Hae PROBABILITY FOR

ENGLNE E RING
WIT =
APPLPCALON;
TO RELIABILITY

LAVON B. PAGE? |
PROBABILITY FOR ENGINEERING
with Applications to Reliability
ELECTRICAL ENGINEERING, COMMUNICATIONS,
AND SIGNAL PROCESSING
Raymond L. Pickholtz, Series Editor

Computer Network Architectures

Anton Meijer and Paul Peeters

Spread Spectrum Communications, Volume I

Marvin K. Simon, Jim K. Omura, Robert A. Scholtz, and Barry K. Levitt

Spread Spectrum Communications, Volume II

Marvin K. Simon, Jim K. Omura, Robert A. Scholtz, and Barry K. Levitt

Spread Spectrum Communications, Volume III

Marvin K. Simon, Jim K. Omura, Robert A. Scholtz, and Barry K. Levitt

Elements of Digital Satellite Communication: System Alternatives, Analyses and

Optimization, Volume I
William W Wu
Elements of Digital Satellite Communication: Channel Coding and Integrated
Services Digital Satellite Networks, Volume II
William W Wu
Current Advances in Distributed Computing and Communications
Yechiam Yemini
Digital Transmission Systems and Networks, Volume II: Applications
Michael J. Miller and Syed V. Ahamed
Transmission Analysis in Communication Systems, Volume I
Osamu Shimbo
Transmission Analysis in Communication Systems, Volume IT
Osamu Shimbo

Spread Spectrum Signal Design: LPE and AJ Systems

David L. Nicholson

Digital Signal Processing Design

Andrew Bateman and Warren Yates
Probability for Engineering with Applications to Reliability
Lavon B. Page

OTHER WORKS OF INTEREST

Local Area and Multiple Access Networks
Raymond L. Pickholtz, Editor

Telecommunications and the Law: An Anthology

Walter Sapronov, Editor
PROBABILITY FOR ENGINEERING
with Applications to Reliability

Lavon B. Page
North Carolina State University

COMPUTER SCIENCE PRESS

Library of Congress Cataloging-in-Publication Data
Page, Lavon B.
Probability for engineering with applications to reliability
Lavon B. Page.
p.. cm.
Bibliography: p.
Includes index.
ISBN 0-7167-8187-5
1. Probabilities. 2. Engineering mathematics. I. Title.
QA273.P19 1988 88-29926
519.2—dc19 CIP

Copyright © 1989 by Computer Science Press, Inc.

Printed in the United States of America
All rights reserved. No part of this book may be reproduced in any form
including photostat, microfilm, xerography, and not in information storage
and retrieval systems, without permission in writing from the publisher, except
by a reviewer who may quote brief passages in a review or as provided in the
Copyright Act of 1976.
Computer Science Press
1234567890 RRD 7654321089
 f 3 =

fe 2 Ese ss ee
a S2aea ae of 5 - 2
ean oa =
_ ‘ =
my =e i 2 =. z a
x o>
4 a =< F tt we = me i a * Se ee L 3

ot s ——— ae Se a
o uP x — : 4 .
4 i —- - +

eI = a ile Lar
= eee a ae rs
a, paeee
a z 2a - _
“ =

om = Ef sn - ee a a

wd os : a
a Zs

———s : a 2
» q = >)
oe coh Pw Av
; ,
Sta Ge payee eee ae
for Jo Ellen—
, . & , I
4 Peg
e a eee
8 .

yy Te

‘ ‘ = iy
Seana ; -
a
; = ca - | -
re na ra i « i
a
ca Aree “ay
oe Ranch can ye. FaGipe: dl

F = in
aS
 Contents

Chapter : The Basics

Sets and Set Operations
The Sample Space
Basic Properties of Probabilities
Conditional Probability
Tree Diagrams and Bayes’ Theorem
More on Independence
ERE
5
UL
US
SEE
STa
enO
co
IS Infinite Sample Spaces
“EACwOTL
Problems

Chapter 2: Applications
rea Circuits
aie Networks
2.3 Case Study: Two Network Reliability Problems
2.4 Fault Trees
Problems

Chapter 3: Random Variables

3.1 Discrete Random Variables and Mass Functions
3.2 Continuous Random Variables and Density Functions
3.3 Distribution Functions
3.4 Expected Value and Variance
3.0 Functions of a Random Variable
Problems

Chapter 4: Discrete Models

4.1 The Binomial Distribution
4.2 The Geometric Distribution
4.3 Discrete Uniform Random Variables
4.4 Poisson Random Variables
4.5 Hypergeometric Random Variables
4.6 Probabilities Based on Observed Data
Problems

Vii
Vili Contents

Chapter 5: Continuous Models 103

5.1 Continuous Uniform Random Variables 103
5.2 Exponential Random Variables 104
5.3 Normal Random Variables 107
5.4 Evaluating the Standard Normal Distribution Function Wee
Problems TAS

Chapter 6: Joint Distributions 127

6.1 Joint Probability Mass Functions 127
6.2 Joint Density Functions 130
6.3. Functions of Two Random Variables 139
6.4 Sums of Random Variables 143
6.5 Conditional Probabilities and Random Variables 152
6.6 The Central Limit Theorem 160
Problems 162

Chapter 7: Stochastic Processes VW

7.1 Independent Trials Processes Twa
7.2 A One-Dimensional Random Walk 173
7.3 Poisson Processes 174
7.4 Poisson or Binomial? 182
7.5 Sample Functions 184
7.6 The Autocorrelation Function 186
7.7 Stationary Stochastic Processes 188
7.8 Ergodic Properties 193
Problems 195

Chapter 8: Time-Dependent System Reliability 199

8.1 Reliability that Varies with Time 199
8.2 Systems with Repair 205
Problems 209

References 213

Appendix A: Answers, Partial Solutions, and Hints

to Selected Problems 21,5

Appendix B: Values of Normal Distribution Function 229

Index 230
 Preface

This book is an attempt to merge the best of two worlds. It brings precise
mathematical language and terminology into an arena of modern engineering
reliability probems. These include such topics as simple analysis of circuits, design
of redundancy into systems, reliability of communication networks, mean time to
failure, time-dependent system reliability, and uses of probability in recursive
solutions to real-world problems.
All books reflect the point of view of the author. Texts on probability written
by engineers usually move as quickly as possible to the applications of most interest
to the writer. The mathematical foundations are presented casually and quickly, if at
all, and understanding is gleaned from examples rather than from definitions and
theorems. Mathematicians, on the other hand, tend to write probability books that
dwell on topics such as combinatorics, labeling problems, allocation schemes, and
limit theorems. Such books present little that seems relevant to the world of an
undergraduate engineering student.
In teaching probability to engineering students for more than a decade, I have
seen the difficulties caused by both of these kinds of textbooks. Students with
good intuition are often handicapped from never having come to grips with the basic
concepts. On the other hand, the importance of fundamental concepts escapes the
student unless some application is in sight. For this reason, significant applications
appear earlyin this text. Chapter 2, for example, illustrates how the idea of
conditional probability can be blended with algorithmic problem solving to develop
tools for reliability analysis of complex systems such as circuits, communication
networks, and chemical reactors. These contemporary problems mix probability
and discrete mathematics, and they require a solid understanding of the basics. But
the student is rewarded with a sense of relevance that isn’t matched by drawing
balls out of an urn.
The essentials of an introduction to probability are found in Chapters 1, 3, 4,
the first three sections of Chapter 5, and the first four sections of Chapter 6. The
remaining sections of Chapter 6 introduce conditional density functions, conditional
expectation, and the central limit theorem. (The De Moivre-Laplace version of the
central limit theorem appears in Chapter 5.) Chapter 7 gives a brief introduction to
stochastic processes, with heavy emphasis on the Poisson process. Chapter 2
x Preface

applies the basic ideas of probability to reliability problems involving a variety of

complex systems. Chapter 8 shows how to extend the concept of reliability to
systems having components whose reliabilities vary with time.
Proper mathematical language and detail are important in formulating correct
mathematical models, and this book reflects that fact. However, the book is
practical rather than theoretical, and logical explanations or proofs are given only
where they are an aid to intuitive understanding.
The real world confronts us with some easy problems and some hard ones.
So does this book. Asa result, it does not seem appropriate to treat all problems the
same with regard to hints given or answers provided. This has prompted the
inclusion of the section of answers, partial solutions, and hints to selected problems
that appears as Appendix A. My hope is that students will give serious thought to
problems before consulting this appendix, and that after consulting it they will
think seriously about alternative solutions suggested or fill in missing details.
My own research in recent years has shifted into the area of reliability
analysis and risk assessment. And, just as with other books, this one reflects the
author’s own interests. The common theme in this book is the search for reliability
in an increasingly complex world. News reports on topics as diverse as arms
control, strategic defense initiatives, hazardous waste dumps, and nuclear reactors
all remind us of the fact that uncertainty (or lack of reliability) is central to many of
the pressing issues of our day.

Lavon B. Page
September 1988
PROBABILITY FOR ENGINEERING
with Applications to Reliability
 ca seneOR
LA fired
x
ar! Oe
. 7

Ore i¢ Pecake
shcong

ies ‘is
aie
Peal :
> Basse hsThis)
= a)
ue
iu
me, Ria

- hi -
*
Pa)
4 we
: »

‘:

a -
wire)
es,
 Chapter 1: The Basics

In January 1986 the space shuttle Challenger exploded in midair. The space
shuttle had previously been considered so safe by NASA that plans were afoot to
send plutonium powered modules into orbit. The previous year an accident at the
Union Carbide plant in Bhopal, India, had killed thousands of people in what was
at the time the worst industrial accident in history. A few months later, the Soviet
reactor at Chernobyl was to run amok and burn out of control for days, while
spewing radioactivity into the atmosphere. The Soviets had thou ght the probability
of such a major accident to be extremely low.
As long as such accidents continue, and there is every indication that they
will, there will be a lot of interest in trying to determine the reliability of things like
space vehicles, nuclear and chemical reactors, and such ordinary devices as
automobiles, garage door openers, or heating systems. Skepticism now greets
claims that something or other is “less likely to happen to you than being hit by
lightning,” or has “only 1 chance in 10,000 of occurring in the next 50 years.”
Often such claims have been based more on hope than on science.
Estimates of the reliability of equipment or complex systems depend heavily
on the field of mathematics known as probability. Probability can be abused, just
as can most tools. The best mathematical model can’t produce true answers if
incorrect or naive assumptions are fed into it. Even at a fairly elementary level,
however, probability opens the door to the investigation of complex systems and
situations. If we want to answer such questions as “What were the chances of that
happening?” or “How much do we expect to gain if we make that decision?”, the
answer will have to be expressed in the language of probability. The purpose of
this book is to present the basics of that language and to show its application to a
variety of meaningful examples, with an emphasis on the idea of reliability.
Interest in probability blossomed around the gambling tables of Europe
hundreds of years ago, though much earlier references can be found in Hebrew and
Chinese. Many games can be analyzed by looking at the possible outcomes of an
experiment, such as rolling a pair of dice or dealing some cards from a deck.
Frequently something about the situation suggests that the various possible
@ Chapter 1: The Basics

outcomes should be considered equally likely. For example, the symmetric shape
of a six-sided die suggests that the six outcomes are equally probable, and the
purpose in shuffling a deck of cards before dealing is to try to approximate a
situation in which one arrangement of the deck is just as likely as another.
The concept of equally likely outcomes leads to a natural concept of
probability. For example, since there are 13 hearts in a deck of cards, there are 13
chances out of 52 that an arbitrary card drawn from a deck will be a heart.
Considering the ratio of these two numbers gives the intuitively satisfying
conclusion that the probability of drawing a heart when a card is drawn from a deck
should be 13/52, or 1/4.
A more skeptical person might argue that the only way to test probabilities is
to experiment. For example, given a coin of unknown characteristics, the only way
to determine the probability of the coin coming up heads is to toss it many times and
see what happens. A mathematician might look at the situation in this way: If H,,
is the number of heads obtained in the first n tosses of a sequence of tosses, the
probability of heads might be taken as
ah
noo N

Of course there’s no way actually to toss a real coin an infinite number of times to
evaluate the limit, but the intuitive idea is that such a limit ought to exist and should
define whatever it is we mean by the probability of the coin coming up heads.
A third way that probabilities are tossed about in everyday conversation
involves subjective considerations. Someone might say, for example, “Notre Dame
is a 2 to 1 favorite to beat Michigan.” This statement has a clear meaning as far as
probabilities go. The speaker is saying that Notre Dame’s chances of winning are 2
chances out of 3, which means a probability of 2/3. Such a statement is the
speaker’s quantitative pronouncement of his or her or somebody’s opinion on the
matter. Another such illustration is a weather forecaster announcing a 30% chance
of rain. Presumably such a statement would be based on existing weather data and
would not be purely subjective. It may be, however, that some kind of subjective
guesswork went into building the weather model from which the 30% figure is
obtained.
A mathematically useful treatment of probability must lay a common
groundwork so that everyone is speaking the same language. Much of this
groundwork consists of the elementary language of sets and set operations.
1.1 Sets and Set Operations 3

1.1 Sets and Set Operations

Intuitively, a set is simply a collection of objects. This is one of the most
primitive of mathematical concepts, and thus we cannot define sets in terms of yet
more elementary concepts. The common practice is to denote sets by capital letters.
Equality of two sets means that the sets consist of exactly the same elements.
For example, A = {1, 2, 3} and B = {2, 3, 1} are equal. (There is no order
associated with the elements of a set. A set is simply an unordered collection of
objects.) If every element of set A is also an element of set B, then A is a subset
of B and we write ACB. Set membership is denoted by the symbol €. For
example, 1 € A but 4 ¢ A in this example.
Perhaps the most useful method of defining sets is by describing the rule for
set membership. For example,
S = {x : x is an integer and x > 0}
is simply a way of describing S as the set of positive integers.
The three basic set operations are union, intersection, and complementation.
Union and intersection are operations that are performed on collections of sets,
whereas complementation is performed on a single set.
The union of two sets A and B is denoted by A UB and defined by
AU B= {x:xe Aorxe B}
It is important to understand that “‘or” allows the possibility of membership in both.
Thus the condition for membership in A U B is simply that the candidate be an
element of at least one of the two sets A and B.
In the definition of the intersection of two sets, “‘or’”’ becomes “and.” Thus,
AQB = {x:xe Aandxe B}
Intersection thus represents the “overlap” of the two sets. If this intersection is
empty, then the sets are said to be disjoint or mutually exclusive. Both the
definition of union and intersection extend naturally to any collection of sets (even
an infinite collection). The union of a collection of sets consists of all elements that
belong to at least one set of the collection, whereas the intersection of the collection
consists of the elements that all the sets have in common.
In any discussion about sets, the sets under consideration will all be subsets
of some universal set lurking in the background. (What the universal set is should
always be clear from context.) The complement of a set A is denoted by A‘ and
is defined as the set of all elements in the universal set that do not belong to A.
For example, in the context of a discussion of the real number system, the
4 Chapter 1: The Basics

complement of A = {x:x>1} 1s
Ao= x ess
The difference of two sets, A —B, is defined as A M BS.

Some of the important elementary laws governing the set operations are as
follows:

Associative law for union AU(BUO=LUB) UC

Associative law for intersection AA (B AVG) Arr by eve
Commutative law for union AUB=BUA
Commutative law for intersection AAB=BOA
Distributive laws AOC BNSC)} = (A Cre) (A ee)
AW(B A.C) =A.UB)ACAUC)
De Morgan's laws (ALB)> = Aa Be
(A 0B) = AS UBS

A simple way to verify some elementary set identities is to use Venn

diagrams. The idea is to visualize a set as being represented by a region in the
plane. Figure 1.1 illustrates the use of this concept with regard to the first
distributive law above. The shaded area in the figure is the region corresponding to
the sets on either side of the equation in the first distributive law. One way to check
equality of sets is to construct a Venn diagram for each set and to observe that the
Venn diagrams coincide.

Figure 1.1 Venn diagram representing the set Aq (BU C) = (AN B)U(AN ©).
1.1 Sets and Set Operations 5

Example 1.1. A group of 9 men and 7 women are administered a test for
high blood pressure. Among the men, 4 are found to have high blood pressure,
whereas 2 of the women have high blood pressure. Use a Venn diagram to
illustrate this data.

Solution:

The circle labeled H represents the 6 people having high blood pressure, and
the circle labeled W represents the 7 women. The numbers placed in the various
regions indicate how many people there are in the category corresponding to the
region. For example, there are 4 people who have high blood pressure and who are
not women. Such people are in set H but not in set W; that is, they belong to the
set HO W*°. The number 5 in the lower right corner indicates the number of men
without high blood pressure.
The decision to use circles to represent “high blood pressure” and “women”
was quite arbitrary. We could just as well use circles for “low blood pressure” and
Sinem. a(seebropleny 1225,)

Example 1.2. If A, B, and C are sets, draw a Venn diagram and shade the
region corresponding to the set (A UB‘) NC.

Solution: The best way to arrive at the following figure is in a step-by-step

manner. First, decide what region corresponds to the set AUB‘. This will
include all the space inside the A circle and all the space outside the B circle.
Then, the final step is to observe what part of this region lies inside of C (since
the final operation is intersection). If you like you can think of spray painting
everything inside the A circle, then spray painting everything outside the B circle,
and then looking to see what part of the inside of circle C is painted. The region
that represents the set (A U B‘) 1 C is shown in the figure.
6 Chapter 1: The Basics

Example 1.3. If A,B, and C are the following sets of characters, then
determine the set (A UB‘) AC. Here we will consider the universal set to be
all 26 letters of the alphabet.
A={
B= {d,
9. 0, t}
C = {d, a, g}

Solution: While it is possible to represent this information in a Venn

diagram, it certainly isn’t necessary. Simply observe that B° consists of all letters
except d,g,b, and t. Therefore, A UB* consists of all letters except d, g,
and b. So the only letter that AU B® has in common with C is a. Conclusion:
(A U B°) NC = {a}. Notice that Problem 1.27 asks you to observe this in the
context of a Venn diagram.

1.2 The Sample Space

The sample space is roughly “the set of all possible observations or

outcomes” of whatever is under discussion. This idea is best illustrated through
examples.

Example 1.4. Ifa coin is tossed, the sample space could be taken to be the
set S = {H, T}. If an ordinary six-sided die is rolled, the sample space could be
taken to. be. the set.S = (1, 273. 4) S46).
1.2 The Sample Space ri

Example 1.5. A card is drawn from a standard deck of 52. Here one could
take the sample space S to be
S = (2%, 24,24, 24, 3m%,3¢,---,Ka, Am, Ao, Av, Aa}
A simpler convention would be to agree to think of the cards as being identified
with the numbers 1,2, - -- , 52 and to simply think of S as consisting of these 52
numbers. It is important to realize that the particular bookkeeping scheme used is
not very important compared to the conceptual understanding of what kind of set is
an appropriate model. Whatever notation is used here, the sample space is a set of
52 elements.

Example 1.6. Suppose a simple electric circuit has two components, say A
and B. Either component can be “good” or “bad” in the sense that the component
may or may not be in working order. If we are interested in all possible states of
the circuit, the sample space used could be S = {GG, GB, BG, BB}, where the
convention might be that “GB,” for example, means that component A is good and
component B is bad.

Example 1.7. A pair of dice is rolled. Let’s refer to them as “red” and
“green.” An appropriate choice for the sample space for this experiment would be
ine set: S_={(1, 1), -G, 2), 3)- > 2,)46-5)> (6, 6)} swhere for instance, swe
might agree that (3, 5) represents the outcome of 3 on red and 5 on green. The set
S contains 36 elements since either die can come up 6 different ways and 6 x 6 is
36. (There’s a basic underlying principle here that some elementary texts call the
multiplication principle. When one task can be performed in m different ways and
another task can be performed in n different ways, then the number of different
ways of performing the two operations together is mn.)
In playing many games, one is not interested in the individual numbers that
appear on the dice, but rather in the sum. In this case it might be tempting to take
the sample space to be S = {2, 3, 4,---, 11,12}. This is not necessarily wrong,
but it does sacrifice information. For example, if this sample space is used, one can
no longer answer questions such as “Did the red die show an even number?” The
result on the red die is not even being recorded. Another reason for caution is that
the outcomes of this set of possible sums are not equally likely. For example, a
sum of 2 occurs only if both dice show the number 1, whereas a sum of 7 occurs in
six different ways. It is often necessary to consider sample spaces in which
individual outcomes don’t all have the same probability. One needs, however, to
8 Chapter 1: The Basics

be aware when this is the situation. A frequent naive mistake is to assume that
outcomes are equally likely when they are not. It is common to refer to a sample
space in which all the elements are considered equally probable as a uniform
sample space.

Example 1.8. When binary data is transmittéd, the output can be thought of
as a string of 0’s and 1’s. (In electrical transmission, a voltage above a certain level
could be defined as 1 and below a certain level as 0.) If 4 bits are transmitted, what
would be an appropriate sample space to represent the possibilities?

Solution: The logical choice would be to take the sample space to be all
ordered quadruples of binary digits. In other words, S = {0000, 0001, 0010,
0011, 0100, 0101, 0110, 0111, 1000, 1001, 1010, 1011, 1100, 1101, 1110,
1111}. These 16 outcomes are simply the numbers from 0 to 15 written in binary
form. There are 16 elements of the sample space because 24 = 16. Eight bits
(commonly referred to as a byte) can represent 2° = 256 different possibilities.
This is equivalent to saying that if 8 bits are transmitted, then the sample space for
all possible outcomes has 256 elements.

1.3 Basic Properties of Probabilities

The word event is commonly used in everyday language, and often we speak
of a particular event “occurring.” In probability discussions, you should think of
an event as a subset of the sample space. For example, if we say when a die is
rolled that “an even number occurs,” we are saying that the observed outcome lies
in the set E = {2, 4, 6}. We are describing E verbally by saying E is the event
“that an even number occurs,” but this is simply an alternate way of saying that E
is the given set of outcomes. The reason for talking about events occurring is that
this language is so common and useful in everyday speech. In talking about
probability, to say that an event occurs is simply to say that the observed outcome is
one of the elements of the event. This mathematical language is consistent with
common speech. You should keep in mind, though, that in the language of
probability there is a precise mathematical meaning to such expressions.
There are three basic axioms that characterize what is meant by a probability
measure. The notation P(A) represents the probability that the event A occurs,
and the assumption is that probabilities always behave according to the these rules.
1.3 Basic Properties of Probabilities i<o)

The Basic Axioms for a Probability Measure

1. For each event A, P(A) 2 0.
2. P(S) = 1, where S denotes the sample space.
3. IfA and B are disjoint, then P(A U B) = P(A) + P(B).

The intuitive basis for all three axioms should be obvious. When we say that
something has probability 1/10, we mean that there is one chance in 10 that it will
occur. A negative probability has no conceivable meaning. Similarly, a probability
of 1 represents absolute certainty, and probabilities greater than 1 would be
meaningless. The last axiom is very important in that it allows the probability of
events (in the case of a finite sample space) to be computed in terms of the
probabilities of the individual elements that make up the events. This will be
demonstrated shortly. (See Equation 1.1.)

Additional Properties of Probabilities

1. P(@) =0, where @ denotes the empty set.
P(A) < P(B) if ACB, and P(A) < 1 always.
P(A-—B) =P(A)-P(A OB).
P(A UB) = P(A) + P(B) — P(A AB) always.
WN If A,, Az,---+,A, are disjoint events (no two events having any
nA

elements in common), then

P(A; VA ,U-++UA,) = P(A}) +--+ +P(A,)

Property 1 is an immediate consequence of Axiom 3. (Simply take A and B

both to be the empty set in Axiom 3, and you have P(@) = 2 P(@), which implies
P(@) = 0.)
If A CB then B =A U (B —A), and this is a disjoint union. Thus, from
Axiom 3 we know that P(B) = P(A) +P(B-—A). The right side here is
greater than or equal to P(A) because P(B — A) 2 0 (Axiom 1), and this proves
Property 2.
Property 3 is a result of the fact that A =(AMB)U(CA —B), and the
union here is disjoint.
10 Chapter 1: The Basics

Property 4 may be obtained by first noting that

P(A UB) = P(A) + P(B-A)
This is a special case of Axiom 3. From Property 3, however, we know that
P(B —A) =P(B)-PA OB)
The truth of Property 5 is a consequence of mathematical induction. In fact, if
P(A, VA, U---UA,) = P(A;) +--+ +PA,)
then it follows that
PAA, UU A= PAp Poe Hr ae)
simply by letting A; UA,U---UA, and A,,, play the roles ofA and B in
Axiom 3.

For finite sample spaces it is always possible to compute the probability of an

event by focusing on the individual elements that make up the event. For if the
sample space is S = {5,,5,-+-,5,}, then, according to Property 5, the
probability of any event A CS may be computed via

P(A) = >, PCs) ) (1.1)

SyEA

In other words, the probability of any (finite) event may be computed by simply
adding up the probabilities of the individual elements of the event.
Later in this book we will see that probabilities of events often must be
approached from a different perspective when the sample space is infinite. The idea
of computing probabilities of all events via a sum, as in Equation 1.1, must be
abandoned. In Chapter 5 we will see that in “continuous” models, integration takes
the place of summation.
Continuous models arise, for instance, when measurements are being made
on some kind of continuous scale and one wishes to think of an interval of possible
values. For example, think of the experiment of “selecting a random number
between 0 and 1.” Clearly it would be intuitively satisfying to think that the
probability is 1/2 that the number should come from the subinterval [0, 1/2], or 1/5
that the number chosen should be in the interval [3/5, 4/5]. In fact, we would like
to know that the probability of the number falling in any particular subinterval is
simply the length of the subinterval. Simple as this situation sounds, the actual
demonstration that there is a probability measure on the interval [0, 1] that has these
1.4 Conditional Probability 11

pleasant properties was a milestone in mathematics around the turn of the century
and made possible the field of study known as real variables. An interesting
consequence of the Axioms | through 3 is that such a probability measure must
necessarily assign probability zero to every individual number. To understand
why, simply think of a number x sitting in the interval Bs = [x — 6, x + 8]. The
length of this interval is 25. But in light of Property 3, P({x}) < P(Bs) for
every 6 > 0, and so P({x}) must be zero. (To say that an event has probability
zero is not to say that the event is impossible. For instance, if a random number is
selected between zero and one, the probability that the number would be 1/2 is zero
as just indicated. Yet it is not impossible for the number selected to be 1/2.)
For the sake of honesty, it is advisable to admit at this point that there are still
other pathological things that occur with infinite sample spaces. With finite sample
spaces, an event is simply a subset of the sample space. With infinite sample
spaces, it is not always possible to consider every subset of the sample space to be
an event. The problem is that for infinite sample spaces it simply isn’t usually
possible to have our probability measures defined for all subsets of the sample
space and still to satisfy the desired axioms. You needn’t lose sleep over this,
however, because in all the frequently encountered sample spaces, the subsets that
must be avoided are ones that you would never encounter anyway because they
cannot be expressed in terms of elementary sets. In other words, you can ignore
this conceptual difficulty and never run into computational problems because of it.

1.4 Conditional Probability

New information changes one’s notion of probabilities. For example, a card

drawn from a deck has probability 1/4 of being a spade since 1/4 of the cards in the
deck are spades. However, if it is known that the card is a black card, then the
conditional probability that the card is a spade becomes 1/2 since half of the black
cards are spades. Effectively, what has happened is that the additional information
(that the card drawn is a black card) has eliminated some of the original
possibilities. This forces one to reconstruct the sample space to form a new
“reduced” sample space consisting of only 26 of the original 52 possible outcomes.
The definition of conditional probability given below represents a slightly different
view. The examples that follow indicate that the definition is consistent with the
intuitive view of a “reduced sample space.”
12 Chapter 1: The Basics

Definition 1.1: If A and B are events in a sample space and P(A) #0, then
the conditional probability of B, given A, is denoted by P(B |A) and is defined
by
P(B OA)
SS ae

Figure 1.2 shows a graphical interpretation of the idea of conditional

probability. Both probability and area can be thought of as measures of size. In
fact, P(B |A) is a measure of the size of P(B OA) in comparison to P(A).
In Figure 1.2, the analogous idea would be to compare the area of B MA to the
area of A. In both the probability setting and the geometric figure, we are in some
sense measuring what fraction of A happens to also lie in the set B.
When considering P(A |B), the size of P(B 7m A) is compared to P(B)
rather than to P(A). Quite possibly P(B 1A) and P(A 1B) may be quite
different in magnitude. In the figure the region B A constitutes a much larger
fraction of A than it does of B.

Figure 1.2 How probable is Am Bin comparison to A?

This is the idea behind the conditional probability P(B| A).

Example 1.9. A card is drawn from a 52 card deck. Let A be the event that
the card is black and B the event that the card is a spade. Thus AM B = B, so
P(A OB) = 13/52, and P(A) = 26/52. This means that
1.4 Conditional Probability 13

13/52 1
P(B|A) = 26/52 2

Example 1.10. A pair of dice is rolled (one red and one green). The sample
space is as in Example 1.7. Let A be the event that the sum on the two dice is 9,
and let B be the event that the red die shows the number 5. Then A 7B contains
the single outcome (5, 4) and has probability 1/36, and P(A) = 4/36 because A =
{(3, 6), (4, 5), (5, 4), (6, 3)}. Thus, P(B 1A) = 1/36 + 4/36 = 1/4. This is
intuitively correct, for knowing that the sum is 9 guarantees that the red die must
show one of the four numbers 3, 4, 5, or 6.

Example 1.11. In the sample space of Example 1.5, let B denote the spades
and now let C denote the aces. Then P(B ™C) = 1/52 since there is only one
card that is the ace of spades, and so P(B | C) = 1/52 + 4/52 = 1/4. Notice that in
this case P(B | C) = P(B). It can be easily verified that in this example it is also
true that P(C |B) =P(C). These two equations indicate that the knowledge that
one of the events occurs does not affect the probability of the other occurring. Two
events that have this relation to each other are called independent events (or
Statistically independent events). The concept of independence is very important in
mathematical models that involve probability. In many situations it will be apparent
for one reason or another that two events can have no possible effect on one
another, and in the mathematical model this leads to an assumption that the events
are independent.

Example 1.12. Two bits of binary data are transferred. Let’s consider the
sample space to be S = {00, 01, 10, 11}. So 01 represents the case in which a 0 is
transmitted first and a 1 is transmitted as the second bit. We will consider the
transmission to be random in the sense that each of these arrangements has
probability 1/4.
Now let’s consider the following two events:
A = {10, 11}
B= "(0134 Lb}
Event A can easily be described as the event “the first digit transmitted is the
digit 1,” andB as the event “the second digit is the digit 1.”
Notice that P(B) = 1/2, and furthermore that P(B |A) = 1/2. (Why?) The
fact that P(B) = 1/2 says that the probability that the second digit is a 1 is equal to
14 Chapter 1: The Basics

1/2. That P(B |A) = 1/2 says that if we know that the first digit is a 1, then the
conditional probability (based on this information) that the second digit is a 1 is
still 1/2. This is another illustration of independence, the concept introduced in
Definition 1.2 below. Problem 1.28 asks you to show that the three conditions are
in fact equivalent (that is, if one of them is true, then all of them have to be true).

Definition 1.2: Given events A and B with nonzero probability, it is easy to

check that the following three conditions are equivalent:
1. P(A|B) = P(A)
2. P(B 1A) =P(B)
3. PAOB)= P(A) P@)
When these conditions are true, the events A and B are said to be independent.

Conditional probabilities have many uses. Strangely enough, one of the

easiest and most useful is in the computation of unconditional probabilities. In fact,
the definition of P(B |A) can be rewritten as

P(A OB) = P(A) P(B IA) (1:2)

This idea can be extended to compute the occurrence of more than two events. For
example,

P(ANBOC)=P(A)
P(BIA) P(CIAQB) (13)
Now for some examples that illustrate the use of these simple but useful little
formulas.

Example 1.13. A box contains 10 transistors, of which 7 are good and 3 are
defective. If 2 transistors are randomly taken from the box, what is the probability
that both are good?

Solution: Think of the transistors as being drawn one at a time. (This just
means that we're going to label them “first” and “second.” Whether they are
actually drawn one at a time or simultaneously doesn’t matter in the least. You
should convince yourself that this little mental trick is legitimate.) Let A be the
event that the first is good, and B the event that the second is good. Then the
probability that both are good is
1.4 Conditional Probability 15

et sG ¥
P(A MB)=P(A) P(B|\A)= T0'O = 15
The reason that P(B |A) = 6/9 is that knowing that the first one is good means
that there are 6 good left among the 9 possible ones that might be chosen second.
In a similar fashion, the probability that all would be good if 3 were selected
could be computed as
TORI ae
10 oS
The last factor, 5/8, is the conditional probability that the third would be good given
that the first two are both good.

Example 1.14. Consider the experiment of rolling a red and a green die.
Let A be the event that the red die shows a 5, and B be the event that the green die
shows a 3. Most people’s intuition says that the two dice should not exercise any
control over each other; that is, the outcome on one die should be “independent” of
the outcome on the other. If the sample space of Example 1.7 is used, then

AMES A(S55) jp eSOur (A.D = =

puso. b-=-{ (1-3), C73)4G673), 44, 3) 6e3)6, 3))}s cand so -P(B) =-6/86.

Thus,

136 _ 1
Ee(AdkB) = 636 ars = P(A)

This shows that A and B are independent according to the Definition 1.2,
just as our intuition would indicate. (A somewhat subtle but important point should
be made here. One cannot prove mathematically whether two real-world dice
behave independently or not. All we are illustrating is that if the sample space
introduced in Example 1.7 is used to model a pair of dice, with the assumption that
the outcomes are equally likely, then the theoretical dice of the model behave
independently.)

Example 1.15. A manufacturing company has two plants, 1 and 2. Plant 1

produces 40% of the company’s output, and plant 2 produces the other 60%. Of
the devices produced at plant 1, 95% are good and 5% are defective. The output of
plant 2 is 90% good and 10% defective. If a device is randomly selected from the
output of this company, what is the probability that the device will be good?
16 Chapter 1: The Basics

Solution: Let B denote the event that the randomly selected item is good,
and let A, and A, be the events that it comes from plants 1 and 2 respectively.
Then P(B) = P(B 1 A,) + P(BO Az) because A; and Ag are disjoint
events and every element of B must be in either A; or Aj. But then
P(B) = P(BOA,) + P(BOA)).
P(A,) P(B1A,) + P(A2) P(B 1A2)
(.4)(.95) + (.6)(.9) = .92
(We are using Equation 1.2 here. An alternative is to use a tree diagram for this
computation. Tree diagrams are explained in Section 1.5.)

Example 1.16. Suppose it is known that 3% of cigarette smokers develop

lung cancer, whereas only .5% of nonsmokers develop lung cancer. Furthermore,
let’s suppose that 30% of adults smoke. If it is known that a randomly chosen
adult has developed lung cancer, what is the probability that the person is a smoker?

Solution: Let S = smokers and L = people who develop lung cancer. The
question is what is P(S|1L). By definition P(S|L)= P(S ML)/P(L).
However,
P(S QL) =P(S) P(L 1S) = (.3)(.03) = .009
Moreover,
P(L) = P(SSQL)+P(S° OL) since SAL and S° OL are disjoint
=-P(S) PLS) + PS Peis
= (.3)(.03) + (.7)(.005) = .009 + .0035 = .0125.
Thus P(S1L) = .009/.0125 =.72.

The technique used in this computation may be incorporated in a tree diagram

or Bayes’ theorem, both introduced in the next section. (See Problem 1.21.)

1.5 Tree Diagrams and Bayes’ Theorem

The techniques used in Examples 1.15 and 1.16 can be formalized in two
different ways. One formulation is commonly referred to as Bayes’ theorem and is
included in Proposition 1.1. Another approach utilizes a simple technique called
tree diagrams which will be introduced shortly.
1.5 Tree Diagrams and Bayes’ Theorem qi

Bayes’ theorem comes into play whenever the elements of the sample space
under consideration are divided up into mutually exclusive categories. Specifically,
suppose that S is the sample space and that A,, A>,---,A, are events in S$
having the property that A,,A>,---,A, are disjoint and that A; U---UA,
= S$. (This means that each element of the sample space S belongs to exactly one
of the events A,, Ay,---,A,, and for that reason such a collection of sets is
often referred to as a partition of the sample space.) Then for any eventBCS,
the additivity property of probabilities implies that
P(B) = P(BOA,) +---+P(BOA,)
If each of the terms on the right is rewritten using the definition of conditional
probability, the first equation in Proposition 1.1 is obtained (the multiplicative law).
Bayes’ theorem itself is an immediate consequence of the multiplicative law. The
definition of conditional probability says that
P(A, OB)
P(A, |B) = pear
If the definition of conditional probability is used now to rewrite the numerator of
this fraction, and if the multiplicative law is used to replace P(B) in the
denominator by the sum on the right side of Equation 1 in Proposition 1.1, then
Equation 2 (Bayes’ theorem) is obtained.

Proposition 1.1: Suppose that B and A, A,---,A, are events in a sample

space S, that A,, A7,---,A, are disjoint, and thatA,; U---UA, =S.

1. (multiplicative law) Then for any event B in S,

P(B) = >, P(A) PB 1A)

i=l

2. (Bayes’ theorem) For any one of the sets A,,

P(A,) P(BIA
P(A, |B) = Bal cata
S'P(4) PBA)
i=1
18 | Chapter 1: The Basics

The solution shown in Example 1.16 is exactly the one that would be
produced by using Bayes’ theorem. To make the identification, simply let the event
L in Example 1.16 correspond to B in Bayes’ theorem, and let S and S°
correspond to A, and A, (with n = 2 and k = 1) in Bayes’ theorem.
Many simple problems that can be done with Proposition 1.1 can also be done
with tree diagrams. The concept of a tree diagram is easily understood by looking
at a few examples.

Example 1.17. To check for mercury contamination, a group of 25 dolphins

is being monitored. Of the 25 dolphins, 6 have been determined to contain a high
concentration of mercury contamination in their fatty tissues, whereas the other 19
do not. If 2 dolphins are randomly chosen from this group of 25, what is the
probability that both will have high mercury concentrations?

Solution: The purpose of this example is to introduce the concept of a tree

diagram. The tree that is constructed will show show more information than is
required to answer the question being posed.

05 19 19 57

5/24 19/24 6/24 18/24

High Low

6/25 19/25

At the lower level in the tree, the nodes labeled “High” and “Low” represent
the two possibilities (high or low mercury concentration) for the first dolphin.
(Again, it doesn’t really matter whether they are chosen one at a time or
simultaneously. We can simply think of one of them as being labeled “first”? and
the other as being labeled “‘second.”) The probabilities on the branches are 6/25 and
19/25 because of what is known about the makeup of the group of 25. At the
1.5 Tree Diagrams and Bayes’ Theorem 19

second level in the tree, the possible conditions of the second dolphin are recorded.
The numbers on the second level of branches are conditional probabilities. For
example, the leftmost number 5/24 at the second level is the conditional probability
that the second dolphin has a high concentration of mercury given that the first one
does. The reason this conditional probability is 5/24 is that once we know that the
first one has a high concentration, we know that 5 of the remaining 24 also have
high concentrations. The four ovals at the top of the figure represent the four
possible outcomes of the experiment. From left to right they may be described as:
both high, first high and second low, first low and second high, both low. Thus
the tree diagram is consistent with thinking of the sample space as something like S$
= {HH, HL, LH, LL}. Finally, the numbers at the top of the figure are the
probabilities of the four outcomes corresponding to the ovals. These are computed
by multiplying the numbers on the branches connecting the root of the tree to the
ovals. For example, the probability of the outcome HH is computed as 6/25 x 5/24
= .05. This use of multiplication is an application of Equation 1.2.

Example 1.18. Two bad resistors have been inadvertently installed in a

device having five resistors altogether. It is not known which two are bad.
Resistors are tested one at a time until the two bad ones or the three good ones have
been found. Draw a tree diagram to represent this process, and find the probability
that exactly two good resistors are tested.

Solution: In the following figure “G” represents selection of a good resistor

and “D” represents selection of a defective one. The first level of the tree represents
the first resistor selected, and the probabilities shown there are 3/5 and 2/5,
corresponding to the makeup of the set of five resistors. Above the lowest level the
probabilities are conditional probabilities. The three shaded circles represent the
outcomes in which exactly two good resistors are tested. The desired probability is
then just the sum of the probabilities of these three outcomes. As before the
probabilities of outcomes may be computed by multiplying the numbers on the
branches between the root and the top. In this example, all outcomes turn out to
have probability 1/10, and so the probability that exactly three good resistors are
tested is 3/10.
Since the tree diagram shows all the relevant information about the
experiment, conditional probabilities also can be computed from the tree. For
example, suppose A is the event that both defective resistors are tested, and B is
the event that exactly three resistors are tested altogether. In the diagram there are
20 Chapter 1: The Basics

three outcomes in which exactly three resistors are tested (the circles at the
intermediate level). Since each has probability 1/10, P(B) = 3/10. An B
consists of two of these outcomes; so P(A A B) = 2/10. Therefore,
2/10 2
P(A1|B) = 3/10 3.
that is, if it is known that exactly three resistors are tested, then the probability that
both defectives are tested is 2/3.

3/5 2/5

It is very important to understand clearly the difference between the

conditional probability P(A |B) and P(A B). Beginning students sometimes
confuse these two probabilities. Problem 1.11 at the end of the chapter is a good
check on this concept.

1.6 More on Independence

Sometimes it is necessary to consider the independence of more than two
events. For three events, it would be tempting to want to consider events A, B,
and C to be independent if P(A XN Bm C) = P(A) P(B) P(C). It is easy to
see that this view is inadequate, however, by considering the special case where C
is empty. For then both sides of the equation are zero, and so the equation is
necessarily valid even though A and B may be highly dependent. The following
definition is what is needed.
1.6 More on Independence 21

Definition 1.3: A collection of events is independent provided that for every

finite subcollection A, ---, A,, it is true that
P(A, +++ OA,) = P(Ay) -- + P(A,)

For example, three events A, B, and C are independent provided all of the
following are true:
1. PANBOC) =P(A) PC) PCO)
2. P(A QB) = P(A) P(B)
3. P(AQC)=P(A)PC)
4. P(BOC)=P@)42OQ

Example 1.19. A coin is tossed twice. A natural sample space is given by

S = {HH, HT, TH, TT}, with each outcome having probability 1/4. Consider the
events:

A = {HH, HT} = event that heads occurs on first toss

B = {HH, TH} = event that heads occurs on second toss
C = {HH, TT} = event that same result occurs on the two tosses

It is routine to verify that equations 2 through 4 above are valid when A, B,

and C are these three events. This simply says that any two of these three events
are independent. However,
1
PANBOAC)=PANB)= a

whereas
Pett 1

So the collection of three events does not form an independent collection. One way
to understand this is to observe, for example, that if we know that A and B occur,
then we know with absolute certainty that C occurs. Intuitively, independence of
the collection of events would indicate that no knowledge involving only A and B
could influence the probability of C occurring. (A more mathematically precise
statement is the following: If A,B, and C are independent, then C will be
22 Chapter 1: The Basics

independent of any event that can be formed from A and B using set union,
intersection and complementation. This can be proved using Definition 1.3.) In
particular, in a collection of independent events, any of the events may be replaced
by their complements and the resulting collection of events will still be
independent. This will be extremely important to us in Chapter 2. Most of the
applications will feature the concept of independent events in one form or another.
In Chapter 2, we will constantly be performing calculations such as
P(AANBOC = P(A) PB) [1-P(O)]J
if A, B, and C are independent.

There is a common situation called an independent trials process that is

closely connected with the concept of independent events. An independent trials
process is a sequence of experiments in which the result of each trial is one of only
two outcomes commonly called success and failure. Furthermore, it is required
that the probability of the two outcomes remain the same on every trial. The easiest
illustration is repeated coin tossing, where we might agree to call heads “success.”
(The choice of which outcome to call success is purely arbitrary.) We will look at
an easy formula for the probability of a certain number of successes in a given
number of trials. First, however, it is useful to present an elementary counting
technique.

Proposition 1.2 (Combinations formula): If 7 is a positive integer and if

0 <k <n, then the number of different possible sets of k objects that can be
chosen from a set of n objects is given by

tn)
a eee

The notation n! is for n factorial, that is, n! = n(n —1)(n-—2)---1. For

example, 4!=4x 3x 2x 1=24. By convention 0! = 1. The proof of the above
counting principle is not difficult, but we will omit it because counting techniques
are not central to our purpose. Example 1.20 presents a simple situation to which
this formula can be applied, and Example 1.21 shows a related situation in which it
is not directly applicable.
1.6 More on Independence 23

Example 1.20. If you have offered to give a friend his choice of any 3 books
from 10 that you own, the number of different ways he can make his selection is
10!
C0, 3) = 3077 = 120
Notice that the selection of 3 books to be given away also completely determines
which 7 will be left behind. In other words, C(10, 3) = C(10, 7). More
generally, if 0 <r <n, then C(n, rf) = C(n, n—7r).

