Math Stats Text

To HARRIET A.M.M.
To my GRANDCHILDREN F.A.G.
To JOAN, LISA, and KARIN D ..C.B.
Library of Congress Cataloging in Publication Data

Mood, Alexander McFar1ane, 1913-
Introduction to the theory of statistics.
(McGraw-Hi1l series in probability and statistics)
Bibliography: p.
1. Mathematical statistics. I. Graybill.
Frank1in A., joint author. II. .Boes, Duane C.,
joint author. III. Title.
QA276.M67 1974 519.5 73-292
ISBN 0-{)7-042864-6
INTRODUCTION
TO THE THEORY
OF STATISTICS
Copyright © 1963. t 974 by McGraw-Hill, Inc. All rights reserved.
Copyright 1950 by McGraw-Hili, Inc. All rights reserved.
Printed in the United States of America. No part of this pubJication may be reproduced,
stored in a retrieval system, or transmitted, in any form or by any means, electronic.
mechanical, photocopying, recording, or otherwise, without the prior written
permission of the publisher.
6789 10 KPKP 7832109
This book was set in Times Roman.

The editors were Brete C. Harrison and Madelaine Eichberg;
the cover was designed by Nicholas Krenitsky;
and the production supervisor was Ted Agrillo.
The drawings were done by Oxford Il1ustrators Limited.
The printer and binder was Kinsport Press, Inc.
CONTENTS
Preface to the Third Edition

...
Xlll
Excerpts from the First and Second Edition Prefaces xv
I Probability 1
1 Introduction and Summary 1

2 Kinds of Probability 2
2.1 Introduction 2
2.2 Classical or a Priori Probability 3
2.3 A Posteriori or Frequency Probability 5
3 Probability-Axiomatic 8
3.1 Probability Models 8
3.2 An Aside-Set Theory 9
3.3 Definitions of Sample Space and Event 14
3.4 Definition of Probability 19
3.5 Finite Sample Spaces 25
3.6 Conditional Probability and Independence 32
vi CONTENTS
II Random Variables, Distribution Functions, and

Expectation 51
2 Random Variable and Cumulative Distribution Function 52
2.1 Introduction 52
2.2 Definitions 53
3 Density Functions 57
3.1 Discrete Random Variables 57
3.2 Continuous Random Variables 60
3.3 Other Random Variables 62
4 Expectations and Moments 64
4.1 Mean 64
4.2 Variance 67
.4.3 Expected Value of a Function of a Random Variable 69
4.4 Ch~byshev Inequali ty 71
4.5 Jensen Inequality 72
4.6 Moments and Moment Generating Functions 72
ill Special Parametric Families of Univariate Distributions 85

2 Discrete Distributions 86
2.1 Discrete Uniform Distribution 86
2.2 Bernoulli and Binomial Distributions 87
2.3 Hypergeometric Distribution 91
2.4 Poisson Distribution 93
2.5 Geometric and Negative Binomial Distributions 99
2.6 Other Discrete Distributions 103
3 Continuous Distributions 105
).( Uniform or Rectangular Distribution 105
# .
(3.2 ) Normal Distribution 107

'--
3.3 Exponential and Gamma Distributions 111
3.4 Beta Distribution 115
3.5 Other Continuous Distributions 116
4 Comments 119
4.1 Approximations 119
4.2 Poisson and Exponential Relationship 121
4.3 Contagious Distributions and Truncated Distributions 122
CONTENTS vii
IV Joint and Conditional Distributions, Stochastic

Independence, More Expectation 129
2 Joint Distribution Functions 130
2.1 Cumulative Distribution Function 130
2.2 Joint Density Functions for Discrete Random Variables 133
2.3 Joint Density Functions for Continuous Random
Variables 138
3 Conditional Distributions and Stochastic Independence 143
3.1 Conditional Distribution Functions for Discrete Random
Variables 143
3.2 Conditional Distribution Functions for Continuous Random
Variables 146
3.3 More on Conditional Distribution Functions 148
3.4 Independence 150
4 Expectation 153
4.1 Definition 153
4.2 Covariance and Correlation Coefficient 155
4.3 Conditional Expectations 157
4.4 Joint Moment Generating Function and Moments 159
4.5 Independence and Expectation 160
4.6 Cauchy-Schwarz Inequality 162
5 Bivariate Normal Distribution 162
5.1 Density Function 162
5.2 Moment Generating Function and Moments 164
5.3 - Marginal and Conditional Densities 167
V Distributions of Functions of Random Variables 175

2 Expectations of Functions of Random Variables 176
2.1 Expectation Two Ways 176
2.2 Sums of Random Variables 178
2.3 Product and Quotient 180
3 Cumulative-distribution-function Technique 181
3.1 Description of Technique 181
3.2 Distribution of Minimum and Maximum 182
3.3 Distribution of Sum and Difference of Two Random
Variables 185
3.4 Distribution of Product and Quotient 187
viii CONTENTS
4 Moment-generating-function Technique 189

4.1 Description of Technique 189
4.2 Distribution of Sums of Independent Random Variables 192
5 The Transformation Y = g(X} 198
5.1 Distribution of Y g(X) 198
5.2 Probability Integral Transform 202
6 Transformations 203
6.1 Discrete Random Variables 203
6.2 Continuous Random Variables 204
VI Sampling and Sampling Distributions 219

2 Sampling 220
2.1 Inductive Inference 220
2.2 Populations and Samples 222
2.3 Distribution of Sample 224
2.4 Statistic and Sample Moments 226
3 Sample Mean 230
~Mean and Variance 231
3.2 Law of Large Numbers 231
.3 Central-limit Theorem 233
3.4 Bernoulli and Poisson Distributions 236
3.5 Exponential Distribution 237
3.6 Uniform Distribution 238
3.7 Cauchy Distribution 238
4 Sampling from the Normal Distributions 239
4.1 Role of the Normal Distribution in Statistics 239
4.2 Sample Mean 240
4.3 The Chi-square Distribution 241
4.4 The F Distribution 246
4.5 Student's t Distribution 249
5 Order Statistics
5.1 Definition and Distributions
5.2 Distribution of Functions of Order Statistics
~
254
5.3 Asymptotic Distributions 256
5.4 Sample Cumulative Distribution Function 264
CONTENTS ix
VII Parametric Point Estimation 271

2 Methods of Finding Estimators 273
2.1 Methods of Moments 274
2.2 Maximum Likelihood 276
2.3 Other Methods 286
3 Properties of Point Estimators 288
3.1 Closeness 288
3.2 Mean-squared Error 291
3.3 Consistency and BAN 294
3.4 Loss and Risk Functions 297
4 Sufficiency 299
4.1 Sufficient Statistics 300
4.2 Factorization Criterion 307
4.3 Minimal Sufficient Statistics 311
4.4 Exponential Family 312
5 Unbiased Estimation 315
5.1 Lower Bound for Variance 315
5.2 Sufficiency and Completeness 321
6 Location or Scale Invariance 331
6.1 Location Invariance 332
6.2 Scale Invariance 336
7 Bayes Estimators 339
7.1 Posterior Distribution 340
7.2 Loss-function Approach 343
7.3 Minimax Estimator 350
8 Vector of Parameters 351
9 Optimum Properties of Maximum-likelihood Estimation 358
vm Parametric Interval Estimation 372

2 Confidence Intervals 373
2.1 An Introduction to Confidence Intervals 373
2.2 Definition of Confidence Interval 377
2.3 Pivotal Quantity 379
X CONTENTS
3 Sampling from the Normal Distribution 381

3.1 Confidence Interval for the Mean 381
3.2 Confidence Interval for the Variance 382
3.3 Simultaneous Confidence Region for the Mean and Variance 384
3.4 Confidence Interval for Difference in Means 386
4 Methods of Finding Confidence Intervals 387
4.1 Pivotal-quantity Method 387.
4.2 Statistical Method 389
5 Large-sample Confidence Intervals 393
6 Bayesian Interval Estimates 396
IX Tests of Hypotheses 401

2 Simple Hypothesis versus Simple Alternative 409
2.1 Introduction 409
2.2 Most Powerful Test 410
2.3 Loss Function 414
3 Composite Hypotheses 418
3.1 Generalized Likelihood-ratio Test 419
3.2 Uniformly Most Powerful Tests 421
3.3 Unbiased Tests 425
3.4 Methods of Finding Tests 425
4 Tests of Hypotheses-Sampling from the Normal Distribution 428
4.1 Tests on the Mean 428
4.2 Tests on the Variance 431
4.3 Tests on Several Means 432
4.4 Tests on Several Variances 438
5 Chi-square Tests. 440
5.1 Asymptotic Distribution of Generalized Likelihood-ratio 440
5.2 Chi-square Goodness··of-fit Test 442
5.3 Test of the Equality of Two Multinomial Distributions
and Generalizations 448
5.4 Tests of Independence in Contingency Tables 452
6 Tests of Hypotheses and Confidence Intervals 461
7 Sequenti'al Tests of Hypotheses 464
CONTENTS xi
7.2 Definition of Sequential Probability Ratio Test 466

7.3 Approximate Sequential Probability Ratio Test 468
7.4 Approximate Expected Sample Size of Sequential Probability
Ratio Test 470
X Linear Models 482

2 Examples of the Linear Model 483
3 Definition of Linear Model 484
4 Point Estimation~Case A 487
5 Confidence Intervals-Case A 491
6 Tests of Hypotheses-Case A 494
7 Point Estimation-Case B 498
XI N onparametric Methods 504

2 Inferences Concerning a Cumulative Distribution Function 506
2.1 Sample or Empirical Cumulative Distribution Function 506
2.2 Kolmogorov-Smirnov Goodness-of-fit Test 508
2.3 Confidence Bands for Cumulative Distribution Function 511
3 Inferences Concerning Quantiles 512
3.1 Point and Interval Estimates of a Quantile 512
3.2 Tests of Hypotheses Concerning Quantiles 514
4 Tolerance Limits 515
5 Equality of Two Distributions 518
5.2 Two-sample Sign Test 519
5.3 Run Test 519
5.4 Median Test 521
5.5 Rank-sum Test 522
Appendix A. Mathematical Addendum 527

1 Introduction 527
xii CONTENTS
2 Noncalculus 527
2.1 Summation and Product Notation 527
2.2 Factorial and Combinatorial Symbols and Conventions 528
2.3 Stirling's Formula 530
2.4 The Binomial and Multinomial Theorems 530
3 Calculus 531
3.1 Preliminaries 531
3.2 Taylor Series 533
3.3 The Gamma and Beta Functions 534
Appendix B. Tabular Summary of Parametric Families

of Distributions 537
1 Introduction 537
Table 1. Discrete Distributions 538
Table 2. Continuous Distributions 540
Appendix C. References and Related Reading 544

Mathematics Books 544
Probability Books 544
Probability and Statistics Books 545
Advanced (more advanced than MGB) 545
Intermediate (about the same level as MGB) 545
Elementary (less advanced than MGB, but calculus
prerequisite) 546
Special Books 546
Papers 546
Books of Tables 547
Appendix D. Tables 548

1 Description of Tables 548
Table 1. Ordinates of the Normal Density Function 548
Table 2. Cumulative Normal Distribution 548
Table 3. Cumulative Chi-square Distribution 549
Table 4. Cumulative F Distribution 549
Table 5. Cumulative Student's t Distribution 550
Index 557
PREFACE TO THE
THIRD EDITION
The purpose of the third edition of this book is to give a sound and self-con-
tained (in the sense that the necessary probability theory is included) introduction
to classical or mainstream statistical theory. It is not a statistical-methods-
cookbook, nor a compendium of statistical theories, nor is it a mathematics
book. The book is intended to be a textbook, aimed for use in the traditional
full-year upper-division undergraduate course in probability and statistics,
or for use as a text in a course designed for first-year graduate students. The
latter course is often a "service course," offered to a variety of disciplines.
No previous course in probability or statistics is needed in order to study
the book. The mathematical preparation required is the conventional full-year
calculus course which includes series expansion, mUltiple integration, and par-
tial differentiation. Linear algebra is not required. An attempt has been
made to talk to the reader. Also, we have retained the approach of presenting
the theory with some connection to practical problems. The book is not mathe-
matically rigorous. Proofs, and even exact statements of results, are often not
given. Instead, we have tried to impart a "feel" for the theory.
The book is designed to be used in either the quarter system or the semester
system. In a quarter system, Chaps. I through V could be covered in the first
xiv PREFACE TO THE THIRD EDITION
quarter, Chaps. VI through part of VIII the second quarter, and the rest of the
book the third quarter. In a semester system, Chaps. I through VI could be
covered the first semester and the remaining chapters the second semester.
Chapter VI is a " bridging" chapter; it can be considered to be a part of" proba-
bility" or a part of" statistics." Several sections or subsections can be omitted
without disrupting the continuity of presentation. For example, any of the
following could be omitted: Subsec. 4.5 of Chap. II; Subsecs., 2.6, 3.5, 4.2, and
4.3 of Chap. III; Subsec. 5.3 of Chap. VI; Subsecs. 2.3, 3.4, 4.3 and Secs. 6
through 9 of Chap. VII; Secs. 5 and 6 of Chap. VIII; Secs. 6 and 7 of Chap. IX;
and all or part of Chaps. X and XI. Subsection 5.3 of Chap VI on extreme-value
theory is somewhat more difficult than the rest of that chapter. In Chap. VII,
Subsec. 7.1 on Bayes estimation can be taught without Subsec. 3.4 on loss and
risk functions but Subsec. 7.2 cannot. Parts of Sec. 8 of Chap. VII utilize matrix
notation. The many problems are intended to be essential for learning the
material in the book. Some of the more difficult problems have been starred.
ALEXANDER M. MOOD
FRANKLIN A. GRAYBILL
DUANE C. BOES
EXCERPTS FROM THE FIRST
AND SECOND EDITION PREFACES
This book developed from a set of notes which I prepared in 1945. At that time
there was no modern text available specifically designed for beginning students
of matpematical statistics. Since then the situation has been relieved consider-
ably, and had I known in advance what books were in the making it is likely
that I should not have embarked on this volume. However, it seemed suffi-
ciently different from other presentations to give prospective teachers and stu-
dents a useful alternative choice.
The aforementioned notes were used as text material for three years at Iowa
State College in a course offered to senior and first-year graduate students.
The only prerequisite for the course was one year of calculus, and this require-
ment indicates the level of the book. (The calculus class at Iowa State met four
hours per week and included good coverage of Taylor series, partial differentia-
tion, and multiple integration.) No previous knowledge of statistics is assumed.
This is a statistics book, not a mathematics book, as any mathematician
will readily see. Little mathematical rigor is to be found in the derivations
simply because it would be boring and largely a waste of time at this level. Of
course rigorous thinking is quite essential to goocf statistics, and I have been at
some pains to make a show of rigor and to instill an appreciation for rigor by
pointing out various pitfalls of loose arguments.
XVI EXCERPTS FROM THE FIRST AND SECOND EDITION PREFACES
While this text is primarily concerned with the theory of statistics, full
cognizance has been taken of those students who fear that a moment may be
wasted in mathematical frivolity. All new subjects are supplied with a little
scenery from practical affairs, and, more important, a serious effort has been
made in the problems to illustrate the variety of ways in which the theory may
be applied.
The probl~ms are an essential part of the book. They range from simple
numerical examples to theorems needed in subsequent chapters. They include
important subjects which could easily take precedence over material in the text;
the relegation of subjects to problems was based rather on the feasibility of such
a procedure than on the priority of the subject. For example, the matter of
correlation is dealt with almost entirely in the problems. It seemed to me in-
efficient to cover multivariate situations twice in detail, i.e., with the regression
model and with the correlation model. The emphasis in the text proper is on
the more general regression model.
The author of a textbook is indebted to practically everyone who has
touched the field, and I here bow to all statisticians. However, in giving credit
to contributors one must draw the line somewhere, and I have simplified matters
by drawing it very high; only the most eminent contributors are mentioned in
the book.
I am indebted to Catherine Thompson and Maxine Merrington, and to
E. S. Pearson, editor of Biometrika, for permission to include Tables III and V,
which are abridged versions of tables published in Biometrika. I am also in-
debted to Professors R. A. Fisher and Frank Yates, and to Messrs. Oliver and
Boyd, Ltd., Edinburgh, for permission to reprint Table IV from their book
" Statistical Tables for Use in Biological, Agricultural and Medical Research."
Since the first edition of this book was published in 1950 many new statis-
tical techniques have been made available and many techniques that were only in
the domain of the mathematical statistician are now useful and demanded by
the applied statistician. To include some of this material we have had to elim-
inate other material, else the book would have come to resemble a compendium.
The general approach of presenting the theory with some connection to prac-
tical problems apparently contributed significantly to the success of the first
edition and we have tried to maintain that feature in the present edition.
I
PROBABILITY
1 INTRODUCTION AND SUMMARY
The purpose of this chapter is to define probability and discuss some of its prop-
erties. Section 2 is a brief essay on some of the different meanings that have
been attached to probability and may be omitted by those who are interested
only in mathematical (axiomatic) probability, which is defined in Sec. 3 and
used throughout the remainder of the text. Section 3 is subdivided into six
subsections. The first, Subsec. 3.1, discusses the concept of probability models.
It provides a real-world setting for the eventual mathematical definition of
probability. A review of some of the set theoretical concepts that are relevant
to probability is given in Subsec. 3.2. Sample space and event space are
defined in Subsec. 3.3. Subsection 3.4 commences with a recall of the definition
of a function. Such a definition is useful since many of the words to be defined
in this and coming chapters (e.g., probability, random variable, distribution,
etc.) are defined as particular functions. The indicator function, to be used
extensively in later chapters, is defined here. The probability axioms are pre-
sented, and the probability function is defined. Several properties of this prob-
ability function are stated. The culmination of this subsection is the definition
of a probability space. Subsection 3.5 is devoted to examples of probabilities
2 PROBABIUTY I
defined on finite sample spaces. The related concepts of independence of u

events and conditional probability are discussed in the sixth and final subsection.
Bayes' theorem, the mUltiplication rule, and the theorem of total probabilities
are proved or derived, and examples of each are given.
Of the three main sections included in this chapter, only Sec. 3, which is
by far the longest, is vital. The definitions of probability, Pfobabiijly space,
conditional probability, and independence, along with familiarity with the
properties of probability, conditional and unconditional and related formulas,
are the essence of this chapter. This chapter is a background chapter; -it intro-
duces the language of probability to be used in developing distribution theory,
which is the backbone of the theory of statistics.
2 KINDS OF PROBABILITY
2.1 Introduction
One of the fundamental tools of statistics is probability, which had its formal
beginnings with games of chance in the seventeenth century.
Games of chance, as the name implies, include such actions as spinning a
roulette wheel, throwing dice, tossing a coin, drawing a card, etc., in which th~
outcome of a trial is uncertain. However, it is recognized that even though the
outcome of any particular trial may be uncertain, there is a predictable' ibng-
term outcome. It is known, for example, that in many throws of an ideal
(balanced, symmetrical) coin about one-half of the trials will result in heads.
It is this long-term, predictable regularity that enables gaming houses to engage
in the business.
A similar type of uncertainty and long-term regularity often occurs in
experimental science. For example, in the science of genetics it is uncertain
whether an offspring will be male or female, but in the long run it is known
approximately what percent of offspring will be male and what percen't will be
female. A life insurance company cannot predict which persons in the United
States will die at age 50, but it can predict quite satisfactorily how many people
in the United States will die at that age. ~
First we shall discuss the classical, or a priori, theory of probability; then
we shall discuss the frequency theory. Development of the axiomatic approach
will be deferred until Sec. 3.
2 KINDS OF PROBAMUl'Y 3
2.2 Classical or A Priori Probability

As we stated in the previous subsection, the theory of probability in its early
stages was closely associated with games of chance. This association prompted
the classical definition. For example, suppose that we want the probability of
the even.t·1hat an ideal coin will turn up heads. We argue in this manner: Since
_ ~ >-a(
there are only two ways that the coin can fall, heads or tails, and since the coin
is well balanced, one would expect that the coin is just as likely to fall heads as
tails; h~nce, the probability of the event of a head will be given the value t·
This kind of reasoning prompted the following classical definition of prob-
ability.
Definition 1 Classical probability If a random experiment can resul t

in n mutually exclusive and equally likely outcomes and if nA. of these
outcomes have an attribute A, then the probability of A is the fraction
nA./n. 1/1/
We shall apply this definition to a few examples in order to illustrate its meaning.
~ If an ordinary die (one ofa pair of dice) is tossed-there are six possible out-
comes-anyone of the six numbered faces may turn up. These six outcomes
are mutually exclusive since two or more faces cannot turn up simultaneously.
··And if the die is fair, or true, the six outcomes are equally likely; i.e., it is expected
that each face will appear with about equal relative frequency in the long run.
blow suppose that we want the probability that the result of a toss be an even
number. Three of the six possible outcomes have this attribute. The prob-
ability t~at an even number will appear when a die is tossed is therefore i, or t.
-l;imilarly, the probability that a 5 will appear when a die is tossed is 1. The
probability that the result of a toss will be greater than 2 is i-
To consider another example, suppose that a card is drawn at random from
an ordinary deck of playing cards. The probability of drawing a spade is
readily seen to be ~ ~, or i· The probability of drawing a number between 5
and lO~ inclusive, is ; ~, or t'3·
The application of the definition is straightforward enough in these simple
cases, but it is not always so obvious. Careful attention must be paid to the
q~<!1ifications "mutually exclusive," equally likely," and random." Suppose
H H
that one wishes to compute the probability of getting two heads if a coin is
tossed twice. He might reason that there are three possible outcomes for the
two tosses: two "heads, two tails, or one head and one tai 1. One of these three
·.
4 PROBABIUTY' I
outcomes has the desired attribute, i.e., two heads; therefore the probability is
-1. This reasoning is faulty because the three given outcomes are not equally
likely. The third outcome, one head and one tail, can OCCUr in two ways
since the head may appear on the first toss and the tail on the second or the
head may appear on the second toss and the tail on the first. Thus there are
four equally likely outcomes: HH, HT, TH, and TT. The first of these has
the desired attribute, while the others do not. The correct probability is there-
fore i. The result would be the same if two ideal coins were tossed simul-
taneously.
Again, suppose that one wished to compute the probability that a card
drawn from an ordinary well-shuffied deck will be an ace or a spade. In enu-
merating the favorable outcomes, one might count 4 aces and 13 spades and
reason that there are 17 outcomes with the desired attribute. This is clearly
incorrect because these 17 outcomes are not mutually exclusive since the ace of
spades is both an ace and a spade. There are 16 outcomes that are favorable to
an ace or a spade, and so the correct probability is ~ ~, or 143'
We note that by the classical definition the probability of event A is a
number between 0 and 1 inclusive. The ratio nAln must be less than or equal to
1 since the total number of possible outcomes cannot be smaller than the
number of outcomes with a specified attribute. If an event is certain to happen,
its probability is 1; if it is certain not to happen, its probability is O. Thus, the
probability of obtaining an 8 in tossing a die is O. The probability that the
number showing when a die is tossed is less than 10 is equal to 1.
The probabilities determined by the classical definition are called a priori
probabilities. When one states that the probability of obtaining a head in
tossing a coin is !-, he has arrived at this result purely by deductive reasoning.
The result does not require that any coin be tossed or even be at hand. We say
that if the coin is true, the probability of a head is !, but this is little more than
saying the same thing in two different ways. Nothing is said about how one
can determine whether or not a particular coin is true.
The fact that we shall deal with ideal objects in developing a theory of
probability will not troub1e us because that is a common requirement of mathe-
matical systems. Geometry, for example, deals with conceptually perfect
circles, lines with zero width, and so forth, but it is a useful branch of knowl-
edge, which can be applied to diverse practical problems.
There are some rather troublesome limitations in the classical, or a priori,
approach. It is obvious, for example, that the definition of probability must
be modified somehow when the total number of possible outcomes is infinite.
One might seek, for example, the probability that an integer drawn at random
from the positive integers be even. The intuitive answer to this question is t.
2 KINDS OF PROBABIUTY 5
If one were pressed to justify this result on the basis of the definition, he might
reason as fol1ows: Suppose that we limit ourselves to the first 20 integers; 10
of these are even so that the ratio of favorable outcomes to the total number is
t%, or 1· Again, if the first 200 integers are considered, 100 of these are even,
and the ra~io is also 1_ In genera!, the first 2N integers contain N even integers;
if we form the ratio N/2N and let N become infinite so as to encompass the whole
set of positive integers, the ratio remains 1_ The above argument is plausible,
and the answer is plausible, but it is no simple matter to make the argument
stand up. It depends, for example, on the natural ordering of the positive
integers, and a different ordering could produce a different result. Thus, one
could just as well order the integers in this way: 1,3,2; 5, 7,4; 9, 11,6; ___ ,
taking the first pair of odd integers then the first even integer, the second pair
of odd integers then the second even integer, and so forth. With this ordering,
one could argue that the probability of drawing an even integer is 1-. The
integers can also be ordered so that the ratio will oscillate and never approach
any definite value as N increases.
There is another difficulty with the classical approach to the theory of
probability which is deeper even than that arising in the case of an infinite
num ber of outcomes. Suppose that we toss a coin known to be biased in
favor of heads (it is bent so that a head is more likely to appear than a tail).
The two possible outcomes of tossing the coin are not equally likely. What is
the probability of a head? The classical definition leaves us completely helpless
here.
Still another difficulty with the classical approach is encountered when we
try to answer questions such as the following: What is the probability that a
child born in Chicago will be a boy? Or what is the probability that a male
will die before age 50? Or what is the probability that a cookie bought at a
certain bakery will have less than three peanuts in it? All these are legitimate
questionswhichwewantto bring into the realm of probability theory. However,
notions of symmetry," "equally likely," etc., cannot be utilized as they could
H
be in games of chance. Thus we shall have to alter or extend our definition to

bring problems similar to the above into the framework of the theory. This
more widely applicable probability is called a posteriori probability, or!requency,
and will be discussed in the next subsection.
2.3 A Posteriori or Frequency Probability

