Comment
There are many sources for computer solutions to LP, e.g. in Excel.

Exercise 5.11
Structural systems.

(a) An engineer is considering the structural system in Figure (a) on which

the loads F1 , F2 and F3 represent storage loads which will be imposed for
a short duration. The objective is to maximize the total storage subject
to the capacity of the columns. These correspond to buckling loads of 200
tonnes for column 1, 100 tonnes for columns 2 and 3, and 150 tonnes for
column 4. Treat all beams as being simply supported at their ends.
(b) The structural system in Figure (b) is comprised of horizontal tension
members (for example wires) supported by vertical beams. The loads
A, B and C are applied as shown in the figure, together with the capacities
of the tension members. Assume that the beams can support the loads.
What is the maximum load (A + B + C) that can be supported?

Solution
F2
(a) 2 ≤ 100

7 5 1
F1 + F2 + F3 ≤ 200
8 8 4
1 3 3
F1 + F2 + F3 ≤ 150
8 8 4
Using units of 100’s of tonnes:

F2 + x1 = 2
7F1 + 5F2 + 2F3 + x2 = 16
F1 + 3F2 + 6F3 + x3 = 12

Max (F1 + F2 + F3 ) = z 0 −→ Min (−F1 − F2 − F3 ) = z; z = z0

@seismicisolation
@seismicisolation
 F1 F2 F3 x1 x2 x3 Basis
2 1 1 x1
16 7 5 2 1 x2
12 1 3 6 1 x3
z−0 −1 −1 −1
2 1 1 F2
6 7 2 −5 1 x2
6 6 −3 1 x3
z+2 −1 −1 1
2 1 1 F2
4 6 23 −4 −1/3 x2
−1
1 1/6 1 2 1/6 F3
z+3 −5/6 1/2 1/6
2 1 1 F2
3
5 1 −3/5 −1/20 F1
0.9 1 −4/10 21/100 F3
z + 3.5 −1/2 1/8
→ Multiple solution; can move to
F2 = 0, F1 = 180, F3 = 170, z 0∗ = 350 tonnes
z ∗ = −3.5
z 0∗ = 3.5 −→350 tonnes
F1 = 70 
F2 = 200 above × 100

F3 = 90

10

@seismicisolation
@seismicisolation
 (b)
2 5 3
A+ B+ C ≤ 600
7 7 7
5 2 4
A+ B+ C ≤ 600
7 7 7
2 2
B+ C ≤ 200
3 9
1 4
B+ C ≤ 200
3 9
2
C ≤ 100
3
1
C ≤ 100
3
Maximize (A + B + C) subject to the above set of constraints. Two opti-
mal solutions are,

A = 720N ; B = 300N ; C = 0N
and
A = 620N ; B = 250N ; C = 150N

Exercises 5.12 and 5.13

5.12 Show that the dual of the dual problem in LP is the primal problem.
5.13 Complementary Slackness Theorem. There is a natural relationship be-
tween the excess or slack variables in either the primal or dual program
and the decision variables in the other. Whenever an excess or slack vari-
able is not equal to zero (at optimality), the associated decision variable
in the dual program is equal to zero. When a decision variable is greater
than zero, the associated excess or slack variable is zero. The solution is
such that
Xn
aij x∗j = bi , or yi∗ = 0 or both, for i = 1, 2, . . . , m;
j=1
Xm
aij yi∗ = cj or x∗j = 0 or both, for j = 1, 2, . . . , n.
i=1

Introduce the slack and excess variables

Xn
xn+i = −bi + aij xj , for i = 1, 2, . . . , m, and
j=1
Xm
ym+j = cj − aij yi , for j = 1, 2, . . . , n,
i=1

to show that

x∗n+i = 0 or yi∗ = 0 or both;

∗
ym+j = 0 or x∗j = 0 or both.

11

@seismicisolation
@seismicisolation
 Xm
aij yi∗ = cj whenever x∗j > 0;
i=1
Xn
yi∗ = 0 whenever aij x∗j > bi .
j=1

Comment
There are many excellent texts on LP and applications, for example Jones (1980)
and Chvátal (1980), which may be referred to.

Exercise 5.14
Consider the following product-mix problem. A company manufactures two
products. Let x1 and x2 be the amounts of the products, say A and B, made.
Various resources such as labour and different kinds of raw materials are needed
to make the products. In the present problem, there are 3 resources required.
The amounts available are bi = 5, 3, 4 units respectively for resources labelled
i = 1, 2 and 3. The units just mentioned would typically be person-days, tonnes
of steel, m3 of fuel, and so on. Each unit of product A requires 2 units of the first
resource (i = 1), while product B requires one unit. The corresponding values
for the second resource (i = 2) are unity for both products while for resource 3
(i = 3) they are 0 and 2. The profit is 1.5 per unit of A and 1 per unit of B
produced. The profit is to be maximized subject to the availability of resources.
(a) Solve using the simplex method
(b) Find the dual of the linear program.
(c) Use the complementary slackness theorem to deduce the solution of the
dual problem.
(d) Provide an interpretation of the dual problem in terms of “shadow prices”
(see for example Monahan, 2000).

Solution
(a)
Max (1.5x1 + x2 = z 0 )
subject to 2x1 + x2 ≤ 5
x1 + x2 ≤ 3
2x2 ≤ 4
x1 , x2 ≥ 0. Standard form:
Min (−1.5x1 + x2 = z)
subject to 2x1 + x2 + x3 = 5
x1 + x2 + x4 = 3
2x2 + x5 = 4

12

@seismicisolation
@seismicisolation
 all x0 s > 0. This is also the canonical form.
b x1 x2 x3 x4 x5 Basis
5 2 1 1 0 0 x3
3 1 1 0 1 0 x4
4 0 2 0 0 1 x5
z−0 −1.5 −1
5
2 1 1/2 1/2 0 0 x1
1
2 0 1/2 −1/2 1 0 x4
4 0 2 0 0 1 x5
z + 15/4 0 −1/4 3/4
2 1 0 1 −1 0 x1
1 0 1 −1 2 0 x2
2 0 0 2 −4 1 x5
z+4 0 0 1/2 1/2
¾ x1 = 2, x2 = 1
z ∗ = −4
Optimal solution: x5 = 2
z 0∗ = 4
x3 = x4 = 0

(b) Dual
Min (5y1 + 3y2 + 4y3 = w)

Subject to  
· ¸  y1  ½ ¾
2 1 0 1.5
y ≥
1 1 2  2  1.0
y3

13

@seismicisolation
@seismicisolation
 or
2y1 + y2 ≥ 1.5 Value of 1 unit of A
y1 + y2 + 2y3 ≥ 1.0 Value of 1 unit of B

(c) Since x5 = 2, y3 = 0. Since decision variables x1 and x2 are > 0, excess

variables in dual are zero.

Hence
¾ y1 = 0.5
2y1 + y2 = 1.5
y2 = 0.5
y1 + y2 = 1.0
w∗ = 4

(d) See Monahan (2000); web search will be fruitful.

Exercise 5.15
A town produces a minimum of 200,000 tonnes of solid waste per year. There
are two possible ways of disposing of the waste, first to dump the untreated
waste in a landfill and second to incinerate it. Let x1 be the mass of untreated
waste and x2 be the mass of untreated waste to be incinerated; use units of
1000 tonnes. The landfill will accept at most 160,000 tonnes of solid waste
comprising untreated material, and 24,000 tonnes of ash from the incinerator.
An 80% reduction of mass results during incineration, leaving a mass equal to
20% of the untreated waste. The incinerator produces 2500 units of energy per
tonne of waste, which can be sold for $0.08 per unit. The table summarizes the
costs associated with shipping and treatment of the waste.

Cost per tonne

Landfill–untreated material $100
1.
Landfill–incinerator ash $160
Incinerator $60

(a) Formulate two objective functions, one that minimizes all costs (including
sale of energy) and another that minimizes the mass of material dumped
in the landfill. Solve the problem for both objective functions.
(b) Reformulate the problem considering other objectives in waste disposal,
for example pollution due to incineration, and recycling of fibre, plastics
and metals.

Solution
Let x1 = mass of untreated waste (1000 tonnes), and x2 = mass of treated waste
(1000 tonnes)

x1 + x2 = 200

14

@seismicisolation
@seismicisolation
 subject to x1 ≤ 160
(0.20)(x2 ) ≤ 24, or x2 ≤ 120
x1 , x2 ≥ 0

Min cost (in $1000’s):

Min [100x1 + 200(0.20x2 ) + 60(x2 ) − (0.08)(2500)x2 ]

i.e. Min (100x1 − 100x2 = z1 )

Min mass:
Min (x1 + 0.20x2 = z2 )

A is the optimal solution for either objective function (objectives are com-
patible) 
x∗1 = 80 tonnes 


x∗2 = 120 tonnes
∗ × 1000
z1 = $8000 − 12000 = −$4000 (profit)  

z2∗ = 80 − ( 120
5 ) = 56 tonnes
Simplex
x1 + x2 − x3 = 200
x1 + x4 = 160

15

@seismicisolation
@seismicisolation
 x2 + x5 = 120

z1 = 100x1 − 100x2
ω = x6 = 200 − x1 − x2 + x3 (Inf easibility f orm)

b x1 x2 x3 x4 x5 x6 Basic Var.
200 1 1 -1 1 x6
160 1 1 x4
120 1 1 x5
z1 − 0 50 -50
ω − 200 -1 -1 1
40 0 1 -1 -1 1 x6
160 1 1 x1
120 1 1 x5
z1 − 8000 -50 -50
ω − 140 - -1 1 1
40 0 1 -1 -1 1 x2
160 1 1 x1
80 0 1 1 1 -1 x5
z1 − 6000 -50 -50 50
ω−0

etc.

Exercise 5.16
Two software companies–A and B–are planning to market a product. Mar-
keting strategists have investigated four strategies open to B, and five open to
A. They have prepared the following table, showing the payoffs for the various
{σ, τ } pairs. The payoffs are from A’s perspective, with 0 = failure of the prod-
uct, 5 = satisfactory, 10 = maximum success. Use a maximin (game theory)
approach to investigate whether there is a saddle point. Is there an optimal
pure strategy or is a mixed strategy appropriate? Under what conditions would
you opt for a different approach from the maximin?

Company B
τ1 τ2 τ3 τ4
σ1 9 4 3 2
1. Company A σ2 3 1 3 8
σ3 7 6 10 7
σ4 4 4 5 3
σ5 6 3 4 4

16

@seismicisolation
@seismicisolation
Solution
FIRM 2 Row
PLAN τ1 τ2 τ3 τ4 minimum
σ1 9 4 3 2 2
σ2 3 1 3 8 1
FIRM 1 σ3 7 6 10 7 6 (maximin)
σ4 4 4 5 3 3
σ5 6 3 4 4 3
Column Maximum 9 6 (minimax) 10 8

The saddle point is at {σ 3 , τ 2 } as noted. Value of game = 6. Pure strategies

σ 3 for firm 1, τ 2 for firm 2 are optimal.
I might opt for another strategy if I thought that the other player was not
out to "get" me; if I thought that there was a reasonable probability that he
or she would not opt for the maximin philosophy. The possibility of obtaining
payoff of 10 ("bonanza") is attractive.

Exercise 5.17
Solve the following two-person zero-sum game using mixed strategies if neces-
sary. Plot the expected values for each player to determine the gain floor and
the loss ceiling as in Figure 5.14.
τ1 τ2
σ1 1 −2
σ 2 −1 2

Solution
B
(q) (1-q) Row
A τ1 τ2 min
(p) σ1 1 -2 -2
(1-p) σ2 -1 2 -1 (Maximin)
Column Max. 1 (Minimax) 2
Maximin and Minimax donot coincide - game has no saddle point. Use
mixed strategy

Player A
If B plays τ 1 , hV i = p − (1 − p) = 2p − 1
If B plays τ 2 , hV i = −2p + 2(1 − p) = 2 − 4p

17

@seismicisolation
@seismicisolation
 © ª
Strategy for A is σ 1 , 12 , σ 2 , 12
Player B
If A plays σ1 , hV i = q − 2(1 − q) = 2q − 2
If A plays σ2 , hV i = −q + 2(1 − p) = 2 − 3q

18

@seismicisolation
@seismicisolation
 3q − 2 = 2 − 3q, or
2
p =
3
© ª
Strategy for B is τ 1 , 23 , τ 2 , 13
Expected value = 0 (Both Players) = value of game

Exercise 5.18
Solve the following two-person zero-sum game using mixed strategies if neces-
sary. Player A plays the rows and B plays the columns. We now denote the pay-
offs (from B to A, as per convention) as v. Plot the values in a two-dimensional
graph as pairs (vi |σ 1 ) versus (vi |σ 2 ) for all τ i , i = 1, . . . , 5. The points show the
payoffs to A for the pure strategies open to B. Describe possible mixed strategies
open to B. Interpret player B’s optimal strategy.
τ1 τ2 τ3 τ4 τ5
σ1 1 2 2 2 4
σ2 3 1 4 3 3

Solution
There is no saddle point; use a mixed strategy solution.
Player A
 
1 3
 2 1 ½ ¾
  p1
 2 4  ≥ v;
  p2
 2 3 
4 3
p 1 + p2 = 1
pi
put xi = to obtain LP
v
Min (x1 + x2 )
Solution through Maple: x1 = 25 ; x2 = 15 ; v = 53 ; =⇒ p1 = 23 ; p2 = 1
3 (optimal
strategy for A)
Player B
 

 q1 

¸ 
·  q2 
 
1 2 2 2 4
q3 ≤ t;
3 1 4 3 3  


 q4 

 
q5
X
qi = 1

19

@seismicisolation
@seismicisolation
 qi
Put yi = to obtain LP
t
³X ´
M ax yi

Solution x1 = 15 ; x2 = 25 ; x3 = x4 = x5 = 0; t = 53 ; =⇒ q1 = 13 ; q2 = 23 ; q3 = q4 =
q5 = 0 (optimal strategy for B)
The strategies are plotted below. The individual points (including the
interior) of the convex set represent mixed strategies, while the vertices represen
the pure strategies (τ 4 is in the interior). See Luce and Raiffa (1957).
Player B uses the minimax principle. B wishes to find a value v ∗ which
represents the smallest amount B has to pay out, no matter what A plays.
The minimum point is shown on the figure with values {v ∗ , v ∗ }; where v ∗ = 53 .
The position of this point confirms the previously obtained solution. (Luce and
Raiffa).

Exercise 5.19
For the game of rock-paper-scissors–mentioned in the introductory section on
Conflict–construct the payoff matrix assuming payments of one unit. Is there
a saddle point? Solve the game.

Solution
Rock-paper-scissors. As er convention, B pays A

20

@seismicisolation
@seismicisolation
 B
τ 1 Rock τ 2 Paper τ 3 Scissors
A σ 1 Rock 0 -1 1
σ 2 Paper 1 0 -1
σ3 ©
Scissors ª -1 1 0
Solution 13 , 13 , 13 in each case.

Exercise 5.20
The following represents a payoff matrix, with A representing the row, and B
the column, player. Payoffs are, according to convention, from B to A. Imagine
that you are player B; the values in the table can be interpreted as (negative)
utilities for you.

(a) Consider the situation where A represents a non-antagonistic opponent,

as in “design as a game against nature”. Find B’s optimal strategy if it
is estimated that Pr(σ 1 ) = 0.2 and Pr(σ 2 ) = 0.8. Plot the values of the
various strategies open to B in a two-dimensional graph with axes “value
given σ 1 ” and “value given σ2 ”. Find the ranges of probability of σ 1 and
σ2 for which various strategies open to B are optimal, and identify those
that are never optimal.
(b) Analyze the problem as a two-person zero-sum game, considering player
B’s game. Set up, but do not solve the linear program.
(c) Again consider the problem as a game. Plot the expected values of
player B’s strategies against Pr(σ1 ) = p. Determine from this the optimal
(mixed) strategy for A.