Example 1.21. In aclub with 10 members, how many ways are there to fill
a Slate of officers consisting of a president, secretary, and treasurer? The answer is
10x9x8 =720. The difference in this example and the last is that here we are not
just selecting 3 people from 10, but specifically 3 people for 3 separate positions.
Utilizing the multiplication principle, you can think of this as 10 choices for
president, then (after that choice has been made) 9 possibilities for secretary, and
then 8 for treasurer. The assumption is that no one occupies two offices.

It is standard practice to refer to problems such as Example 1.20 as

combinations problems. These are problems in which a set of objects is being
selected but no ordering of the objects is being considered. Problems such as
Example 1.21 are called permutations problems. In such problems the order of
selection does come into play. In Example 1.21, we are essentially assigning an
order to the three officers by matching them up to the three specific offices.
Generally, the term permutation is used to refer to an ordered arrangement of some
collection of objects. Permutations and combinations problems are part of a field of
mathematics known as combinatorics.

Proposition 1.3 (Independent Trials Formula): Suppose n trials are

conducted of an experiment in which the probability of success on each trial is p
and the probability of failure is q=1-p. Then for any integer k that satisfies
0 <k <n, the probability of exactly k successes in the n trials is
C(n, k) p* Go

For example, suppose a regular unbiased six-sided die is rolled 4 times.

What is the probability of obtaining two 6’s? Considering a 6 to be a success, then
24 Chapter 1: The Basics

p = 1/6 and q = 5/6. Here n = 4 and k = 2 in Proposition 1.3. Therefore, the

probability of exactly two 6’s is

o.(§) (5) = as
You should verify the validity of the formula in this simple case by looking at the
same experiment in terms of a tree diagram, with one level of branching for each
roll of the die (with the outcomes for each roll shown as “6” or “not 6”). Doing
this, in fact, leads to understanding why the formula is correct. For in a tree
diagram to represent n independent trials, there would be exactly C(n, k)
outcomes across the top corresponding to k successes in the n trials, and each of
these outcomes would have probability p* q”“*.
In the tree diagram for an independent trials process, the probability for
success is the same at all locations in the tree. It is important to keep in mind that
much more general kinds of experiments can be viewed in terms of a tree diagram
than can be viewed as an independent trials process. The relation between an
independent trials process and the concept of independent events is this: In an
independent trials process, if A is an event that may be described in terms of the
outcome of one trial and B is an event that may be described in terms of the
outcome of another trial, then A and B will be independent events. This statement
may be generalized to more than two events. For example, if a coin is tossed 10
times, the experiment may be viewed as 10 trials in an independent trials process.
If A, is the event that “heads occurs on the first toss,” A> that “heads occurs on
the second toss,” and so forth, then Aj, - - - , Ajg are independent events.

1.7 Infinite Sample Spaces

When dealing with infinite sample spaces, it is often necessary to consider

unions of infinite collections of events. For this reason, Axiom 3 is inadequate for
infinite sample spaces and must be strengthened to Axiom 3.1 following. The
reason is that mathematical induction cannot extend the property of finite additivity
to the infinite series needed in the case of an infinite sample space. The right side of
the equation in Axiom 3.1 is the sum of an infinite series. (The series will
necessarily be convergent if the events are disjoint as in Axiom 3.1.)
1.7 Infinite Sample Spaces 25

Axiom 3.1: If Aj, Ay, --- is a sequence of disjoint events in a sample space,
then
P(A, U Ap U =++) = P(Ay) + P(g) +++

Axiom 3.1 simply says that even with a countably infinite collection of
disjoint events, the probability of the union is still the “sum of the probabilities”
provided that the latter is interpreted as the sum of an infinite series.
The concept of a geometric series will be useful in the next example and in
other applications later. ;

Proposition 1.4 (Geometric Series): If Irl <1 then

1
ltrtretrP+... =
1l-r

The derivation of the sum of a geometric series is easy, because

dar) (1+ ptr? 46047 =I!
Simply dividing both sides of this equation by 1 —r and taking the limit as r > ©
gives the sum of the infinite series, since r™1 _ +0 asm 00 if Irl< 1.

Example 1.22. A coin is to be tossed repeatedly until a head occurs, at

which time the tossing stops. What is the probability that an even number of tosses
is required?

Solution: Since there is no upper bound that can be placed on how many
tosses might be required, an infinite sample space is required to model this
experiment. In fact, we can simply think of the sample space S as being the set of
positive integers, where the integer 4, for example, would represent the outcome
that the head occurs on the fourth toss.
How should the probabilities be assigned? In order for the head to occur on
the fourth toss, tails would have been necessary on the first three tosses. The
probability of getting 3 tails followed by a head is 54 when computed by any of
our standard methods. In general, an integer ke S represents the outcome in
26 Chapter 1: The Basics

which a head follows k-1 tails and hence has probability .5*. To check
consistency, observe that the sum of the probabilities of all the outcomes in S is
Dy he Speke Sok
gg ee ig ee
Factoring out 1/2 from each term here one is left with the geometric series found in
Proposition 1.4 with r = 1/2.
The question that is asked is what is the probability of the event
A= {2, 4,6, +. —}
that is, that the head occurs on an even numbered toss. Then

P(A)
G) GG)
7 a5 > =e
Gasoy es Wa mn at A mn cae

Gi lores)+} eee
at t+(s)+(4) +04) = (GQ) = 3
In this example we have been interested in the number of the toss on which a
head will first appear. We have viewed the sample space as simply being the
positive integers,
Sa tly 2 dee seuy
For example, the integer 17 represents the outcome in which the first head appears
on the seventeenth toss.
Someone night raise the question, “But isn’t it also possible that a head will
never appear?” You might be inclined to include some other element in the sample
space to represent that possibility, for example, the symbol oe. Is it correct to do
this?
Actually, it doesn’t matter. Since the sum of the probabilities that 1, 2, 3, - -.
occur is equal to 1, we would have no choice but to assign probability 0 to any
additional outcome that we included in the sample space, including an outcome to
represent the possibility of heads never occurring. So including it is pointless, but
not necessarily incorrect.
This example involves a countably infinite sample space. A countably
infinite set is one that can be put in one-to-one correspondence with the positive
integers. Notice that the techniques used are almost identical to those used with
finite sample spaces. We needed only to replace the idea of a finite sum by that of
an infinite series.
In Chapter 5 we will be studying continuous models. As mentioned earlier,
in continuous models the computations involve integrals rather than sums. Often
Problems 27

these are, in fact, easier to work with. This is because the standard methods of
calculus often are sufficient to evaluate integrals, whereas evaluating sums (even
finite ones) can be pretty tedious at times. There is not much that can be said in this
introductory chapter about continuous models, however, because the groundwork
that is laid in Chapter 3 must come first.

Problems

ae Assume that P(A) = .6, P(B) = .3, and A and B are independent.
Give the following probabilities:
(a)P(ANB) (b)P(ATB). (c) P(A UB) (d) P(A UB‘)

£2 Two devices are tested from a batch of 6 items of which 4 are good and two
are defective. Find the probability that
(a) both are good.
(b) both are good given that at least one is good.

13 John figures the probability he will make a grade of C or better in his

science course is .6 and the probability he will make a C or worse is .7.
What is the probability he will make exactly a C?

1.4 Peter has 4 blue socks, 3 brown socks, and 2 white socks in his drawer. If
he randomly reaches in and pulls out a pair of socks, what is the probability
that they will match?

1.5 Suppose A and B are events with the properties that P(AS M B°) = .2,
that P(A > B) = .2, and that P(B°) = .4. Find P(A).

1.6 A device has 5 components. Assume that each one has a probability .1 of
being defective and that whether a given one is defective does not depend on
the condition of any of the other components. Thus the 5 components may
be viewed as an independent trials process with n = 5 = number of trials.
(a) For k = 0, 1, 2, 3, 4, and 5, find the probability that exactly k are
good.
(b) What is the probability that at least 3 are good?
28 Chapter 1: The Basics

Ley In a county with 11 incorporated towns, 7 have an approved water supply

and 4 do not. An inspector intends to make a random spot check of the
water supply of 2 of the towns in the county.
(a) What is the probability that both of the towns that are checked will have
an approved water supply?
(b) What is the probability that neither will?”
(c) What is the probability that at least one will?

1.8 A and B are going to play a tennis match. The winner must win two sets
in order to win the match. (They stop playing when someone has won two
sets.) In each set, the probability that A wins the set is .6.
(a) Draw a tree diagram for the tennis match.
(b) What is the probability that A wins the match?
(c) What is the probability that B wins exactly one set?
(d) What is the probability that B wins at least one set?
(e) What is the probability that A wins the match given that B wins the
first set?

1-9 Assume the following to be true: Thirty percent of the members of a club
play tennis. Of those who play tennis, 70% also play badminton. Of those
who do not play tennis, only 10% play badminton.
(a) What is the probability that a randomly chosen person from the club
plays badminton?
(b) If you know that a randomly chosen person plays badminton, what
then is the probability that he or she plays tennis?

1.10 A system has five components, two of which are known to have a special
marking stamped on the side. (The other three are known not to have the
marking.) Components are checked one at a time until two of the
unmarked components have been found.
(a) Draw a tree diagram for this experiment.
(b) What is the probability that both of the marked components are
checked?
(c) What is the probability that exactly one of the marked components is
checked?
(d) What is the probability that at least one marked component is checked?
Problems 29

1.11 In a certain population, 5% have a disease. A diagnostic test for the disease
returns a positive result 90% of the time when it is used on a person who
actually does have the disease. (The other 10% of the time the test fails to
detect the disease.) However, when the test is used on a person who
actually does not have the disease, the test returns a positive result 8% of the
time.
(a) What is the probability that a randomly chosen person actually has the
disease and will test positive?
(b) If the test returns a positive result for a randomly chosen person, what
is the probability the person actually has the disease?

In a certain high school, 10% of the male students weigh 200 lb or more.
Of those who do weigh at least 200 lb, 30% play football. Of those who
weigh less than 200 lb, only 5% play football. If a randomly chosen male
student is known to play football, what is the probability that he weighs at
least 200 1b? Do this problem with a tree diagram and again using Bayes’
theorem.

A red and a green die are rolled.

(a) Find the probability that the number appearing on red is greater than the
number appearing on green.
[Hint: Let B be the event under consideration, and let A,,---,A¢
be the events representing occurrence of the numbers 1 through 6,
respectively, on the green die. Then
P(B) = P(A;) P(B1\A,) +--+ + P(A¢) P(B |Ag)
This equation comes from Proposition 1.1.]
(b) Use Bayes’ theorem to compute the conditional probability that the
green die shows a 3 given that the number on red exceeds the number
on green.

A coin is tossed and a die is rolled.

(a) Construct a natural sample space and probability measure for this
experiment.
(b) Show that the events A = “a head appears on the coin” and B = “an
even number appears on the die” are independent events.
30 Chapter 1: The Basics

Suppose S; and S$, are finite sample spaces associated with two different
experiments. The sample space for the combined experiment consisting of
both of these experiments could be taken to be the set of all ordered pairs
(w1,W 2) where w;eS, and we 5. Do you see a natural way of
defining a probability measure on this sample space that will make the
results of the two experiments independent of each other? (Consider this
question in light of Problem 1.14.)

An electrical device has two components, A and B. Each component can

be “good” or can be in a “shorted” or “open-mode” failure state. Let’s
assume that component A has probability .6 of being good and probability
.2 of being in each of the failure states, and that component B has
probability .7 of being good, .2 for “open” failure, and .1 for “shorted”
failure. Furthermore, we will assume that the two components function
independently.
(a) What is the probability that at least one of the components is shorted?
(b) What is the probability that neither is good?
(c) If the two components are wired in series, what then is the probability
that current can pass through?
(d) If they are wired in parallel, what is the probability that current can pass
through.
[Current can pass through a device if it is “good” or if it is in the “shorted”
failure state.]

Among the products of a certain manufacturer, 30% are defective. If you

pick 10 off the assembly line for testing, what is the probability that exactly
2 are defective? (Treat this as an independent trials process.)

1.18 A system has 100 components, 30 of which are defective. If you pick out
10 different components for testing, what is the probability that exactly 2 of
them are defective? (Use the combinations formula.) Make sure you
understand the difference in the situation described here and the situation of
Problem 1.17. This is a bit subtle, for in Problem 1.17 you presumably are
considering the testing to take place on 10 different items as well. Yet in
this problem, the sampling would have to be considered to be sampling
with replacement in order to produce the same answer as in 1.17.
Problems 31

(Sampling with replacement means that an item tested once may be tested
again. In other words, each item is replaced in the original pool from which
the samples are being drawn immediately after it is tested. For example, if 2
cards are drawn with replacement from a deck of cards, they both could be
the ace of spades.)

A device is put into service on a Monday and operates seven days each
week. Each day there is a 10% chance that the device will break down.
(This includes the first day of operation.) The maintenance crew is not
available on weekends, and so the manager hopes that the first breakdown
does not occur on a weekend. What is the probability that the first
breakdown will occur on a weekend. [Hint: Make use of the concept of a
geometric series. View each day as a separate trial in an independent trials
process in which you are waiting for the first success.]

£20 Show by induction that for any collection of events A,, A2,---,A,, it
is true that P(A; U---UA,) S$ P(A,) +--+ +P(A,).

2 1 Do the problem given in Example 1.16 using a tree diagram.

22 A random digit generator outputs random digits. Each time a digit is

generated, the digit is equally likely to be any one of the digits 0,1,2,...,9.
How many digits must be generated in order for us to be 95% certain that at
least one 7 is generated? [Hint: This means that the probability of getting at
least one 7 should be greater than or equal to .95.]

b23 Show that if a pair of events A and B is both independent and mutually
exclusive, then at least one of the events must have probability zero.

1.24 Suppose that A, B, and C are independent events.

(a) Show that A and B A C are independent.
(b) Show thatA and B U C are independent.

1.25 Draw a Venn diagram for the data in Example 1.1, using circles to
correspond to the categories “low blood pressure” and “men” as opposed to
“high blood pressure” and “women,” as is shown in the text.
32 Chapter 1: The Basics

1.26 In the Venn diagram that follows, A, B, and C represent the sets of letters
given in Example 1.3. For each of the 26 letters of the alphabet, place the
letter in the correct one of the eight regions shown in the Venn diagram.
For example, the letter x would go in the outer region (outside the 3 circles)
as shown in the figure because it doesn’t belong to any of the three events.

P27 In a Venn diagram with three sets, the figure is divided into eight connected
regions.

AaABoac

One of the eight regions is shaded in the figure, and the set that is
represented by the region is described in terms of A, B, C, and the
elementary set operations. Give a similar description for each of the other 7
regions in the figure.

1.28 To check whether two given events are independent, any of three different
equations can be used. (See Definition 1.2.) Prove that the three equations
Problems SiS}

in Definition 1.2 are equivalent. In other words, show that if one of them is
true, then all of them must be true. (This is very easy. For example, you
might first show that equations 1 and 3 are equivalent. By symmetry, an
identical argument would show that Equations 2 and 3 are equivalent.)

1°29 Example 1.22 shows that if a coin is tossed repeatedly until a head occurs,
the probability is 1/3 that the first head will occur on an even-numbered
toss. Give a similar argument to show that the probability is 2/3 that the
first head will occur on an odd-numbered toss.
 Chapter 2: Applications

This section will give further engineering applications of the topics discussed
in Chapter 1. Some such problems involve reliability of systems like circuits and
communication networks. The examples given in this chapter will be elementary,
but they will show the usefulness of the properties and techniques that have already
been encountered.

2.1 Circuits

In Figure 2.1, each box represents a component of a circuit. It is assumed

that current can flow along the lines connecting the components.

Figure 2.1 Three circuits. (a) A series circuit,

(b) a parallel circuit, and (c) a series-parallel circuit.

In the circuits we are interested in determining the probability that the circuit
will carry current, and we assume we know the probability that current can flow
through each of the components. It is common practice to call the probability that
the circuit will carry current the reliability of the circuit, and similarly for the

34
2.1 Circuits BD

individual components. (We will sometimes shorten “reliability” to “rel.”) Let’s

denote by A,B,C, and D the events that components A,B,C, and D
(respectively) in Figure 2.1 can carry current. Furthermore, we will assume that
these events are independent. It is very important to keep this assumption
constantly in mind, because not all circuits actually behave this way (though many
certainly do). For example, failure of one component may cause another to
overheat, thereby making failure of another component more likely.

Series Circuits

Example 2.1. The series circuit in Figure 2.1 (a) can carry current only if all
components are functioning. Thus the reliability of the circuit is

rel=P(ANBOC) =P(A) P@) P(C) (2.1)

More generally, it should be clear that for any series circuit, the reliability is
simply the product of the reliabilities of the individual components, assuming that
the events corresponding to the components being good are independent events.
(Engineers usually describe this assumption by saying that “the components fail
independently.’’)

Parallel Circuits

Example 2.2. The parallel circuit in Figure 2.1 (b) can carry current if at
least one of the components is functioning. Thus rel = P(AUBUC). The
easiest way to evaluate this probability is to recall De Morgan’s law:
(A.GB-U.C)
= ALERB? A.C
The fact that A, B, and C are independent guarantees that A‘°, B°, and C* are
also independent. Thus,
rel. =21 -P( Aw BOC T= 1 Pa4a Pb) PC)

Sa aS (2.2)
where p PR and Pc denote P(A), P(B), and P(C), respectively.
In general, a parallel circuit with n components would have reliability
36 Chapter 2: Applications

rel = 1= C= peep.) Cl ae
where p, denotes the reliability of the ith component, and the components are
assumed independent.

In Examples 2.1 and 2.2 we have expressed the reliability of simple series
and parallel circuits in terms of the reliabilities of the individual components. If
there is more reason to be interested in failure of the circuit than in reliability, the
situation can be turned around and the unreliability of the circuit can be expressed
in terms of the unreliability of the components. The unreliability of a circuit or
component is simply the probability of failure; that is, the probability that the device
cannot carry current. (See Problem 2.6.)

Series-Parallel Circuits

Example 2.3. A series-parallel circuit is a circuit built up from smaller

pieces (or individual components) by hooking modules together in series or
parallel. Figure 2.1c is a simple example. By visual inspection, it is apparent in
this example that the circuit works only if A and B function and if either C or D
does. That is,
rel=P[ANBA(CUD)]
Because A, B, C, and D are independent events, it follows that the events A,
B, and C UD are also independent. Thus
rel P(A) P(B) P(C UD)

P(A) PB) [ P(C) + PD) - P(C) PCD) J

=Pr~PRPc +P,PRPp —PaPpRPcPp

In general, there is no formula to represent the reliability of an arbitrary

series-parallel circuit. However, a simple manual technique is always available in
that the problem can always be done by a “bottom-up” approach. In other words, a
group of components that form a series or parallel module can be treated as a single
“super” component. The reliability of the super component is computed using the
series or parallel formula, and the module is then treated as a single component,
thereby creating a simpler problem to be solved. In Figure 2.1c this would amount
to treating the module consisting of C and D as a super component, computing its
2.2 Networks W/

and the new component.

2.2 Networks

From a mathematical standpoint, there is little to distinguish between a circuit

and a network. Figure 2.2 illustrates this fact.
Networks often have two special nodes called the source and the sink, and
this is true of the ones pictured in Figure 2.2. The links of the network are thought
of as devices that may fail, and as such are analogous to the components in Figure
2.1. The nodes where the links join are analogous to the “wires” that connect the
components of the circuits. It should be clear that mathematically the problem of
computing the reliability of the networks in Figure 2.2 is identical to the problem of
computing the reliability of the circuits in Figure 2.1. Thus the computations
shown above for the circuits apply without modification to the networks.

A B C
source @——————_@—_@—__® sink

source sink

Figure 2.2 These series-parallel networks are

mathematically identical to the circuits of Figure 2.1.

In the case of the circuits, it was assumed that only the components can fail;
that is, we did not consider the possibility of any of the “wires” connecting the
38 Chapter 2: Applications

components failing. Similarly, in the case of the networks we are assuming that the
nodes are perfectly reliable. (In Figure 2.2 the nodes are the black dots at the ends
of the links.)
One current important class of network problems has to do with
communication networks. Visualize a large group of computers or terminals
(nodes) linked together by a variety of communication lines (edges). Since things
like power outages or equipment failures can cause communication links to fail,
there is a clear need to study the reliability of such systems. It is also possible to
study systems in which both the vertices and the edges can fail. We did not allow
for that possibility in treating Figures 2.1 and 2.2 above.

Graphs

Both networks and circuits, when viewed from a purely mathematical point of
view, are instances of what are called graphs. A graph consists of “edges” and
“vertices.” When drawn, the edges are usually lines or curves, and the vertices are
the junction points where edges meet. Thus graphs are visually similar to the
networks of Figure 2.2. To identify a circuit with a graph, one identifies the
components of the circuit with the edges of the graph. The individual connected
clumps of “connecting wires” that connect the components in the circuit correspond
to the vertices of the graph. Graph theory is a branch of mathematics which has
important applications to many types of engineering problems.

Networks That Are Not Series-Parallel

Example 2.4. The nice thing about series-parallel networks (or circuits) is
that reliability calculations can be manually reduced to simpler problems in a
straightforward way. More general kinds of networks do not necessarily lend
themselves to such direct treatment. Figure 2.3 illustrates this. In this figure we
assume that the edges A, B, C, D, and E have known probabilities of being in a
working condition and that these events are independent.
If w and y are the source and the sink, then the network is series-parallel
and the reliability is the probability of the event (A 7B) VU EU(CaD).
If x and z are source and sink, then the problem becomes substantially more
difficult. It is no longer possible to merge series or parallel components. There are
a variety of approaches, however, that can be used. Perhaps the simplest is to
2.2 Networks 39

“partition” the problem depending upon whether link E is working or not. To see
what this means, let’s let G represent the event that a communication route from x
to zis available. Then
P(G) = P(E) P(G| E) + P(E‘) P(GIES
(In this equation the event E denotes the event that link E is functioning.) This is
Proposition 1.1, the multiplicative law. Since P(E) and P(E*) are assumed
known, we yet need to know how to evaluate P(G | E) and P(G| E°).

Figure 2.3 If wand y are viewed as the source and the sink, this is a series-
parallel network. If x and z are source and sink, the network is not series-parallel.

But if edge E is good, we might as well think of w and y as a single vertex

(since we know they are connected). In order to get to this single vertex from x
and from z, at least one of the edges A and B has to be good and at least one of
C andD must be good. This event is (A UB) N(C UD). So,
PG) EY = PL AVUB)A(CUD)]) PC GB)xP(GeD)
= (P, + Pp — PyPp) X Oc + Pp — PcPp)
On the other hand, if edge E is bad, then we can simply remove it from the
picture. In this case, a path from x to z will be open only if A and C are good or
if B and D are good. This eventis (A NC) U(B OD). Thus,
PGIE) = PIAAC)UBAD))
= P(AANC) + PBAD)-PANBOCOD)
Each of these terms can now be expanded as a product because of the assumed
40 Chapter 2: Applications

independence.
There is a conceptual point here that is more important than the nitty-gritty
details. It is that by “pivoting” on the edge EZ, that is, by splitting the problem into
two cases depending upon whether edge E is good or bad, the problem can be split
into two simpler problems. This idea plays a key role even in very sophisticated
techniques and in current research. ;
Even very important contemporary problems needn’t involve highly abstract
concepts. The network reliability problems introduced above illustrate this fact.
Suppose that a group of telephone users are hooked together by a network of
telephone lines of known reliabilities and that whether a given line is down is
independent of whether any of the other lines are down. It is hypothetically
possible to determine the probability that all individuals can communicate with each
other or that any particular subgroup can communicate with each other. This
calculation can be done using a variety of the techniques developed in Chapter 1.
Some of the techniques can be refined into quite powerful tools, while others have
to be discarded as too crude. All methods of determining network reliability are
very “inefficient” algorithms in that the time required for the calculation grows
exponentially with the size of the network. The amount of time or the number of
computations required to perform a calculation is referred to as the computational
complexity of the method. The development of computer algorithms for
performing such analyses is an important area of contemporary research.

2.3 Case Studies: Two Network Reliability Problems

Network reliability problems provide a useful illustration of several of the
basic notions of probability and conditional probability that were discussed in
Chapter 1. In performing parallel and series reductions in a circuit or a network,
we use fundamental properties of union and intersection since a series reduction
corresponds to an intersection of two events and parallel reduction corresponds to a
union of two events. The idea of independence is involved at every step of the
way.
In this section we will examine how using the concept of conditional
probability can help with more complicated systems than series-parallel
configurations. As mentioned earlier, problems such as this are in the forefront of
present-day research because of the increasing importance of communication and
transportation hookups of one kind or another around the world.
2.3 Case Studies: Two Network Reliability Problems 41

Example 2.5. In Figure 2.4, the five circles in the network Go at the top
represent five people linked together via the seven communication lines shown in
the picture as edges e€,,---, 7. We will assume that the reliability of each of the
links is known; that is, that the probability the link is working is known.
Furthermore, let’s assume that the links are independent. As a convenient notation,
we denote by p, the probability that link e, is good.
The problem we are going to investigate is the problem of determining the
reliability of communication between the two people represented by the black
nodes. What are some possible approaches that could be used?

Tree Diagrams

Here we are looking at seven links, each of which is going to be either good
or bad. To represent all states of this system in a tree diagram, we would need
to have seven layers of branching in the tree, one layer for each line. Since the
number of outcomes shown in the tree will double for each layer, we will wind up
with 27 = 128 outcomes in the tree.
In reality it isn’t necessary to draw the entire tree in order to do this problem.
For instance, suppose we were to start out by considering edge e3 first and edge
€ second as we start building the tree. If both these links are bad, then certainly it
is going to be impossible for the two black nodes to communicate. Therefore, there
isn’t any point in continuing to draw the part of the tree that would continue on
from the assumption that that these two links have both failed, if all we are
interested in getting from our tree diagram is the probability that the black nodes can
communicate. Even so, this problem clearly taxes the methodology of a tree
diagram to its limit. While it may be (barely) feasible to do this problem with a tree
diagram, given a very large sheet of paper, it wouldn’t be much fun.

Finding All Possible Communication Paths

There are clearly many routes that could be used to send a message from one
black node to the other. A few are e€3 e4 és and é2 &€¢ and é3 €; €7 €s5. For
any one of these possible paths, we can easily compute the probability of the path
being available because of the independence we are assuming for the links. For
example, the probability that the path e3e4e5 is a working path is simply the
42 Chapter 2: Applications

P11 = P4a+ P7—-P4P7

P12= P2+ P3- P2 P3
Ps= P3 Pg
Gs

P9= Pst P2— Pg Po

Pi0= Pe + Ps— Pe Ps

Figure 2.4 Network reliability using

recursion and conditional probabilities.
2.3 Case Studies: Two Network Reliability Problems 43

product p3 x p4 x ps, Since it depends on all three of the three edges involved
being good. Therefore, we could list all the possible paths and compute the
probability for each path. The problem with this approach, however, is that the
different paths aren’t mutually exclusive. For example, consider the two paths
€3€4€5 and €3€,€7€5. Let’s think of event A as being the event that the first
of these two paths is good and event B as the event that the second is good.
Clearly A and B aren’t mutually exclusive, because it’s quite possible that both
paths might be good, and so we can’t compute P(A U B) just by adding P(A)
and P(B). The best we can do is
P(A UB) =P(A) + P(B)-P(A MB) (223)
=P3P4P5 + P3P1P7P5 — P3P4P5P1P7
This computation is easy, since A B, the event that both paths are good,
depends only upon é3, €4, €5, €;, and e7 being good, and the probability of this
happening is given by p34Ps P P7-
So if there were only two possible paths connecting the black nodes, the
computation we have just described would do the trick. The problem is that there
are many such paths, not just two. Let’s think of n as representing the number of
paths between the black nodes and A,,---,A, as being the events “path 1 is
good,” “path 2 is good,” and so forth. Since what is needed in order for
communication to be possible is for at least one of the paths to be good, the task is
to compute the probability P(A; U---UA,).
There is a standard way to do this using what is sometimes called the
inclusion-exclusion principle. In fact, Equation 2.3 above is the special case of
this principle when n = 2. For n = 3 the inclusion-exclusion principle says that
PACE OC) = PA) EPGy+PC)
—P(ANB)—-PAANC)-P(BOC)
+P(AANBAC)
For the union A, U- - - U A,, the inclusion-exclusion principle says that
the probability P(A; U - - - U A,) can be computed by adding the probabilities of
each of the individual events, then subtracting off the probability of the intersection
of all possible pairs of the events, then adding on the probabilities of all
intersections of three events, subtracting off the probabilities of all intersections of
four, and so forth. While it’s not terribly difficult to see why this principle is
correct (in fact, it can be derived from Equation 2.3 using mathematical induction),
it clearly isn’t a recipe that’s going to be very pleasant to use if n is large.
44 Chapter 2: Applications

Network Reductions Using Conditional Probabilities

In earlier examples we have seen how to use series and parallel

simplifications to determine the reliability of a series-parallel circuit or network.
Even in networks that are not of a totally series-parallel structure, it’s still often
possible to make such simplifications. The problem with the network Gg in
Figure 2.4, however, is that there aren’t any edges in the network that can be
reduced this way, and so there is no apparent way to get started.
A useful starting point is to focus on a single edge (any one will do) and
consider the two possibilities: Either the link is good or it is bad. We can use the
concept of conditional probability as embodied in the multiplicative law of
Proposition 1.1.
To make things specific, let’s fix our attention on edge e,. Denote by E the
event that this edge is good so E* is the event that edge e, is bad. We also need a
name for the event whose probability we are trying to compute, so let’s designate
A as the name of the event that the two black nodes in network Gg can
communicate. Proposition 1.1 says that we can determine the probability of this
event by
P(A) = P(E) P(A | E) + P(E°) P(AIE*)

The usefulness of this little formula lies in the fact that there is a very natural
interpretation to each of the conditional probabilities contained in it. The
conditional probability P(A | £) is the probability that communication is possible
given that the link e, is good. If e, is good,jhowever, then we know for certain
that the two nodes at the ends of e, will be able to communicate, and so any
message that can get to one can automatically get to the other. Why not then think
of them as a single node? In other words, take e; out of the picture altogether and
“collapse” the two nodes at the ends of e, into a single logical node, because being
able to get to one of them is equivalent logically to being able to get to the other one
given that e, is good. What this observation boils down to is the fact that the
conditional probability P(A | E) is the same as the unconditional probability that
the two black nodes can communicate in network G, of Figure 2.4. On the other
hand, the conditional probability P(A | E°) has an even simpler interpretation.
For if €, is bad, we might as well remove it from the picture altogether. The
conditional probability P(A | E°) is the same as the probability that the two black
nodes could communicate if e, weren’t there at all, and this is what is shown in
network G, in the figure. The conditional probability P(A | E°) is the same as
2.3 Case Studies: Two Network Reliability Problems 45

the unconditional probability that the black nodes can communicate in the network
G,.
It is common to refer to G; as the network (or graph) obtained from Go by
deleting the edge e, and to refer to Gy as the network obtained from Gog by
contracting the edge e,. Notationally this is often written as G, = Go — e; and
G» = Go * ej, as is indicated in the figure.
You may not have noticed, but a very important thing has just happened. We
have described how to solve our original problem by expressing the solution in
terms of two simpler problems. Specifically, we have passed from having to deal
with network Go to having to deal with networks G, and Gy. Furthermore, each
of these networks can be further simplified by series and parallel reductions. For
example, e3 and e, in G; are series edges, and G; shows them being replaced
by a single edge eg whose reliability is given by pg = p3x py. Similarly, G,
has €y and é3 as parallel edges, as well as e4, and e7, and G4 shows these pairs
each being replaced by a single edge with the appropriate probability. G7 and Gg
continue this process, and in fact G, can be completely solved by series and
parallel reductions.
In solving G,, however, we are stuck again when we get to G3, because
G; has no series or parallel edges. For that reason we have to repeat the trick we
used to start with, that is, to pick another edge and to delete and contract it leading
to two further problems. In the figure, the edge chosen is e, (the particular choice
of edge is not critical), and deleting and contracting e7 leads to the two networks
G6 and Gs, respectively. Gs is then reduced to Gg and G¢ to Gyo.
All known methods for doing problems such as this one require a number of
computational steps that grows exponentially with the size of the network.
Reduction methods similar to what is shown in Figure 2.4, however, work about
as well as any known method. In fact, it is possible to develop algorithms using
such reduction methods and to run the algorithms on a microcomputer.
Figure 2.5 shows a much more complicated network than the one we have
been considering. In fact it is based on a computer network that existed in the
1970s with source in California and sink in New England. The probability that the
two black nodes can communicate in the network of Figure 2.5 can be calculated on
a microcomputer in less than a minute using an algorithm that implements the ideas
of Figure 2.4. Since the network has 22 links, a complete tree diagram for all
states of the network would show 222 = 4,194,304 outcomes.
46 Chapter 2: Applications

source sink

Figure 2.5 The network reliability problem is to determine the probability that
communication is possible between the source and the sink.

The lengthy example we have been treating is an instance of the two-

terminal, or source-to-sink, communication problem, since we are interested in
determining the probability that two given members of the network can
communicate. Another variation is the all-terminal problem in which we wish to
determine the probability that a// members can communicate with each other.

Example 2.6 (All-terminal reliability). The network G, that is shown in

Figure 2.6 has five nodes, all of which must be able to communicate with each
other. Here there is no concept of a source and a sink, but the concept of series and
parallel edges is still meaningful. For example, the edges e, and ep are in series
in the network on the left. The network G, on the right shows these two edges
being replaced by a single edge e3. Let’s examine the relationship between the two
networks G, and G, in the figure.
First, think about the node v, at the top of network G,, the node where
edges e€, and €2 meet. Such a node is called a degree-two node or a degree-two
vertex because there are exactly two edges coming into the node. In order for this
node to be able to communicate with any of the others, at least one of the edges e,
and @ must be good. Let’s introduce some names for events so as to be able to
write this down as an equation. We will let A denote the event that all nodes can
communicate in the network G, and C denote the event that at least one of the
edges e, and e is good. Then
P(A) = P(C) P(AIC) + P(C P(AIC®)
= P(C) P(AIC)
The last expression comes from the fact that P(A | C*) = 0 in the first equation.
2.3 Case Studies: Two Network Reliability Problems 47

A Hee
SS eeP1 Po a
P1 + P2— Py Po
al €2
V. e3
Vo Vo 2 ¥3

V4 V5
V,4 G
Vi5
G, 2
rel(G;) = (p; ain Obey 1
OF Po) rel(Go)

Figure 2.6 An all-terminal reliability problem. All nodes in network

G, must be able to communicate with all other nodes in Gy.

Clearly P(C) =p, + pz —PiP2, where p, and p> are the reliabilities of
€, and é€9, respectively. What about the term P(A 1C)? Given that C occurs,
we know that the node v, is not cut off from all other nodes. Furthermore, it is
possible that v. and v3 can communicate through v, if both e, and e, are
good. Here’s where things become a bit subtle. If we knew for certain that
exactly one of the edges e, and e, were good, we could simply remove eé), é9,
and v, from the network, and the reliability of the reduced network would be the
same as the original. However, since both might be good, what we need to do is
replace them by link e3 in the reduced network and to assign the proper probability
to e3. The reason for the presence of e3 is to take care of the possibility that e,
and é can provide a communication route from v2 to v3. The probability of this
is the probability that both e; and ey are good, which is p; pz. Given the
information that at least one of them is good, the conditional probability that both
are good is the probability p3 shown in the figure.
The relationship between the two networks is described by the equation at the
bottom of the figure. The reliability of G, is the conditional probability that G,
is good given that at least one of the two original edges e, and e, is good. The
important point though is that the equation
rel(G,) = (Pp, + P2—P}P2) Tel(G2)
enables us to describe the solution to our original problem in terms of the solution
48 Chapter 2: Applications

to a simpler problem. In other words, what is being demonstrated is a simple

algorithm for reducing two edges in series when the problem being solved is
determination of the probability that all terminals can communicate.
Parallel reductions are no different in this example than in previous examples.
So the reliability of the network G, in Figure 2.6 can be determined by continuing
this process. For example, the degree-two vertex v3 in Gz can be removed in a
similar fashion, and this will create parallel edges between vz and v5. After the
parallel edges are reduced to a single edge, we are left with a triangle of three
degree-two vertices. Any of them can be removed in the same way that v, was,
and that will leave us with an elementary network consisting of two parallel edges.
By collecting all the computations, this gives us the reliability of network G, as
some multiplying factor times the reliability of this elementary network.

So What?

Examples 2.5 and 2.6 have been used to show how solutions to important
contemporary problems can be formulated in such a way that the probability
concepts involved are basically elementary. Some other very important problem-
solving techniques have come up also. In fact, it might be worthwhile to list them.

1. Aconditional probability described in one sample space is often simplified by

viewing it as an unconditional probability in another sample space. For
example, if a red and a green die are tossed, the conditional probability that
the sum is 8 given that the red die shows a 5 is the same as the unconditional
probability that the green die shows a 3.
In Examples 2.5 and 2.6, conditional probabilities involving networks
have been logically replaced by unconditional probabilities involving smaller
networks.

2. The concept of recursion is very important in problem solving, and it has

been illustrated here. To illustrate recursion, think about the simple function
fin) =n! for non-negative integers n. It is possible to define f by saying

eae ifn>0O
oS 1 ifn=0
2.4 Fault Trees 49

While this definition looks redundant, it actually is quite usable. For

example, to compute f(5) it says you should first find f(4), but to do that you
have to evaluate f(3), and so forth, until you get down to f(0), at which time
you can wrap up all your calculations.
The equation rel(G,) = (p, + pz — P; Pz) rel(G) in Figure 2.6 is
similarly recursive in that it tells how to solve the all-terminal reliability
problem for network G, in terms of the all-terminal reliability problem for
network G», which is simpler.

3. Different kinds of physical systems can be modeled with the same

mathematical model. Figures 2.1 and 2.2 illustrate the equivalence of some
small circuit and network problems. Figures 2.4 and 2.5 are drawn as
networks rather than as circuits, but they could provide mathematical models
for either. For example, each of the edges in the network could be interpreted
as being a resistor, and the source and sink could be incoming and outgoing
wires to be connected to the terminals of a battery. The white nodes are then
junction points where the wires coming out of the resistors are soldered
together. A resistor is considered good if current can pass through it.
The reliability problem we discussed in Example 2.5 then would be the
determination of the probability that current will flow in the circuit when the
source and sink are connected to the battery terminals.

2.4 Fault Trees

Frequently it is highly desirable to know the likelihood that a complex system

will fail. One tool that is sometimes used to analyze complex systems is a fault
tree. Figure 2.7 shows a hypothetical fault tree that has been constructed to
analyze a pressure tank. The event that is of interest (and which is represented by
the top node in the fault tree) is the event in which the tank explodes.
There are three kinds of nodes that appear in the fault tree. The events
represented by circles are the “leaves” in the tree and represent basic events that
may contribute to failure of the tank. The AND gates and OR gates represent
logical intersections or unions. The very top node in the tree (which represents the
event that the tank explodes) is an AND gate. This means that the tank explodes if
the events corresponding to nodes 2 and 3 occur, that is, if there is high pressure
and valve failure. Similarly, the valve fails (the event corresponding to node 3) if
the valve is defective or if the electricity supply is interrupted.
50 Chapter 2: Applications

tank AND gate (intersection)

explodes

(a OR gate (union)

(a) primary event

high
pressure

pipe loss of valve electricity

plugged coolant defective failure

pipe water supply

rupture cut off

Figure 2.7 Fault tree for failure of a pressure tank.

Sometimes it is possible to construct the fault tree in such a way that the
primary inputs will be independent events. (Some of the most serious defects in
safety analyses, however, have centered around the assumption of independence of
events that, in fact, were later found not to be independent. So independence
assumptions should be made with a great deal of caution. For instance, in Figure
2.7 if water is supplied by an electric pump, then nodes 9 and 7 will not be
independent unless the electricity supplying the pump comes from a different
source than that described in node 7.)
If the primary inputs are assumed to be independent, then it is easy to
compute the probability of the top event by a simple bottom-up hand calculation.
To make things concrete, let’s introduce some hypothetical probabilities:

Event Probability
node 4 pipe plugged at
node 6 valve defective O01
2.4 Fault Trees 51

node 7 electricity failure 05

node 8 pipe rupture .05
node 9 water supply cutoff .02
Then,
P(node 5) = .05 + .02 — (.05)(.02) = .069
P(node 2) = (.1)(.069) = .0069
P(node 3) = .01 + .05 — (.01)(.05) = .0595
P(node 1) = (.0069)(.0595) = .00041055

Normally fault trees require much more subtle approaches than is shown
here. Suppose, for instance, that there were more than one way in which electricity
failure could contribute to failure of the tank. In this case we would have the same
primary input (node 7) appearing at more than one location in the fault tree. This
means that the kind of bottom-up approach used here would not work because
subtrees beneath certain nodes would fail to be independent.
Fault tree analysis was developed primarily within the aerospace and nuclear
industries during the 1960s and 1970s. Better ways of calculating or estimating the
probability of the top event for very large and complex trees (sometimes having
hundreds or even thousands of nodes) are still being sought.

Example 2.7. Draw a fault tree for the circuit of Figure 2.1c. The top event
of the fault tree should represent the event that the circuit can transmit circuit.

Solution: There is not a unique way of constructing such a fault tree. Either
of the following trees would be correct, and, in fact, it should be easy to convince
yourself that these two fault trees are logically equivalent.

ifs a
ae (2) & on
Sao OieViOn "©
Notice that if we wish the top event to represent failure of the circuit, then a
different fault tree is called for. Such a tree is the “dual” of a tree that represents
“non-failure” of the circuit. Simply convert all the AND gates to OR gates and all
the OR gates to AND gates. For example, here are the duals of the two fault trees
above:
52 Chapter 2: Applications

In either of these fault trees the top event represents failure of the circuit if we now
interpret the primary inputs A, B, C, and D to represent failures of the
respective components.

Problems

px (a) What is the reliability of 5 items in parallel if each has reliability .6?
(b) What is the reliability of 5 items in series if each has reliability .6?

2.2 Find the reliability of the circuits shown below. The reliability of each
component is shown with the component.

2.3. Find the reliability of the circuits shown below. The reliability of each
component is shown with the component.
Problems 53

2.4 The following figure is a highway network. Consider each of the 5 links to
function independently of the others.
(a) Assume that for each link the probability is .9 that the link is open.
What is the probability that you can get from “start” to “finish”? (As an
instructive exercise and a check of your answer, it might be worthwhile
to list a sample space for this network. Since there are 5 links (edges),
there are 2° = 32 possible states (outcomes). One such state, for
example, would be to have link A be good, B good, C bad, D bad,
and E good. By the assumed independence of the links, the
probability of this state .
=PAANBAC ADS OE)
= P(A) P(B) P(C*) PD P(E)
The desired probability then can be obtained by just adding up the
probabilities of all of the 32 states for which it is possible to get from
“start” to “‘finish.’’)
(b) Now drop the assumption that each link has reliability .9 and simply
write the reliability of the network in terms of the reliabilities of the 5
links.

start finish

Die Find the probability that a route is open from the source to the sink in the
network that follows. The reliability of each edge is listed alongside the
edge.

source sink

2:6 Examples 2.1 and 2.2 show how the reliability of simple series and parallel
circuits are expressed in terms of the reliabilities of the individual
54 Chapter 2: Applications

components. Show how the failure probability (the unreliability) for the
circuits in Figure 2.1(a, b) can be expressed in terms of the failure
probabilities of the individual components. Let q4, gg, and dc denote the
failure probabilities of components A, B, and C, respectively.

aa, Show how the unreliability of the circuit in Figure 2.1(c) can be expressed
in terms of the unreliabilities of the failure probabilities of the individual
components. (If you have already done Problem 2.6, the meaning of this
should be clear.)