A coin which seemed to be well balanced and symmetrical was tossed 100 times,
and the outcomes recorded in Table 1. The important thing to notice is that the
relative frequency of heads is close to 1. This is not unexpected since the coin
6 PROBABILITY I
was symmetrical, and it was anticipated that in the long run heads would occur
about one-half of the time. For another example, a single die was thrown 300
times, and the outcomes recorded in Table 2. Notice how close the relative
frequency of a face with a I showing is to !; similarly for a 2, 3, 4, 5, and 6.
These results are not unexpected since the die which was used was quite sym-
metrical and balanced; it was expected that each face would occur with about
equal frequency in the long run. This suggests that we might be willing to use
this relative frequency in Table 1 as an approximation for the probability that
the particular coin used will come up heads or we might be willing to use the
relative frequencies in Table 2 as approximations for the probabilities that
various numbers on this die will appear. Note that although the relative fre-
quencies of the different outcomes are predictable, the actual outcome of an
individual throw is unpredictable.
In fact, it seems reasonable to assume for the coin experiment that there
exists a number, label it p, which is the probability of a head. Now if the coin
appears well balanced, symmetrical, and true, we might use Definition 1 and
state that p is approximately equal to 1. It is only an approximation to set p
equal to 1- since for this particular coin we cannot be certain that the two cases,
.-
heads and tails, are exactly equally likely. But by examining the balance and
symmetry of the coin it may seem quite reasonable to assume that they are.
Alternatively, the coin could be tossed a large number of times, the results
recorded as in Table 1, and the relative frequency ofa head used as an approxima-
tion for p. In the experiment with a die, the probability P2 of a 2 showing
could be approximated by using Definition I or by using the relative frequency
in Table 2. The important thing is that we postulate that there is a number P
which is defined as the probability of a head with the coin or a number P2
which is the probability of a 2 showing in the throw of the die. Whether we use
Definition 1 or the relative frequency for the probability seems unimportant in
the examples ci ted.
Table 1 RESULTS OF TOSSING A COIN 100 TIMES
Long-run expec~ed
Observed Observed relative relative frequency
Outcome Frequency frequency of a balanced coin
H 56 .56 .50
T 44 .44 .50
Total 100 1.00 1.00

2 KINDS OF PROBABIUTY 7
Suppose, as described above, that the coin is unbalanced so that we are

quite certain from an examination that the two cases, heads and tails, are not
equally likely ,to happen. In these cases a number p can still be postulated
as the probabilit y that a head shows, but the classical definition will not help us
to find the value of p. We must use the frequency approach or possibly some
physical arialysis of the unbalanced coin.
In many scientific investigations, observations are taken which have an ele-
ment of uncertainty or unpredictability in them. As a very simple example, sup-
pose that we want to predict whether the next baby born in a certain locality will
be a male or a female. This is individually an uncertain event, but the results of
groups of births can be dealt with satisfactorily. We find that a certain long-
run regularity exists which is similar to the long-run regularity of the frequency
ratio of a head when a coin is thrown. If, for example, we find upon examination
of records that about 51 percent of the births are male, it might be reasonable to
postulate that the probability of a male birth in this locality is equal to a number
p and take .51 as its approximation.
To make this idea more concrete, we sha11 assume that a series of observa-
tions (or experiments) can be made under quite uniforlJ1 conditions. That is,
an observation of a random experiment is made; then the experiment is repeated
under similar conditions, and another observation taken. This is repeated
many times, and while the conditions are similar each time, there is an uncon-
trollable variation which is haphazard or random so that the observations are
individually unpredictable. In many of these cases the observations fall into
certain classes wherein the relative frequencies are quite stable. This s~ggests
that we postulate a number p, called the probability of the event, and approximate
p by the ~elative frequency with which the repeated observations satisfy the
Table 2 RESULTS OF TOSSING A DIE 300 TIMES
Long-run expected
Observed Observed relative frequency
Outcome Frequency relative frequency of a balanced die
1 51 .170 .1667
2 54 .180 .1667
3 48 .160 .1667
4 51 .170 .1667
5 49 .163 .1667
6 47 .157 .1667
Tota1 300 1.000 1.000
8 PROBABILITY I
event. For instance, suppose that the experiment consists of sampling the
population of a large city to see how many voters favor a certain proposal.
The outcomes are" favor" or "do not favor," and each voter's response is un-
predictable, but it is reasonable to postulate a number p as the probability that
a given response will be "favor." The relative frequency of" favor" responses
can be used as an approximate value for p.
As another example, suppose that the experiment consists of sampling
transistors from a large collection of transistors. We shall postulate that the
probability of a given transistor being defective is p. We can approximate p by
selecting several transistors at random from the collection and computing the
relative frequency of the number defective.
The important thing is that we can conceive of a series of observations or
experiments under rather uniform conditions. Then a number p can be postu-
lated as the probability of the event A happening, and p can be approximated by
the relative frequency of the event A in a series of experiments.
3 PROBABILITY-AXIOMATIC
3.1 Probability Models

One of the aims of science is to predict and describe events in the world in which
we live. One way in which this is done is to construct mathematical models
which adequately describe the real world. For example, the equation s = tgt 2
expresses a certain relationship between the symbols s, g, and t. It is a mathe-
matical model. To use the equation s = !gt 2 to predict s, the distance a body
falls, as a function of time t, the gravitational constant g must be known. The
latter is a physical constant which must be measured by experimentation if the
equation s = tgt 2 is to be usefuL The reason for mentioning this equation is
that we do a similar thing in probability theory; we construct a probability
model which can be used to describe events in the real world. For example, it
might be desirable to find an equation which could be used to predict the Sex of
each birth in a certain locality. Such an equation would be very complex, and
none has been found. However, a probability model can be constructed which,
while not very helpful in dealing with an individual birth, is quite useful in
dealing with groups of births. Therefore, we can postulate a number p which
represents the probability that a birth will be a male. From this fundamental
probability we can answer questions such as: What is the probability that in
ten births at least three will be males? Or what is the probability that there will
be three consecutive male births in the next five? To answer questions such as
these and many similar ones, we shall develop an idealized probability model.
3 PROBABILlTY-AXIOMATIC 9
The two general types of probability (a priori and a posteriori) defined

above have one important thing in common: They both requir~ a conceptual
experiment in which the various outcomes can occur under somewhat uniform
conditions. For example, repeated tossing of a coin for the a priori case, and
repeated birth for the a posteriori case. However, we might like to bring into
the realm of probability theory situations which cannot conceivably fit into the
framework of repeated outcomes under somewhat similar conditions. For
example, we might like to answer questions such as: What is the probability my
wife loves me? Or what is the probability that World War III will start before
January 1, 1985? These types of problems are certainly a legitimate part of
general probability theory and are included in what is referred to as subjective
probability. We shall not discuss subjective probability to any great extent in
this book, b~t we remark that the axioms of probability from which we develop
probability theory are rich enough to include a priori probability, a posteriori
probability, and subjective probability.
To start, we require that every possible outcome of the experiment under
study can be enumerated. For example, in the coin-tossing experiment there are
two possible outcomes: heads and tails. We shall associate probabilities only
with these outcomes or with collections of these outcomes. We add, however,
that even if a particular outcome is impossible, it can be included (its probability
is 0). The main thing to remember is that every outcome which can occur
must be included.
Each conceivable outcome of the conceptual experiment under study will be
defined as a sample point, and the totality of conceivable outcomes (or sample
points) will be defined as the sample space.
Our .object, of course, is to assess the probability of certain outcomes or
collections of outcomes of the experiment. Discussion of such probabilities
is conveniently couched in the language of set theory, an outline of which
appears in the next subsection. We shaH return to formal definitions and
examples of sample space, event, and probability.
3.2 An Aside-Set Theory

We begin with a collection of objects. Each object in our collection will be
called a point or element. We assume that our collection of objects is large
enough to include all the points under consideration in a given discussion.
The totality of all these points is called the space, universe, or universal set.
We will call it the space (anticipating that it will become the sample space when
we speak of probability) and denote it by n. Let (1) denote an element or point
i"n n. Although a set can be defined as any collection of objects, we shall
10 PROBABIliTY I
assume, unless otherwise stated, that all the sets mentioned in a given discussion
consist of points in the space n.
EXAMPLE 1 Q = R 2 , where R2 is the collection of points co in the plane and

co = (x, y) is any pair of real numbers x and y. IIII
EXAMPLE 2 n = {all United States citizens}. IIII
We shall usually use capital Latin letters from the beginning of the
alphabet, with or without subscripts, to denote sets. If co is a point or element
belonging to the set A, we shall write co E A; if co is not an element of A, we
shall write co ¢; A.
Definition 2 Subset If every element of a set A is also an element of a

set B, then A is defined to be a subset of B, and we shall write A c B or
B => A; read" A is contained in B" or " B contains A." II1/
Definition 3 Equivalent sets Two sets A and B are defined to be equiva-
lent, or equal, if A c Band B c A.. This will be indicated by writing
A =B. /111
Definition 4 Empty set If a set A contains no points, it \'vill be called
the null set, or empty set, and denoted by </>. /1//
Definition S Complement The complement of a set A with respect to
the space n, denoted by A, A or n - A, is the set of all points that are in
C
,
Q but not in A. 1//1

Definition 6 Union Let A and B be any two subsets of n; then the
set that consists of all points that are in A or B or both is defined to be
the union of A and B and written A vB. 1///
Definition 7 Intersection Let A and B be any two subsets of n; then
the set that consists of all points that are in both A and B is defined to be
the intersection of A and B and is written A (\ B or AB. / / //
Definition 8 Set difference Let A and B be any two subsets of n. The

set of all points in A that are not in B wHl be denoted by A - B and is
defined as set difference. II//
3 PROBABILITY-AXIOMATIC 11
EXAMPLE 3 Let n = {(x, y): 0 <x ~ 1 and 0 <y< I}, which is read the
collection of all points (x, y) for which 0 < x < 1 and 0 < y < 1. Define
the following sets:
Ai = {(x, y): 0 < x < 1; 0 < Y < t},

A2 = {(x, y): 0 < x <!; 0 <y< I},
A3 = {(x, y): 0 < x < y < l},
A4 = {(x, y): 0 <x<!; 0 <y< !}.
(We shall adhere to the practice initiated here of using braces to embrace
the points of a set.)
The set relations below follow.
A1 (\ A2 = A1A2 = A4;
A1 ={(x,y):O<x< l;!<y< I};
Al - A4 = {(x, y): !<x< I; 0 < Y < !}. IIII
EXAMPLE 4 Let n, A 1, A 2 , and A3 be as indicated in the diagrams in Fig. I

which are called Venn diagrams. IIII
The set operations of complement, union, and intersection have been

defined in Definitions 5 to 7, respectively. These set operations satisfy quite a
number of laws, some of which follow, stated as theorems. Proofs are omitted.
Theorem 1 Commutative laws A u B = B u A and A (\ B = B (\ A.

IIII
Theorem 2 Associative laws A u (B u C) = (A u B) u C, and
A (\ (B (\ C) = (A (\ B) (\ C. IIII
Theorem 3 Distributive laws A (\ (B u C) = (A (\ B) u (A (\ C), and

A u (B (\ C) = (A u B) (\ (A.u C). IIII
Theorem 4 (Ac)C = (1) = A; in words, the complement of A comple-

ment equals A. IIII
Theorem 5 An = A; A u n = n; A4> = 4>; and A u 4> = A. IIII
Theorem 6 AA = 4>; A u A = n; A (\ A = A; and A u A = A. IIII
12 PROBABILITY I
n is the entire square
FIGURE 1
Theorem 7 (A u B) = A n 13, and (A n B) = A u 13. These are known

as De Morgan's laws. IIII
Theorem 8 A - B = AB. Jill
Several of the above laws are illustrated in the Venn diagrams in Fig. 1.
Although we will feel free to use any of the above laws, it might be instructive
to give a proof of one of them just to illustrate the technique. For example,
let us show that (A u B) = A n B. By definition, two sets are equal if each is
contained in the other. We first show that (A u B) cAn 13 by proving that if
OJ E A u B, then OJ E A n 13. Now OJ E (A u B) implies OJ ¢ A u B, which implies
that OJ ¢ A and OJ ¢ B, which in turn implies that OJ E A and OJ E 13; that is,
OJ E A n 13. We next show that A nBc (A u B). Let OJ E A n 13, which means
OJ belongs to both A and 13. Then OJ ¢ A u B for if it did, OJ must belong to at
least one of A or B, contradicting that OJ belongs to both A and 13; however,
OJ ¢ A u B means OJ E (A u B), completing the proof.
We defined union and intersection of two sets; these definitions extend

immediately to more than two sets, in fact to an arbitrary number of sets. It
is customary to distinguish between the sets in a collection of subsets of n by
assigning names to them in the form of subscripts. Let A (Greek letter capital
lambda) denote a catalog of names, or indices. A is also called an index set.
For example, if we are concerned with only two sets, then our index set A
includes only two indices, say I and 2; so A = {I , 2}.
Definition 9 Union and intersection of sets Let A be an index set and

{AA,= XE A} = {AA,}' a collection of subsets of n indexed by A. The set
of points that consists of all points that belong to AA, for at least one A is
called the union of the sets {AA,} and is denoted by UAA,' The set of points
A. E A
that consists of all points that belong to A A, for every A is called the inter-
section of the sets {AA,} and is denoted by
A.
n
AA,' If A is empty, then define
A
nAA, = n.
E
UAA, = ljJ and

A.EA A.EA
IIII
EXAMPLE 5 If A = {1, 2, ... , N}, i.e., A is the index set consisting of the
first N integers, then U
AA, is also written as
A.eA
N
U An = A 1 U
n=l
A2 U ... U AN' IIII
14 PROBABILlTY I
One of the most fundamental theorems relating unions, intersections, and

complements for an arbitrary collection of sets is due to De Morgan.
Theorem 9 De Morgan's theorem Let A be an index set and {A;J a

collection of subsets of il indexed by A. Then,
IIII
We wi1l not give a proof of this theorem. Note, however, that the special
case when the index set A consists of only two names or indices is Theorem 7
above, and a proof of part of Theorem 7 was given in the paragraph after
Theorem 8.
Definition 10 Disjoint or mutually exclusive Subsets A and B of Q are

defined to be mutually exclusive or disjoint if A (') B = l/J. Subsets
AI, A 2 , ••• are defined to be mutually exclusive if AjAj = l/J for every i =# j.
IIII
Theorem 10 If A and B are subsets of Q, then (i) A = ABu AB, and

(ii) AB (') AB = l/J.
(i) A = A () Q
PROOF = A () (B u B) = AB u AB. (ii) AB (') AB
=AABB=Al/J = l/J. M
Theorem 11 If A c B, then AB = A, and A u B = B.
PROOF Left as an exercise. IIII
3.3 Definitions of Sample Space and Event

In Subsec. 3.1 we described what might be meant by a probability model.
There we said that we had in mind some conceptual experiment whose possible
outcomes we would like to study by assessing the probability of certain outcomes
or collection of outcomes. In this subsection, we win give two important
definitions, along with some examples, that will be used in assessing these
probabilities.
Definition 11 Sample space The sample space, denoted by il, is the

collection or totality of all possible outcomes of a conceptual experiment.
IIII
One might try to understand the definition by looking at the individual

words. Use of the word "space" can be justified since the sample space is the
total collection of objects or elements which are the outcomes of the experiment.
This is in keeping with our use of the word U space" in set theory as the collec-
tion of all objects of interest in a given discussion. The word ~'sample" is
harder to justify; our experiment is random, meaning that its outcome is un-
certain so that a given outcome is just one sample of many possible outcomes.
Some other symbols that are used in other texts to denote a sample space,
in addition to Q, are S, Z, R, E, X, and A.
Definition 12 Event and event space An event is a subset of the sample

space. The class of all events associated with a given experiment is
defined to be the event space. 1//1
The above does not precisely define what an event is. An event will
always be a subset of the sample space, but for sufficiently large sample spaces
not all subsets will be events. Thus the class of all subsets of the sample space
will not necessarily correspond to the event space. However, we shall see that
the class of all events can always be selected to be large enough so as to include
all those subsets (events) whose probability we may want to talk about. If the
sample space consists of only a finite number of points, then the corresponding
event space will be the class of all subsets of the sample space.
Our primary interest will not be in events per se but will be in the prob-
ability that an event does or does not occur or happen. An event A is said to
occur if the experiment at hand results in an outcome (a point in our sample
space) that belongs to A. Since a point, say OJ, in the sample space is a subset
(that subset consisting of the point OJ) of the sample space n, it is a candidate to
be an event. Thus OJ can be viewed as a point in n or as a subset of n. To
distinguish, let us write {OJ}, rather than just OJ, whenever (J) is to be viewed as a
subset of n. Such a one-point subset will always be an event and will be called
an elementary event. Also if> and n are both subsets of Q, and both will always
be events. n is so metimes called the sure event.
We shall attempt to use only capital Latin letters (usually from the begin-
ning of the alphabet), with or without affixes, to denote events, with the excep-
tion that if> will be used to denote the empty set and n the sure event. The event
space will always be denoted by a script Latin letter, and usually d. ;?4 and /F,
as well as other symbols, are used in some texts to denote the class of all events.
The sample space is basic and generally easy to define for a given experi-
ment. Yet, as we shall see, it is the event space that is really essential in de-
fining probabiHty. Some examples follows.
16 PROBABILITY I
EXAMPLE 6 The experiment is the tossing of a single die (a regular sjx-sided

polyhedron or cube marked on each face with one to six spots) and noting
which face is up. Now the die can land with anyone of the six faces up;
so there are six possible outcomes of the experiment;
n = { [:J, G:J, I.··', 1::1, (:-:" nil}.

Let A = {even number of spots up}. A is an event; it is a subset of n.
A = {0, @], [ll)}. Let Ai = {i spots up}; i = ], 2, ... , 6. Each Ai is an
elementary event. For this experiment the sample space is finite; hence
the event space is all subsets of n. There are 26 = 64 events, of which only
6 are elementary, in J7I (including both q, and n). See Example 19 of
Subsec. 3.5, where a technique for counting the number of events in a finite
sample space is presented. 1///
EXAMPLE 7 Toss a penny, nickel, and dime simultaneously, and note which
side is up on each. There are eight possible outcomes of this experiment.
n ={(H, H, H), (H, H, T), (H, T, H), (T, H, H), (H, T, n,
(T, H, n,
(T, T, H), (T, T, T)}. We are using the first position of (', " .), called a
3-tuple, to record the outcome of the penny, the second position to record
the outcome of the nickel, and the third position to record the outcome ot
the dime. Let Ai = {exactly i heads}; i = 0, I, 2, 3. For each i, Ai is an
event. Note that Ao and A3 are each elementary events. Again all
subsets of n are events; there are 2 8 = 256 of them. IIII
EXAMPLE 8 The experiment is to record the number of traffic deaths in the

state of Colorado next year. Any nonnegative integer is a conceivable
outcome of this experiment; so n = {O, I, 2, ... }. A = {fewer than 500
deaths} = {O, I, ... , 499} is an event. Ai = {exactly i deaths}, i = 0, 1,
... , is an elementary event. There is an infinite number of points in the
sample space, and each point is itself an (elementary) event; so there is an
infinite number of events. Each subset of n is an event. III/
EXAMPLE 9 Select a light bulb, and record the time in hours that it burns
before burning out. Any nonnegative number is a conceivable outcome
of this experiment; so n = {x: x >O}. For this sample space not an
subsets of Q are events; however, any subset that can be exhibited will be
an event. For example, let
A = {bulb burns for at least k hours but burns out before m

hours}
= {x: k < x < m};
then A is an event for any 0 < k < m. II1I
EXAMPLE J0 Consider a random experiment which consists of counting

the number of times that it rains and recording in inches the total rainfall
next July in Fort Collins, Colorado. The sample space could then be
represented by
Q = {(I, x): i = 0, 1, 2, . .. and 0 x},
where in the 2-tuple (', .) the first position indicates the number of times
that it rains and the second position indicates the total rainfall. For
example, OJ = (7, 2.251) is a point in Q corresponding to there being seven
different times that it rained with a total rainfall of 2.251 inches. A =
{(i, x): i = 5, ... , 10 and x 3} is an example of an event. IIII
EXAMPLE 11 In an agricultural experiment, the yield of five varieties of

wheat is examined. The five varieties are all grown under rather uniform
conditions. The outcome is a collection of five numbers (Yl ,Y2 ,Y3 'Y4 ,Ys),
where Yi represents the yield of the ith variety in bushels per acre. Each
Yi can conceivably be any real number greater than or equal to O. In this
example let the event A be defined by the conditions that Y2, Y3' Y4' and
Ys are each 10 or more bushels per acre larger than Yl, the standard
variety. In our notation we write
Our definition of sample space is precise and satisfactory, whereas our

definitions of event and event space are not entirely satisfactory. We said that
if the sample space was" sufficiently large" (as in Examples 9 to 11 above), not all
subsets of the sample space would be events; however, we did not say exactly
which subsets would be events and which would not. Rather than developing
the necessary mathematics to precisely define which subsets of Q constitute OUf
18 PROBABILITY I
event space d, let us state some properties of d that it seems reasonable to

.
requIre:
(i) QEd.
(ii) If A E d, then A E d.
(iii) If Al and A2 Ed, then Al U A2 Ed.
We said earlier that we were interested in events mainly because we would

be interested in the probability that an event happens. Surely, then, we would
want d to include Q, the sure event. Also, if A is an event, meaning we can
talk about the probability that A occurs, then A should also be an event so that
we can talk about the probability that A does not occur. Similarly, if Al and
A2 are events, so should Al U A2 be an event.
Any collection of events with properties (i) to (iii) is caned a Boolean
algebra, or just algebra, of events. We might note that the collection of all
subsets of Q necessarily satisfies the above properties. Several results follow
from the above assumed properties of d.
Theorem 12 t/> E d.
PROOF By property (i) QEd; by (ii) QEd; but Q = t/>; so t/> Ed.
1111
Theorem 13 If Al and A2 Ed, then Ai () A2 Ed.
PROOF Al and A2 Ed; hence Al U A2 , and (AI U A 2) Ed, but

(-=A:--1-U-A='2) = Al () 12 = Al () A2 by De Morgan's law. 1111
U Ai and In1 Ai• E d.

n n
Theorem 14 If AI' A2 , ... , An Ed, then
1=1
PROOF Follows by induction. 1111
We will always assume that our collection of events d is an algebra-

which partially justifies our use of d as our notation for it. In practice, one
might take that collection of events of interest in a given consideration and
enlarge the collection, if necessary, to include (i) the sure event, (ii) all comple-
ments of events already included, and (iii) all finite unions and intersections of
events already included, and thIS will be an algebra d. Thus far, we have not
explained why d cannot always be taken to be the collection of all subsets of Q.
Such explanation will be given when we define probability in the next subsection.
3 P:aOBABILITY-AXIOMATIC 19
3.4 Definition of Probability

In this section we give the axiomatic definition of probability. Although this
formal definition of probability will not in itself allow us to achieve our goal :>f
assigning actual probabilities to events consisting of certain outcomes of random
experiments, it is another in a series of definitions that will ultimately lead to
that goal. Since probability, as well as forthcoming concepts, is defined as a
particular function, we begin this subsection with a review of the notion of a
function.
The definition of a function The following terminology is frequently used

to describe a function: A function, say f('), is a rule (law, formula, recipe) that
associates each point in one set of points with one and only one point in another
set of points. The first collection of points, say A, is caned the domain, and the
second collection, say B, the counterdomain.
Definition 13 Function A function, say f( .), with domain A and coun-

terdomain B, is a collection of ordered pairs, say (a, b), satisfying (i) a E A
and b E B; tii) each a E A occurs as the first element of some ordered pair
in the collection (each bE B is not necessarily the second element of some
ordered pair); and (iii) no two (distinct) ordered pairs in the collection
have the same first element. IIII
If (a, b) Ef( .), we write b = f(a) (read" b equals f of a") and call f(a)
the value of f(·) at a. For any a E A, f(a) is an element of B; whereasf( -) is
a set of Qrdered pairs. The set of all values of f( .) is called the range of f(· );
i.e., the range of f(') = {b E B: b = f(a) for some a E A} and is always a subset
of the counterdomain B but is not necessarily equal to it. f(a) is also called the
image of a under f( +), and a is called the pre image of /(a).
EXAMPLE 12 Let.ft(-) andf2(') be the two functions, having the real line
for their domain and counterdomain, defined by
fi ( .) = {(x, y) : y = x 3 + X + 1, - 00 < x < oo}
and
i
f2(' ) = {(x, y): y = x 2, - 00 < x < oo}.

The range offi ( .) is the counterdomain, the whole real line, but the range
of f2(') is all nonnegative real numbers, not the same as the counter-
domain_ Jill
20 . PROBABILITY I
Of particular interest to us will be a class of functions that are called

indicator functions.
- Definition 14 Indicator function Let Q be any space with points W

and A any subset of Q. The indicator function of A, denoted by [A('),
is the function with domain Q and counterdomain equal to the set consist-
ing of the two real numbers 0 and 1 defined by
if WE A
if W ¢ A.
[A(') clearly U indicates" the set A. IIII
Properties of Indicator Functions Let Q be any space and d any collection
of subsets of Q:
(i) [A(W) = 1 - [A(W) for every A E d.
(ii) IAIA2'''A,,(W) = IAt(w)' I A2(w)'" IA,,(w) for'Aj, ... , An E d.
(iii) I Al VA2V ... VAn(W) = max [lAt( w), I A2( w), •.. , I An(W)] for AI' ... ,
An E d.
(iv) [~(w) [A(W) for every A E d.
Proofs of the above properties are left as an exercise.
The indicator function will be used to "indicate" subsets of the real line;
e.g.,
I([o.I))(X) = lto •• lx) = {~ if 0 <x < 1

otherwise,
and if [+ denotes the set of positive integers,
if x is some positive integer
otherwise.
Frequent use of indicator functions will be made throughout the remainder
of this book. Often the utility of the indicator function is just notational
efficiency as the following example shows.
EXAMP LE ] 3 Let the function I( .) be defined by

o for x<O
for O<x<l
I{x) = x
2-x for l<x<2
o for 2 < x.
3 PROBABIUTY-AXIOMATIC 21
By using the indicator function, f(x) can be written as

f(x) = x/(o, l](x) + (2 - x)/(l, 2lx ).
or also by using the absolute value symbol as
f(x) = (1 -II - x 1)/(0. 2](X). IIII
Another type of function that we will.have occasion to discuss is the set

function defined as any functon which has as its domain a collection of sets
and as its counterdomain the real line including, possibly, infinity. Examples
of set functions follow.
EXAMPLE 14 Let 0 be the sample space corresponding to the experiment of

tossing two dice, and let d be the collection of all subsets of O. For
any A E d define N(A) = number of outcomes, or points in 0, that are
in A. Then N( <p) = 0, N(O) = 36, and N(A) = 6 if A is the event con-
taining those outcomes having a total of seven spots up. If1/
The size-oJ-set function alluded to in the above example can be defined. in

general, for any set A as the number of points in A, where A is a member of an
arbitrary collection of sets d.
EXAMPLE 15 Let 0 be the plane or two-dimensional euclidean space and

d any collection. of subsets of 0 for which area is meaningful. Then
for any A Ed define Q(A) = area of A. For example, if A {(x, y); 0
< x <1,0 <y < I}, then Q(A) = 1; if A = {(x, y): x 2 + y2 r2}, then
Q(A) = rcr 2 ; and if A = {(O. 0), (1, I)} then Q(A) = O. IIII
The probability function to be defined will be a particular set function.
Probability function Let 0 denote the sample space and d denote a collec-
tion of events assumed to be an algebra of events (see Subsec. 3.3) that we shall
consider for some random experiment.
22 PROBABILITY I
(I Definition 15 Probability function A probability function P[· ] is a set

function with domain d (an algebra of events)* and counterdomain the
interval [0, 1] which satisfies the following axioms:
(i) P[A] > 0 for every A Ed.