τ1 τ2 τ3 τ4
σ1 2 3 1 5
σ2 4 1 6 0

Solution
(a)

τ 1 : hU i = (0.2)(−2) + (0.8)(−4) = −3.6

τ 2 : hU i = (0.2)(−3) + (0.8)(−1) = −1.4
τ 3 : hU i = (0.2)(−1) + (0.8)(−6) = −5.0
τ 4 : hU i = (0.2)(−5) + (0.8)(0) = −1.0
τ 4 is optimal.

21

@seismicisolation
@seismicisolation
 22

@seismicisolation
@seismicisolation
 Ranges
Pr(σ 1 ) Pr(σ2 ) Optimal Strategy
1 − 57 0 − 27 τ3
5 1 2 2
7 − 3 7 − 3 τ2
1 2
3 − 0 3 − 1 τ4
τ 1 is never optimal.
(b)
Max w = y1 + y2 + y3 + y4
Subject to 2y1 + 3y2 + y3 + 5y4 ≤ 1
4y1 + y2 + 6y3 ≤ 1
all yj ≥ 0
qj
where yj = , j = 1, ...., 4
t
(c)
τ1 : 2p + 4(1 − p) = 4 − 2p
τ2 : 3p + (1 − p) = 1 + 2p
τ3 : p + 6(1 − p) = 6 − 5p
τ4 : 5p = 5p

23

@seismicisolation
@seismicisolation
Gain floor:

1 + 2p = 6 − 5p
5
7p = 5; p =
7
©5 2
ª 17
Strategy 7, 7 ;Value = 7

24

@seismicisolation
@seismicisolation
Solutions for Chapter 6
Exercise 6.1
Rare events, those events with a low probability p, are said to cause greater sur-
prise when they occur and therefore contain greater information. The functions
“surprisal” and “information” su are defined as
1
su ∝ ln = − ln p;
p

see Tribus (1969) and Lathi (1998). Discuss this concept in terms of events such
as

• aircraft crashes
• suicidal aircraft crashes (Twin Towers)
• frequent letters in the English language, e.g. t, e, a as against rare ones
x, q, z (use of code words)
• events that cause surprise generally.

Discuss entropy as expected information

X
hSu i = pi (su )i

and the relevance to source encoding (Lathi, 1998).

Suggestion
See the texts quoted: Tribus (1969) and Lathi (1998).

Exercise 6.2
Consider an event E with probability p. Plot the entropy associated with this
event (and its complement) against p. What would the maximum-entropy as-
signment of probability be, in the absence of constraints other than coherence?
Explain and contrast why the maximum entropy solution becomes very
likely, with a “sharp” distribution while the plot of H is quite smooth.

@seismicisolation
@seismicisolation
Solution

p = 1 − p = 0.5, H = 0.693
p = 0.75, 1 − p = 0.25, H = 0.562
p = 1.0, 1 − p = 0, H = 0

H = −p ln p − (1 − p) ln(1 − p)
dH
= − ln p − 1 − [− ln(1 − p) − 1]
dp
p 1
= − ln = 0if p =
1−p 2
The "most likely" solution corresponds to very large N in:
w
= exp [N (S − T )]
w0
(Equation 6.85), where S = Hmax and T = Hnot max ; see p.298 and Figure 6.6.

Exercise 6.3
Solve Example 6.4, obtaining the probabilities in terms of the mean µ. Plot the
three pi against µ, the latter ranging from 1 to 3.

@seismicisolation
@seismicisolation
Solution
As in text:
µ X3 ¶
Max H = − pi ln pi
i=1

subject to
X3
pi = 1
i=1
and X3
xpi = hXi = µ,
i=1
which becomes
0 · p1 + 1 · p2 + 2 · p3 = hXi i = µ
or
p2 + 2p3 = µ.
In the last three equations, µ stands for the mean energy level.
The Lagrangian is
X3 µX ¶
¡ ¢ 3
L p1 , p2 , p3 ; λ00 , λ1 = − pi ln pi −λ00 pi − 1 −λ1 (p2 + 2p3 − µ) .
i=1 i=1

Differentiating the Lagrangian:

∂L 0 
∂p1 = (− ln p1 − 1) − λ0 = 0 

∂L 0 

∂p2 = (− ln p2 − 1) − λ0 − λ1 = 0


∂L 0
∂p3 = (− ln p2 − 1) − λ0 − 2λ1 = 0
∂L 

∂λ00 = − (p1 + p2 + p3 − 1) = 0


∂L


∂λ1 = − (p2 + 2p3 − µ) = 0.

As before, is convenient to write

−λ0 = −λ00 − 1,

in which we have borrowed (−1) from the derivative of H. >From the first three
derivatives of the Lagrangian, we can write

p1 = exp (−λ0 ) , p2 = exp (−λ0 − λ1 ) , p3 = exp (−λ0 − 2λ1 ) .

The first of the two remaining conditions, which is the constraint giving the
normalization of probabilities, reduces to

exp (−λ0 ) [1 + exp (−λ1 ) + exp (−2λ1 )] = 1

or
exp λ0 = 1 + exp (−λ1 ) + exp (−2λ1 ) .

@seismicisolation
@seismicisolation
(the partition function Z).
X
Z (λ1 ) = exp λ0 = exp (−λj x) , j = 0, 1, 2.
j

The second of the constraints yields the following equation:

exp (−λ0 ) [exp (−λ1 ) + 2 exp (−2λ1 )] = µ.
Usual to leave above in parametric form; in the present case, it is easy to solve
for the p0i s (Jaynes, p.181). Find

¡ ¢
e−λ1 + 2(e−λ1 )2 = µeλ0 = µ 1 + e−λ1 + e−2λ1
(1 − µ)e−λ1 + (2 − µ)e−2λ1 = µ
(2 − µ)(e−λ1 )2 + (1 − µ)e−λ1 − µ = 0;
let y = e−λ1 ; then
(2 − µ)y 2 + (1 − µ)y − µ = 0.
This is a quadratic equation that can be solved for y, and hence the probabilities
can be found.
√ p
−b ± b2 − 4ac −(1 − µ) ± 1 + µ2 − 2µ + 4µ(2 − µ)
y= =
2a 2(2 − µ)
p1 = e−λ0 ; p2 = e−λ0 −λ1 ; p3 = e−λ0 −2λ1

@seismicisolation
@seismicisolation
Exercise 6.4
Complete the proof of equations (6.61) to (6.64). Also show that
∂S
λj = ,
∂ν j
where S represents Hmax . Form the matrix of second order partial derivatives
of λ0 with respect to λ1 , . . . , λj , . . . , λm (the Hessian matrix). Show that this is
the variance-covariance matrix. Since this is positive definite, what conclusions
can be reached?

Solution
See Jaynes (2003, p.188) and Kapur and Kesavan (1992, p. 45-46)
X
vj = gj (xi )pi
i
P
exp − [λj gj ]
pi =
z
1X h X i
vj = gj (xi ) exp − λj gj (xi )
z
X h X i
z= exp − λj gj (xi )
i j

∂ ln z ∂λ0 1X
= = exp [] gj (xi ) = vj
∂λj ∂λj z i

At Hmax n X o
pi = exp −λ0 − [λj gj ]
j
n X o
ln pi = −λ0 − [λj gj ]
j

X X n X o X X³ X ´
− pi ln pi = − exp { } · −λ0 − [. . .] = ( pi )λ0 + pi [λj gj ]
X
= λ0 + λj · vj

Differentiate exp λ0 ∂λ
∂λj from above with respect to λj
0

µ ¶2
∂ 2 λ0 ∂λ0 ­ ® ∂ 2 λ0
+ = gj2 (Xi ) =⇒ = σ 2Gj
∂λ2j ∂λj ∂λ2j

Differentiating exp λ0 ∂λ
∂λj with respect to λk
0

∂ 2 λ0
= σ Gj Gk
∂λj ∂λk

@seismicisolation
@seismicisolation
(covariance - see Kapur (1992)).
Expression for S = Hmax above; differentiate with respect to vj
∂S
= λj
∂vj
Hessian matrix is

 ∂ 2 λ0 ∂ 2 λ0

∂λ21 ∂λ1 ∂λ2 ...
 ∂ 2 λ0 ∂ 2 λ0 
=  ... 
∂λ1 ∂λ2 ∂λ22
... ... ...
 
σ 2G1 σ G1 G2 ...
=  σ G2 G1 σ 2G2 ... 
... ... ...
This is positive definite so that the Hessian is positive definite −→ λ0 is a
convex function of the λj .

Exercise 6.5
Show that the maximum-entropy distribution subject to no constraints other
than coherence is a uniform distribution.

Solution
n possibilities; pi = exp λ0 ; λ0 = ln n
See Tribus (1969, p.128).

Exercise 6.6
Find maximum-entropy solutions for the gamma and beta distributions.

Solution
Gamma: P P
R.Q.X.: fix mean µ = hXi ; pi xi = µ; fix ln X also pi ln Xi = τ , a
constant
(Kapur suggests fixing the geometric mean hln Xi)
p
n
X1 X2 ....Xn = τ1
X1 X2 ..... = τ n1
ln X1 + ln X2 + · · · = const.
Beta: Fix ln X and ln(1 − X)
See Tribus (1961, p.144, 150) and Kapur (1992).

@seismicisolation
@seismicisolation
Exercise 6.7
Show that the maximum-entropy distribution in the case where the first and sec-
ond moments (means, variances and covariances) of a set of random quantities
X1 , X2 , . . . , Xi , . . . , Xn are fixed is the multivariate normal distribution. Note
that here we use the subscript i here to denote individual random quantities
and not particular values of a single random quantity as in (6.55).

Solution
(mean)
hXi i = µi
(variance) ­ ®
σ2Xi = σ2i = (Xi − µi )2
(covariance)
­ ¡ ¢®
σ Xi ,Xj = ρi,j σi σ j = (Xi − ui ) Xj − µj

Maximize
X X
− ··· fX1 ,X2 ,... (x1 , x2 , . . . , xi, . . . xj , . . .) ln fX1 ,X2 ,... (x1 , x2 , . . . , xi, . . . xj , . . .)

subject to constraints above. From Lagrangian, differentiate and one finds an

exponential form with all terms involving xi , xj such as x2i , x2j , xi xj appearing
as sum in exponent.
· hXi−1 ¸
1 1 t
fX (x) = n/2 P 12
exp − (x − µ) (x − µ)
(2π) | | 2
where
   

 x1 
 
 µ1 
   
: :
x= and µ = and

 
: 
 : 
   
 xn µn 
2
σ1 ρ12 σ 1 σ 2 · · ·
P  ρ σ σ σ 22 ··· 
=
21 2 1 

··· ··· ···
··· ··· ···
See Kapur (1992, p. 69).

Exercise 6.8
Show the equivalence (6.94) directly starting with equations (6.77) and (6.78).
Confirm in examples that the occupancy corresponds to the information
given in section 2.6.3 and Table 2.3.

@seismicisolation
@seismicisolation
Solution
Equation (6.78): X
Z (µ, T ) = exp [(µ i − ei ) /kT ] ,
i

(ei = · ) ;
i− summed over all , and for each , all the -particle states of the system
∞
" #
X X X (µ − i )
Z= exp k A
kT
→ k
|k {z }

Second summation: all combinations of occup. #’s k that add up to .; see

next figure for examples with = 1, 2, 3, . . .
Sum rather over the simple-particle occup. #’s:
· ¸ · ¸ · ¸
X (µ − 1 ) 1 X (µ − 2 ) 2 X (µ − 3 ) 3
Z = exp · exp · exp ···
kT kT kT
1 2 3
YX · ¸
(µ − i ) i
= exp
i
kT
i

P Y
···
(Note: e in A = e··· )
But · ¸
X (µ − i ) i 1
exp =
kT 1 − exp · · ·
i

1
Use 1 + x + x2 + ... = 1 −
x
· ¸−1
(µ − 1 )
ξ i = 1 − exp
kT
Y
Z= ξi
i

Hence result li = 0, 1, 2, 3... for bosons

Exercise 6.9
Q
Again let Z = i ξi. Obtain the function ξ i (6.99) for the Fermi-Dirac gas;
result is
ξ i = 1 + exp [(µ − i ) /kT ]
since not more than one particle can go to any energy state.

@seismicisolation
@seismicisolation
Solution
Fermions,
P i = 0, 1; two terms only
= 1 + exp ... = ξ i
li
(just two terms above)

2
3 4 5
(2+4-1)= 5! =5
simple particle states 4 4!*1!

0 1 2 3
0

3
3 3

3 energy
levels 3
3
2 3
2 3

2 3
(n+r-1)
r 3 3

3 6 10

Exercise 6.10
Derive the Bose-Einstein distribution using Jaynes’ widget method. It concerns
a salesperson who is in charge of a factory producing differently painted widgets.

@seismicisolation
@seismicisolation
The question is which colour should be produced each day. Constraints are
concerned with the average number of widgets ordered each day and the average
size of individual orders for the different colours. Details can be found in Jaynes
(1963a).

Solution
See Jaynes (1963a)

Exercise 6.11
Show that the mean number of particles for the partition function of (6.94) is
given by
∂ ln Z
hLi = kT , (1)
∂µ
and in state i
∂ξ
hLi i = kT i , (2)
∂µ
and that this results in
X exp [(µ − i ) /kT ]
hLi =
i 1 − exp [(µ − i ) /kT ]
X 1
= . (3)
i {exp [( i − µ) /kT ]} − 1

and
1
hLi i = . (4)
{exp [( i − µ) /kT ]} − 1
−1
Use the result ξ i = {1 − exp [(µ − i ) /kT ]} .

Solution
X
Z(λ1 , λ2 ) = exp λ0 = exp(−λ1 ei − λ2 i )
i
Now we know (exercise 6.4) that
∂ ln Z
vj = − ;
∂λj

∂ ln Z
hLi = − ;
∂λ2
µ ∂ ln Z ∂µ
λ2 = − =− ·
kT ∂µ ∂λ2
∂ ln Z
= kT
∂µ

10

@seismicisolation
@seismicisolation
 X X· ½ (µ − i
¾¸
ln Z = + ln ξ i =− 1 − exp X
i i
kT
µX ¶ X
∂ ln Z ∂ ln ξ i
kT = kT = hL1 i + hL2 i + .... = hLi i
∂µ µ
µ : individual contributions of the mean of each simple particle state.
Hence,
∂ ln ξ i
hLi i = kT
∂µ
Expressions for hLi and hLi i follow from differentiation of X above.

Exercise 6.12, 6.13

Derive the Fermi-Dirac distribution using an urn model.
Derive the Maxwell-Boltzmann distribution using an urn model.

Solution
Follow procedure in chapter (other facets)
See for example L.A. Girifalco, "Statistical Physics of Materials" Wiley
(1973)

Exercise 6.15
Derive from first principles the expression for the rate of a spontaneous reaction
using the theory of activation energy,
³ ´
A
ν exp −
kT
where A is the energy barrier and ν the number of runs per second at the
barrier.

Solution
The maximum-entropy distribution of energy amongst particles has been derived
in the Chapter. The classical Maxwell-Boltxmann distribution will be used
(Equations 6.44 to 6.48):

exp (−ei /kT )

pi = X .
exp (−ej /kT )
j

In activation energy analysis, it is supposed that a particle that acquires an

energy equal to or greater than the activation energy eA is able to surmount
an energy barrier and move into a position of lower potential energy. These

11

@seismicisolation
@seismicisolation
movements are observed macroscopically, e.g. in deformation of a specimen, or
some other application. Let E be the event of a particle surmounting the energy
barrier eA in "run" at the barrier. Noting that ei = ihυ, where h = Planck’s
constant, ν = frequency, and i = 1, 2, 3, . . . .
P
ei >eA exp (−ei /kT )
Pr (E) = P
exp (−ei /kT )
P i
exp [− (eA + ihν) /kT ]
= P
i

i exp [− (ihν) /kT ]

P
exp [− (eA ) /kT ] i exp [− (ihν) /kT ]
= P
i exp [− (ihν) /kT ]
= exp [− (eA ) /kT ] .