2.8 The circuit in the following picture shows a battery, a light, and two
switches for redundancy. The two switches are operated by different
people, and for each person there is a probability of .9 that the person will
remember to turn on the switch. The battery and the light have reliability
.99. Assuming that the battery, the light, and the two people all function
independently, what is the probability that the light will actually turn on?

battery
switch 2 light

switch 1

Dye} In the figure below, the probability that the battery is good is py, the
probability that the light bulb is good is ps, and the probabilities that the
resistors R,, Ry, and R3 are good are pj, p>, and p3, respectively.

battery R3 cy) light

Problems 55

Assuming that the 5 components function independently, what is the

probability that the bulb will light? (For a resistor to be “good” here means
that current passes through it, and it is assumed that the bulb will light if the
bulb is good and current passes through the bulb.)

2.10 A complex system requires that a certain function be performed with

reliability .9999. Devices that perform the function have reliability only .8,
and so it is necessary to build in redundancy. Several of the devices are to
be used, and if at least one works then the function will be performed.
Furthermore, we are willing to assume that the redundant devices function
independently of each other. How many of the devices must be installed in
order that the system will have the desired reliability?

This problem is a simple example of the all-terminal reliability problem.

Three computers are all linked together over telephone lines. (If you like
you can think of this as a network that is a simple triangle with the edges
being the phone lines and the vertices the computers.) We’ll assume that the
computers are completely reliable (an unreasonable assumption) but that
each of the phone lines has probability .9 of being good. Also assume the
phone lines are independent. What is the probability that each of the
computers will be able to communicate with the other two? [Hint: This
situation is simple enough to be treated via basic principles without recourse
to the ideas presented in Example 2.6.]

202 Suppose A, B, and C are events having the properties that

PAMBNC)= PAC) PBA) and
PAsABIG), =PAIC) PBC)
It would make sense then to call A and B conditionally independent
relative to C and C°. For either C does or does not occur, and the two
equations say that in either case A and B are independent relative to the
conditional probability measure induced by the information that C does or
does not occur.
(a) Box 1 contains 2 red and 1 white items and box 2 contains 1 red and 2
white items. One of the boxes is randomly selected (each with
probability 1/2), and two items are chosen from the box one at a time
and without replacement. Let A be the event that the first item drawn
56 Chapter 2: Applications

is white, let B be the event that the two items drawn are of the same
color, and let C be the event that the items are drawn from box 1.
Show that A and B are independent events, but not conditionally
independent relative to C and C° in the above sense. (This is easy to
check with a tree diagram, but you should also try to understand
intuitively why A and B are independent and why they cease to be
independent once we know what box we are drawing from.)

(b) Suppose again that box 1 contains 2 red and 1 white items and box 2
contains 1 red and 2 white items. Again randomly choose one of the
boxes and draw two items, only this time replace the first item before
drawing the second. Let A be the event that the first item is red, let B
denote the event that the second item is red, and let C be the event that
the items are being drawn from box 1. Show (via a tree diagram) that
A and B are conditionally independent relative to C and C*, but that
A and B are not independent. This says that A and B become
independent once we know which box we are drawing from, but in the
absence of this knowledge, A and B are dependent.

2.13 Fill in the details of the calculation in Example 2.5. Suppose that each of
the links shown in the original network Gg has reliability .9. Determine
then, with the assumption that the links function independently, what the
probability is that the source and sink shown as the black nodes in network
Go will be able to communicate.

2.14 Complete Example 2.6. Assume that all links have reliability .9 and that the
links function independently. What is the probability that every node will
be able to communicate with every other node?

ould Draw a fault tree for each of the circuits of Problems 2.2 and 2.3. First
construct a fault tree in which the top event is the event that the circuit is in
working order, and then give the dual fault tree in which the top event
represents the failure of the circuit. (See Example 2.7.)
 Chapter 3: Random Variables

The study of calculus begins with an investigation of properties of the real

number system, but it does not progress very far before the concept of a function
must be introduced. Similarly, in the study of probability we start with sample
spaces and probability measures, but modeling real problems leads very quickly to
the need to use functions as a way of describing measurements being made or
processes being observed. Generally these functions are real-valued (just like in
calculus), but they measure some relevant information about an experiment or a
process having probabilities associated with the different outcomes or states of the
process. In essence this simply says that the object we are usually interested in is a
function defined on some sample space where there is a probability measure.

Definition 3.1: A real-valued function defined on a sample space S is called a

random variable.

One of the essential things that random variables do for us is to enable us to

avoid dealing with unnecessarily complicated sample spaces. Suppose, for
example, you were doing quality-control analysis of a particular product such as a
line of microcomputers. Some quantities you might be interested in could include
number of repairs needed during the first year, time to the first breakdown, average
number of hours in use per week, cost of repairs, and so on. What is the sample
space here? Conceptually we could perhaps think of the sample space as consisting
of all existing items of the particular type being studied. However, it’s really the
quantities we are measuring rather than the physical items themselves that we are
interested in. The measurements can be thought of as numerical values associated
with each item. For example, the time until the first repair is needed might be 6
months for one item and 3 years for another. In any case we are associating a
number with each item, and that is what a function does. We will see that by
phrasing our discussions in terms of random variables, we can often bypass

Sif
58 Chapter 3: Random Variables

complicated sample spaces and instead do all our work with real numbers where
everything is simpler.
The common practice in studying probability is to use uppercase letters from
near the end of the alphabet such as X, Y, or Z to denote random variables.
While this may seem strange at first, it does enable us to reserve common letters
such as f, g, and so on, for use in the conventional calculus sense, that is, as real-
valued functions of a real variable.

Example 3.1. A pair of dice is rolled. Let X denote the random variable
that indicates the sum on the two.dice. Using the usual 36 element sample space to
represent the experiment, the value that X takes on for the outcome (2, 4) is 6.
The domain of the random variable X is the 36-element sample space, and the
range of X is the set of numbers {2,---, 12}. One might indicate the action of
X on (6, 4), for example, by writing
K(6,4) => 10
(One can of course think of functions in terms of inputs and outputs. In the case of
the random variable X, if we input the element (6, 4) from the sample space S,
then X outputs the value 10.)

Example 3.2. A room full of people is considered as the sample space. The
random variable Y gives the weight of each person; that is, when a given person is
“input” to Y, the random variable Y “outputs” the weight of the person.

Example 3.3. A single die is tossed. Consider the random variable that
indicates the number that appears on the die. Here it is natural to identify the
outcomes of the experiment with the numbers 1, - - - , 6; that is, it is reasonable to
consider the sample space to be the set of numbers S = {1, 2, 3, 4, 5, 6}. If one
does this, then X is simply the identity function on S; that is, X(w) = w for
each element we S. This makes sense in this case, because S is a set of
numbers.

Example 3.4. A coin is tossed twice. Let X be the number of heads that
occurs. If we take S = {HH, H1I,_TH, TT}, then X(HH) = 2, X(HT) ='1,
X(TH) = 1, and X(TT) =0. This is similar to Example 3.1.
Someone may take the following view, however. If the number of heads is
all we are interested in, why not record just that information; that is, why not
Chapter 3: Random Variables 59

consider the sample space to be just S = {0, 1, 2}, where 0 represents the outcome
of no heads, 1 represents the outcome of 1 head, and 2 represents the outcome of 2
heads? In this way we can, in effect, consider the random variable that counts the
number of heads to be simply the identity function X(w) = w on this sample
space.
This latter viewpoint is not really as different as it might first seem. We shall
soon see that what is really needed in order to understand a random variable is
knowledge of the probabilities of the random variable assuming specific values.
Furthermore, the above interpretations are consistent in this respect if consistent
probability measures are used on each of the two possible sample spaces. For
example, the probability of the event {w: X(w) = 1} is 1/2 in either case. In the
former sample space, this event is {HT, TH}. In the latter it is {1}. The latter
sample space is not a uniform one since two of the three outcomes listed have
probability 1/4 and the other 1/2.

Example 3.5. Four bits of binary data are transmitted, so the sample space is
that of Example 1.8. Consider Z to be the random variable that assigns to any 4-
bit pattern the number of 1’s appearing among the 4 bits. So
Z:0110 42
If we identify the sample space here with the numbers 0, - - - , 15, then the random
variable is simply assigning to each number the number of 1’s that appear in its
binary representation.

Example.3.6. Five coins are tossed. The random variable Y indicates how
many heads are obtained. Here Y acts on a uniform sample space of 32 = 2°
possible outcomes for the experiment and Y assumes values {0, 1, 2, 3, 4, 5}.

Example 3.7. Lines of text are being transmitted. Each line is 80 characters.
We will suppose that for formatting purposes it is necessary to know how many
blank spaces are included in each line. Furthermore, we assume that characters
may consist of the 26 letters, 10 digits, or 6 different punctuation symbols, for a
total of 42 different characters.
The number of different possible lines that could be transmitted would then be
equal to 8042. If we take this set of all possible lines as our sample space and if we
wish to count the blank characters, then we should think of the random variable
defined on this sample space that simply assigns to each possible line the number of
60 Chapter 3: Random Variables

blank characters within the line. Clearly the values that this random variable can
assume are restricted to the counting numbers from 0 (no blanks in a line) to 80 (the
whole line being made up of blank spaces).

Example 3.8. A resistor is put into a circuit and the length of time X it lasts
before burning out is measured. This situation is a bit more vague since it is not
clear what the “possible outcomes” are. As a first effort one might try to visualize
some hypothetical batch of resistors as the sample space. Since it is “time to
burnout” that we are interested in, however, it is perhaps more useful to think of the
sample space as a set of numbers representing possibilities for the time to burnout.
In fact, one could construe any non-negative length of time as being hypothetically
possible, in which case the random variable could take on all non-negative values.
With this point of view, the random variable X is simply the identity function, in
other words X(w) = w for all w 20. Clearly in this example we have no guess
as to how to treat probabilities on such a sample space. That will come later.
It is useful to have a short notation for certain events associated with random
variables. For instance, in Example 3.1, the event {w : X(w) = 4} is simply the
set {(1, 3), (2, 2), (3, 1)}.. The notation {w : X(w) = 4} is often shortened to
(X = 4). Similarly, {w:X(w) < 9} is shortened to (X < 9). Other similar
situations are handled in a like manner. And finally, the probabilities of the two
events (X = 4) and (X < 9) are usually written as P(X = 4) and P(X < 9).

3.1 Discrete Random Variables and Probability Mass Functions

Random variables come in two primary “flavors,” discrete and continuous.
A random variable that assumes only finitely or countably infinitely many different
values is a discrete random variable. (A countably infinite set is a set that can be put
in one-to-one correspondence with the positive integers.) The most common
discrete random variables are integer-valued and generally arise in situations where
some kind of counting is being performed. The important information about a
discrete random variable is carried by its probability mass function.

Definition 3.2: Given a discrete random variable X, the probability mass

function py forX is defined for all numbers t by py(t) = P(X = 2).
3.1 Discrete Random Variables and Probability Mass Functions 61

Example 3.9. IfX is the sum when a pair of dice is rolled, then X has
range equal to the set {2,---, 12}. So py(t) takes on non-zero values only
when fe {2,---, 12}. For example, py(5) = 4/36 = 1/9, whereas px(13) = 0.

Example 3.10. If X is the number of 1’s that appear when 6 random binary
digits are transmitted, then py(t) # 0 only when fe {0, 1, 2, 3,4, 5, 6}. If we
think of the stream of digits as an independent trials process, then the values of the
probability mass function are easily obtained from the independent trials formula.
For example, the probability of exactly two 1’s being transmitted among the 6 digits
is py(2) = C(6, 2) (1/2) (1/2)".

Example 3.11. If X is the number of heads to occur in two tosses of a coin,

then py(0) = px(2) = 1/4 and py(1) = 1/2. Notice in Example 3.4 that this is true
for either proposed sample space.

One of the important attributes of the probability mass function is that it

incorporates the important information about the values the random variable
assumes, and the probabilities of these values, without reference to the sample
space at all. Questions about the random variable can be answered by using just the
mass function.

Example 3.12. If 10 random binary digits are transmitted, what is the

probability that more than seven 1’s are included among them?

Solution: If we letX denote the number of 1’s among the 10 digits, then
what we have to compute is

P(X >7) = P(X=8)+PQ@=9)

+P = 10)

Px(8) + px(9) + px(10)

cao. (3) (E)-+cua. ($Y (2)eu. (3)”

1 \10 56
(45 +10 +1) ( ) = 7024
62 Chapter 3: Random Variables

3.2 Continuous Random Variables and Density Functions

One can intuitively think of a continuous random variable as one that takes on
values in some interval of the real line. For example, many continuous random
variables have as their range of possible values the non-negative real numbers. The
concept of the probability mass function does not apply to random variables of this
type. The analogous concept is that of the density function of the random variable.
Fundamentally, when dealing with continuous random variables, one has to ask a
different kind of question. Rather than asking for the probability that the random
variable assumes a specific value, we ask instead what is the probability that it
assumes a value in some interval of numbers. The role that the probability mass
function plays in the discrete case is played by the density function in the
continuous case.

Definition 3.3: A density function is a function f defined on the real numbers

and having the following properties:
1. f(t) 20 forallt

Zz (ftbat =)

Since the endpoints of the integration in Equation 2 are infinite, we must be

prepared to deal with improper integrals. Often, however, these integrals may be
reduced to integrals over an interval of finite length. For example, if a density
function happens to have the property that f(t) = 0 when ¢ lies outside of some
interval [a,b] of numbers, then

[toa = [fio at
The analogy between this definition and the definition of the probability mass
function in the discrete case is that the requirement that the integral equals 1 in
Definition 3.3 corresponds to the fact that the sum of the (nonzero) values of a
probability mass function is 1. Whereas each discrete random variable has a
probability mass function, every continuous random variables has a density
function. The relationship between a continuous random variable and its density
function is given in Definition 3.4.
3.2 Continuous Random Variables and Density Functions 63

Definition 3.4: A random variable X is said to have density function f provided

that f is a density function related to X in the following way:
For all numbers a and b witha < b,
b
PUG AC ap) = | f( at

The values a =—ce and b = o are permitted here, in which case this
becomes an improper integral.

If X is a random variable with a density function, it is common to denote the

density function by fy. Sometimes when there can be no confusion from context,
we will drop the subscript and just write f/ For the purposes of this book, it is
sufficient to think of a continuous random variable as simply a random variable
that has a density function.

Example 3.13. Perhaps the most intuitively natural density function is the
function f defined by f(t) = 1 forO <t< 1 and f(t) = 0 otherwise. If a and b
are two numbers with 0 sa<b< 1, then
eo b
Pa<X<b)= | fO ar=| ldt =b-a

Thus the probability of X assuming a value in a given subinterval of the interval

[0, 1] is simply the length of that subinterval.
This provides a continuous model for the experiment of “choosing a random
number between 0 and 1.”

Example 3.14. Figure 3.1 shows the density functions for three random
variables. Suppose X, has f, as its density function and X, has fj. Since both
functions vanish off the interval [0,1], we know that P(X, < 0) = 0 and also that
P(X, > 1) = 0 and similarly for X2. In other words, X, and Xz are random
variables that assume values between 0 and 1. The graph of f,, however, is higher
toward the right end of this interval, and the graph of f, is higher toward the left
end. This tells us that values for X; are more probable near 1 than near 0, and vice
versa for X>. As a sample computation to illustrate this, let’s compute the
 64 Chapter 3: Random Variables

probabilities P(.8 < X, < 1) and P(O <X, <.2).

57
PO <X, <.2)= i 2t dt = .04
0
whereas,
1 ?
P(8<X, <= | 2tdt= 36
8

The situation regarding X, is exactly reversed.

Example 3.14 illustrates that a simple glance at the density function gives
much information about the random variable. A random variable having the
function g at the bottom of Figure 3.1 as its density function would be one that
assumes values between 2 and 5 with values around 3 being most probable.

2 Z

{ {

{ 1
frit
or eR 2-21 if0<te<1
dite O otherwise
id=
2
0 otherwise

1 2 3 + 5
Figure 3.1 Three density functions.

An important observation is that any random variable X having a density

function automatically has the property that for every number lo, P(X = to) = 0.
3.3 Distribution Functions 65

To see this, simply think of taking a small interval of radius 6 centered at fp. Then
tot6
P(X =1) S$ P(ty-8<X <ty+8) = |x xO dt
t,

and the fact that fis an integrable function guarantees that the integral on the right
tends to zero as 6 > 0. Since P(X = fp) is less than or equal to this quantity for
every positive 6, P(X = fy) must equal zero.
Proposition 3.1 includes the observation we have just made as well as a
useful additional consequence of this fact.

Proposition 3.1: If X is a random variable having a density function, then for

any number fp, P(X = to) = 0, and for any numbers a and b,
Pia <X $b) =Pasx <b) =P(a@<XxX<b)=P(asx sb)

The explanation for the latter part of Proposition 3.1 is simple. For example,
P(asx <b)=(P(a@<X <b) +P?C =a)
= P(a<Xi<b)+0=-Pia<X-<b)
The other parts may be proved in a similar way.
Other examples of commonly occurring density functions will be given in
Chapter 5. Before looking at continuous models, however, it is a good idea to get a
better understanding of the mathematical implications of the relation between a
random variable and its density function. It is important to understand some
parallels between discrete and continuous random variables, and one way to
understand the parallelism is in terms of a third way of representing the “probability
distribution” of a random variable: the cumulative distribution function.

3.3 Distribution Functions

Definition 3.5: For any random variable X (discrete, continuous, or neither),

the cumulative distribution function of X (or simply the distribution function of
X) is the function Fy defined on the real numbers by
Fy(t) =P(X $0) forallt
66 Chapter 3: Random Variables

It is useful to understand what the distribution function for a discrete random

variable looks like and how it relates to the probability mass function. Figure 3.2
shows the graph of the distribution function for the random variable X that counts
the number of heads in three coin tosses. This figure illustrates the fact that for a
discrete random variable, the distribution function.is a step function with a jump
discontinuity at each value of t such that py(¢) #0. Furthermore, the magnitude
of the jump discontinuity is equal to py(?).

jumps by 1/8

————SsSsSS

ss jumps by 3/8

Figure 3.2 Distribution function for the random variable that

counts the number of heads in 3 coin tosses.

Proposition 3.2: If a function F is the distribution function for some random

variable, then F has the following properties:
1. FQ 7 last,
2. F(t) > 0 as t > —29,
3. F is right-continuous on the real line; that is,
pee (t) =F (to) for every fo

4. F is a nondecreasing function; that is, if t) < fj then F(t,) $ F(t).

3.3 Distribution Functions 67

A proof of Proposition 3.2 may be based on the basic properties of

probabilities found in Chapter 1. The first three parts of Proposition 3.2 are limit
properties and require use of Axiom 3.1 in Chapter 1.
One should be very aware that the word distribution is often used in a very
loose sense. In light of Proposition 3.3, it is straightforward to pass from the
density function to the distribution function, and vice versa. The term distribution
is sometimes used to describe a density function and sometimes a distribution
function. For example, one might say that Definition 4.1 in the next chapter
describes the “binomial distribution,” although literally it is the probability mass
function of a binomial random variable that is being described. Similarly, it might
be said that Definition 5.2 encountered in Chapter 5 describes the “exponential
distribution,” though in fact it is the density function that is defined there.
In the case of a continuous random variable, the fundamental theorem of
calculus gives the relation between the density function and the distribution
function. The relation is shown in Proposition 3.3. The important point here is
that the distribution function and the density function for a continuous random
variable may easily be derived from each other. If either one is known, the
computation of the other boils down to an integration or a differentiation. This
frequently makes calculations with continuous random variables easier than
calculations with discrete ones.
®

Proposition 3.3: Suppose X is a random variable having density function fy

and distribution function Fy. Then these two functions are related as follows:
i)
t F x(to) = { eo) dt

2. < [Fy(t)] =fy( for all values of t where fy is continuous

The similarity between the discrete case and the continuous case is illustrated
by Figure 3.3. The integrals encountered when working with continuous random
variables replace the sums that appear in the discrete case, and the density function
replaces the probability mass function.
68 Chapter 3: Random Variables

Expression Discrete case Continuous case

b
P(asX <b) Pee | Felt) dt
ast< b a

ty
F x(t) DY pxlty [feo at
iat ny

Figure 3.3 Integration replaces summation in the

case of continuous random variables.

Example 3.15. Suppose X is a random variable having density function

defined by f(f) = 2t for 0 <t < 1, with f(t) = 0 otherwise. (This is the function
f;, of Figure 3.1.) Find the distribution function Fy.

Ss

Solution: Fyy(s)= [ Od = 05 s=<0

F,(s) = { ar dt=s? if0<s<1

S

0
1
Fys)= | 2r dt=1 ifs>1
 3.4 Expected Value and Variance 69

Notice that the derivative of Fy is fy at all points where fy is continuous,

which in this example is everywhere except at t= 1. Furthermore, notice that Fy
is a continuous function, whereas fy is not (because of its discontinuity at t = 1).
_ The distribution function for a continuous random variable will always be a
continuous function. The density function may or may not be.

A little should be said about the “‘not quite uniqueness” of the density function
for a random variable. Since the relation between a random variable and its density
function depends solely on integrating the density function, strictly speaking the
density function is not unique. (Two different functions can be density functions
for the same random variable.) To see this, simply think about changing the value
of the density function at some finite set of numbers. Making such a change would
not change the value of the integral of the function over any interval. For example,
in the case of the uniform density on the interval [0,1], it doesn’t matter whether we
say f(t) = 1 on the closed interval 0 <¢ <1 or the open interval 0 <t<1. The
only difference in the two involves the value of f(t) at the endpoints 0 and 1, and
the value assigned at these two places will not affect the value of the integral of f
over any subinterval of the real line. These observations notwithstanding, it is a
common practice to speak of the density function of a random variable with the
realization that this is a slight abuse of language.
When dealing with continuous random variables, Proposition 3.1 says that
one can disregard the endpoints of intervals since the probability of the random
variable assuming a value in a given interval is the same regardless of whether the
endpoints are included or not. With discrete random variables this is not the case.
For a discrete random variable X, P(a < X <b) and P(a < X <b) are not the
same in general, the reason being simply that there is no longer any guarantee that
P(x. = a). and P(X = b) are 0. In tact, it is easy to see that P(a@=<X <b) =
P(a<X <b) if and only if P(X =a) = P(X =b) = 0.

3.4 Expected Value and Variance

The expected value, or expectation or mean, of a random variable is a
weighted average of the values assumed by the random variable. The notation
commonly used for the expected value of a random variable X is E(X). For
example, if X is the number obtained when a fair die is rolled, then the expected
70 Chapter 3: Random Variables

value is E(X) = 3.5, which is just the average of the numbers 1 to 6 that constitute
the values of X.
If Y is the number of defects in a sample of two devices drawn from a batch
of 5 of which 2 are defective, then it is easy to check (using a tree diagram) that
P(Y:=:0) = 3, PY =) = Omen =2)-= ai ts Caccrr ae anes
defined as :
EQ y=" 02S) 1X6 2 oe as

This example illustrates that the expected value of a random variable is indeed a
weighted average, where the weights are simply the probabilities of the respective
outcomes. The rationale is that if an experiment were conducted a large number of
times, each leading to an “observed value” of the random variable and if each value
assumed by the random variable were to be “observed” exactly as frequently as
predicted by the probability mass function, then this theoretical weighted average
would indeed match the true average of all the observed results of the sequence of
experiments. The fact that E(Y) = .8 in this simple illustration reflects the fact that
if the experiment were repeated a large number of times and if no defects were
found 30% of the time, 1 defect was found 60% of the time, and 2 were found 10%
of the time, then the average number found would be .8. This idea is incorporated
in Definition 3.6.

Definition 3.6: If X is a discrete random variable, then the expected value, or

mean, of X is defined by

where X1, X2,--- represent the values X assumes.

In case X assumes infinitely many values, then the sum on the right is to be
interpreted as the sum of an infinite series.

If X assumes infinitely many values, then there is no guarantee that the series
in the above definition converges. If the series fails to converge, then the mean of
the random variable is not defined. As a matter of fact, when one speaks of a
random variable having an expected value, it is generally understood that not only
does the series above converge, but that it converges absolutely.
If we keep in mind that for continuous random variables integration replaces
 3.4 Expected Value and Variance 71

summation, the following definition should be expected (no pun intended).

Definition 3.7: If X is a continuous random variable with density function f,

_ then the expected value of X is defined by

E(X) = irtf(t) dt

Just as with the sum in the discrete case, there is no guarantee that the
improper integral here will converge. Thus when one says that a given continuous
random variable has a mean or expected value, the statement implies that the integral
converges, and again common usage is to interpret this as meaning that the integral
actually converges absolutely.

Example 3.16. Suppose X is a random variable having as its density

function the function /;, of Figure 3.1. The expected value of X then is given by

Ew =| dt = [rend = 2x5], = 3
co 1 P 1

A very important property of expected value is the property of being a linear

operation. Proposition 3.4 states this linearity property.

Proposition 3.4: IfX and Y are random variables (with finite expected value)
defined on the same sample space and if a and D are real numbers, then
E(aX + bY) = a EQ) + DEY)

The proposition asserts that the expected value of aX + bY will exist

automatically (the sum or integral will converge) if X and Y have expected values.
In this case, the proposition says that the expected value of a linear combination of
X and Y is the corresponding linear combination of their expected values.
This fact is actually deeper than it might appear. For a finite or countably
infinite sample space, the proposition can be proved by first showing that
72 Chapter 3: Random Variables

E(Z) = Y, Zowy) PCOwp))

for any random variable Z, where the sum extends over all elements w, in the
sample space. For then,

E(axX
+ bY) DY) aX + bYVGH) PCCw4))
a >, XC) P{wy}) +b >, Yow, PC(wp))
=a E(X) + bDE(Y)

For continuous random variables there is no elementary proof of Proposition

3.4. In higher level treatments of probability theory (treatments preceded by a
study of real variables, including measure theory), Proposition 3.4 arises naturally
in the theory of integration. In any case, it should be apparent that Proposition 3.4
can be used repeatedly to apply to any linear combination of a finite number of
random variables.

Example 3.17. A construction project requires 30 components. Of these, 10

have been acquired from manufacturer A and 20 from manufacturer B. If 8 of the
components are randomly chosen to be used in the first stage of construction, what
is the expected value for the number of these that come from manufacturer A?

Solution: If we consider X to be the number among these 8 that come from

manufacturer A, it is not terribly difficult to determine the probability mass
function for X. Clearly X can assume values 0,1, --- , 8, and the probability of
each of these values can be computed using Proposition 1.2, the combinations
formula. The number of different ways that 8 components can be chosen from the
30 is C(30, 8). To compute the probability that, for example, 3 of these come
from manufacturer A, we need to know how many of these C(30, 8) ways of
choosing the components for stage one would involve using 3 of manufacturer A’s
components and 5 of manufacturer B’s. The number of ways to pick 3 of A’s is
C(10, 3), and the number of ways to pick 5 of B’s is C(20, 5). Thus the number
of ways to use 3 of A’s and 5 of B’s is C(10, 3) x C(20, 5). Dividing this
number by the total number of possible outcomes, C(30, 8), would give
P(X=3). If this method is used to compute P(X = 0), ---, P(X = 8), then
3.4 Expected Value and Variance 73

E(X) may be determined from Definition 3.6.

Proposition 3.4, however, offers a simple alternative. Let’s think of the 8
components to be used in stage one as being chosen one at a time. (This is a mental
trick; it makes no difference whether they are actually chosen one at a time or
simultaneously. Is that clear to you?) We will consider X, to be 1 if the first one
chosen comes from manufacturer A and 0 if it comes from manufacturer B.
Similarly X, ---, Xg tell whether the others among the 8 being used come from
manufacturer A. So X,, Xz,---,Xg are very simple random variables.
Each has probability 1/3 of assuming value 1 and 2/3 of assuming value 0 since 1/3
of the total come from A and 2/3 from B. So E(X,) = E(X) =... = 1/3.
Furthermore, the relation they have to X is that

X=X,+X2+---+Xe

In light of the linearity property of expected value,

E(X) = E(X)) +--+. + EX)

Each of the expected values on the right is 1/3, and so this means E(X) = 8/3. A
little reflection will tell you that this is the only reasonable answer. For if 1/3 of the
total available components are from manufacturer A, then surely the expected
number in our chosen lot of 8 should be 1/3 of these, which is 8/3.

A trick encountered in this example is so useful that it deserves special

comment. Often it is helpful to have a random variable that serves as a flag to
indicate whether a particular event occurs or not. The procedure used in the last
example was to introduce a random variable that takes on value 1 if the event occurs
and value 0 if it does not. While other values could be used besides 0 and 1, the
nice thing about that choice in the preceding example was that to find how many of
the events occurred we then needed only to add the values of the corresponding
random variables. Given an event E, it is common practice to refer to the random
variable that takes on value 1 if E occurs and value 0 if E does not occur as the
indicator random variable for the event E. We will have occasion to use this trick
from time to time in future chapters.
For a given random variable X, E(X) is a number that gives a useful
characteristic of the random variable. An alternate notation that is frequently used
for E(X) is Ly. In the study of statistics, Ly is often referred to as the mean of
X. Another useful number is the variance of the random variable.
74 Chapter 3: Random Variables

Definition 3.8: IfX is a random variable with expected value Ly, then the
variance of X is defined by
var(X) = El (X — Ly)"]
The standard deviation of X is the square root of var(X). It is common practice

to denote the standard deviation of X by oy and the variance by of

Whereas the mean represents the theoretical average value, the variance
indicates the extent to which values of the random variable tend to concentrate
around the average (leading to a small variance) or fluctuate greatly (leading to a
large variance). The variance or standard deviation is used as a measure of how
“spread out” the values of a random variable are. If the values of X tend to cluster
tightly around the mean Ly, then the variance of X will be small. On the other
hand, if the values of X vary widely with high probability, then the variance will
be large. It is of course possible that (X — Ly)? not have finite expected value,
which is just to say that not all random variables have a finite variance.
Definition 3.8 is useful in that it gives a good intuitive understanding of what
it is that the variance measures. There is an alternate way of computing the
variance, however, that is often more convenient. This alternative is given in
Proposition 3.5. Proposition 3.5 says that if we know the mean, then all we need
to know additionally is E(X?).

Proposition 3.5: The variance of a random variable X is given by

var(X) = E(X’) - [E(x)]} and_ var(aX) = a? var(X)

Proof: E{(X -py)"] = E(X? — 2pyX + py?)

= E(X?) — 2uyE(X) + py?
E(X*) — Quy? + py?
E(X?) — [E(X)?
Also,

var(aX) = E[(aX)"] -[E(ax))

= E(a*X*) — [aE(X)|*
3.4 Expected Value and Variance 75

= a’ E(X”) — a°E(X)*
= q’ var(X)

While most random variables encountered in practice fall into one of the
categories discrete or continuous, it is not hard to visualize situations where a
mixture of the two might occur. First, let’s examine a situation that leads to a mix
of two continuous random variables of the type found in Problem 3.14.

Example 3.18. Two different manufacturers supply a component with an

exponentially distributed lifetime; that is, the length of service the component gives
until it fails is an exponentially distributed random variable with a density function
of the type found in Problem 3.14. Manufacturer A’s device has expected lifetime
4 months and manufacturer B’s has 10 months. A particular user has a batch of
devices of which 40% came from manufacturer A and 60% from manufacturer B.
If a randomly selected device from this batch is used, what is the probability
distribution for its lifetime and what is the expected value for its lifetime?

Solution: We will denote by X the random variable that gives the time to
failure for the randomly selected device. What we must do is to partition the sample
space according to the manufacturer of the chosen device. If we think of E and F
as the following two events,
E = event that device is made by manufacturer A
_F =event that device is made by manufacturer B
then for any t> 0,
P(X > t) P(E) PX >tlE)+P(F) PX >tlF)
Set a 6 eo
The exponential terms in this expression come from the simple calculation
called for in Problems 3.14 and 3.15. (By Problem 3.15, the expected value of an
exponential random variable is the reciprocal of the parameter A in the distribution.
The reciprocal of 4 is .25 and the reciprocal of 10 is .1). Thus the distribution
function forX is given by
Fy) = 1-.4e°7>' - 6e°" whent>0
The density function then is
fx) = 167? + .06e7" fort >0
76 Chapter 3: Random Variables

Problem 3.20 asks you to find the expected value for the random variable X,
that is, the expected time to failure for the randomly chosen component under
discussion.

Example 3.19. We will now consider a variation on the problem just

considered. Suppose that manufacturer B leases its components rather than selling
them. What one buys then is a 10-month service contract rather than a single
component. If the original fails during that period, the manufacturer replaces it at
no additional charge, but whatever unit is in service is removed at the end of 10
months. In this case, we know exactly how much service will result from a
purchase from manufacturer B.
If we continue to assume that 40% of a given user’s purchases are
components from manufacturer A and 60% are 10-month service contracts from
manufacturer B, what now is the probability distribution for the amount of service
obtained from a randomly chosen purchase?

Solution: Again we can let E and F be described by the following:

E = event that the purchase is a component from manufacturer A
F = event that the purchase is a service contract from manufacturer B
Then the length of time of service provided can be thought of as a random
variable X, and for any number ¢ > 0,
P(X >t) = P(E)
PX >tl£E)+PC)
PX >tlF)
The first term on the right is still .4 e~?5! for the same reasons as in the
previous example. However, the term P(X >1f1F) will be 1 if t < 10 and will
be Oif t2 10. So we have
Aer 6 HES 10
P(X>t) =
atl hit, if
r>10
The distribution function for X is then given by

Ad -—e?") ift<10
Fy(t)=1-PX>f = u
A) lest anes if
7>10
The graph of this distribution function is shown in the following picture.
From the form of the equations defining Fy(2), it is apparent that there is going to
be a jump discontinuity of magnitude .6 at t = 10.
3.5 Functions of a Random Variable LF

5 10

If you bear in mind that a discrete random variable always has a step function
as its distribution function and that a continuous random variable always has a
continuous function as its distribution function, it is apparent here that we are
looking at a distribution function that is of neither type. In particular, there will be
no probability mass function or density function for this random variable. For
distributions of this “mixed” type, the cumulative distribution function is the only
tool available in computations.

3.5 Functions of a Random Variable

A random variable is simply a real-valued function defined on some sample

space S. This means that if g is some ordinary real-valued function defined on
some part of the real line that includes the range of the random variable X, we can
consider the random variable g(X) which is the composite of the two functions.

Example 3.20. If X is the number obtained when a die is rolled and g is

the function g(t) = 72, then g(X) = X2 is the random variable (defined on the
same sample space as X) that gives the square of the number obtained when the die
is rolled. The values that g(X) assumes are 1, 4, 9, 16, 25, and 36, each with
probability 1/6.

Example 3.21. If X is the random variable that represents a random number

drawn from the unit interval [0, 1], and g(t) = 12, then g(X) = X? is the
random variable that would be used to model “squaring a random number between
0 and 1.”

Example 3.22. Suppose S = {w,,w 2, w3} and X(w;) = 3, X(w) =

5, and X(w3) = —3. Furthermore let’s assume that S is a uniform sample space
in which each of the elements of S$ has probability 1/3.
78 Chapter 3: Random Variables

Now consider the composite random variable Y = g(X), where g(t) = th

that is, Y= X*. It is easy to see that Y assumes only two values, 9 and 25.
Thus, according to the definition of expected value,
Zz, i 43
EY) =9 P(Y = 9) +25 P(Y =25) =9 x a 25x abies
This computation involves the probability mass function of Y. It is very useful to
notice that E(Y) can be computed by using the mass function of X. For in fact

E(Y) = 9P&X =3) + 9 P(X =-3) + 25 P(X =5)

Recall that the definition of E(Y) requires us to know the probability mass
function for Y itself. The last expression for E(Y) given here depends very
heavily on the fact that Y is a function of X. This method, as shown in
Proposition 3.6, can be applied when one random variable is expressed as a
function of another.

Proposition 3.6: If X is a discrete random variable and if Y = g(X), then

EW) = >, gap) P& =x)

where the sum extends over all values x, assumed by the random variable X.

The importance of this proposition is in situations where the probability mass

function of X is well understood, and so the sum appearing in Proposition 3.6 can
be computed to obtain E(Y). If Definition 3.6 were used instead to compute
E(Y), it would first be necessary to know the probability mass function for Y
itself. This can all be summed up by saying the following: Jf Y is a function of X,
then E(Y) can be computed using py without the necessity of knowing py.
As shown in the discussion preceding Proposition 3.6, the proof of the
proposition depends only upon rearranging and combining terms in the sum which
defines E(Y). Each term of the form yjPY = y;) in the sum that defines
E(Y) can be split into a sum of terms of the form g(x,) P(X=x,) where
8(X,) = yj.
In the continuous case there is a similar fact that is stated in Proposition 3.7.
The proof depends on a fairly sophisticated use of knowledge from advanced
calculus or real variables.
3.5 Functions of a Random Variable 79

Proposition 3.7: If X is a continuous random variable and if Y = g(X), then

the expected value of Y is given by

BM= | eOfo at

Again, as in the discrete case, the importance of this proposition is that it

allows computation of E(Y) without the necessity of knowing the probability
density for Y itself.

Example 3.23. Let’s consider taking the square root of a random number
between 0 and 1. Denote by X the random variable representing the random
number between 0 and 1, and the assumption is that the probability distribution of
X is the uniform distribution on [0, 1] as in Example 3.13. Let’s denote Y = VX
, in which case Y = g(X), where g is the function g(t)=Vr .
According to Proposition 3.7,
a 1
BO) =| Vi fy(t) at’ = [ Vt dt= :

In this example it is not difficult to work out the density function for Y and
find the value of E(Y) from Definition 3.7. In fact, for any t 2 0,
Fy) =P st) =PVX st) =PK sr) =Fy(’)
In Problem 3.3 the distribution F y is determined. From that exercise it follows
that
0 ifrs0
Fy(t) = + Pif0<t<1
Lire
Differentiation of Fy gives the density function for Y:
Zeb <7 <I
fy) = |
OQ otherwise

Now E(Y) can be computed from Definition 3.7:

80 Chapter 3: Random Variables

u 5 :
E(Y) =| tfy(t) dt = I i(2t) dt = =
Except in elementary cases, such as the above, it is often not easy to
determine the density function for Y = g(X) even if g is simple and X has an
easily described density function. Proposition 3.7 is very valuable in enabling
computation of E(Y) without needing to know the density function for Y.

Example 3.24. Suppose X is a random variable having density function

equal to the unbounded density function

= ye eee |
fo=
0 otherwise
of Problem 3.19(a). What is the variance of X?

Solution:

po= fra =z and Bx = f PL a= t

G 2Ne 3 Oa NE. 5

Therefore var(X) = 1/5-—1/9 = 4/45.

Problems

3.1 Graph the cumulative distribution function for the random variable that
counts the number of heads in four coin tosses.

3.2 Often it is necessary to construct a probability distribution based on a set of

data. For example, suppose that 100 people have been checked by a
dentist, and the breakdown on the number of cavities found is as follows.
No. of cavities No. of people with this many cavities
0 37
1 22
2. ry
3 ja
4 2
a 2
6 0
a 2
Problems 81

Let’s consider the experiment described as “randomly choosing a person

from this group and counting the number of cavities.” If X denotes this
random variable, then clearly the probability mass function forX should be
P(X = 0) = .37, P(X =1) = .22, and so on. Sketch a graph of the
distribution function for this random variable.

eye) Suppose X is a random variable with density function equal to the uniform
density function on the interval [0, 1] as in Example 3.13. Determine the
distribution function of X and sketch its graph.

3.4 A continuous random variable X has density function defined by f(1) =

1/2 for -1 <t <1, with f(t) = 0 for other values of t.
(a) Compute P(0 <X <.5).
(b) Find the value Fy(.5).
(c) Now find a formula for Fy(‘) for all t with-1 <t<1.
(d) Sketch a graph of the distribution function of X.

ees) Determine the probability mass function and sketch a graph of the
distribution function for the random variable that gives the largest number
appearing when a pair of dice is rolled.

3.0 Sketch a graph of the distribution function Fy for a random variable X

whose probability mass function is given by py(x) = .05 if x = 0, .1, .2,
3, +++ 1.8, 1.9 and py(x) = 0 for other values of x.

S:7 If X has the density function f(t) = 3e-*! for t = 0, with f(t) = 0
whenever t < 0, find P(X > 1) and P(1 < X < 2).

3.8 Show that the function defined by f(t) = (1/2)e~"" for all t is a density
function, and for a random variable X having this density function, find
P(X! < 1).

3.9 Suppose f(t) = c(4 — #7) for -2 <t < 2, with f(t) = 0 otherwise.
Determine the value that c must have in order for f to be a density
function.

3.10 Sketch a graph of the distribution function of the random variable of

Example 3.9.
82 Chapter 3: Random Variables

By A word is chosen at random from the following list of words: horse, dog,
cow, elephant, pig. (“At random” in this context means that each word
has an equal chance of being chosen.) Let X be the random variable that
tells the number of vowels in the word. Determine the probability mass
function for X, sketch the graph of Fy, and find E(x).

S42 A factory produces transistors. Of those produced, 10% are defective.

Four transistors produced by the factory are tested, and X is the random
variable that tells how many defects are found. Find the probability mass
function for X, and graph the distribution function. (Consider this an
independent trials process in which each transistor tested constitutes a trial.)

3513 Suppose X is a continuous random variable having density function f(f) =

2e for t 2 0, with f() = 0 forr< 0. Let Y= 4X.
(a) Compute E(X) from Definition 3.7.
(b) Based on Proposition 3.4, what is E(Y)?
(c) Use Proposition 3.7 to compute E(Y) using the density for X.
(d) Determine the density function for Y and then compute E(Y) directly
from Definition 3.7.

3.14 Show that if X is a random variable having the density function

re for r>0
AR eer er
where A > 0, then P(X >t) = e™.

3:15 Show that the expected value of a random variable having the density
function of Problem 3.14 is E(X) = 1/h.

3.16 Suppose X is uniform on [0, 1]. Determine and sketch a graph of the
density function and the distribution function of Y = 2X + 3.

a7 Suppose X is a continuous random variable and that Y = aX +b, where

a and b are real numbers with a > 0. Show how to write the density
function for Y in terms of the density function for X. [Hint: First
determine the relation between the distribution functions, and then
differentiate.] You will use this technique in exercises in Chapter 5.
Problems 83

3.18 A device is tested. If it is found to be defective, then two more devices

from the same lot are tested. However, if the first device is found to be
okay, then only one additional item is tested. Assume that the devices are
independent and that they come from a batch in which 10% are defective.
Let X denote the random variable that gives the total number of defective
devices encountered during this testing procedure.
(a) Draw a tree diagram to represent this experiment.
(b) Find the probability mass function for X.
(c) Find the expected value of X.

Bek? Verify that each of the following is a density function:

Pere
zara ifQ<r<1 ll ifld<l
(a) f(®) = : (b) f(t) = .
0 otherwise 0 otherwise

(Notice that the first of these two functions is unbounded, so the integral in
Definition 3.3 must be interpreted as an improper integral.)

3:20 Verify that the function defined by f(t) = .1e7?>! + .06e7" for all
numbers t¢ > 0, with f(t) = 0 for t < 0, is a density function and find the
expected value of a random variable having such a density function. (This
is the final step in Example 3.18).

Brat Find var(X), where X is the number of heads in 4 coin tosses.

(a) Use Definition 3.8 and Proposition 3.6.
(b) Now do it using Proposition 3.5.

Ore If X is a random variable whose density function is given by f(t) = 1/4

whenever —1 < t < 1 and f(t) = 3/4 when 1 < ft < 2, sketch a graph of the
cumulative distribution function for X.
 Chapter 4: Discrete Models

This chapter will demonstrate how some of the most common discrete
probability distributions are used to model real-world phenomena. Two of these
have been presented in Chapter 1, though not by name. They are the binomial
distribution and the geometric distribution. Both are intimately involved with
independent trials processes.

4.1 The Binomial Distribution

Definition 4.1: A random variable X is said to be binomial with parameters n

and p (where n is a positive integer and 0 < p < 1) provided that X assumes
values 0, 1, 2, ---,7, and
P(X =k =C(n, k) p* (1 —p)"*
for k= 0, 1,2,---,n

Clearly the natural way in which a binomial random variable arises is in

counting the number of “successes” in an independent trials process, since the
expression for P(X =k) is simply the independent trials formula of Chapter 1.
For example, if X is the number of days in a year in which a certain electrical grid
experiences a power outage, and if the probability of an outage occurring on any
given day is .01, then assuming this to be an independent trials process means that
X is binomial with parameters n = 365 and p =.01.
Frequently an independent trials process is called a Bernoulli process or a
sequence of Bernoulli trials. (Jakob Bernoulli was a seventeenth century
mathematician who wrote one of the earliest treatises on probability.) The special
case of a binomial random variable for which the parameter n is equal to 1 is called
a Bernoulli random variable.
The mean and the variance of a binomial random variable are given in
Proposition 4.1. This information will be used in many applications.