(ii) prO] = L
(iii) If AI. A 2 • ••• is a sequence of mutually exclusive events in d
(that is, Ai (l Aj = 4> for i =1= j; i,j = 1,2, ... ) and if Ai U A2 U ... =
IIII
These axioms are certainly motivated by the definitions of classical and

frequency probability. This definition of probability is a mathematical defini-
tion; it tells us which set functions can be called probability functions; it does not
tell us what value the probability function P[·] assigns to a given event A. We
will have to model our random experiment in some way in order to obtain values
for the probability of events.
P[A] is read" the probability of event A" or "the probability that event A
occurs," which means the probability that any outcome in A occurs.
We have used brackets rather than parentheses in our notation for a
probability function, and we shall continue to do the same throughout the
remainder of this book.
*In defining a probability function, many authors assume that the domain of the set
function is a sigma-algebra rather than just an algebra. For an algebra d, we had the
property
if A1 and A2 E d, then Al U A2 E d.
A sigma-algebra differs from an algebra in that the above property is replaced by
00
if A tt A 2 ..... A n •... Ed, then U AnEd.
If'" 1
It can be shown that a sigma-algebra is an algebra, but not necessarily ronversely.

If the domain of the probability function is taken to be a sigma-algebra then axiom
(iii) can be simplified to
if A 1 , A 2 • ••• is a sequence of mutually exclusive events in d. p[ ~ At] = f P[AtJ.

1~1 1=1
A fundamental theorem of probability theory. called the extension theorem. states that
if a probability function is defined on an algebra (as we have done) then it can be
extended to a sigma-algebra. Since the probability function can be extended from an
algebra to a sigma-algebra, it is reasonable to begin by assuming that the probability
function is defined on a sigma-algebra.
3 PROBABILITY-AXlOMATIC 23
EXAMPLE 16 Consider the experiment of tossing two coins, say a penny and
a nickel. Let n {(H, H), (H, n,
(T, H), (T, T)} where the first com-
ponent of (', .) represents the outcome for the peJ}ny. Let us model this
random experiment by assuming that the four points in n are equally
likely; that is, assume P[{(H, H)}] = P[{~fI, T)}] = P[{(T, H)}] =
P[{(T, T)}]. The following question arises: Is the P[· ] function that is
implicitly defined by the above really a probability function; that is, does
it satisfy the three axioms? It can be shown that it does, and so it is
a probability function.
In our definitions of event and .PI, a collection of events, we stated that .PI
cannot always be taken to be the collection of all subsets of n. The reason for
this is that for" sufficiently large" n the collection of all subsets of n is so large
that it is impossible to define a probability function consistent with the above
aXIOms.
We are able to deduce a number of properties of our function P[· ] from
its definition and three axioms. We list these as theorems.
I t is in the statements and proofs of these properties that we will see the
convenience provided by assuming.PI is an algebra of events. .PI is the domain
of P[·]; hence only members of .PI can be placed in the dot position of the
notation P[· ]. Since .PI is an algebra, if we assume that A and BE .PI, we know
that A, A u B, AB, AB, etc., are also members of .PI, and so it makes sense to
talk about P[A], P[A u B], P[AB], P[AB], etc.
Properties of P[·] For each of the following theorems, assume that nand
.PI (an algebra of events) are given and P[·] is a probability function having
domain .PI.
Theorem 15 P[4>] = o.
PROOF Take Al = 4>, A2 = 4>, A3 = ¢, ... ; then by axiom (iii)
P[q,] =P [,Q Ai] = f~P[Ai] = .t,P[q,],

which can hold only if P[4>] = o. 111/
Theorem 16 If A., ... , An are mutually exclusive events in .PI, then
n
P[A 1 U ... U An] = L P[Ad.
t::::: 1
24 PROBABILITY I
II <Xl
PROOF Let A II + 1 = ¢, A II + 2 = ¢, ... , ; then U Ai = U Ai Ed,

i=l i=::l
and
IIII
Theorem 17 If A is an event in d, then

P[A] =1- P[A].
PROOF A u A= Q, and A (l A = ¢; so
P[Q] = P[A u A] = P[A] + P[A].
But P[!l] = 1 by axiom (ii); the result follows. IIII
Theorem 18 If A and BEd, thenP[A] = P[AB] + P[AB], and P[A - B]

= P[AB] = P[A] - P[AB].
PROOF A = AB u AB, and AB (l AB = ¢; so P[A] = P[AB] +

P[AB]. Ilfl
Theorem 19 For every two events A and BEd, P[A u B] = P[A]

+ P[B] - P[AB]. More generally, for events Ab A 2 , ... , All E d
II
P[A 1 U A2 U •. , u All] = L P[A j ] - Li<jL P[AiAj]

j== 1
+ L L L P[AiAjAk] - ... + (- 1)"+ 1 P[A 1 A 2 ... An]·

i<j<k
PROOF A u B =A u AB, and A (l AB = ¢; so

P[A u B] = P[A] + P[AB]
= P[A] + P[B] - P[AB].
The more general statement is proved by mathematical induction. (See

Problem 16.) Ilfl
Theorem 20 If A and BEd and A c: B, then P[A] < P[B].
PROOF B = BA u BA, and BA = A; so B = A u BA, and A ( l BA =

4>;
hence P[B] = P[A] + P[BA]. The conclusion follows by noting that
P[BA] > 0 . ' fill
3 PROBABILITY-AXIOMATIC 2S
Theorem 21 Doole's inequality If Ab A z , ... , A" E,d, then

P[A 1 u' A2 U ... u A,,] < P[Ad + P[A z ] + ... + P[A,,].
PROOF P[A 1 U A 2] = P[A 1 ] + P[A z ] - P[A 1A z ] < P[Ad + P[A z ]'
The proofis completed using mathematical induction. IIII
We conclude this sUbsection with one final definition.
Definition 16 Probability space A probability space is the triplet

(0, d, P[·]), where 0 is a sample space, d is a collection (assumed to be
an algebra) of events (each a subset of 0), and P[·] is a probability func-
tion with domain d. IIII
Probability space is a single term that gives us an expedient way to assume

the existence of all three components in its notation. The three components are
related; d is a collection of subsets of 0, and P[·] is a function that has d as its
domain. The probability space's main use. is in providing a convenient method
of stating background assumptions for future definitions, theorems, etc. It also
ties together the main definitions that we have covered so far, namely, definitions
of sample space, event space, and probability.
3.5 Finite Sample Spaces

In previous subsections we formally defined sample space, event, and probability,
culminating in the definition of probability space. We remarked there that these
formal definitions did not in themselves enable us to compute the value of the
probability-for an event A, which is our goal. We said that we had to appro-
priately model the experiment. In this section we show how this can be done
for finite sample spaces, that is, sample spaces with only a finite number of
elements or points in them.
In certain kinds of problems, of which games of chance are notable
examples, the sample space contains a finite number of points, say N = N(O).
[Recall that N(A) is the size of A, that is, the number of sample points in A.]
Some of these problems can be modeled by assuming that points in the sample
space are equally likely. Such problems are the subject to be discussed next.
Finite sample space with equa1ly likely points For certain random
experiments there is a finite number of outcomes, say N, and it is often realistic
to assume that the probability of each outcome is liN. The classical definition
of probability is generally adequate for these problems, but we shall show how
26 PROBABILITY I
the axiomatic definition is applicable as well. Let W h W 2 , ••• , wN be the N

sample points in a finite space O. Suppose that the set function P[·] with
domain the collection of all subsets of 0 satisfies the following conditions:
(i) P[{w t }] = P[{w 2 }] = ... = P[{WN}]'
(ii) If A is any subset of 0 which contains N(A) sample points [has size
N(A)], then P[A] = N(A)IN.
Then it is readily checked that the set function P[·] satisfies the three axioms
and hence is a probability function.
Definition 17 Equally likely probabIlity fUDction The probability func-

tion P[· ] satisfying conditions (i) and (ii) above is defined to be an equally
likely probability function. IIII
Given that a random experiment can be realistically modeled by assuming

equally likely sample points, the only problem left in determining the value of
the probability of event A is to find N(O) = Nand N(A). Strictly speaking this
is just a problem of counting-count the number of points in A and the number
of points in O.
EXAMPLE 17 Consider the experiment of tossing two dice (or of tossing one
die twice). Let 0 = {(ii' i 2): il = 1,2, ... , 6; i2 = 1,2, ... , 6}. Here i1 =
number of spots up on the first die, and ;2 = number of spots up on the sec-
ond die. There are 6'6 = 36 sample points. It seems reasonable to attach
the probability of l6 to each sample point. 0 can be displayed as a lattice
as in Fig. 2. Let "A 7 = event that the total is 7; then A7 = {(I, 6), (2, 5),
(3,4), (4, 3), (5, 2), (6, I)}; so N(A7) = 6, and P[A 7] = N(A 7 )/N(O) = 36(j =
7;. Similarly P[Aj] can be calculated for Aj = total ofj;j = 2, ... , 12. In
this example the number of points in any event A can be easily counted,
and so P[A] can be evaluated for any event A. IIII
If N(A) and N(O) are large for a given random experiment with a finite
number of equally likely outcomes, the counting itself can become a difficult
problem. Such counting can often be facilitated by use of certain combinatorial
formulas, some of which will be developed now.
Assume now that the experiment is of such a nature that each outcome
can be represented by an n-tuple. The above example is such an experiment;
each outcome was represented by a 2-tuple. As another example, if the ex-
periment is one of drawing a sample of size n, then n-tuples are particularly
3 PROBABIUTY-AXIOMATIC 27
6 • • • •
5 • • •
4 • • •
3 • • •
: : : _.__________~
:1~- ·;1
FIGURE 2 1 2 3 4 5 6
useful in recording the results. The terminology that is often used to describe
a basic random experiment known generally by sampling is that of balls and urns.
It is assumed that we have an urn containing, say, M balls, which are numbered
1 to M. The experiment is to select or draw balls from the urn one at a time
until n balls have been drawn. We say we have drawn a sample of size n. The
drawing is done in such a way that at the time of a particular draw each of the
balls in the urn at that time has an equal chance of selection. We say that a
ball has been selected at random. Two basic ways of drawing a sample ~are
with replacement and without replacement, meaning just what the words say. A
sample is said to be drawn with replacement, if after each draw the ball drawn
is itself returned to the urn, and the sample is said to be drawn without replace-
ment if the ball drawn is not returned to the urn. Of course, in sampling without
replacement the size of the sample n must be less than or equal to M, the original
number of balls in the urn, whereas in sampling with replacement the size of
sample may be any positive integer. In reportjng the results of drawing a sample
of size n, an n-tuple can be used; denote the n-tuple by (Zh ••. , zn), where Zi
represents the number of the ball drawn on the ith draw.
In general, we are interested in the size of an event that is composed of
points that are n-tuples satisfying certain conditions. The size of such a set can be
compute.d as follows; First determine the number of objects, say N I , that may be
used as the first component. Next determine the number of objects, say N 2 ,
that may be used as the second component of an n-tuple given that the first com-
ponent is known. (We are assuming that N2 does not depend on which
object has occurred as the first component.) And then determine the number of
objects, say N 3 , that may be used as the third component given that the first
and second components are known. (Again we are assuming N3 does not
28 PROBABILITY I
depend on which objects have occurred as the first and second components.)
Continue in this manner until N n is determined. The size N(A) of the set A of
n-tuples then equals Nl • N2 ... N n •
EXAMPLE 18 The total number of different ordered samples of n balls that

can be obtained by drawing balls from an urn containing M distinguish-
able balls (distinguished by numbers 1 to M) is M n if the sampling is done
with replacement and is M(M - 1) ... (M - n + 1) if the sampling is
done without replacement. An ordered sample can be represented by
an n-tuple, say (Zl,"" zn), where zJis the number of the ball obtained on the
jth draw and the total number of different ordered samples is the same as
the total number of n-tuples. In sampling with replacement, there are M
choices of numbers for the first component, M choices of numbers for the
second component, and finally M choices for the nth component. Thus
there are M n such n-tuples. In sampling without replacement, there are
M choices of numbers for the first component, M - 1 choices for the
second, M - 2 choices for the third, and finally M - n + 1 choices for
the nth component. In total, then, there are M(M - 1)(M - 2) ...
(M - n + 1) such n-tuples. M(M - 1) ... (M - n + 1) is abbreviated
(M)n (see Appendix A). II11
EXAMPLE 19 Let S be any set containing M elements. How many subsets

does Shave? First let us determine the number of subsets of size n that
S has. Let Xn denote this number, that is, the number of subsets of S of
size n. A subset of size n is a collection of n objects, the objects not
arranged in any particular order. For example the subset {Sl1 ss, J7} is
the same as the subset {s 5, S1, S7} since they contain the same three objects.
If we take a given subset of S which contains n elements, n! different
ordered samples can be obtained by sampling from the given subset
without replacement. If for each of the Xn different subsets there are n!
different ordered samples of size n, then there are (n !)xn different ordered
samples of size n in sampling without replacement from the set S of M
elements. But we know from the previous example that this number is
(M)n; hence (n!)xn = (M)n, or
Xn (M)n = (M) = number of subsets of size n that may be formed

n! n
from the elements of a set of size M. (1)
3 PROBABlUTY-AXIOMATIC 29
.
~ n .
The total number of subsets of S, where S is a set of SIze M, IS i..J . (M)
"=0
This includes the empty set (set with no elements in it) and the whole set,
both of which are subsets. Using the binomial theorem (see Appendix
A)
with a = b = 1,
we see that
2M = JJ~); (2)
thus a set of size M has 2M subsets. 11/1
EXAMPLE 20 Suppose an urn contains M balls numbered 1 to M, where the

first K balls are defective and the remaining M - K are nondefective.
The experiment is to draw n balls from the urn. Define AA; to be the event
that the sample of n balls contains exactly k defectives. There are two
ways to draw the sample: (i) with replacement and (ii) without replace-
ment. We are interested in P[A k ] under each method of sampling. Let
our sample space Q = {(z., . .. , zJ: zJ = number of the ball drawn on the
jth draw}. Now
P[A] _ N(AJ
k - N(Q)'
From Example 18 above, we know N(Q) = M"under (i) and N(Q) = (M)n
under {ii). AA; is that subset of Q for which exactly k of the z/s are ball
numbers 1 to K inclusive. These k ball numbers must fall in some subset
of k positions from the total number of n available positions. There are
(:) ways of selecting the k positions for the ball numbers 1 to K inclusive to
fall in. For each of the (~) different positions, there are K.(M - K)·-k
different n-tuples for case (i) and (K)k(M - K)n A; different n-tuples for
case (ii). Thus A. has size (Z)Kk(M - K)" • for case (i) and size
(Z)(Kl.{M - K).-. for case (ii); so,

30 PROBABIUTY I
in sampling with replacement, and
(~)(K).(M - K)._. (4)

P[A k ]=
. .
)
(M n
in sampling without replacement. This latter formula can be rewritten

as
(!)(~=:)
P[A.l = ( '!:) . (5)
It might be instructive to derive Eq. (5) in another way. Suppose

that our sample space, denoted now by 0', is made up of subsets of size
n, rather than n-tuples; that is, 0' = {{Zb ... , zn}: Zl' ..• , Zn are the numbers
on the n balls drawn}. There are ('!:) subsets of size n of the M balls;
so N(fi') = (~) If it is assumed that each of these subsets of size n is

just as likely as any other subset of size n (one can think of selecting all n
balls at once rather than one at a time), then P[Ak] = N(A~)/N(O'). Now
N(A~) is the size of the event consisting of those subsets of sizenwhichcon-
tain exactly k balls from the balls that are numbered I to K inclusive.
The k balls from the balls that are numbered I to K can be selected in
(!) ways, and the remaining n - k balls from the balls that are numbered
K + I to M can be selected in (M-K)

n_ k ways; hence N(AiJ =
(!)(~ := n, and finally
P[A.] = N(AJ/N(fi') = (!) (~:= n.

(:J
We have derived the probability of exactly k defectives in sampling
without replacement by considering two different sample spaces; one
sample space consisted of n-tuples, the other consisted of subsets of size
n.
To aid in remembering the formula given in Eq. (5), note that
K + M - K = M and k + n - k = n; i.e., the sum of the "upper" terms
3 PROBABlUfV-AXIOMATlC 31
in the numerator equals the" upper" term in the denominator, and the
sum of the" lower" terms in the numerator equals the" lower" term
in the denominator. // / /
EXAMPLE 21 The formula given in Eq. (5) is particularly useful to calculate

certain probabilities having to do with card games. For example, we
might ask the probability that a certain 13-card hand contains exactly
6 spades. There are M = 52 total cards, and one can model the card
shuffling and dealing process by assuming that the 13-card hand represents
a sample of size 13 drawn without replacement from the 52 cards. Let
A6 denote the event of exactly 6 spades. There are a total of 13 spades
(defective balls in sampling terminology); so
13) (52 - 13)

(6 13 - 6
PIA
6
] = (ii)
by Eq. (5). IIII
Many other formulas for probabilities of specified events defined on finite

sample spaces with equally likely sample points can be derived using methods of
combinatorial analysis, but we will not undertake such derivations here. The
interested reader is referred to Refs. 10 and 8.
Finite sample space without equally likely points We saw for finite sample
spaces with equally likely sample points that P[A] = N(A)IN(o.) for any event A.
For finite sample spaces without equally likely sample points, things are not
quite as simple, but we can completely define the values of P[A] for each of the
2N (Q) events A by specifying the value of P[· ] for each of the N = N(o.) elemen-
tary events. Let 0. = {WI' ... , WN}, and assume Pj =P[{Wj}] for j = 1, ... , N.
Since
I = p[n] pL0/en j}]

= = it/[{en)].
N
L Pj =
j="'l
1.
For any event A, define P[A] = J:.Pj' where the summation is over those Wj
belonging to A. It can be shown that P[· ] so defined satisfies the three axioms
and hence is a probability function.
32 PROBABILITY I
EXAMPLE 22 Consider an experiment that has N outcomes, say (01; (02' ••• ,
(ON' where it is known that outcome (OJ+l is twice as likely as outcome
(OJ' wherej = 1, ... , N - 1; that is, Pj+ 1 = 2Pi' where Pi P[{(Ol)]' Find
prAll, where A.k = {(Oh (02' ••• , (Ok}' Since
N N
L Pj = j:::::l
j=1
L 21- 1
PI = Pl(1 + 2 + 22 + ... + 2N - 1
) = Pl(2 N - 1) = 1,
1
Pi = 2N -1
and
hence
C> 3.6 Conditional Probability and Independence

In the application of probability theory to practical problems it is not infrequent
that the experimenter is confronted with the following situation: Such and such
has happened; now what is the probability that something else will happen?
For example, in an experiment of recording the life of a light bulb, one might
be interested in the probability that the bulb will last 100 hours given that it
has already lasted for 24 hours. Or in an experiment of sampling from a box
containing 100 resistors of which 5 are defective, what is the probability that
the third draw results in a defective given that the first two draws resulted in
defectives? Probability questions of this sort are considered in the framework
of conditional probability, the subject that we study next.
Conditional probability We begin by assuming that we have a probability

space, say (n, d, P[· ]); that is, we have at hand some random experiment for
which a sample space n, collection of events d, and probability function
P[ . ] have all been defined.
Given two events A and B, we want to define the conditional probability
of even"t A given that event B has occurred.
Definition 18 Conditional probability Let A and B be two events in d

of the given probability space (n, d, P[· ]). The conditional probability
of event A given event B, denoted by P[A IB], is defined by
P[A IB] = ~AB] if P[B] > 0, (6)

P[B]
and is left undefined if P[B] = O. 1111
Remark A formula. that is evident from the definition is P[AB] =

P[A IB]P[B] = P[BI A]P[A] if both P[A] and P[B] are nonzero. This
formula relates P[A IB] to P[BI A] in terms of the unconditional prob-
abilities P[A] and P[B]. IIII
We might note that the above defmition is compatible with the frequency
approach to probability, for if one observes a large number, say N, of occur-
rences of a random experiment for which events A and B are defined, then
P[A IB] represents the proportion of occurrences in which B occurred that A also
occurred, that is,
P[A IB] = NAB,
NB
where N B denotes the number of occurrences of the eve~t B in the N occur-
rences of the random experiment and NAB denotes the number of occurrences
of the event A n B in the N occurrences. Now P[AB] = NABIN, and P[B] =
NBIN; so
P[AB] = NABIN = NAB = P[AIB]
P[B] NBIN NB '
consistent with our definition.
EXAMPLE 23 Let 0 be any finite sample space, d the collection of all subsets
of 0, and P[·] the equally likely probability function. Write N = N(O).
For events A and B,
P[AIB] = P[AB] = N(AB)IN
P[B] N(B)! N '
where, as usual, N(B) is the size of set B. So for any finite sample space
with equally likely sample points, the values of P[A IB] are defined for any
two events A and B provided P[B] > O. IIII
EXAMPLE 24 Consider the experiment of tossing two coins. Let 0 =

{(H, H), (H', T), (T, H), (T, T)}, and assume that each point is equally
likely. Find (i) the probability of two heads given a head on the first
coin and (ii) the probabiUty of two heads given at least one head. Let
Al = {head on first coin} and A2 = {head on second coin}; then the prob-
ability of two heads given a head on the first coin is
P[A A IA = =!
] = P[AIA2Ad = P[A 1 A 2] t
1 2 1 P[A d P[Ad! 2.
34 PROBABILITY I
The probability of two heads given at least one head is

P[AIA2 (\ (A~l U 04 2)] P[A 1 A 2] ! 1
P [ AiA2 1 Ai U A2 ] = = -
P[A 1 U A 2] P[A 1 u A 2] i 3
We obtained numerical answers to these two questions, but to do so we
had to model the experiment; we assumed that the four sample points
were equally likely.
When speaking of conditional probabilities we are conditioning on some

given event B; that is, we are assuming that the experiment has resulted in some
outcome in B. B, in effect, then becomes our" new" sample space. One ques-
tion that might be raised is: For given event B for which P[B] > 0, is P['I B] a
probability function having d as its domain? In other words, does P[· IB]
satisfy the three axioms? Note that:
(i) P[A IB] = P[AB]IP[B] > 0 for every A Ed.
(ii) P[OI B] = P[OB]IP[B] = P[B]IP[B] = 1.
(iii) If AI' A 2 , ••• is a sequence of mutually exclusive events in .91 and
00
U Ai Ed, then
i=:l
Hence, P['I B] for given B satisfying P[B] > 0 is a probability function, which
justifies our calling it a conditional probability. P[· IB] also enjoys the same
properties as the unconditional probability. The theorems listed below are
patterned after those in Subsec. 3.4.
Properties of Pl· IB] Assume that the probability space (n, .91, P[·]) is given,
and let BEd satisfy P[B] > O.
Theorem 22 P[q,IB] O. IIII
Theorem 23 If At, ... , An are mutually exclusive events in .91, then

n
PlAt U ... U Ani B] = i;;;;;;I 1
P[Ad B]. IIII
Theorem 24 If A is an event in d, then

P[AI B] = 1- P[A I B]. IIII
3 PROBABIUTY-AXlOMATIC 3S
Theorem 25 If At and A2 E d, then

P[Ad B] = P[A1A2IB] + P[A t A21B]. ////
Theorem 26 For every two events At and A2 E d,

PlAt u A2IB] = P[A1IB] +P[A 2IB] - p[A1A2IB]. /1//
Theorem 27 If A1 and A2 E d and Al C A2 , then

P[AII B] < P[A21 B]. 1//1
Theorem 28 If A 1, A 2 , ... , An Ed, then

II
P[A I U A2 U ••• U AnI B ] < LP[AIIB]. 1/11

)=1
Proofs of the above theorems follow from known properties of P[·] and
are left as exercises.
There are a number of other useful formulas involving conditional prob-
abilities that we will state as theorems. These will be followed by examples.
Theorem 29 Theorem of total probabilities For a given probability

space (n, d, P[· ]), if B1, B 2 , ••• , Bn is a collection of mutually disjoint
n
events in d satisfying n= UB
)=1
j and P[Bj ] > O· for j = 1, "" n, then
n
for- every A E d, P[A] = L P[A IBj]P[Bj ].
j=1
n
PROOF Note that A = U ABj
j= 1
and the AB/s are mutually disjoint;
hence
1//1
Corollary For a given probability space (n, d, P[·]) let BEd

satisfy 0 < P[B] < 1; then for every A Ed
P[A] = P[AIB]P[B] + P[AIB]P[B]. //1/
Remark Theorem 29 remains true if n = 00. 1///

36 PROBABILITY I
Theorem 29 (and its corollary) is particularly useful for those experiments

that have stages; that is, the experiment consists of performing first one thing
(first stage) and then another (second stage). Example 25 provides an example
of such an experiment; there, one first selects an urn and then selects a ball
from the selected urn, For such experiments, if Bj is an event defined only in
terms of the first stage and A is an event defined in terms of the second stage,
then it may be easy to find P[Bj ]; also, jt may be easy to find peA IBj ], and then
Theorem 29 evaluates peA] in terms of P[Bj ] and peA IB j ] for j = 1, ... , n, In
an experiment consisting of stages it is natural to condition on results of a first
stage.
Theorem 30 Bayes' formula For a given probability space (0, .91, Pl· D,
if B 1 , B 2 , ••• , Bn is a collection of mutually disjoint events in .91 satisfying
n
n = U Bj and P[Bj ] > 0 for j = 1, ... , n, then for every A Ed for which
j:::l
peA] >0
P[B"IA] = nP[AIBk]P[Bk ] •
2: P[AIBj]P[B
j= 1
j]
PROOF
, ..
by using both the definition of conditional probability and the theorem of

total pro babi Ii ties. IIII
Corollary For a given probability space (0, d, P[·]) let A and BE d
satisfy peA] > 0 and 0 < PCB] < I ; then
P B/ _ P[A/B]P[B]
[ A] - P[A' B}P[B] + peA IB]P[B]' 1I11
Remark Theorem 30 remains true if n = 00. 11II

As was the case with the theorem of total probabilities, Bayes' formula is
also particularly useful for those experiments consisting of stages. If Bj ,
j = 1, ... , n, is an event defined in terms of a first stage and A is an event defined
in terms of the whole experiment including a second stage, then asking for
P[BkIA] is in a sense backward; one is asking for the probability of an event
3 PROBABILITY-AXJOMATIC 37
defined in terms of a first stage of the experiment conditioned on what happens

in a later stage of the experiment. The natural conditioning would be to con-
dition on what happens in the first stage of the experiment, and this is precisely
what Bayes' formula does; it expresses P[Bkl A] in terms of the natural con-
ditioning given by P[AIB j ] and P[Bj],j= 1, ... , n.
Theorem 31 Multiplication rule For a given probability space

(0, d, P[· D, let AI,"" An be events belonging to d for which
P[A I •• , An- tl > 0; then
P[A 1 A 2 ••• An] = P[AdP[A 2 1 AtlP[A 3 1 A 1 A 2 ] ••• PlAn IAl ... An-d·
PROOF The proof can be attained by employing mathematical

induction and is left as an exercise. If/I
As with the two previous theorems, the multiplication rule is primarily

useful for experiments defined in terms of stages. Suppose the experiment has
n stages and A J is an event defined in terms of stage j of the experiment; then
.f~f[Ajl AIA2 .. , Aj-d is the conditional probability of an event described in
. terms of what happens on stage j conditioned on what happens on stages
1, 2, ... , j - 1. The multiplication rule gives PlAt A 1 ..• An] in terms of
the natural conditional probabilities P[A j IA 1A 2 '" Aj-tl forj= 2, ... , n.
EXAMPLE 25 There are five urns, and they are numbered I to 5. Each
urn contains 10 balls. Urn i has i defective balls and 10 - i nondefective
balls, i = 1, 2, ... , 5. For instance, urn 3 has three defective balls and
seven nondefective balls. Consider the following random experiment:
First an urn is selected at random, and then it ball is selected at random
from the selected urn. (The experimenter does not know which urn was
selected.) Let us ask two questions: (i) What is the probability that a
defective ball will be selected? (ii) If we have already selected the ball
and noted that it is defective, what is the probability that it came from
urn 5?
SOLUTION Let A denote the event that a defective ball is selected and
B t the event that urn i is selected, i = I, ... , 5. Note that P[B,l = ~,
i 1, ... ,5, andP[AIBi ] = i/lO, i= 1, ... , 5. Question (i) asks, What is
P[A]? Using the theorem of total probabilities, we have
5 5 ill 5. 1 5·6 3
peA] = i~lP[ArB,]p[Ba = l~ 10'"5 = 50 t~l' = 50"2 = 10'
38 PROBABILITY I
Note that there is a total of 50 balls of which 15 are defective! Question

(ii) asks, What is P[Bsl A]? Since urn 5 has more defective balls than
any of the other urns and we selected a defective ball, we suspect that
P[Bsl A] > P[B i IA] for i = 1, 2, 3, or 4. In fact, we suspect P[Bsl A]
> P[B41 A] > ... > P[BII A]. Employing Bayes' formula, we find
P[BsIA] = ;[AIBs]P[Bs] =t~!=~.