For ν runs per second at the barrier, the probability of crossing the barrier is

ν exp [− (eA ) /kT ]

.

12

@seismicisolation
@seismicisolation
Solutions for Chapter 7
Exercise 7.1
Confirm the use of (7.10) in an example of your choice. For example, one might
transform the number of successes X in the hypergeometric distribution by
g (X) = X k
where k is a positive constant.

Solving the problem

The method consists of moving the probability masses according to equation
given, from points x to points xk .

Exercise 7.2
Transform the uniform distribution on [0, 1] , (fX (x) = 1 for 0 ≤ x ≤ 1, = 0
otherwise) using the function
µ ¶
1
g (X) = ln .
1−X
How may this be used with the probability integral transform in Monte Carlo
simulation?

Solution
· ¸
1
y = g(x) = ln = − ln(1 − x)
1−x
x = 1 − e−g(x) = 1 − e−y = g −1 (y)
¯ ¯ · ¸
¯ dx ¯ £ ¤ dg −1 (y)
fY (y) = fX (x) ¯¯ ¯¯ = fX g −1 (y)
dy dy
= 1 · e−y for y ≥ 0.
Used with uniform [0, 1] to generate independent realizations of the expo-
nential distribution. If X is uniform [0, 1] then y = − ln(1 − x).

Exercise 7.3
Obtain the standardized form of the Weibull distribution in terms of the para-
meter k.
µ ¶k−1 " µ ¶k #
k x−ξ x−ξ
fX (x) = exp − for x ≥ ξ.
α α α

@seismicisolation
@seismicisolation
Solution
Make the transformation
x−ξ
y = g(x) = ,
α
giving the pdf :
k
fY (y) = ky k−1 e−y for y ≥ 0

Exercise 7.4
Find the standardized Rayleigh distribution in terms of µ.

2 (x − a) h i
2
fX (x) = exp − (x − a) /µ .
µ

Solution
Make the transformation
y =x−a
giving
2y −y 2
fY (y) = exp( )
µ µ

Exercise 7.5
Find the location and scale parameters for the standardization of the uniform
distribution
1
fX (x) = for a ≤ x ≤ b
b−a
into the distribution
fY (y) = 1 for 0 ≤ y ≤ 1 .

Solution
x−a
y=
b−a
(a is the location parameter, and (b − a) represents the scale).
b−a
fy (y) = =1
b−a

@seismicisolation
@seismicisolation
Exercise 7.6
If X is normally distributed, consider the transformation

X = ln Y.

Then Y has a lognormal distribution.

(a) Write down the distribution of Y.

(b) Write down the mean and variance of the distribution.
(c) Show that 1/Y also has a lognormal distribution. Write down the mean
and variance of this distribution.
(d) If Y1 and Y2 are two independent random quantities with lognormal dis-
tributions, show that YP = Y1 Y2 also has a lognormal distribution. De-
termine the parameters of YP in terms of the parameters of Y1 and Y2 .
(e) Discuss the lognormal distribution as a “model of products” in the light
of the central limit theorem as the “model of sums”.

Solution
(a)
dx 1
x = ln y = g −1 (y) −→ =
dy Y
y = ex = g(x)
¯ ¯
£ −1 ¤ ¯ dg −1 (y) ¯
fY (y) = fx g (y) ¯¯ ¯
dy ¯
1 1 x−µ 2
fX (x) = √ e− 2 ( σ )
2πσ
· ¸
1 −1 1 2
fY (y) = √ y exp − 2 (ln y − µ) , for y>0
σ π 2σ
Note that
µ = µX , σ = σ X
¡ ¢
(b) Mean and Variance. Let a = exp µ, b = exp σ 2
√
µY = a b
σ 2Y = a2 b (b − 1)

@seismicisolation
@seismicisolation
 (c) Let T = 1/Y, then from Equation (7.11)
¯ ¯
£ ¤ ¯ dg −1 (t) ¯
fT (t) = fY g −1 (t) ¯¯ ¯
dt ¯
£ ¤ 1
= fY t−1 · 2
t · ¸
1 −1 1
= = √ t exp − 2 (− ln t − µ)2 ,
σ π 2σ

Which is also lognormal??(Ochi p.117)

(d)
(e)

Exercise 7.7
Exponential growth. Let the size Z of a population be given by

Z = c exp (Xt)

for given time t; c is a constant. If X is random with the density

2
fX (x) = d (1 − x) for 0 < x < 1,

and 0 otherwise, find fZ (z) .

Solution
First,
Z
d£ ¤1
fX (x)dx = 1 =⇒ − (1 − x)3 0 = 1;
3
d = 3

z = gX (x) = cext

1 z dx 1
x= ln( ); =
t c dy tz

· ¸2
1 z 1
fZ (z) = 3 1 − ln( ) , for c < z < cet
t c tz
= 0, otherwise

@seismicisolation
@seismicisolation
Exercise 7.8
Given that fX1 ,X2 (x1 , x2 ) = 1 for 0 < x1 < 1 and 0 < x2 < 1, find fY1 ,Y2 (y1 , y2 )
for the transformation
½ ¾ · ¸½ ¾ ½ ¾
y1 1 −1 x1 2
= + .
y2 2 1 x2 1
Why is this transformation termed affine?

Solution

y1 = x1 − x2 + 2
y2 = 2x1 + x2 + 1

1
x1 = (y1 + y2 ) − 1
3
1
x2 = (−2y1 + y2 ) + 1
3
¯ ¯ ¯ ¯
¯ ∂x1 ∂x1 ¯ ¯ 1 1 ¯ 1 2
¯
¯ ∂y1 ∂y2 ¯ ¯
¯=¯ 3 3 ¯= + = 1
¯ ∂x2 ∂x2
¯ −2 −2 ¯ 9 9 3
∂y1 ∂y2 3 3

Therefore
1 1
fY1 Y2 (y1 y2 ) = 1 · =
3 3

@seismicisolation
@seismicisolation
(Uniform on parallelogram).
Affine because parallelism is preserved (see figure above)

Exercise 7.9
Consider Y = X1 + X2 . Use equations (7.76) and (7.77) to obtain fY (y) if X1
and X2 are iid with exponential distributions exp (−x) for x ≥ 0, = 0 otherwise.

Solution
Z y
−(y−x1 )
fY (y) = e−x1 e dx1
0
Z y
= e−y dx1 = ye−y , for y ≥ 0.
0

Exercise 7.10
Consider two iid random quantities X1 and X2 that are uniformly distributed
on (0, 1) . Use the transformations
p
y1 = −2 ln x1 cos (2πx2 ) and
p
y2 = −2 ln x1 sin (2πx2 )

to show that Y1 and Y2 are independent random quantities with standard normal
distributions.
This algorithm can be used to generate normal random quantities from the
uniform distribution (Chapter 11).

Solution

−y12
= ln x1 cos2 (2πx2 )
2
−y22
= ln x1 sin2 (2πx2 )
2
Sum of above:
µ ¶
y12 y22
ln x1 = − + ;
2 2
1 2 2
x1 = e− 2 (y1 +y2 ) .

Division of above, one by the other:

1 y2
x2 = arctan
2π y1

@seismicisolation
@seismicisolation
 · 2
¸· 2
¸
1 y1 1 y2
|J|+ = + √ e− 2 √ e− 2
2π 2π
and the result follows.

Exercise 7.11
Two independent random quantities Z1 and Z2 are modelled using the standard
normal distribution. The quantities R and Θ are polar coordinates related to
Z1 and Z2 as follows for 0 ≤ R < ∞ and 0 ≤ Θ ≤ 2π.
¡ 2 ¢1/2
R Z1 + Z22
= and
µ ¶
Z2
Θ = arctan .
Z1

Determine fR (r) and fΘ (θ) .

Solution
¾
z1 = r cos θ
z2 = r sin θ
¯ ¯
¯ cos θ −r sin θ ¯¯
|J|+ = ¯¯ =r
sin θ r cos θ ¯+
r2
fR (r) = re− 2 chi-square (2)

1
fΘ (θ) =
2π

Exercise 7.12
Derive equation (7.109).

For solution
See Feller (1968, p. 502, vol. II)

Exercise 7.13
Gamma and beta functions. It is a good idea to become familiar with the gamma
function Z ∞
Γ (s) = us−1 exp (−u) du.
0

@seismicisolation
@seismicisolation
Show that
Γ (s) = (s − 1) Γ (s − 1)
√
and that for positive integers s, Γ (s) = (s − 1)!. Also show that Γ (1/2) = π,
and that the beta function
Z 1
m−r−1
B (r, m − r) = ar−1 (1 − a) da
0

used to normalize the beta distribution is given by

Γ (s) Γ (t − s)
B (s, t − s) = .
Γ (t)

Solution
The gamma function is given by
Z ∞
Γ (s) = us−1 exp (−u) du, (1)
0

for s > 0. check. This is used in the gamma distribution

³ ´
xα−1 exp − βx
fX (x) = , for x ≥ 0, = 0 otherwise. (2)
β α Γ (α)

The gamma function serves in the normalization to unity of this equation. The
parameter α is the shape parameter, and β is the scale parameter.
Integration by parts will give the recurrence formula. Thus
Z ∞
£ ¤∞
Γ (s) = −us−1 exp (−u) 0 + (s − 1) us−2 exp (−u) du
0
= (s − 1) Γ (s − 1) . (3)

Note that for positive integers s, Γ (s) = (s − 1)!. This can be seen using the
preceding equation, and since
Z ∞
Γ (1) = exp (−u) du = 1.
0
R∞
Now make the transformation u = v 2 in the equation Γ (s) = 0
us−1 exp (−u) du.
Then Z ∞
¡ ¢
Γ (s) = 2 v 2s−1 exp −v 2 dv (4)
0

is an equivalent form. We can use this to evaluate, for example, Γ (1/2) . Thus
Z ∞
¡ ¢
Γ (1/2) = 2 exp −v 2 dv;
0

@seismicisolation
@seismicisolation
we can also write Z ∞ ¡ ¢
Γ (1/2) = 2 exp −w2 dw,
0
and multiplying these last two equations find:
·Z ∞ ¸ ·Z ∞ ¸
2 ¡ ¢ ¡ ¢
[Γ (1/2)] = 4 exp −v 2 dv exp −w2 dw
0 0
Z ∞
£ ¡ 2 2
¢¤
= 4 exp − v + w dvdw.
0

Now use polar coordinates {r, θ} :

Z π/2 Z ∞ ¡ ¢
[Γ (1/2)]2 = 4 r exp −r2 drdθ
0 0
= π.
√
Therefore Γ (1/2) = π.
Assume that A is represented by the beta distribution with parameters:
m−r−1
ar−1 (1 − a)
fA (a) = , 0 ≤ a ≤ 1. (5)
B (r, m − r)

In order that the probability density normalises,

Z 1
B (r, m − r) = ar−1 (1 − a)m−r−1 da. (6)
0

Here the Beta function is given by

Γ (s) Γ (t − s)
B (s, t − s) = , (7)
Γ (t)

where s, t > 0. To show this, first we write this equation in the following nota-
tion:
Γ (x) Γ(y)
B(x, y) = .
Γ(x + y)
R1
Now, from B (r, m − r) = 0 ar−1 (1 − a)m−r−1 da,
Z 1
B(x, y) = ax−1 (1 − a)y−1 da
0
Z π/2
= 2 sin2x−1 θ cos2y−1 θdθ, (8)
0
R∞ ¡ ¢
which is found by the substitution a = sin2 θ. From equation Γ (s) = 2 0
v 2s−1 exp −v 2 dv,
Z ∞
¡ ¢
Γ (x) = 2 v 2x−1 exp −v 2 dv,
0

@seismicisolation
@seismicisolation
and Z ∞ ¡ ¢
Γ(y) = 2 w2y−1 exp −w2 dw.
0
Multiplying these two integrals:
Z ∞Z ∞
£ ¡ ¢¤
Γ (x) Γ(y) = 4 v 2x−1 w2y−1 exp − v 2 + w2 dvdw.
0 0

We may again introduce polar coordinates, and we find that

à Z !µ Z ¶
π/2 ∞ ¡ 2¢
2x−1 2y−1 2(x+y)−1
Γ (x) Γ(y) = 2 sin θ cos θ dθ 2 r exp −r dr ,
0 0
R∞ ¡ ¢
since the integrals can be separated. From equations Γ (s) = 2 0 v 2s−1 exp −v 2 dv
R π/2
and B (x, y) = 2 0 sin2x−1 θ cos2y−1 θdθ above, we can see that
Γ (x) Γ(y) = B(x, y)Γ(x + y),
or
Γ (x) Γ(y)
B(x, y) = . (9)
Γ(x + y)

Exercise 7.14
Human error. In a study of the effect of human error in design, Nessim used
two independent gamma distributions with the same scale factor (β in equation
(7.210) in the book) to model the distributions of undetected and detected
errors (M and N respectively) (see Nessim and Jordaan (1985)). Show that the
probability of detecting an error,
N
P =
M +N
is given by the beta distribution.
Note that these distributions of error rates are consistent with the idea of
gamma distributions modelling the uncertainty in rates and the errors them-
selves occurring as a Poisson processes. In this context the gamma distributions
are conjugate prior distributions for the rates of the Poisson process. This situ-
ation is modelled in Chapter 8. There are other applications using this idea; for
example in arctic shipping, the two processes could correspond to successfully
detected and avoided ice features on the one hand, and collisions with ice, on
the other; see Jordaan et al. (1987).

Solution
Let
P2 = M and
M
P1 = (≡ P in question)
M +N

10

@seismicisolation
@seismicisolation
 £ ¤
fP1 P2 = fM,N g −1 (p1 , p2 ) |J|+

M = P2
P1 P2
N =
1 − P1
· ∂M ∂M ¸ · ¸
∂P1 ∂P2 0 1
J= ∂N ∂N = p2 p1
∂P1 ∂P2 (1−p1 )2 1−p1

p2
|J|+ =
(1 − p1 )2
gamma for M −→ α1 , β
gamma for N −→ α2 , β

³ ´α1 −1 p1 p2 1
p2
p2α2 −1 e− β
p1 p2
1−p1 e− 1−p1 · β p2
fP1 P2 = · ·
β α2 Γ(α2 ) β 1 Γ(α2 ) (1 − p1 )2
³ ´
p1α1 −1 1
−β
p2
= const. × p2α1 +α2 −1 · ·e 1−p1

(1 − p1 )α1 +1
Integrate out p2 :
Z ∞
1
− ·p
p2α1 +α2 −1 e β(1−p1 ) 2 dp2 = Γ (α1 + α2 ) [β(1 − p1 )]α1 +α2
0
using integral for gamma distribution.

fP1 P2 (p1 , p2 ) ∝ p1α1 −1 (1 − p1 )α2 −1

Final result:
p1α1 −1 (1 − p1 )α2 −1
fP1 (p1 ) = .
B(α1 , α2 )
Γ(α1 )Γ(α2 )
B is beta function = Γ(α1 +α2 ) .

Exercise 7.15
For the binomial distribution of equation (3.3),
µ ¶
n x
pX (x) = p (1 − p)n−x ,
x

(a) Find the pgf.

(b) Use the result to show that the sum of two independent binomial random
quantities X1 and X2 , both with parameter p, but with n1 and n2 trials
respectively is another binomial distribution.