84
4.1 The Binomial Distribution 85

Proposition 4.1: Suppose X is a binomial random variable with parameters n

and p. Then
1. E(x) = np and
2. var(X)=npq whereg=1-—p

The first part of this proposition is extremely plausible. If 10 coins are tossed,
the expected number of heads is 5. (Here n = 10 and p = 1/2.)
Some texts provide proofs for this proposition that depend on rather tedious
observations involving binomial coefficients. It is more instructive to focus on the
relationship between the first equation E(X) = np and the linearity of expected
value, Proposition 3.4.
Consider nv trials in an independent trials process with success probability p.
It is useful to introduce random variables X,,---,X, to indicate whether
success occurs on each of the n trials. Specifically, let X, be the random variable
that assigns value 1 to any outcome of the sequence of trials in which success
occurs on the first trial and assigns value 0 to any outcome in which failure occurs
on the first trial. Similarly, X, = 1 if success occurs on the second trial and X, =
0 if failure occurs on the second trial, etc.
Now if we look at X =X, +--.-+X,, the random variable X will tell how
many successes occur in the v trials. If you are at all confused about what is going
on here, it might be a good idea to consider a special case, such as n = 3. In that
case the sample space can be represented by a tree diagram or can be viewed as
{SSS, SSF, SES, SFF, FSS, FSF, FFS, FFF} where S is success and F is
failure. Give the value of X,, Xz, and X3 for each of the eight elements of the
sample space, and make sure you understand that X =X, + Xz + X3 is indeed
the number of successes in the three trials.
By the linearity property of expected value (Proposition 3.4), the expected
value of X is E(X) = E(X,) +---+E(X,). However, for each k from 1 to
n, X, assumes only the values 0 and 1, and E(X;,) = P(X, = 1) =p. Thus
E(X), being a sum of n terms each of which is equal to p, is in fact np.
A short proof of the second part of Proposition 4.1 will be given in Chapter 6
as an application of the fact that variance is additive for “independent” random
variables. (The concept of independence of random variables is introduced in
Chapter 6.)
86 Chapter 4: Discrete Models

4.2 The Geometric Distribution

The geometric distribution is rooted in the concept of geometric series

introduced in Chapter 1. Like binomial random variables, random variables having
such a distribution arise in observing independent trials processes.

Definition 4.2: A random variable X that assumes positive integer values is said
to be geometric with parameter p provided that for each positive integer k
PX =k) =q*"p
where g=1-—-p.

Geometric random variables are used to model discrete “waiting time”

phenomena. For example, consider an independent trials process with success
probability on any given trial being p. If X is the number of the trial on which a
success first occurs, then it is not hard to see that P(X =k) = g*1p. The
reason for this is simply that in order for the first success to occur on the kth trial,
it must be true that the first kK—1 trials result in failure. Assuming that the process
being observed is an independent trials process means that the probability of this
happening is gp.

Example 4.1. Suppose that in a given floodplain, the probability of a flood in

any given year is .2. Assuming that a structure has just been built in the floodplain,
let's consider the random variable X that indicates the number of the year in which
the structure is first subjected to a flood. Thus, saying X = 1 means that a flood
occurs during the first year of existence of the structure, and so forth.
Assuming the occurrence of floods to form an independent trials process (some
meteorologists might quibble with this assumption) makes X geometric with
parameter p= .2. Thus the probability that the structure is first subjected to a
flood in its fifth year is P(X = 5) = .2x.84 = .08192. It is implicit in the model
here that a flood eventually must occur. This is seen by simply noting that the sum
of the probabilities P(X = 1) + P(X =2)+---=1. (See Problem 4.16. )
This is a discrete time model because time is being measured in terms of whole
years rather than on a continuous scale. Clearly one could observe the same
phenomena and choose to record time on a continuous rather than a discrete scale.
To do so leads to consideration of a continuous rather than a discrete random
4.2 The Geometric Distribution 87

variable. This alternate way of modeling the waiting time for a flood to occur will
be considered in the next chapter.

One of the most interesting properties of geometric random variables is the

“lack of memory” property stated in Proposition 4.2.

Proposition 4.2: If X is a geometric random variable, then

PX =n+k\iX>n) = P(X=k)
for all positive integers k and n.

Proof: Suppose X is geometric with parameter p. Then

P(X =k)=pxqt!
where g = 1-p. For a given positive integer n, it is easy to see that P(X >n)
= q". The reason is that this probability is simply the sum of the infinite series
P(X =n+1) + P(X =n+2) +... =pq"+pq™!+...
=pq'l+q+q°t--:)

An alternative and intuitively helpful way to observe that P(X >n) =q” is
to think of X as modeling waiting time for success in an independent trials
process. Then P(X >n) is simply the probability that more than n trials are
required, that is, that the first trials result in failure. However, the probability
that the first n trials result in failure is just q”.
If k > O then the intersection of the events (X >n) and (X =n +k) is the
event (X =n +h), since (X =n +k) is a subset of (X > 7). Thus
P(X=n+k) _ pqr**7!
P(X
(oie =n+kiX>n) ole = apn) o
pq! =P(X =k)
Let’s interpret the meaning of this property in light of Example 4.1. Suppose
in that example that a flood does not occur during the first five years. The
conditional probability that the first flood occurs in year 12, based on this
88 Chapter 4: Discrete Models

information, is then exactly the same as the initial (unconditional) probability that a
flood would occur in year 7. In other words, there is no penalty for the 5 years of
good luck. The conditional probabilities after the 5 years of good luck are the same
as if we think of the process as starting all over again.
An example that is simpler still is the independent trials process consisting of
repeated tosses of a coin. If we think of the random variable X that indicates the
number of the trial on which the first head occurs, then X is geometric (with
parameter p = 1/2 if the coin is unbiased). The lack of memory property says, for
example, that if tails occur on the first two tosses, then the probability that the first
head occurs on the fifth toss is the same as the original probability that the first
head occurs on the third toss. In other words, you can forget about the first two
tosses that have resulted in tails and think of the whole process as starting all over
again. This mathematical model is consistent with most people’s intuitive feeling
that a coin doesn’t have a memory, and it provides a little more evidence supporting
the choice of an independent trials model as the “correct” one for repeated coin
tossing.

Proposition 4.3: If X is a geometric random variable with parameter p, then

1
E(X)(X) = —:

Proof: By definition,

E(X) = ¥ Px =n) = » npqr! = oe

alt ned n=l

It is a pleasant fact from calculus that power series may be differentiated term by
term. This means that if

f(x) = Ya,"
n=0
then
4.2 The Geometric Distribution 89

In particular, if f is defined for lvl < 1 by the geometric series

f@=1+xt+xr+...= 1-x
then

1
Ff = + 0+ 3x +. ;
(=x)
In particular, this implies that the infinite series that defines E(X) in the first
line of the proof does sum to p x (1 — quer = plp” = I/p.

Example 4.2. A quality-control engineer has possession of a box of devices.

He is going to test the devices one at a time until he encounters one that does not
meet design specifications. How long should this process take?

Solution: The easiest model here is an independent trials model. In such a

model we would consider each device tested as an independent trial for which the
results of the test are pass or fail (success or failure). It’s not entirely clear that this
model is perfect for this situation. For example, if the devices are aligned in the
box in the order in which they come off the assembly line and if the assembly line
goes through periods of malfunctioning, then it is possible that the “independence”
required by an independent trials process might be lacking. For instance, if a given
device in the box is bad, then it may be an indicator that there was something wrong
on the assembly line as the device was going through, and in turn this might mean it
is more likely that the others that came off the assembly line at about the same time
are also bad.
On the other hand, it’s also possible to visualize a situation where the defects
are caused by some random phenomena, or where the devices have become all
mixed up later, so that the testing of different devices does approximate independent
trials. If an independent trials model is used, then the expected number of devices
that would be tested in order to find a defective one is E(X) = 1/p, where p is the
fraction of the devices that are defective. For example, if 1% are defective, the
expected number that would have to be tested in order to find a defective one is then
E(X) = 100.

Example 4.3. Proposition 1.1 in Chapter 1 is valid even when the sample
space is partitioned into an infinite sequence of mutually exclusive events
90 Chapter 4: Discrete Models

A}, A2, A3, - +>

In that case we simply interpret the sums as infinite series. Let’s look at a simple
situation in which this idea is useful. Suppose a nickel and a dime are tossed
simultaneously and that this experiment is repeated over and over. What is the
probability that we will obtain a heads on the nickel before we obtain a heads on
the dime, that is, that the first head on the nickel will appear on an earlier toss than
the first on the dime?

Solution: Here we are actually interested in two quantities that correspond to

geometric random variables: the number of the toss on which a head first appears
on the nickel and the number of the toss on which a head first appears on the dime.
Let’s denote by X the number of the first toss when a head is obtained on the
nickel, and by Y the number of the first toss in which a head is obtained on the
dime. The problem then is to compute P(X < Y). The events
(X = 1), X=2),.--:
can play the roll of A;, A,,--- in Proposition 1.1. Then from the multiplicative
law of Proposition 1.1,

PRGcY)i= PWis1) PX <= Vx = 1) oP =D) PX <V Xea 2) ee

= (s\a)+lala)+(e)G)
STD Ca gay
=rtr¢PeAtis where r= >
=ril¢+rt+rt:--)

Ee MEone ee, See

Tho. Pehle
A)Ata Anan oe
We have made use of the facts that

PX <VIX=1)=5

PX <YIX=2)=4

PIX <Y1X=3)= 5
and so on, in the derivation of the above series. Do you understand the basis for
this? See Problem 4.17.
4.3 Discrete Uniform Random Variables 91

4.3 Discrete Uniform Random Variables

Suppose S is any set with n elements. A uniform probability measure on S

is one which assigns the same probability to each element of S. So to make S into
a uniform sample space all that is required is to assign probability 1/n to each
element of S. This idea was heavily used in Chapter 1, and in Chapter 5 the close
relation between this situation and the continuous uniform distribution (introduced
in Chapter 3) will be examined.
Imagine the experiment of “choosing a random number from the interval
{O, 1].”. The continuous model for this would be to assume that the random
variable X denoting the number chosen has as its density the uniform density on
[O, 1].

Distribution function for

continuous random variable with
uniform density on interval [0,1]

a Distribution function for

Tee discrete uniform random variable
— used to model same phenomenon

Figure 4.1. Comparison of the distribution functions for continuous

and discrete models for choosing a random number between 0 and 1.
92 Chapter 4: Discrete Models

Suppose, however, that someone prefers to truncate the chosen number to a

single decimal place. Let’s denote by Y the truncated value. Thus, Y assumes
values 0, .1, .2,---, .9 each with probability 1/10. It is highly instructive to
compare the distribution function for Y, which is a discrete random variable, with
that of the continuous random variable X. These are both shown in Figure 4.1.

4.4 Poisson Random Variables

The Poisson distribution is less easily motivated than those considered to this
point. It is commonly used to model situations that involve some kind of random
phenomena such as malfunctions of equipment, calls coming in to a switchboard,
cars entering a parking lot, etc. In each of these situations, the Poisson distribution
is used to model the number of times the phenomena occurs during a fixed time
interval. For example, the number of calls received at a telephone switchboard
during a 30-minute time interval might be assumed to be a Poisson random
variable. A more thorough discussion of these applications will come after the
discussion of Poisson processes in Chapter 7.

Definition 4.3: A random variable X that assumes non-negative integer values

is said to have a Poisson distribution with parameter 4 (A > 0) provided that for
each non-negative integer k

To note that these probabilities sum to 1 requires remembering the usual series
expansion for the exponential function:
n

eos >.=< for all real numbers x (4.1)

aR

Replacing the x in this expansion by A, we can see that the reason that the e~*
factor is present in the definition of the Poisson mass function is simply to make the
probabilities sum to 1.
4.4 Poisson Random Variables 93

Example 4.4. Suppose the number of cars entering a certain parking lot
during a 30-second time period is known to be a random variable having a Poisson
mass function with parameter 1 = 5. What is the probability that during a given 30
second period exactly 7 cars will enter the lot? What is the probability that more
than 5 cars enter the lot during this time period?

Solution: P(X =7) =e x 5’/7! = .104445. Note that Problem 4.9

shows that in this example the average number of cars entering the parking lot
during a 30-second time interval is also equal to 5; that is, the expected value is the
same number as the parameter 4 that appears in the mass function.
P(X >5) =1-P(X =0) - P&® = 1) -----—P( = 5). Each of these
terms on the right side of the equation is computed in the same way that P(X = 7)
was just computed.

A further consideration of Example 4.4 gives more insight. Suppose it is

known that during a time period of 100 minutes, exactly 1000 cars entered the
parking lot. (This is again an average of 5 cars every 30 seconds.) A particular
subinterval of 30 seconds duration constitutes 1/200 of the total time. So we might
think of the 1000 cars as 1000 independent trials, with each car having probability
1/200 of entering the lot during the given 30-second subinterval. From this point of
view, the number of cars entering during the 30-second subinterval would be a
binomial random variable with parameters n = 1,000 and p = .005. In that case,
the probability P(X = 7) would be
C(1000, 7) .0057 .995°°? = .104602
Notice how close this answer is to the computation in Example 4.4 where an
apparently quite different assumption is made about the probability distribution of
the random variable in question. Numerically what is happening is summarized by
saying that if n is “large” and p is “small,” then the binomial distribution with
parameters n and p is approximately equal to the Poisson distribution with
parameter A = np. A good understanding of precisely why this is so will have to
wait until a discussion of Poisson processes in Chapter 7.

Proposition 4.4: If X is a Poisson random variable with parameter A, then

1. E(X)=X and
2. oy =A
94 Chapter 4: Discrete Models

The proof of the first part of Proposition 4.4 depends only on the series
representation for the exponential function, Equation 4.1. (See Problem 4.9.) The
derivation of the variance of the Poisson distribution requires more manipulation
with power series and is a more difficult computation. (See Problem 4.18.)

4.5 Hypergeometric random variables

Suppose 10 items of type A and 20 items of type B have become all mixed
up, and 7 items from the total are to be chosen at random. What is the probability
that 3 will be of type A and 4 of type B?
This question can be treated with the combinations formula (Proposition 1.2)
from Chapter 1. The number of possibilities when 7 items are chosen from the total
of 30 items is C(30, 7). If 3 are of type A, this could occur in C(10, 3) different
ways. The 4 of type B can be any 4 of the 20 of type B, and there are C(20, 4)
different 4-element subsets of set B. Therefore, the number of ways of choosing 3
of type A and 4 of type B is the product C(10, 3) C(20, 4). The probability of
obtaining this distribution of items then, when 7 are randomly selected, is
C(10, 3) C(20, 4)
C(30, 7)
This is an example of the hypergeometric distribution. It arises when
sampling is performed from a batch of objects that have been grouped into
categories.

Definition 4.4: A random variable X has a hypergeometric distribution with

parameters n, M, and N (positive integers with n < N and M <N) if
C(M, k) C(N
- M,n—-k)
Baas eae C(N, n)

Familiarity with the hypergeometric distribution is useful in understanding the

two basic kinds of sampling. These two sampling methods can be described as
sampling with replacement and sampling without replacement. Specifically, the
hypergeometric distribution arises in sampling without replacement. In the
language of sampling theory, the parameters in Definition 4.4 are as follows: n is
the sample size, N is the population size, and M is the number of items in the total
4.5 Hypergeometric Random Variables 95

population of a particular type, say type A. The random variable X then

corresponds to the number of type A items that are chosen when n items are
chosen from a total population of N objects. The rationale is basically the same as
that used in the example that precedes the definition. The denominator is the
number of possibilities for the sample drawn, n items being drawn from the
population of N objects. C(M, k) is the number of ways that k items could be
chosen from the M items of type A. And C(N —M,n-—k&) is the number of
ways that the other n —k items in the sample could come from the N — M items
that are not of type A.

Example 4.5. Of the items produced on an assembly line, 5% are defective.

From a group of 17 items taken from the line, what is the probability that exactly 2
of the 17 will be defective?

Solution: We will assume an independent trials model. (Any other approach

would require more information than we are given.) Assuming the 17 items to be
independently produced means that the probability of 2 defects will be viewed as
the probability of 2 successes in 17 trials where p = .05. This would give
PO = 2) = C7.2).05..05 = 1575
where X is a binomial random variable with parameters n = 17 and p =.05.

Example 4.6. Assume now that 17 items are to be tested from a batch of 100
items of which 5% are defective. What is the probability that exactly 2 in the
sample of 17 will be defective?

Solution: The question that has to be decided first is this: What are we going
to do with an item after we test it? Are we going to lay it aside, or are we going to
put it back in with the others so that the same item might get tested again? (In the
latter case we would not be testing 17 different items necessarily.)
First let’s suppose that we lay items aside once they are tested, so that there is
no chance of testing the same item twice. In this case 17 items are being chosen
from the batch of 100, of which 95 are good and 5 defective. If X is the number
of defectives in the sample of 17, then X is a random variable having a
hypergeometric distribution with N = 100, n = 17, and M=5. And
CS, 27 C5, 15)
P(X =2)=
CUO ae La
96 Chapter 4: Discrete Models

This constitutes sampling without replacement.

Suppose that we do replace each item after testing. In that case each item
tested can be any of the 100 items, of which 5 are defective. The results of any
given test are independent of previous tests, since tested items have been replaced in
the population. This puts us back in an independent trials situation, and so the
probability of 2 defects among the 17 tests would again be
CUT N52 95 s= 1515
as in Example 4.5. This is an example of sampling with replacement. Sampling
with replacement leads to the binomial distribution, whereas sampling without
replacement leads to the hypergeometric distribution.

4.6 Probabilities Based on Observed Data

Often it is useful to study the distribution of values in some fixed data set. The
following example illustrates the relation between this idea and the concept of a
probability distribution. One can, of course, divorce the data set analysis from
probability altogether, but the idea of a histogram, or frequency distribution, is in
fact quite similar to the ideas of the probability mass function and the density
function.

Example 4.7. Failure data is being maintained for a particular type of

equipment. For 250 items that have been in service for a year or more, the
manufacturer has compiled the data which appears in the table below.
If we wish to assign probabilities for number of failures, a direct method of
doing so is obvious. Since 93 of the 250 pieces of equipment experienced no
failure during the year, we would say that the probability of no failure based on
this data is 93/250 = .372. This leads to a sample space consisting of 7 elements
with nonzero probability, specifically S$ = {0, 1, 2, 3, 4, 6, 8}. The probability of
8 failures in a year is only 1/250 = .004.

No. of failures Observed frequency

0 93
1 98
2 44
2 id
4 2
2) 0
4.6 Probabilities Based on Observed Data 97

6 1
i 0
8 1

There are two important observations to make about this simple situation. One
is that the decision to include or exclude the outcomes 5 and 7 shown in the table is
arbitrary. If they are included, they will be assigned probability 0. However,
having them included with probability 0 serves no useful purpose. Rather than to
include them in the sample space with assigned probability 0, it is simpler to leave
them out. In reality, there is no more reason to include the outcomes 5 and 7 in the
sample space with assigned probability zero than there is to include the values 9 and
10 or any other numbers. The only thing special about 5 and 7 is that they happen
to be included in the table of data above.
The second important point in this example is that a probability model based on
this data does not necessarily have predictive value for other data. For example, if
this data were accumulated for items made at one plant, it does not mean necessarily
that similar equipment made at another plant would have the same failure rates. Nor
does it mean that equipment from the same plant made during a different time
period would have similar failure rates. The fact that 5 or 7 failures were never
observed during the period that the table covers does not mean that the next piece of
equipment made at the plant will not fail 5 or 7 times during the first year.
Different data sets may or may not reflect a similar distribution. An analysis of
the “correlation” between data sets is a question for statistics. The important
probabilistic concept is that the probabilities introduced in this example to describe
this data set do only precisely that: The probabilities describe the probability
distribution of the given data set.
The following figure is a histogram. Histograms are frequently used to
display frequency distributions. In the figure the values 5 and 7 are shown only to
maintain the linearity of the horizontal scale. Notice that the idea of a histogram is
basically the same as the idea of displaying the probabilities of the various
outcomes. If the values on the vertical scale were divided by 250, the heights of the
bars would be the probabilities of the outcomes. This would then be a visual
representation of the probability mass function of the random variable that gives the
number of failures during the first year of operation for the equipment described by
the data set. (This would not technically be a “graph” of the probability mass
function because the mass function satisfies p(x) = 0 except for a few integer
values of x. However, it could be construed as the graph of the density function
98 Chapter 4: Discrete Models

of a closely related continuous random variable. See Problem 4.15.)

Problems

4.1 Three cards are drawn without replacement from the 13 spades in a deck of
cards. Let X denote the number of face cards drawn. (The ace, king,
queen, and jack are the face cards.) Determine the probability mass function
for X.

4.2 If a coin is tossed repeatedly, what is the probability

(a) that the first head occurs after the Sth toss?
(b) that exactly 3 heads occur in the first 10 tosses?

4.3 A random variable X assumes values 1, 2, 3, ---, 100, and for each
integer k where 1< k < 100, P(X =k) = .01. Using & notation, give a
sum that would evaluate each of the following:
(a) E(X) (b) var (X) (c) E(Y) where Y = 2*

4.4 A coin is tossed 3 times. Let X denote the number of heads obtained. Let
Y be the absolute value of the number of heads minus the number of tails.
Compute E(Y) in each of the following ways:
(a) Write Y as a function of X; that is, determine g such that Y = g(X)
and then use Proposition 3.6.
Problems 99

(b) Determine the probability mass function for Y and compute E(Y)
directly from Definition 3.6.

4.5 Suppose S = {w,,---,w,} is a uniform sample space, and suppose X

is a random variable defined on S. For 1<k <n, let Xn =X(w,).
Show that E(X) is simply the arithmetic average of the numbers x, --- ,
X,, that is, show that
1
E(X) = — x (4 + +++ + Xp)

4.6 Suppose X is a Poisson random variable with parameter A. What is the

“most probable” value for X to assume? In other words, what value of n
has probability such that P(X =n) > P(X =k) for all other positive
integers n #k? Show that if A is itself an integer, then there is a tie for the
“most probable value” of X.

4.7 Suppose that it is known that the number of components to fail in a complex
electrical device during a one-day time period is a random variable X having
a Poisson distribution with parameter 2 = 5. What is the probability that
there will be exactly 3 failures next Tuesday?

4.8 Suppose it is known that a particular electrical device had 50 components fail
during the past 10 days. What is the probability that exactly 3 should have
failed yesterday? (The model you should use is that of an independent trials
process. Each of the 50 breakdowns could have occurred yesterday or on 1
of the other 9 days. Considering the days equally likely means that a
particular breakdown had a 1/10 chance of occurring yesterday and a 9/10
chance of occurring one of the other days.)
Notice that the answers to this problem and to Problem 4.7 are almost
identical. In each case we are looking at a random variable for which E(X)
= 5. (For the Poisson case, this is a consequence of Problem 4.9.) In fact
there is a very close relationship between the Poisson and the binomial
distributions which these problems illustrate. This relationship will be
explored in Chapter 7.

4.9 Show that if X is a Poisson random variable with parameter 1, then E(X)
=}. [Hint: This is easy. Just write out the series that defines E(X) and
100 Chapter 4: Discrete Models

make use of Equation 4.1.]

4.10 Determine the variance of the discrete uniform random variable that assumes
values 1, 2, 3, 4, and 5, each with probability .2.

Generalize Problem 4.10 as follows: Suppose X is a discrete uniform

random variable assuming values x,,--+-,X,, each with probability 1/n.
Use the expression for var(X) given in Proposition 3.5 to derive an
expression for the variance of X. (Use notation.)

4.12 A random variable X gives the voltage output from an acoustical transducer.
The voltage varies from —3.5 to +3.5 volts, but X measures the voltage as
‘rounded off to the nearest integer. Assuming that the probability distribution
forX is the discrete uniform distribution and that X assumes the values
{—3, -2, -1, 0, 1, 2, 3}, find the cumulative distribution function for X and
sketch its graph.

4.13 A “black box” transmits binary data in the form of 0’s and 1’s which go into
the box and are transmitted out according to the following probabilities.
Let’s denote by X and Y the random variables that give the input digit and
the output digit, respectively. The relation between X and Y is as follows:
P(Y =01\X =0)=pp and PY =11X =0) =1-pp.. On the other
hand, if a 1 is input then the probability is p, that a 1 is output and 1 —p,
that a 0 is output.
(a) If the probability is 1/3 that a given input digit is 0 and 2/3 that it is 1,
find the probability that the output digit is a O and the probability that it
isa 1. These will be expressed in terms of pg and p;.
(b) If an output digit 1 is observed, what is the probability that the input
digit was aQ? (This is a conditional probability question. Continue to
assume that the unconditional probability of 0 as the input digit is 1/3.)

4.14 The number of letters mailed out from a business office in a day was
surveyed over a period of time. Results of the survey were as follows:

No. of letters No. of days observed

0 3
1 4
v2 9
3 14
Problems 1014

= 12
5 9
6 5
I 5
8 4
9 1
10 or more 4

Construct a probability distribution based on this data with the outcome “10
or more” considered as one of the elements of the sample space. Construct a
histogram for this data set, and sketch a graph of the distribution function of
the random variable that corresponds to the number of letters leaving the
office in a day.

4.15 A discrete random variable X assumes values 1, 2, and 3 with probabilities

1/2, 1/3, and 1/6, respectively. Do you see a relation between X and the
continuous random variable Y having density function equal to the step
function that takes on values 1/2, 1/3, and 1/6 and whose graph is shown
below?

Specifically, show that Fy(t) = Fy(t) if tis an integer. Now compare the
graphs of Fy and Fy.

4.16 Use the idea of a geometric series to show that for any geometric random
variable X, P(X =1)+ P(X =2)+ P(X =3)+---=1.

4.17 In Example 4.3, give an explanation for why P(X < Y|1X = 1) = 1/2,
P(X <Y1X =2) = 1/4, and PX < Y1X =3) = 1/8, and so on.

4.18 Show that if X¥is a Poisson random variable with parameter A, then

Oy =A
[Note: This problem should be attempted only if you are comfortable with
102 Chapter 4: Discrete Models

E(X?). One way to do this is to write E(X*) in the form Ae%g(a),

where g(A) is the sum of a power series. While it is not possible to evaluate
the series defining g(A) directly, the sum of the series can be found by
looking instead at the antiderivative of g(A), by applying Equation 4.1 to
this series, and then differentiating to obtain g(A).]

4.19 A motel owner has bought 7 television sets from UltraView TV Corporation.
If 26% of UltraView’s televisions have to be returned for repair during the
first year of operation, what is the probability that the motel owner will have
to return more than two of his sets? (Make sure that you know how to frame
your solution in terms of the binomial distribution.)

4.20 The motel owner of Problem 4.19 put 4 of his 7 televisions in private rooms
and 3 in the lobby. If 3 of them have to be returned for repairs during the
first year, what is the probability that 2 of the 3 to be repaired are ones that
were in the lobby? [Hint: Assume that the seven televisions are all equally
likely to fail. Frame your model here in terms of the hypergeometric
distribution. What are n, M, and N? Can you summarize explicitly what
assumption you are making about this question when you decide to model it
using the hypergeometric distribution?]

4.21 Suppose the motel owner of Problems 4.19 and 4.20 has 3 instances during
the year of a guest damaging a TV so that it must be repaired. What is the
probability that two of the repairs are to televisions in the lobby?
This question is ambiguous in that it is not explicitly stated whether the
same television may be damaged more than once. Common sense would
suggest that it might be; so let’s assume that each case of vandalism is
equally likely to happen to any of the 7 televisions. Which of the probability
distributions discussed in this chapter is your model then based on? What is
the relation between this problem and Problem 4.20?

4.22 Suppose you were going to sample the opinion of 200 registered voters in
the United States as a gauge of public opinion on some issue. Would it make
a significant difference whether you chose sampling with replacement or
sampling without replacement? Why or why not?
 Chapter 5: Continuous Models

Continuous random variables arise when quantities are being observed that
may assume values throughout some interval of numbers. The necessary
information to do computations involving such random variables is carried by the
density function. In Chapter 3 we saw a few elementary examples of density
functions, and in this chapter we will see how some of the most common are used
to model real-world phenomena. Just as was the case with discrete models, the
skill that the modeler must bring to a problem involving continuous random
variables is the experience and intuitive understanding necessary to judge what kind
of distribution “fits” the situation.

5.1 Continuous Uniform Random Variables

We introduced the uniform density on [0, 1] in Chapter 3. This concept is

just as natural for an arbitrary interval. Keep in mind that we will have to adjust the
value of the function in order to keep the area under the graph equal to 1.

Definition 5.1: The uniform density on an interval [a,b] is the function

defined by
1
ifa <t<b
fossa =
QO otherwise

A random variable that has this density function is said to be uniform (or
uniformly distributed) on the interval [a, b].
The reason f(t) must assume the value 1/(b — a) on the interval [a, b], of
course, is that the area under the graph is required to be 1, in accordance with
Definition 3.3. It should be apparent from the symmetry of this density function

103
104 Chapter 5: Continuous Models

that the mean of a uniform random variable is simply the midpoint of the interval
and that the variance is greater the longer the interval. (See Problem 5.1.)
Figure 5.1 shows two uniform density functions, the uniform density on [0,
1/2] and the uniform density on [3, 5]. The area under the graph (which is required
to equal 1) is shaded in each instance.

SS

N uniform density on [0, 1/2]

N
uniform density on [3, 5]

Figure 5.1 Two uniform density functions.

5.2 Exponential Random Variables

Definition 5.2: A random variable X is said to be exponential with parameter A

provided that X has density function
rne™ for r= 0
fe
1 0 fort<0O
Here A is required to be a positive constant in order to make f(t) integrable.

Exponential random variables were encountered in Chapter 3, for example in

Problems 3.13, 3.14, and 3.15. The effect that the specific value of A has on the
shape of the density function is illustrated by Figure 5.2. When A is large, the
associated random variable will have a high probability of assuming a value near 0.
When A is small, the values assumed by the random variable are much more spread
5.2 Exponential Random Variables 105

out. Figure 5.2 shows two scenarios corresponding to A = 2 and A = 1/2.

If X is an exponential random variable with parameter 4, then E(X) = 1/2
(see Problem 3.15). This characteristic, that the expected value is the reciprocal of
the parameter in the distribution, is shared with the discrete geometric distribution;
if a discrete random variable X is geometric with parameter p, then E(X) = 1/p
(Proposition 4.3). In fact, the similarity between geometric and exponential
random variables is substantial. The exponential distribution shares the same lack
of memory property as does the geometric, and the exponential distribution is used
to model continuous “waiting time” situations in the same way that the geometric is
used to measure discrete “waiting time” for success in an independent trials
process. This application of the exponential distribution will be treated in depth in
the investigation of Poisson processes in Chapter 7.

exponential density
with parameter A =2 ; :
exponential density
with parameter 2 = 1/2

1 2 3 1 2 3 4 5
Figure 5.2 Two exponential density functions.

Proposition 5.1: If X is an exponential random variable, then for any positive

numbers s and ¢,
PX >st+thx>o=PX
>s)

Proof: First observe that for any positive number w,

P(X>w)= A i ea =e
Ww

(This computation has appeared previously as Problem 3.14.) From this, it follows
106 Chapter 5: Continuous Models

_X(s+t)
that P(X >s+t)=e , and so
—s+t)
P(X>sttiX>)= -M
=e
= P(X >s)
é

Example 5.1. Suppose that time to first failure of an electrical component is

an exponential random variable with expected value 25 months. If X denotes the
random variable that measures time to failure and if X is assumed to have an
exponential density function, then the knowledge that E(X) = 25 months requires
that the parameter A in the distribution be A = .04 (so that 1/A will be 25).
In the context of these assumptions, what is the probability that the device
lasts longer than 30 months?

Solution: From Problem 3.14, if s >0 then P(X >s) = eS. Thus

P(X > 30) = 6 0% 4 = 6! = 30119

Now by way of comparison, consider a discrete model for the same
experiment. Suppose someone wants to measure the month in which first failure
occurs as an integer value. Let Y be this integer-valued random variable. If the
sequence of months is considered to be independent trials in which “success”
means that the device fails and “failure” means it doesn’t, then Y will be the
number of the month in which the first failure of the device occurs. Therefore Y
will have a geometric distribution, and if E(Y) = 25 then the parameter p in the
geometric distribution must be p = .04. So P(Y > 30) = (1 —p)°9 (as observed
in the proof of Proposition 4.2) = .96°° = .29385. (One could, of course, identify
the failure of the device with the outcome called “failure” in the independent trials
process. While the wording seems more natural this way, we would then be
considering “waiting time for failure” rather than “waiting time for first success,”
which was the terminology we used in discussing geometric random variables. It
should be clear that the differences are purely linguistic and have no effect on the
underlying mathematics.)
In the continuous model for this “waiting time” experiment, P(X > 30) and
P(X 2 30) would of course be the same since P(X = 30) = 0. In the discrete
model, however, P(Y = 30) = P(Y > 29) = .9629 = .30610. Notice that when
the discrete and continuous models are compared, the discrete model yields a
slightly higher or slightly lower probability than does the continuous model,
depending on whether we interpret the question as being “‘what is the probability
5.3 Normal Random Variables 107

that the first failure occurs after month 30” or as being “what is the probability that
the first failure occurs during or after month number 30?”
You might be interested to know that Proposition 5.1 characterizes
exponential random variables. In other words, a continuous random variable that
has the lack of memory property exhibited in Proposition 5.1 necessarily has an
exponential density function.

5.3 Normal Random Variables

The normal distribution is widely used in statistics and plays an important

theoretical role in probability. This distribution is commonly used to model
phenomena in which a continuous random variable assumes values symmetrically
weighted around some central value which is the mean of the random variable.
Whereas the exponential density functions have only one parameter (referred to as A
in Definition 5.2), there are two parameters in normal density functions. The shape
of normal density functions is that of “bell-shaped” curves. The location of the
curve along the horizontal axis is determined by the mean, and the “steepness” of
the bell is determined by the variance.
Before considering the most general types of normal density functions, we
must first understand the standard normal density function.

Definition 5.3: The standard normal density function is the function

1 _f2

fo) aa Von ee

A random variable X that has this density function is called a standard normal
random variable.

It is not elementary to show that the standard normal density is a density

function at all. To show that

(sed = n.
requires a special trick from multivariable calculus. Finding the mean and the
108 Chapter 5: Continuous Models

variance, however, are straightforward. Proof of the Proposition 5.2, which

follows, is left as Problem 5.19.

Proposition 5.2: If X is standard normal, then E(X) = 0 and var(X) = 1.

More general normal random variables have density functions of the

following type.

Definition 5.4: The normal density function with parameters 1 and © (where
6 >0) is the function
1 t= wpe?
1D Sain
If a random variable has this density function, it is said to be a normal random
variable with parameters UX and o.

Unfortunately, the antiderivatives of the normal density functions cannot be

expressed in terms of known elementary functions. This means that one cannot do
computations involving normal density functions and evaluate the integrals using
the familiar fundamental theorem of calculus, that is, by finding an antiderivative
and evaluating at the endpoints of integration. Various numerical approximation
methods can be used, and this is commonly done to determine values of the
distribution function for a standard normal random variable. Further discussion of
ways to do this (including the use of tables of values) appears later in this chapter.
Fortunately, it is sufficient to have a method for evaluating the standard normal
distribution function. Once the standard normal distribution function is known,
Proposition 5.3 can be used to compute probabilities involving any normal random
variable. The key idea behind the proposition is that every normal random variable
is a linear function of a standard normal one. The proof of this proposition is based
on the change of variables s = (t — )/o, which converts the normal density with
parameters LL and o to the standard normal density. (See Problem 5.20.)
For a normal density function, the parameter 1 turns out to be the mean of the
corresponding random variable and o is the standard deviation. (See Proposition
5.3 Normal Random Variables 109

5.3.) Thus the value of 1 determines where the “bell” is located on the ¢ axis.
This is shown in Figure 5.3. It is important to realize that the general shape of the
bell is determined by the parameter o. A small value of o produces a tall skinny
bell, whereas a large value of o causes the values to be much more spread out with
a correspondingly lower peak.

TT

Figure 5.3 The graph of a typical normal density function.

The tremendous help that Proposition 5.3 provides is that it enables us to

convert problems about arbitrary normal random variables into problems about
standard normal random variables. This means that we can frame all computations
in terms of the standard normal distribution function.

Proposition 5.3:
1. If X is normal with parameters wp and o, then X has mean wu and
variance 0°.
2. If X is normal with parameters wu and o, then the random variable
(X—\)/o is standard normal.
3. IfX is standard normal and uy and o are numbers with o > 0, then the
random variable Y = oX + is normal with parameters Lt and o.

Comment: Remember that for any random variable X,

P(a<X <b) =Fy(b)-Fx@
This says that all probabilities of interest can be computed once the distribution
110 Chapter 5: Continuous Models

function Fy is known. Since all normal random variables can be expressed as a

function of a standard normal one (in light of part 2 of Proposition 5.3), this means
that values of the standard normal distribution function are all that is needed.
Because of the importance of the standard normal distribution function, it is
common to devote a special symbol to serve as a name for it. Often the symbol ®
is used to denote the standard normal distribution function, and we will follow that
convention. (Other symbols sometimes used for the normal distribution function
are Nz and Q.)

Example 5.2. A particular type of lightbulb has a lifetime that is a normal

random variable with expected value 1,000 hours and standard deviation 200
hours. What is the probability that a lightbulb of this type actually lasts longer than
900 hours?

Solution: Let X denote the lifetime of the bulb. Then X is normal with
parameters = 1,000 and o = 200. Furthermore, if we let Y denote (X — )/o,
then Y is standard normal.
We wish to know P(X > 900). However,
P(X > 900) = P( (X — )/o > (900 — L)/o )
P(Y >-.5) =1-P(¥ <—.5)
1 — Fy(-.5) = 1 — .3085 = .6915
Thus the probability that the bulb lasts more than 900 hours is approximately
.6915. (See Problem 5.14.)

Why would we choose a normal distribution to model the lifetime of a

lightbulb? There is at least one logical inconsistency in doing so. It is that all
normal density functions are positive for all real values of t, meaning that it is
theoretically possible for any normal random variable to assume a value in any
interval whatsoever. To assume a normal distribution with mean pt = 1,000 and
standard deviation o = 200 means we are necessarily going to come up with a
positive probability that the lightbulb has a negative lifetime! While this may seem
troubling at first, it is of little practical significance. Remember that most mathe-
matical models are only approximate. If you actually compute P(X < 0), where
X is a normal random variable with mean pt = 1,000 and standard deviations =
200, you will find that the probability is negligibly small. In fact, a normal
distribution may be quite accurate for such a device if indeed the manufacturing
5.3 Normal Random Variables At

process is consistent enough that observed lifetimes do tend to cluster around some
mean value.
Both the normal and the exponential distributions are used to model “time to
failure” for various devices. The normal distribution fits situations such as the
lightbulb example above where some kind of actual “wearing out” or “burning out”
is taking place. Clearly, a bulb that has burned for 1,000 hours is not physically
the same as a new bulb. There is no “lack of memory” property in this situation.
On the other hand, something like a piece of electrical cable may not be undergoing
any physical deterioration during use and may be no more likely to fail after 10
years of use than the day it was installed. Instead, it might fail only because of
some random occurrence having nothing to do with any aging process, and thus its
lifetime might well have the “lack of memory” property exhibited by the exponential
distribution.

Example 5.3. Even such a clearly discrete phenomenon as coin tossing is

related to the normal distribution. In fact, if X is any binomial random variable for
which the parameter n is “large,” then (X — Ly)/oy is “almost” standard normal
if [ly and oy denote the expected value and the standard deviation, respectively,
for X. This fact, Proposition 5.4, shows a very important way the normal
distribution can be used in approximations.
Let’s compute, for example, the approximate probability that in 1,600 tosses
of a fair coin, less than 775 heads occur. Let X denote the number of heads in
1,600 tosses. Then X is binomial with n = 1,600 and p = .5. From Proposition
4.1, we know that the values of Ly and Oy are Hy = 800 and oy = Vnpq =i).
Therefore,

X — py ete kn az 800
Oy 20
is “approximately” standard normal. Thus,
P(X < 775) = P(X — 800 < —25)
= P[(X — 800)/20 < -1.25]
= PCY <=1.25)
= Fy(-1.25) = @O(-1.25) = .1056
Some ways of obtaining this value of @(—1.25) are discussed in the next
section. All require use of some kind of numerical method.
What we have done in this example is to use the normal distribution as an
112 Chapter 5: Continuous Models

approximation to the binomial. If great accuracy is desired in such a calculation,

there is a trick called the continuity correction that can be brought into play. Recall
that our original task was to determine P(X < 775). Since the number of heads in
reality can assume only integer values, this is the same as P(X < 774). However,
once we move to the normal distribution, it makes no difference whether we use <
or <in the inequality. (Remember that for continuous random variables, endpoints
may be disregarded.) So, should we use 775 or 774? The result of using one is
that our approximation is a little too high, and with the other it is a little too low. To
compromise and use 774.5 would give a more accurate result, though the result
already obtained, .1029, is quite good. Problem 5.29 asks you to redo the
calculation using 774.5 rather than 775 and to observe how much this changes the
answer. Problem 5.16 shows that the continuity correction can provide a
substantial improvement when the parameter n appearing in the binomial
distribution is relatively small.

The method used in Example 5.3 is based on Proposition 5.4. This is a

special case of an important theoretical result called the central limit theorem. A
more general form of the theorem is stated in Chapter 6.

Proposition 5.4: If X is binomial with parameters n and p and if n is “large,”

then the random variable

is approximately standard normal. (In other words, the distribution function of Y

is approximately the same as the standard normal distribution function if n is
large.)

Bear in mind that the larger n is, the better the approximation. This is
fortunate, because with large values of n, the binomial distribution is cumbersome
to use. So the approximation provided by the normal distribution is most accurate
just where it is most needed, for large values of n.
The relation between the binomial and the normal distribution is of more than
just theoretical interest. It provides a useful tool for estimating the reliability of
“sampling” strategies. The following example illustrates this.
5.3 Normal Random Variables 1413

Example 5.4. A pharmaceutical firm wishes to test the effectiveness of a

new drug. In each patient, the drug either will or will not achieve a desired result.
Different patients being administered the drug may be viewed as independent trials
of an experiment with success probability p, where p is the probability that the
drug will have the desired effect.
The problem, of course, is that the value of p is not known. The company
wishes to conduct a test to estimate p. The idea of the test is simple. Since
probabilities should reflect a theoretical “relative frequency” of occurrence, a large
number of patients can be tested with the drug and the fraction on whom the drug
works should then approximate p. A strategy, though, is needed in order to decide
how many patients the drug should be tested on.
What the company would like in the end is a conclusion similar to this: “Our
drug has the desired effect on 84% of all patients, plus or minus 3%. Furthermore,
the probability that this claim is correct is at least 95%.”
What kind of mathematical basis could there be for such a claim?

Solution: Suppose the drug is tested on n people, assumed to be

independent trials, and that the probability that the drug works on any given person
is p (unknown). The number of people on whom the drug has the desired effect
then is a binomial random variable with parameters n and p. We will denote this
random variable by X.
The fraction of people on whom the drug works is then X/n, and this is
what will be used to estimate p. Specifically, let’s concentrate on the question as
to how likely it is that this approximation is not within .03 accuracy. In other
words, let’s look at

P(I= —p| > 03)

n
Now
ee
P(|= pl > 03) I (a
n
> 03)

q ofLX — np fen |
——— —_
Vnpq Vpq
\X — npl —
Se
Ss = > Io n } (5.1)

The reason this latter inequality is true depends on a little elementary calculus.
Check that the function f(p) = p(1 — p) takes on its maximum value at p = 1/2,
114 Chapter 5: Continuous Models

where f(p) = 1/4. This means that pq, where g = 1 — p, is always less than or
equal to 1/4, and therefore Vpq < 1/2 for all possible values of p and g = 1 —p.
Therefore,

Because of this inequality, the event whose probability forms the left side of
Inequality 5.1 above is a subset of the event whose probability is on the right. So
the inequality is a special case of the fact that P(A) < P(B) if A CB.
Here’s where the normal distribution comes in. Recall that the random
variable
X —np
Vnpq
is approximately standard normal. Problem 5.30 asks you to show that if Z is a
standard normal random variable, then P(IZ| > d) = 2 — 2®(d) for any positive
number d. Therefore, the rightmost term in Inequality 5.1 is approximately equal
to 2—2@(.06V7).
The company would like to be 95% confident of their claim, that is, the goal
is that this term be no greater than .05. So let’s just set 2 — 2(.06 Vn ) equal to
.05 and solve for n. Then, since 2 — 2(.06 Vn) = .05, we know ®(.06 Vn) =
.975. In the next section we will discuss ways of evaluating the function ® and its
inverse. If @(.06 Vn )= .975, then .06 Vn ~ 1.96 (use the table in Appendix B if
you like), and solving for 2 gives n = 1067.
Conclusion: If the firm wishes to claim with 95% confidence that they have
estimated the probability p to within .03 by testing n cases and then estimating p
by the fraction of the cases on which the treatment works, it will be necessary for
them to test at least 1067 patients with the drug.