I P[A IBtJP[BtJ To 3
i= I
Similarly,
P[B IA] = (kilO) .t = !5... k = 1, ... , 5,
k ~ 15'
10
substantiating our suspicion. Note that unconditionally all the B/s were
equally likely whereas, conditionally (conditioned on occurrence of event
A), they were not. Also, note that
s skI s 1 5·6
IP[BkIA]
k=l
= k=115
I -=- Ik=--=l.
15k=l 152
IIII
EXAMPLE 26 Assume that a student is taking a multiple-choice test. On a

given question, the student either knows the answer, in which case he
answers it correctly, or he does not know the answer, in which case he
guesses hoping to guess the right answer. Assume that there are five
multiple-choice alternatives, as is often the case. The instructor is con-
fronted with this problem: Having observed that the student got the
correct answer, he ,wishes to know what is the probability that the student
knew the answer. Let p be the probability that the stydent will know the
answer and 1 - P the probability that the student guesses. Let us assume
that the probability that the student gets the right answer given that he
guesses is t.
(This may not be a realistic assumption since even though the
student does not know the right answer, he often would know that certain
alternatives are wrong, in which case his probability of guessing correctly
should be better than ~.) Let A denote the event that the student got the
right answer and B denote the event that the student knew the right answer.
We are seeking P[BI A]. Using Bayes' formula, we have
P[A IB]P[B] 1 .p
P[BI A] = P[A IB]P[B] + P[A IB]P[B] = 1 . P + t(1 - p)"
Note that
p+t~I-P)";::.P. IIII
3 PROBABlUTY-AXIOMATIC 39
EXAMPLE 27 An urn contains ten balls of which three are black and seven
are white. The following game is played: At each trial a ball is selected
at random, its color is noted, and it is replaced along with two additional
balls of the same color. What is the probability that a black ball is
selected in each of the first three trials? Let Bi denote the event that a
black ball is selected on the ith trial. We are seeking P[B1 B 2 B3]' By the
mul tiplication r"ule,
P[B 1 B 2 B 3] = P[B 1 JP[B 2 I'BdP [B 3 IB 1B 2 ] = 130 • /2 • 174 = /6' IIII

\ '
EXAMPLE 28 Suppose an urn contains M balls of which K are black and

M - K are white. A sample of size n is drawn. Find the probability
that the jth ball drawn is black given that the sample contains k black
balls. (We intuitively expect the answer to be kin.) We have to con-
sider sampling (i) with replacement and (ii) without replacement.
SOLUTION Let Ak denote the event that the sample contains exactly
k black balls and Bj denote the event that the jth ball drawn is black.
We seek P[Bjl Ak]' Consider (i) first.
P[AkJ =
n) Kk(M - K)"-k
(k and P[A kIBJ.J =
(n - 1) K k- 1
(M _ K)"-k
1
M(I k- 1 M"-
by Eq. (3) of Subsec. 3.5. Since the balls are replaced, P[Bj ] = KIM for
any j. Hence,
For case (ii),
(~)(~ ~ ~) (K- 1)(M - K)

and P[A IB.J = k - 1 n - k
P[ A.J = ( '::) k J(M _ 1)
n- 1
j-1
by Eq. (5) of Subsec. 3.5. P[Bj ] = L P[Bjl CdP[Ci ] , where C i denotes
i == 0
the event of exactly i black ba~ls in the firstj - 1 draws. Note that
40 PROBABILITY I
and
K-i
P[B., C.]
J t
=- --
M-j+l'
and so
Finally,
PCB _I A]
,'P[A kI Bj]P[BJ =
[(K -
k- 1
l)(M -
n- k
K)j(M -
n- 1
1)] K
Ai = ~.
J k P[Akl (~)('~ =f) j(~) .
n
Thus we obtain the same answer under either method of sampling. "t-'/IIF, "
Independence of events If P[A IB] does not depend on evem: B, that is,
P[A, B] = P[A], then it would seem natural to say that event A is independent
of event B. This is given in the following definition.
Definition 19 Independent events For a given probability space

(0, &/, P[·]), let A and B be two events in.!il. Events A and Bare
defined to be independent if and only if anyone of the following conditions
is satisfied:
(i) P[AB] = P[A]P[B].
(ii) P[A, B] = P[A] if P[B] > O.
(iii) P[B IA] P[B] if PtA] > O. IIII
Remark Some authors use" statistically independent," or "stochasti-

cally independent," instead of" independent." /111
To argue the equivalence of the above three conditions, it suffices to show
that (i) implies (il), Oi) implies (iii), and (iii) implies (i). If P[AB] P[A]P[B],
then P[A, B] = P[AB]IP[B] = P[A]P[B]fP[B] = P[A] for P[B] > 0; so (i) implies
(ii). If PtA IB] = P[A], then P[BI A] = P[A, B]P[B]IP[A] = P[A]P[B]IP[A] =
P[B] for P[A] > 0 and P[B] > 0; so (ii) implies (iii). And if P[BI A] = P[B],
then P[AB] = P[B,A]P[A] = P[B]P[A] for P[A] > O. Clearly P[AB] =
P[A]P[B] if P[A] = 0 or P[B] = o.
EXAMPLE 29 Consider the experiment of tossing two dice. Let A denote the
event of an odd total, B the event of an ace on the first die, and C the
event of a total of seven. We pose three problems:
(i) Are A and B independent?
Oi) Are A and C independent?
(iii) Are Band C independent?
We obtain P[A IB] = 1 = P[A], P[A I C] = I :# P[A] = 1, and P[CI B] =

1; = P[C] = !; so A and B are independent, A is not independent of C,
and Band C are independent. IIII
The property of independence of two events A and B and the property that
A and B are mutually exclusive are distinct, though related, properties. For
example, two mutually exclusive events A and B are independent if and only if
P[A]P[B] = 0, which is true if and only if either A or B has zero probability.
Or if P[A] :# 0 and P[B] =F 0, then A and B independent implies that they are
not mutually exclusive, and A and B mutually exclusive implies that they are not
independent. Independence of A and B implies independence of other events
as wel1.
Theorem 32 If A and B are two independent events defined on a given

probability space (0, d, P[·]), then A and 13 are independent, A and B
are independent, and .If and 13 are independent.
PROOF
P[ABJ = P[A] - P[AB] = P[A] - P[A]P[B] = P[A](l - P[BD =

P[A]P[B].
Similarly for the others. IIII
The notion of independent events may be extended to more than two
events.
Definition 20 Independence of several events For a given probability

space (0, d, Pl']), let A., A 2 , ••• , An be n events in d. Events Ab
A 2 , ••• , All are defined to be independent if and only if
P[A,A j ] = P[Ai]P[A~] for i :# j
P[Ai1 j A d = P[A f ]P[A j ]P[A k l for i:# j,j:# k, i =F k
..
P[.OI AI] = tIl PtA,]. 1111
42 PROBABILITY I
One might inquire whether all the above conditions are required in the
definition. For instance, does P[A I A 2 A 3] = P[AtlP[A 2 ]P[A 3 ] imply P[A I A 2 ]
= P[AtlP[A 2 ]? Obviously not, since P[A 1A 2 A 3 ] = P[AdP[A 2 ]P[A 3 ] if P[A 3]
= 0, but P[A I A 2 ] #: P[AdP[A 2 ] if At and A2 are not independent. Or does
pairwise independence imply independence? Again the answer is negative,
as the following example shows.
EXAMPLE 30 Pairwise independence does not imply independence. Let Al

denote the event of an odd face on the first die, A2 the event of an odd face
on the second die, and A3 the event of an odd total in the random experi-
ment that consists of tossing two dice. P[AdP[A 2 ] = ! . ! = P[A 1 A 2],
P[AtlP[A 3] = 1 '1 = P[A31 AtlP[Ad = P[A I A 3], and P[A 2 A 3 ] = i =
P[A 2 ]P[A 3 ]; so Ab A 2 , and A3 are pairwise independent, However
P[A 1 A 2 A 3 ] = 0 #: ! = P[AdP[A 2 ]P[A 3 ]; so Ah A 2 , and A3 are not
independent. IIII
In one sense, independence and conditional probability are each used to find
the same thing, namely, P[AB], for P[AB] = P[A]P[B] under independence and
P[AB] = P[A IB]P[B] under nonindependence. The nature of the events A and
B may make calculations of P[A], P[B], and possibly P[A IB] easy, but direct
calculation of P[AB] difficult, in which case our formulas for independence or
conditional probability would allow us to avoid the difficult direct calculation
of P[AB]. We might note that P[AB] = P[A IB]P[B] is valid whether or not A
is independent of B provided that P[A IB] is defined.
The definition of independence is used not only to check if two given events
are independent but also to model experiments. For instance, for a given
experiment the nature of the events A and B might be such that we are willing
to assume that A and B are independent; then the definition of independence gives
the probability of the event A n B in terms of P[A] and P[B]. Similarly for
more than two events.
EXAMPLE 31 Consider the experiment of sampling with replacement from

an urn containi ng M balls of which K are black and M K white. Since
balls are being replaced after each draw, it seems reasonable to assume that
the outcome of the second draw is independent of the outcome of the
first. Then P[two blacks in first two draws] =
P[black on first draw]P[black on second draw] = (KIM)2, IIII
PROBLEMS 43
PROBLEMS
To solve some of these problems it may be necessary to make certain assumptions,
such as sample points are equally likely, or trials are independent, etc., when such
assumptions are not explicitly stated. Some of the more difficult problems, or those
that require'special knowledge, are marked with an *.
lOne urn contains one black ball and one gold ball. A second urn contains one
white and one gold ball. One ball is selected at random from each urn.
(a) Exhibit a sample space for this experiment.
(b) Exhibit the event space.
(c) What is the probability that both balls will be of the same color?
(d) What is the probability that one ball will be green?
2 One urn contains three red balls, two white balls, and one blue ball. A second
urn contains one red ball, two white balls, and three blue balls.
(a) One ball is selected at random from each urn.
(i) Describe a sample space for this experiment.
(ii) Find the probability that both balls will be of the same color.
(iii) Is the probability that both balls will be red greater than the prob-
ability that both will be white?
(b) The balls in the two urns are mixed together in a single urn, and then a sample
of three is drawn. Find the probability that all three colors are represented,
when (i) sampling with replacement and (ii) without replacement.
3 If A and B are disjoint events, P[A] =.5, and P[A u B] = .6, what is P[B]?
4 An urn contains five balls numbered 1 to 5 of which the first three are black and
the last two are gold. A sample of size 2 is drawn with replacement: Let Bl
denote the event that the first ball drawn is black and B2 denote the event that the
second ball drawn is black.
(a) Describe a sample space for the experiment, and exhibit the events B 1 , B 2 ,
and B 1 B 2 •
(b) Find P[B1], P[B 2 ], and P[B1B2]'
(c) Repeat parts (a) and (b) for sampling without replacement.
5 A car wit~ six spark plugs is known to have two malfunctioning spark plugs.
If two plugs are pulled at random, what is the probability of getting both of
the malfunctioning plugs ?
6 In an assembly-line operation, 1 of the items being produced are defective. If
three items are picked at random and tested, what is the probability:
(a) That exactly one of them will be defective?
(b) That at least one of them will be defective?
7 In a certain game a participant is allowed three attempts at scoring a hit. In the
three attempts he must alternate which hand is used; thus he has two possible
strategies: right hand, left hand, right hand; or left hand, right hand, left hand.
His chance of scoring a hit with his right hand is .8, while it is only .5 with his
left hand. If he is successful at the game provided that he scores at least two hits
in a row, what strategy gives the better chance of success? Answer the same
44 PROBABILITY I
question if .8 is replaced by PI and .5 by P2. Does your answer depend on PI

and P2? .
8 (a) Suppose that A and B are two equally strong teams. Is it more probabfe
that A will beat B in three games out of four or in five games out of
, -
seven? ~
(b) Suppose now that the probability that A beats B in an individual game is p.
Answer part (a). Does your answer depend on p?
9 If P[A] = t and P[B] = !, can A and B be disjoint? Explain.
10 Prove or disprove; If P[A] =P[B] =p, thenP[AB] <p2.
11 Prove or disprove: If P[A] = P[B] then A = B.
12 Prove or disprove: If P[A] = 0, then A = rp.
13 Prove or disprove: If P[A] = 0, then P[AB] =0.
14 Prove: If P[A] = IX and P[B] = fl, then P[AB] > 1 - IX -fl.
15 Prove properties (i) to (iv) of indicator functions.
16 Prove the more general statement in Theorem 19.
17 Exhibit (if such exists) a probability space, denoted by (0, d, P[· D, which satisfies.
the following. For Al and A2 members ofd, if P[Atl = P[A 2], then Al = A 2 .
18 Four drinkers (say I, II, III, and IV) are to rank three different brands of b&r
(say A, B, and C) in a blindfold test. Each drinker ranks the three beers as.-l.
(for the peer he likes best), 2, and 3, and then the assigned ranks of each brand
of beer,are summed. Assume that the drinkers really cannot discriminate between
beers so that each is assigning his rankings at random.
(a) What is the probability that beer A will receive a total score of 4?
(b) What is the probability that some beer will receive a total score of ~?
(c) What is the probability that some beer will receive a total score of 5 or less?
19 The following are three of the classical problems in probability.
(a) Compare the probability of a total of 9 with a total of 10 when trrreefair
dice are tossed once (Galileo and Duke of Tuscany).
. .,. the,
(b) Compare the probability of at least one 6 in 4 tosses of a fair die witli
probability of at least one double-6 in 24 tosses of two fair dice (Chevalier
de Mere).
(c) Compare the probabiJity of at least one 6 when six dice are rolled witi- the
probability of at least two 6s when twelve dice are rolled (Pepys to Newton).
20 A seller has a dozen small electric motors, two of which are faulty .. Acust~mff is
interested in the dozen motors. The seller can crate the motors with all twelve in ,/L
•
one box or with six in each of two boxes; he knows that the customer will inspect
twO of the twelve motors if they are all crated in one box and one motor from each'
of the two smaller boxes if they are crated six each to two smaller boxes. He
has three strategies in his attempt to sell the faulty motors: (i) crate all twelve
in one box; (ii) put one faulty motor in each of the two smaller box~; or (iii) put
both of the faulty motors in one of the smaller boxes and no faulty motor~ in the
other. What is the probability that the customer will not inspect a faulty motor
under each of the three strategies?
".
PROBLEMS 45
21 A sample of five objects is drawn from a larger population of N objects (N 5).

Let Nw or N wo denote the number of different samples that could be drawn
depending, respectively, on whether sampling is done with or without replacement.
Give the values for N w and Nwo • Show that when N is very large, these two values
are approximately equal in the sense that their ratio is close to 1 but not in the
sense that their difference is close to O.
22 • Out of a .!¥oup of 25 persons, what is the probability that all 25 will have different
birthdays? (Assume a 365-day year and that all days are equally likely.)
23 A bridge player knows that his two opponents have exactly five hearts between
the two of them. Each opponent has thirteen cards. What is the probability
that there is a three-two split on the hearts (that is, one player has three hearts
and the other two)?
24 (a) If r balls are randomly placed into n urns (each ball having probability lin
of going into the first urn), what is the probability that the first urn will
contain exactly k balls?
" (b) Let n -'1-Cl) and r -7 Xl while r/n = m remains constant. Show that the
",-W.
~ probapility you calculated approaches e .. m k /k!.

lll
,}S' )(:{biased coin has probability p of landing heads. Ace, Bones, and Clod toss the
coin successively, Ace tossing first, until a head occurs. The person who tosses
the first head wins. Find the probability of winning for each.
*26 It is told that in certain rural areas of Russia marital fortunes were once told in the
following way: A girl would hold six strings in her hand with the ends protruding
abQve and below; a friend would tie together the six upper ends in pairs and then
tie together the six lower ends in pairs. If it turned out that the friend had tied
~he six strings into at least one ring, this was supposed to indicate that the girl
''Would get married within a year. What is the probability that a single ring will
,be formed when the strings are tied at random? What is the probability that at
'least one ring will be formed? Generalize the problem to 2n strings.
27 Mr. Bandit, a well-known rancher and not so well-known part-time cattle rustler,
l¥ls twenty head of cattle ready for market. Sixteen of these cattle are his own
tfiilnd consequently bear his own brand. The other four bear foreign brands. Mr.
Bandit knows that the brand inspector at the market place checks the brands of
.20 percent of the cattle in any shipment. He has two trucks, one which will haul
all twenty <!attle at once and the other that will haul ten at a time. Mr. Bandit'
feels that he has four djfferent strategies to follow in his attempt to market the
cattle without getting caught. The first is to sell all twenty head at once; the
others are to sell ten head on two different occasions, putting all four stolen cattle
in one set of ten, or three head in one shipment and one in the other, or two head in
each of the shipments of ten. Which strategy will minimize Mr. Bandit's prob-
a~ility 6i" getting caught, and what is his probability of getting caught under each
strategy?
28 Show that the formula of Eq. (4) is the same as the formula of Eq. (5).
46 PROBABIliTY I
29 Prove Theorem 31.

30 Either prove or disprove each of the following (you may assume that none of the
events has zero probability):
(a) If P[A IB] > P[A], then P[BI A] > P[B].
(b) If P[A] > P[B], then P[A IC] > p[BI C].
31 A certain computer program will operate using either of two subroutines, say A
and B, depending on the problem; experience has shown that subroutine A will be
used 40 percent of the time and B will be used 60 percent of the time. If A is
used, then there is a 75 percent probability that the program will run before its
time limit js exceeded; and jf B is used, there is a 50 percent chance that it will do
so. What is the probability that the program will run without exceeding the time
limit?
32 Suppose that it is known that a fraction .001 of the people in a town have tuber-
culosis (TB). A tuberculosis test is given with the following properties: If the
person does have TB, the test will indicate it with a probability .999. If he does
not have TB, then there is a probability .002 that the test will erroneously indicate
that he does. For one randomly selected person, the test shows that he has
TB. What is the probability that he really does?
*33 Consider the experiment of tossing two fair regular tetrahedra (a polyhedron with
four faces numbered 1 to 4) and noting the numbers on the downturned faces.
(a) Give three proper events (an event A is proper if 0 < P[A] < 1) which are
independent (if such exist).
(b) Gjve three proper events which are pairwise independent but not independent
(if such exist).
(c) Give four proper events which are independent (if such exist).
34 Prove or disprove:
(a) If A and B are independent events, then P[ABI C] = P[A IC]p[BI C].
(b) If p[A IB] =P[B], then A and B are independent.
(a) If P[A IB] > P[A], then P[BI A] > P[B].
(b) If p[BI A] =P[BI A], then A and B are independent.
(c) If a = P[A] and b =P[B], then P[A IB] > (a + b - 1)/b.
36 Consider an urn containing 10 balls of which 5 are black. Choose an integer n
at random from the set 1, 2, 3, 4, 5, 6, and then choose a sample of size n without
replacement from the urn. Find the probability that all the balls in the sample
will be black.
37 A die is thrown as long as necessary for a 6 to turn up. Given that the 6 does not
turn up at the first throw, what is the probabjlity that more than four throws will
be necessary?
38 Die A has four red and twO blue faces, and die B has two red and four blue faces.
The following game is played: First a coin is tossed once. If it falls heads, the
game continues by repeatedly throwing die A; if it falls tails, die B is repeatedly
tossed.
PROBLEMS 47
(a) Show that the probability of red at any throw is i.