11

@seismicisolation
@seismicisolation
Solution
(a)
gx (t) = htx i
Xn µ ¶
n x
= tx p (1 − p)n−x
0
x
X µn¶
= (pt)x (1 − p)n−x
x
= [pt + (1 − p)]n
using binomial expansion.
(b)
[pt + (1 − p)]n1 · [pt + (1 − p)]n2
[pt + (1 − p)]n1 +n2
binomial distribution; parameter p with (n1 + n2 ) trials.

Exercise 7.16
Use of transforms for expected utility. Keeney and Raiffa (1976) note that
transforms can be used for expected utility. Use (7.100), (7.102) or (7.137) to
obtain the expected utility if the utility function is given by (4.18):
u (x) = − exp (−cx) .
Obtain the equations of expected utility for various distributions such as the
binomial and normal.

Solution
The easiest way is via the Laplace Transform
Z
L(f ) = e−st f (t)dt

Can also use the mgf. We want

Z ∞ Z ∞
u(x)fX (x)dx = − e−cx fX (x)dx;
−∞ −∞

but −c = s in mgf and we have

Z
− esx fX (x)dx = − hesx i = −mX (s)

= −mX (−c)
mX is mgf. Obtain values for various distributions (p.202 Keeney and Raiffa)

12

@seismicisolation
@seismicisolation
Exercise 7.17
Show that the cf for a random quantity with a uniform distribution on [0, 1] is
¡ ¢ ¡u¢
exp iu
2 sin 2 exp (iu) − 1
φX (u) = u =
2 iu

Solution
For random quantity
fX (x) = 1 for 0 ≤ x ≤ 1

Characteristic function
iu
e 2 sin u2
φX (u) = u ←− 1
2
eiu − 1
= ←− 2
iu
Can be proved by substituting
eiw − e−iw
sin w = −i sinh w =
2i
into 1. Thus
à iu iu
!
2 e 2 − e− 2 iu 1 iu
φX (u) = e2 = (e − 1)
u 2i iu

13

@seismicisolation
@seismicisolation
which is 2
See Johnson et al. V II, p.276

Exercise 7.18
Deduce from first principles the cf of the gamma density

xα−1 exp (−x/β)

fX (x) =
β α Γ (α)

to give equation (7.124)

φX (u) = (1 − iuβ)−α .

Solution
See Feller II, p.502

Exercise 7.19
Find the cf of the bilateral exponential distribution:

fX (x) = exp (− |x|) for − ∞ < x < ∞.

Solution
Answer
1
1 + u2
See Feller II, p.502 (or from exponential distribution)

Exercise 7.20
Find the characteristic function of the Cauchy distribution (introduced in exer-
cise 3.14),
a
fX (x) = , for − ∞ < x < ∞,
π (x + a2 )
2

using the inversion formula. Show that the distribution is infinitely divisible
and stable.

14

@seismicisolation
@seismicisolation
Solution
e−a|u|
Also see Feller II, p.502. Start with a = 1 −→ e−|u| is characteristic function.
³ ´n
e−|u| = e−|nu| for n > 0

for
k
fX = ¡ x ¢2
1+ n
Since fX is invariant apart from changes in scale, it is also stable de Finetti, I,
p.292

Exercise 7.21
Ball bearing problem. The diameter X of a spherical ball bearing taken from a
lot is normally distributed with a mean of 10.0 mm and a standard deviation
of 0.5 mm. The ball is supposed to fit into a steel tube taken from another
lot. The tube is circular in cross section, and the diameter Y is also normally
distributed but with a mean of 11.3 mm and a standard deviation of 1.2 mm.
Use the result of Example 7.19 to find the probability that a randomly chosen
ball will fit into a randomly chosen tube. Five randomly chosen balls are paired
off with five randomly chosen tubes. What is the probability that all five balls
will fit inside their tubes?

Solution
X is N (10.0, 0.52 ), Y is N (11.3, 1.22 )

V =X −Y

σ2V = 0.52 + 1.22 = 1.69

σV = 1.3mm

0 − (−1.3)
Pr(f it) = Pr(X − Y ) < 0 = Pr(z < )
1.3
= Fz (1.00) = 0.8413

Pr(all 5 pairs fit) = (0.8413)5 = 0.4214 · · ·

15

@seismicisolation
@seismicisolation
Exercise 7.22
Show that, for the binomial random quantity X,
Pr (X = x + 1) n−x p
= · .
Pr (X = x) x + 1 p̃

Solution
px+1 n! x!(n − x)! px+1 q n−x−1
=
px (x + 1)!(n − x − 1)! n! px q n−x
n−x p
= ·
x+1 q
de Finetti, II,p.23

Exercise 7.24
Prove the relation between (7.139) and (7.140):
g (t) ⇐⇒ G (ω)
then
G (t) ⇐⇒ 2πg (−ω) .
Give an illustration of its use.

Solution
Z ∞
1
g(t) = G(ω)eiwt dω Equation (7.138)
2π −∞
w −→ write as v
Z
2πg(t) = G(v)eivt dv
Z ∞
2πg(−t) = G(v)e−ivt dv
−∞

Now change t to ω, and then v to t!

G(t) ⇔ 2πg(−ω)
For illustration seeLathi, p.87.

Exercise 7.25
Deduce equation (7.161) using convolution formulae and mathematical induc-
tion.

16

@seismicisolation
@seismicisolation
Solution
See text for n = 1, 2, 3
See Johnson et. al. V.2, p.296
Feller, II, p.27 P
Xi
Related material. Consider Y = n

 µ ¶ µ ¶ 
n−1 n n−1 n
 (ny) − (ny − 1) + (ny − 2)n−1 − · · · 
n  1
µ ¶ 2 
fY (y) =
(n − 1)!  n 
(−1)r (ny − r)n−1 , for nr < y ≤ r+1
n
r

Y 0Y :

fY 0 (y 0 )dy 0= fY (y)dy
y0 1
fY (y 0 ) = fY ( ) ·
n  n µ ¶ µ ¶ 
0 n−1 n 0 n−1 n
 (y ) − (y − 1) + (y 0 − 2)n−1 − · · · 
1  µ ¶ 1 2 
=
(n − 1)!  n 0 
(−1)r (y 0 − r)n−1 , for nr < yn ≤ r+1 0
n or r ≤ y ≤ r + 1
r

for r = 0, 1, ...., n − 1

17

@seismicisolation
@seismicisolation
Exercise 7.26
Using the result of Exercise 7.17 to show that the means and standard deviations
of the distributions of Equation (7.161) are given by
n n
µY = and σ 2Y =
2 12
respectively.

Solution
Expand equation 2 in solution to Exercise 7.17 as series:
à ! µ ¶
1 iu (iu)2 (iu)3 iu (iu)2
φX (u) = + + + ··· = 1 + + + ··· ;
iu 1! 2! 3! 2 6

P
∞ (iu)j
equivalent to
j=0 (j + 1)!

"µ ¶µ ¶ µ ¶2 #
n in (iu)2 n(n − 1) in (iu)2
φY (u) = φnX (u)
= 1+ + + ··· + + + ···
1 2 6 2! 2 6
µ ¶
in (in)2 n(n − 1) (in)2
= 1+n +n + + ··· + ···
2 6 2 4
µ ¶
in n n(n − 1)
= 1+n + + (in)2 + · · ·
2 6 8
µ 2 ¶
in 3n + n 2
= 1+n + (in) + · · ·
2 24

1 0 n
µY = hY i = φ (0) =
i 2
­ 2® 1 00 3n2 + n n(3n + 1)
Y = 2 φ (0) = =
i 12 12
­ ® n 2
3n2
+n n
σ 2Y = Y 2 − hY i2 = − + =
4 12 12
Note that σ 2Y −→ ∞
n µY σ 2Y
1 1
1 2 12
1
2 1 6
3 1
3 2 4
1
4 2 3

18

@seismicisolation
@seismicisolation
Exercise 7.27
Continuity correction. In approximating discrete distributions by continuous
ones, as noted in section 7.4.2, probability masses in the discrete distribution
are approximated by areas, for example by small rectangles or trapezoids. The
areas can be better approximated using the “continuity correction”. In practice,
this merely means approximating the probability of a particular value i in the
discrete distribution by the area under the continuous curve from i − 1/2 to
i + 1/2.
The arrival of loaded carriers at a work station in a conveyor system occurs
with probability p = 0.3, the remaining occurrences being unloaded (empty)
carriers. The arrival of 20 carriers is being studied using the binomial distrib-
ution. Let X be the number of loaded carriers. Use the normal approximation
to the binomial distribution to calculate
(a) Pr(X = 5) , and
(b) Pr(2 ≤ X ≤ 10) .
Compare the results to the values from the binomial distribution.

Solution

(a)
µ = np = 6;
σ 2 = 4.20
σ = 2.049

19

@seismicisolation
@seismicisolation
 µ ¶
4.5 − 6.0 5.5 − 6.0
Pr(X = 5) ≈ Pr <Z<
2.049 2.049
= Pr(−0.7319 < Z < −0.2440)
= 0.4036 − 0.2321 = 0.1715

b(5, 20, 0.3) = 0.1789

(b)
µ ¶
1.5 − 6 10.5 − 6
Pr(2 ≤ X ≤ 10) = Pr <Z<
2.049 2.049
= Pr(−2.20 < Z < 2.20) = 0.9722

Binomial value is 0.9753.

Exercise 7.28
For the multidimensional normal distribution of Example 7.9, determine the
characteristic functions of the forms given in equations (7.57) and (7.68).

Solution
Part I.
Feller (1968, p. 522), Fraser (1976, p. 235), Lloyd (1980, p. 275)
Standard normal · ¸
1
exp − uT u
2
or · ¸
1
exp − uT [I] u
2
or with covariance matrix
· ¸
1 hXi
exp − uT u
2
P
—see Equation (7.69), covariance matrix. Note, not [ ]−1 here.
 

 u1 
 u2 
 
u= ..

 
 . 
 

un
Generally, · ¸
1 hXi
exp − uT u + iuµ
2

20

@seismicisolation
@seismicisolation
  

 µ1 


 µ2 

where µ is a vector ..

 . 


 

µn

Part II.
If X is normal
σ 2 u2
φX = eiµu− 2

1
ψ X = iµu − σ 2 u2
2

κ1 = µ κ2 = σ 2

X∞ (s)
µ
ψ X = ln φX = ln[1 + (iu)s X ]
s=1
s!

(1) iu £ ¤ (iu)2
= . . . (Ochi, 1990, p.85) = µX + µ2X − (µ1X )2 + ...
1! 2!
.
(2) (1)
σ 2X = µX − (µX )2

Exercise 7.29
Sums of random numbers of random quantities. Very often we are interested in
the sum
Y = X1 + X2 + · · · + Xi + · · · + XN ,
where the Xi ’s and N are random. Assume that the Xi ’s are iid.

(a) Give examples of practical interest of such a situation.

(b) Find an expression for the mgf of Y in terms of the mgf of X and the
probability distribution of N.
(c) A mathematical model of the travel of tourists is to be set up for a group
of islands. The number of tour groups per day, N, visiting a particular
island is random with a Poisson distribution and a mean equal to µ. The
size X of a tour group is also Poisson-distributed with a mean equal to ν.
Find the mgf of Y, the number of visitors per day at the specified island.

21

@seismicisolation
@seismicisolation
Solution
(a) • Damage caused by (random) no. of earthquakes, lightning, etc. is a
R.Q. Total damage fits this model.
• Telephone calls—no. is a random quantity. Duration is a R.Q. Total
time exchange is in use fits the model.
P P
• Waiting time at a bus stop = (individual entry times). is over
a random no. of people. Or, better, for travel
P time - no. of stops is a
R.Q.; −→ total time for boarding/leaving is (R.Q.) over a random
number of stops.
(b) If n is known, the mgf of the sum of n independent, identically distributed
RQ’s is:
n
MX (t)
where MX (t) is the mgf of X. Now
∞
X
fY (y) = fN (n)fXn (y) = fN (0)fX0 (y)+fN (1)fX1 (y)+fN (2)fX2 (y)+· · ·
n=0

Note: fXn (y) =Pr(X1 + .... + Xn ) = n-fold convolution of fX (x) with

itself. Hence ∞
X
n
MY (t) = fN (n)MX (t) (1)
n=0

Note: if n = 0, X1 + .....Xn = 0 = y; fX0 (0) must = 1, fX0 (y) = 0 for y 6=

n
0, and MX (t) = 1 for n = 0, further confirmed by [MY (t)]n=0 = E(etY )
using fX0 as given above; E(ety ) = 1 in this case; hence the first term in
equation (1) above is fN (0)
(c)
X ∞
e−µ µn νn(et −1)
MY (t) = e
n=0
n!
h in
e−µ µn t t
where n! = fN (n) and eνn(e −1) = eν(e −1) .
· ¸
d
E(Y ) = MY (t)
dt t−→0

X∞
d e−µ µn νn(et −1)
MY (t) = ·e · νnet
dt n=0
n!

E(Y ) is found by letting t −→ 0; hence

X∞
e−µ µn
E(Y ) = ν · n = νE(N ) = µν
n=0
n!

22

@seismicisolation
@seismicisolation
 e−µ µn h νn(et −1) ¡ i
X∞
d2 ¢
t 2 νn(et −1) t
M Y (t) = e νne + e νne
dt2 n=0
n!
Letting t −→ 0, we find

e−µ µn h i
X∞
E(Y 2 ) = (νn)2 + νn
n=0
n!
¡ ¢
= ν 2 E(N 2 ) + νE(N ) = ν 2 µ2 + µ + νµ

Hence

σ 2Y = E(Y 2 ) − E 2 (Y ) = ν 2 (µ2 + µ) + νµ − ν 2 µ2
= ν 2 µ + νµ = νµ(1 + ν)

£ ¤
MX (t) = E etx

1  σ =1.3m
σ02 = 1 3 = 0.1120 σ = 0.6
+ 
1.69 0.36 µm = 4.9
σ 0 = 0.3347

µ0 = 0.1120(2.8994 + 8.333x̄) = 0.3247 + 0.9333x̄

8
Pr = Pr(M > 5) =
15
0 02
N (µ , σ )
x̄ − (0.3247 + 0.9333x̄
Pr(Z > ) = 0.533
0.3347
x̄ − 0.3247 − 0.9333x̄ ∼
=⇒ = 0.08
0.3347

x∼
= 5.27

Exercise 7.30
There are 35 components in an assembly. Each component has a mean life of 5
years with a standard deviation of 2 years. Their lifetimes may be considered
to be independent in the application considered. Find the probability that the
mean life of all 35 components is greater than 6 years.

23

@seismicisolation
@seismicisolation
Solution
Use central limit theorem P35
i=1 Xi
X=
35
where x̄ is approximately normal, with µ = 5, σ x̄ = √2 = 0.3381
35

6−5
Pr(x̄ > 6) ∼
= Pr(z > ) = Pr(z > 2.96)
0.3381
= 1 − 0.9985 = 0.0015

Exercise 7.31
The strong law of large numbers cannot be proved using the Chebyshev inequal-
ity. Find the proof.

For solution
See Feller (1968, II, p. 237)
de Finetti (2, p. 33)

Exercise 7.32
Research the following concepts: convergence in the quadratic mean; weak con-
vergence (convergence in probability); strong convergence (almost sure conver-
gence). Which implies the others and why?

For solution
see de Finetti (1, p.264-266)
Papoulis (1965, p. 188)

24

@seismicisolation
@seismicisolation
Solutions to Chapter 8
Exercise 8.1
In a problem involving the binomial distribution, the uniform distribution is
(i)
judged an appropriate measure of prior knowledge, i.e. fA (a) = 1, 0 ≤ A ≤ 1.
Ten observations of the process are made with x successes. Plot the pdf of the
posterior distribution for all possible results, i.e. with x = 0, 1, . . . , 10 successes.