Clearly there are numerous situations to which the techniques of the above
example would be applicable. The most familiar is that of the public opinion
research firms that make statements such as, “If the election were held today,
candidate A would get 42% of the vote, plus or minus 3%.” The 95% confidence
level is so often used that it has become something of a standard and is often left
unsaid in statements such as this.
There is another useful tool for estimating the probability that a random
variable differs from its mean by more than a given amount. It is called
5.3 Normal Random Variables HES

Chebyshev’s inequality and appears as Proposition 5.5. The estimate that

Chebyshev’s inequality provides is necessarily going to be rather crude since the
hypotheses are so weak. Nothing is assumed about the nature of the distribution of
X other than that X has finite mean and variance. Nevertheless it is a powerful
theoretical tool that is often useful for the qualitative rather than the quantitative
information that it provides.

Proposition 5.5 (Chebyshev’s Inequality): Suppose that X is a random

variable having finite mean 1 and variance o”. Then for any positive number €
2
Co)
Px Sie &) Ss ,
€
When expressed in terms of the complementary event this says
?
PUK
= lee) et e2

Comment on Proposition 5.5: Let’s examine what the proof looks like in the
continuous case. (The discrete case may be proved similarly by replacing the
integrals by sums.)
First notice that it is sufficient to prove the proposition for the special case in
which E(X) =0. The reason is that the random variables X and X — u have the
same variance. Therefore, if the proposition is true for X — LL, it will be true for X
also.
So now suppose that X is a continuous random variable with mean u = 0 and
with density function f; then
P(xXl2e) = PC 2 €)+ POX s=-€)
a =
ith + iS(t) dt
ee, : oo (2
lA | ait dt + ) ait at

ea ch iri.
oe 9 d
said f(t) dt + —Bi t2 f(t) d at
116 Chapter 5: Continuous Models

Example 5.5. A lot of 2,000 items has been produced on an assembly line
on which 2% of the items produced are defective. Treating the assembly line as an
independent trials process with each item produced having probability .02 of being
defective, use Chebyshev’s inequality to give a bound on the probability that in the
batch of 2,000 items the number of defects is between 30 and 50.

Solution: Let X denote the number of defects in the 2,000 items. The
assumption is that X is binomial with parameters n = 2,000 and p = .02. This
enables us to compute w and o: = np = 40 and o” = npq = 39.2. If we now
take € = 10 in Chebyshev’s inequality, we have
392
P(30<X <50) = P(X-pul<10) = 1- 00 = .608
This should be viewed not so much as an estimate of the actual probability as a
bound on the probability. The probability that X assumes a value within 10 of its
mean is at least as great as .608. Since Chebyshev’s inequality is based on
nothing more than the mean and the variance of the random variable involved, it
can’t possibly give accurate estimates.
By way of comparison, we can relate this to the estimate provided by the
central limit theorem (Proposition 5.4):

P(X — ul < 10) pease ati

Vnpq Vnpq

10
v 20( )-1
Vnpq.
=2~x 9441-1
= 888

As an illustration of the theoretical usefulness of Chebyshev’s inequality, let’s

look further at what it has to say about an independent trials process. Denote by
X,, the number of successes to occur during the first n trials of an independent
trials process with success probability p. Chebyshev’s inequality says then that
5.4 Evaluating the Standard Normal Distribution Function mid7

P(X, —npl<e)>1- os
The value of this information is that it is true for whatever choice of € > 0 that we
wish to make. In particular, let’s replace € by nS, where 6 is thought of as
“small.” Then

{ [=I <5) a
i wd nd
Now notice that for any value of 8, the right side converges to 1 as n > o»,
What does this mean? It means that for any choice of 6, no matter how small, the
probability that the fraction of trials resulting in success differs from the theoretical
probability p by less than 6 tends to 1 as the number of trials increases without
bound. In other words, it is guaranteed that the observed relative frequency of
successes will converge to the theoretical relative frequency (as measured by p) as
the number of trials tends to ce. This is a special case of an important theoretical
result called the law of large numbers.

5.4 Evaluating the Standard Normal Distribution Function

Since the antiderivatives of the standard normal density function cannot be
expressed in terms of elementary functions, it is necessary to use means other than
the fundamental theorem of calculus to evaluate the integrals that define values of
the standard normal distribution function.
The relation between the standard normal distribution function ® and the
standard normal density function is given by

a= | fod
where f is the standard normal density function

Thus, if values for the standard normal distribution function are needed, what is
necessary is a good way to approximate this improper integral. In fact, the
symmetry of the integrand f(¢) and the fact that f is a density function mean that
oo 0
{ f()dt = 1 and } f(t) dt =5

So if we want to compute D(s) for some s > 0, then

118 Chapter 5: Continuous Models

Os) = 5 - iyf() dt
On the other hand, if s > 0, then the symmetry of f(t) about the y-axis implies
that
=S, oo Ss

M(-s) = i f@) dt = i f@@ dt = | fi) dt = 1-QD(s)

—oo Ss —co

AY

Thus all we really need is a way of evaluating | @) distorts 0.

0

Problems about normal random variables, of course, don’t have to be phrased

in terms of the normal distribution function at all. Rather, they can be expressed in
terms of the normal density function. For example, if X is standard normal and
we need to know P(.5 < X < 1), we might just as well think of integrating (by
some numerical method) the standard normal density from .5 to 1 rather than
thinking of P(.5 <X <1) as ®(1) — B(.5).
In summary, all that is needed to get approximate values for normal random
variables is some means of approximating definite integrals of the standard normal
density function over finite intervals. It is not necessary to approximate improper
integrals.

Simpson’s Rule

Most calculus students learn Simpson’s rule as a numerical method for

approximating definite integrals. The idea behind Simpson’s rule is very simple:
Given a function defined on an interval, plot the value of the function at the two
endpoints and at the midpoint and then draw the unique parabola through the three
points. The area under the parabola is used to approximate the area under the
curve. To obtain accuracy, the original interval may be split into smaller
subintervals and this procedure used on the subintervals.
Crude as it sounds, this procedure works beautifully for smooth functions
such as the standard normal density. For example, if we wish to approximate the
probability P(O0 < X < 1) using this method, where X is standard normal, and if
we use the two subintervals, [0, 1/2] and [1/2, 1], then the approximation obtained
is 341355. The true value, correct to 6 digits, is .341344. The arithmetic of
Simpson’s rule in this particular scenario boils down to computing
Problems 119

1
TD * { f(O) + 4 f(.25) + 2 f(.5) + 4 f(.75) + f(A) J
where f is the standard normal density function. Such a calculation is easy with a
scientific calculator and even easier with a programmable one. Students who know
Simpson’s rule are encouraged to try it out as a method of getting numerical
answers in exercises involving normal random variables.

Other Available Numerical Methods

Since the mid 1970s many scientific calculators have included the normal
distribution function as a built-in function. Some program must then reside within
the calculator for computing the values. The one described in Problem 5.14 has
been used in Texas Instruments™ calculators for a number of years. It is easily
programmed on any programmable calculator or pocket computer.

Tables

Before calculators were available, it was common practice to consult tables

that give values of the normal distribution function. This practice is rapidly
becoming as outdated as the practice of looking up square roots in a table.
Nevertheless, a table is included as Appendix B at the end of this book.

Problems

5.1. Show that if X is uniform on [a, b], then

2
at+b (b — a)
EX) = 5) and var(X) = Seep

5.2 A mechanical device is known to have an expected lifetime of 10 years.

Assume that the lifetime for such devices is known to have an exponential
distribution.
(a) What is the probability that a device of this type will last more than 5
years?
(b) What is the probability that such a device will fail during the first 10
120 Chapter 5: Continuous Models

years?
(c) If such a device has already lasted 10 years, what is the probability that
it will fail during the next 10 years?

Bin) Suppose we have a random number generator that generates numbers

between 0 and 1 according to the uniform density on the interval [0, 1]. If
we generate such a random number X and then consider Y = 1 —-X, what
kind of density function does Y have? (Your intuition might lead you to
guess the answer before doing the calculation.)

5.4 Suppose you have just bought a piece of equipment that has been advertised
to have an expected lifetime of 2 years. Is it likely to actually last that long?
(In other words, how probable is it that the device will actually last that
long?)
(a) Assume that the lifetime is normally distributed.
(b) Assume that the lifetime is exponentially distributed.

Oye) Assume that X has an exponential density with parameter A = 1.

(a) Find Fy.
(b). If Y = 2X + 1, find Fy.
(c) Find the density for Y.
(d) Using Definition 3.7, compute E(Y).
(e) Compute E(Y) again, this time using Proposition 3.7.

5.6 Suppose X is a uniform random variable on [0, 1]. Let Y =X’. Compute
the distribution function for Y and the density function for Y. Notice that
fy is an unbounded function, but that part 2 of Definition 3.3 is satisfied by
fy if the integral is interpreted as an improper integral. Compute E(Y)
using Definition 3.7 and again using Proposition 3.7.

SVP Assume X is uniform on [0, 2], but now suppose Y = LX — 1|. Determine
the distribution function, the density function, and the expected value of Y.
[Hint: The biggest difference between this and Problem 5.6 is that you will
have to be more careful and clever in trying to write Fy in terms of Fy.]
One interpretation of this problem is to think of it as a model for choosing a
random number between 0) and 2 and then to consider how close the number
chosen is to 1.
Problems 421

5.8 A wire is 2 ft long. A random point on the wire is picked, and the wire is cut
at that point.
(a) What is the expected value for the length of the shortest piece?
(b) If Y is the random variable that denotes the length of the shortest piece
of wire obtained when the original is cut, find the distribution function
and the density function for Y.
[Hint: This problem is closely related to Problem 5.7. Think of the wire as
lying on the interval [0, 2] and X as being the point where the cut is made.
Assume X is uniform on [0, 2]. Now write Y as a function of X.]

53 A device has a lifetime which is known to be an exponential random variable

X with E(X) = 10 years. Find the value of to for which the probability is
exactly 1/2 that the device lasts at least fo years; that is, P(X > fg) = 1/2.

5.10 Suppose X is an exponential random variable with parameter A = 1.

Determine the distribution function for the random variable Y = X?.

Sell If X is normal with parameters uw = 5 and o = 2, find P(2 < X < 4). You
can use Simpson’s rule, the numerical method of Problem 5.14, or the table
in Appendix B.

aes Show that if X is an exponential random variable with parameter 1, then the
variance of X is 1/47.

5.13 Use the normal approximation to the binomial (Proposition 5.4) to

approximate the probability of between 290 and 305 successes in 1,000 trials
of an independent trials process with p =.3.

5.14 It used to be commonplace to consult long and tedious tables when values of
the standard normal distribution function were needed. Many scientific
calculators now have this function built in. In addition, there are many
common and simple numerical methods that may be used to evaluate it. The
following peculiar but useful formula is given in the Handbook of
Mathematical Functions of the National Bureau of Standards.
If ® denotes the standard normal distribution function and
ogre Py
fO aa
denotes the standard normal density function, then values of ® may be
122 Chapter 5: Continuous Models

approximated via the numerical approximation

O(1) = 1 —f(t) (4,8 + ay8? + a;8? + a48* + a5B°)
where 8 = 1/(1 + .2316419 t) and where a, = .31938153, a, =
~.356563782, az = 1.781477937, aq = —1.821255978, and as =
1.330274429. This approximation is easy to program (see Appendix B) and
quite accurate when t > 0. Try it out with a scientific calculator in Example
5.2, where ®(—.5) is needed. Compute ®(.5) using the method described
here. Then ®(—.5) = 1 — ®(.5).

5.15 Assume that the number of defects in a batch of 10,000 manufactured items
is anormal random variable with mean ut = 300 and standard deviation o =
100. (Clearly the actual number is an integer-valued random variable. By
now you should be getting used to the idea of using a continuous model for a
‘situation that technically is discrete.) Find the probability that the actual
number of defects found in a batch will be between 225 and 275.

5.16 If 100 items are tested from a batch in which 20% are bad, what is the
probability that the number of bad ones found will be between 15 and 25
inclusive, that is, P(15 < X < 25), where X = number of defectives
found?
(a) If each one tested is considered an independent trial with probability p
= .2 that the item is bad, then the number of bad items found is binomial
with n = 100 and p = .2. Compute P(15 < X < 25) assuming that
X has this probability distribution.
(b) Now approximate this probability using the fact that (X — np) npq is
approximately standard normal, and compare your answer to the exact
answer in part a. (If you are really interested in great accuracy in your
approximation here, you might want to consider further the fact that X
is actually integer-valued. This means that, in fact,
P(USSX<25) =P(14 <X < 26)
If one “averages” these two intervals and uses
P(14.5 < X < 25,5)
in calculating the normal approximation, then the approximation of the
discrete binomial with the continuous normal will be very good.)

5.17 Suppose that a flagpole is erected and that it is assumed that the “lifetime” of
the flagpole is a random variable having an exponential density function and
Problems 123

having an expected value of 50 years. Find

(a) the probability that the flagpole lasts more than 25 years,
(b) the probability that it lasts more than 50 years, and
(c) the probability that it lasts more than 100 years.

a18 We will approach Problem 5.17 from another viewpoint. Instead of

measuring time “continuously,” we will measure it “discretely.” Suppose
that each year there is a probability .02 that the flagpole will fall during that
year, and suppose that this is viewed as an “independent trials process.” Let
X denote the number of the year in which the flagpole finally falls, so X
could assume values 1, 2, 3, - - - (X is a discrete random variable.) If we
consider the years to be independent trials, then X can be viewed as a
geometric random variable with p = .02. Under these assumptions,
compute
(a) the probability that the flagpole lasts at least 25 years,
(b) the probability that it lasts at least 50 years, and
(c) the probability that it lasts at least 100 years.

alo Prove Proposition 5.2. [Hint: The mean is easy. To compute the variance,
use integration by parts with u = and dy =te~/2. Then to evaluate
Jv du
in the integration-by-parts formula, where
fudv = uw -Jvdu
use the fact that the standard normal density is a density function.]

5.20 Prove Proposition 5.3. [Hint: In part 1 do a change of variables to convert

to the standard normal density. In part 2 show that the distribution function
of (X — 1)/o is the standard normal distribution function. In part 3 a similar
change of variables is useful.]

21 Suppose that the number of cases of a certain disease per 100,000 population
is a normally distributed random variable with mean 300 and standard
deviation 100. What is the probability that a particular community of
100,000 people would have more than 500 cases of the disease? (Notice that
this is another instance where we are using a continuous model for what is
actually a discrete phenomenon.)
124 Chapter 5: Continuous Models

Deed Suppose a random number is generated (uniform on [0, 1]) and then
truncated after the first digit so the result is one of the numbers 0, .1, Phe
, 9. Compute the expected value for this truncated result, and understand
your computation in the context of Proposition 3.7. One interesting feature
of this situation is that we are looking at an instance of the relation Y =
g(X), where X is continuous and Y is discrete.

See: Suppose that a > 0 and that X is a random variable that is uniformly
distributed on the interval [0, 1]. Find the distribution function and the
density function for the random variable

y=-— In (1 —X)
(This problem has very useful applications, because many devices such
as calculators and computer routines are equipped with random number
generators that produce numbers between 0 and 1 according to a uniform
density on [0, 1]. If one wants a “random” value for an exponentially
distributed random variable and has such a “uniform on [0, 1]” random
number generator available, this exercise shows how to get one.)

5.24 A random number between 0 and 1 is produced using the uniform density on
[0, 1]. What is the probability that the first digit in the decimal representation
of the square root of the number is a 7?

ee) In the circuit pictured below, the resistance of the resistor labeled R
fluctuates between 1 and 2 ohms. Assume a uniform distribution for the
resistance between 1 and 2 ohms. If the battery output is a constant 12 volts,
what is the probability distribution of the voltage drop across this resistor?
Find the distribution function and the density function for this voltage drop.
What is the expected value for this voltage drop across R?
R

12 volts

1 ohm
Problems 125

5.26 Resistor R, in the following figure has a constant resistance of 1 ohm.

Resistor Ry has resistance which varies uniformly over the range of 1 to 2
ohms. Find the density function for the resistance in the circuit, that is, the
resistance of the two resistors wired in parallel.

R
2 12 volts

eG (a) A random variable X gives the voltage output from an acoustical

transducer. The voltage varies from —3.5 to +3.5 volts. Assume that
the probability distribution for X is the uniform distribution on the
interval [-3.5, 3.5]. Sketch a graph of the cumulative distribution
function for X.
(b) Suppose now that for some reason it is desirable to round off the
voltage to the nearest integer. In this case, the voltage will be between
—3 and +3 volts. This integer-valued voltage random variable can be
written as Y = round(X), where round is the round-off function; that
is, round(x) = the value of x rounded off to the nearest integer for any
real number x. Determine the probability mass function for this
discrete random variable Y, and then observe that this is the same
probability distribution as the one encountered in Problem 4.12.

5.28 A manufacturer wishes to estimate the proportion of defects in a batch of

items. If n items are to be tested and X is the number of defects found,
then X/n will be used as the estimate of p, the fraction of the lot that
consists of defective items. Using techniques parallel to those of Example
5.4, determine how large n must be, that is, how many must be tested in
order to be 90% certain that the observed value of X/n is actually within .05
of the true value of p.

a9: Redo the calculation in Example 5.3 using the continuity correction (as
described in that example) to see how much it changes the answer.
126 Chapter 5: Continuous Models

5.30 Show that if Z is a standard normal random variable, then P(IZ| > d) =
2=—2(d) for any number d > 0.

Supplemental Exercises on the Normal Distribution

See, If X is a standard normal random variable, find P(X < 1.5).

Soe If X is standard normal, find P(X > .22).

O92 If X is normal with mean yp = 15 and standard deviation o = 3, find the

probability P(X < 17).

5.34 If X is a normal random variable, find the probability that the value assumed
by X lies more than 2 standard deviations away from the mean. [Hint: Part
of the problem here is to convince yourself that the answer to this question is
the same regardless of what the mean and variance of X happen to be.]

5:35 If X is standard normal, find (approximately) the value of t for which it is

true that P(X >t) = .95. [Hint: If you look at this problem carefully, you
should see that this question is equivalent to the problem of evaluating the
inverse of the standard normal distribution function at .05. You can either
use a table, or if you have a routine for calculating values of the standard
normal distribution, you can by trial and error zero in on an approximate
answer to this question. ]

5.36 If X is normal with mean u = 3 and standard deviation o = 1, find a 95%

confidence interval for X centered at 3. This is a fancy way of saying the
following: Find the number 6 > 0 having the property that
PG-5 <X <3+8) =.95
(See the hint in Problem 5.35. This question can also be phrased in
terms of finding values of the inverse of the standard normal distribution
function.)
 Chapter 6: Joint Distributions

Many common mathematical models involve more than one quantity. For
example, a model for a gas is concerned with the interaction of temperature and
pressure. When dealing with a probabilistic model, this means that we must be
concerned with the way that two or more random variables interact. Knowledge of
the individual behavior of the two will not be adequate because we may need to
know how certain values assumed by one affect the probability of the other
assuming given values. In the discrete case the interaction is studied via the joint
probability mass function, and in the continuous case via the joint density function.
There is additionally a natural way to extend the concept of the distribution function
to this “joint” setting, but the joint distribution function is less useful for
computation and will not be emphasized in this book.
When studying the relation between two random variables, we often will need
to consider the intersection of two events of the form (X =a) and (Y=b). A
useful way to denote the intersection of these two events is to write the intersection
as (X =a, Y=b). The probability of this event is written P(X =a, Y =b).
Expressions such as P(a <X <b,c <Y <d) have a similar meaning, that is,
the event whose probability is being represented is the intersection of the two
events separated by commas.

6.1 Joint Probability Mass Functions

Definition 6.1: If X and Y are discrete random variables, then their joint
probability mass function py y is the function of two variables defined by

Py y% Y) = PX=x,Y=y)

for each pair of real numbers x and y.

er
128 Chapter 6: Joint Distributions

Example 6.1. Four items are labeled with the numbers 0, 1, 2, and 3, and
two of the items are randomly selected one at a time without replacement. We will
denote the number on the first item selected by X, and the number on the second
by Y.
It is sometimes useful, in examples such as this where there are a small
number of possible outcomes, to exhibit all values of the joint mass function in a
table such as the one shown in Figure 6.1, which presents everything there is to say
about the joint mass function forX and Y.

2 1/12 1/12 0 1/12 1/4

Sie. SUA. Ave 0 1/4

Yo atha? Sa? Sata ah Tia
Figure 6.1 Joint probability mass function for the
random variables X and Y of Example 6.1.

The 16 numbers in the table that are not in boldface give all values of py y.
For example the first 1/12 in the top row gives the probability P(X = 1, Y = 0).
The boldface numbers at the bottom and at the right edge are the values for the mass
function of X (bottom row) and Y (rightmost column). Notice that these
probabilities may be obtained simply by adding the values of the joint mass function
in the corresponding row or column. The reason is elementary. For example, the
event (X = 1) is the disjoint union
A= 1YeQURCeLYaH1l)vuG
a1 Y=2) 0s, res)
and therefore,

PX =1)=PXK=1,Y=0)+PX=1,¥=1)+P%=1, Y=2)+P4%=1, Y =3)

6.1 Joint Probability Mass Functions 129

Proposition 6.1 summarizes the relation between py, py, and Pxy- The
proof of the proposition is based on the same ideas that are illustrated in the
example above. Conceptually, it is important to understand that the joint probability
mass function carries all the information necessary to do computations related to the
interaction of X and Y. In particular, Proposition 6.1 summarizes how the mass
functions of X and Y individually may be determined from the joint mass
function. Often the probability mass functions of X and Y are referred to as the
marginal probability mass functions of X and Y to distinguish them from the
Joint mass function for the pair X, Y.

Proposition 6.1: The relationship between the marginal mass functions py and
Py and the joint mass function py y can be described in general as follows:

1. For any number x, py(x) = BY Px y(%, y,) where the sum extends

over all values y, assumed by Y.

2. For any number y, py(y) = yy Px,y(*,, y) where the sum extends

over all values x, assumed by X.

We have encountered the idea of independence of events in Chapter 1. It is

also useful to have an understanding of what independence means in the context of
random variables.

Definition 6.2: Two discrete random variables X and Y are independent if it is

true that
PQX=4k,¥ =y) = PC=x) PY =)
for all choices of x and y.

The basic idea is that two random variables are independent if events
described in terms of one random variable are independent from events described in
terms of the other. Literally, the definition says that X and Y will be considered
independent random variables provided that events of the form (X = x) and
(Y =y) are independent events for all choices of numbers x and y.
130 Chapter 6: Joint Distributions

Example 6.2. One 4-ohm resistor and two 8-ohm resistors are in a box. A
resistor is randomly drawn from the box and inspected; then it is replaced in the box
and a second is drawn. We will denote by X the resistance of the first one drawn
from the box and by Y the resistance of the second. Since this is sampling with
replacement, the result of one test has no influence on the result of the other test.
This is indicated in the following table of values of the joint mass function for X
and Y. In this case X and Y are independent random variables.
When X and Y are independent, the values of the joint mass function are
determined by the values of the mass functions of X and Y. For this example,
this is clearly visible in Figure 6.2. The circled entries illustrate that 2/9 = 1/3 x 2/3.
Similarly, each of the four values of the joint mass function shown in Figure 6.2 is
the product of the appropriate value of the mass function of X and the mass
function of Y. In general (when independence is lacking), it is not possible to
construct the joint probability mass function from a knowledge of the mass
functions of X and Y individually. See Problem 6.31(a) for an interesting
illustration of this.

Figure 6.2 Joint mass function for the independent random

variables of Example 6.2. The circled entries illustrate that
P(X=8, Y=4) = P(X=8) P('Y=4).

6.2 Joint Density Functions

The joint density for a pair of continuous random variables will also
necessarily be a function of two variables. It plays the same role that the joint mass
function plays for a pair of discrete random variables. That the probability of the
entire sample space be 1 requires the double integral over the xy-plane be 1, just as
the integral in Definition 3.3 in Chapter 3 is required to be 1.
6.2 Joint Density Functions 1341

Definition 6.3: A function f of two variables is said to be a (two-dimensional)

density function provided that
1. f(x,y) 20 for allx andy, and

De iei (9) axidy=1

In order to motivate the relation between a pair of continuous random

variables and their joint density function, it is worth observing that in the discrete
case the probability P(a< X <b,c < Y <d)is, in fact, equal to the sum
X Px y(%}s Ye)
where the sum here extends over all pairs (Xj ¥x) where x; and y, satisfy the
conditions a < xjs b andc <yz <d. In fact, if B is any region in the plane
and if we used the notation (X, Y)eB as a short way to denote the event
{weS : (X(w), Y(w))eB} in the underlying sample space S, then we can
write the probability of this event as

PUK NeB) = YD pyyopyo (6.1)

GpyPEB

Problem 6.10 is a simple exercise which will check your understanding of

what Equation 6.1 says.
Perhaps.now you can guess the definition of the joint density function for a
pair of continuous random variables by simply thinking of summation in the
discrete case being replaced by integration in the continuous case.

Definition 6.4: Given a pair of random variables X and Y, to say that a

function f (of two variables) is the joint density function forX and Y means that
1. the function f is a density function (see Definition 6.3), and

22 Pie YeaBbius dil f(x,y) dx dy whenever B is a region in

the xy-plane.
132 Chapter 6: Joint Distributions

Among the easiest to understand examples of joint density functions are the
joint uniform densities.

Definition 6.5: Given a region T in the xy-plane, the uniform joint density
function on the region T is the function f defined by
i
on if (x,y) € T
fy) =
0 if &y) ¢T
where A denotes the area of the region T.

The reason that the constant value that fassumes on T must be the reciprocal
of the area of T is that the integral of the constant function 1 over the region T is
precisely the area of T. So this requirement makes the joint uniform density satisfy
condition 2 of Definition 6.3.

Example 6.3. Suppose X and Y have as joint density function the uniform
density’ on the square T = {@)y) 70 s2 =1,0 sys 1}. What is:the
probability that X takes on a value larger than Y? In other words, what is the
probability of the event (X > Y)?

Solution: The main step that is necessary to grasp this example is to

understand what the event (X > Y) means. It means the set of all elements w in
whatever sample space X and Y are defined on which have the property that the
pair of number X(w) and Y(w) satisfy the condition X(w) > Y(w). In order
to relate this to Definition 6.4, it is necessary to rephrase this as follows: We are
considering the event that consists of all elements of the sample space for which
(X(w), Y(w)), when viewed as a point in the xy plane, lies to the right of the
line y=x. So we are interested in P{(X, Y)e B}, where B is the half-plane
that is shaded in the following picture.
From Definition 6.4, what is needed in order to compute P(X > Y) is to
integrate the joint density over the region B. In this example, however, since the
joint density is 0 off the square T, the integral over the region B can be reduced to
the integral over B.A T.
6.2 Joint Density Functions 133

This train of thought is summarized in the following steps:

P{(X, Ye B} = ilef(x,y) dx dy

: (pki f(x,y) dx dy
I 1 dx dy
Baw

a
2
(Since we are integrating the constant function 1 over the triangle B 1 T, the value
of the integral is just the area of the triangle.)

Recall how, in the discrete case, the mass functions for X and Y can be
constructed by adding the appropriate values of the joint mass function. A similar
method is available in the continuous case for obtaining the density functions for X
and Y from the joint density. This method is described in Proposition 6.2, which
is similar to Proposition 6.1 except that (as usual) summation is replaced by
integration. Often the density functions of X and Y are referred to as the
marginal densities to distinguish them from the joint density for X and Y.
134 Chapter 6: Joint Distributions

Proposition 6.2: If X and Y have joint density functionf, then the densities
forX and Y may be obtained from f as follows:

f=
| fonray and fo=
| fon» ax

Proof: Given any numbers a and b with a <b, P(a <X <b) can be
computed from the joint density function using Definition 6.4 by taking B to be the
region
{(%, y):a<x<b,--o<y<oo}

that is, B is an infinite vertical strip in which no restriction is placed on the second
coordinate. Then
Pia<X <b) =P{(X, YeB} = J fo dx dy

~ f He f(x, y) dy dx = feo dx
where g(x) is the function

sa)=
| fee» ay
But knowing that
b
Pia<X <b) =| g(x)ax

for all choices of a and b with a < b is exactly what we need to know in order to
conclude that g is the density function for X. (See Definition 3.4.)
Clearly the proof of the second statement involving the density of Y can be
given in a similar fashion.

Example 6.4. In a certain device, X represents the voltage drop across a

component and Y represents the current in another part of the device. The two are
related probabilistically, however. Let’s suppose that X and Y have a joint
density function f(x, y) =c in the part of the first quadrant bounded by the
curve y=4~—x* and the x axis, with f(x, y) = 0 outside this region.
6.2 Joint Density Functions 135

(a) Find the value c must have in order to make this a joint density.
(b) Find the marginal densities fy and fy.
(c) What is the probability that 3X is greater than Y?

co co 2 p4— x?
Solution: (a) | { SX, yak dy = [{ ec dyax=1
eee 0~0

Carry out this computation and solve for c, obtaining c = 3/16.

(b) If 0 <x <2 then

Ree : : 2
co
B(4e
( re )
fx) = [ fx, y) dy= I =3 dy =

Furthermore, if x is not between 0 and 2, then fy) = 0. If0<y <4, then

?
BO Nee eyV4 igiS 3V4—y
eee
Moreover, fy(y) = 0 if y is not between 0 and 4.

(c) P(Y < 3X) is computed by integrating the joint density function over
the shaded region in the picture that follows.

YA

1|

If the double integral is evaluated using iterated integration in the order shown, the
136 Chapter 6: Joint Distributions

endpoints of integration for the integration with respect to y are 0 and 3. This
reason is that the y coordinate of the point where the line y = 3x and the parabola
y = 4-—x? intersect is 3.
3 pV4—y
py <3x) = || c dx dy
0 *y/3 :
3
-bd |
A pareSe yey {ee
is

What should be meant by independence of continuous random variables? The

most natural guess might be to require something like

P(G:2Xh-<0
CY 30), = P@<X
<b) PE <a dD)

for all values of a,b,c, and d. This approach would be satisfactory (see Problem
6.11), but it is a bit easier to take the following definition, which is essentially
equivalent and is easier to use.

Definition 6.6: Two continuous random variables X and Y are independent

provided that the function f defined by
FO, y) = fx@) fro)
is a joint density function for X and Y.

Example 6.5. Let’s reconsider X and Y in Example 6.3. With Proposition

6.2 it is easy to check that both X and Y have as their density functions the
uniform density on the interval [0, 1]. Furthermore, it is quite apparent now that
the joint density f does satisfy f(x, y) = fy(x) fy(), for both sides are equal
to 1 when0 <x <1 and0<y< 1, and both sides are equal to 0 otherwise.
This example is a model for the following experiment: Suppose you have a
random number generator that will generate two random numbers between 0 and 1
in such a way that the numbers produced are independent of each other and so that
the probability distribution for each is the uniform distribution on the interval [0, 1].
A variety of interesting questions can be asked about the two numbers obtained as
output from the random number generator. Problems 6.8 and 6.12 are two such
questions, and a further application of the same model is found in Example 6.9.
6.2 Joint Density Functions 137

Example 6.6. Two devices manufactured by different companies are known

to have exponentially distributed lifetimes. One has an expected lifetime of 5 years
and the other 8 years. What is the probability that the one with the longer expected
lifetime will actually last more than twice as long as the other? (Assume that the
two devices function independently.)

Solution: Let X be the lifetime of the device with the expected lifetime of 5
years and Y the lifetime of the device with expected lifetime of 8 years. This
simply says E(X) = 5 and E(Y) = 8. Furthermore, since both have exponential
distributions, we know from Problem 3.15 that X is exponential with parameter A
= .2 and Y is exponential with parameter 4 = .125.
Since X and Y are to be assumed independent, we know that the function
I(x, y) =fx(x) fy) will be a joint density function for the pair. The question
posed is what is the probability that Y > 2X. The picture guides the calculation.

1 2

The event (Y > 2X) can be viewed as {(X, Y)eB} where B is the region
consisting of the half-plane above the line y = 2x in the picture. So the
probability of this event can be computed by integrating the joint density over this
region. However, the joint density function itself is equal to 0 except in the first
quadrant. Therefore the integral to be computed actually can be reduced to an
integral over the wedge shaped slice of the first quadrant shown in the picture. The
computation then goes as follows:
P{(X, Ye B} - || f(x,y) dx dy
wedge
138 Chapter 6: Joint Distributions

=| (oP (1250) cay. ax

02%

Problem 6.16 asks you to complete this calculation.

There is not a great conceptual hurdle in going from independence of two

random variables to independence of more than two. It is simply a matter of
passing from two dimensions to n dimensions, where n is the number of random
variables being considered. For example with continuous random variables one
first has to extend Definition 6.3 so as to define an n-dimensional density function
with the analogous properties. (The integral in that case becomes an integral over
n-dimensional space.) Definition 6.4 could then be extended to define the joint
density of random variables X,,---,X,, and the integral in that definition
would pertain to regions B in n-dimensional space. Finally, an extended form of
Definition 6.6 would say that X, ..., X, are independent provided that their joint
density function is given by

fli 5) = Vee)
In this book we will not do computations involving such n-dimensional
density functions. In the next section, however, we will be examining some
situations where more than two independent random variables interact. While the
technical definitions involve concepts such as those illustrated in the above
paragraph, the most important idea you need to grasp is the intuitive meaning of
independence of more than two random variables. The idea is that knowledge
about values assumed by certain ones is independent (in the sense of independent
events) from knowledge about others. For example, if X,, X2, X3, and X4 are
independent, this means, for instance, that the values assumed by X, and X3
would have no effect on the values assumed by X, and X4.
The reason this intuitive understanding is important is that in constructing a
mathematical model, one decision that often must be made is the decision as to
whether certain random variables can be considered to be independent.
Historically, some of the most serious errors in constructing complex probability
models (for example, safety analyses of complicated systems such as aircraft or
nuclear reactors) have come about because certain random variables were assumed
to be independent when, in fact, they were not. Section 6.4 will introduce various
scenarios where the concept of independence is important and useful.
6.3 Functions of Two Random Variables 139

6.3 Functions of Two Random Variables

Just as we found it helpful in Chapter 3 to consider composite random

variables of the form Y = g(X), it is now useful to consider composite random
variables of the form Z = g(X, Y), where g is a function of two real variables.
Such random variables have occurred earlier in examples. For instance, if two dice
are rolled and if X and Y represent the numbers on the respective dice and Z the
sum, then clearly Z =X + Y. This simply says Z = g(X, Y) where g is the
function of two variables g(x, y) =x+y.
If Z = g(X, Y), it is useful to be able to compute the expected value of Z
directly from knowledge of the function g and the joint distribution of X and Y.
Propositions 3.6 and 3,7 have equally powerful versions for functions of two
random variables provided that the two are discrete or, in the continuous case, that
they have a joint density function. While sometimes it is possible to determine the
probability distribution of Z from a knowledge of g, X, and Y, Proposition 6.3
offers an alternative that is usually easier if it is only the expected value that is
needed.

Proposition 6.3:
1. IfX and Y are discrete random variables and if Z = g(X, Y), then
the expected value of Z is given by

E(Z) = Y 8%» Ye) Px YX Yd

XjVe

where the sum extends over all ordered pairs x;, y, where x; is a
value assumed by X and y, is a value assumed by Y.

2. IfX and Y are continuous random variables having joint density

function f and if Z = g(X, Y), then

B@=| [sty fey aay

140 Chapter 6: Joint Distributions

Example 6.7. Two signals will occur at a random time during a one-hour
period. Assume that they are timed independently and that the time of occurrence of
each is uniformly distributed throughout the one-hour interval. What is the
expected length of time between the occurrences of the two signals?

Solution: LetX and Y denote the times when the two signals occur. The
assumption is that both X and Y are uniform random variables on [0, 1] and that
they are independent. Their joint density then is the uniform density on the unit
square, as in Example 6.3. The length of time between occurrence of the two
signals is the random variable Z = IX — Yl = g(X, Y) where g is the function
g(x,y) =lx-yl. So the expected value of Z can be computed by performing
the integration

B@=[ [ &-y fey) aray = fel n—yl dedy

The trick to evaluating the integral on the right is to get rid of the absolute
value signs by splitting the integral into two parts. The line x = y separates the
square into two triangles. Integrate over the unit square by integrating over the two
triangles. This enables the absolute value signs to be removed, for on one of the
triangles x > y and so lx -yl=x-—vy, and on the other triangle x < y and
therefore Ix — yl =y—x. When you carry out these details, you conclude that
E(Z) = 1/3. ‘Gee Problem6. 138.)

Example 6.8 (Redundant safety systems). The picture below shows a pair of
redundant safety switches that are installed to cut off the power to a device if a
dangerous situation should quickly develop.

Power Source
Process that may
need to be
interrupted

Switch 1

Poet edie ee
Switch 2
6.3 Functions of Two Random Variables 141

Different operators control the two switches, and so the two switches are
viewed as functioning independently. There is a brief time period between the time
that the dangerous situation arises and the time the switch is actually opened. We
will assume that for both switches the delay time is uniformly distributed over a 5
second time interval. Clearly, if there were only one safety switch, the expected
value for the length of time that elapses from the time that danger develops until the
switch is actually thrown is 2.5 seconds. How much safety is actually achieved via
the use of a backup operator controlling a second switch? In other words, using the
two-switch system, what is the expected value for the elapsed time before a switch
is opened to shut down the system?

Solution: We can let X denote the waiting time until switch 1 is opened and
Y the waiting time until switch 2 is opened. The assumption is that X and Y are
each uniformly distributed on the time interval [0, 5] and that they are independent.
If Z is the waiting time until the current is actually cut off to the device, then Z =
g(X, Y), where g is the function of two variables
g(x, y) = min{x,y}
that is, g is the function that selects the minimum of the two numbers input.
The relation between Z,X, and Y is that Z = min{X, Y}. For any
number 5s,
PCS SFA SS. 1 2 Se Pees)
hae Ss)
the last expression coming from the independence of X and Y. (See Problem
6.11.) In particular, if 0 < s <5, this probability is
ee ok
3 P(Z>s)= ise Bee
(Check this detail. It depends only upon understanding what the uniform
distribution on the interval [0,5] looks like.)
Therefore,

(5-5)
55
F7(s) = P(ZS$s)=1-P(Z>s)=1-
whenever 0 < s <5. The density function f7 for Z may be obtained by
differentiating Fz. The two functions are as follows:
Onis

if (S45) 1
F7(s) = ia ae nOssS5

i eyes
142 Chapter 6: Joint Distributions

pay etal if0<s<5

fps) = 42 25
0 otherwise

Once the density function for Z is known, E(Z) carrbe calculated:

E(Z) = [ t fz(t) at

. (pou a
<J9 285
5
04 | 10r—2¢2 dt
0

2
3
So the addition of the backup safety switch controlled by an independent operator
reduces the expected time until shutdown from 2.5 seconds to 1.67 seconds.
Of course, we can also compute E(Z) from Proposition 6.3, and this will be
the easiest way to get the expected value if we have no particular interest in the
probability density. Since Z = min{X, Y},

B=] [| ecyfxrey) aray

5 p5
= 4 [ } min{x,y} dx dy
0 “0

The factor .04 in the last integral is the product of fy(x) and fy(y), each of which
is 1/5. The way to evaluate the integral on the right is to realize that the value of the
function g(x, y) = min{x, y} is equal to y if (x, y) is a point lying below the
line y = x and is equal to x if (x, y) lies above the line y = x. So we split the
double integral into two pieces as in the equation below, and then we compute
S px 5S 5
BZ) ies i } O4y dydx + }| 04x dy dx
0 “0 0 #x
These integrals are easily evaluated. Each is equal to 5/6, so the calculation again
confirms that E(Z) = 5/3.
6.4 Sums of Random Variables 143

Proposition 6.4: For any two independent continuous random variables X and
Y, the distribution and density functions for Z = max{X, Y} are given by
Fo) =Fy() x Fy) and f7(t)=fy(t) Fy) + fp Fy)

The explanation for this observation is simple. To say that Z < tis to say
that X <tand Y <z, that is, (Z <1) = (X <1) A (Y <D), and so
F7(t)=P(ZSH=PXSt,Yson=PXs)PY Sd) =F yy) Fy
This equation can be differentiated with respect to ¢ to obtain the density function
CS Ge

6.4 Sums of Random Variables

Often models for common situations lead to consideration of sums of random

variables. This means that it is often necessary to know how to determine the
probability distribution for a random variable that is described as a sum of two or
more other random variables. A few frequently encountered scenarios which
involve sums of random variables are listed here.

Scenario 1; An experiment is conducted which results in a numerical reading

of some kind. Since experiments, when repeated, often do not produce identical
results, it may be desirable to conduct the experiment several times and compare the
results of the separate “trials.” If the experiment is conducted n times and the n
resulting readings are averaged, this numerical average is the random variable

= x Ky +X, +---+X,)

where X,,---,X,, are the random variables that give the readings on the n
individual trials.
Often it is desirable to reproduce experiments for the purpose of verification
of the results. In such cases it is important to know that the experiments are being
conducted “independently” in the sense that the outcome of one experiment doesn’t
affect the outcome of the others. In this context, the random variables X,,---,
X,, are independent random variables.
144 Chapter 6: Joint Distributions

Scenario 2: A piece of equipment is put in place, and a “cold standby”

backup piece of equipment is available to replace it when it fails. So service is
available from the two pieces of equipment for a period as long as the sum of their
individual operational lifetimes.

Scenario 3: There are various sources for a particular commodity. (For

example, there may be several manufacturers of a particular device.) The total
available supply is then the sum of the amounts available from the various sources.
If these fluctuate probabilistically, we are then talking about a sum of random
variables.

Sums of independent random variables are especially important. One tool for
working with such sums is provided by the “convolution” integral.

Definition 6.7: Iffand g are functions defined on the entire real number line,
then the convolution of f and g is the function denoted by f*g and defined by

feet) = | ft-s) els) as

Convolution is a commutative operation; that is, f*g = g*f. The change of

variables u = t—s in the integral demonstrates this fact. (See Problem 6.30.)
The convolution integral is useful in other mathematical settings as well. For
example, when Laplace transform techniques are used to solve differential
equations, the convolution integral comes into play because the inverse Laplace
transform of the product of two functions is the convolution of their inverse Laplace
transforms.
In the area of probability, the convolution integral comes into play when sums
of independent continuous random variables are encountered.

Proposition 6.5: If X and Y are independent continuous random variables and

if Z =X + Y, then the density function for Z is given by
fz = fx*fy
6.4 Sums of Random Variables 145

Proof: To qualify as a density function for Z, a function f must have the

property
t

Ft) = } f(s) ds for every real number t

By definition,

F(t) EAZS t) = P(X +Y Ste | fxy@, y) dx dy

X+ySt

frefc“fyl) yy) dy dx = i [Ae t beads

I

The last expression is obtained by performing the change of variables s = x + y

in the innermost of the integrals in the iterated integration. Now, if the order of
integration is reversed, we have

FA) = [| fx@ofils-2 ads

The reason that this form is desirable is that now the inner integral is the
convolution of fy and fy, which shows that fy*fy satisfies the criteria to be a
density function for Z.

Example 6.9. Two random numbers between 0 and 1 are independently

produced by a random-number generator. Find the probability distribution of the
sum of the two numbers.