(b) If the first two throws of the die resulted in red, what is the probability of red
at the third throw?
(c) If red turns up at the first n throws, what is the probability that die A is
being used?
39 Urn A contains two white and two black bans; urn B contains. three white and
two black bans. One ball is transferred from A to B; one ball is then drawn
from B and turns out to be white. What is the probability that the transferred
ball was white?
*40 It is known that each of four people A, B, C, and D tells the truth in a given
instance with probability 1-. Suppose that A makes a statement, and then D says
that C says that B says that A was telling the truth. What is the probability
that A was actually telling the truth?
4J In a T maze, a laboratory animal is given a choice of going to the left and getti ng
food or going to the right and receiving a mild electric shock. Assume that
before any conditioning (in trial number I) animals are equally likely to go to the
left or to the right. After having received food on a particular trial, the prob-
abilities of going to the left and right become .6 and .4, respectively, on the follow-
ing trial. However, after receiving a shock on a particular trial, the probabilities
of going to the left and right on the next trial are .8 and .2, respectively. What
is the probability that the animal will turn left on trial number 2? On trial
number 3?
*42 In a breeding experiment, the male parent is known to have either two dominant
genes (symbolized by AA) or one dominant and one recessive (Aa). These two
cases are equally likely. The female parent is known to have two recessive genes
(aa). Since the offspring gets one gene from each parent, it will be either Aa or
aa, and it will be possible to say with certainty which one.
(a) If we suppose one offspring is Aa, what is the probability that the male
parent is A A ?
(b) If we suppose two offspring are both Aa, what is the probability that the
male parent is AA ?
(c) If one offspring is aa, what is the probabjJity that the male parent is Aa 1
4! The constitution of two urns is
three black four black

I II
two white six white
A draw is made by selecting an urn by a process which assigns probability p to the

selection of urn I and probability 1 p to the selection of urn II. The selection
of a ball from either urn is by a process which assigns equal probability to all
balls in the urn. What value of p makes the probability of obtaining a black
ball the same as if a single draw were made from an urn with seven black and
eight white balls (all balls equally probable of being drawn)?
48 PROBABILITY I
44 Given P[A] = .5 and P[A v B) = .6, find P[B] if:

(a) A and B are mutually exclusive.
(b) A and B are independent.
(c) P[A IB] = .4.
45 Three fair dice are thrown once. Given that no two show the same face:
(a) What is the probability that the sum of the faces is 7?
(b) What is the probability that one is an ace?
46 Given that P[A] > 0 and P[B] > 0, prove or disprove:
(a) If P[A] = P[B], then P[A IB] = P[BI A].
(b) If P[A IB) = p[BI A), then P[A] = P[B].
47 Five percent of the people have high blood pressure. Of the people with high
blood pressure, 75 percent drink alcohol; whereas, only 50 percent of the people
without high blood pressure drink alcohol. What percent of the drinkers have
high blood pressure?
48 A distributor of watermelon seeds determined from extensive tests that 4 percent
of a large batch of seeds will not germinate. He sells the seeds in packages of 50
seeds and guarantees at least 90 percent germination. What is the probability
that a given package will violate the guarantee?
49 If A and B are independent, P[A] = t, and P[E] = 1, find P[A u B].
50 Mr. Stoneguy, a wealthy diamond dealer, decides to reward his son by allowing
him to select one of two boxes. Each box contains three stones. In one box two
of the stones are real diamonds, and the other is a worthless imitation; and in the
other box one is a real diamond, and the other two worthless imitations. If the
son were to choose randomly between the two boxes, his chance of getting two
real diamonds would be!. Mr. Stoneguy, being a sporting type, allows his
son to draw one stone from one of the boxes and to examine it to see if it is a real
diamond. The son decides to take the box that the stone he tested came from if
the tested stone is real and to take the other box otherwise. Now what is the
probability that the son will get two real d!amonds?
51 If P[A] =P[B]=P[BI A) = !, are A and B independent?
52 If A and B are independent and P[A] = P[B] = !, what is P[AE v AB]?
53 If P[B) = P[A IB] = P[CI AB] = i, what is P[ABC]?
54 If A and B are independent and P[A) =P[BIA) = i, what is P[A u B]?
55 Suppose Bt, B 2 , and B3 are mutually exclusive. If P[Bil = ! and P[A IB j ] = j/6
for j = 1, 2, 3, what is P[A]?
*56 The game of craps is played by letting the thrower toss two dice until he either wins
or loses. The thrower wins on the first toss if he gets a total of 7 or I I ; he loses
on the first toss if he gets a total of 2, 3, or 12. If he gets any other total on his
first toss, that total is called his point. He then tosses the dice repeatedly untH he
obtains a total of 7 or his point. He wins if he gets his point and loses if he gets a
total of 7. What is the thrower's probability of winning?
57 In a dice game a player casts a pair of dice twice. He wins if the two totals
thrown do not differ by more than 2 with the following exceptions: If he gets a
PROBLEMS 49
3 on the first throw, he must produce a 4 on the second throw; if he gets an ) I on

the first throw, he must produce a 10 on the second throw. What is his prob-
ability of winning?
58 Assume that the conditional probability that a child born to a couple will be
~ is! mEl fEz ~ where El and Ez are certain small constants, m is the
nUmber of male children already born to the couple, andfis the number of female
children already born to the couple.
(a) What is the probability that the third child will be a boy given that the first
two are girls?
(b) Find the probability that the first three children will be all boys .
. (c) Find the probability of at least one boy in the first three children.
(Your answers will be expressed in terms of 81 and 82.)
*59 A network of switches a, b, c, and d is connected across the power lines A and B
as shown in the sketch. Assume that the switches operate electrically and have
independent operating mechanisms. All are controlled simultaneously by the
same impulses; that is, it is intended that on an impulse all switches shall close
&l;imultaneously. But each switch has a probability P of failure (it will not close
when it should).
A B
~--------4C~·----~----~~-------4
------------ Power lines
(a) What is the probability that the circuit from A to B will fail to close?
(b) If ~a line is added on at e, as indicated in the sketch, what is the probability
that the circuit from A to B will fail to close?
(c) 'If a line and switch are added at e, what is the probability that the circuit from
A to B will fail to close?
II
60 Let Bt, B 2 , ••• , Bn be mutually disjoint, and let B = U BJ • Suppose P[Bj ] > 0
J=1
and P(A I B J] = P for j = I, ... , n. Show that peA 1 B] = p.
61 In a laboratory experiment, an attempt is made to teach an animal to turn right
in a maze. To aid in the teaching, the animal is rewarded if it turns right on a
given trial and punished if it turns left. On the first trial the animal is just as
likely to turn right as left. If on a particular trial the animal was rewarded, his
probability of turning right on the next trial is PI > !, and if on a given trial the
animal was punished, his probability of turning right on the next trial is P2 > Pl.
(a) What is the probability that the animal will turn right on the third trial?
(b) What is the probability that the animal wHi turn right on the third trial,
given that he turned right on the first trial?
50 PROBABILITY I
*62 You are to play ticktacktoe with an opponent who on his turn makes his mark by
selecting a space at random from the unfilled spaces. You get to mark first.
Where should you mark to maximize your chance of winning, and what is your
probability of winning? (Note that your opponent cannot win, he can only
tie.)
63 Urns I and II each contain two white and two black balls. One ball is selected
from urn I and transferred to urn II; then one ball is drawn from urn II and turns
out to be white. What is the probability that the transferred ball was white?
64 Two regular tetrahedra with faces numbered 1 to 4 are tossed repeatedly until a
total of 5 appears on the down faces. What is the probability that more than two
tosses are required?
65 Given P[A] = .5 and P[A v B] = .7:
(a) Find P[B] if A and B are independent.
(b) Find P[B] if A and B are mutually exclusive.
(c) Find P[B] if P[A IB] =.5.
66 A single die is tossed; then n coins are tossed, where n is the number shown on the
die. What is the probability of exactly two heads?
*67 In simple Mendelian inheritance, a physical characteristic of a plant or animal is
determined by a single pair of genes. The color of peas is an example. Let y and
9 represent yellow and green; peas will be green if the plant has the color-gene
pair(g, g); they will be yellow if the color-gene pair is (y, y) or (y, g). In view of
this last combination, yellow is said to be dominant to green. Progeny get one
gene from each parent and are equally likely to get either gene from each parent's
pair. If (y, y) peas are crossed with (g, g) peas, all the resulting peas will be (y, g)
and yellow because of dominance. If (y, g) peas are crossed with (g, g) peas, the
probability is .5 that the resulting peas will be yellow and is .5 that they will be
green. In a large number of such crosses one would expect about half the result-
ing peas to be yellow, the remainder to be green. In crosses between (y, g) and
(y, g) peas, what proportion would be expected to be yellow? What proportion
of the yellow peas would be expected to be (y, y)?
*68 Peas may be smooth or wrinkled, and this is a simple Mendelian character.
Smooth is dominant to wrinkled so that (s, s) and (s, w) peas are smooth while
(w, w) peas are wrinkled. If (y, g) (s, w) peas are crossed with (g, g) (w, w) peas,
what are the possible outcomes, and what are their associated probabilities? For
the (y, g) (s, w) by (g, g) (s, w) cross? For the (y, g) (s, w) by (y, g) (s, w) cross?
69 Prove the two unproven parts of Theorem 32.
70 A supplier of a certain testing device claims that his device has high reliability
inasmuch as P[A IB] = P[A IB] = .95, where A = {device indicates component is
faulty} and B = {component is faulty}. You hope to use the device to locate the
faulty components in a large batch of components of which 5 percent are faulty.
(a) What is P[B IA]?
(b) Suppose you want p[BIA] =.9. Let p=P[AIB]=P[AIB]. How large
does p have to be?
II
RANDOM VARIABLES, DISTRIBUTION
FUNCTIONS, AND EXPECTATION

The purpose of this chapter is to introduce the concepts of random variable,
distribution and density functions, and expectation. It is primarily a "definitions-
and-their-understanding" chapter; although some other results are given as
well. The definitions of random variable and cumulative distribution function
are given in Sec. 2, and the definitjons of density functions are given in Sec. 3.
These definitions are easily stated since each is just a particular function. The
cumulative distribution function exists and is defined for each random variable;
whereas, a density function is defined only for particular random variables.
Expectations of functions of random ,-ariables are the underlying concept of all
of Sec. 4. This concept is introduced by considering two particular, yet
extremely important, expectations. These two are the mean and variance,
defined in Subsecs. 4.1 and 4.2, respectively. Subsection 4.3 is devoted to the
definition and properties of expectation of a function of a random variable.
A very important result in the chapter appears in Subsec. 4.4 as the Chebyshev
inequality and a generalitation thereof. It is nice to be able to attain so famous
a result so soon and with so little weaponry. The Jensen inequality is given in
52 RANDOM VARIABLES, DISTRIBUTION FUNCTIONS. AND EXPECTATION II
Subsec. 4.5. Moments and moment generating functions, which are expecta-
tions of particular functions, are considered in the final subsection. One major
unproven result, that of the uniqueness of the moment generating function, is
given there. Also included is a brief discussion of some measures, of some
characteristics, such as location and dispersion, of distribution or density
functions.
This chapter provides an introduction to the language of distribution
theory. Only the univariate case is considered; the bivariate and multivariate
cases will be considered in Chap. IV. It serves as a preface to, or even as a
companion to, Chap. III, where a number of parametric families of distribution
functions is presented. Chapter III gives many examples of the concepts
defined in Chap. II.
2 RANDOM VARIABLE AND CUMULATIVE

DISTRIBUTION FUNCTION
2.1 Introduction
In Chap. I we defined what we meant by a probability space, which we denoted

by the triplet (n, d, P[ . ]). We started with a conceptual random experiment;
we called the totality of possible outcomes of this experiment the sample space
and denoted it by n. d was used to denote a collection of subsets, called
events, of the sample space. Finally our probability function P[ . ] was a set
function having domain d and counterdomain the interval [0, 1]. Our object
was, and still is, to assess probabilities of events. In other words, we want to
model our random experiment So as to be able to give values to the probabilities
of events. The notion of random variable, to be defined presently, will be used
to describe events, and a cumulative distribution function will be used to give the
probabilities of certain events defined in terms of random variables; so both
concepts will assist us in defining probabilities of events, our goal. . One advan-
tage that a cumulative distribution function will have over its counterpart, the
probability function (they both give probabilities of events), is that it is a
function with domain the real line and counterdomain the interval [0, 1]. Thus
we will be able to graph it. It will become a convenient tool in modeling
random experiments. In fact, We will often model a random experiment by
assuming certain things about a random variable and its distribution function
and in so doing completely bypass describing the probability space.
2 RANDOM VARIABLE AND CUMULATIVE DISTRIBUTION FUNCTION 53
2.2 Definitions
We commence by defining a random variable.
Definition 1 Random Variable For a given probability space (n, .91,

P[ . ]), a random variable, denoted by X or X(·), is a function with
domain nand counterdomain the real line. The function X( . ) must be
such that the set A r , defined by Ar = {w: X(w) < r}, belongs to .91 for-every
real number r. IIII
If one thinks in terms of a random experiment, n is the totality of out-

comes of that random experiment, and the function, or random variable, X( . )
with domain n makes some real n umber correspond to each outcome of the
experiment. That is the important part of our definition. The fact that we
also require the collection of w's for which X(w) < rto be an event (i.e., an element
of d) tor each real number r is not much of a restriction for our purposes
'since our intention is to use the notion of random variable only in descri.bing
events. -.We will seldom be interested in a random variable per se; rather we
will be interested in events defined in terms of random variables. One might
note that the P[ . ] of our probability space (n, .91, P[ . ]) is not used in our
definition.
The use of words" random" and" variable" in the above definition is
unfortunate since their use cannot be convincingly justified. The expression
"random variable" is a misnomer that has gained such widespread use that it
would be foolish for us to try to rename it.
In our definition we denoted a random variable by either X(· ) or X.
Although X( . ) is a more complete notation, one that emphasizes that a random
variable is a function, we will usually use the shorter notation of X. For many-
experiments, there is a need to define more than one random variable; hence
further notations are necessary. We will try to use capital Latin letters with or
without affixes from near the end of the alphabet to denote random variables.
Also, we use the corresponding small letter to denote a value of the random
variable.
EXAMPLE 1 Consider the experiment of tossing a single coin. Let the

random variable X denote the number of heads. n = {head, tail}, and
X(w) = 1 if w = head, and X(w) = 0 if w = tail; so, the random variable X
associates a real number with each outcome of the experiment. We
called X a random variable so mathematically speaking we should show
54 RANDOM VARIABLES, DISTRIBUTION FUNCTIONS, AND EXPECTATION II
2d dice
6 • • • • • •
5 • • • • • •
4 • • • • • •
3 • • • • • •
2 • • • • • •
1 • ,. • • • •
1st dice
FIGURE 1 1 2 3 4 5 6
that it satisfies the definition; that is, we should show that {w: X(w) < r}
belongs to d for every real number r. d consists of the four subsets:
4>, {head}, {tail}, and n. Now, if r < 0, {w: X(w) < r} = 4>; and if
° <r < I, {w: X(w) < r} = {tail}; and if r > 1, {w: X(w) < r} = n = {head,
tail}. Hence, for each r the set {w: X(w) < r} belongs to d; so X( . ) is a-
random variable. IIII
EXAMPLE 2 Consider the experiment of tossing two dice. n can be de-

scribed by the 36 points displayed in Fig. 1. n = {(i, j): i = I , ... , 6 and
j = 1, ... , 6}. Several random variables can be defined; for instance, let
X denote the sum of the upturned faces; so X(w) = i + j if w = (i, j). Also,
let Y denote the absolute difference between the upturned faces; then
Y(w) = ]i - j] if w = (i, j). It can be shown that both X and Yare ran-
dom variables. We see that X can take on the values 2, 3, ... , 12 and Y
can take on the values 0, 1, ... , 5. IIII
In both of the above examples we described the random variables in terms

of the random experiment rather than in specifying their functional form; such
will usually be the case.
• Definition2 Cumulative distribution function Thecumulativedistribution

function of a random variable X, denoted by Fx ('), is defined to be
that function with domain the real line and counterdomain the interval
2 RANDOM VARIABLE AND CUMULATIVE DISTRIBUTION FUNCTION 55
[0, 1] which satisfies Fx(x) P[X x] P[{w: X(w) < x}] for every real
numberx. I1II
A cumulative distribution function is uniquely defined for each random
variable. If it is known, it can be used to find probabilities of events defined
in terms of its corresponding random variable. (One might note that it is in
this definition that we use the requirement that {w: X(w) < r} belong to d for
every real r which appears in our definition of random variable X.) Note that
different random variables can have the same cumulative distribution function.
See Example 4 below.
The use of each of the three words in the expression" cumulative distri-
bution function" is justifiable. A cumulative distribution function is first of
all a/unction; it is a distribution function inasmuch as it tells us how the values
of the random variable are distributed, and it is a cumulative distribution func-
tion since it gives the distribution of values in cumulative form. Many writers
omit the word "cumulative" in this definition. Examples and properties of
cumulative distribution functions follow.
EXAMPLE 3 Consider again the experiment of tossing a single coin. Assume

that the coin is fair. Let X denote the number of heads. Then,
if x <0
if 0 x < 1
if 1 x.
Or Fx(x) 1/ [0, .)(x) + 1[1, oo)(x) in our indicator function notation. III1
EXAMPLE 4 In the experiment of tossing two fair dice, let Y denote the
absolute difference. The cumulative distribution of Y, Fy( . ), is sketched
in Fig. 2. Also, let X k denote the value on the upturned face of the kth
die for k 1, 2. Xl and X 2 are different random variables, yet both
have the same cumulative distribution function, which is FXk(x) =
5 •
·1
l~ 6Ip, i+ I)(X) + 1[6, oo)(X) and is sketched in Fig. 3. IIII
Careful scrutiny of the definition and above examples might indicate the
following properties of any cumulative distribution function Fx( . ).
S6 RANDOM VARIABLES, DISTRIBUTION FUNCTIONS, AND EXPECTATION II
Fr(y)
34/36
------------
• •
30/36 •
24/36 •
16/36 T
I
I
6/36
I
! L. I I I .. y
FIGURE 2 1 2 3 4 5
Properties of a Cumulative Distribution Function F x (')

(i) Fx( ex)) = lim Fx(x}
x ..... -oc
= 0, and Fx( ex)) lim Fx(x) = 1.
x-+oo
(ii) Fx (') is a monotone, nondecreasing function; that is, Fx(a) < Fx(b}
for a < b.
(iii) Fx (') is continuous from the right; that is,
lim Fx(x + h) = Fx(x}.
o <h ..... O
Except for Oi), we will not prove these properties. Note that the event
{w: X(w) b} {X < b} {X < a} u {a < X b} and {X a} n {a < X <b}
= 4>; hence, Fx(b} P[X < b] = P[X < a] + P[a < X b] P[X < a] = Fx(a)
which proves (ii). Property (iii), the continuity of Fx( . ) from the right, results
from our defining Fx(x) to be P[X < xl lf we had defined. as some authors do,
Fx{x) to b~ P[X < x], then Fx( . ) would have been continuous from the left.
Definition 3 Cumulative distribution function Any function F( . } with

domain the real line and counterdomain the interval [0, J] satisfying the
above three properties is defined to be a cumulative distribution/unction.
IIII
This definition allows us to use the term" cumulative distribution func~

tion" without mentioning random variable.
After defining what is meant by continuous and discrete random variables
in the first two subsections of the next section, we will give another property
that cumulative distribution functions possess, the property of decomposition
into three parts.
3 DENSITY FUNCTIONS 57
1 •
•
•
•
•
•
FJGURE 3
The cumulative distribution functions defined here are univariate; the

introduction of bivariate and multivariate cumulative distribution functions
will be deferred until Chap. IV.
3 DENSITY FUNCTIONS
Random variable and the cumulative distribution function of a random variable
have been defined. The cumulative distribution function described the distri-
bution of values of the random variable. For two distinct classes of random
variables, the distribution of values can be described more simply by using
density functions. These two classes, distinguished by the words "discrete"
and" continuous," are considered in the next two subsections.
3.1 Discrete Random Variables

Definition 4 Discrete random variable A random variable X will be
defined to be discrete if the range of X is countable. If a random variable
X is discrete, then its corresponding cumulative distribution function
Fx( . ) will be defined to be discrete. IIII
By the range of X being countable we mean that there exists a finite or

denumerable set of real numbers, say XI' X 2 , X 3 , ..• , such that X takes on values
only in that set. If X is discrete with distinct values x I, X2, •.. , x"' ... , then
0= U {w: X(w) = x,,} = U{X = x,,}, and {X = xJ n {X = x) = 4> for i # j;
" "
hence 1 = prO] = L P[X = x,,] by the third axiom of probability.
"
Definition 5 Discrete density function of a discrete random variable If

X is a discrete random variable with distinct values X h Xl' ••. , X n , ••• ,
then the function, denoted by fx( . ) and defined by
if X Xj' j = I, 2, ... , n, ...

(1)
if x =I: Xj
is defined to be the discrete density function of X. IIII
The values of a discrete random variable are often called mass points; and,
fx(xj) denotes the mass associated with the mass point x j • Probability mass
function, discrete frequency function, and probability function are other terms
used in place of discrete density function. Also, the notation px( .) is some-
times used intead of fx( . ) for discrete density functions. fx(') is a function
with domain the real line and counterdomain the interval [0, 1]. If we use the
indicator function,
00
fx(x)
n
L P[X = xn]/{x,,}(x),
I
(2)
Theorem 1 Let Xbe a discrete random variable. Fx (') can be obtained

fromfx('), and vice versa.
PROOF Denote the mass points of Xby Xl, Xl, • ••• Suppose fx(')
is given; then Fx(x) = L
fx(x). Conversely, suppose Fx(') is given;
{j: xr';x}
then fx(xj) = Fx(xj ) - lim Fx(xj - h); hence fx(x) can be found for
O<h ..... O
each mass point X j; however, fx(x) = 0 for X =I: X J' j 1, 2, ... , so fx(x) is
determined for all real numbers. IIII
EXAMPLE 5 To illustrate what is meant in Theorem 1, consider the experi-

ment of tossing a single die. Let X denote the number of spots on the
upper face:
fx( x) = (~) 1(1.2 • .••• 6)(X)'

and
5
Fx(X) = L (i16)/[i,
t:= I
i+ 1)(x) + 1[6. oo)(X).
-A
-h ~ i
~ 36
f6 16
-h io
0 1 2 3 4 5 6 7 8 9 10 11 12
FIGURE 4
According to Theorem 1, for given fx('), Fx(x) can be found for any x;
for instance, if x = 2.5,
And, if Fx (') is gjven, fx(x) can be found for any x. For example, for
x = 3,
fx(3) = Fx(3) - o~~o Fx(3 - h) = W-W=~. IIII
The cumulative distribution function of a discrete random variable has

steps at the mass points; that is, at the mass point x j ' Fx (') has a step of size
fx{x J) , and Fx {') is flat between mass points.
EXAMPLE" 6 Consider the experiment of tossing two dice. Let X denote

the total of the upturned faces. The mass points of X are 2, 3, ... , 12.
lx(') is sketched in Fig. 4. Let Y denote the absolute difference of
the upturned faces; thenfy(') is given in tabular form by
Y 0 I 2 3 4 5
6 10 8 6
fly) 36 J6 36 36
4
36
..L
36 11I1
The discrete density function tells us how likely or probable each of the
values of a discrete random variable is. It also enables one to calculate the
probability of events described in terms of the discrete random variable X.
For example, let X have mass points XI' x 2 , ••• , X n , ••• ; then P[a < X b] =
L !x(Xj) for a < b.
j:{ a< XJ :;;b}
Definition 6 Discrete density function Any function f( . ) with domain

the real line and counterdomain [0, 1] is defined to be a discrete density
function if for some countable set Xl' x 2 , ••• , X n , ••• ,
(0 f(x) > 0 for j = 1, 2, ....
(ii) f(x) = 0 for X #= Xj;j = 1,2, ....
(iii) L f(x)) = 1, where the summation is over the points XI' x2 , •• "
Xn , • • • • IIII
This definition allows us to speak of discrete density functions without
reference to some random variable. Hence we can talk about properties that
a density function might have without referring to a random variable.
3.2 Continuous Random Variables

Definition 7 Continuous random variable A random variable X is
x
called continuous if there exists a functionfx(') such that Fx(x) = J fx(u)du
-00
for every real number x. The cumulative distribution function Fx( .) of a

continuous random variable X is called absolutely continuous. - IIII
Definition 8 Probability density function of a continuous random variable

If X is a continuous random variable, the function fx(') In Fx(x) =
x
f fx(u) du is called the probability density function of X. I111
-00
Other names that are used instead of probability density function include
density function, continuous density function, and integrating density function.
, Note that strictly speaking the probability density function fx(') of a
random variable X is not uniquely defined. All that the definition requires is
that the integral of fx(') gives Fx(x) for every x, and more than one function
fx(') may satisfy such requirement. For example, suppose Fx(x) = x/[o, I)(x) +
x
1[I.oo)(x); then fx(u) = 1(0, I)(u) satisfies Fx(x) = J fx(u)
-00
du for every x, and
so fx(') is a probability density function of X. However fx(u) ::=: 1(0. tlu) +
x
69/(t}(u) + l<t, I)(u) also satisfies Fx(x) = J
-00
fx(u) duo (The idea is that ,if the
value of a function is changed at only a "few" points, then its integral is
unchanged.) In practice a unique choice of fx(') is often dictated by continuity'
considerations and for this reason we will usually allow ourselves the liberty of
3 DENSITY fUNCTIONS 61
speaking of the probability density when in fact a probability density is more

correct.
One should point out that the word "continuous" in "continuous
random variable" is not used in its usual sense. Although a random variable
is a function and the notion of a continuous function is fairly well established
in mathematics, .. continuous" here is not used in that usual mathematical
sense. In fact it is not clear in what sense it is used. Two possible justifica-
tions do come to mind. In contrasting discrete random variables with contin-
uous random variables, one notes that a discrete random variable takes on a
finite or denumerable set of values whereas a continuous random variable takes.
on a nondenumerable set of values. Possibly it is the connection between
" nondenumerable" and continuum" that justifies use of the word "contin-
U
uous." All the continuous random variables that we shall encounter will take
on a continuum of values .. The second justification arises when one notes that
the absolute continuity of the cumulative distribution function is the regular
mathematical definition of an absolutely continuous function (in words, a
function is called absolutely continuous if it can be written as the integral of its
derivative); the "continuous," then, in a corresponding continuous random
variable could be considered just an abbreviation of" absolutely continuous."
Theorem 2 Let X be a continuous random variable. Then Fx (') can

be obtained from an/x {'), and vice versa.
PROOF If X is a continuous random variable and an /x( .) is given,

:x
then Fx(x) is obtained by integrating/x ('); that is, Fx(x) = J
-00
ix(u) duo On
the other hand, if Fx(') is given, then an/x(x) can be obtained by differ-
entiation; that is, /x(x) = dFx(x)/dx for those points x for which Fx(x) is
differentiable. II11
The notations for discrete density function and probability density func-
tion are the same, yet they have quite different interpretations. For discrete
random variables /x(x) = P[X = xl, which is not true for continuous random
variables. For continuous random variables,
- dFx(x) _ l'
- 1m Fx(x
+ ~x) - Fx(x - ~x)
f x (x ) - .
dx 4:x-O 2~x '
hence fx(x)2~x ~ Fx(x + ~x) - Fx(x - ~x) = P[x - Ax < X < x + ~xl; that
is, the probability that X is in a small interval containing the value x is approxi-
mately equal to /x(x) times the width of the interval. For discrete random
variables fx(') is a function with domain the real line and counterdomain the
interval [0,.1]; whereas, for continuous random variables fx(') is a function with
domain the real line and counterdomain the infinite interval [O~ (0) .
. Remark We will use the term" density function" without the modifier
of" discrete" or "probability" to represent either kind of density. 1111
EXAMPLE 7 Let X be the random variable representing the length of a

telephone conversation. One could model this experiment by assuming
that the distribution of X is given by Fx(x) = (1 - e AX)/[O.oo,(x), where
l is some positive number. The corresponding probability density func-
tion would be given by fx(x) = le-AX/[O, oo)(x). If we assume that
telephone conversations are measured in minutes, P[5 < X < 10] =
n°
le-xAdx = e- SA - e- 10A = e- 1 - e- 2 ~.23forl = t,orP[5 < X < 10]
= P[X < 10] - P[X::; 5] = (I - e AI0) - (I - e A.s) = e- 1 - e- 2 for
l = t· /1//
The probability density function is used to calculate the probability of

events defined in terms of the corresponding continuous random variable X.
For ex~mple, P[a < X < b) = S!fx(x) dx for a < b.
Definition 9 Probability density function Any function f( . ) with domain

the real line and counterdomain [0, (0) is defined to be a probability
density function if and only if
(0 f(x) >
IX)
°
for all x.
(ii) J f(x) dx =
-IX)
1. III/
With this definition we can speak of probability density functions without

reference to random variables. We might note that a probability density func-
tion of a continuous random variable as defined in Definition 8 does indeed
possess the two properties in the above definition.
3.3 Other Random Variables

Not all random variables are either continuous or discrete, or not all cumulative
distribution functions are either absolutely continuous or discrete.
______ ~ ______ ~- ___________ --~---- __ x

~
FIGURE 5
EXAMPLE 8 Consider the experiment of recording the delay that a motorist

encounters at a one..way traffic stop sign. Let X be the random variable
that represents the delay that the motorist experiences after making the
required stop. There is a certain probability that there will be no oppos-
ing traffic so that the motorist will be able to proceed with no delay.' On
the other hand, if the motorist has to wait, he may have to wait for any of
a continuum of possible times. This experiment could be modeled by
assuming that X has a cumulative distribution function given by Fx{x)
= (1 - pe AX)/[O,OO)(x). This Fx(x) has a jump of I - p at x = 0 but is
continuous for x > O. See Fig. 5. IIII
Many practical examples of cumulative distribution functions that are

partly discrete and partly absolutely continuous can be given. Yet there are
still other types of cumulative distribution functions. There are continuous
cumulative distribution functions, called singular continuous, whose derivative
is 0 at alm<?st all points. We will not consider such distribution functions other
than to note the following result.
Decomposition of a cumulative distribution function Any cumulative

distribution function F(x) may be represented in the form
where PI > 0, i = 1,2,3. (3)

3
L Pi =
i= 1
I, and F d ( • ), F ac( • ), and pc ( . ) are each cumulative distribution functions
with F d (.) discrete, Fac{.) absolutely continuous, and P SC( . ) singular continuous.
Cumulative distributions studied in this book will have at most a discrete
part and an absolutely continuous part; that is, the P3 in Eq. (3) will always be 0
for the F(') that we will study.
64' RANDOM VARIABLES, DISTRIBUTION FUNCTIONS, AND EXPECTATION II
EXAMPLE 9 To illustrate how the decomposition of a cumulative distribu-

tion function can be implemented, consider Fx(x) = (l - pe-Ax)/[o, oo)(x)
as in Example 8. Fx(x) = (l - p)Fd(X) + ppac(x), where Fd(X) = 1[0, OO)(x)
and Fac(x) = (1 - e-Aj/[o. OO)(x). Note that Fx(x) = (l - p)Fd(x) +
pFac(x) = (l - p)/[o. OO)(x) + p(l - e-Ax)/[o, oo)(x) = (l - pe-AX)I[o, oo)(x).
IIII
A density function corresponding to a cumulative distribution that is

partly discrete and partly absolutely continuous could be defined as follows:
If F(x) = (l - p)Fd(X) + pFac(x), where 0 < p < I and F d(.) and F ac (.) are,
respectively, discrete and absolutely continuous cumulative distribution func-
tions, let the density function f(x) corresponding to F(x) be defined by f(x)
= (l - p)~(x) + pfac(x), where r(') is the discrete density function corre-
sponding to F d (.) andfac(.) is the probability density function corresponding to
F ac (.). Such a density function would require careful interpretation; so when
considering cumulative distribution functions that are partly discrete and
partly continuous, we will tend to work with the cumulative distribution func-
tion itself rather than with a density function.
Remark In future chapters we will frequently have to state that a

random variable has a certain distribution. We will make such a state-
ment by giving either the cumulative distribution function or the density
function of the random variable of interest. IIII
4 EXPECTATIONS AND MOMENTS

An extremely useful concept in problems involving random variables or distri-
butions is that of expectation. The subsections of this section give definitions
and results regarding expectations.
4.1 Mean
Definition 10 Mean Let X be a random variable. The mean of X,

denoted by /-lx or G[X], is defined by:
(i) (4)
if X is discrete with mass points Xl' x2 , •.. , xj ' .•..