Solution
Let a = proportion of "successes" in the population. A sample of size n is
available (with replacement or from a large population). Let X = # of successes
in sample and let A = random quantity representing the unknown a; then

Pr (X = x | A = a) = b(x; n, a)

We shall start by modelling A as a discrete distribution. Let our prior belief

about a be represented by the probability mass function Pr(A = a) = p0 (a).
Then the posterior distribution, obtained by updating our belief with the result
from the sample, by Bayes’ theorem is

Pr (A = a) ∧ (X = x)
Pr (A = a | X = x) =
Pr (X = x)
Pr (X = x | A = a) · Pr (A = a)
= P
a Pr (X = x | A = a) · Pr (A = a)
= pU (a)

Then

b(x; n, a) · p0 (a)
pU (a) = P
a b(x; n, a) · p0 (a)

To let A become continuous, we let the probability point mass f0 (a) spread
out over an infinitesimal interval da, with mass f0 (a)da, where f0 (a) is now a
probability density function. Then

b(x; n, a) · f0 (a)da
fU (a)da = Pr [(a ≤ A < a + da) | (X = x)] = R 1
0
b(x; n, a) · f0 (a)da

Suppose now that we consider any one value of a to be just as likely as any
other in the interval 0 ≤ a ≤ 1. (See pp. 396 to 398 and section 8.6.6 of the
book for a discussion of this assumption.) Then

@seismicisolation
@seismicisolation
 ½
1, for 0 ≤ a ≤ 1
f0 (a) =
0, otherwise
and
b(x; n, a)f0 (a) =n Cx ax (1 − a)n−x

Z 1 Z 1
b(x; n, a) · f0 (a)da = n Cx ax (1 − a)n−x da
0 0
n! x!(n − x)!
= ·
x!(n − x)! (n + 1)!
1
=
n+1
Upon observing x successes in n trials, our assessment of the value of a is
updated to

fU (a) = (n + 1)n Cx ax (1 − a)n−x for (0 ≤ a ≤ 1)

The mean hAi for this distribution is

Z ∞ Z 1
(n + 1)n!
a · fU (a)da = · ax+1 (1 − a)n−x da
−∞ x!(n − x)! 0
(n + 1)! (x + 1)!(n − x)! x+1
= · =
x!(n − x)! (n + 2)! n+2
The most likely value of A (the mode) is found by solving
d
fU (a) = 0
da
giving the mode at
x
a=
n
(which is the classical "point estimate" for A).
If a sample of size n = 10 is drawn and used to update f0 (a), then the
updated pdf for A is:
Observed value of x fU (a) hAi Mode
1
0 11(1 − a)10 12 0
9 2 1
1 110a(1 − a) 12 10
2 8 3 2
2 495a (1 − a) 12 10
3 7 4 3
3 1320a (1 − a) 12 10
4 6 5 4
4 2310a (1 − a) 12 10
5 5 6 5
5 2772a (1 − a) 12 10
6 4 7 6
6 2310a (1 − a) 12 10
8 7
7 1320a7 (1 − a)3 12 10
8 2 9 8
8 495a (1 − a) 12 10
9 10 9
9 495p (1 − a) 12 10
10 11
10 11a 12 1

@seismicisolation
@seismicisolation
 Figure 1:

Exercise 8.2
Show that (8.32) reduces to (8.18) for large r0 , n0 .

Solution
We want to show that
µ ¶
n B (x + r0 , n + n0 − x − r0 )
Pr(X = x|n) =
x B (r0 , n0 − r0 )
µ ¶
n Γ (n0 ) Γ (x + r0 ) Γ (n + n0 − x − r0 )
= 0 0 0
·
x Γ (r ) Γ (n − r ) Γ (n + n0 )

reduces to
µ ¶ µ ¶
n x n−x n x n−x
Pr(X = x|n) = pX (x) = p (1 − p) = p q ,
x x
¡ ¢
The term nx is common to both so we deal with the terms to the right of
this. Treat the gamma functions as factorials (Γ (s) = (s − 1)!) (see Exercise
¡ ¢
7.13), so that we can write the first of the equations (omitting the term nx ) as

(n0 − 1)! (x + r0 − 1)!(n + n0 − x − r0 − 1)!

·
(r0 − 1)!(n0 − r0 − 1)! (n + n0 − 1)!
Note that x, n ¿ r0 , n0 . This can be rearranged as follows

(n0 − 1)! (x + r0 − 1)!(n + n0 − x − r0 − 1)!

0
·
(n + n − 1)! (r0 − 1)!(n0 − r0 − 1)!

@seismicisolation
@seismicisolation
Now observe that
(n0 − 1)! 1
0
→ 0 n,
(n + n − 1)! (n )
and
(x + r0 − 1)!(n + n0 − x − r0 − 1)!
(r0 − 1)!(n0 − r0 − 1)!
(x + r − 1)! (n + n0 − x − r0 − 1)!
0
= ·
(r0 − 1)! (n0 − r0 − 1)!
x n−x
→ (r0 ) · (n0 − r0 ) .

The initial expression

µ ¶
n Γ (n0 ) Γ (x + r0 ) Γ (n + n0 − x − r0 )
0 0 0
· →
x Γ (r ) Γ (n − r ) Γ (n + n0 )
µ ¶ 0 x 0 µ ¶
n (r ) (n − r0 )n−x n x
= a (1 − a)n−x ,
x (n0 )n x

where
r0
a= .
n0

Exercise 8.3
Show that ω x,n in the case of Bayes’ scholium is equal to 1/(n+1) by integrating
(8.43) with a uniform prior distribution. Also use (8.32) to verify this.

Solution
¡n¢
Recall that ω r,n = r pr,n . Then
µ ¶Z 1
n n−x
ω x,n = ax (1 − a) dx
x 0
µ ¶
n
= B (x + 1, n − x + 1)
x
µ ¶
n Γ (x + 1) Γ (n − x + 1)
=
x Γ (n + 2)
n! x!(n − x)!
= ·
x!(n − x)! (n + 1)!
1
= .
n+1
(See also the solution to Exercise 8.1 above.)
See Raiffa and Shlaifer, page 237, for second part. Put r0 = 1 and n0 = 2 in
equation 8.32.

@seismicisolation
@seismicisolation
Exercise 8.4
The voltage that can be carried by an integrated circuit is known to be nor-
mally distributed with a standard deviation of 0.10 volts. This is the “system
capacity” (see also Chapters 9 and 10). The mean capacity is to be studied in
an application. Past measurements lead to the conclusion that the mean ca-
pacity is normally distributed (µ0 = 5.50 volts, σ 0 = 0.05 volts) . Six tests are
undertaken on prototype circuits, with the results of the measurements giving
that x̄ = 5.80 volts.

(a) Update the distribution of the mean capacity using the prior information
and the results of the measurements.
(b) Find the 99% interval estimate based on the Bayesian distribution found
in (a). (This is the so-called “Bayesian interval estimate”.)
(c) Contrast the result of (b) with the classical confidence intervals. Sketch
all distributions.
(d) How many measurements need to be taken to reduce the standard devia-
tion of the mean to 0.01? Use the result of (a) as the basis of calculation.

Solution
(a)
1
w0 = = 20 = 400
(0.05)2
n
wd = = 100 = 600
(0.1)2
(400)(5.50) + (600)(5.80)
µf = = 5.68
1000
1
σf = √ = 0.0316 . . .
1000

(b)
5.68 ± (2.57583)(0.0316...) → 5.60 < A < 5.76)
with 99% probability.
(c) Classical result:
0.1
5.80 ± (2.57583)( √ ) → 5.69 < A < 5.91
6
with 99% probability.

@seismicisolation
@seismicisolation
 14

12

Posterior

10
Prior

8
fV(v)

2 S.D.
= 0.05 S.D. = 0.0316

0
5.2 5.3 5.4 5.5 5.6 5.7 5.8 5.9 6
Voltage (V)

Figure 2:

Intervals are quite different: importance of prior information (see sketch

above).

(d)
1
σf =0.01 = √
400 + 100n
→ 1002 = 400 + 100n

Then

9600
n= = 96
100
Classical result:
σ2
= (0.01)2
n
0.01
→ n= = 100
(0.01)2

@seismicisolation
@seismicisolation
Exercise 8.5
Assume now that the measurements on the prototype circuits in question 8.4
have not been taken. It is being decided whether or not to take these mea-
surements. Voltage surges occur in the circuit in practical applications. The
maximum of these in a time period under consideration is termed the “demand”,
which is to be compared to the capacity of question 8.4. If demand exceeds ca-
pacity, failure occurs.
The basic question to be decided is whether to include the circuits in a
design to be marketed. Draw a decision tree that models the problem. Include
the option “not to experiment” versus “experiment with n prototype circuits”.
Note on the decision tree the following consequences in units of utility.
Design accepted and survival: 10 units
Design accepted and failure: −10 units
Design rejected: −3 units
Cost of testing: 0.5 units per circuit tested.
The failure probability can be worked out (in closed form for two normal
or two lognormal distributions–see Chapter 10). This could be done with
representative values of the parameters, and the problem solved fully. In the
calculation of failure probability, a future value of capacity is required (see
exercise 8.6).

Solution

Consider
D − C;
this is normally distributed
¡ ¢
N µD − µC , σ2D + σ 2C ,

assuming independence. We want

Pr (D > C) = Pr (D − C) > 0;

Finding this probability is straightforward since we are dealing with a normal

distribution. See also Chapter 10 of the probability of failure.

Exercise 8.6
Future values. In many problems, we are interested in a future value. For exam-
ple, we might have estimates of the parameters for a batch of steel specimens.
Rather than the distribution of the parameter “mean”, we might wish to es-
timate the uncertainty in a future single value of strength, corresponding to a
particular specimen. Describe how this might be accomplished.

@seismicisolation
@seismicisolation
D

Figure 3:

Figure 4:

@seismicisolation
@seismicisolation
Solution
We consider the normal distibutions as in sections 8.5.1, 8.5.2 and 8.5.3. Use
σ f as in Bayesian analysis (sections 8.5.2 and 8.5.3) or σ if variance is known.
The standard deviation on the future value is then
r
1
σf 1 +
n
or r
1
σ 1+ .
n
Data-based result is r
1
s 1+
n
and typical confidence intervals are
r
1
x̄ ± zα/2 · σ 1 +
n
and r
1
x̄ ± tα/2 · s 1 + .
n
(Uncertainty associated with σ f must be taken into account—hence the use of
the t-distribution.)
The estimation of uncertainty on future values is dealt with in regression as
well; see section 11.3 and page 593 of the text. We consider regression of Y on
x. One can show that (Y |w, x0 ) is
" #
2
0 2 σ2 2 (x0 − x̄)
N â + b̂ (x0 − x̄) , σ + +σ .
n sxx

The terms
h in square i brackets are the mean and variance. The third term of the
2
2 (x0 −x̄)
latter σ sxx is related to the uncertainty in the slope of the regression
³ 2
´
line; the other terms σ 2 + σn are as above. Note that in Chapter 11 the
symbol σ 2 is treated as uncertain, represented
¡ ¢ there by W. Taking into account
the uncertainty in W, standing for 1/σ 2 , we found that
h i
Y − â0 + b̂ (x0 − x̄)
T = q 2
s 1 + n1 + (x0s−x̄)
xx

follows a t-distribution with (n − 2) degrees of freedom.

Exercise 8.7
Show how to obtain (8.89) from the prior distribution and the data.

@seismicisolation
@seismicisolation
Solution
v0 −w
(i) −1
fW (w) ∝ w 2 e 2w0
n w 2
l(w) ∝ w 2 e− 2 sd
Note that
w 1 w
( + s2d ) =
2 w0 2wf
v0 +n w
(f ) − 2w
fW (w) ∝ w 2 e f

Form is as in fW 0 (w0 ); see next exercise. Governed by exponential term

−w
− 2w
e ref .

Exercise 8.8
Show that W in ν ref ³ 0´
−1
(w0 ) 2
exp − w2
fW 0 (w0 ) = ν ref ¡ν ¢
2 2 Γ ref 2

with W 0 = W/wref has a gamma distribution of the form (8.82) with parameters
α = ν/2, and β = 2wref . The subscript “ref ” refers to the parameter of either
the prior or posterior distribution (subscripts “0” or “f ” respectively). The chi-
square random quantities in equations (8.85) and (8.89) are of the form (8.171)
and the result shows how to model W using either the chi-square or gamma
forms.

Solution
w
w0 = ;
wref
fW dw = fW 0 dw0 ;
1
fW = fW 0
wref

v w0
0 (w0 ) 2 −1 e− 2
fW 0 (w ) = v
2 2 Γ( v2 )

10

@seismicisolation
@seismicisolation
 1 w
fW (w) = fW 0 ( )
wref wref
w
v −
1 (w) 2 −1 e 2wref
= v v
wref (wref ) 2 −1 2 2 Γ( v2 )
w
v −
w 2 −1 e 2wref
= v
(2wref ) 2 Γ( v2 )
w
wα−1 e− β
=
(β)α Γ(α)
v
This is a 2-parameter gamma distribution with α = 2 and β = 2wref

Exercise 8.10
Geometric interpretations.
(a) Show that equation (8.105) represents the distance from the origin to the
centre of mass of the combined masses n0 and n.
(b) Show that equation (8.107) represents the second moment of the dis-
tributed masses about the centre in (a). We have assumed, in writing
ν 0 = n0 − 1, that the prior information is in the form of, or analogous to,
a set of data with n0 members. If this is not the case, one can think of the
parameters n0 and (ν 0 σ20 ) as the mass and moment of inertia of the prior
information, without there being necessarily any link between ν 0 and n0 .

Solution
(a) See figure
µ0 n0 + x̄n
x0 =
n0 + n
P(b) The 2nd moment of distributed masses = (2nd moment about centroid)
+ (mass)(d2 ),
where d = distance to axis from centroid
(Parallel axis theorem)
The first two terms in (8.107) correspond to
X
(xi − x̄)2 ,

where the xi are masses and x̄ the centroids. The final second moment is
first two terms (ν 0 σ 20 + νs2 )
µ ¶2 µ ¶2
µ0 n0 + x̄n µ0 n0 + x̄n
+n0 − µ0 + n − x̄ ;
n0 + n n0 + n

11

@seismicisolation
@seismicisolation
 Figure 5:

these second two terms are equal to

2 2
n0 n2 (x̄ − µ0 ) + n20 n (x̄ − µ0 )
(n0 + n)2
n0 n (n0 + n) (x̄ − µ0 )2
= 2
(n0 + n)
n0 n (x̄ − µ0 )2
= ,
(n0 + n)
as in equation (8.107).

Exercise 8.11
Find the constant in the proportionality relationships (8.102) and (8.103).

Solution
Take (8.103) for example:
n wh ¡ io
(u) ¡ ¢ ¢2
νf 1
fA,W a, w|nf , µf , ν f , σ f ∝ w 2 − 2 exp − nf a − µf + ν f σ 2f .
2
We want C such that
n wh ¡ io
(u) ¡ ¢ ¢2
νf 1
fA,W a, w|nf , µf , ν f , σ f = C · w 2 − 2 exp − nf a − µf + ν f σ2f ;
2

12

@seismicisolation
@seismicisolation
this is obtained from
ZZ
(u) ¡ ¢
fA,W a, w|nf , µf , ν f , σ f dadw = 1.