Solution: It is intuitively clear that the sum will assume values between 0 and
2. Is it also clear to you that the values of the sum will be more concentrated near 1
than toward the endpoints of the interval?
Denote the two numbers generated independently by X and Y, and denote
their sum by Z=X+Y.
From Proposition 6.5, we need only compute the convolution of fy and fy,
each of which is the uniform density on [0, 1]. This is not difficult if we get the
proper geometry in mind. This is shown in the figure below.
Here the functions g and /f in parts a and b of the figure are simply two
copies of the uniform density on [0, 1]. It is the convolution of g and f that we
need to determine. In each part of the figure, the shaded area is the region between
the graph and the s axis. In part c we have a similar representation of the function
146 Chapter 6: Joint Distributions

f(—s), where f is the uniform density on [0, 1]. And in parts d and e we have
graphs of f(t—s). In part d the graph of f(t—s) is shown for a typical value of
t such that 0 <t < 1, and part e shows a similar graph for a f value in the interval
D<p< 2.

XS) f(s)

Now the integral to be evaluated is

fg) = | fe-s) g(s) as

The integrand will be nonzero only when both factors in the integrand are nonzero.
When 0 <¢ < 1, this means we should look at the overlap of the shaded regions in
parts a and d. This will be the region whose area the convolution integral
computes, and we get in that case f*g(t) = t (because the overlap is a rectangle of
height 1 and width t). Next, when 1 <t <2, we look at the overlap of the shaded
regions of parts a and e and see that the area of the overlap now is 2 —t. [Observe
that the base of the overlap rectangle goes from t—1 to 1, so the length of the
base of the rectangle is 1 - (t-— 1) =2-1.] The same pictures should convince
you that f*g(4) =O if <0 ort22.
Summary: The density function for Z = X + Y is the function

0 t<0
V<T<S1
Feat) =
Z-t. Leape2
0 P22
6.4 Sums of Random Variables 147

Notice that the result of this calculation does indeed show that values of the random
variable Z = X + Y will more concentrated near the center of the interval [0, 2].
(Graph the density function if this isn’t obvious to you.)

Proposition 6.6: If X and Y are independent, then

E(XY) = E(X) E(Y)

Proof: We will check the validity of this proposition in the continuous case.
The discrete case is similar if one replaces the pee and the density functions by
sums and the mass functions.
Since the product XY is expressed as a simple function of the random
variables X and Y, its expected value is given by Proposition 6.3 as

eon=[ [| ofa» ad =] [ oh@no aa

‘ax fy (x) [>A dy dx = (2x f(x) E(Y) dx

EM) | xfya) dx = EO) EW)

The key idea in this simple calculation is that the independence of X and Y causes
the joint density function to split into a product which is a function of x times a
function of y. Once this happens, the double integral splits into a product of two
single integrals that evaluate to E(X) and E(Y), respectively.

The converse of Proposition 6.6 is not true. Knowing just the fact that
E(XY) = E(X)E(Y) is not enough to conclude that X and Y are independent.
There is, in fact, a special term that describes a pair of random variables for which
E(XY) = E(X)E(Y). IfX and Y satisfy this condition, they are said to be
uncorrelated. So Proposition 6.6 states that if two random variables are
independent, then they are uncorrelated.
Recall that for any pair of random variables X and Y defined on the same
sample space, E(X + Y) = E(X) + E(Y). While expected value is additive in
this sense, in general variance is not. In the case of independent random variables,
however, the variance is additive. 3
148 Chapter 6: Joint Distributions

Proposition 6.7: If X and Y are independent, then

var (X + Y) = var (X) + var (Y)

Proof: If we denote E(X) by Ly and E(Y) by Ly, then

E(X + Y) = Uy + by
Therefore from Proposition 3.5 we know that
var (X + ¥)= EL (X + Y)*] - (ty + by)?
But then
var (X + Y) E(X? + 2XY¥ +¥7) — (uy? + yyy + My?)
Ea + By ly EIB) 2 oye
var (X) + var (Y)
since E(XY) = Uyby.
Using mathematical induction, one can extend Proposition 6.7 to any finite
collection of independent random variables. If X,,---,X, are independent,
then
var (X, +... + X,) = var (X;) +... + var (X,,)

A useful and easy application of this fact is in proving that

Oy = Vnpg
whenever X is binomial with parameters n and p. (This was the second part of
Proposition 4.1.) Think of X as the number of successes in n trials of an
independent trials process, so that
X = X,+X,+---+X,,
where X, is O or 1 depending on whether failure or success occurs on the kth
trial. It is easy to see that for k = 1,2,---,n, var(X,) = pq. The reason is that
E(X,2) =p, and that means
var(X;) = E(X?) - E(X,)" = p -p? = pq
From the additivity of the variance of independent random variables,
var(X) = var(X,) + - +. + var(X,,) = npq

The most “elegant” probability distributions as far as sums of independent

random variables are concerned are the binomial distribution and the normal
6.4 Sums of Random Variables 149

distribution. These are the two that reproduce themselves when independent
random variables are summed. Let’s consider sums of normal random variables
first.

Proposition 6.8: If X and Y are normal random variables and are independent,
then Z = X + Y is also normal with
Uz=Uxy+Uy and var (Z) = var (X) + var (Y)

Comment on proof: That Uz = Uy + Uy and var(Z) = var(X) + var(Y) is

already known; the latter depends on the fact that X and Y are independent. What
is new and important here is that Z inherits the distribution shared by X and Y;
that is, Z is normal. There are a variety of ways to prove this, all of which are
moderately tedious. One way is to use a powerful tool called the Fourier
transform. For continuous random variables, certain properties can be established
by considering the Fourier transforms of the density functions.
Another way to derive the fact that Z is normal is to use the convolution
formula of Proposition 6.5. This, however, leads to some quite messy algebra
which doesn’t shed much light on the proposition under discussion. For that
reason we will omit the proof and go on to illustrate its importance in situations to
which the proposition applies.

Example 6.10. A component in a system is known to have a normally

distributed lifetime with mean 2 years and standard deviation 6 months. A backup
component is available to be placed in service when the original fails. The backup
is identical to the original. What is the probability that the cumulative service
provided by the two components totals more than 5 years?

Solution: Let’s consider the length of time the original remains in service to
be X and the length of time the replacement remains in service to be Y. The
assumption then is that X and Y are normal random variables with mean 2 years
and standard deviation 1/2 year.
The combined lifetime of the two then is the random variable Z =X + Y,
which is normal with mean
EZ) =2+2=4
years, and variance
150 Chapter 6: Joint Distributions

var (Z) = var (X) + var (Y) = 6,’ + 6,” = T+ =F year

Therefore, 6, =V1/2 ~ .7071.

Now that we know that Z is normal with pz = 4 and 07 = .7071, we have
all the information we need to answer questions about Z. In particular,

ZS) ipa 3 ay t:(— > 14142)

1 — (1.4142) = .07865
(As usual, ® in this equation denotes the standard normal distribution function.
This calculation is based on Proposition 5.3.)

Proposition 6.9: If X and Y are independent binomial random variables with

X having parameters n, and p and Y having parameters ny and p, then
Z =X +Y is binomial with parameters n, + 12 and p.

Comment on Proposition 6.9: Proposition 6.9 says that the binomial

‘distribution is the discrete distribution that shares the inheritance property that the
normal distribution displays. A little reflection on how the binomial distribution
arises leads to an intuitively simple understanding as to why it is true. Consider an
independent trials process with success probability p in which X counts the
number of successes during the first n, trials and Y counts the number of
successes during the next n, trials. The nature of an independent trials process
guarantees that X and Y will be independent random variables, and X + Y will
give the number of successes during the first n, + n trials and so will be binomial
with parameters n, + n2 and p.

Often it is assumed that errors inherent in laboratory procedures cause a

normally distributed pattern of results when repeated measurements are made of a
given quantity. In fact, the normal density function is sometimes called the “error”
function. One technique that is often used in making measurements where errors are
expected is to take several measurements and average the results. There is a clear
theoretical basis for this. Suppose, for example, that a measurement made in an
experiment is assumed to result in a normal random variable with mean u and
variance 62, If the experiment is repeated n times independently, we can think of
6.4 Sums of Random Variables 1534

the n resulting measurements as random variables X 1>°**>X,, which all have

mean | and variance o? and which are independent. The average Y is then given
by Y = (1/n)(X, +---+X,,), and the mean and variance of Y are given by
1
et aD (it-.-+[) =p
and

var VY) S=-

lene var(X= = xno? =
2
&
Tea | n Me

The fact that the variance of the average of the measurements is appreciably
smaller than the variance of an individual measurement means that it is much more
probable that the average will be near uy than will an arbitrary single measurement.
A common assumption is that the measuring technique is not biased, that is, that the
expected value u of any given measurement agrees with the true value of whatever
quantity is being measured.

Example 6.11. The voltage in a circuit is 115 volts. A particular technique

for measuring the voltage gives readings which are normally distributed with mean
ut = 115 volts and standard deviation o = 5 volts. If four readings are taken and the
results averaged, the resulting average is normally distributed with mean t = 115
volts and variance o*/4. So the standard deviation of the average is 5/2 = 2.5 volts.
Let’s do a brief calculation to indicate the value of this. If Y is the average of
four readings, then

Y= 115
2-—20(1.2) = .2301

whereas for an individual reading X,

IX — 115!
P(X —1151>3) = pa > 6) 2 —20(.6) = .5485

Summary: The probability that an individual reading will differ from the
actual voltage by more than 3 volts is 5485, whereas if four readings are taken and
the average used then the probability that the average will differ from from the true
value by more than 3 volts is only .2301.
152 Chapter 6: Joint Distributions

6.5 Conditional Probabilities and Random Variables

In Chapters 1 and 2 we found the concept of conditional probability to be very

useful. The idea helps to describe how likely one event is to occur if we know that
some other event has occurred.
There is a similar situation regarding pairs of random variables. Suppose we
have two discrete random variables X and Y with their associated probability
mass functions py and py. The mass function py carries the necessary
information about the probability with which X assumes its various values. But
what if we know what value Y assumes? For instance, let’s suppose that we
know that Y takes on a particular value b. If X and Y are not independent
random variables, then in general events of the form (X = a) and (Y = b) are not
independent events. This means that P(X =a) and P(X =alY =b) will not
be the same. The conditional probability P(X =a1Y =b) is the conditional
probability that X takes on the value a given the knowledge that Y takes on value
b. This leads to the idea of the conditional probability mass function. The
conditional probability mass function will be a useful tool for studying the ways in
which one random variable influences another.

Definition 6.8: IfX and Y are discrete random variables and if b is a number
having the property that P(Y = b) # 0, then the conditional probability mass
function Pyyy_, is defined by
Pry = ae eo
for each real number x.

Recall from the definition of conditional probability that

P(X=x,¥ =b) = PY =b)
PX =x1Y=5d)
This equation relates the joint probability mass function of X and Y to the
conditional probability mass function pyy_,, for it can just as well be written as

Px yx, b) = py(b) Pyiyap)

Often situations involving more than one random variable are most naturally
described in terms of conditional probability mass functions. The following
6.5 Conditional Probabilities and Random Variables 153

example illustrates this.

Example 6.12. A particular device is used in various kinds of systems. We

will assume that all systems employ either 1, 2, 3, or 4 of these devices and that
each of these four possibilities is equally likely to be the case. Each device
employed in a system has probability p = .1 of failing, and the devices function
independently. This means that once we know how many devices are present, the
probability distribution of the number of failures will then be known. For example,
if it is known that a system employs 3 of the devices, then the number that fail will
be binomial with parameters n = 3 and p =.1.
We will denote by X the number of failures of such devices in the system
and by Y the total number of devices employed in the system. What we have
observed is that for b = 1, 2, 3, and 4, the conditional probability mass function
Pxiy=p 1S the binomial mass function with parameters n = b and p =.1.
As a sample computation that can be based on these observations, we will try
to determine the probability that there are exactly two failed devices in the system.

Solution:
4

P(X =2) >) PX =2,¥=5)

b=r1
4

>) PY =b) PX =21Y =5)

b=]

PY =2) P& =21Y =2)+ PY =3)P4 =21Y¥ =3)

+ P(Y = 4) P(X =21Y =4)

Lea) wick 2
me (.1)* + Z C32) EL) FCO) + 7 C42
A Signe
T)-C9)

0214

Determining the expected value for the number of failed devices in the system
is not a difficult computation if we proceed along the lines begun above. Having
computed P(X = 2), it is clear that P(X = 0), P(X = 1), P(X = 3) and
P(X = 4) can also be determined in a similar manner. Moreover, once all the
values of the mass function of X are known, the expected value is easy.
154 Chapter 6: Joint Distributions

There is another point of view, however, that makes this computation still
easier. It involves the idea of conditional expectation. Just as the expectation of a
discrete random variable utilizes the probability mass function, the conditional
expectation utilizes the conditional probability mass function.

Definition 6.9: Suppose that X and Y are discrete random variables and that
P(Y =b)>0. The conditional expectation ofX given that Y = b is denoted
by E(X | Y = b) and defined by

E(X|Y=b) = >. x, P(X =x, |Y =b)

k
where the sum extends over all values x, assumed by the random variable X.

Notice that all that is happening in this definition is that the probability mass
function in Definition 3.6 is now being replaced by the conditional probability mass
function.
Proposition 6.10 will now demonstrate a way in which the conditional
expectation can be used to determine the unconditional expectation. Depending on
what kind of information is available, this may in fact be the most attractive
approach to get the expected value. The continuation of Example 6.12 that appears
after Proposition 6.10 will illustrate this.

Proposition 6.10: If X and Y are discrete random variables, then

EX) = >, PW=y) EXIY=y)

j
The sum extends over all values y; assumed by Y.

Proof:

Se PY =y) EXIY=y,) = py PY =y) >, Xp P(X =x, 1Y =y,)

a J k
6.5 Conditional Probabilities and Random Variables 15

= py2 x, PY = yj) PX =x, 1Y =y,)

= » SS x4 P(X =xX,,
Y =yj)
jek

= Dd)ted PR =tpE=y)
k j

= So Xx P(X = Xz)
i ,

= E(X)

Notice that this proof depends only on rearranging the terms of the sum and using
elementary properties of conditional probabilities.

Example 6.12 (continued). Proposition 6.10 gives an easy way to determine

the expected number of failed devices in the system we were examining in Example
6.12. The conditional probability distribution of X, given the information Y = 4,
is the binomial distribution with parameters n = 4 andp =.1. So E(X |Y =4)
=o 4% I= tA. Ineaddition, EGY
= 3)i= 3 xb =23.,and ECW =)
Soul sande x NL) = 1. de)... Dherefore,

BD16.¢ wr axa + +x 3 + ax2 + -x.1 = 25

Now let’s turn to the case of continuous random variables. What will be
meant by the conditional probability density function in the case of two continuous
random variables? We cannot proceed as in Definition 6.8 because the two events
(X =x) and (Y = b) will both have probability 0 if X and Y are continuous
random variables. Remember, however, that even in the discrete case

P(Xs xyV=b)
PX =x1¥=b) = —Sysy
Both the numerator and the denominator of this expression on the right have parallel
concepts in the continuous case. The numerator is the joint probability mass
156 Chapter 6: Joint Distributions

function of X and Y, which, of course, corresponds to the joint density function

when continuous random variables are considered. Moreover, the denominator is
just the mass function for Y, which corresponds to the density function of Y in
the continuous case. The analogy then leads to Definition 6.10.

Definition 6.10: Suppose X and Y are continuous random variables. For all
numbers y where the density function fy satisfies fy(y) # 0, the conditional
density function of X given that Y = y is denoted by fyjy-y and defined by

Fyyy)
Fxiyay) = 7,0)
Similarly,

fyyy)
frx=x) = “F.0)

Example 6.13. A closer look at an example considered previously is in order

now. Suppose a point is randomly chosen in the following shaded triangle.

The assumption will be that the coordinates of the point selected, which we
denote by X and Y, have as their joint density function the uniform density on the
triangle.
In this context, what is meant by the conditional density fyy_3/4? From the
definition, fy)y-3/4(x) = 2/fy(3/4) provided that the point (x, 3/4) lies inside the
triangle, and fy,y-3/4(x) = 0 otherwise. (Recall that the joint density has the
constant value 2 throughout the triangle and 0 outside the triangle.) Furthermore,
the point (x, 3/4) will be inside the triangle precisely when 3/4 <x < 1.
It is easy to compute (using Proposition 6.2) that fy(y) = 2 — 2y whenever
O<y <1. So fy(3/4) = 1/2. Therefore fyiy-3;4(x) = 2 + .5 = 4 whenever
3/4 <x <1. Notice that this is simply the uniform density function on the interval
6.5 Conditional Probabilities and Random Variables 157

[3/4, 1].
One final look at the picture should make all of this fall into place. Once we
know that Y = 3/4, this guarantees that if (X, Y) is in the triangle then X must
satisfy the inequality 3/4 < X <1. The “uniformity” of the joint density on the
triangle is passed down to X when we are conditioning on the information that Y
= 3/4 in the sense that the conditional density for X then is the uniform density on
the interval [3/4, 1].

Definition 6.11: If X and Y are continuous random variables and fy(y) # 0,

then E(X | Y =y) denotes the conditional expectation of X given that Y = y
and is defined by
co

EQIY=y) = | thayay(0 at

Notice that the only way in which the above definition differs from the
definition of E(X) is that the density function fy is replaced by the conditional
density function. Also it differs from Definition 6.9 only in that summation of the
conditional mass function turns to integration with the conditional density function.
In the continuous case, just as in the discrete case, the conditional expectation
can be used to compute the unconditional expectation. The summation and mass
function in Proposition 6.10 are replaced by integration and the density function in
the continuous case.

Proposition 6.11: IfX and Y are continuous random variables having a joint
density function, then

BO) = | fo) E&Y =y) dy

Proof:

[nor eciv=ydy = [fo] they ae

158 Chapter 6: Joint Distributions

(t,y)
fyy
ify) =a Cy
-{. Ie ty fy)

= | | theres) at ay
The last integral above is equal to E(X). In fact, this integral is just the
special case of the integral in part 2 of Proposition 6.3 in which g is the function
g(x,y) =x. (You may, of course, change the variable ftto x in the double
integral here if it helps you to see the connection between this and Proposition 6.3.)

Just as in the discrete case, with models based on continuous random

variables, it is often most natural to state assumptions in terms of conditional
densities. Example 6.14 illustrates this.

Example 6.14. An electrical cable has a time to failure which is exponentially

distributed. The parameter 4 in the distribution, however, varies from one
manufacturer to another. We will assume that for a given supply of cables of
different origins, the parameter in the exponential distribution is uniformly
distributed between A = .5 and 4 = 1. We will denote by Y the random variable
whose value will be the parameter A in the exponential distribution corresponding to
the time to failure for a randomly selected cable from the given supply. And we
denote by X the time to failure for such a randomly selected cable, where time is
measured in years. This means that for any number y, if we know that Y = y,
then we know precisely the characteristics of X; that is, the conditional density
function for X based on the knowledge that Y = y is the exponential density with
parameter A = y. So in precise mathematical language, the assumptions are that Y
is uniform on the interval [.5, 1] and that for a given value of y between .5 and 1,
fxiy=y 18 the exponential density with parameter A = y. This means that the
conditional expectation E(X | Y = y) will be 1/A = 1/y.
What is the expected time to failure for a cable randomly selected from this
batch?
Solution:
oo
2 1
E(X) = | fy) E(X1Y =y) dy -| 5 dy =2In2
=1.386
275s a
6.5 Conditional Probabilities and Random Variables 159

Does this make sense? The worst cables (where A = 1) have expected time to
failure 1 year. The best ones (where 4 = .5) have expected time to failure 2 years.
The parameter in the distribution for the available supply of cables is assumed to be
uniformly distributed between .5 and 1, so the average would be .75. In other
words, E(Y) = 3/4. Notice that E(X |Y = 3/4) = 4/3 since the conditional
density fyy —3/4 is the exponential density with parameter X = 3/4. If it seems
paradoxical to you that this 4/3 doesn’t agree with the answer 1.386 above, the
explanation is that two different kinds of averaging are being compared.

Example 6.15. A person will arrive at work between 9 and 10 o’clock in the
morning. Sometime before 10 o’clock an important phone call must be placed.
Assume that the time of arrival is uniformly distributed between 9:00 and 10:00,
and assume that the time that the call is placed is uniformly distributed between the
time of arrival and 10:00. What is the probability distribution of the time at which
the call is placed, and when is the expected time for the call to be placed?

Solution: This situation ties together several of the concepts of this chapter
because there are two random variables that interact. One is the time of arrival (let’s
call it X), and the other is the time at which the call is placed (which we’ll call Y).
If we agree to measure time in hours starting at 9:00, then the assumption
regarding X is thatX is uniformly distributed on the interval [0, 1]. But what
about Y? Here the information is conditional. It is that given the information X =
x, then Y is uniformly distributed on the interval [x, 1]. Or to put it more
succinctly, fyy-, is the uniform density on the interval [x, 1] for each value of
x between 0 and 1.
We can solve for the joint density by simply looking back at Definition 6.10.

fy.y@, y) =f) fyixex)

Since fy(x) = 1 whenever 0 < x < 1 and since fyjy_,(y) = 1/-x) whenever
x<y<1, this means that

i if O<x <1 andx<y <1

fy y@ y) = ae
0 otherwise

The (unconditional) density function for Y can now be calculated from

Proposition 6.2. We find that fy(y) =—In (1 — y) for 0 <y < 1, with fy) = 0
160 Chapter 6: Joint Distributions

otherwise. (Check the details of this calculation.)

For the expected value of Y, we can of course compute E(Y) from its own
density function:
b, 1 1
EW) = } tf.)dt = }-tin(l-1)dt= | a-oinear
0
by a simple change of variables. Therefore, using integration by parts we can now
derive the expected value.
2 1
eA t 1 ue > ¥ es
EW)-= (¢—)Ine| + i 7 (t—t/2) C= q

The first term here is 0. (You have to look carefully at the limit as t-0 to see
that this is true since In t + —ce.) The integral on the right is equal to 3/4. This
means that the expected time for the call to be placed is 9:45.
However, if it’s the expected value we want, then the easy route is to use
Proposition 6.11. Since fyy_, is the uniform density on [x, 1], this means that
E(Y |X =x) = (x + 1)/2, simply the midpoint of the interval [x, 1]. But then
Proposition 6.11 says that
co 1
E(Y) - | f(x)
EW IX =x) dx -| 1x dx = 5

6.6 The Central Limit Theorem

Proposition 5.4 is the original version of the central limit theorem and is often
referred to as the De Moivre-Laplace theorem. It states that the binomial distribution
with parameters n and p is closely approximated by the normal distribution if the
parameter n is large.
Recall that every binomial random variable is a sum of Bernoulli random
variables. (A Bernoulli random variable is a binomial random variable in which the
parameter n is equal to 1.) For example, if Y is the number of successes in 4
trials of an independent trials process, then
y= xq + Xo +X3 +X4
where X, is either 0 or 1, depending upon whether success occurs on the first
trial, and X2, X3, and X4 similarly indicate whether success occurs on the
second, third, and fourth trials. Furthermore, the random variables X,,-- -, X4
6.6 The Central Limit Theorem 161

all have the same distribution function and they are independent.
If the parameter n in the binomial distribution of Y were large, then Y
would be the sum of a large number of independent, identically distributed random
variables. A more general form of the central limit theorem says that this is all that
is required in order for “convergence” to the normal distribution to take place.
Specifically, suppose that X,, X2, X3,--- are random variables which are
independent and have a common distribution function. Furthermore, suppose that
they have finite mean 1 and variance o”. (Since they all have the same distribution,
they must all have the same mean and variance.) For each positive integer n, let
S, =X, +++. +X, (6.2)
Then S,, has mean nu and variance no”. (The variance of a sum is the sum of the
variances. Proposition 6.7 states this fact for two random variables, but it is
equally true for any finite number of independent random variables.)
The De Moivre-Laplace version of the central limit theorem, Proposition 5.4,
requires that we subtract the mean of the binomial random variable and divide by
the standard deviation in order to make the binomial random variable approximately
standard normal. The effect of subtracting the mean and dividing by the standard
deviation is to give the “adjusted” random variable mean 0 and variance 1, just as
the standard normal distribution has mean 0 and variance 1. The corresponding
adjusted form of the random variable S,, in Equation 6.2 is the random variable
5), Hy
a (6.3)

Proposition 6.12 (Central Limit Theorem): If X;, X2, X3,--- are

independent random variables having the same distribution and having finite mean
ut and variance o%, and if for each positive integer n
Sot XK norkNy
then
oy — nA

oVn
is approximately standard normal when vis large. (More precisely, the distribution
functions converge to the standard normal distribution function as n > ©.)
162 Chapter 6: Joint Distributions

A good model to consider in order to understand the implications of the

central limit theorem is that of a numerical-valued experiment conducted over and
over again. The result of each experiment can be considered an observation of a
random variable. For example, X; gives the outcome of the first experiment, X2
the second, and so forth. If the experiment is conducted in the same manner each
time, and if the various experiments are conducted independently of one another,
then the random variables X,, X2, X3,--- will be independent and will have
the same distribution function. However, there is no restriction necessary on just
what that distribution might be. For example, each of these random variables could
be uniform on [0, 1]. Nevertheless, if the experiment is repeated a large number of
times and if S, is the sum of the results of the first n experiments, then the
distribution of the adjusted form of S, in Expression 6.3 (with mean 0 and
variance 1) will be approximately standard normal.
The central limit theorem indicates why the normal distribution is so very
important. Independent observations of any type of random variable will “average
out” to a normal distribution if the results are averaged in the sense of Equation 6.2
and Expression 6.3.

Problems

6.1 A coin is tossed twice. Let X denote the number of heads on the first toss
(0 or 1) and let Y denote the total number of heads on the two tosses (0, 1,
or 2). Draw a table similar to Figure 6.1, showing all values of the joint
probability mass function for X and Y.

6.2 In Example 6.1, let Z denote the sum of the numbers on the two items
selected. Draw a table similar to Figure 6.1, showing all values of the joint
mass function for X and Z. As in Figure 6.1, sum the rows and columns
to show the values for the mass functions of X and Z along the bottom
and right side.

6.3. A device contains three transistors and three resistors. One transistor and
one resistor are defective. Two of the six components are randomly
selected. Let X denote the number of transistors selected and Y denote the
number of defective components selected. Draw a table that shows all the
values of the joint probability mass function for X and Y.
Problems 163

6.4 A card is drawn from a standard deck of 52. Let X be the number of hearts
drawn and Y the number of red cards drawn. (So X and Y assume
values 0 or 1.) Show all values of the joint probability mass function in a
table.

6.5 A coin is tossed three times. The random variable X is the number of
heads occurring on the first two tosses, and Y is the number of heads
occurring on the last two tosses. Draw a table like Figure 6.1 for the joint
probability mass function for X and Y. (Observe that X and Y are not
independent.)

6.6 Let T be the triangle bounded by the x and y axes and the linex+y=1.
Suppose that f is the function defined by f(x, y) = cxy for (x, y) in T,
and that f(x, y) = 0 when (x, y) is not in T.
(a) Find the value that the constant c must have in order for f to be a joint
density function.
(b) Suppose that this is the joint density for random variables X and Y.
Find the density fy.
()-LeeZ a maxixe Yeakind F(Z <l/2):

6.7 Suppose X and Y have the joint density function f(x, y) = 24xy inside
the triangle bounded by the line x + y = 1 and the coordinate axes, with
f(x, y) = 0 for points (x, y) not in this triangle. Find the expected
values of X, Y, and XY.

6.8 Two random numbers are independently generated using a random number
generator which generates numbers according to the uniform density on
[0,1]. Find the expected value for the square of the difference of the two
numbers.

ae) Two dice are rolled. Let X denote the maximum of the two numbers that
appear, and let Y denote the minimum of the two numbers that appear.
(a) Ina table show all values of the joint mass function py y.
(b) Find E(X) and E(Y).

6.10 This exercise is designed to check your understanding of what Equation 6.1
says. Suppose two dice are rolled and X and Y are the numbers that
appear on the two dice. Let B denote the region in the xy plane consisting
 Chapter 6: Joint Distributions

of all points (x, y) such that x2 + y2 < 15. Compute the probability
P{(X, Y)eB} and verify that Equation 6.1 is valid in this specific case.

Suppose X and Y are independent continuous random variables. Show

that Definition 6.6 implies that
P@<X <b, ¢<Vi<d) =P@<sX <b) Re<Y-<a)
whenever a <b andc<d. [Hint: This is elementary; just split the
double integral into a product.] Show that this is also true when a or c is
—eo or when b or dis +09.

6.12 Suppose two random numbers between O and 1 are independently

generated. (Assume a uniform distribution for each.) What is the
probability that the first is more than twice as large as the second? (See
Example 6.5.)

6.13 Suppose that X and Y have as joint density function the uniform density
on the triangle in the xy plane having vertices at (0, 0), (1, 0), and (1, 1).
Determine the density functions of X and Y. Are X and Y independent?

6.14 An experiment is conducted four times. On each attempt, the probability

that the experiment succeeds is 1/2. Let X denote the total number of
successes, and Y the longest run of successes; that is, the greatest number
of successes to occur consecutively during the four experiments. Construct
a table showing all values of the joint mass function for X and Y as well
as (by summing rows and columns) the mass function of X and the mass
function of Y.

6,15 Let f(x, y) = ye” whenever x > 0 and y > 0, with f(x, y) = 0
otherwise.
(a) Show that fis a density function.
(b) If fis the joint density function for a pair of random variables X and
Y, find the density function for X and similarly for Y.

6.16 Complete the calculation of the probability that the device with expected
lifetime 8 years lasts more than twice as long as the device with expected
lifetime 5 years in Example 6.6.
Problems 165

Oly A friend says he will call you on the phone between 8 and 9 o’clock.
(Assume a uniform distribution for the time of the call during this interval.)
Other phone calls occur on your phone according to an exponential
distribution; that is, the waiting time for a call is an exponential random
variable. Let’s assume the expected waiting time for a call from someone
other than the friend is 30 minutes. If you arrive at home at precisely 8
o’clock, how likely is it that your friend will be the first person to call you?
Comment on this problem: Remember the lack-of-memory property that
exponential random variables have. This means that if you walk in at 8:00,
it really doesn’t matter when the last call occurred. You might as well
consider the whole process starting from 8:00; that is, the waiting time as
measured from 8:00 until the next call (from someone other than the friend
who promised to call) can be assumed to be exponentially distributed with
expected value 30 minutes.

6.18 Fill in the details in calculating that E(Z) = 1/3 in Example 6.7.

6.19 A room is lighted with two 100-watt bulbs and one 60-watt bulb. During
the course of a week, two of the bulbs burn out. Let X be the wattage of
the first bulb to burn out and Y the wattage of the second. Assume that
bulbs are not replaced when they burn out and that the bulbs are equally
likely to burn out.
(a) Draw atree diagram to represent the possibilities.
(b) Ina table similar to Figure 6.1, show all values of the joint mass
function of X and Y.
(c) What is the expected value of X + Y?

6.20 Suppose the random variables X and Y represent the lifetimes of two
devices (measured in hours) and that the joint density function for X and
Y is given by
fx, y) = .02 elk —.2y

(a) Find fy and fy.

(b) Are X and Y independent?
(c) What is the probability that both devices last longer than 10 hours?
[Hint: This is simply P(X > 10, Y > 10).]
(d) Now find the probability that the sum of the lifetimes of the two
devices is greater than 10 hours. [Hint: This is just P(X + Y > 10),
166 Chapter 6: Joint Distributions

but you may find it easier to compute this expression by looking at the
complement and using P(X + Y>10) = 1-P(*+Y<10).]

6.21 Answer parts c and d of Problem 6.20 under the assumption that X and Y
are independent and uniformly distributed. This time assume that the
expected lifetime of each device is 10 hours.

6.22 A device is known to have an exponentially distributed lifetime with

expected value 5 years. Your plan is to replace it immediately with a similar
device when it stops functioning. So if X and Y are the lifetimes of the
devices, then X and Y are independent and each has an exponential
distribution with parameter A = .2.. What is the probability distribution for
the length of service you will get from the two devices together? You
know, of course, that E(X+Y) = E(X) + E(Y). The question is:
What is the probability distribution of X + Y? [Hint: Since X and Y are
independent, you know their joint density. For any real number f,
Fy,y(t) = P(X+Y <t). This can be computed by looking at the
correct double integral. Once the distribution function is known,
differentiate to find the density function for X + Y. An alternate solution
method would be to compute the convolution of the two exponential
densities. See Problem 6.34.]

6.23 Two random numbers between 0 and 1 are independently generated.

Assume each is uniformly distributed on [0, 1]. What is the probability
distribution and the expected value for the larger of the two numbers
obtained? [Hint: Use Proposition 6.4.]

6.24 Two components of a system have exponentially distributed lifetimes, one

with expected lifetime 1 year and the other 2 years. Assume the two
function independently. What is the probability distribution and the
expected value for the length of time until both devices have burned out?

Gi29 A measurement is made, and the result is a random variable having mean 10
and standard deviation 1. How many times would the experiment have to
be repeated independently until the average for all the measurements
obtained would have mean 10 and standard deviation less than 1/5?

6.26 Two houses are built in separate flood plains. House A is in a 100-year
flood plain, and house B is in a 50-year flood plain. This means that the
Problems 167

expected waiting time fora flood in the two areas is, respectively, 100 years
and 50 years. Assume these waiting times both to be exponentially
distributed random variables. Assume furthermore that these two waiting
times are independent random variables. This means that you know the
joint density function since you know the distribution of each. What then is
the probability that house A will be washed out by a flood before house B
is?

All Suppose both X and Y are uniform on [0, 1] and are independent.
(a) Find P(X +Y <1).
(b) Find P(X +Y <1), assuming that 0<r< 1.
(c) Find P(X +Y <2), assuming now that 1<1t<2.
(d) Now put all this information together and sketch a graph of the
cumulative distribution function for W =X + Y.

6.28 Suppose X and Y are random variables having as joint density function
the uniform density on the square S = {(x, y) |0<x<1,0<y<1}.
(a) Find the probability P(Y < X?).
(b) Find E(XxY?).

6.29 In Example 6.8, greater safety (in the sense of a shorter expected time to
shutdown) is achieved by putting the two switches in series and having
them independently controlled. If they were instead placed in parallel, as in
the picture below, the effect would be to make it more difficult to shut down
the system.

Power source
Process that may
need to be
interrupted
Switch 1 nt P

Switch 2

With this kind of configuration, the system would not be shut down
until both operators had opened their switches. This kind of configuration
168 Chapter 6: Joint Distributions

could be desirable if time is not terribly critical (from the standpoint of the
danger involved) and if a shutdown is expensive enough that we elect not to
shut down unless both operators agree that a shutdown is necessary. With
this configuration, the waiting time for the system to be shut down is the
random variable Z = max{X, Y}. What is the expected waiting time for
the system to be shut down if this configuration is used? (Assume that the
assumptions aboutX and Y remain the same as in Example 6.8.)

6.30 Show that f*g = g*f. (See the comment following Definition 6.7.)

6.31 (a) What does the joint probability mass function of X and Y look like if
X and Y are the same random variable? Make up a simple example if
this situation sounds confusing to you. For example, let X and Y
both be equal to the number obtained when a die is rolled. If we were
to roll the die twice and let X be the number on the first roll and Y the
number on the second, then X and Y are independent and certainly
not equal. (Their probability mass functions are equal.) Now,
however, you need to think of X and Y as both corresponding to the
same roll of the die.
(b) As a follow up to part a, can you intuitively see why in the continuous
case if Y = X, then it is impossible for X and Y to have a joint
density function? The reason is that P{(X, Y)eB} would then have
to be 0 unless the region B intersects the line y = x. From this it can
be proved that the joint density would have to be 0 except on the line y
=x, and such a function could not possibly satisfy

J [feo dxdy = 1
6.32 Many devices, such as lightbulbs, for example, have time to failure which is
approximately normally distributed. (Notice that the exponential
distribution is not a good model for a device that is physically wearing out
when in use. The lack-of-memory property would be totally out of place
here. For a device with the lack-of-memory property, a used device is
always just as good as a new one.) Let’s suppose that 75-watt lightbulbs
have normally distributed time to failure with mean 750 hours and standard
deviation 150 hours. If you buy three bulbs, what is the probability that
Problems 169

you get more than 2,000 hours use from the three?

6.33 The lifetime of a device is a normal random variable with expected value
1,000 hours and standard deviation 100 hours.
(a) If an identical backup device is available to be placed in service when
the original fails, what is the probability that the total length of service
they provide is more than 1,800 hours?
(b) If three such devices are available, what is the probability that the
cumulative service they provide exceeds 2,700 hours?
(Assume in all cases that the devices function independently; that is, the
lifetimes of the devices are independent random variables.)

6.34 Use the convolution formula to derive the distribution of X + Y, where X

and Y are independent and each has the exponential density with parameter
X = 1. This computation is not difficult; just be careful about your
endpoints of integration.

6.35 A person’s blood contains 70 parts per million (ppm) of a certain substance.
When a particular technique is used to measure the concentration of the
substance in the person’s blood, the result is a normal random variable with
mean 70 ppm and standard deviation 10 ppm. Assuming that it is possible
to reproduce the tests independently, what is the probability that the average
of four tests conducted independently would be between 65 and 75 ppm?

6.36 In Example 6.1, find the conditional probability mass functions pyjy_7 and
Pxiy=3- -Then use these to compute E(X | Y = 2) and E(X | Y = 3).

6.37 If X and Y are independent discrete random variables and if P(Y =b) #
0, then show that the mass function py and the conditional mass function
Pxiy=p are identical as functions.

6.38 If X and Y are as in Problem 6.3, find the conditional mass functions
Pxiya=1 and pyly-). Then find E(X | Y = 1) and E(Y |X = 1).

6.39 In Example 6.14, what is the joint density function for the pair of random
variables X and Y?

6.40 In Example 6.4, what is the conditional density function fyy.9? Now
find the conditional expectation E(Y | X = 0).
170 Chapter 6: Joint Distributions

6.41 A random number generator is used to generate 100 random numbers

uniformly from the interval [0, 1]. Use the central limit theorem to estimate
the probability that the sum of the numbers lies between 50 and 52. [Hint:
Sio0 =X, +++-+Xj09, where the random variables on the right are
independent and uniform on [0, 1]. If Si09 1s adjusted so as to have mean
0 and variance 1, the resulting random variable is then approximately
standard normal.]
 Chapter 7: Stochastic Processes

Sometimes a situation is best described by a large family of related random

variables rather than by a single random variable or a few jointly distributed ones.
A stochastic process is a family of random variables indexed on some set of real
numbers. One common situation is that of a sequence X,, X,,--- of random
variables. In this case the subscripts are positive integers. Another common
situation is to have a family of the form {X;,: t= 0}, in which case the subscripts
may be any non-negative real numbers. The former is an example of a discrete
parameter stochastic process since the positive-integer subscripts form a discrete
set of numbers. The latter is a continuous parameter stochastic process because the
parameter (subscript) ¢ is measured on a continuous scale. In the most common
examples of continuous stochastic processes, ¢ represents time.

7.1 Independent Trials Processes

Without using the language of stochastic processes, we have already
repeatedly encountered one of the most common ones. An independent trials
process is an example of a stochastic process. Here the random variables of
interest are the random variables X,, X>,--- where, for each positive integer n,
X,, is the number of successes to occur during the first n trials of the independent
trials process. The entire process cannot be described via a single random variable
if the number of trials is not fixed in advance, but rather the interaction of the
random variables describes the process.
In an independent trials process, it is easy to see that the only way that there
can be exactly k successes in the first n trials is for one of two things to happen.
Either there must be k successes during the first n — 1 trials and failure on the nth
trial, or else there must be k — 1 successes during the first n — 1 trials and success
on the nth trial. This leads to the observation stated next as Proposition 7.1.
Equation 2 of the proposition is just an application of the familiar multiplicative law
from Proposition 1.1 in Chapter 1.

171
WZ Chapter 7: Stochastic Processes

Proposition 7.1: In an independent trials process with success probability p, if

X,, denotes the number of successes to occur during the first n trials, then the
random variables X,, X, -- - are related as follows:
1. P(X; =1) = p, P(X, =0) = 1-p and -
2. PQ, = k) = pPX1 = kelp ed apy Pas *)

It is, in fact, true that Equations 1 and 2 in Proposition 7.1 characterize the
binomial distribution. In other words, if all one knows about X,, X9, - - - is that
Equations 1 and 2 in Proposition 7.1 are valid, it is easy to show that in fact each of
the random variables X,, does have the binomial distribution with parameters n
and p. This argument uses mathematical induction. Suppose that Equations 1 and
2 are known to be true. Equation 1 says that X, is binomial with parameters n =
1 and p. Assume now that X,_; is binomial with parameters n — 1 and p. Then
from Equation 2,
P(X, =k) = p P(X), =k-1) + (1 -p) P&,4 =4)

But since X,,_; 1s binomial with parameters n— 1 and p, this means that

P(X, =k) =p C(n-1,k-1)p*? A-py* + (-p)C-1,4) pd —p)

=C@-Lk-Dp (op + Caen Gone.
= {C(n—1,k—1)+C(n—1, 8} p* (1 —p)*
= C(n, k) p* (1 -p)"™
The last line uses the fact that C(n-—1,k —1)+C(n—-1,k) =C(n, k).
There is an easy way to see why this little identity is true by just remembering that
C(n, k) is the number of ways of choosing k elements from a set of n elements.
If we separate out one object from a set of n objects, then we can choose k objects
from the entire group by either (1) choosing k objects from the n — 1 still grouped
together, or (2) choosing k — 1 from the n — 1 and then including the one sitting
off to the side. Since (1) can be performed in C(n — 1, k) different ways and (2)
can be performed in C(n — 1, k — 1) different ways, this gives C(n — 1, k) +
C(n — 1, k—1) ways of choosing k objects from the n objects.
7.2 A One-Dimensional Random Walk 473

7.2 A One-Dimensional Random Walk

Example 7.1. A computer network serves many users simultaneously. Each

time there is a change in the number of current users, we will suppose that the
probability is p that a new user is entering the system and that the probability is ¢
= 1 —p that an old user is leaving the system.
Question: What is the probability distribution for the net gain in the number
of users of the system after a given number of users have either entered or left the
system?

Model: If we think of Y, as the net gain after n “additions” or

“subtractions” from the system, then the random variables Y,, Y2,--+ form the
stochastic process of interest. Though they do not have a binomial distribution,
they are closely related to a sequence of binomial random variables. Denote by X,,
the number of “‘additions” to the system during the first n changes. This means
that n—X,, will be the number of “subtractions” from the system during these n
changes. The advantage in looking at X,, rather than Y,, is that X,, is binomial
with parameters n and p, and Y, can now be described in terms of X,. For in
fact the net gain Y,, after n changes in the number of users on the system will be
equal to the number of new additions minus the number of people leaving the
system. The net increment then is X, —(n-—X, ) =2X, —n. Thus for a given
integer k,

PW, = =POX, n=) = P(X, as

;

One thing to notice is that the right side here is clearly 0 unless n+k is an
even integer. This is because after an even number of changes in the number of
users (that is, when 7 is even), the net gain must be an even number (that is, Y,, is
even). Thus P(Y, =k) will necessarily be 0 if n is even and k is odd. For
similar reasons, P(Y,, =k) will be 0 if n is odd and k is even. Since X, is
known to be binomial with parameters n and p, we can now describe the
probability distribution of Y,,.
If a =(n+k)/2 is an integer, then

PY, =k) =PX,=0a)=C(n,a)p*q?

For instance, suppose that the number of users of the system initially was 10,
that the number of users has changed (because of persons joining or leaving the
system) 5 times, and that each time there is a change the probability is 2/3 that
174 Chapter 7: Stochastic Processes

someone is entering the system and 1/3 that someone is leaving. Then,
5+3
P(Y5 = 3) = P(X = )= P(X; = 4)
2
4

= C(5, 4) p*q =5 p4q = 5 (2) G % a

The probability that there will be exactly 13 users after there have been exactly 5
changes in the number of users on the system is 80/243.
We should pause to acknowledge that, as is often the case with mathematical
models, this model requires that certain restrictions be placed on the inputs if we
want to represent a realistic situation. The restriction is that a negative number of
users of the computer network is meaningless. This means that the computation
performed actually gives the correct probabilities for the net gain or loss to the
system only as long as the number of users is always positive. For example, if the
system had 4 users to start with, then the fact that P(Y; = 3) = 80/243 does not
mean that after 5 changes in the number of users the probability is 80/243 that there
will be 8 users. The reason is that if the first four “occurrences” are people leaving
the system, at that point the number of users has dropped to 0. At that point the
assumptions we used in the model become invalid because it is impossible for
anyone else to leave the system. Many mathematical models, whether they are
deterministic (as in differential equations models) or probabilistic, are only
approximations of real phenomena and are valid for only a restricted range of
inputs. The model of this example is accurate as long as the number of users is
larger than the number of people entering and leaving.