4 EXPECTATIONS AND MOMENTS 65
(ii) G[X] = f- oo xfx (x)dx

00
(5)
if X is continuous with probability density functionfx(x).
(iii) G[X] = fo
00
[1 - Fx(x)] dx -
fO_ooFx(X) dx (6)
for an arbitrary random variable X. IIII
In 0), G[X] is defined to be the indicated series provided that the series is
absolutely convergent; otherwise, we say that the mean does not exist. And in
(ii), G[X] is defined to be the indicated integral if the integral exists; otherwise,
we say that the mean does not exist. Final!y, in (iii), we require that both
integrals be finite for the existence of G[X].
L
Note what the definition says: In xjfx(xj), the summand is thejth value
j
of the random variable X multiplied by the probability that X equals that jth
value, and then the summation is overall values. So G[X] is an" average" of the
values that the random variable takes on, where each value is weighted by the
probability that the random variable is equal to that value. Values that are
more probable receive more weight. The same is true in integral form in (ii).
There the value x is multiplied by the approximate probabjlity that X equals
the value x, namely fx(x) dx, and then integrated over all values.
Several remarks are in order.
Remark In the definition of a mean of a random variable, only density

functions [in (0 and (ii)] or distribution functions [in (iii)] were used;
hence we have really defined the mean for these functions without reference
to random variables. We then call the defined mean the mean of the
cumulative distribution function or of the appropriate density function.
Hence, we can and will speak of the mean of a distribution or density
function as well as the mean of a random variable. IIII
Remark G[X] is the center of gravity (or centroid) of the unit mass that
is determined by the density function of X. So the mean of X is a meas-
ure of where the values of the random variable X are" centered." Other
measures of "location" or "center" of a random variable or its corre-
sponding density are given in Subsec. 4.6. 7Tn '
66 RANDOM VARIABLES; DISTRIBUTION FUNCI10NS; AND EXPECTATION II
Remark (iii) of the definition is for all random variables; whereas,

(i) is for discrete random variables, and (ii) is for continuous random
variables. Of course, tC[X] could have been defined by just giving (iii).
The reason for including (i) and (ii) is that they are more intuitive for
their respective cases. It can be proved, although we will not do it, that
(i) follows from (iii) in the case of discrete random variables and (ii) follows
from (iii) in the case of continuous random variables. Our main use
of (iii) will be in finding the mean of a random variable X that is neither
discrete nor continuous. See Example 12 below. IIII
EXAMPLE 10 Consider the experiment of tossing two dice. Let X denote

the total of the two dice and Y their absolute difference. The discrete
density functions for X and Yare given in Example 6.
5
tC[Y] = LYjfY(yj) = L ifrCi)
i=O
= o· 366 + 1 • ~~
+ 2·-1L+3·~+4·--±"'+5·--L=70
36 36 36 36 36·
12
tC[X] = L ifx(i) =
1=2
7.
Note that tC[ Y] is not one of the possible values of Y. IIII
EXAMPLE 11 Let X be a continuous random variable with probability

density functionfx(x) = le- Ax/[o. oolx).
oo = foo xle- Ax dx = ,.
1
tC[X] =
f
-00
xfx(x) dx
0 I\.
The corresponding cumulative distribution function is
IIII
EXAMPLE 12 Let X be a random variable with cumulative distribution

function given by Fx<.x) = (1 - pe-A,/[o. oolx); then
tC[X]
oo
= f0 [1 - Fx(x)] dx -
fO
-00 Fx(x) dx =
foo0 pe- AX dx =;:.
P
Here, we have used Eq. (6) to find the mean of a random variable that is
partly discrete and partly continuous. IIII
EXAMPLE 13 Let X be a random variable with probability density function

given by fx(x) = x- 2 /[1,(0)(x); then
oo dx .
S[X] =
f1
x 2" = lIm loge b =
X b ...... oo
00,
so we say that S[X] does not exist.' We might also say that the mean of X
is infinite since it is clear here that the integral that defines the mean is
~~ @
4.2 Variance
The mean of a random variable X, defined in the previous subsection, was a
measure of central location of the density of X. The variance of a random vari-
able X will be a measure of the spread or dispersion of the density of X.
Definition 11 Variance Let X be a random variable, and let Ilx be

S[x]. The variance of X, denoted by O'i or var [X], is defined by
(i) var [X] = L (Xj -
j
Ilx)2fx(xj) (7)
if X is discrete with mass points Xh x2 , ..• , xj ' ••••
(ii) var [X] = f_ 00
00 (x - Ilx)2fx(x) dx (8)
if X is continuous with probability density functionfx(x).
(iii) var [X] = fo

00
2x[1 - Fx(x) + Fx( - x)] dx - Ili (9)
for an arbitrary random variable X. IIII

The variances are defined only if the series in (i) is convergent or if the
integrals in (ii) and (iii) exist. Again, the variance of a random variable is
defined in terms of the density function or cumulative distribution function of
the random variable; hence variance could be defined in terms of these functions
without reference to a random variable.
Note what the definition says: In (i), the square of the difference between
the jth value of the random variable X and the mean of X is multiplied by the
probability that X equals the jth value, and then these terms are summed.
More weight is assigned to the more probable squared differences. A similar
comment applies for (ii). Variance is a measure of spread since if the values
of a random variable X tend to be far from their mean, the variance of X will
be larger than the variance of a comparable random variable Y whose values
tend to be near their mean. It is clear from (i) and (ii) and true for (Hi) that
variance is nonnegative. We saw that a mean was the center of gravity of a
density; similarly (for those readers familiar with elementary physics or me-
chanics), variance represents the moment of inertia of the same density with
respect to a perpendicular axis through the center of gravity.
Definition 12 Standard deviation If X is a random variable, the

standard deviation of X, denoted by ax, is defined as + Jvar [Xl. IIII
The standard deviation of a random variable, like the variance, is a meas-

ure of the spread or dispersion of the values of the random variable. In many
applications it is preferable to the variance as such a measure since it will have
the same measurement units as the random variable itself.
EXAMPLE 14 Let X be the total of the two dice in the experiment of tossing
two dice.
var [X] = L(Xj - JlX)2/X(Xj)
= (2 - 7)2l6 + (3 - 7) 2l6 + (4 - 7)2 336 + (5 - 7)2 346
+ (6 - 7)2 356 + (7 - 7)2 366 + (8 - 7)2 356 + (9 - 7)2 346
+ (10 - 7)2 /6 + (11 - 7) 2l6 + (12 _7)2 316 = 2l6°. IIII
EXAMPLE 15 Let X be a random variable with probability density given by

/x(x) = ).e-;,xI[o, oo)(x); then
Var [X] = foo (x - Jlx)2fx(x) dx
r
- 00
AX
= (x - W.<e- dx
1
- A?' IIII
EXAMPLE 16 Let X be a random variable with cumulative distribution

given by Fx(x) = (1 - pe-;'X)I[o. OO)(x); then
Var [X] = foo 2x[1 - F(x) + F( -x)] dx - Jli
r
o
= 2xpe- AX dx - (~)'
= 2 ~_
).2
(E)). 2 _
-
p(2 - p)
).2 . IIII
4.3 Expected Value of a Function of a Random Variable

We defined the expectation of an arbitrary random variable X, called the mean
of X, in Subsec. 4.1. In this subsection, we will define the expectation of a
function of a random variable for discrete or continuous random variables.
Definition 13 Expectation Let X be a random variable and g(.) be

a function with both domain and counterdomain the real line. The
expectation or expected value of the function g(') of the random variable
X, denoted by ..&'[g(X)], is defined by:
(i) ..&'[g(X)] = L g(xj)fx(x j)
j
(10)
if X is discrete with mass points Xb x2 , .•. , Xj' .•. (provided this

series is absolutely convergent).
(ii) ..&'[g(X)] = f_oog(x)/x(x) dx

00
(11)
if X is continuous with probability density function fx(x) (provided

S~oolg(x)lfx(x) dx < (0).* IIII
Expectation or expected value is not really a very good name since it is

not necessarily what you "expect." For example, the expected value of a
discrete random variable is not necessarily one of the possible values of the
discrete random variable, in which case, you would not" expect" to get the
expected value. A better name might be " average value" rather than" expected
value."
Since -..&'[g(X)] is defined in terms of the density function of X, it could be
defined without reference to a random variable.
Remark If g(x) = x, then ..&'[g(X)] = ..&'[X] is the mean of X. If g(x) =

(x - f.lx)2, then ..&'[g(X)] = ..&'[(X - f.lX)2] = var [Xl IIII
* tf[g(X)] has been defined here for random variables that are either discrete or
continuous; it can be defined for other random variables as well. For the reader
who is familiar with the Stieltjes integral, C[g(X)J is defined as the Stieltjes integral
J~ oog(x) dFx(x) (provided this integral exists). where F x(·) is the cumulative distribu-
tion function of X. If X is a random variable whose cumulative distribution fUnction is
partly discrete and partly continuous. then (according to Subsec. 3.3) Fx(x) =
(l - p)r(x) + pPC(x) for some 0 < p < 1. Now tf[g(X)J can be defined to be tf[g(X)J
= (1- p) 2:g(X))fd(X)) + p J~ rz;)g(x)fac(x) dx, where fd(.) is the discrete density func-
tion corresponding to Fd(.) and r C ( . ) is the probability density function corre-
sponding to F ac (.).
Theorem 3 Below are properties of expected value:

0) S[c] = c for a constant c.
(ii) S[cg{X)] = cS[g(X)] for a constant c.
(iii) S[Clg1(X) + C2g2(X)] = ClS [gl{X)] + C2S[g2(X)].
(iv) S[gl(X)] S[g2{X)] if gl(X) < g2(X) for an x.
PROOF Assume X is continuous. To prove (i), take g(x) = c, then
S[g{X)] = S[c] = foo cfx(x) dx = c foo fx(x) dx = c.

-00 -00
00 00
S[cg(X)] = f_ 00 cg(x)/x(x) dx = c f_ 00 g(x)/x{x) dx = cS[g(X)],
which proves (ii). (iii) is given by
00 00
= Cl f- oo gl(X)/X(x) dx + Cz f- oo gix)/x(x) dx
= C1 S[gl(X)] + Cz S[g2(X)]'
Finally,
which gives (iv).

Similar proofs could be presented for the discrete random variable
~~. ffll
Theorem 4 If X is a random variable, var [X] = S[(X - S[XD2] =

S[X2] - (S[XD 2 provided 8[X2] exists.
PROOF CNe first note that if S[X2] exists, then S[X] exists.)* By
our definitions of variance and S[g(X)], it follows that var [X] =
S[eX - S[X])2]. Now S[(X - S[X])2] = S[X2 - 2XS[X] + (S[XD2] =
S[X2]- 2(S[X])2 + {S[X])2 = S[X2] - {S[X])2. 1//1
The above theorem provides us with two methods of calculating a vari-

ance, namely S[(X - IlX)l] or S[X2] Ili. Note that both methods require Ilx.
* Here and in the future We are not going to concern ourselves with checking existence.
8[g(X)] is used in each of the following three subsections. In Subsec. 4.4

and 4.5 two inequalities involving 8[g(X)] are given. Definitions and examples
of 8[g(X)1 for particular functions g(') are given in Subsec. 4.6.
4.4 Chebyshev Inequality

Theorem 5 Let X be a random variable and g(' ) a nonnegative function
with domain the rea11ine; then
c P[g(X) > k] < 8[g!X)] for every k > O. (12)

-.::.{.
~y.
" 1I'~·t.:.
PROOF Assume that X is a continuous random variable with

probability density functionfx('); then
8[g(X)] = f«>
-00
g(x)/x(x) dx f =
~g~)~~
g(x)/x<x) dx
+f {x:g(x)<t}
g(x)/x(x) dx > f g(x)/x{x) dx
{x: g(x)~t}
>f k/X<x) dx = kP[g(X) ~ k].

(x: .(x)~"}
Divide by k, and the result follows. A similar proof holds for X discrete.
1111
Corollary Chebyshev inequality If X is a random variable with finite
vanance,
for every r > O. (13)
PROOF Takeg(x) = (x - Px)2 and k = r 2 ui in Eq.(12) of Theorem 5.

1/11
Remark If X is a random variable with finite variance,
1
P[I X - Pxl < rux] ~ 1 - "2' (14)
r
which is just a rewriting of Eq. (13). 1/1/
The Chebyshev inequality is used in various ways. We will use it later to

prove the law of large numbers. No~e what Eq. (14) says:
P[Px - rux < X < Px + rux] > I - ..;.;

r
72 RANDOM VARIABLES, DISTRIBUTION FUNCTIONS. AND EXPECTATION II
that is. the probability that X falls within ru x units of Jlx is greater than or
equal to I - llr2. For r = 2, one gets P[Jtx - 2ux < X < Jlx + 2u x ] > t, or
for any random variable X having finite variance at least three-fourths of the
mass of X falls within two standard deviations of its mean.
Ordinarily, to calculate the probability of an event described in terms of
a random variable X, the distribution or density of X is needed; the Chebyshev
inequality gives a bound, which does not depend on the distribution of X, for the
pro bability of particular events described in terms of a random variable and
its mean and variance.
4.5 Jensen Inequality
Definition 14 Convex function A continuous function g( . ) with domain

and counterdomain the real line is called convex if for every Xo on the
real line, there exists a line which goes through the point (xo, 'g(xo)) and
lies on or under the graph of the function g('). IIII
Theorem 6 Jensen inequality Let X be a random variable with mean

tS'[X], and let g(') be a convex function; then tS'[g(X)] ~ g(tS'[X]).
PROOF Since g(x) is continuous and convex, there exists a line, say
/(x) = a + bx, satisfying /(x) = a + bx < g(x) and /(tS'[X]) = g(tS'[X]).
/(x) is a line given by the definition of continuous and convex that goes
through the point (tS'[X], g(tS'[X])). Note that tS'[/(X)] = tS'[(a + bX)] =
a + btS'[X] = /(tS'[X]); hence g(tS'[X]) = /(tS'[X]) = tS'[/(X)] < tS'[g(X)] [using
property {iv} of expected values (see Theorem 3) for the last inequality].
IIII
The Jensen inequality can be used to prove the Rao-Blackwell theorem to

appear in Chap. VII. We point out that, in general, tS'[g(X)] ¥- g(tS'[X]); for
example, note that g(x) = x 2 is convex; hence tS'[X2] > (tS'[X])2, which says
that the variance of X, which is tS'[X2] - (tS'[X]) 2 , is nonnegative.
4.6 Moments and Moment Generating Functions
The moments (or raw moments) of a random variable or of a distribution are

the expectations of the powers of the random variable which has the given
distribution.
Definition 15 Moments If X is a random variable, the rth moment of

X, usually denoted by Il;, is defined as
Il; = G[X'] (15)
if the expectation exists. IIII
Note that Il~ = G[X] = Ilx, the mean of X.
Definition 16 Central moments If X is a random variable, the rth

central moment of X about a is defined as G[(X - aYl If a = Ilx, we
have the rth central moment of X about Ilx, denoted by Il" which is
Il, = G[(X - IlxYl (16)
IIII
Note that III = G[(X - Ilx)] = 0 and 112 = G[(X - IlX)2], the variance of X.
Also, note that all odd moments of X about Ilx are 0 if the density function of X
is symmetrical about Ilx, provided such moments exist.
In the ensuing few paragraphs we will comment on how the first four
moments of a random variable or density are used as measures of various
M
, .characteristics of the corresponding density. For some of these characteristics,
'.' other measures can be defined in terms of quantiles.
Definition 17 Quantile The qth quantile of a random variable X or of

e
its corresponding distribution is denoted by q and is defined as the
e
smallest number satisfying Fx(e) > q. IIII
If X is a continuous random variable, then the qth quantile of X is given as

e
the smallest number satisfying Fx(e) = q. See Fig. 6.
Definition 18 Median The median of a random variable X, denoted by

medx , med (X), or (so, is the .5th quantile. IIII
Remark In some texts the median of X is alternatively defined as any

number, say med (X), satisfying P[X ~ med (X)] > t and P[X > med (X)]
~ t· IIII
If X is a continuous random variable, then the median of X satisfies
mea (X)
f L
c£)
fx(x) dx =t = fx(x) dx;

- c£) med (X)
Fx(x)
1.0
1 Fx(x)
.75
.50
.25
x
FIGURE 6 0
So the median of X is any number that has half the mass of X to its right and
the other half to its left, which justifies use of the word median."
H
We have already mentioned that8[X], the first moment, locates the" center"
of the density of X. The median of X is also used to indicate a central
location of the density of X. A third measure of location of the density of X,
though not necessarily a measure of central location, is the mode of X, which is
defined as that point (if such a point exists) at whkh fx(') attains its maximum.
Other measures of location [for example, t('.25 + '.75)] could be devised, but
three, mean, median, and mode, are the ones commonly used.
We previously mentioned that the second moment about the mean, the
variance of a distribution, measures the spread or dispersion of a distribution.
Let us look a little further into the manner in which the variance characterizes
the distribution. Suppose that It (x) and f2(x) are two densities with the same
mean f.l such that
p+a
{-a [fleX) - f2(X)] dx > 0 (17)
for every value of a. Two such densities are illustrated in Fig. 7. It can be
shown that in this case the variance ai in the first density is smaller than the
FIGURE 7
FIGURE 8
variance CT~ in the second density. We shall not take the time to prove this in
detail, but the argument is roughly this: Let
g(x) = It (x) - f2(X) ,
co
where It (x) and f2(X) satisfy Eq. (17). Since S g(x) dx = 0, the positive area
- co
between g(x) and the x axis is equal to the negative area. Furthermore, in
view of Eq. (17), every positive element of area g(x') dx' may be balanced by a
negative element g(x") dx" in such a way that x" is further from J-l than x'.
When these elements of area are multiplied by (x - J-l)2, the negative elements
will be multiplied by larger factors than their corresponding positive elements
(see Fig. 8); hence
JCO (x - J-l)2 g (X) dx < 0
-00
unless It (x) and f2(X) are equal. Thus it follows that ui < u~ . The converse
of these statements is not true. That is, if one is told that ui < u~ , he cannot
conclude that the corresponding densities satisfy Eq. (17) for all values of a;
although it can be shown that Eq. (17) must be true for certain values of a.
Thus the condition ui < u~ does not give one any precise information about
the nature of the corresponding distributions, but it is evident that It (x) has
more area near the mean thanf2(x), at least for certain intervals about the mean.
We indicated above how variance is used as a measure of spread or
dispersion of a distribution. Alternative measures of dispersion can be defined
in terms of quantiles. For example 7S -e. e.
2S , called the interquartile range,
is a measure of spread. Also, p - e el- p for some -!<p < I is a possible
measure of spread.
The third moment J-l3 about the mean is sometimes called a measure of
asymmetry, or skewness. Symmetrical distributions like those in Fig. 9 can be
shown to have J-l3 = O. A curve shaped likelt(x) in Fig. 10 is said to be skewed
to the left and can be shown to have a negative third moment about the mean;
one shaped like f2{X) is called skewed to the right and can be shown to have a
positive third moment about the mean. Actually, however, knowledge of the
--~--~------------~--------------~~-------x
FIGURE 9
third moment gives almost no clue as to the shape of the distribution, and we
mention it mainly to point out that fact. Thus, for example, the density f3(x)
in Fig. 10 has /13 = 0, but it is far from symmetrical. By changing the curve
slightly we could give it either a positive or negative third moment. The ratio
/13/(13, which is unitiess, is called the coefficient of skewness.
The quantity 11 = (mean - median)/(standard deviation) provides an
alternative measure of skewness. It can be proved that -1 < 11 < 1.
The fourth moment about the mean is sometimes used as a measure of
excess or kurtosis, which is the degree of flatness of a density near its center.
Positive values of /14-/(14 - 3, called the coefficient 0/ excess or kurtosis, are
sometimes used to indicate that a density is more peaked around its center than
the density of a normal curve (see Subsec. 3.2 of Chap. III), and negative values
are sometimes used to indicate that a density is more flat around its center than
the density of a normal curve. This measure, however, suffers from the same
failing as does the measure of skewness; namely, it does not always measure
what it is supposed to.
While a particular moment or a few of the moments may give little
information about a distribution (see Fig. 11 for a sketch of two densities having
the same first four moments. See Ref. 40. Also see Prob. 30 in Chap. Ill),
the entire set of moments (J1~, /1~, f.L;, ...) will ordinarily determine the distri-
~X~) __LL__ ~ ~=-____________~_

__
FIGURE 10
.7
.6
.5
.4
.3
.2
.1
-2
FIGURE 11
bution exactly, and for this reason we shall have occasion to use the moments
in theoretical work.
In applied statistics, the first two moments are of great importance, as
we shall see, but the third and higher moments are rarely useful. Ordinarily
one does not know what distribution function one is working with in a practical
problem, and often it makes little difference what the actual shape of the distri-
bution is. But it is usually necessary to know at least the location of the
distribution and to have some idea of its dispersion. These characteristics can
be estimated by examining a sample drawn from a set of objects known to have
the distribution in question. This estimation problem is probably the most
important problem in applied statistics, and a large part of this book will be
devoted to a study of it.
We now define another kind of moment,/actorial moment.
Definition 19 Factorial moment If X is a random variable, the rth

/actorial moment 0/ X is defined as (r is a positive integer):
CrX(X - 1) ... (X - r + I)]. (18)
I111
For some random variables (usually discrete), factorial moments are
78 RANDOM VARlABLES, DISTRIBUTION FUNCTIONS, AND EXPECTATION
easier to calculate than raw moments. However the raw moments can be
obtained from the factorial moments and vice versa.
The moments of a density function play an important role in theoretical
and applied statistics. In fact, in some cases, if all the moments are known,
the density can be determined. This will be discussed briefly at the end of
this subsection. Since the moments of a density are important, it would be
useful if a function could be found that would give us a representation of all
the moments. Such a function is called a moment generating function.
Definition 20 Moment generating function Let X be a random vari-

able with density fx(')' The expected value of etX is defined to be the
moment generating function of X if the expected value exists for every
value of t in some interval - h < t < h; h > O. The moment generating
function, denoted by mx(t) or met), is
met) = G[etX ] = foo etxfx(x) dx

-00
(19)
if the random variable X is continuous and is

met) = 8[etX ] = Lx etxfx(x)
if the random variable is discrete. IIII
One might note that a moment generating function is defined in terms of
a density function, and since density functions were defined without reference
to random variables (see Definitions 6 and 9), a moment generating function
can be discussed without reference to random variables.
If a moment generating function exists, then met) is continuously differ-
entiable in some neighborhood of the origin. If we differentiate the moment
generating function r times with respect to t, we have
d'
- , met)
dt
= f
-00
00
x'e x 1x(x)dx, (20)
and letting t -+ 0, we find

d'
dt' m(O) = 8[X1 = Il;, (21)
where the symbol on t~e left is to be interpreted to mean the rth derivative of
met) evaluated as t -+ O. Thus the moments of a distribution may be obtained
from the moment generating function by differentiation, hence its name.
If in Eq. (19) we replace ext by its series expansion, we obtain the series
expansion of met) in terms of the moments of fx( •); thus
'4 EXPECTATIONS AND MoMENrS 79
met) = 4[1 + Xt + ~(Xt)2 + ~ (Xt)3 + ... ]