Then
½Z h w¡ Z h w ¡ ¾
νf
− 12 2
¢i ¢2 i
C w 2 exp − ν f σf exp − nf a − µf da dw = 1;
2 2
the integral over a is of the form of the normal distribution; thus
 Ã !2 
Z √ −1/2
1 a − µf
exp − √  da = 2π w√
2 w −1/2 / nf nf

and ½Z ¾
νf h w¡ ¢i √ w−1/2
− 12
C w 2 exp − ν f σ 2f 2π √ dw = 1
2 nf
or √ ½Z ¾
C 2π νf h w¡ ¢i
−1 2
√ w 2 exp − ν f σf dw = 1. (A)
nf 2
Now the gamma distribution
³ ´
xα−1 exp − βx
fX (x) = , for x ≥ 0, = 0 otherwise,
β α Γ (α)
normalizes upon integration to unity; the integral in the equation (A) is of this
form so that √
C 2π 1 ν
³ν ´
√ ·³ ´ νf · 2
2 · Γ = 1.
nf 2 2 2
ν f σf
The final result, omitting the subscript "f ", is
r
−( ν2 + 12 ) n ¡ 2 ¢ ν2 1
C=2 νσ ¡ ¢.
π Γ ν2

Exercise 8.14
Verify (8.120).

Solution
(u) −( βa +an)
fA (a) ∝ aαi −1 anr e i

−a( β1 +n)
= a(αi +nr)−1 e i

1 1
where (αi + nr) = αf and βi +n= βf

13

@seismicisolation
@seismicisolation
Exercise 8.15
The occurrence of flaws in solid bodies can be modelled using the Poisson dis-
tribution, using length, area (see exercise 3.8) or volume as the “exposure”
parameter. This can be used in failure analysis; for example the safety of steel
pressure vessels in nuclear reactors, pipelines, ships, offshore rigs, and in aircraft,
to name a few.

(a) The occurrence of flaws in a welded joint is modelled as a Poisson process

with a mean occurrence rate of Λ flaws per metre of weld. Prior informa-
tion has show that this parameter has a mean of 0.5 flaws per metre with a
coefficient of variation of 40%. Model this as a prior density using the ap-
propriate conjugate family. Update the distribution using measurements
with a reliable device indicating 5 flaws in a weld of length 9.2 metres.
Calculate the posterior mean, variance and coefficient of variation.
(b) Structure a decision problem–such as that in (a)–using the Poisson
process involving initial information regarding the expected value or rate
of the process, which might be updated using Bayesian methods. Show
decision trees in which the decision “whether to experiment” is included
(preposterior analysis). The problem could be concerned with events other
than flaws in a material; arrivals of messages or packets of information
could be considered or any of the numerous uses of the Poisson process.

Solution
(a) Use the gamma prior distribution of the form

−a
aα−1 e β
fA (a) =
β(α)Γ(α)

Mean = αβ

¾ √
Variance =√β 2 α β α 1
c. of v. = =√
s.d. = β a αβ α

1 1
√ = 0.40; αi = = 6.25
αi (0.40)2
1
6.25(β i ) = 0.5; = 12.5
βi

)
αf = 6.25 + nr = 6.25 + 5 = 11.25
1 Defines new gamma distribution
β = 12.5 + 9.2 = 21.7
f

14

@seismicisolation
@seismicisolation
 Figure 6:

1
Posterior Mean = αf β f = 11.25 · = 0.518
21.7
1
Posterior Variance = 11.25 · = 0.0239; s.d. = 0.155
(21.7)2
1
Posterior c. of v. = √ = 0.298
11.25
(b) Typical decision problem is shown in figure:

Exercise 8.16
Show that the reference prior (8.127) for the scale parameter, applied to the
standard deviation, also applies to the variance and to the weight W .

Solution
Consider
y = g(x) = xn ;
1
fx (x) ∝
x
x is the s.d.; −→ x2 is the variance and x−2 is the weight w

fy dy = fx dx

1 dx 1 1 1 1−n
x = yn; = y n −1 = y n
dy n n
1 dx 1 1 1 1
fY ∝ · = 1 · n−1 = 1 1 =
x dy y n ny n ny n y 1− n ny

15

@seismicisolation
@seismicisolation
and
1
fy ∝ .
y

Exercise 8.17
What is the reference prior for Example 8.18? How would you interpret the
prior µ ¶
1 λ
fΛ (λ) = exp − .
λ̄ λ̄
Is there a maximum-entropy interpretation of this prior distribution?

Solution
1
fΛ (λ) ∝ ;
λ
µ ¶
1 λ
fΛ (λ) = exp −
λ̄ λ̄
implies one observation in specified time period td . Maximum entropy solution
subject to mean rate λ̄ known.

Exercise 8.18
Linear loss problems. In problems of decision, there is often a linear loss asso-
ciated with the parameter in the probability distribution. Raiffa and Schlaifer
(1961) provide substantial background to this problem. For example, these au-
thors describe a manufacturing process in which objects are mass-produced in
lots of size n. The proportion of defectives is a binomial process with parameter
A. The lots can be sent directly for assembly, in which case the defective parts
have to be replaced at a cost c1 each. Or the lots can be screened at cost c2 per
lot and the defectives replaced at cost c3 < c1 .

(a) Write the utility functions in terms of cost for the two strategies, assuming
utility to be linear with money, in terms of the parameter A = a. Find
the break even point for a in terms of the constants.
(b) How would you take into account randomness in A? Research the lin-
ear loss integrals, integrated over two parts of the parameter space, as
presented by Raiffa and Schlaifer. How could they be applied in case (a)?
(c) Sketch a Bayesian approach to the problem of determining the parameter
A at minimum cost.
(d) Suggest conditions under which the cost of testing and experimentation is
additive to the utility and independent of the current value of utility.

16

@seismicisolation
@seismicisolation
 (e) Propose an application of the theory above in an area of interest to you.
Note that Raiffa and Schlaifer have developed linear loss integrals for a
variety of processes other than the binomial one.

Solution
Linear Loss
(a) Let α1 and α2 be the strategies.
¾
α1 u1 = −c1 na
;
α2 u2 = −c2 − c3 na
equate for cutoff "a"= ac . Then
c2
ac =
(c1 − c3 )n

α1 optimal if a < ac
α2 optimal if a > ac
(b) Integrate to obtain the expected utility:
Z
u(a)pA (a) da;

u(a) is linear in a, so that the result is in terms of the expected value hAi .
Linear loss: Z ac
...
0
for the first strategy α1
Z 1
...
ac

for the second strategy α2 .

(c) see figure 7
(d) If utility is linear with money in the range under consideration. See also
Raiffa and Schlaifer (1961), p.79. R
(e) See Raiffa and Schlaifer (1961), p.93; p. 97 - definition of linear loss ,
p.99 - table.

Exercise 8.19
Verify equations (8.163) and (8.164).

17

@seismicisolation
@seismicisolation
 Figure 7:

18

@seismicisolation
@seismicisolation
Solution
Equation (8.89) gives directly that
µ ¶
2 W 2
Pr χ1− α2 ≤ ≤ χ α2 = 1 − α;
wf
hence
à !
1 1 1
Pr 2 ≤ ≤ =1−α
wf χ α W wf χ21− α
2 2

with 1/wf = (n−1)s2 in classical case (equation 8.144). Hence equation (8.164).

Exercise 8.20
Find 90% confidence intervals for the parameter W and for the corresponding
standard deviation in example 8.16.

Solution
10 d. of f.; χ20.95 = 3.94; and χ20.05 = 18.31

W
Pr(3.94 ≤ ≤ 18.31) = 0.90;
wf
wf = 61.07

Pr(240.6 ≤ W ≤ 1118.2)
1
Pr(0.0008943 ≤ ≤ 0.004156)
W
Note that W stands for ( σ12 ) (random); then

Pr(0.030 ≤ σ ≤ 0.064)
where σ is random.

Exercise 8.21
Show that a random quantity that is the square of a standard normal distribu-
tion is chi-square with 1 degree of freedom.

19

@seismicisolation
@seismicisolation
Solution
Let

Y = Z2
where
1 1 2
fZ (z) = √ e− 2 z
2πσ
1 2
∝ e− 2 z

¯ ¯ ¯ ¯
¯ dz ¯ dy ¯ dz ¯ 1
fY 2 ¯
= fZ (with z = y) · ¯ ¯ ;¯ = 2z; ¯¯ ¯¯ = 1
dy dz dy 2y 2
1 1
fY (y) ∝ e− 2 y · y − 2

χ2 with 1 d.f.

Exercise 8.22
We know (example 7.21) that the gamma family is closed under convolution.
Show that the distribution of χ2 as given by equation (8.166) is chi-square with
n degrees of freedom.

Solution
Follows from "closed under convolution" and Exercise 8.21. See also Feller II
p.48

Exercise 8.23
Why in the classical approach are there (n − 1) degrees of freedom in the chi-
square distribution of (8.168) using (8.167)?

0.1 Solution
Nicely proved for example in DeGroot, M. H. and Schervish, M.J. 2002. p.397
ff.
Loosely, one degree of freedom lost in using X as mean from {X1 , . . . , Xn } .

20

@seismicisolation
@seismicisolation
Exercise 8.24
Bayesian fly in the ointment: inverted gamma distribution. Use the notation of
Exercise 8.8. Make the transformation
1
Y =
W
where Y models the variance to show that the distribution of Y follows the
“inverted gamma” (or “inverted chi-square”) distribution. Show that

σ2ref ν ref
hY i = ,
ν ref − 2

which differs from the “unbiased” result of (8.170).

Solution
We start with the result of Exercise 8.8, showing that W in
ν ref ³ 0´
−1
(w0 ) 2 exp − w2
fW 0 (w0 ) = ν ref ¡ ν ¢
2 2 Γ ref 2 (A)

with W 0 = W/wref has a gamma distribution of the form (8.82) with parameters
α = ν ref /2, and β = 2wref . The chi-square random quantities in equations
(8.85) and (8.89) are of the form (A) and the result shows how to model W
using either the chi-square or gamma forms. We write generally
³ ´ ν ref
2 −1
h ³ ´i
w
wref exp − 12 ww
ref
fW 0 (w0 ) = ν ¡ν ¢ .
2 2 Γ ref
2
The mean of this is
¿ À
W
= ν ref or
wref
1
hW i = νwref = . (B)
σ 2ref

since
1
ν ref σ 2ref = .
wref
The subscript “ref ” refers to the parameter of either the prior or posterior
distribution (subscript “0” or “f ”).
There are various versions of the distribution of equation (8.89). Lindley
(1965, v.2, p26) uses the variance σ2 as the random quantity; prior knowledge

21

@seismicisolation
@seismicisolation
is expressed via ν ref , as above, and σ 2ref ; the chi-squared random quantity is
(ν ref σ 2ref /σ 2 ); the mean is
* +
ν ref σ 2ref
= ν ref or
σ2
¿ À
1 1
= ,
σ2 σ 2ref
which is the same as (B). These results do not imply that
­ 2®
σ = σ 2ref .
­ ®
Then let us consider σ 2 . Lindley (1965, v.2, pp 28-29) uses the transformation
ν ref σ2ref
X=
2σ2
¡ ν ref ¢
which is Γ 2 , 1 . He gets
­ ® σ 2ref ν ref
σ2 = . (C)
ν ref − 2
One can make the transformation
1
Y = .
W
Then Y models the variance and
³ ´ ν ref
2 −1
h ³ ´i
1
ywref exp − 12 yw1ref 1
fY (y) = ν ref ¡ ν ref ¢ · 2
2 2 Γ 2 y
· µ ¶¸
ν ref 1 1
∝ y 1− 2 −2 exp −
2 ywref
· µ ¶¸
ν ref 1 1
= y −(1+ 2 ) exp − . (D)
2 ywref
Compare Y wref with χ−2 on p 88 Box, G.E.P., and Tiao, G.C., 1973, eq 2.3.12.
Equation (D) is the “inverted chi-square” distribution. See Lindley (1965, v.2,
pp 28-29) Box and Tiao pp 29-30 and pp 86-87 regarding the inverted chi-square
distribution; also Bernardo, J.M., and Smith, A.F.M., 1994, p119 shows that
the inverted gamma gives the following mean.
1
2wref 1 σ 2ref ν ref
hY i = ν ref = = (E)
2 −1 wref (ν ref − 2) ν ref − 2
since
1
ν ref σ 2ref = .
wref
We can see that (E) is the same as (C) above.

22

@seismicisolation
@seismicisolation
Exercise 8.25
Write notes expressing your opinion on estimation from the point of view of
decision-making, considering point estimates, centroids (mean values) and bias.
We gave examples in section 8.7.3 of cases where bias in the estimate is a
desirable property of the estimate, and in particular an overestimate. Give
further examples of this situation and give an example of the case where we
might wish to underestimate.

Solution
Underestimate the bid price on a strategic contact so as to increase probability
of success (if sufficient resources are available).

Exercise 8.26
Write notes expressing your opinion of the best estimator for a binomial process.
Consider cases where there is strong and weak prior information. In particular,
discuss the values r0 and n0 in (8.33):
r + r0
hAi = .
n1 + n0

Solution
The prior distribution can be considered as corresponding to r0 succeses in n0
trials. Weak priors are generally in the following range:

r0 = 0 −→ r0 = 1
n0 = 0 −→ n0 = 2
with mean nr00 . Use this mean, and other parameters such as standard deviation;
parameters in the range shown can express various opinions. See also section
8.3.2, 8.3.3 and 8.6.6. Weak priors will often be dominated by susequent data.
Strong priors can generally be covered by the above especially the mean nr00 .

Exercise 8.27
Research basis for members of the set of “Common Classical Inference Prob-
lems”, section 8.7.6 above.

Comment
My opinions are given at the beginning of sections 8.7.5 and 8.7.6. In the first of
these, possible defects in the statement of the null and alternative hypotheses are
described. An example is the proposing of hypotheses that have a probability

23

@seismicisolation
@seismicisolation
of being true equal to zero. This will not be dealt with any further in the
following solutions but should be borne in mind, especially when teaching the
methodology. A further fundamental point is concerned with the choice of the
level of significance. The common 5% significance level, or any other, is quite
arbitrary, with the 5% value resulting more from the 2-standard deviation value
for the normal distribution (1.960 ' 2) and the convenience of the number
"2". In engineering design, for example, in which safety is paramount, loads
at the 10−2 and 10−4 annual exceedance levels are commonly used (extreme
and abnormal events in ISO terminology), not arbitrarily decided but based on
detailed analyses of safety and probability of failure.
The methods are permeated with the attempt to convert uncertainty to
uncertainty, some of which will be pointed out below.

General comment on classical hypothesis testing

It is important to ponder in each case the relevance of the level of significance α
and the consequences of actions. It is suggested that decision trees be drawn in
each case. In real problems, possible prior opinion should also be conjectured.

Exercise 8.28
Classical inference 1. A manufacturer claims that the mean strength of a new
type of strand for cable manufacture is more than 20 kN. The distribution of
actual strengths is known to be normal to a good approximation, with a standard
deviation of 1.4 kN. A random sample of five of the new cables has a mean of
22.5 kN.

(a) Is there sufficient evidence to support the claim?

(b) Now suppose that the standard deviation σ is unknown and that the
measured standard deviation of the random sample is s = 1.4 kN. Is there
sufficient evidence to support the claim?

Solution
The following is a standard solution of the classical kind. Note the language
“burden of proof” that is sometimes used. This is the same kind of thinking
that leads a judge to state that a person is innocent of a crime when he is in
reality not at all sure that this is true.

(a) Let X be the tensile strength of a randomly selected cable. Then the mean
of a random sample of n such cables has a probability distribution
µ ¶
σ2
N x̄, ;
n

24

@seismicisolation
@seismicisolation
 see section 8.6.1. Summary statistics:

n = 5, x̄ = 22.5kN

The burden of proof is on the claim that µ > 20kN. There are several
different ways of solving this question.