7.3 Poisson Processes

While the Poisson and exponential distributions have been introduced earlier,
it is in the context of modeling phenomena via the concept of a Poisson process that
the intricate relationship between Poisson and exponential random variables is most
visible.
At an intuitive level, the essence of a Poisson process is the idea of “random
phenomena” occurring intermittently in accordance with certain descriptive
assumptions. Real phenomena that might be modeled via a Poisson process are
things such as the following: (1) calls arriving at a telephone switchboard, (2)
breakdowns of a piece of equipment, (3) traffic entering a parking lot, (4)
7.3 Poisson Processes 75

emissions of alpha particles from a quantity of radioactive substance, and (5) traffic
accidents in a city. Some of these examples are idealized. For example, if a traffic
light affects the flow of traffic into a parking lot, the movement of cars into the lot
will not satisfy the axioms of a Poisson process.
A Poisson process involves a family {X,:t>0} of random variables. The
basic idea is that for a given time t, X, counts the number of occurrences during
the time interval [0, ¢] of whatever phenomena is being modeled. So time is
measured on a continuous scale relative to some fixed “starting time” referred to as
time t= 0. For example, if we are observing cars entering a parking lot and if time
is being measured in minutes, then X;9 would be the number of cars that enter the
parking lot during the first 10 minutes. Notice that X, — X,;, would represent the
number of occurrences of the phenomena between times ft, and ty.
Let’s look at the precise requirements that a Poisson process must satisfy and
try to understand intuitively what each of the conditions is all about. The
requirements are given in Definition 7.1.

Definition 7.1: A Poisson process is a family {X,:t20} of non-negative

integer-valued random variables related in the following ways:
1. Xo = 0, and, if t) St.-<Stz St, then X,, aoe y. and X,, ao Ss are
independent random variables.
2. There is aconstant u > 0 such that for any ¢ > 0,

1 = PX ar= Xp) i
(a) At>0 At =)

(b) P(X t+At Ne)

t teens)
At>0 A t

PQ&par— X= 1)
(c) At>0 At
=

Property 1 simply says that we start counting at time ¢ = 0 and that the
number of occurrences in two disjoint time intervals should be independent of each
176 Chapter 7: Stochastic Processes

other. This is clearly intuitively plausible for many situations. If we are modeling
traffic accidents, the number of accidents between 10 o’clock and 11 o’clock would
have no apparent reason to affect the number of accidents between 1 o’clock and 2
o’clock.
Property 2(c) says that for small time intervals, the probability of exactly one
occurrence of the phenomena during the time interval is approximately proportional
to the length of the time interval. The constant p is the constant of proportionality.
Property 2(b) says that for short time intervals the likelihood of more than one
occurrence is negligible, that is, negligible in comparison to the length of the time
interval.
Notice that the expression 1 — P(X;+at = X;) in the numerator of Property
2(a) is 1 minus the probability that no occurrence has taken place during the time
interval from ¢ttot+At. But since 1—P(E) = P(E°) for any event E, this
is simply the probability of at least one occurrence between time ¢ and time ft + Af.
So Property 2(a) says that for small time intervals the probability of at least one
occurrence is approximately proportional to the length of the time interval, with u
being the constant of proportionality. Property 2(a) is a logical consequence of
Properties 2(b) and 2(c), but it is listed separately for reference later.

Investigation of the Poisson Process

If {X,: t 20} is a Poisson process as described by the above properties, it

is interesting to investigate the properties of the function f,(r) defined for all t > 0
by
folt) = P(X, = 0)
that is, f(t) is the probability that no occurrences have yet taken place at time t.
Then,

fot
+ At) = P(X t+At =0) = P(X,=0,X t+At =0) = P(X, =0) P(X,,,,t+At =0)

Being able to write this last product depends on Property 1; that is, the random
variables X, and X;4a; — X; are independent. But now

folt+ At) —fol) — P(X;=0) PX,,a,=X,) — P(X, =0)

At At
7.3 Poisson Processes 177

PX = ae) Sl
= P{(X,= laa > -ufp(t) as At > 0
t
[This uses Property 2(a).] However, the value of this limit is, by definition, the
derivative fp'(t), and so we have shown that fy'(t) = -1 fo(t).
This simple differential equation is easily solved to give fo(t) =Ae™,
where A is constant. If we take into account now that f9(0) = P(X9 = 0) = 1
(because no occurrences have occurred yet at time ¢ = 0), this tells us that A = 1,
and so

fo(t) = P(X, =0) =e" (7.1)

Now we need to consider the probabilities of X, assuming positive values.
For each positive integer k, let’s denote by f, the function

Ft) = PX, =k)

that is, f,(¢) is the probability that exactly k occurrences have taken place by time
t. We will need to make use of the following decomposition for f,(t+An):

fift+At) = PXia1 =k)

= P(X,=k,
Xi44, -X,=0) + P(X, =k-1, Xp44,-X, = YD

Py 6<a3a Nibdiay der g Miranda! g 10 Gh dinesOMY

ori 9
= P(X,=k) PX; —X,=0) + PX, =k- YD PAXua
—X; = 1)
IPC uke 2) Cue
km Dct ie deh ROG SOP, ees SO
felt + Ad) — fit)
Therefore —————— is given by
At

1
ee { P(X, =k) P(X pyar —Xz = O)+ P(X, =k - 1) P(X pyar —X; = 1D)

+--+ +P(X, =0) P(Ki4q;-X; =k) — P(X,

= *) }

PO oe On 1 PU =k DP Oa XD)
Sept )) _ ooo + ——————
ed

: At At

+ other terms
178 Chapter 7: Stochastic Processes

where the sum of the other terms is dominated in absolute value by

P(X. ar- X,> 1)
At
which converges to 0 as At > 0, by Property 2(b).
Taking the limit as At > 0 gives us the information that

KO = PAO + UA AO (7.2)
Since we already know what fo is, we can use Equation 7.2 with k = 1 to
find f,;. The function f; must satisfy the differential equation

AiO =vf\O+ne
This differential equation is easily solved to give f,(t) = (ut + c)e’, where c
is a constant. However, f;(0) = P(X9 = 1) = 0, since there have been no
occurrences yet at time ¢=0, and this tells us that the constant c = 0, which
allows us to conclude that

fi = pte™
Now, using Equation 7.2 again, it is possible to compute f/>, since we now
know f;. Equation 7.2 says that

fy) = BAW +UAO =A + u{ute™)}

This differential equation, together with the initial condition f,(0) = P(X, =0) =
0, can be easily solved as a first-order linear differential equation to obtain

y(t) = 5beers
The method that has been used to find f; and f, from Equation 7.2 can be
continued, since each function f, is recursively defined in terms of f;_; by
Equation 7.2. It is not difficult to show via mathematical induction that for any
positive integer k,
1
fi) = P&p=k) = FF (uyte™ (7.3)
Recall now from Chapter 4 the definition of the Poisson distribution. A
random variable X is a Poisson random variable with parameter 4 provided
1
P(X = ky = e* a
The computation we have just finished has demonstrated Proposition 7.2.
7.3 Poisson Processes 179

Proposition 7.2: If {X,:t 20} is a Poisson process as in Definition 7.1, then

for each positive number f, X, is a Poisson random variable with parameter
N= ut

The constant t in Definition 7.1 is referred to as the average intensity or

simply the intensity of the process. The reason is that ut represents the average
number of occurrences of the phenomena per unit of time. To see this simply
remember that for a Poisson random variable X, E(X) =i, where A is the
parameter in the distribution. (See Problem 4.9.) Therefore, since X, has a
Poisson distribution with parameter A = wt, it follows that E(X,) = ws; that is,
the expected number of occurrences of the phenomena being studied during the time
interval [0, ¢] is uz, the intensity of the process times the length of the time
interval.

Example 7.2. Suppose that the calls coming into a switchboard constitute a
Poisson process with intensity u = 4 calls per minute. Then for any value of f, the
random variable X,, which counts the number of calls that have arrived by time f,
has the Poisson distribution with parameter A = 4t. So if we wish to measure the
number of calls coming into the switchboard during a 10-minute time period, we
can think of this as a Poisson process in which X19 is the random variable of
interest, and it will be Poisson with the parameter’ = 4 x 10 = 40. Since the
probability distribution is now explicitly known, any probabilities of interest can be
computed. For example, the probability that 30 or fewer calls come in during a 10-
minute time interval would be
30
-+0 40°
k=0

Connection between the Poisson Process and Exponential Distribution

As we have just seen, if we count the number of occurrences of the

phenomena being observed in a Poisson process during a fixed time interval, the
random variable that does the counting has a Poisson distribution. Instead of
counting the number of occurrences during a fixed time interval, what if we
180 Chapter 7: Stochastic Processes

consider the waiting time for the first occurrence in a Poisson process? Let Y; be
this random variable; that is, Y,; gives the length of time from the starting time
(time t = 0) to the time of the first occurrence. In Example 7.2, Y; would be the
length of time that passes during the observed time period before the first call comes
in to the switchboard.
The key relation between Y, and the Poisson process {X,:t = 0} is that
P(Y, >t) = P(X; = 0). This is because (Y; >t) and (X, = 0) represent
precisely the same event; to say that the waiting time for the first occurrence is
greater than ftis the same as saying that at time ¢ there have been no occurrences.
However, Equation 7.1 gives P(X, =0) ase 4’. Therefore P(Y, >?) =
e! and P(Y, <t)=1-—e 4". We can differentiate to find that for t > 0, the
density function of Y, is the exponential density fr, @ = le for f >"0 30
Y,, the waiting time for the first occurrence, is exponential with parameter [L.
It might be a good idea at this time to review the analogy between an
independent trials process and a Poisson process. The binomial distribution arises
when we fix the number of trials in an independent trials process and count the
number of successes. The Poisson distribution arises when we fix the time interval
in a Poisson process and count the number of occurrences of the observed
phenomena. The geometric distribution arises in an independent trials process
when we consider “waiting time for first success” in terms of the number of trials
required. The exponential distribution arises when we consider “waiting time for
first occurrence” in a Poisson process in which time is measured on a continuous
scale. Both of these “waiting time” phenomena, described by geometric or
exponential random variables, have the lack-of-memory property discussed in
Chapters 4 and 5.
Proposition 7.3 summarizes the main features of a Poisson process. Property
3 says, for example, that if we let Y, denote the length of time required for two
occurrences (in other words the total waiting time from time t = 0 until the second
occurrence), then Y,—Y, has the same distribution as Y,;. What is of perhaps
more interest is that Y, — Y, is independent from Y,. All this is summarized by
saying that waiting time between the first and second occurrence is independent of
the time of the first occurrence and has the same probability distribution as the
waiting time for the first occurrence. This is a consequence of the lack-of-memory
property. In fact it can be shown that the entire Poisson process has a certain lack-
of-memory feature in that the choice of starting time is irrelevant. In other words,
for any to >0, the random variables X, and X,.,; — X,, have the same
7.3 Poisson Processes 181

distribution. The former counts occurrences during the time interval {0, ¢] and the
latter during the interval [tp, t9 + t]. The distributions turn out to be the same
because the intervals are of the same length. A stochastic process having this
property is said to have stationary increments. There are other senses in which
some stochastic processes are unaffected by time shifts. Stationary stochastic
processes are introduced and discussed in Section 7.7.

Proposition 7.3 (Properties of Poisson Processes): Suppose some kind of

random phenomena occur intermittently. For each positive integer t, let X,denote
the number of occurrences during the time interval [0, t]. Also, for each positive
integer n, let Y,, denote the time at which the nth occurrence takes place.
If {X,} is a Poisson process with parameter L, then all the following are true:

1. For every tg > 0, the random variable Xt)+t - Xt, (which gives the
number of occurrences during the time interval from fp to fg + ¢) is a
Poisson random variable with parameter A = Ut.

2. The number of occurrences during disjoint time intervals constitute

independent Poisson distributed random variables.

3. The random variables, Y;, Y,2-Y,, Y3-—Y>,--- are all exponentially

distributed with parameter . This says that the waiting time for the first
occurrence, the waiting time between the first and the second occurrence,
between the second and the third, and so on, are all exponential with
parameter u. Furthermore, these random variables are mutually
independent.

4. The constant Lt can be interpreted as the average number of occurrences per

unit time.

5. If there is exactly one occurrence between times ¢, and f, then the time of
that occurrence is uniformly distributed on the interval [, fq].
182 Chapter 7: Stochastic Processes

7.4 Poisson or Binomial?

Example 7.3. Suppose a snowstorm has begun and a grid such as the one
shown in Figure 7.1 has been laid out on the ground. The “intensity” of the
snowfall is 3 flakes per square inch per minute. The experiment we are going to do
is to count the number of snowflakes hitting the one-square-inch shaded section of
the grid during a specified 1-minute time interval. Let’s call this random variable
X. What kind of model is appropriate here?

Solution: The first thing to consider is that there are at least two different
interpretations to the statement that “snowflakes are falling at a rate of three
snowflakes per square inch per minute.”

One-square-inch section

Grid is 10 inches square

Figure 7.1 Snowflakes are falling on this grid at a

rate of three snowflakes per square inch per minute.

Model 1: Suppose we interpret this as meaning that 300 snowflakes fall on

the entire 100-square-inch grid during each minute. We can think then of each
snowflake as an independent trial that may or may not fall on the one square inch
section where we are doing the count. All sections being the same size, we have p
= .01 as the probability that a given snowflake hits the shaded square. From this
point of view, the number of snowflakes hitting the shaded square during a 1-
minute time interval is binomial with n = 300 and p = .01. The expected value is
np = 3, which is, of course, to be expected (no pun intended) since the average is
7.4 Poisson or Binomial? 183

3 snowflakes per square inch per minute.

Let’s do a simple computation using this model. What is the probability that
exactly two snowflakes hit the shaded square during the observed minute?
P(X = 2) = C(300, 2) p? q298 = C(300, 2) .012 .99298 = .2244

Model 2: Now let’s model this situation as a Poisson process. Since the
intensity of the snowfall is three snowflakes per square inch per minute, this means
that X, the number of snowflakes to hit the shaded square in a 1-minute interval,
will now be considered to have a Poisson distribution with parameter A = intensity
x length of time interval = 3 x 1 = 3. Since X is Poisson now, E(X) =A, by
Problem 4.9, so again we have E(X) = 3, just as it should be.
In this model, what is the probability that exactly two snowflakes hit the
shaded square during the 1-minute interval?

P(X =2) = e* — = e?x < = .2240

Isn’t it astounding that two apparently very different models should give such
nearly identical answers for this calculation? The reason this happens can be
explained mathematically by saying that if 4 is fixed, then the binomial distribution
with parameters n and p =A/n converges to the Poisson distribution with
parameter A asn— co. A less precise but intuitively more pleasant way of
describing this is to say that if n is “large” and p is “small,” then the binomial
distribution with parameters n and p is approximately the same as the Poisson
distribution with parameter A = np.
But which model is the correct one? A little further insight is gained by
considering one more scenario. Suppose that the 10 by 10 grid is replaced by a 100
by 100 grid and that the binomial model is used. Now n = 30,000 and p = .0001.
However, np = 3 still. In this model,
P(X = 2) = C(30000, 2) .00012 .999829998 = .2240
This calculation is consistent with the Poisson model to at least four significant
digits. Furthermore, it sheds additional light on the question as to which model is
correct. If one interprets the average rate of 3 snowflakes per square inch per
minute as being an exact rate measured over a finite grid, then the correct model is a
binomial model based on the idea of an independent trials process. If, however,
one imagines an infinite plane on which snow is falling at a rate of 3 snowflakes per
square inch per minute, then the correct model is the Poisson model based on the
184 Chapter 7: Stochastic Processes

idea of a Poisson process. (Of course, the very idea of snowflakes falling at a rate
of 3 per square inch per minute over an infinite plane inherently involves the limit
concept.)

7.5 Sample Functions

The mathematical definition of a stochastic process is precise enough. It is
useful also, however, to know a few productive ways to think about them. This is
especially true when the parameter f in the process {X,} varies over a continuum
of real numbers rather than a discrete set. Let’s illustrate with an example.

Example 7.4. Consider a stochastic process of the form

X,=A cos (Wt + 8)
where @ is constant, but where A and @ are random variables. Suppose for
definiteness that A is uniformly distributed on the interval [0, 1] and @ is uniformly
distributed on the interval [0, 27]. Since w is a constant, the frequency of the wave
is absolutely determined. The amplitude and phase angle are probabilistic,
however, and will depend on the values of the random variables A and 0. Each
pair of values for A and @ (0 <A <1 and 0 <@ < 2m) leads to a unique curve.
For example, the special case of A = .3 and 8 = 2.6 would give one possible
observation of this process. Common terminology is to call each of these possible
functions a sample function. In a sense you can think of the sample functions as
the possible outcomes of the process being observed. This is an extension of the
way in which we used the term possible outcomes in Chapter 1 to refer to the
possible results of a simple experiment with a finite sample space.

It is also worthwhile to compare the way one looks at a stochastic process

{X,} to the way one looks at values of a single random variable. Suppose Y is a
single random variable defined on a sample space S. Then for every possible
outcome se S, Y(s) is simply a real number. Now in the case of a stochastic
process {X,}, each of the random variables X, is defined on a common sample
space S, so for each se § and each ¢, X,(s) is a real number, the observed
value of X,, Often it is useful to think of s as fixed and to consider how X,(s)
looks as a function of t. For each fixed se S, the function t > X,(s) is simply
a real-valued function of a real variable.
7.5 Sample Functions 185

To make this idea still more concrete, let’s return to the example
X,=A cos (Wt+ 0)
Here @ is a fixed constant. We’ll pretend that the values of the random variables A
and 8 are produced independently by random-number generators so that0 <A < 1
and 0 <@<2n. The “possible outcomes” of the experiment then correspond to
pairs of random numbers produced, that is, the values of A and @. For each such
“possible outcome,” there results a sample function. The sample function
Jd) = .2 cos (@t + 5)
being the result, for instance, if the values produced by the random-number
generator for A were .2 and for 6 were 5.
The stochastic process X, =A cos (Wt + @) is somewhat specialized in the
sense that every random variable X, here is described in terms of the random
variables A and @. Generally the random variables that make up a stochastic
process do not have such a simple relationship connecting them with each other.
Even so, the concept of a “sample function” remains that of fixing an outcome
se S in the underlying sample space and considering how X,(s) varies as a
function of ¢.
As a further illustration, suppose we are monitoring a Poisson process such
as calls coming into a telephone switchboard. One “possible outcome” can be
partially described by saying that the first call arrives at time ¢t, = 1.3 minutes, the
second at time ft, = 3.2 minutes, and the third at time ¢; = 3.8 minutes. In this
case, a portion of the graph of the sample function that corresponds to such an
“observation” of the process would be as shown in Figure 7.2. This is a graph of
X, corresponding to the observation just partially described, where X, is the total
number of calls to have arrived by time t.

1 2 3 4
Figure 7.2 Typical sample function for a Poisson process.
186 Chapter 7: Stochastic Processes

There are some very subtle points related to the study of stochastic processes
that require sophisticated and advanced mathematics. This can be illustrated, for
example, even with regard to an independent trials process. How does one think of
the sample space for an infinite sequence of coin tosses? A finite sequence of
tosses is no problem, but when one envisions the stochastic process {X n}» where
X,, is the number of heads to occur in the first n tosses of an infinite sequence of
tosses, things get a bit confusing. While it’s clear what the probability distribution
of X,, is going to be, it’s not so clear just how one can envision a sample space
(with probability measure) on which all of the X,,’s are defined. Fortunately, one
can do a lot of worthwhile modeling without dealing with such riddles, but it is
worth realizing that the complexities required to do a rigorous mathematical
treatment are substantial.

7.6 The Autocorrelation Function

For certain types of stochastic processes, the autocorrelation function is an

important concept. Applications of this idea will be made in the next section.

Definition 7.2: Given a stochastic process {X,}, the autocorrelation function

for the process is the function of two variables Ry defined by

Ryx(t, to) = E(X, X;,)

The subscript X in the expression Ry is, of course, the name of the

stochastic process and is used simply to identify the stochastic process whose
autocorrelation function is being discussed.
There are various situations in which the autocorrelation function for a
random process is a useful tool. One which we will mention later has to do with
what is called a “stationary” random process. Crudely speaking, a stationary
process is one for which the choice of time to be considered as “time t = 0” is
immaterial. One of the senses in which the term stationary is used involves the
autocorrelation function for the process, and this idea will be featured in the next
section. Meanwhile, we will analyze a few concrete examples.
7.6 The Autocorrelation Function 187

Example 7.5. Consider the sine wave stochastic process

X,=Ycoso@t
where @ is a constant and Y is a random variable which is uniform on the interval
[O, 1]. This can be viewed as a sine wave in which the frequency and phase angle
are fixed but in which the amplitude is given only in probabilistic terms. Each
sample function will be simply a constant multiple (between 0 and 1) of the cosine
function cos wt. What are the expected value and the autocorrelation function of
this stochastic process?

Solution:

E(X,) =E(Y cos wt) = E(Y) cos Mt = = cos wt

2
Ry(ty, t2) = E(X,, X,,) = E(Y? cos wt, cos ty)
1
= COS Wr, 1 COS Wl, 24 E(Y2) = =e: cos Wt, 1 cos Wt 2

[That E(Y2) = 1/3 was Problem 5.6.]

Example 7.6. Now let’s consider a sine-wave process in which the

amplitude and frequency are fixed but the phase angle is known only in a
probabilistic sense. Specifically, we will consider X,=A cos (@t + ®), where
A and @ are constants with A > 0 and where @ is a random variable uniform on
the interval [—1, 7]. What will the expected value and autocorrelation function of
this stochastic process be?
Notice that in this example all the sample functions are simply translates of
each other; that is, they all have the same amplitude and frequency and vary only in
phase angle.

Solution: Each of the random variables X, is now a function of the random

variable @. Specifically, X, = g(6) where g is the function of one variable g(s)
= A cos (@t+5). This mean that we can use Proposition 3.7 and determine the
expected value of X, by using the density function for 0:
TT
T
1 Aliy,
E(X,) = i A cos (t+ S) aa ds = =—
on sin (@t+ o|
Tt n —t
188 Chapter 7: Stochastic Processes

ere (—sin @t + sin wt) = 0

rege

Now for the autocorrelation function:

Ry(t,, ta) E[ A cos (wt, + s) A cos (@t, + S) ]

us
1
A2 cos (Wt, + S) cos (Mt + 5) a ds
—1

Aig ea
sa z [ cos (Wt, + Mt, + 2s) + Cos (Mt, — Wry) ] ds
—

This last expression comes from the trigonometric identity

2 cosx cosy = cos (x + y) + cos (x-y)

The last integral is easily evaluated. Notice that one of the cosine terms does
not contain s at all and therefore may be treated as a constant as far as integration
with respect to s is concerned. The other term has a corresponding sine term as its
antiderivative with respect to s, and since the endpoints of integration are —m and
mt, the periodicity of the sine function makes that term integrate to 0. Thus, the
answer is simply
AZ

Ry(ty, t2) = > COS (Wt, — @t2)

7.7 Stationary Stochastic Processes

Keep in mind that in most common models, the parameter f in a stochastic

process {X,:t 20} represents time as measured on some discrete or continuous
scale. So time t = 0 represents the “starting time” for the process. In many
contexts, the starting time really is immaterial. For example, if you walk into a
room where someone is tossing a coin, the probability that a head occurs on the
next toss does not depend on how long the process has been going on before your
arrival. It doesn’t matter at what time you consider the process to have begun.
In a Poisson process, the random variables X, and Xt +t - Xt, have the
same distribution for any value of fp. In terms of the phenomena being observed,
this says that the number of occurrences during a time interval depends only on the
length of the time interval and not on when the time interval begins. An even
stronger condition is described in Definition 7.3.
7.7 Stationary Stochastic Processes 189

Definition 7.3: A stochastic process {X, } is said to be a stationary stochastic

process provided that for any values ¢,, t),--+-,¢,, and any ¢t > 0, the joint
probability distribution of the random variables X,> X,,,+++,X, is the same as
it would be if the subscripts 1), t2,---,t, were replaced by t) +t, t) +t,---,
t, +t. In other words, if all of the times were shifted by the same amount, this
joint distribution of the random variables would remain the same.

Definition 7.3 treats a manner in which time shifts are irrelevant in some
stochastic processes. This definition says, for instance, that the probability
distribution for X; is the same for every value of t. However, it says much more
than this. Remember that the joint distribution contains all the information as to
how the random variables interact with each other. The definition requires that all
the details of this interaction be preserved if there is a time shift.
We can illustrate this definition by considering exactly what is and what is not
stationary with regard to an independent trials process.
When one talks about an “independent trials process,” there are really two
stochastic processes lurking in the background. One is the process {X,,}, where
X,, is the number of successes to occur during the first n trials. For each n, X,,
is binomial with parameters n and p, where p is the success probability
associated with the process. Clearly, if n #m then X, and X,, have different
distributions, and so this stochastic process isn’t stationary.
Rather than looking at the accumulated number of successes during the first
n trials (which is what X,, tells us), we could consider the process {Y,,}, where
Y,, is 1 or 0 depending upon whether success does or does not (respectively) occur
on the nth trial. In other words, Y,, is the number of successes (0 or 1) that occur
on the nth trial. This stochastic process {Y,,} is stationary. This is an easy
consequence of two facts: One is that P(Y, =1)=p and P(Y,=0) = 1-p
for every n (so the Y,,’s all have the same probability distribution), and the other
is that Y,, Y,,--- are independent. The fact that they are independent means that
the joint probability mass function of any finite collection of them is just the product
of their individual mass functions which are all the same. The stochastic process
{Y,,} then clearly meets the criteria of Definition 7.3.
Standard terminology is to refer to {X,,} as the binomial process and {Y,,}
as the Bernoulli process. Either can be defined in terms of the other quite easily:
190 Chapter 7: Stochastic Processes

Y, =X,, and forn >1,Y, =X, —X,,7, and

In practice it is usually very difficult to show that a given stochastic process

satisfies the conditions of Definition 7.3. The reason is that the joint distributions
referred to may be very difficult or impossible to compute.
Stationary stochastic process have properties that are very useful in a variety
of areas of application, such as signal processing. While Definition 7.3 is very
difficult to check in practice, there is a weaker condition that is much easier to work
with and is often used instead. It is introduced in Definition 7.4.

Definition 7.4: A stochastic process {X,} is said to be a stationary in the wide

sense provided that
1. the expected value E(X,) remains the same for all values of the parameter
t, and
2. the value of the autocorrelation function, Ry(t,, t2), remains the same if
t, and f, are both shifted by the same amount, that is,

Ry(tyt2) = Ry(t, +t, t,+1) for every time shift r

Notice that the computation shown in Example 7.6 was precisely the
computation to show that the stochastic process X,=A cos (ot + 6), where A
and @ are constant and @ is uniform on [—z, 1], is stationary in the wide sense.
Stochastic processes that satisfy Definition 7.3 are often called strictly stationary to
contrast them with the wide-sense stationary processes of Definition 7.4. It is easy
to see that any strictly stationary stochastic process is also wide-sense stationary,
but the converse isn’t true. (See Problem 7.16.)

Proposition 7.4: Ifa stochastic process {X;,} is strictly stationary, then it is

also stationary in the wide sense.
7.7 Stationary Stochastic Processes 191

Comment on Proof: From Definition 7.3, any two random variables Xt,
and X;, in the process must have the same probability distribution. It follows then
that they must have the same expected value.
But why must Rx(t,,t.) = Rx(t, + t,t, +t)? We will illustrate in the
case in which X,, and X,, are continuous random variables with a joint density
function. This will enable us to view the computation in terms of Proposition 6.3.
Since Ry(t;, ty) = E(X,, X,,), it is first necessary to recall how this
expected value can be computed in terms of the joint density function for X,, and
X,,. The means for doing this is provided by Proposition 6.3:

BK,%,) =| [wifey dedy (7.4)

where fis the joint density function for X,, and X tp:
The significant point now is that X,,,, and X,,, have exactly the same
joint density function as do X,, and X ty: (This is what Definition 7.3 says.) So
the expected value E(X,,,;X;,4,) would be described by exactly the same
integral as the one that appears in Equation 7.4 above, and it is this that defines
Rx(t, +t, 2 + 2).

Example 7.7 (Random telegraph signal). Consider binary transmission of

data sent along an electrical line. The two possibilities we will analyze are whether
the voltage of the line is zero or nonzero relative to a ground voltage. Let’s
represent zero voltage by the value 0 and nonzero voltage by 1. At any time ft, we
can then think of the random variable X, as indicating the state of the line at that
time; that is, X, = 0 if the voltage is zero and X, = 1 if the voltage is nonzero at
time ¢.
Furthermore, we will assume that there is some random “background
phenomena” occurring which causes the line to toggle from one state to the other.
An example could be something like cars entering a parking lot. As each car enters,
it trips a switch which in turn sends a signal by changing the state of the
transmission line being monitored. The situation we will consider is the case in
which the background phenomena constitutes a Poisson process. So in the parking
lot example, this means that for any given time interval the number of cars entering
during the interval is a Poisson random variable with parameter 1 = Ut, where ¢
is the length of the time interval and p is the average number of cars entering the lot
per unit time. Specifically, this leads to the following assumptions about the
192 Chapter 7: Stochastic Processes

stochastic process {X;}:

1. For any time t, P(X, = 0) = .5 and P(X, = 1) =.5.

2. During any time interval of length t, the probability that the line switches
from the state X, = 0 to the state X, = 1 or vice versa exactly k times is
given by .

was
k

The constant Lt in condition 2 is the “intensity” of the underlying Poisson

process, that is, the rate of the background phenomena that is causing the line to
toggle back and forth between its two states. Condition 1 is a symmetry condition
on the two possible states of the line. If the two voltage states are assumed equally
likely to start with, the symmetry of the process suggests that the two states will
always be equally likely.
Let’s try to show that this “telegraphic” process {X,} is stationary in the
wide sense.
First, for any 1,
EX, = 0x P&,=0)+1xP@&,=1) = 3
according to assumption (1).
Next we have to determine what the autocorrelation function Ry(tj, ty) is
for the process. Since X,X 1, can only assume values 0 or 1,

E(X, X,,) = 0xP(%,X,,=0) + 1x PX, X,=))

ll P(X,= 1,X,=)
The reason for the last step is simply that the only way that the product of X,, and
X,, can equal 1 is for each of them to be 1.
For definiteness, let’s suppose that t,; < f,. The best way then to determine
the probability P(X,,= 1, X;,= 1) is to use the idea of conditional probability:
P(B OA) =P(A) P(B|\A). This gives

POS 2 Age) S Payer regetixpeD

Think about the conditional probability on the right side of this equation. If
we know that X,,= 1, then the only way that X,, can equal 1 is for the line to
toggle an even number of times between times ¢, and fj. Since assumption 2 of
7.8 Ergodic Properties 193

this example gives the probability for a given number of changes in the state of the
line during a given time interval, we can write the probability of an even number of
changes during a time interval of length fas

2 4 6

Since 10. 1) =.5 by assumption 1, we now know that

1 u Cee et ne Pe)
Ry(t1,t) = E(X,X,) = 7X @ ae + — sp ae + aaa +

where t = fy — f, is the length of the time interval between times ¢, and fp.
The remaining question is how to evaluate the sum of this infinite series on
the right. This is not difficult if we make use of the standard power series
representation for the exponential function. The sum of this series is
1
> (e + eH)

(See Problem 7.17.) This gives us

1 i s 1 =
ee 1 os
pes les cee u(t, ) )
Ryx(ty, ty) = ze Wee 5 (ey
=| S|

Notice that we have shown that this stochastic process is stationary in the
wide-sense. This is a consequence of the fact that Ry(t,,t.) depends here only
on the difference t. —t,. Consequently, if t; and t, were translated by the same
amount, there would be no change in Ry(fj, fy).

7.8 Ergodic Properties

Example 7.8. Let’s return once more to the simple illustrative stochastic
process given by X,=A cos (mt +6), where A and @ are assumed to be
constants and @ is assumed to be a uniformly distributed random variable on the
interval [-1, x]. From Example 7.6 we already know that E(X,) = 0 for all ¢.
Let’s now take a look at
1 T
zh, X, dt

This is the time average of the process X, for the time interval [0, 7]. The value
194 Chapter 7: Stochastic Processes

will depend on what “observation” of the process we are looking at. In other
words, since X,=A cos (wt +6), the value we get for the time average will
depend on what value between —1 and m we use for the random variable 0. If we
carry out this averaging treating 0 for the time being as just an unspecified real
number, we get

Lf oar Seances
Ti
Tap ea dt = ell
T
cos (wt+8) dt

1 oT+0
= ari, Acosz dz (by achange of variables z= wt + @)

A : oT+0
or Sinz 1,

A ; :
or [ sin (oT + 8) — sin
6]

Observe that the time average obtained here does depend on the value that the
random variable 6 assumes. However, if we take the limit of this as T >0°, we
get 0, which is the expected value of X, This is true no matter what value @
assumes. So in fact we have
Be es Bae
E(X,) = ah X, dt (7.5)

A stochastic process that satisfies this condition is said to be ergodic in the mean.
Notice that in order for this condition to be met, a few curious things have to
be happening. Normally one would expect E(X,) to be different for different
values of t. However, the right side of this equation is not dependent on f¢ since it
is a time average over the whole positive real line. Thus E(X,) must necessarily
be the same for all values of t when Equation 7.5 is valid. Secondly, the right side
would normally be expected to be dependent upon what “observation” of the
process we are averaging, that is, what sample function. In this simple example,
this means that we would normally expect the right size of Equation 7.5 to be
different for different values assumed by the random variable 6. What we saw in
the above computation is that this is, in fact, true for any finite interval. But when
we average over the entire positive real line, such differences wash out.
In theory, the two kinds of averaging being performed on the two sides of
Equation.7.5 are very different. The left side averages the value of a fixed random
variable X, over the sample space, and the right side averages all X,’s over time
Problems 195

but corresponding to a fixed “outcome” in the sample space. There is an intuitive

idea lurking behind all this. It is that over a long period of time the “average”
behavior of the process is quite predictable.
Certainly not all stochastic processes have this ergodic property. Moreover,
there are other forms of ergodicity that are of similar interest, such as ergodicity of
the variance or of the autocorrelation function, for example. One characteristic that
some models possess that leads to ergodic properties is that the behavior of the
process during a given time interval may be “almost” independent of the behavior
of the process during a distantly removed time interval. Assumptions of ergodicity
frequently come into play when constructing models for real-world phenomena in
which limited data is available. In modeling climatic phenomena, for example, one
might have only one 50-year record of observations for a particular location. A
reasonable assumption could be, however, that long-term cause and effect is
negligible and therefore some future 50-year record may have similar statistics but
be approximately independent from the existing record.

Problems

7.1. Use the Poisson distribution to approximate the probability of exactly 3

successes in 100 trials in an independent trials process with success
probability p = .04.

7.2 Assume incoming calls at a switchboard to be a Poisson process with

average intensity = .5 calls per minute. What is the probability that no
more than 2 calls come during a 2-minute time interval? What is the
probability that no more than 2 calls come during a 3-minute time interval?

7.3 For an independent trials process with success probability p, derive the
probability mass function for the random variable X that gives the number
of the trial on which the second success occurs.

7.4 Suppose X and Y are independent geometrically distributed random

variables, each with parameter p. Derive the probability mass function for
Z=X+Y. Can you relate this exercise to Problem 7.3?

7.5 Using Equations 7.1 and 7.3, derive the probability density function for the
196 Chapter 7: Stochastic Processes

random variable that measures the “waiting time until the second occurrence”
in a Poisson process.

TG Suppose X and Y are independent exponentially distributed random

variables, each with parameter A. Derive the density function for the random
variable Z = X + Y. Can you relate this to Problem 7.5?

Ce Assume that breakdowns of a particular device constitute a Poisson process

with an intensity of 10 breakdowns per year. What is the probability of
getting 7 or more breakdowns during a 6-month period?

7.8 Calls are coming into a switchboard at a rate of 30 calls per hour. Assume
this to be a Poisson process, and let X be the random variable that gives the
number of incoming calls between 1:30 and 1:40, and Y the random variable
that gives the number of incoming calls between 1:40 and 1:50. What kind
of probability distribution does X have? How about Y? How about their
sum X + Y? [Hint: Use the properties of Poisson processes described in
this chapter. It would be possible to compute the mass function from basic
principles based upon knowledge of the mass function of X and of Y and
the fact that they are independent. In a Poisson process, however, the
number of occurrences during any interval is a Poisson random variable
where the parameter in the distribution is determined by the length of the time
interval and the intensity of the process.].

tt Assume that automobile accidents in a city constitute a Poisson process with

intensity . = 5 accidents per day. Compute the probability of only one
accident occurring during a 1-day period, and compare this with the
probability of two accidents occurring during a 2-day period. Do you
intuitively understand why these probabilities are so different in magnitude?

7.10 Consider the sine-wave process X,=A cos (ot +6), where @ is a
constant and where A and 6 are independent random variables with A
uniform on [0, 1] and @ uniform on [0, 27]. Find the expected value of X,
as a function of t. (The function t— E(X,) can be viewed as the average
value of the process as a function of ¢.)

Ald Let Y,=A cos wt + B sin wt, where A and B both are uniform on
[0, 1] and are independent and where @ is a fixed constant. Find E(Y,).
Problems 197

Recall that every function of the form

g(t)=acosat+bsinwt
where a, b, and @ are constants, can be written in the form
g(t) =c cos (@t+ 8)
for the proper choice of c and 8. In fact,

and c is the amplitude of the wave.

For the stochastic process Y,=A cos ot + B sin wt, this means that
the amplitude is

Z=a) A? +B?
Notice that this is a random variable that does not depend on t. The
expected value of this random variable, E(Z), is then the expected
amplitude. Write down an integral that would give the value of E(Z).
(Don’t try to evaluate the integral.)

ThC2 Let X,=A+Bt, where A and B are independent and where both A
and B have mean 1 and variance 1. Find the expected value and the
variance of the stochastic process X,. In other words, express E(X,) and
var(X,) as functions of tf.

Ue Find E(X,) and var(X,), where X,=A cost and where A is standard
normal.

7.14 A quantity of radioactive substance is emitting alpha particles at a rate of 2

particles per microsecond. Assume that this constitutes a Poisson process.
(a) What is the probability that more than 4 particles will be emitted during a
fixed time interval of 2 Us?
(b) What is the probability that more than 6 particles will be emitted during a
fixed time interval of 3 Us?
(c) IfX is the number of particles emitted during a 1-LUs time interval, and
Y is the number emitted during a nonoverlapping 2-Us time interval,
what is the probability that X + Y will be greater than 6? Relate this
to part b.

dls Suppose that Y is a non-negative continuous random variable. (This means

that Y has a density function and that the density function is identically 0 on
198 Chapter 7: Stochastic Processes

the negative part of the real line.) Consider the stochastic process
X,=e" for r20
What is the relation between E(X,) and the Laplace transform?

Lelio Suppose Xj, X>, «++ is a stochastic process as follows:

(1) When k is even, X; assumes value 2 or =2, each with probability 1/2.
(2) When k is odd, X, assumes value 1 with probability 4/5 and value —4
with probability 1/5.
(3) The random variables X,, X>,--- are independent.
Show that this stochastic process is stationary in the wide sense but not in
the strict sense. [Hint:~It is easy to show that E(X,) = 0 for all & and that
Ry(G,k) =O if j #k and RyV,k) =4 if j =k. This is enough to
establish that the process is stationary in the wide sense. On the other hand,
for the process to be stationary would require that each of the random
variables X,, X>,---have the same distribution. (This is the specific case
that n = 1 in Definition 7.3.)]

Vee! Show that

2 4 6
ut (ut) (ut) (it) he Pe (eM + eH)
é (1 + SOY SP Abs ar 26E ata =
Nl]
This fact is used in Example 7.7.
[Hint: Let f(x) be the sum of the series
Xf alan
f(x) = Dap a
Differentiate the series term by term, and show that
f+ f@ =e
Solve this differential equation to show that f(x) = (1/2) (e* + e~).]

TAS Flow of traffic into a parking lot has been cited as an example of a Poisson
process. If the flow of traffic into the lot is regulated by a stop light, why
then will the entrance of cars into the lot not be a Poisson process? [Hint:
Which of the properties listed in Proposition 7.3 are not now satisfied?]
 Chapter 8: Time-Dependent
oystem Reliability

It is often assumed that “waiting time for a device to fail” is an exponentially

distributed random variable. Chapter 7 sheds light on this in the context of Poisson
processes. If the distributions of such “waiting time” random variables are
assumed to be known, then the reliability analysis of simple systems as performed
in Chapter 2 can be put in a time-dependent framework. In other words,
probability calculations can give not just a raw number as the absolute reliability of
a system; the reliability can instead be expressed as a function of time.

8.1 Reliability That Varies with Time

In Chapter 2 we talked about system reliability as a static concept. Each
component of a system is assumed to have a known probability of failure, and the
probability of system failure is then determined using these probabilities as inputs.
Clearly one can envision situations when the reliability of components varies with
time. New components may have a high failure rate because of “burn in” failures,
and old ones may simply wear out. Engineers like to frame this discussion in terms
of a function called the hazard function or hazard rate of acomponent. Let’s relate
this concept to.some things that we already know.
Suppose that for a particular device, the time (from some reference time
t=0) until the device fails is a continuous random variable X. So X has a
density function fy and a distribution function Fy. The reliability function for the
device is the function
R(t) = 1-Fy() = P(X > 2) (8.1)

This means that R(t) is simply the probability that the device is still functioning at
time ¢. Since Fy(t) > 1 as t > ©, it follows that R(t) > 0 as t > ©.
The hazard function h(2) for the device is defined then by

199
200 Chapter 8: Time-Dependent System Reliability

Ft) %, fy
h(t) = ce ee eet 2
1-Fy(t) = Rd) Cf

To see why this is a useful concept, notice first that the conditional

probability P(¢< X <1+ArlX > 2) is the probability that the device fails
between time f and time ¢ + At, given that the device is still functioning at time t.
The hazard rate h(t) is
P(tsX<t+AtlxX>2)
~ At
—>0+ At

The reason for this is that

A
im PESH SteDO
Fy(t + At) — Fy(2) Pre MeL
ft) = Fy) = tim SAEE At > 0+ At At > 0+ At

Thus,
PUSK
Sit AtlxX >t) PosxX=<1eA)
At At P(X
> 2)

< 1 PUSsX
sit At)
Pars!) At

fx)
——— asdAtr>0
PRE

Equation 8.2 describes the hazard rate h(t) in such a way that it is easily
computed if the density function of X is specified. It is also possible to turn the
situation around and describe the density function or distribution function in terms
of the hazard rate. The key relationship is
t
R(t) = 1-Fy(t) =exp {- j h(s) ds } (8.3)

where exp denotes the exponential function. The derivation of this expression is as
follows. If we integrate Equation 8.2, we have
t
fxs) ‘
I h(s) ds = 1 Fy) ds = -—In{l -Fy(o1 |= —In{1—Fy(0)]
8.1 Reliability That Varies with Time 201

The last step here assumes that Fy(0) = 0; that is, that at time t = 0 we
know the device is in working condition. (Remember that fy is the derivative of
Fy. That’s what the integration here is based on.) Applying the exponential
function to both sides of this equation gives Equation 8.3.
A common scenario with real-world devices is a hazard rate something like
what is shown in Figure 8.1. The curve is relatively high initially because of
potential failure during the break-in period, and the curve rises again as the device
ages. Clearly, not all devices behave this way, however. The most intuitively
pleasant and easily understood situation is that of a constant failure rate; that is,
h(t) is a constant function. In this case Equation 8.3 becomes simply
Ri) =F) =e"
which is the familiar case of the exponential distribution. This is important because
it is the canonical example for this concept.

Failures Failure rate

Burn-in period increasing because
ps High failure rate of age NI!

Time

Figure 8.1 Graph of a typical hazard function.

Summary: To say that the failure rate of a device as given by Equation 8.1
is constant is precisely the same as saying that the “time to failure” for the device is
exponentially distributed, that is, that the random variable that measures elapsed
time until the device fails has an exponential density function.
One widely used statistic that is an indicator of the reliability of a device is the
mean time to failure (mttf). Intuitively, this is just the average waiting time from
the time that the device is put into service until it first experiences failure. If X is
the random variable that gives the time at which first failure occurs, then the mean
202 Chapter 8: Time-Dependent System Reliability

time to failure for the device is simply E(X).

Since the reliability of a device is often specified in terms of the reliability
function, however, it is useful to be able to determine the mean time to failure in
terms of the reliability function. Proposition 8.1 indicates how that is done.