2! 3.
1
= 1 + r1 t + -2! r2 t 2 + ...
II' II'
oo
1 , j
(22)
= L0 I. -., rJ t ,
II.
j=
from which it is again evident that J.l; may be obtained from met); J.l; is the co-
efficient of trlr!.
EXAMPLE 17 Let X be a random variable with probability density function

given by fx(X) = Ae- AX1[0, oo)(x).
oo l
mx(t) = 4[lX] =
fo lXle-h dx = --
1-t
for t < 1.
dm(t) 1 , 1
m'(t) = = hence m (0) = 8[X] = -.
dt (1 - t)2 1
And " 21
m (t) = (1 _ t)3 ' so m"(O) = 8[X2] = 1~' IIII
EXAMPLE 18 Consider the random variable X having probability density

functionfx(x) = x- 2 /[1, oolx). (See Example 13.) If the momentgenerat-
ing function of X exists, then it is given by x- 2e tx dx. It can be Jf
shown, however, that the integral does not exist for any t > 0, and hence
the moment generating function does not exist for this random variable X.
1II1
As with moments, there is also a generating function for factorial mODlents.
Definition 21 Factorial moment generating function Let X be a ran-

dom variable. The factorial moment generating function is defined as
8[t X ] if this expectation exists. IIII
The factorial mODlent generating function is used to generate factorial
moments in the same way as the raw moments are obtained from 8[e tX ] except
that t approaches 1 instead of O. It sometimes simplifies finding moments
of discrete distributions.
80 RANDOM VARIABLES, DISTRIBUTION FUNCTIONS, AND EXPECTATION IJ
EXAMPLE 19 Suppose X has a discrete density function given by
e-;'A:~
fx(x) = - for x = 0, 1, 2, ....
x!
Then
d
hence - B[ tX] = A. 1111
dt t=)
In addition to raw moments, central moments, and factorial moments,

there are other kinds of moments, called cumulants, or semi-invariants. Cumu-
lants will be defined in terms of the cumulant generating function. We will not
make use of cumulants in this book. .,
Definition 22 Cumulant andcumulant generating function The logarithm

of the moment generating function of X is defined to be the cumulant
generating function of X. The rth cumulant of X, denoted by "r( X) or "r'
is the coefficient of trlr! in. the Taylor series expansion of the cumulant
generating function. 1111
A moment generating function is used, as its name· suggests, to generate

moments. That, however, will not be its only use (or us. An important use
will be in determining distributions.
Theorem 7 Let X and Y be two random variables with densities fx(')

andfy(-), respectively. Suppose that mx(t) and my(t) both exist and are
equal for all t in the interval - h < t < h for some h > O. Then the two
cumulative distribution functions Fx (') and Fy(-) are equal. 111I
A proof of the above theorem can be obtained using certain transfornl

theory that is beyond the scope of this book. We should note, however, what
the theorem asserts. It says that if we can find the moment generating function
of a random variable, then, theoretically, we can find the distribution of the
random variable since there is a unique distribution function for a given moment
geneniting function. This theorem will prove to be extremely useful in finding
the distribution of certain functions of random variables. In particular, see
Sec. 4 of Chap. V.
PROBLEMS 81
EXAMPLE 20 'Suppose that a random yariab'le ,x has a moment generating

function mx(t) = 1/(1 - t) for -}. <' t < 1; then we know that the den-
sity of X is given by /x(x) = e- x I[o,d:J) (x) since we showed in Example
17 above that Ae-.ax I[o,d:J)(x) has A/(A - t) for its moment generating
function. / //1
Problem of moments We have seen that a density function determines a set

of moments Ill, ~;, . .. when they exist. One of the im'portant problems in
theoretical statistics is this: Given a set of moments, what is the density function
from which these moments came, and is there only one density function that
has these particular moments? We shall give only partial answers. First,
there exists a sequence of moments for which there is an infinite (nondenumer-
able) collection of different distribution functions having these same moments.
In general, a sequence of moments Il~, Il;, ... does not determine a unique
distribution function. However, we did see that if the moment generating
function of a random variable did exist, then this moment generating function
did uniquely determine the corresponding distribution function. (See Theorem
7 above.) Hence, there are conditions (existence of'the moment generating
function is a sufficient condition) under which a sequence of moments does
uniquely determine a distribution function. The general problem of whether or
not a distribution function is determined by its sequence of moments IS
referred to as the problem 0/ moments and will not be discussed further.
PROBLEMS
J (a) Show that the following are probability density functions (p.d.f.'s):
hex) = e-x/(o.OO)(x}
f2(X} = 2e- 2X/(0. OO)(x}
f(x) = «() + l}};(x) - ()f2(X)
Prove or disprove: If hex) and f2(x) are p.d.f.'s and if ()1

(b) ()2 = 1, then
()lh(x) + ()2f2(x) is a p.d.f.
2 Show that the following is a density function and find its median:
a;2(a; + 2x) x(2ac + x) ,
f(x) = x2(a; +X)2 I«I.oo)(x)+ ac(ac + X)2 /(0. «](x), for a; > 0
3 Find the constant K so that the following is a p.d.f.

f(x) = Kx 2 I(_K, KJ(X).
4 Suppose that the cumulative distribution function (c.d.f.) Fx(x) Can be written
as a function of (x - a.)/fJ, where a. and fJ > 0 are constants; that is, x, a., and fJ
appear in Fx( .) only in the indicated form.
(a) Prove that if a. is increased by aa., then so is the mean of X.
(b) Prove that if fJ is multiplied by k(k > 0), then so is the standard deviation
of X.
5 The experiment is to toss two balls into four boxes in such a way that each ball
is equally likely to fall in any box. Let X denote the number of balls in the first
box.
(a) What is the c.d.f. of X?
(b) What is the density function of X?
(c) Find the mean and variance of X.
6 A fair coin is tossed until a head appears. Let X denote the number of tosses
required.
(a) Find the density function of X.
(b) Find the mean and variance of X.
(c) Find the moment generating function (m.g.f.) of X.
*7 A has two pennies; B has one. They match pennies until one of them has all
three. Let X denote the number of trials required to end the game.
(a) What is the density function of X?
(b) Find the mean and variance of X.
(c) What is the probability that B wins the game?
8 Let fx(x) =(1!fJ)[1 -I (x - a.)/fJl ]/(II-p. a+plx), where IX and (3 are fixed con-
stants satisfying - 00 < a. < 00 and fJ > O.
(a) Demonstrate that fx(') is a p.d.f., and sketch it.
(b) Find the c.d.f. corresponding to fx(·).
(c) Find the mean and variance of X.
(d) Find the qth quantile of X.
9 Letfx(x) = k(l/fJ){1 - [(x - a.)/{3]2}I(rJ-p. Cl+p,(X), where - 00 < IX < 00 and (3 > O.
(a) Find k so that/xC') is a p.d.f., and sketch the p.d.f.
(b) Find the mean, median, and variance of X.
(c) Find 8[1 X - a.1]·
(d) Find the qth quantile of X.
10 Let fx(x) = t{O/(o. 1 ,(x) + 1[1. 2)(X) (1 - 0)/(2.3,(X)}, where 0 is a fixed constant
satisfying 0 0 ~ 1.
(a) Find the c.dJ. of X.
(b) Find the mean, median, and variance of X.
J1 Let f(x; 0) =' Of(X; 1) + (1 - O)f(x; 0), where 0 is a fixed constant satisfying
o < 0 ~ 1. Assume that/(·; 0) andf(·; 1) are both p.d.f.'s.
(a) Show that f( . ; 0) is also a p.d.f.
(b) Find the mean and variance of f(· ; 0) in terms of the mean and variance of
f(' ; 0) and f(' ; 1), respectively.
(c) Find the m.g.f. of/('; 0) in terms of the m.g.f.'s of/('; 0) andf(·; 1).
PROBLEMS 83
J2 A bombing plane flies directly above a railroad track. Assume that if a large
(small) bomb falls within 40 (15) feet of the track, the track will be sufficiently
damaged so that traffic will be disrupted. Let X denote the perpendicular
distance from the track that a bomb falls. Assume that
100-x
/x(x) = l[o.loo)(x).
5000
(a)Find the probability that a large bomb will disrupt traffic.

(b)If the plane can carry three large (eight small) bombs and uses all three
(eight), what is the probability that traffic will be disrupted?
13 (a) Let X be a random variable with mean f' and variance a 2 • Show that
cf[(X - b)2], as a function of b, is minimized when b = 1-'.
*(b) Let X be a continuous random variable with median m. Minimize cf[ I X b I]
as a function of b. HINT: Show that 8[ IX - b I] = 8[ IX - m I] +
2 S; (x - b}fx(x) dx.
14 (a) If X is a random variable such that 8[X] = 3 and
8[X2] = 13, use the
ChebYShev inequality to determine a lower bound for P[ - 2 < X 8].
(b) Let X be a discrete random variable with density
/x(x) =l/(_l)(x) + f/(o)(x) + i/{1}(x).

For k = 2 evaluate P[I X -I-'xl > kax]. (This shows that in general the
Chebyshev inequality cannot be improved.)
(c) If X is a random variable with 8[X] = I-' satisfying P[X ~ 0] = 0, show that
P[X> 21-'] ~ l.
15 Let X be a random variable with p.d.f. given by
/x(x) = 11 x I1[0. 2J(X).
Find the mean and variance of X.

16 Let X be a random variable having c.d.f.
Fx(x) = pH(x) (1 - p)G(x),
where p is a fixed real number satisfying 0 < p < 1,
and
(a) Sketch Fx(x) for p = i.

(b) Give a formula for the p.d.f. of X or the discrete density function of X,
whichever is appropriate.
. (c) Evaluate p[X ~.I X ~ 1].
17 Does there exist a random variable X for which P[p.x - 2ax X s;, I-'x + 2a x] = .6?
18 An urn contains balls numbered 1, 2, 3, First a ball is drawn from the urn,
and then a fair coin is tossed the number of times as the number shown on the
drawn ball. Find the expected number of heads.
19 If X has distribution given by P[X = 0] = P[X = 2] = p and P[X = 1] = 1 - 2p
for 0 <p< i, for what p is the variance of X a maximum?
20 If X is a random variable for which P[X 0] = 0 and S[X] fL < 00, prove that
P[X fLt] > 1 -l/t for every t 1.
21 Given the c.d.f.

Fx{x) = 0 for x < 0
forO <x <.5
= x for.5 x< I
= I for I <X.
(a) Express Fx{x) in terms of indicator functions.
(b) Express Fx{x) in the form
aPC(x) + bP(x),
where Fac(.) is an absolutely continuous c.d.f. and P{·) is a discrete c.dJ.
(c) Find P[.25 < X .75].
(d) Find P[,25 < X < .5],
22 Let/{x) = Ke-lIx(l - e-U)/(o. ro){x).
(a) Find K such that /( .) is a density function.
(b) Find the corresponding c.d.f.
(c) Find P[X > 1],
23 A coin is tossed four times. Let X denote the number of times a head is followed
immediately by a tail. Find the distribution, mean, and variance of X.
24 Let /x(x; () = ({)x + 1)/(- J. u(x), where () is a constant.
(a) For what range of values of () is /x{· ; () a density function?
(b) Find the mean and median of X.
(c) For what values of () is var [X] maximized?
25 Let X be a discrete random variable with the nonnegative integers as values.
Note that S[tX] = L tJp[X =
J=O
j]. Hence,S[t X] is a probability generating function
of X, inasmuch as the coefficient of t J gives P[X = n. Find S[t X] for the random
variable of Probs. 6 and 7.
III
SPECIAL PARAMETRIC FAMILIES OF
UNIV ARIA TE DISTRIBUTIONS

The purpose of this chapter is to present certain parametric families of univariate
density functions that have standard names. A parametric family of density
functions is a collection of density functions that is indexed by a quantity called
a parameter. For example, letf(x; l) le-·Ilx/(o. ro)(x), where A. > 0; then for
each l > 0, f(· ; l) is a probability density function. l is the parameter, and as
A. ranges over the positive numbers, the collection {f(· ; l): A. > O} is a parametric
family of density functions.
The chapter consists of three main sections: parametric families of dis..
crete densities are given in one; parametric families of probability density func-
tions are given in another, and comments relating the two are given in the final
section. For most of the families of distributions introduced, the means,
variances, and moment generating functions are presented; also, a sketch of
several representative members of a presented family is often included. A
table summarizing results of Sees. 2 and 3 is given in Appendix B.
86 SPECIAL PARAMETRIC FAMIUES OF UNIVARIATE DlSTRIBUnONS III
2 DISCRETE DISTRIBUTIONS
In this section we list several parametric families of univariate discrete densities.
Sketches of most are given; the mean and variance of each are derived, and usually
examples of random experiments for which the defined parametric family
might provide a realistic model are included.
The parameter (or parameters) indexes the family of densities. For
each family of densities that is presented, the values that the parameter can
assume will be specified. There is no uniform notation for parameters; both
Greek and Latin letters are used to designate them.
2.1 Discrete Uniform Distribution
Definition 1 Discrete uniform distribution Each member of the family

of discrete density functions
1
for x = 1, 2, ... , N
1
f(x) =f(x; N) = N -
-N- I {l.Z ••.•• N} (x) , (1)
o otherwise
where the parameter N ranges over the positive integers, is defined to have
a discrete uniform distribution. A random variable X having a density
given in Eq. (1) is called a discrete uniform random variable. 1111
l/Nlu~
o 2 3
I ---~ __ _
FIGURE 1
Density of discrete uniform.
Theorem 1 If X has a discrete uniform distribution, then 8[X] =

(N + 1)/2,
(N Z -1) N 't 1
' and mx(t) = 8[e ] = j~l e' N'
tX
var [X] = 12
2 DISCRETE DISTRIBUTIONS 87
PROOF
var [X] = S[X2] - (S[X])2

N 1
= jf./2 N -
(N 2+ 1)2
N(N + 1)(2N + 1) (N + 1)2 (N + l)(N - 1)
- -
6N 4 ]2
11II
Remark The discrete uniform distribution is sometimes defined in
density form as I(x; N) = [1/(N + l)]I{o. I. ...• N} (x), for N a nonnegative
integer. Jf such is the case, the formulas for the mean and variance have
to be modified accordingly. /11I
2.2 Bernoulli and Binomial Distributions
Definition 2 Bernoulli distribution A random variable X is defined

to have a Bernoulli distribution if the discrete density function of X is
given by
for x 0 or I} ,
= pX(l - p)1 -xI lO • l}(x), (2)
otherwise
where the parameter p satisfies 0 <p < 1. p is often denoted by q.

111/
FIGURE 2
Bemou1li density.
88 SPECIAL PARAMETRIC FAMILIES OF UNIVARIATE DISTRIBUTIONS III
Theorem 2 If X has a Bernoulli distribution, then

8[X] = p, var [X] = pq, and mx(t) = pet + q. (3)
PROOF 8[X] = 0 . q + 1 . p = p.
var [X] = 8[X2] - (8[X])2 = 0 2 • q + 12 • p _ p2 = pq.
mx(t) = 8[e tX ] = q + pet. IIII
EXAMPLE 1 A random experiment whose outcomes have been classified
into two categories, called" success" and" failure," represented by the
letters d and I, respectively, is called a Bernoulli trial. If a random
variable X is defined as 1 if a Bernoulli trial results in success and 0 if
the same Bernoulli trial results in failure, then X has a Bernoulli distribu-
tion with parameter p = P[success]. 1III
EXAMPLE 2 For a given arbitrary probability space (0, d, P[·]) and for A
belonging to d, define the random variable X to be the indicator function
of A; that is, X(w) 1.4.(w); then X has a Bernoulli distribution with
parameter p = P[X 1] = P[A]. II1I
Definition 3 Binomial distribution A random variable X is defined to
have a binomial distribution if the discrete density function of X is given
by
n)
fx(x) =fx(x; n, p) = {( ~ p q
x n-x for x = 0, I, ... , n
otherwise (4)
11 10. p =.25 11 "" lO,p .5
o 1 2
-l~.~, Illll
4 5 6 7 8 9 10
I
o• •I 2 3 4 5 6
I • •
7 8 9 10
.. x
11 '" 5, P .2 11 '" 5. P =.4 11 -= 5. p ,6
012345
x L
o I
FIGURE 3
Binomial densities.
where the two parameters nand p satisfy 0 <p ~ 1, n ranges over the
positive integers, and q = I - p. A distribution defined by the density
function given in Eq. (4) is called a binomial distribution. IIII
Theorem 3 If X has a binomial distribution, then

8[X] = np, var [X] = npq, and (5)
PROOF
mx(t) = 8[etX ] =
x=o
± etx(n)pxqn-x =
X
t
x~o
(n)(petyqn-x
X
= (pet + qt.
Now
and
hence
S[X] = m~(O) = np
and
var [X] 8[X2] - (8[X])2
= mi(O) - (np)Z = n(n - l)p2 + np - (np)2 = np(l - p). I1II
Remark The binomial distribution reduces to the Bernoulli distribution
when n = I. Sometimes the Bernoulli distribution is called the point
binomial. IIII
EXAMPLE 3 Consider a random experiment consisting of n repeated inde-

pendent Bernoulli trials when p is the probability of success a at each
individual trial. The term "repeated" is used to indicate that the prob-
ability of 6 remains the same from trial to trial. The sample space for
such a random experiment can be represented as follows:
Q = {(ZI' zz, ... , zn): Zi = 6 or Zi =/}.
Zt indicates the result of the "ith trial. Since the trials are independent,
the probability of any specified outcome, say {(/' /' a, /, a, 6, ... , / , a)},
is given by qqpqpp ... qp. Let the random variable X represent the num-
ber of successes in the n repeated independent Bernoulli trials. Now
P[X = x] = P[exactly x successes and n - x failures in n trials]
(:) p'q"- x for x = 0, 1, ... , n since each outcome of the experimentthat has
exactly x successes has probability p"q'-x and there are (:) such outcomes.
Hence X has a binomial distribution. I1I1
EXAMPLE 4 Consider sampling with replacement from an urn containing

M balls, K of which are defective. Let X represent the number of defec-
tive balls in a sample of size n. The individual draws are Bernoulli trials
where "defective" corresponds to "' success," and the experiment of
taking a sample of size n with replacement consists of n repeated inde-
pendent Bernoulli trials where p P[success] = KIM; so X has the
binomial distribution
for x = 0, 1, ... , n, (6)
which is the same as P[Ad in Eq. (3) of Subsec. 3.5 of Chap. I, for x = k.
1III
The sketches in Fig. 3 seem to indicate that the termsfx{x; n, p) increase

monotonically and then decrease monotonically. The following theorem states
that such is indeed the case.
Theorem 4 Let X have a binomial distribution with density fx(x; n, p);

then fX<x - 1; n, p) <fx(x; n, p) for x < (n + l)p; fx(x - 1; n, p) >
fx(x; n, p) for x > (n + l)p, andfx<x - 1; n,p) = fx(x; n, p) if x = (n + l)p
and (n + l)p is an integer, where x ranges over 1, ... , n.
PROOF
ix(x; n,p) n- x +1 p (n + 1)p - x

- '- 1+ ,
fx(x - 1; n,p) x q xq
which is greater than 1 if x < (n + l)p, smaller than 1 if x > (n + l)p,

and equal to 1 if the integer x should equal (n + l)p. I111
2.3 Hypergeometric Distribution

Definition 4 Hypergeometric distribution A random variable X is
defined to have a hypergeometric distribution if the discrete density function
of X is given by
for x = 0, 1, ... , n
fx(x; M, K, n) =
o otherwise (7)
Iro.I, .... n}(x)
where M is a positive integer, K is a nonnegative integer that is at most M,

and n is a positive integer that is at most M. Any distribution function
defined by the density function given in Eq. (7) above is called a hyper-
geometric distribution. III1
Theorem 5 If X is a hypergeometric distribution, then
&[X] = n ' -
K
and var [Xl = n.- . K M-K . -
M-n
- (8)
M M M M-t
PROOF
(K)(M - K) (K - l)(M - K)
8[ Xl = t x x n- x =n. K t x-I n - x
FO (~) M x=, (~~n
(Ky- 1) (M n-l-y
Kn-I
- 1- K+ 1)
=n .- L ...:.----....;..---:~~-:-----.:;...~
M (M - 1)
y=O
n-l
K
-- n 'M'
-
using f(a)( m b- i )=(a+b)

i=O i m
given in Appendix A.
M ]0; K = 4; n = 4 M = ]0; K 4; n =5
~ __~__~~_~I__~.____. x
o 2 3 4
FIGURE 4
Hypergeometric densities.
8[X(X - 1)]
= Ix(x-l) (~)(~=~)
.=0 (~)
K- 2)(M - K)
=n(n_1)K(K-1)
M(M - l)x=2
± x-2 n-x
(
2) (M -
n-2
(K- 2) (M - 2- K+ 2)
= n(n _ 1) K(K - 1)
I
M(M - 1) y""o
n 2 n- 2- Y
y
(M - 2)
= n(n _ 1) K(K - 1) .
M(M -1)
n-2
Hence
var [X] = 8[X2] - (8[X])2 = 6"[X(X - 1)] + S[X] - (6"[X])2
= n(n _ 1) K(K - 1) +n K _ n 2 K2
M(M-1) M M2
~]
K [ K-1
=n- (n-l). +1
M M-l
= nK [(M - K)(M - n)].

M M(M -1) 1111
Remark If we set KIM = p, then the mean of the hypergeometric dis-

tribution coincides with the mean of the binomial distribution, and the
variance of the hypergeometric distribution is (M - n)/(M - 1) times the
variance of the binomial distribution. 1111
EXAMPLE 5 Let X denote the number of defectives in a sample of size n

when sampling is done without replacement from an urn containing M
balls, K of which are defective. Then X has a hypergeometric distribution.
See Eq. (5) of Subsec. 3.5 in Chap. I. IIII
2.4 Poisson Distribution
Definition S Poisson distribution A random variable X is defined to

have a Poisson distribution if the density of X is given by
e-A.li X
= fx(x; A) =~
I x!
for x = 0, 1, 2, ... 1
e -A.'X
),
fx(x)
I
r=
1 xl I (D. I .... }(x), (9)
(0 otherwise
where the parameter Ii satisfies ), > O. The density given in Eq. (9) is
called a Poisson density. IIII
Theorem 6 Let X be a Poisson distributed random variable; then

6"[X] = Ii, var [X] = Ii, (10)
.607
.1.=t .368 .368

.303 A.-]
0
G
.~--~
I 2
.0I3
3
.002
4• -e--... x
5
I_J
0
.184
I
2
.06]
t
3
.015
•
4
.003
•5
..x
.073
_~.0!8 t .005 .002 .001
o I . •
IO
•
11
---4---+ X
12
FIGURE 5
Poisson densities.
PROOF
hence,
and
So,
G[X] = m~(O) = A
and
The Poisson distribution provides a realistic model for many random

phenomena. Since the values of a Poisson random variable are the nonnega-
tive integers, any random phenomenon for which a count of some sort is of
interest is a candidate for modeling by assuming a Poisson distribution. Such
a count might be the number of fAtal traffic accidents per week in a given state,
the number of radioactive particle emissions per unit of time, the number of
telephone calls per hour coming into the switchboard of a large business, the
number of meteorites that collide with a test satellite during a single orbit,
the number of organisms per unit volume of some fluid, the number of defects
per unit of some material, the number of flaws per unit length of some wire,
etc. Naturally, not all counts can be realistically modeled with a Poisson dis-
tribution, but some can; in fact, ifcertain assumptions regarding the phenomenon
under observation are satisfied, the Poisson model is the correct model.
Let uS assume now that we are observing the occurrence of certain happen-
ings in time, space, region, or length. A happening might be a fatal traffic
accident, a particle emission, the arrival of a telephone call, a meteorite col-
lision, a defect in an area of material, a flaw in a length of wire, etc. We will
talk as though the happenings are occurring in time; although happenings
occurring in space or length are appropriate as well. The occurrences of the
happening in time could be sketched as in Fig. 6. An occurrence of a happen-
ing is represented by x; the sketch indicates that seven happenings occurred
between time 0 and ti.me t I' Assume now that there exists a positive quantity,
say v, which satisfies the following:
~-~~----------~--~~-------*--*-~--------~--------x
o
FIGURE 6
(i) The probability that exactly one happening will occur in a small
time interval of length h is approximately equal to vh, or prone happening
in interval of length h] = vh + o(h).
(ii) The probability of more than one happening in a small time interval
of length h is negligible when compared to the probability of just one
happening in the same time interval, or P[two or more happenings in
interval of length h] = o(h).
(iii) The num bers of happenings in nonoverlapping time intervals are
independent.
The term o(h), which is read some function of smaller order than 17,"
H
denotes an unspecified function which satisfies

lim o(h) = 0.
h ..... O h
The quantity v can be interpreted as the mean rate at which happenings occur per
unit of time and is consequently referred to as the mean rate of occurrence.
Theorem 7 If the above three assumptions are satisfied, the number of

occurrenceS of a happening in a period of time of length t has a Poisson
distribution with parameter ). = vt. Or if the random variable Z(t)
denotes the number of occurrenceS of the happening in a time interval
of length t, then P[Z(t) = z] = e vt(vt)z/z! for z = 0, 1,2, ....
We will outline two different proofs, neither of which is mathemati-
cally rigorous.
For convenience, let t be a poi nt in time after time 0; so the