Confidence Interval:
A (1 − α) × 100% confidence interval estimate for the unknown µ has end-
points
σ
x̄ ± z α2 × ( √ )
n
for example
1.4
22.5 ± z0.025 × √
5
If α = 5% is chosen: If α = 1% is chosen:
= 22.5 ± 1.960 × 0.626 . . . = 22.5 ± 2.576 × 0.626 . . .
= 22.5 ± 1.227 . . . = 22.5 ± 1.612 . . .
= (21.27, 23.73) = [20.89, 24.11]
Either way, this interval does not contain (and lies entirely above) µ = 20kN.
Therefore there is sufficient evidence to conclude, with 95% of 99% confidence,
that µ > 20kN.
Note that a one-sided confidence interval is more appropriate here, but if a
parameter value lies below a two sided confidence interval then it will certainly
lie below the corresponding one-sided confidence interval.
One sided hypothesis test:
Hypotheses:
H0 : µ = 20kN
vs.
H1 : µ > 20kN √
c = µ0 + zα × σ/ n
If α = 5% is chosen: √ If α = 1% is chosen: √
c = 20 + z0.05 × 1.4/ 5 c = 20 + z0.01 × 1.4/ 5
= 20 + 1.645 × 0.626 . . . = 20 + 2.326 × 0.626 . . .
= 20 + 1.0299 . . . = 20 + 1.456 . . .
∴ c = 21.03 (to 2 d.p.) ∴ c = 21.46 (to 2 d.p.)

x̄ = 22.5 > c x̄ = 22.5 > c

Reject H0 in favour of H1 : µ > 20kN
or
Variation of this method (in line with recipe in the book, p. 436)
1. H0 : µ = 20 kN
2. H1 : µ > 20 kN (one-tailed)
3. Level of significance: 1% and 5%
4. Critical region for acceptance of H0 :
z0.05 < 1.645 or z0.01 < 2.326

25

@seismicisolation
@seismicisolation
 √ √
5. x̄ = 22.5 kN; z = (x̄ − µ0 )/(σ/ n) = (22.5 − 20.0)/(1.4/ 5) = 3.992 . . .
6. Reject H0 , accept H1

(b) s = 1.4 kN. Here we use the t-distribution rather than the normal one. In
confidence intervals, the t0.025,4 value is 2.132 rather than 1.96 (note both
close to the magic value of 2) and we use for the 5% level
σ
x̄ ± t α2 × ( √ )
n

which gives the range

{21.17, 23.83} .
The one sided hypothesis test at the 1% level is as follows

1. H0 : µ = 20 kN
2. H1 : µ > 20 kN (one-tailed)
3. Level of significance: 1%
4. Critical region for acceptance of H0 :
t0.01,4 < 3.747 √ √
5. x̄ = 22.5 kN; z = (x̄ − µ0 )/(σ/ n) = (22.5 − 20.0)/(1.4/ 5) = 3.992 . . .
6. Reject H0 , accept H1
The p-value can be calculated as 0.000 033...
Note in this solution the inappropriate use of language "burden of proof " in
the above. If the cable is to be used in an important application involving safety,
such as a cable in mining headgear, one would need much better characterization
of the strength properties before using the results in design.

Exercise 8.29
Classical inference 2. Measurements of the output of power from a generator
are known to be normally distributed with a standard deviation of σ = 2.89
kW. Find the smallest size of sample needed so that the mean of the “random
sample” lies within 0.10 kW of the unknown mean of the distribution 95% of
the time (treat the sample as future unknown values).

Solution
The half-width of the 95% confidence interval estimate for µ is

σ
e = z0.025 × √
n
σ 2
=⇒ n > (z0.025 × )
e
= (1.960 × 2.89/0.10)2
= (56.644)2 = 3208.5 · · ·

26

@seismicisolation
@seismicisolation
 Therefore the minimum sample size is 3210.
Repeated measurement is often an effective way to reduce measurement error.
Why 95%?

Exercise 8.30
Classical inference 3. A boat manufacturer claims that a new type of hull
increases the average speed of a speedboat by more than 2 km/h over that of
a competitor’s hull design. Eleven speedboats, six with the new design and
five with the old, are tested and their speeds (in km/h) are measured on a test
course under identical environmental conditions.
New design (XA ): 36 42 38 36 37 39
Old design (XB ): 31 33 30 37 34
Conduct an appropriate hypothesis test at a level of significance of 5%. Is
there sufficient evidence to accept the designer’s claim? What assumptions have
you made?

Solution
Calculation of summary statistics:
New design speeds xA Old design speeds xB x2A x2B

42 30 1764 900
36 37 1296 1369
38 33 1444 1089
37 31 1369 961
39 34 1521 1156
36 1296
228 165 8690 5475
nA = 6 nB = 5
vA = 5 vB = 4
x̄A = 38 x̄B = 33
s2A = 5.2 s2B = 7.5

The designer’s claim is that µA − µB > 2 (which is where the "burden of

proof" lies).
Test
H0 : µA − µB = 2 vs. H1 : µA − µB > 2
We assume that the two populations are independent, nearly normal and
have a common population variance.
Then ­ ®
X A − X B = µA − µB
and
£ ¤ σ2 σ2
V XA − XB = A + B
nA nB

27

@seismicisolation
@seismicisolation
σ 2 is unknown, so that we use the pooled sample variance

ν A s2A + ν B s2B 5 × 5.2 + 4 × 7.5 56

s2p = = = = 6.2̇
νA + νB 5+4 9
The sample standard error is
s µ ¶ s µ ¶ p
2
1 1 1 1
sp + = 6.2̇ + = 2.28̇14̇
nA nB 6 5

The total number of degrees of freedom is v = vA + vB = 9

Hypothesis test
1. H0 : µA − µB = 2
2. H1 : µA − µB > 2
3. Upper tail test. 5%
4. Critical region for acceptance of H0 : t0.050,9 < 1.833
5. t = (xA −x B )−∆0
s.e. = (38−33)−2
√ 3
= 1.5104... ' 1.986
2.2814
6. Reject H0 , accept H1
The designer’s claim is supported (incorrectly stated as "is true" in some
sources).
[Note: if we had taken α = 0.01, then we would have reached the opposite
conclusion, because the p-value is .032 (to 3 d.p.).]
Note the absurdity of the statement of H0 .

Exercise 8.31
Classical inference 4. Samples are drawn at random from two separate popula-
tions, with no members in common, resulting in the following summary statis-
tics:
X : n = 50, x̄ = 643, sX = 26
Y : n = 90 ȳ = 651, sY = 32
Are these data consistent with the hypothesis that the two population means
are equal? Calculate the p-value and comment.

Solution
The "burden of proof" (!!) is on the alternative hypothesis that the two popu-
lation means are not equal. Therefore test
H0 : µX − µY = 0 vs. H1 : µX − µY 6= 0
>From the question, the random quantities X and Y are independent, but
we cannot assume equality of the variances. A test for this can be conducted
(see item 8 on page 441 of book). The data are consistent with σ 2X = σ2Y .
The Central Limit Theorem leads to
µ ¶ µ ¶
σ2 σ2
X ∼ N µX , X , Y ∼ N µY , Y
nX nY

28

@seismicisolation
@seismicisolation
 and µ ¶
σ2 σ2
X −Y ∼N µX − µY , X + Y .
nX nY
Approximate the unknown σ 2X and σ2Y by the point estimates s2X and s2Y
respectively. If H0 is "true" (!!), then to an acceptable approximation,
µ ¶
262 322 ¡ ¢
X − Y ∼ N 0, + = N 0, 24.897 .
50 90
Hypothesis test
1. H0 : µX − µY = 0
2. H1 : µX − µY 6= 0
3. α = 5%, α/2 = 2.5%
4. Accept H0 if −1.96 < z < 1.96
5.
(x − y) − ∆o (643 − 651) − 0
zobs = q 2 = √ = −1.603 . . .
sX s2Y 24.897
nX + nY

6. Accept H0 , reject H1
The p-value is

Pr [|Z| > |zobs |] = 2Pr [Z < −1.603 . . .]

' 2Φ(−1.60) = 2 × 0.0548
= 0.110 > 0.10

The data are consistent with equal population means, (although the p-value
suggests that the agreement is not close).

Exercise 8.32
Classical inference 5. Measurements of the torque output from an electric motor
have been found to follow the normal distribution. The output is governed by
a controller, and a new controller is being tested. Seven different motors are
chosen at random and subjected to tests with the new and old controllers giving
the following results. Carry out an hypothesis test on the null hypothesis that
the torques using the new method are on average at least 2 units less for the
new method, as compared to the old.
Motor 1 2 3 4 5 6 7
New Method 6.01 3.16 4.43 5.25 6.12 5.75 2.21
Old Method 8.88 6.22 7.46 7.83 8.83 8.19 5.64
See also exercise 11.10 for the use of linear regression in this kind of analysis.

29

@seismicisolation
@seismicisolation
Solution
Let X and Y represent the random quantities for the two methods respectively.
x and y in the table are pairs of measurements of different methods applied to
the same seven machines. The random quantities X, Y are therefore likely to
be linearly dependent. A quick visual scan of the data supports the suspicion
of strong correlation between the two samples.
A paired two sample t test should be used. Let di = xi − yi ; then the
hypotheses to be tested are
H0 : µD = −2 vs. H1 : µD < −2
The observed values of D are
di : −2.58, −3.06, −3.03, −2.71, −2.44, −3.43, −2.87
Summary statistics:

n = 7
X
d = −20.12
X
d2 = 58.5004
v = 6
d = −2.874 . . .
2
sD = 0.1116 . . .

Hypothesis test
1. H0 : µD = −2
2. H1 : µD < −2
3. α = 0.01
4. Reject H0 if t < −t0.01,6 = −3.143
5.
d−µ −2.874 . . . − (−2) −0.874 . . .
t = ³ ´o = q = = −6.923
s
√D 0.1116... 0.126 . . .
n 7

6. Reject H0 , accept H1
The p-value is 0.000225, small enough to reject H0 , and accept H1 at any
reasonable value of α.
There is sufficient evidence to conclude that the new method has reduced
the range by more than 2 units.
Simple linear regression of X on Y . The summary statistics are:
P1 = n2 = 7
n
P x = 32.93 =⇒ x = 4.704285 . . .
P y 2= 53.05 =⇒ y = 7.57857
2
...
P 2x = 168.6941 =⇒ s X = 2.296995249 ...
2
P y = 411.3579 =⇒ sY = 1.552447619 . ..
xy = 260.7758
P P
nSxx = n P x2 − ( P x)2 P = 7 × 168.6941 − (32.93)2 = 96.4738
nSxy = n xy − ( x)( y) = 7 × 260.7758 − 32.93 × 53.05 = 78.4941

30

@seismicisolation
@seismicisolation
 P P
nSyy = n y 2 − ( y)2 = 7 × 411.3579 − (53.05)2 = 65.2028
Sxy
β̂ 1 = Sxx = 78.4941...
96.4738... = ¢0.813631265 . . .
1
¡P P
β̂ 0 = n y − β 1 x = 17 (53.05 − 0.813 . . . × 32.93) = 3.75101749
ˆ

Therefore Pthe line ofPbest fit by P least squares is y = 3.751 + 0.8136x

SSE = y 2 −ˆ β 0 y −ˆ β 1 xy = 411. . . . − 3.75 . . . × 53.05 − 0.813 . . . ×
260. . . .
= 0.1910782
SST = Syy = 65.2028/7 = 9.314685714 . . .
SSR = SST − SSE = 9.123607514
=⇒ R2 = SSR 9.123...
SST = 9.314... ≈ 97.95%
=⇒ r = .98969 . . .
ANOVA Table:
d.f. SS MS f
R 1 9.123607514 9.123607514 238.74. . .
E 5 0.1910782 0.03821564
T 6 9.314685714
or
S2 (78.4941)2
r2 = Sxxxy Syy = 96.4738×65.2028 = 0.979486
Almost 98% of the variation in Y is explained by the simple linear regression
model of part (c) above.
Very strong correlation =⇒ should not use the unpaired t-test

Exercise 8.33
Classical inference 6. In the manufacture of a member carrying tension, two
types of alloy are tested. It was found that 22 out of 40 specimens made of alloy
A had strengths greater than a specified value whereas 29 specimens out of 42
made with alloy B had strengths greater than the specified value. Do a test at
the 5% level of significance to examine whether the difference between the two
proportions is significant.

Solution
Let pA = proportion of alloy A that has ultimate strength ≥ 34 ksi
and pB = proportion of alloy B that has ultimate strength ≥ 34 ksi.
Then pbA = 22
40 = 0.55 and
29
h pbB = 42 i' 0.6905
The point estimate of V PbA − PbB is

pbA qbA pbB qbB

+
nA nB
1 22 18 1 29 13
= × × + × ×
40 40 40 42 42 42
396 377
= + ' 0.01128.
64000 74088

31

@seismicisolation
@seismicisolation
 A 95% confidence interval for pA − pB is

pbA − pbB ± z0.025 (s.e.)

µ ¶
22 29 √
= − ± 1.960 0.0112 . . .
40 42
= −0.1405 ± 0.2081

The confidence interval is −0.349 < pA − pB < +0.068 (to 2 d.p.)

Not significant because pA = pB is included in the 95% confidence interval.
Why is 5% used? This might be an important design issue meriting more
detailed analysis and testing.

Exercise 8.34
An engineer is studying the fatigue life, Y, in cycles of stress, of a steel connector
device. It is convenient to consider the quantity

X = log Y.

Previous measurements have shown that X is a normally distributed random

quantity with a standard deviation of 0.20. The mean A of this distribution is
the subject of study in the present investigation. Past experience has resulted
in the assignment of a normal distribution to A with a mean of 5.80 and a
standard deviation of 0.10. The following ten measurements have subsequently
been taken of X.
5.61, 5.24, 5.12, 5.40, 5.14, 5.38, 5.40, 5.46, 5.41, 5.67
Obtain the posterior distribution of A, and sketch the prior and posterior
distributions. What is the name for the distribution of Y ?

Solution
Dr. I. J. Jordaan’s solution
Model for the random quantity A:
Prior: N (5.8, (0.10)2 )
Measurements:
p x = 5.383
s = (0.28981/9) = 0.1794 (consistent with σ = 0.20).
Prior weight: wo = σ12 = (0.1)1
2 = 100
o
n 10
Data weight: wd = σ2 = (0.2)2 = 250
The sum of the prior and data weights is wf = wo + wd = 350
σ2f = w1f = 0.002857
σf = 0.05345
µf = wd x+w
wf
o µo
= 250×5.38+100×5.8
350 = 5.50

32

@seismicisolation
@seismicisolation
 8

6
Posterior

5
Prior
fX(x)

1 S.D. S.D.
= 0.06 = 0.10

0
5.2 5.4 5.6 5.8 6 6.2
X = log (Y)

Figure 8:

Dr. G. H. George’s solution

Prior distribution:
³ ´ ³ ´
σ2 2
X ∼ N µo , no = N 5.8, (0.10) 10 = N (5.8, 0.001)
where the prior variance is a measure of the precision with which we believe
we knew µ prior to taking the new random sample.
New evidence:
Sample size = n = 10
Sample mean = x = 5.383 p
Sample standard deviation = s = (0.28981/9) = 0.1794,(consistent with
σ = 0.20).
Therefore use σ = 0.20 in the calculation for the posterior mean.
Calculate2
nxσ +µ σ 2 2
+5.8×(0.20)2
µ∗ = nσo2 +σo2 = 10×5.383×(0.10)
10×(0.10) 2 +(0.20)2 = 0.7703/0.14 = 5.5021 . . .
o
Posterior distribution:
X ∼ N (µ∗ , (σ ∗ )2 ) = N (5.502 . . . , 0.002 . . .)
The mean has shifted down by 0.3 and √ the precision (as measured by the
standard error) has improved by a factor of ( 3.5) (more than an 87% improve-
ment).
(b) see figure 8

(c)Log-normal.