Proposition 8.1: Suppose a device has reliability function R(t). The mean
time to failure, or mttf, for the device is given by

mttf = I,R(t) dt

Proof: If X is the random variable that indicates time to failure for the
device, then by definition mttf = E(X). However,

[ro dt = fi —Fx() dt = fu - io ds] dt

co

ie [ [x asdt = [ fro dt ds = [sts ds = E(X)

Time-dependent system reliability studies can be performed based on these

concepts. Assumptions for the system might be, for example, that the various
components of the system function independently and that they each have a known
hazard function.
For a few examples, we will re-examine some systems analyzed in Chapter 2
with these ideas in mind. Figure 8.2 shows circuits that were discussed in Chapter
2. Now, however, we will assume that each component has a lifetime that is
known to be an exponentially distributed random variable; that means that each
component has a constant failure rate. For the components A, B, C, and D in
the figure, we will denote by p, Pz, P3, and p, the parameters in the exponential
distribution for the lifetime of the respective devices; that is, the time to failure for
device A is assumed to be an exponential random variable with parameter 1 = Py>
and so on. (A bit more insight into understanding why this constant p, might be
called the failure rate comes from remembering that if a random variable has
exponential density with parameter A, then the expected value is 1/A. Thus the
expected value for the lifetime of component A is 1/p,, and so “on the average”
8.1 Reliability That Varies with Time 203

that would correspond to p, failures per unit time. For example, if p, =2 and time
is measured in years, then the expected value for time until failure would be 1/2
year, which corresponds to a failure rate of two failures per year.)

Figure 8.2 Three circuits.

In the circuits of Figure 8.2, let’s denote by X, Y, Z, and W the times to

failure of the components A, B, C, and D, respectively. So the assumption is
that X, Y,Z, and W are exponential with parameters p,,p,,P3, and py,
respectively. Furthermore, we will assume that the lifetimes are independent; that
is, X, Y, Z, and W form a set of independent random variables.
Since it is the system rather than an individual component that we are now
considering, let’s denote by Rgys(t) the reliability function for the system. In
other words, R,,,(t) is the probability that the system is still functioning at time ¢.
For the series circuit in part a of the picture,
Rsys(t) SPX Sys OZ oe Hnara Spr Sor(Ze oO
= efi!
eoPato Ps! = eo Py +P, + P3)t

From the form of this expression, we know that the “waiting time for the circuit to
fail” is itself an exponential random variable with parameter) =p, +p, +p3. (The
reason is simply that this reliability expression is equivalent to the statement that the
distribution function for “time to failure for the circuit” is given by 1 — e where
XN=P, +P.
+ P3-)
For the parallel circuit in part b of the picture, at least one device must still be
functioning in order for current to flow, and so
204 Chapter 8: Time-Dependent System Reliability

R(t) = PL ASIC DUGe DI

== PASnyse 7s)
1=P OX spr Lsp Pres)
tee ee ee
For the series-parallel circuit in part c of the picture;
Ry) = PLUK > NAW aD eenUW>)T}
=P(X>nPY>H)l1-PZs)PW<S?d)
= ei oP [1 (1 —e C1 —e P)] (8.4)
What are the probability densities of the time to failure for each of the three
systems? For the series circuit shown in part a, the time to failure is exponentially
distributed with parameter A = p, +P, + P3. So we know from Problem 3.15 that
the expected time to failure (mean time to failure) is given by
1
fps es es
st oe +PERoy pe
For the parallel circuit shown in part b, the probability that the circuit will
have failed by time ¢is

1 —Reys(t) = (1 — €PH) —eP2'\(1 — Ps)

Since this is the distribution function for the random variable that represents waiting
time for system failure, we can differentiate to get the density function
=O Ff a =
fO® = pye?" + poe P + pze Ps
= t = t
a P Pre PiP2 _ P1P3e PiP3

= PoP3e Syyorys
PoP + P1P2P3e 4 PiP2P3 t

The mean time to failure is easy to compute here, because each of the exponential
terms is of the form Ae, and

} Ate™ dt = > for any A >0

0
This implies that the mean time to failure for the parallel system is given by

1 i 1 1 1 1 of i

Py Ps suis P1P. P1P3 PoP3 P)PoP3

The density function for waiting time to system failure for the system shown in part
8.2 Systems with Repair 205

c of the figure can be computed similarly. (See Problem 8.1.)

The computations taking place in these examples may appear to be quite a bit
more complicated than the time-independent problems treated in Chapter 2. Let’s
try to understand precisely what the difference is in the two situations. The
problems in Chapter 2 may be solved in a simple fashion because we can start
putting numbers into the calculations whenever we want to. In other words, there
is no necessity to work out a general reliability expression, a “formula” you might
call it, into which we can substitute numbers corresponding to the reliabilities of the
components and get the reliability of the system. However, reliability expressions
have been derived in some examples. For instance, see Example 2.3. In principle,
this can be done for more complex problems as well. Once the reliability
expression is in hand, then we can make use of the fact that for each particular
component the reliability function is known. Substituting these functions into the
reliability expression gives the reliability function for the entire system.
To illustrate this concept, let’s reconsider part c of Figure 8.2. Here the
probability that this circuit is good is

P(A good) P(B good) P(C good orD good) = p, Pp (Po +Pp -PcPp)

Notice that in this expression it doesn’t matter whether we are thinking of the
reliability as being time-dependent or not. Now if we utilize the assumption that
device A has an exponentially distributed lifetime with parameter p,, this means
that at time ¢ the probability that A is good is e~?:’. If we substitute e~?:’ for
p,, in the equation above and do a similar replacement for the other probabilities,
the reliability function for the system is obtained just as in Equation 8.4. As a more
elaborate example, see Problem 8.3. A solution using these ideas is given in
Appendix A.

8.2 Systems with Repair

How can we model a device that is subject to random failures but in which
there is also the capability of repairing the device after a failure occurs? After the
device is repaired it can be placed back in service.
Let’s assume a constant failure rate A. This means that whenever the device
is functioning, the waiting time for the next failure is assumed to be exponentially
distributed with parameter A. (Remember that the lack-of-memory property of the
exponential distribution means that the time at which we begin measuring is
206 Chapter 8: Time-Dependent System Reliability

irrelevant.)
We will similarly assume that the “time to repair” for the device is also
exponentially distributed; that is, the total time that the device is out of service
because of the failure is also an exponential random variable. We will denote by
the parameter in the exponential distribution for the time the unit is out of service.
It is now possible to track the performance of the device over a lon g period
of time. Presumably there would be a number of times that the device fails, and in
each case there would be a certain delay in getting it serviced and put back into
operation. What are some of the things that would be useful for us to know in this
situation? One thing we would be interested in is the percentage of the time that the
device is operational. For example, if breakdowns are so difficult to repair that the
device is unavailable for use 60% of the time, this is certainly something that a
prospective buyer would be interested in knowing. Or we might be interested in
knowing the probability that the device will be functioning at some specific time in
the future. This question is related to the former one. For instance, if we knew that
the device was going to be out of service for 60% of the time during the next
several years, that certainly suggests that over the long haul its probability of being
operational ought to “average out” in some sense to 40%. Whether its probability
of being operational next Tuesday afternoon is 40% is another question, and this
suggests that we look at the situation in more detail.
First, let’s remember some things we have already encountered. Since time
to failure is exponentially distributed with parameter 1, we know that the mean time
to failure for the device is 1/A. Similarly, since the time to repair is exponential
with parameter [1, the mean time required for the device to be repaired is 1/u. For
example, if time is measured in years and A = .5 and uy = 12, then the mean time to
failure will be 2 years and the mean repair time 1 month.
Now let’s consider how we might actually work out the probability that the
device will be functional at some specific time in the future. In order to do that, we
need first to establish the necessary notation. Specifically, we would like to know
about the following two functions:
f(t) I= probability the device is operational at time t
g(t) = 1-—f(t) = probability the device is nor operational at time t
We need to recall a few properties of the exponential function. If we think of
the random variable X as representing time to failure for the device (as measured
from initial time when the device is known to be working), then P(X > At) =
e —At Now remember from the power series representation for the exponential
8.2 Systems with Repair 207

function that
2
(AAr)
et = 1 A+
2!
In particular, if Art is small, then e-*4‘ ~ 1—AAt. (This approximation
amounts to using a linear approximation to evaluate the exponential function near
0.) This means that
P(X > At) = 1—AAt
when At is small, and it means that the probability that a failure will not occur
before time Af is approximately 1 — Az. Conversely, the probability that a failure
will occur before time At is approximately AAt. Applying this same idea to the
time to repair leads to the conclusion that for a small time interval At, if we know
the device to be nonoperational at the beginning of the time interval, then the
probability that it will still be nonoperational after time At has elapsed is
approximately 1 — uWAr, whereas the probability that it will have been repaired
during the time interval is approximately LAr.
We now need to consider f(t + At), where t is some fixed time and At is
considered a small positive number. By definition, f(t + Az) is the probability
that the device is working at time ft + At.
There are two primary ways that the device can be working at time ¢ + At:
(1) It could be working at time ¢ and experience no failure between time ¢ and time
t+ At. (2) It could have been out of service at time ¢ but have been repaired
between time t andt¢+At. There are clearly other possibilities involving more
than one failure and/or repair during the time interval, but if At is small then the
probabilities of such terms will be small compared to the terms above. (If all of this
seems a little fishy, it might be a good idea to go back and review the discussion of
Poisson processes in Chapter 7. The assumptions here are quite similar to those
that describe the Poisson process.)
So we can break f(t + At) down as follows:
f(t+ At) = probability device works at time ¢ and
doesn’t fail during time interval of length At
+ probability device is failed at time t
but is repaired during interval of length At
= f(t) (1—-AAr) + g(x) pat
If we rearrange the terms and replace g(t) by 1 — f(t), we can write this as
208 Chapter 8: Time-Dependent System Reliability

f(t + a fee Af(t)+ ug) = A+WfO+p

Taking the limit as At > 0 leads to the differential equation

FO =-A+wWfO +h
This is a simple first-order linear equation and has the solution

f(D) = cert +et (8.5)

The value of the constant c will of course be determined by the initial
conditions. For instance, if we know for certain that the device is working at time
0, then f(0) = 1, and this leads to the solution

£0. > OES eM teg (8.6)

for the probability that the device is working at time t. Moreover, since g(t) = 1 —
f(d), the probability that the device is not working at time f is
r
OVE reer agt Seree (8.7)
The situation is actually more symmetric than the last two equations make it
appear. The asymmetry stems from the fact that f(0) = 1 whereas g(0) = 0. The
general solution for g is the same as the general solution for f except that the roles
of and A are reversed.
Notice that as time passes the effect of the initial conditions dies out.
r
(0 > arv and AGT wins winey

So if A = .5 and pp = 12 as in the earlier illustration, then as time passes the

probability that the device will be working asymptotically approaches .96 and the
probability that the device is not working approaches .04.
The model which we have just been using for a repairable device is a special
case of an important class of stochastic processes called Markov processes. A
Markov process is a stochastic process {X,} in which the random variable X,
describes the state of a system at time ¢. In the example we have just considered,
there are two states: The device is working or it is not working. Clearly if there
are a finite number of states, we may then think of them as being represented by
integers 1,---,n. Thus the random variables that make up a Markov process
may be viewed as taking values 1, ---,n for some integer n, with the values
representing the different states of some system. The study of Markov processes
revolves around consideration of the transition probabilities associated with
moving from one state into another. In our simple example we could consider
“working” as state 1 and “not working” as state 2. The length of time that we
Problems 209

remain in each state was assumed exponentially distributed in our example, and in
our case there is no doubt about which state we will move to when a change of state
takes place. In a more general setting with numerous states involved, the transition
probabilities specify this information. For example, if one is modeling the number
of items waiting to be repaired at a service shop and if the items are being delivered
and processed individually, then the “state” of the system would be the number of
items present, and a change of state would be caused by (1) finishing work on an
item and shipping it out or (2) a new item being brought to the shop. If the state is
k, then a change of the first type makes the new state of the system k — 1, anda
change of the second type makes the new state k + 1. Whether the transition
probability of going from state k to state k — 1 is greater or less then the
probability of going from state k to k + 1 is then simply a question of whether the
items are being processed faster than they are arriving.
The precise definition of a Markov process is somewhat technical, but the
general idea is that the probability of making a transition from one state (state 7) to
another (state /) in a certain time period should not be dependent on how one came
to be in state i in the first place. In our problem above, this assumption is
implicitly present in the assumed exponential waiting times for a change of state.
From any point in time and from either of the two states, future transitions are
governed by the current state of the device and by the two exponential distributions
involved and are independent of the past history of the device.

Problems

8.1 Determine the mean time to failure (mttf) for the system in Figure 8.2(c)
assuming that the components A, B, C, and D function independently and
have constant failure rates p,, Py, P3, and p, respectively, that is, assuming
that the time to failure for each component is an exponential random variable
with the given parameter.

8.2 Consider a device with hazard function given by A(t) =1+(2- ‘ye
whenever t > 0. If we think of time as measured in years, you can notice
that the instantaneous failure rate is dropping off until the age of the device
reaches 2 years, at which time the failure rate starts to increase again because
of age. Let X denote “time to failure.” Determine the density function for
210 Chapter 8: Time-Dependent System Reliability

X as well as the distribution function. (Actually, the latter should be done

first, using Equation 8.3.) What is the probability that the device fails during
its second year, that is, between time t = 1 and time tf = 2? Notice that the
mttf here is not easily computable because tfy(d) is not readily integrable.

8.3 In the circuits shown in Problem 2.2 of.Chapter 2, assume that each
component has an exponentially distributed lifetime with mean 1 year, and
that the components function independently. Find the reliability function

Rsys(t) = probability that the system is still functioning at time t

8.4 In the highway network shown in Problem 2.4 of Chapter 2, assume that
each link in the network (each edge of the graph) has an exponentially
distributed lifetime with mean 6 weeks. Find the reliability function

Rsys(t) = probability of being able to go from Start to Finish at time t

8.5 Determine the density function for “waiting time to system failure” for the
series-parallel system in the following figure.

Assume that the components function independently, that each has an

exponentially distributed time to failure, and that the mean time to failure for
the individual components are 5 months for component A, 10 months for
component B, and 20 months for component C. What is the mttf for the
system?

8.6 Suppose a device has a hazard function that is a linear function of time; that
is, h(t) =kt, where k is a positive constant. (This says that the rate at
which such devices fail is proportional to the age of the devices.) Find the
reliability function for such a device.

8.7 A piece of equipment averages going 90 days between breakdowns, and the
time requirement to get it serviced averages 3 days. Answer the following
questions using the model developed in Section 8.2.
(a) In the long run, what fraction of the time is the equipment operable?
(b) If itis operating now, what is the probability it will be operating 5 days
Problems 211

from now? [Hint: Treat “now” as time zero and the given information
that it is operating at time zero as an initial condition.]

8.8 A telephone is a two-state device in that the line is either free or busy.
Assume that when the phone is not being used, the time until a shift into the
“busy” state is exponentially distributed with expected value 10 minutes.
And assume that when the phone is busy, the time until a shift into the not-
in-usé state is exponentially distributed with expected value 3 minutes.
Based on the model of Section 8.2, answer the following:
(a) If the phone is busy at time ¢ = 0, find the function f(t) = probability
that the phone will be in use at time t.-
(b) In the long run, what fraction of the time is the phone in use?

8.9 (A modeling problem) A telephone has a “hold” button so that an incoming

call can be placed on hold if the line is tied up. So there are three states for
this phone: (1) the line is free, (2) the line is busy, and (3) the line is busy
and someone is on hold.
Any incoming call that arrives while the phone is in state 3 is turned
away. Thus from state 3 it is possible to go only to state 2, and similarly
from state 1 it is possible to go only to state 2. From state 2 we might go to
either state 1 or state 3, depending on whether and incoming call arrives
before the present caller hangs up. The probabilities will depend on the
relative rate of incoming calls as compared to the length of time that the calls
last.
We will assume that arriving calls form a Poisson process with
intensity A and that the duration of calls is exponentially distributed with
parameter Lt. So the length of time in state 3 (the length of time until the
present caller hangs up) is exponential with parameter LU, and the length of
time in state 1 (the length of time until an incoming call) is exponentially
distributed with parameter A. (Review the properties of a Poisson process if
you need to.) But what about the length of time in state 2? When in state 2,
the phone will stay in state 2 until either the caller hangs up (sending us back
to state 1) or another call arrives (sending us to state 3). Recall that the
minimum of two exponential random variables with parameters Lt and A is
exponential with parameter 1 + 4. So the waiting time for a change of state
while in state 2 is exponentially distributed with parameter 1 + 4. When a
change of state does occur from state 2, what is the probability it will be a
 Chapter 8: Time-Dependent System Reliability

change to state 1? This is simply the question: What is the probability that
the present caller will hang up before the next incoming call arrives? We
solved this problem in Chapter 6. (See Problem 6.26.)
If f(t), g(t), and A(t) represent the probabilities that at time t we
will be in states 1, 2, and 3, respectively, see if you can derive the system of
differential equations that must be satisfied by these three functions. You can
do so by mimicking the approach used in Section 8.2.
 References

Breiman, Leo, Probability and Stochastic Process with a View toward Applications,
Houghton Mifflin, Boston, 1969.

Chung, Kai Lai, Probability Theory with Stochastic Processes, Springer-Verlag,

New York, 1974. A very readable general-purpose text.

Cinlar, Erhan, Jntroduction to Stochastic Processes, Prentice-Hall, Englewood Cliffs,

N.J., 1975. This book presupposes a solid foundation in probability and advanced calculus. The
exposition is aimed, however, at engineers and others who have applications rather than theory as their
objective.

Dwass, Meyer, Probability Theory and Applications, W. A. Benjamin, New York,

1970. This book is a good reference on specific distributions. It includes treatments of the gamma
and beta distributions, the chi-square distribution, and the multinomial and negative binomial
distibutions. The exposition is a bit more theoretical than is found in engineering oriented texts.

Feller, William, An Introduction to Probability Theory and Its Applications, Wiley,

New York, 1968. This is perhaps the most widely referenced work on probability in existence. It
consists of two volumes, the first on discrete distributions, and the second on continuous
distributions. These books include much information not found in general-purpose texts.

Page, Lavon B., and Jo Ellen Perry, “A practical implementation of the factoring
theorem for network reliability,” JEEE Trans. Reliability, R-36, pp. 259-267, 1988.
This article describes a microcomputer algorithm which uses conditional probabilities to treat network
reliability problems in a manner reflecting the ideas encountered in Chapter 2.

Solomon, Frederick, Probability and Stochastic Processes, Prentice-Hall, Inc.,

Englewood Cliffs, N.J., 1987. This isa general-purpose undergraduate probability text.

213
 Appendices

Appendix A: Answers, Partial Solutions, and Hints to Selected

Problems

Appendix B: Values of the Normal Distribution Function

214
Appendix A 2NS

Appendix A: Answers, Partial Solutions, and Hints

to Selected Problems

Chapter 1

Vi *@)at8 = (b) 6c ye72 (0) 88

1.2 This problem can be done easily using a tree diagram or the combinations
formula. If the combinations formula is used the answers can be given by
Cr 2) 2 b 2/5 23
Ge 2) 3 CET CO, DiLG Det
Pe kes

uy C(4, 2) + C(3, 2) + C(2, 2) =

Cor) 18

1.5. -P(A)=

1.6 (a)P(2 good) =C(, 2) COvCL. = .0081, and so on.

(b) P(at least 3 good) = P(3 good) + P(4 good) + P(5 good)

1.8 The answers are easily obtained from following tree diagram.

QD Q®
a io
EI 1a (uae

(b) (.6)? + 2(.6)7(.4) (c) 2(.6)*(.4) (d) 64 (e).36

216 Appendices

1.11 Using the following tree, we see that P(disease and +) = .045 and
P(disease | +) = .045/.121 = .3719.

ee 08 92
Disease . Nodisease

.05 95

1.13 (a) P(B)=P(A,) PIA, + --- +P(Ag) PB 1A¢)

Lees | ERs Leg it $s
a ea te as
(b) Az; is the event that the green die shows a 3. So

P(A3) P(B |A3) 1/6x3/6 1

P\As|
2) =) ae ae SPS AG == -5-
» P(A;)P(B |Aj)
i=1

1.14 @) S ={ (HE) 2) C3) eT, Se CE? G)}

(b) P(head on coin) = 6/12, P(even number on die) = 6/12,
P(head on coin and even no. on die) = 3/12.

1.17 C(10, 2) (.3)°0.7)® = .233474

C30, 2) €U0.S
nibs os ee
1.22 If n is the number of random digits generated, then P(no 7’s) = (.9)”.
Therefore, P(at least one 7) = 1 — (.9)”. So what we want is to have n
sufficiently large that 1 - (.9)” > .95. This means we need (.9)” < .05.
Take the natural logarithm of both sides: To have n In (.9) < In(.05)
requires that n 2 In(.05)/In(.9) = 28.43. Thus we need to generate at least
29 random digits.
Appendix A DAT.

Chapter 2

Deo (a) ook es (b) .95086

2.4 Going from start to finish is possible provided that link E is good and that at
least one of the two pairs A, B or C, D is good. The probability is
P{[ANB)U(CAD)]AE}
= PL(ANB)U(COD)] x P(E)
=[P(ANB)+P(CAD)-P(ANBOACAD)) x P(E)
=PaPBPE*PcPDPE~PAPBPCPDPE

2608 eo10299

2.9 P(battery) x P(R,) x P(light) x P(R, or R3)

= P(battery) x P(R,) x P(light) x {P(R2) + P(R3) — P(R,) x P(R3)}

2.10 Suppose n of the devices are used. The probability they all fail is then .2”,
so the probability that at least one works is 1 — .2”. We wish this probability
to be = .9999. So we need to have n large enough that 1 — .2”> .9999.
This inequality is easily solved for n: The inequality .2”< .0001 will be
true whenever n In (.2) < In (.0001). This requires that n = 5.72. Since a
whole number of devices is required, we should use n = 6 devices.

Chapter 3

3.1. Your graph should have a jump discontinuity at 0, 1, 2, 3, and 4 and should
be constant on any interval not containing one of these numbers.

3.3
0 ift<0
Fylt)rathet df0sisd
Rapes
218 Appendices

3.4 (ay\PO <x <65)=325 (b) Fy(.5) = .75

0 ift<-l

(c) Fy(t) = an if-1<t<1

(Le see ies I

Sud P(X = 1) = 1/36; P(X = 2) =3/36, PC = 3)-= 5/36, and'sojon.

3.6 This random variable could be interpreted as what you’re looking at if you
choose a random number between 0 and 2 and truncate it after the first
decimal place.

Sal P(X >1)=e° and P(1<X <2) =e3-e.

3.8 PU<1) = PCL<x <ty="1-e*

a cC=—=

Dalz Skip ahead and look at Definition 4.1. In Chapter 4 the random variable X
will be called a binomial random variable with parameters n = 4 and p =.1.

Salas E(X) = 1/2 and E(Y) =2.

Appendix A 219

316 Fy(t)= PY <1) = P(2X+3.s tf) = P(X $.@ —3)/2) =F y(t = 3)/2).
You computed Fy in Problem 3.3. Using that information we get
0 ifi<3
t-3.
Fy(t) = ee ito Sts)

iy Mitts5S
To obtain fy simply differentiate Fy. What you will notice is that Y is
uniformly distributed on the interval [3, 5]. This should not be surprising
given the way that Y is defined in terms of X.

3.17 Notice that Problem 3.16 is a special case of this,

3.18 (c) E(X) =.21

Chapter 4

paca = SeBeS
4.1 Forexample,

4.2 (a) e = 35 (b) C(10, 3) a0

100 100

4.3 (a) EX)= o> k (c) E(Y) = o>, 2*

k=1 k=1

4.4 X = number of heads in 3 tosses, Y = number of tails = 3 — X. So the

number of heads minus the number of tails is X — (3 —X) = 2X — 3.
Therefore E(Y) is given by
3)

>) 2e-31P&= = 1.5

k=0
220 Appendices

4.6 Show that

P(XX=kK A
PX=k-1l) k
This shows that P(X =k) > P(X =k-— 1) if and only ifA >k. So if
we look at the probabilities P(X = 1), P(X = 2), --- , the probabilities are
getting larger until k gets to be bigger than 2.
Conclusion: The largest probability is P(X=k) where k is the largest
integer less than X. If A is itself an integer, then there is a tie between the two
numbers P(X =) and P(X =A + 1).

4.7 P(X =3)=.14037

4.8 X is binomial with n = 50, p =.1, and P(X = 3) = .13857.

4.10 var(X) =2

4.19 .2646

4.21 The distribution to use is the binomial distribution.

Chapter 5

5.2% a) 6065 (b) .6321 (c) .6321

5.3. Y=1-X. This means that

Fyt) =PYson=PA-X sp)=PX>1-0
= 1-P(X $<1-N)=1-Fy(1-d
Use your knowledge of what Fy is (Problem 3.3) and be very careful with
your inequalities to show that in fact Fy = Fy, and so Y also is uniform on
(Oly,

5.4 (a) The answer is 1/2. You don’t need a table if you just remember the
symmetry of the density function.
(b) .36788
Appendix A 221

0) The method is similar to Exercise 5.3.

fo.
em OP if tora
(c) ft) = 4 2 (d) E(Y) =3
QO otherwise

>.6 fy is given by

sae THOS fl

0 otherwise

EQ) =5

ayy Y = |X -1I. This means

Fy) = PYSd
= P(x-I|l<f)
=P(-tsX-1<f) = P(l-tsXs1+nf
Since the distribution of X is known, this probability can be easily
evaluated. You probably will want to treat the three cases t< 0,0 <1r< 1,
and 1 <¢ separately. What you should find out is that Fy turns out to be
uniform on [0, 1]. Think about the way Y is defined, and try to understand
intuitively why this is true.

5.8 This problem is really a combination of two problems already done. Think
of the wire as being laid on the x-axis from 0 to 2. If X is the x coordinate
of the point where the cut is made, the assumption is that X is uniform on
{0, 2]. Figure out that the length Y of the shortest piece is the random
variable Y = 1 — IX — 1I. To put this in the context of problems already
done, let’s temporarily use the notation Z = |X — 1]. From Problem 5.7 we
know that Z is uniform on [0, 1]. Since Y = 1-—Z, we know from
Problem 5.3 that Y then is also uniform on [0, 1], and so E(Y) = 1/2. The
expected length of the shortest piece of wire is 6 inches.
Of course it is not absolutely necessary to view this problem in terms of
Problems 5.3 and 5.7. This problem could be done from scratch using the
same techniques as are used in those problems.
222 Appendices

Slay 6.93 years

Sb P(2<X <4) = ®(3/2) — B(1/2) = .2417, where ® is the standard normal

distribution function.

Siz var(X) = E(X’) - EGY. Compute E(X’) as follows using integration

by parts:

E(X’) = { Rre™ dt
0

eile; Let X = number of defects. Then

és 225— Wb 2 aah Sa
PQ25 <X < 275) p( - Se ers

@D(—1/4) — B(-3/4)
(3/4) — (1/4) = .1747

ald (a) .6065 (b) .3679 ~— (c)- .1353

5.18 (a) 98% =~ 6035 (b) 98° = .3642 (c) .98!° ~ 1326

SEO Let X denote the random number, and let Y denoted the value of X
truncated to a single digit after the decimal. Then Y = g(X) where g is the
step function shown in Figure 4.1. This means that E(Y) can be computed
as
co iL

BO) = | sfxo.ar = |gar = 49

5.24 mtie)

bs The equation E =/R is used here, where E = voltage drop, J = current,

and R = resistance. Applying this equation to the entire circuit gives 12 =/
(R + 1). Also from this equation we know that the voltage drop V across
the resistor is given by V=/R. Sol = 12/(R + 1), and this implies that
V=12R/(R+ 1). Therefore,
Fyt) =PVso)=P[12R/(R+1) st] = PL 12R<t(R+1)]
Appendix A 223

= P[((12-n)Rs<t) = PIRS<#e#12-4) = Fpl /12-))

The distribution function Fp is familiar. Figure out what it is, and from
that you can deduce that
0 ift<6
t ;
Fy{t) = ee

1 ift=8

It is easy to obtain the density function for R just by differentiating Fy.

As usual, E(V) can be computed from fy or directly from the density
function for R.

Chapter 6

6.1 The table should look something like this:

6.3 The computations to obtain the table below are a bit tedious, but not terribly
difficult if you get the probabilities from a tree diagram.

0 1/15 4/15 1/15 2/5

| 2/15 4/15 2/15 8/15

2 0 1/15 0 4/15

Y |] 15 3/5 1/5
224 Appendices

6.5 This one is easy. The table is as follows:

6.7 Partial answer: EX) =E(Y) = 2/5.

6.12 Let X and Y be the two numbers. The first step is to realize that the
probability in question can be obtained by integrating the joint density
function Fy y over the half-plane below the line y = x/2. Secondly, the
joint density function is easily obtained from Definition 6.6 since X and Y
are independent. The problem then boils down to integrating the joint density
function over the shaded triangle in the picture. Since the joint density is
simply the uniform density on the unit square, the probability is just the area
of the triangle, which is 1/4.

6.15 The density function for X is: fy(x) = 1/(1 + x)? whenever x > 0, with
fx(x) = 0 when x < 0. The density for Y is simply the exponential density
with parameter A = 1.
Appendix A 225

6.17

Set time t = 0 to be 8 o’clock. Let X be the time when your friend calls
you, so the assumption is that X is uniform on [0, 1]. Let Y be the waiting
time for the first call from someone other than your friend, so Y is
exponential with parameter A = 2. (Remember that in the exponential
distribution the parameter is the reciprocal of the expected value.)
The problem is to find P(X < Y). This is computed by integrating the
joint density function over the region consisting of the half-plane above the
line y =x. The joint density is 0 except when 0 < x < 1 and y > 0, and the
part of this region that lies in this half-plane is the region shaded in the picture.
So the probability is computed by performing the integration
1 poo

) 2e-2y dy dx = .4323
0 4x

6.19
|
E(X+Y)= Se
20

6.20 (a) It is easy to show that both X and Y have exponential densities.
(Dan Y eS:
(Cha ROG SAO ee LO een,
(d) Integrate the joint density over the appropriate triangle and subtract from
1.

6.22 It should be apparent that this computation is going to require integrating the
joint density over a triangle in the first quadrant. The density function works
out to be like this: If Z= X + Y, then
226 Appendices

—2t
tht >0

0 otherwise

6.23 This is an excellent problem, and it will be well worthwhile for you to work it
two different ways. One way is to work out the density function for the
random variable Z = max{X, Y}, and then to compute E(Z) using this
density function. The second way is to simply utilize the joint density, since
Z is already expressed-as a function of X and Y. Since this requires
figuring out how to integrate the function of two variables g(x, y) =
max {x, y}, the picture below should be helpful.

On this triangle On this triangle

max{ x, y}= y max{ x, y}= x

6.24 The expected value is 7/3.

Chapter 7

fie) Denote by Y the number of the trial on which the second success occurs. In
order for the second success to occur on trial n, there must be exactly one
success during the first n — 1 trials. Given any two specific trials among
the first n trials, the probability of success occurring on those 2 and failure
on all others would be p2qg"-?._ How many ways are there in which the
second success could occur on the nth trial? The first must then occur on
one of the trials from trial 1 to trial n — 1, so there are n — 1 ways for the
second to occur on the nth trial. Thus
PY=n) = (n- 1) p2q”
Appendix A 227

Another way of considering this problem is to think of X being the number

of the trial on which the first success occurs, in which case
PY =n)i= P=) PC endl Xea MN) ha:
+ P(X =n-1)P(Y=n1X =n-1)
Each term in the sum on the right side of this equation can be checked to be
equal to p2q”*, and there are n — 1 terms.

Es If Y, and Y> are the times of the first and second occurrences respectively,
then Yy = Y, +(Y,-—Yj,). This simply amounts to looking at Y, as the
time to the first occurrence plus the time between the first and second
occurrence. From part 3 of Proposition 7.3, Y, and Y, —Yj, are
exponential with parameter 1 = intensity of Poisson process, and they are
independent. So Y, is the sum of two independent exponential random
variables with parameter 1. This type of problem was treated in Chapter 6.
See for example Problem 6.27.

7.6 See the comment above regarding Problem 7.5.

vey Let X denote the number of breakdowns to occur during 6 months. Then
X is Poisson with parameter 2 = pt = (10 per year) x (.5 years) = 5. And
the probability of 7 or more occurring is then
P(X >7)=1-{
P(X =0)+---+P(X =6) }

7.8 X is Poisson with parameter A = ut = (30 per hour) x (1/6 hour) = 5. And
Y has the same distribution. X + Y is Poisson with parameter

X = ut = 30perhour x = hour = 10
Thus, using the properties of a Poisson process described in this section,
there is no work to do in this problem. However, there is a basic underlying
fact at work: If X is Poisson with parameter 4, and Y is Poisson with
parameter A. and X and Y are independent, then X + Y is Poisson with
parameter A, +,.
228 Appendices

Chapter8

8.3 First work out a reliability expression for the circuit. The following is one
possibility:
trel=[ (Dp +Pp-PBPp) Pc +Pa-PaPc (Pp +Pc —PBPp) IPE
Now since each of the components is assumed to have exponentially
distributed lifetime with parameter 1, this means that for any given device, at
time ¢ the probability that the device is still functioning is e’. So for every
one of the probabilities on the right side of this equation, we can substitute
e', This gives the reliability of the system as a function of t, and works
out to be
relays(t) = e-7! + 264! Be + e-F
Let’s denote by X the time to failure for the whole circuit. So (X > 2) is
the event that the circuit is still working at time ¢, and this is what rel,,.(¢)
denotes above. Therefore
Fy(t) = P(X St) = 1-rel,,.(f) = 1— { e+ 2e- — et + 8}
Differentiation gives the density function for X, and it is then easy to
compute E(X), which is the mttf for the circuit. Is it clear to you that Fy is
a density function? If not, you should check out that it is. Can you make
that observation based on the fact that1+2-—3+1=1?

85) fwee "4 .05er ese

8.7 (a) H/(A +L) = 30/31 = .968; that is, 96.8% of the time the equipment is
operable.

(b) f(5) = (1- + UL

eso espe
A+U
eer O75
 229 Appendix B

Appendix B: Values of the Normal Distribution Function

The table of values appearing here was generated on a Macintosh micro-
computer using an elementary program written in Turbo™ Pascal to implement
the algorithm shown in Problem 5.14 of this text. Values shown are for (x),
where ® is the standard normal distribution function and x > 0. For x <0, the
identity ®(—x) = 1 — M(x) may be used.

0.00 | 0.5000 0.5040 0.5080 0.5120 0.5160 0.5199 0.5239 0.5279 0.5319 0.5359
0.10 | 0.5398 0.5438 0.5478 0.5517 0.5557 0.5596 0.5636 0.5675 0.5714 0.5753
0.20 | 0.5793 0.5832 0.5871 0.5910 0.5948 0.5987 0.6026 0.6064 0.6103 0.6141
0.30 | 0.6179 0.6217 0.6255 0.6293 0.6331 0.6368 0.6406 0.6443 0.6480 0.6517
0.40 | 0.6554 0.6591 0.6628 0.6664 0.6700 0.6736 0.6772 0.6808 0.6844 0.6879
0.50 | 0.6915 0.6950 0.6985 0.7019 0.7054 0.7088 0.7123 0.7157 0.7190 0.7224
0.60 | 0.7257 0.7291 0.7324 0.7357 0.7389 0.7422 0.7454 0.7486 0.7517 0.7549
0.70 | 0.7580 0.7611 0.7642 0.7673 0.7704 0.7734 0.7764 0.7794 0.7823 0.7852
0.80 | 0.7881 0.7910 0.7939 0.7967 0.7995 0.8023 0.8051 0.8078 0.8106 0.8133
0.90 | 0.8159 0.8186 0.8212 0.8238 0.8264 0.8289 0.8315 0.8340 0.8365 0.8389
1.00 | 0.8413 0.8438 0.8461 0.8485 0.8508 0.8531 0.8554 0.8577 0.8599 0.8621
1.10 | 0.8643 0.8665 0.8686 0.8708 0.8729 0.8749 0.8770 0.8790 0.8810 0.8830
1.20 | 0.8849 0.8869 0.8888 0.8907 0.8925 0.8944 0.8962 0.8980 0.8997 0.9015
1.30 | 0.9032 0.9049 0.9066 0.9082 0.9099 0.9115 0.9131 0.9147 0.9162 0.9177
1.40 | 0.9192 0.9207 0.9222 0.9236 0.9251 0.9265 0.9279 0.9292 0.9306 0.9319
1.50 | 0.9332 0.9345 0.9357 0.9370 0.9382 0.9394 0.9406 0.9418 0.9429 0.9441
1.60 | 0.9452 0.9463 0.9474 0.9484 0.9495 0.9505 0.9515 0.9525 0.9535 0.9545
1.70 | 0.9554 0.9564 0.9573 0.9582 0.9591 0.9599 0.9608 0.9616 0.9625 0.9633
1.80 | 0.9641 0.9649 0.9656 0.9664 0.9671 0.9678 0.9686 0.9693 0.9699 0.9706
1.90 | 0.9713 0.9719 0.9726 0.9732 0.9738 0.9744 0.9750 0.9756 0.9761 0.9767
2.00 | 0.9772 0.9778 0.9783 0.9788 0.9793 0.9798 0.9803 0.9808 0.9812 0.9817
2.10 | 0.9821 0.9826 0.9830 0.9834 0.9838 0.9842 0.9846 0.9850 0.9854 0.9857
2.20 | 0.9861 0.9864 0.9868 0.9871 0.9875 0.9878 0.9881 0.9884 0.9887 0.9890
2.30 | 0.9893 0.9896 0.9898 0.9901 0.9904 0.9906 0.9909 0.9911 0.9913 0.9916
2.40 | 0.9918 0.9920 0.9922 0.9925 0.9927 0.9929 0.9931 0.9932 0.9934 0.9936
2.50 | 0.9938 0.9940 0.9941 0.9943 0.9945 0.9946 0.9948 0.9949 0.9951 0.9952
2.60 | 0.9953 0.9955 0.9956 0.9957 0.9959 0.9960 0.9961 0.9962 0.9963 0.9964
2.70 | 0.9965 0.9966 0.9967 0.9968 0.9969 0.9970 0.9971 0.9972 0.9973 0.9974
2.80 | 0.9974 0.9975 0.9976 0.9977 0.9977 0.9978 0.9979 0.9979 0.9980 0.9981
2.90 | 0.9981 0.9982 0.9982 0.9983 0.9984 0.9984 0.9985 0.9985 0.9986 0.9986
3.00 | 0.9987 0.9987 0.9987 0.9988 0.9988 0.9989 0.9989 0.9989 0.9990 0.9990
3.10 | 0.9990 0.9991 0.9991 0.9991 0.9992 0.9992 0.9992 0.9992 0.9993 0.9993
3.20 | 0.9993 0.9993 0.9994 0.9994 0.9994 0.9994 0.9994 0.9995 0.9995 0.9995
3.30 | 0.9995 0.9995 0.9995 0.9996 0.9996 0.9996 0.9996 0.9996 0.9996 0.9997
3.40 | 0.9997 0.9997 0.9997 0.9997 0.9997 (0:9997% 0.9997 0.9997 0.9997 0.9998
 Index

A computational complexity, 40
conditional density function, 156
all-terminal reliability, 46 conditional expectation
AND gate, 49 continuous random variables, 157
autocorrelation function, 186 discrete random variables, 154
average intensity of Poisson process, conditional probability mass function,
179, 181 152
conditional probability, 11-12
continuity correction, 112
B continuous (random variable), 60, 62—
63
Bayes’ Theorem, 17 continuous parameter stochastic
Bernoulli process, 84, 189 process, 171
Bernoulli trials, 84 contract (an edge of a graph), 45
binomial distribution, 84, 182 convolution, 144
binomial process, 188 countably infinite, 26
cumulative distribution function, 65-66

C
D
central limit theorem, 112, 160-161
Chebyshev’s inequality, 115 De Moivre-Laplace theorem, 160
circuit De Morgan’s laws, 4
parallel, 35 degree-two vertex, 46
series, 35 delete (an edge of a graph), 45
series-parallel, 36 density function, 62, 131
combinations formula, 22 discrete (random variable), 60
combinations, 23 discrete parameter stochastic process,
communication path, 41 171
complement (of a set), 3 discrete uniform distribution, 91

230
Index 231

disjoint, 3
distribution function, 65
distribution, 67 inclusion-exclusion principle, 43
independent
continuous random variables, 136
Es discrete random variables, 129
events,-13, 21
ergodic in the mean, 194 independent trials formula, 23
ergodic, 193 independent trials process, 22, 171—
event, 8 Has 2
expectation (see expected value) indicator (random variable), 73
expected value, 69-71 intensity of Poisson process, 179, 181
exponential distribution, 104, 179, 201 intersection (of sets), 3

FE
J

fault tree, 49 joint density function, 130

frequency distribution, 96 joint probability mass function, 127
function of a random variable, 77
function of two random variables, 139
iG

G lack-of-memory property, 87, 105, 111

Laplace transform, 144
geometric distribution, 86 law of large numbers, 117
geometric series, 25 linearity (of expected value), 71
graphs, 38

M
H
marginal density function, 133
hazard function, 199 marginal mass function, 129
hazard rate, 199 Markov process, 208
histogram, 96 mean (see expected value)
hypergeometric distribution, 94 mean time to failure, 201-202
232 Index

mean, 69-71 standard deviation, 74

multiplication principle, 7 variance, 73-74
multiplicative law, 17 random walk, 173
mutually exclusive, 3 recursion, 48
redundant safety systems, 140
reliability, 34
N

networks, 37 S
normal density function, 108
normal distribution, 107 sample function, 184
sample space, 6
sampling with replacement, 94, 96
O sampling without replacement, 94, 96
sets
observation (of stochastic process), complement, 3
184, 194 difference, 4
OR gate, 49 intersection, 3
union, 3
universal set, 3
Pp
Simpson’s rule, 118
sink, 37
permutations, 23 source, 37
Poisson distribution, 92 source-to-sink communication, 46
Poisson process, 174-175, 181 standard deviation, 74
probability mass function, 60 standard normal distribution, 107
probability measure, 8 state (of a Markov process), 208
stationary stochastic process, 181, 186,
188-189
R stochastic process, 171
strictly stationary stochastic process,
random variable, 57 190
continuous, 62-63
discrete, 60
expected value, 69-71
mean, 69-71
Index 233

telegraph signal, 191

time average (of stochastic process),
193
transition probability, 208
tree diagram, 16, 41
two-terminal reliability problem, 46

uncorrelated random variables, 147

uniform distribution, 103
uniform sample space, 8
union (of sets), 3
universal set, 3
unreliability, 36

variance, 73-74
Venn diagram, 4

waiting time, 86, 105, 199

wide-sense stationary process, 190
 at as sitar 74

a + : ~ 2s Saeed Peg indetian leninreicon

; : ms Se | rf

. , Py r T= v =... -
D7 hae Z . . woes an ae c
ee A 3 oe x
Su wpe. —
« 2 = . ' ~ o
eet

sy Rb Lypoielguts poner Ie
= £5 a) Agie tA. DL weiind igen
aea
e
 Ml PROBABILITY FOR Hi
BaeG LIE ERI NG
WITH
APPLICATIONS
: TO RELIABILITY

LAY ONE. PAGE

North Carolina State-University

PROBABILITY .FOR ENGINEERING introduces the basic tools of

probability that are necessary to investigate the reliability of various
complex systems. Emphasizing practical applications, the book
addresses many contemporary engineering problems—from simple
circuits to complex communications networks—and presents a number
of procedures used to solve these problems.

The author offers substantial engineering applications throughout the

book. He examines many of the latest problem-solving techniques,
including fault trees and recursive algorithms, and includes numerous
worked examples to illustrate how probability tools are used to model
real-life situations.
4.0% -\) Is a professor of mathematics at North Carolina State
University, where he teaches a course on probability designed especial-
ly for students of electrical engineering. Professor Page has published
widely in the area of reliability. His research has appeared in such
journals as EEE Transactions on Reliability, Journal of Microelectronics
and Reliability, and Computers and Chemical Engineering.

COMPUTER SCIENCE PRESS

An imprint of W. H. Freeman and Company
41 Madison Avenue, New York, NY 10010
20 Beaumont Street, Oxford OX1 2NQ, England

ISBN 0-7167-8187-5