33

@seismicisolation
@seismicisolation
Solutions to Chapter 9
Exercise 9.1
Verify that equation (9.15) follows from equation (9.14).

Solution
Xn µ ¶
n
F t (1 − F )n−t
t=r t
n! n!
= F r (1 − F )n−r + F r+1 (1 − F )n−r−1
r!(n − r)! (r + 1)!(n − r − 1)!
n!
+··· + F n (1 − F )0
n!(n − n)!

d n!
→ rF r−1 (1 − F )n−r · f
dz r!(n − r)!
n!(n − r) r n!(r + 1)
− F (1 − F )n−r−1 · f + F r (1 − F )n−r−1 · f
r!(n − r)! (r + 1)!(n − r − 1)!
− · · · + · · · etc.
Apart from the first, the terms cancel in pairs as illustrated in the second line
above. This continues for the remaining pairs.

Exercise 9.2
Basic exercise in extremes to become familiar with the process.
(a) Carry out Monte Carlo simulations (Chapter 11) to reproduce the ex-
tremal distributions obtained in section 9.4. The Monte Carlo procedure
should follow the procedure outlined under “Random Blows”. Histograms
and fitted distributions should be produced.
n
(b) Repeat (a) by using the relationship FZ (z) = FX (z) based on the parent
cdf, thereby obtaining graphical representations of extremal distributions
as in section 9.4.

Comment
It is most helpful to the learning process to carry out these simulations.

Exercise 9.3
Find a second proof of the return period relationship 1/p.

@seismicisolation
@seismicisolation
Solution
We may write equation (9.37) as
X
hV i = p vq v−1
¡ ¢
= p 1 + 2q + 3q 2 + 4q 3 + · · ·
d ¡ ¢
= p q + q2 + q3 + q4 + · · ·
dq
µ ¶
d q
= p
dq 1 − q
1
= p·
(1 − q)2
1
= .
p

Exercise 9.4
Show that equation (9.14) reduces to the same form as (9.6) for the lowest
endpoint (r = 1) .

Solution
Equation (9.14) is as follows
µ ¶
¡ ¢ Xn n t
¡ ¢£ ¡ ¢¤n−t
FX(r) x(r) = FX x(r) 1 − FX x(r) .
t=r t

Put r = 1 and x(r) = w:

Xn µ ¶
n n−t
FW (w) = F t (w) [1 − FX (w)] . (A)
t=1 t X
Now µ ¶
Xn n n−t
F t (w) [1 − FX (w)] = 1,
t=0 t X
the sum of all binomial probabilities. Then
Xn µn¶
[1 − FX (w)]n + F t (w) [1 − FX (w)]n−t = 1. (B)
t=1 t X

Substitute (A) into (B) and we find

[1 − FX (w)]n + FW (w) = 1 or
FW (w) = 1 − [1 − FX (w)]n

which is (9.6).

@seismicisolation
@seismicisolation
Exercise 9.5
Find a closed form solution for the pdf of the extreme wave (section 9.4.3) given
a Rayleigh parent. Plot behaviour with respect to log n.

Solution
Use the dimensionless form
¡ ¢
fX (x) = 2x exp −x2 ,

with £ ¡ ¢¤
FX (x) = 1 − exp −x2 .
Then for the maximum of n occurrences, denoted Zn ,
n
FZ (zn ) = FX (zn )

(9.5) and differentiate to find that

n−1
fZ (zn ) = nfX (zn ) FX (zn )

(generally!). Applying to our example

¡ ¢£ ¡ ¢¤n−1
fZ (zn ) = 2nzn exp −zn2 1 − exp −zn2 .

See page 172, Ochi.

Exercise 9.8
An event (rams of an arctic vessel, voltage surges of a certain magnitude) oc-
curs 20 times in a year. The distribution of its magnitude X is exponential,
λ exp (−λx). Find the magnitude in terms of the mean 1/λ such that the return
period is 100 years. Assume iid.

Solution
(1 − pe )20 = 0.99
Approx. (1 − pe )20 ' 1 − 20pe = 0.99
pe = 5.02 × 10−4
pe = 0.01
20 = 0.0005
e−λx = 5.02 × 10−4
ln 0.0005 = −7.60
ln e−λx = ln(5.02 × 10−4 )
−λx = −7.60
−λx = −7.596
x = 7.60( λ1 )
x ' 7.60( λ1 )

@seismicisolation
@seismicisolation
Exercise 9.9
See Corrigenda
An approximation that is often useful for very small p (¿ 1/n) is
(1 − p)n ' 1 − np.
(a) Apply this in Exercise 9.8.
(b) Equation (9.49) states that
µ ¶n
n 1 1
FZ (xn ) = FX (xn ) = 1− → ,
n e
in other words the n-year load on a structure has a 63% chance of being
exceeded in n years for large n. A designer wishes therefore to choose
a more conservative value. He then considers a value z̆ with a small
probability of exceedance on the n-year load pne such that
pne = 1 − FZ (z̆)
rather than the mode. Then if X is the parent,
[FX (z̆)]n = 1 − pne , or
pne
FX (z̆) = (1 − pne )1/n ' 1 − ,
n
using a variation of the rule (9.172). Obtain pne in terms of an , bn and
α for a Weibull parent distribution based on equation (9.119), with X
replacing Z: · µ ¶α ¸
x − an
FX (x) = 1 − exp − .
bn
An engineer in offshore practice would realize that a load factor is usually
applied to the 100-year load, giving a
result that approaches the situation described.

Solution
(a) See RHS of solution to Exercise (9.8).
(b)
· µ ¶α ¸
x − an
FX (x) = 1 − exp − ,
bn
with a ≤ x < ∞.

n
[FX (z̆)] = 1 − pne
1 pne
Fx (z̆) = (1 − pne ) n ' 1 −
n
pne = n [1 − Fx (z̆)]

@seismicisolation
@seismicisolation
 · ¸
z̆ − an α
pne = n exp (− )
bn
h ³ ´i ½ h ³ ¾
z̆ − an pne α 1
pne ´i α1
= − ln =⇒ z̆ = an + bn − ln
bn n n
For example, if α = 1, we have the exponential distribution. If an = 0, and
bn = µ (the mean), h ³ p ´i
ne
z̆ = µ − ln ,
n
so that h ³ p ´i
ne
z̆ = 2.303µ − log .
n
For n = 100, pne = 1/10
z̆ = (2.303µ) (3) ;
the factor (3) would be (2) if pne = 1.

Exercise 9.10
Find the mode of the Rayleigh distribution
µ 2¶
2x x
fX (x) = exp − 2 .
σ2 σ

Solution
Put
dfX (x)
= 0;
dx
then one finds that
σ
x̂ = √ ,
2
at which point the probability density is
r
1 2
fX (x̂) = .
σ e

Exercise 9.11
Consider the following form of the Pareto distribution
FX (x) = 1 − xk ,
x ≥ 1, k > 0,
which has been used for example to model relative income distributions. Show
that L’Hôpital’s rule does not apply, and that the critical quotient is equal to
k
.
k+1

@seismicisolation
@seismicisolation
Solution
See p.151, Gumbel

Exercise 9.12
Compute and plot the intensity function and critical quotient for the following
distributions: normal, logistic and lognormal distributions. Compare to that of
the exponential distribution.

Solution
See p. 22 Gumbel

Exercise 9.13
Show that the mode of the Gumbel distribution
· µ ¶ ¸
1 z−α z−α
fZ (z) = exp − exp − −
β β β

is ẑ = α. Show also that this is approximately equal to the characteristic extreme

zn .

Solution
1 −z−α −z − α
fz (z) = exp(−e β ) exp( )
β β
1 −z−α −z − α (−e −z−α
β + −z−α )
= exp(−e β + )e β
β β
Let
z−α −y
y= → fy (y) = e−(e +y)
β
dfy (y) −y
= e−(e +y) (e−y − 1)
dy
Put = 0:
zb − α
e−by = 1; yb = 0 → =0
β
zb = α

(Check for max.—use standard method).

Now refer to Example 9.4

@seismicisolation
@seismicisolation
 1
zn = α − β ln[− ln(1 − )]
n
1
' α − β ln for large n
n
= α + β ln n

Max. for n of above:

z−α−β ln n
−y −y −(y−ln n) −
β
FY = [e−e ]n = e−ne = e−e ; −→ e−e
z = α + β ln n

Exercise 9.14
Compare the results of using Table 9.1 with distributions obtained from the
parent using equation (9.5).

Solution
Suggest to use Matlab and numerical methods

Exercise 9.15
Derive the type III (Weibull) distribution from Type I and Type II. Explain
which parameters depend on n.

Solution
See Gumbel p. 157

Exercise 9.16
¡ ¢
Calculate the hazard function for the distribution FX (x) = 1 − exp −xk ,
where k is a constant.

Solution
¡ ¢
kxk−1 exp −xk
µ (x) = = kxk−1
exp (−xk )

@seismicisolation
@seismicisolation
 Figure 1:

Exercise 9.17
Plot equation (9.114)
· ¸
dbn 1 1
= 2 −1
dn n fX (an ) Q (an )
for the distribution in Example 9.10, and discuss the behaviour of bn with an .

Solution
Q and FX (an ) are given in Example 9.10
· ¸
dbn 1 1
= 2 −1
dn n fx (an ) Q(an)
Also see figure 1

Exercise 9.18
Consider Example 9.11. Let the following vector
 
 1 
x= 2
 
3

@seismicisolation
@seismicisolation
represent the the position of the three elements. We took rR as r (the stress in
element 1). Find a matrix φ that transforms this into
 
 1 
1
r ,
 2 
1

giving the relative values of s in each element.

Solution
If φ is the matrix  
1 0 0
 0 1
0 ,
4
1
0 0 3

it will, if multiplied into x, give the desired result.

Exercise 9.19
Consider the failure criterion
µ ¶d µ ¶d µ ¶d
σ 11 σ 22 σ 33
s= + +
σ0 σ0 σ0

where σ 11 , σ 22 and σ 33 are principal stresses, and d and σ 0 are constants. Show
how this can be used in Weibull’s strength distribution using an equation of the
kind (9.122).

Solution
See p. 615 Freudenthal.

Exercise 9.20
Find the reduced volume v∗ , as in Example 9.12, for a cylindrical bar under
torsion; for a prismatic beam simply supported under point load; and for a
prismatic beam, simply supported, loaded symmetrically by two point loads.
Treat the reduced stress as the greatest principal stress.

Solution
Pure bending
R h/2 ¡ 2y ¢α
v∗ bl 0 h dy 1
= =
v blh 2(α + 1)

@seismicisolation
@seismicisolation
 Prismatic beam under point loading (3-point bending)
R b/2 R c R l/2 ¡ 2xy ¢α
v∗ 2 −b/2 0 0 cl dxdydz 1
= =
v 2bcl 2(α + 1)2

Prismatic beam under 2-point loading (4-point bending)

R b/2 R c R l/3 ¡ 3xy ¢α R b/2 R c R l/2 ¡ y ¢α

v∗ 2 −b/2 0 0 cl dxdydz + 2 −b/2 0 l/3 c dxdydz (α + 3)
= =
v 2bcl 6(α + 1)2
Cylindrical bar under torsion
R r ¡ ¢α
v∗ 2πl 0 rc dc 2
= 2
=
v 2πr l r (α + 1)

Exercise 9.21
Consider the following non-homogeneous failure condition. A bar is stressed
under uniform tension. Two aspects are to be considered. The first relates to
a population of surface flaws, and the second to flaws within the volume of the
material. How would you approach this problem?

Solution
For a bar under uniform tension with a population of surface flaws and flaws
within the volume, the strength distribuition function FR (x) would be a func-
tion of both surface area and volume according to:
· ¸
A V
FR (x) = 1 − exp − ln (1 − F1 (x)) − ln (1 − F2 (x))
A0 V0
where:
A = surface area of the bar
A0 = reference surface area corresponding to F1 (x)
F1 (x) = strength distribution function for reference surface area A0
V = volume of the bar
V0 = reference volume for F2 (x)
F2 (x) = strength distribution function for reference volume V0

Recalling that for small arguments we can approximate ln (1 − F (x)) as

µ ¶α
x − x0
ln (1 − F (x)) ≈ −
xc
Based on this representation we can write the strength distribution as:

10

@seismicisolation
@seismicisolation
 h ³ ´ ³ ´ i
x−x01 α1 x−x02 α2
1 − exp − AA0 xc1 − V
, x ≥ max {x01 , x02 }
h ³ V0 ´αxc12i
x−x
FR (x) = 1 − exp − AA0 01
, x01 ≤ x ≤ x02
h ³ xc1 ´α2 i
V x−x02
1 − exp − V0 xc , x02 ≤ x ≤ x01
2

where: =
x01 = lower limit on strength based on reference area A0
x02 = lower limit on strength based on reference area V0
and xc1 , xc2 , α1 , α2 are positive constants.

Exercise 9.22
Show that the extremal distributions do not possess conjugate prior distributions
or sufficient statistics.

Solution

In section 8.2.1, we analyzed the case of ‘repeated trials of the same phenom-
enon’. It was shown that for an example with n = 10, a typical string of
successes and failures is given by string (8.2). The concept of exchangeability
tells us that if the results in this string are exchangeable, the order of 0s and 1s
are of no significance. The only relevant information is the number of successes
and failures, which can be termed sufficient statistics. The sufficient statistics
are therefore the minimum statistics required to describe a statistical model.
If we take as an example the double exponential distribution:
½ · µ ¶¸ µ ¶¾
1 z−a z−a
fZ (z) = exp − exp − −
b b b
The likelihood function associated with this distribution is:

µ ·µ ¶ µ µ ¶¶¸¶
P zi − a zi − a
l(a, b | z1 , z2 , ..., zm ) = −bm exp − + exp − .
i b b

This shows that there are no two-dimensional sufficient statistics for (a, b),
and that consequently, fZ (z) possesses no natural conjugate prior. The same
problem arises for the Weibull distribution and others.

Exercise 9.23
Analyze how you would take into account parameter uncertainty in equation
(9.148) for the scale effect, and in (9.151) for the variance.

11

@seismicisolation
@seismicisolation
Solution
The equation (9.148) for the scale effect,
µ ¶ α1
hRi1 v2
= ;
hRi2 v1

and (9.151) for the variance

"¡ ¢ # 12
Γ 1 + α2
γ= ¡ ¢ −1 ,
Γ2 1 + α1

were derived on the basis that the parameter α is known. Consider the much
more natural assumption that the flaw structure of the material consists of a
series of exchangeable configurations. Taking the analogy of links in chain;
the expression m (r) was used to represent the expression {− ln [1 − FT (r)]}
within equation (9.121), and in particular the power-law equation m (r) =
α
[(r − r0 ) /r1 ] constituted the approximation. For the two expressions above
for scale effect and variance, r0 = 0, so that α is the only parameter in the dis-
tribution. To consider the strengths as exchangeable between elements rather
than iid, α must be treated as a random quantity. Thus the scale effect becomes
Z µ ¶ α1
hRi1 v2
= fA (α) dα.
hRi2 v1

See section 9.22 regarding parameter uncertainty and return periods.

Exercise 9.24
Consider an extremal process that is modelled as a Poisson process. Treat the
arrival rate as being random. Develop a “peak over threshold” (POT) model
that includes parameter uncertainty in the arrival rate and an application of the
model.

Solution
This can be treated as in sections 8.5.4 combined with 9.8. For a given threshold,
the probability distribution of the random quantity exceeding this threshold can
be determined. For example, the exponential distribution will have a simple
exponential tail of the form
exp (−αx) .
This, combined with the Poisson arrival process gives the double exponential
distribution of section 9.8.
Applications of the POT method are common in environmental studies of
wind and wave and are easy to track down.

12

@seismicisolation
@seismicisolation