PROOF
time interval (0, t] has length t, and the time interval (t, t h] has length
h. Let Pn(s) = P[Z(s) = n] = P[exactly n happenings In an interval of
length s]; then
Poet + h) =P[no happenings in interval (0, t + h]]
= P[no happenings in (0, t] andno happenings in (t, til]]
= P[no happenings in (0, t]]P[no happenings in (t, t + h]]
= Po(t)Po(h),
using (iii), the independence assumption.
Now P[no happenings in (t, t + h]] = 1 - prone or more happenings

in (t, t + h]] = 1 - prone happening in (t. t + h]] - P[more than one
happening in (t, t + h]] = I - vh - o(h) - o(h); so Po(t + h) = Po(t)
[1 - vh - o(h) - o(h)], or
poet + h) - poet) ( ) o(h) + o(h)
h = - "~poet) - Po t h '
and on passing to the limit one obtains the differential equation P~Ct) =
- vPo(t), whose solution is Po(t) = e vr, using the condition PoCO) = 1.
Similarly, PtCt + h) = Pt(t)Po(h) + Po(t)Pt(h), or P1(t + h) = P.(t)[l - vh
- o(h)] + PoCt)[vh + o(h)], which gives the differential equation P~(t) =
- vPI(t) + vPo(t), the solution of which is given by PICt) = vte- vr , using
the initial condition PI (0) = 0. Continuing in a similar fashion one
obtains P~(t) = - vPII (t) + vP1I-.(t), for n = 2, 3, ....
It is seen that this system of differential equations is satisfied by
P II (t) = (vt)lI e -vt/n L
The second proof can be had by dividing the interval (0, t) into, say
n time subintervals, each of length h = tin. The probability that k
happenings occur in the interval (0, t) is approximately equal to the prob-
ability that exactly one happening has occurred in each of k of the 11
subintervals that we divided the interval (0, t) into. Now the probability
of a happening, or success," in a given subinterval is vh. Each sub-
H
interval provides us with a Bernoulli trial; either the subinterval has a

happening, or it does not. Also, in view of the assumptions made, these
Bernoulli trials are independent, repeated Bernoulli trials; hence the
probability of exactly k "successes" in the n trials is given by (see
Example 3)
(Z)( vh)k(l - vh)"-t = (~) [~n 1 - ~rk,

which is an approximation to the desired probability that k happenings will
occur in time interval (0, t). An exact expression can be obtained by
letting the number of subintervals increase to infinity, that is, by letting n
tend to infinity;
(kn) [vt]k [1 _ W]II-k = _1 (vtt [1 __vt]lI-k C_n)k -+ _(v_tt_e-_

k
vt
n n k! n n k!
as n -+ 00, . that [1v

noting - t-;;
] 11 -+ e- vr, [1v
- t-;;
] -k -+ 1, and (n)klnk -+ 1.
//1/
Theorem 7 gives conditions under which certain random experiments in-

volving counts of happenings in time (or length, space, area, volume, etc,) can
be realistically modeled by assuming a Poisson distribution. The parameter v
in the Poisson distribution is usually unknown. Techniques for estimating
parameters such as v will be presented in Chap. VI I. .
In practice great care has to be taken to avoid erroneously applying the
Poisson distribution to counts. For example, in studying the distribution of
insect larvae over some crop area, the Poisson model is apt to be invalid since
insects lay eggs in clusters entailing that larvae are likely to be found in clusters,
which is inconsistent with the assumption of independence of counts in small
adjacent subareas.
EXAMPLE 6 Suppose that the average number of telephone calls arriving

at the switchboard of a small corporation is 30 calls per hour. (i) What
is the probability that no calls will arrive in a 3-minute period? Oi) What
is the probability that more than five calls will arrive in a 5·minute interval?
Assume that the number of calls arriving during any time period has a
Poisson distribution. Assume that time is measured in minutes; then 30
calls per hour is equivalent to .5 calls per minute, so the mean rate of
occurrence is .5 per minute. P[no calls in 3-minute period] = e - vt =
e-(·S,(3) = e-I.s ~ .223.
'Xl e- vt( vt)k

P[more than five calls in 5-minute interval] = "
'=-
k-6 k'•
00 e (.S)(S)(2.5)k
=2:
k=6
k! ~ .042. IIII
EXAMPLE 7 A merchant knows that the number of a certain kind of item

that he can sen in a given period of time is Poisson distributed. How
many such items should the merchant stock so that the probability will be
.95 that he will have enough items to meet the customer demand for a time
period of length T? Let v denote the mean rate of occurrence per unit
time and K the unknown number of items that the merchant should stock.
Let X denote the number of demands for this kind of item during the time
period of length T. The solution requires finding K so that P[X K]
K
> .95 or finding K so th.at L [e-VT(vT)klk!] > .95.
k=O
In particular, if the
merchant sells an average of two such items per day, how many should
98 SPECIAL PARAMETRIC FAMIUES OF UNIVARIATE DISTRIBUTIONS III
he stock so that he will have probability at least .95 of having enough

items to meet demand for a 30-day month? Find K so that
K e-(2)(30)60k
k~O k! > .95,
or find K so that
00
L
k==K+ )
The desired K can be found using an appropriate Poisson table (e.g.,

Molina, 1942 [45]). It is K = 73. 1111
EXAMPLE 8 Suppose that flaws in plywood occur at random with an average

of one flaw per 50 square feet. What is the probability that a 4 foot
x 8 foot sheet will have no flaws? At most one flaw? To get a sol-
ution assume that the number of flaws per unit area is Poisson distributed.
P[no flaws] = e- lo32 = e-· 64 ~ .527.
P[at most one flaw] = e-· 64 + .64e-· 64 ~ .865. 1111
A Poisson density function, like the binomial density, possesses a certain
monotonicity that is precisely stated in the following theorem.
Theorem 8 Consider the Poisson density

e-A)..k
for k = 0, 1, 2, ....
k!
e -A~k-I
A e -;',k
1'.
---<-- for k < )..,
(k-l)! k!
e -A'k-I
.I. e A1k
A
for k > )..,
(k - I)! > -k-!-
and
I e -A1k-)
A
--~--~ if ).. is an integer and k = )...
(k - I)! k!
PROOF
e-1).k-1/(k -l)! k
-
e- A)..klk! ).. ,
which is less than 1 if k < l, greater than I if k > )., and equal to 1 if )..
is an integer and k = ).. 1111
2.5 Geometric and Negative Binomial Distributions

Two other families of discrete distributions that play important roles in statistics
are the geometric (or Pascal) and negative binomial distributions. The reason
that we consider the two together is twofold; first, the geometric distribution is a
special caSe of the negative binomial distribution, and, second, the sum of
independent and identically distributed geometric random variables is negative
binomially distributed, as we shall see in Chap. V. In Subsec. 3.3 of this chapter,
the exponential and gamma distributions are defined. We shall See that in
several respects the geometric and negative binomial distributions are discrete
analogs of the exponential and gamma distributions.
Definition 6 Geometric distribution A random variable X is defined

to have geometric (or Pascal) distribution if the density of X is given by
/x(x) =/x(x; p)
= {:(l -p)'
for x = 0, 1, ... }
= p(l - p)XI{o, I, ... }(X), (11)
otherwise
where the parameter p satisfies 0 < p s,: I. (Define q = 1 - p.) IIII
Definition 7 Negative binomial distribution A random variable X

with density
/x(X) = fx(x;r,p)
= {~r +; - l)p'q' = (~r)v( -qt for x = 0, 1, 2, ...
otherwise (12)
= (
r X-I) pq,. xI
+x ( )
{O.I .... }X'
where the parameters rand p satisfy r = I, 2, 3, ... and 0 < p < 1

(q = 1 - p), is defined to have a negative binomial distribution. The
density given by Eq. (12) is called a negative binomial density.
IIII
~ema~k If i? the ne~a~ive binomial distribution r = 1, then the negative

bInOmIal c,tensity speCialIzes to the geometric density. IIII
100 SPEOAL PARAMETRIC FAMILIES OF UNIVARIATE DISTRIBUTIONS III
_l
P -4
Theorem 9 If the random variable X has a geometric distribution,

then
q q p
C[X] =-, var [X] = 2 ' and mx(t) = 1 r (13)
p P -qe
PROOF Since a geometric distribution is a special case of a negative

binomial distribution, Theorem 9 is a corollary of Theorem 11. IIII
The geometric distribution is well named since the values that the geometric
density assumes are the terms of a geometric series. Also the mode of the
geometric density is necessarily o. A geometric density possesses one other
interesting property, which is given in the following theorem.
Theorem 10 If X has the geometric density with parameter p, then

P[X > i + j] X > i] = P[X > j] for i,j = 0, 1,2, ....
., . P[X 2 i + j]
PROOF P[X > I + JI X > I] = P[X 2 i]
00
I p(1 - PY
_ ;:.:...x=_.;....·+...:;.l_--- _
(l _p)i+l
~p(1 - PY (1-
x=i
= (1 - p)l
=P[X> j]. IIII
Theorem 10 says that the probability that a geometric random variable is

greater than or equal to i + j given that it is greater than or equal to i is equal to
the unconditional probability that it will be greater than or equal to j. We will
comment on this again in the following example.
EXAMPLE 9 Consider a sequence of independent, repeated Bernoulli trials

with p equal to the probability of success on an individual trial. Let the
random variable X represent the number of trials required before the first
succesS; then X has the geometric density given by Eq. (11). To See this,
note that the first success will occur on trial x + 1 if this (x + I)st trial
results in a succeSs and the first x trials resulted in failures; but, by in-
dependence, x successive failures followed by a SUCCeSS has probability
(1 - PYp· In the language of this example, Theorem 10 states that the
probability that at least i + j trials are required before the first succeSS,
given that there have been i successive failures, is equal to the uncon-
ditional probability that at least j trials are needed before the first suc-
cess. That is, the fact that one has already observed i successive failures
does not change the distribution of the number of trials required to obtain
the first success. //1 I
A random variable X that has a geometric distribution is often referred

to as a discrete waiting-time random variable. It represents how long (in terms
of the number of failures) one has to wait for a SUCceSs.
Before leaving the geometric distribution, we note that some authors
define the geometric distribution by assuming 1 (instead of 0) is the smallest
mass point. The density then has the form
!(x;p) = p(1 - pY- 1 I{l.2, ... }(x), (14)

and the mean is lip, the variance isqlp2, and the moment generating function is
petl( 1 - qe r ).
Theorem 11 Let X have a negative binomial distribution; then
rq rq
S[X] =
p
, var [X] = 2 ' and m xC t) =[ p
1 - qe
t] "', (15)
P
PROOF
x.:O
f (- r)pr( -qe'Y = [ 1 - pqer]
X
r
[see Eq. (33) in Appendix A],
m~(t) = pr( -r)(l qe t ) r-I( _qe t )
and
hence
rq
8[X] = m~(t)
r=O p
and
var[X] = mit t) 0 - (tf[X)2 = rqp'[qp -,-2(r + 1) + p-'-'] - (;r

rq2 rq rq
=-+-=-
p2 p p2' /11/
The negative binomial distribution, like the Poisson, has the nonnegative
integers for its mass points; hence, the negative binomial distribution is poten-
tially a model for a random experiment where a count of some sort is of interest,
[ndeed, the negative binomial distribution has been applied in population counts,
in health and accident statistics, in communications, and in other counts as
welL Unlike the Poisson distribution, where the mean and variance are the
same, the variance of the negative binomial distribution is greater than its mean.
We will see in Subsec. 4,3 of this chapter that the negative binomial distribution
can be obtained as a contagious distribution from the Poisson distribution.
EXAMPLE 10 Consider a sequence of independent, repeated Bernoulli

trials with p equal to the probability of success on an individual triaL Let
the random variable X represent the number of failures prior to the rth
success; then X has the negative binomial density given by Eq. (12), as the
following argument shows: The last trial must result in a success, having
probability p; among the first x + r - I trials there must be r - I succeSSeS
and x failures, and the probability of this is
(x+r-l)
r-l p
r-I x~
q -
(r+x-l)
x p
r-Iqx
,
which when multiplied by p gives the desired result. ///1
A random variable X having a negative binomial distribution is often

referred to as a discrete waiting-time random variable. It represents how long
(in terms of the number of failures) one waits for the rth success.
EXAMPLE II The negative binomial distribution is of importance in the

consideration of inverse binomial sampling. Suppose a proportion p
of individuals in a population possesses a certain characteristic. Jf
individuals in the population are sampled until exactly r individuals with
the certain characteristic are found, then the number of individuals in
eXcess of r that are observed or sampled has a negative binomial dis-
tribution. 1//1
2.6 Other'Discrete Distributions

In the previous five subsections we presented Seven parametric families of uni-
variate discrete density functions. Each is commonly known by the names
given. There are many other families of discrete density functions. In fact,
new families can be formed from the presented families by various proceSses.
One such process is called truncation. We will illustrate this process by looking
at the Poisson distribution truncated at O. SUppose, as is sometimes the case,
that the zero count cannot be observed yet the Poisson distribution Seems a
reasonable model. One might then distribute the mass ordinarily given to the
mass point 0 proportionately among the other mass points obtaining the family
of densities
for x = 1,2, ...

(16)
otherwise.
104 SPECIAL PARAMETRIC FAMILIES OF UNIVARIATE DISTRIBUTIONS HI
A random variable having density given by Eq. (16) is called a Poisson

random variable truncated at O.
Another process for obtaining a new family of densities from a given
family can also be illustrated with the Poisson distribution. Suppose that a
random variable X, representing a count of some sort, has a Poisson distribu w
tion. If the experimenter is stuck with a rather poor counter, one that cannot
count beyond 2, the random variable that the experimenter actually observes
has density given by
z o I 2
e -A I - e -A - Ae
,-..1.
fez)
The counter counts correctly values 0 and 1 of the random variable X; but if X
takes on any value 2 or more, the counter counts 2. Such a random variable
is often referred to as a censored random variable.
The above two illustrations indicate how other families of discrete densities
can be fonnulated from existing families. We close this section by giving two
further, not so wellwknown, families of discrete densities.
Definition 8 Beta-binomial distribution The distribution with discrete

density function
f(x) = I(x; n, 0:, fJ) = (xn) r(o: + f3) rex + o:)r(n + fJ -

r(o:)r(f3)' r(n + + f3)0:
x)
I{o. I ..... n}(x)
(17)
where n is a nonnegative integer, 0: > 0, and fJ > 0, is defined as the beta w
binomial distribution.
rem) is the well-known gamma function rem) = Io
xm - I e- X dx for
m > O. See Appendix A. The beta-binomial distribution has
net. . no:fJ(n +Cl+ f3)

Mean = f3 and varIance = (Cl + fJ)2(et. + f3 + 1) . (18)
Cl+
It has the same mass points as the binomial distribution. If et. = fJ = 1,

then the beta-binomial distribution reduces to a discrete uniform distribu-
tion over the integers 0, 1, "', n. IIII
3 CONTINUOUS DISTRIBUTIONS 105
Definition 9 Logarithmic distribution The distribution with discrete

density function
for x = 1, 2, ... x
q. I ( )
!(x;p) = - 1 {1,2, ... } x, (19)
-x oge P
o otherwise
where the parameters satisfy 0 < P < 1 and q = 1 - p is defined as the
logarithmic distribution. IIII
The name is justified if one recalls the power-series expansion of loge (1 ~. q).
The logarithmic distribution has .----
. q . q(q + loge p)
M ean =---- and varIance = ( )2 . (20)
-p loge P - plogeP
[t can be derived as a limiting distribution of negative binomial distributions

that have been generalized to include r, any positive number (rather than just
an integer), truncated at O. The limiting distribution is obtained by letting r
approach O.
3 CONTINUOUS DISTRIBUTIONS
In this section several parametric families of univariate probability density
functions are presented. Sketches of some are incl uded; the mean and variance
(when they exist) of each are given.
3.1 Uniform or Rectangular Distribution

A very simple distribution for a continuous random variable is the uniform dis-"
tribution. It is particularly useful in theoretical statistics because it is convenient
to deal with mathematically.
Definition 10 Uniform distribution If the probability density function

of a random variable X is given by
1
/x(x) = /x(x; a, b) = l[a bJ(X), (21)
b-a '
I
b-a • •
---+-----a b - - - - - -........ X
FIGURE 8
Uniform probability density.
where the parameters a and b satisfy - 00 < a < b < 00, then the random
variable X is defined to be uniformly distributed over the interval [a, b],
and the distribution given by Eq. (21) is called a uniform distribution.
IIII
Theorem 12 if X is uniformly distributed over [a, b], then

(b - a)2 e"t _ ~t
var [X] = - - - and mx(t) = (b - a)t' (22)
12 '
PROOF
b 1 b2 - a 2 a +b
8[X] = {x b _ a dx = 2( b - a) = 2 .
var [X] = .8'[X2] - (.8'[X])2 = {X2 b~ adx _ (a ; b)'

b 3 -a 3 (a+b)2 (b-a)2
-
3(b-a) 4 12
IIII
The uniform distribution gets its name from the fact that its density is
uniform, or constant, over the interval [a, b]. It is also called the rectangular
distribution-the shape of the density is rectangular.
The cumulative distribution function of a uniform random variable is
given by
(23)
It provides a useful model for a few random phenomena. For instance, if it

is known that the values of some random variable X can only be in a finite
interval, say [a, b], and if one assumes that any two subintervals of [a, b] of
equal length have the same probability of containing X, then X has a uniform
distribution over the interval [a, b]. When one speaks of a random number
from the interval [0, 1], one is thinking of the value of a uniformly distributed
random variable over the interval [0, 1].
EXAMPLE 12 If a wheel is spun and then allowed to come to rest, the point
on the circumference of the wheel that is located opposite a certain fixed
marker could be considered the value of a random variable X that is
uniformly distributed over the circumference of the wheel. One could
then compute the probability that X will fall in any given arc. 11II
Although we defined the uniform distribution as being uniformly dis-

tributed over the closed interval [a, b], one could just as well define it over the
open interval (a, b) [in which case/x(x) = (b - a) II(a,b)(x)] or over either of the
ha/f-open-ha/f-c1osed intervals (a, b] or [a, b). Note that all four of the possible
densities have the same cumulative distribution function. This lack of unique-
neSS of probability density functions was first mentioned in Subsec. 3.2 of
Chap. II.
3.2 Normal Distribution

A great many of the techniques used in applied statistics are based upon the
normal distribution; it will frequently appear in the remainder of this book.
Definition 11 Normal distribution A random variable X is defined to

be normally distributed if its density is given by
/x(x) = /x(x; p" q) = JI e-(X-Jl,)2/2cr\ (24)

2nq
where the parameters p, and q satisfy - (X) < p, < 00 and q > 0. Any
distribution defined by a density function given In Eq. (24) is called a
normal distribution. IIII
We have used the symbols p, and q2 to represent the parameters because

these paramete~s t,urn .out, as we shan see, to be the mean and variance, respec-
tively, of the distrIbution.
FIGURE 9
Norma] densities.
" '
One can readily check that the mode of a normal density occurs at x = Jl
and inflection points occur at Jl - a and Jl + a. (See Fig. 9.) Since the normal
distribution occurs so frequently in later chapters, special notation is introduced
for it. If random variable X is norma]]y distributed with mean J1 and variance
a 2, we wi]] write X,.... N(J1, ( 2). We will also use the notation </J/1. a2(x) for the
density of X,.... N(Jl, ( 2) and <1>/1. a2(x) for the cumulative distribution function.
If the normal random variable has mean 0 and variance 1, it is called a
standard or normalized normal random variable. For a standard normal ran-
dom variable the subscripts of the density and distribution function notations
are dropped; that is,
</J(x),= J~-
2n
e-
tx2 and <1>(x) = IX
-00
</J(u) duo (25)
Since </J1l. a2(x) is given to be a density function, it is implied that
I_
00
00 </J /1 , a 2 (x) dx = 1,
but we should satisfy ourselves that this is true. The verification is somewhat
troublesome because the indefinite integral of this particular density function
does not have a simple functional expression. Suppose that we represent the
area under the curve by A; then
A= 1 foo e-(x-Il)2j2a2d x,
J2na -00
and on making the substitution y = (x - Jl)/a, we find that
1
A=J~ e
I oo
-ty2 d
y.
2n - 00
FIGURE 10
Norma] cumulative distribution
function.
We wish to show that A = 1, and this is most easily done by showing that A2 is
1 and then reasoning that A = 1 since <P 1l ,a2 (x) is positive. We may put
2
A=J.~
1 foo e
-ty2 d
Y 1-2n foo e -tz 2 d
z
2n -00
J -00
= -1 foo foo e- t (y2+ z 2) dy dz

2n -00 -00
by writing the product of two integrals as a double integral. In this integral

we change the variables to polar coordinates by the substitutions
y= r sin e
z = r cos e,
and the integra] becomes
=1.
Theorem 13 If X is a normal random variable,
&[ X] = J.l, and (26)
PROOF
mx(t) = C[e tX ] = llltS'[l(X-Il)]
oo 1
= e tll
f
-00
-= l(x- ll )e-(1/2a2)(x-Il)2 dx
J2n
= etll 1 foo e-O/2(2)[(x-IlP-2a2t(x-ll)]dx.
J2n -00
I f we complete the square inside the bracket, it becomes

(x - /1)2 - 2(12t(X - J.l) = (x - J.l)2 - 2CT 2t(x - /1) + (14 t 2 _ (14t 2
= (x - J.l - (12t)2 - (14 t 2,
110 SPECIAL PARAMETRIC FAMILIES OF UNIVARIATE DISTRIBUTIONS In
and we have
mx(t) = etJletl2t2/2 1 foo e-(x-Jl-a2t)2/2t12 dx.

J21tu -00
The integra] together with the factor I IJ21tU is necessarily I since it is the
area under a normal distribution with mean Jl + u 2 t and variance a 2 •
Hence,
2 2
mx(t) = eJlt+a t /2,
On differentiating mx(t) twice and substituting t = 0, we find
8[X] = m~(O) = jJ
and
var [X] = 8[X2] - (8[X])2 = mi.-(O) - Jl2 a 2,
thus justifying our use of the symbols Jl and u2 for the parameters. IIII
Since the indefinite integral of 4J Jl • a2(x) does not have a simple functional
form, one can only exhibit the cumulative distribution function as
f- oo 4Jp.,a2(u) duo
x
Jl,a2(x) = (27)
The folIowing theorem shows that we can find the probabiHty that a normally
distributed random variable, with mean Jl and variance a 2 , falls in any interval
in terms of the standard norma] cumulative distribution function, and this
standard normal cumulative distribution function is tabled in Table 2 of
Appendix O.
Theorem 14 If X '" N(p, a2 ), then
PIa < X <hJ = fl> e: 1') - e 1').fl> ~ (28)
PROOF
b 1
P[a < X < bJ = f e- H (x- Jl )/a]2 dx
a J21tU
1
f (b-Jl)/a -tz 2 d
= (a-Jl)/a J21t e z
Remark (x) = 1 - <1>( - x). IIII
The normal distribution appears to be a reasonable model of the behavior

of certain random phenomena. Jt also is the limiting form of many other prob~
ability distributions. Some such limits are given in Subsec. 4.1 of this chapter.
The normal distribution is also the limiting distribution in the famous centra/-
limit theorem, which is discussed in Sec. 4 of Chap. V and again in Sec. 3 of
Chap. VI.
Most students are already somewhat familiar with the normal distribution
because of their experience with "grading on the curve." This notion is
covered in the fol1owing example.
EXAMPLE 13 Suppose that an instructor assumes that a student's final

score is the value of a norma]]y distributed random variable. If the
instructor decides to award a grade of A to those students whose score
exceeds J1 + a, a B to those students whose score fa]]s between Jl and
}J. + a, a C if a score falls between J1 - (j and Jl, a D if a score falls between
J1 - 2a and J1 - a, and an F if the score falls below J1 - 2a, then the pro-
portions of each grade given can be calculated. For example, since
PI X > J1 + 0"] = 1- PI X < Jl + 0"] = 1 - <\l (Jl + ; - Jl)

= 1 - (l) ~ .1587,
one would expect 15.87 percent of the students to receive A's. IIII
EXAMPLE 14 Suppose that the diameters of shafts manufactured by a cer-

tain machine are normal random variables with mean 10 centimeters and
standard deviation .1 centimeter. If for a given application the shaft must
meet the requirement that its diameter fa]] between 9.9 and 10.2 centi-
meters, what proportion of the shafts made by this machine wi]] meet the
requirement?
P[9.9 < X< 10.2] = <\l Co.~ 1- 10) - <\l e·9.~ 10)
= <1>(2) - <1>(-1) ~ .9772 - .1587 = .8185. IIII
3.3 Exponential and Gamma Distributions
Two other families of distributions that play important roles in statistics are the
(negative) exponential and gamma distributions, which are defined in this sub-
section. The reason that the two are considered together is twofold; first, the
exponential is a special case of the gamma, and, second, the sum of independent
identically distributed exponential random variables is gamma-distributed, as
we shall see in Chap. V.
Definition 12 Exponential distribution If a random variable X has a

density given by
/x(x; A) = Ae-AxI[o. oolx), (29)

where A > 0, then X is defined to have an (negative) exponential distribu-
~. M
Definition 13 Gamma distribution If a random variable X has density
given by
°
where r > and A > 0, then X is defined to have a gamma distribution.
r(·) is the gamma function and it is discussed in Appendix A. IIII
Remark If in the gamma density r = 1, the gamma density specializes

to the exponential density. /III
Theorem 15 If X has an exponential distribution, then
1 1 A
4[X] =-, var [Xl = A2 ' and mx(t) = - - for
A A- t
(31)
PROOF The exponentia.l distribution was the distribution used as an

example for some definitions given in Chap. II, and derivations of the above
appear there. Also, Theorem 15 is a corol1ary to the following theorem.
fill
Theorem 16 If X has a gamma distribution with parameters r and A,

then
r
var [X] = A2 ' and mx(t) ( ). )r for t < A.
A-t
(32)
1.0
o~~~~~~~~~~~~7~~8~X
1 2
FIGURE 11
Gamma densities (A I),
PROOF
= (A-
A )r Joo(A - ty x r-I e -(A.-t)x d x-
_ (_A . )r
- t 0 r(r) A- t
m~(t)= rArO. - t)-r-I
and
mHt) = r(r + l)Ar(A - t)-r-2;

hence
4'[X] = m~(O) = i
and
var [X] = 4'[X2] - (~[X])2
= mi(O) _ (!:.)2 = r(r ~ 1) _ (~) 2_ r IIII

1 1 1 A
The exponential distribution has been used as a model for lifetimes of

various things. When we introduced the Poisson distribution, we spoke of cer-
tain happenings, for example, particle emissions, occurring in time. The length
of the time interval between successive happenings can be shown to have an
exponential distribution provided that the number of happenings in a fixed
time interval has a Poisson distribution. We comment on this again in Subsec.

4.2 below. Also, if we assume again that the number of happenings in a fixed
time interval is Poisson distributed, the length of time between time 0 and the
instant when the rth happening occurs can be shown to have a gamma distribu-
tion. So a gamma random variable can be thought of as a continuous waiting-
time random variable. It is the time one has to wait for the rth happening.
Recall that the 'geometric and negative binomial random variables were dis-
crete waiting-time random variables. In a sense, they are discrete analogs of the
negative exponential and gamma distributions, respectively.
Theorem 17 If the random variable X has a gamma distribution with

parameters r and A, where r is a positive integer, then
r - 1 e - ).x ().x)j
Fx(x) = 1 - I
j=O
.,
J.
(33)
PROOF The proof can be obtained by succeSSIve integrations by

parts. IIII
For A = 1, Fx(x) given in Eq. (33) is cal1ed the incomplete gamma/unction

and has been extensively tabulated.
Theorem 18 If the random variable X has an exponential distribution

wi th parameter A, then
P[X> a + bl X> a] =P[X> b], for a > 0 and b > O.
P[X> a + b] e-).(a+b)
PROOF P[X> a + blX> a] = P[X> a] = e-).a
= e-)'b = P[X > b]. IIII
Let X represent the lifetime of a given component; then, in words,

Theorem 18 states that the conditional probability that the component wi1I last
a + b time units given that it has lasted a time units is the same as its initial
probability of lasting b time units. Another way of saying this is to say that an
"old" functioning component has the same 1ifetime distribution as a new" H
functioning component or that the component is not subject to fatigue or to

wear.
3.4 Beta Distribution

A family of probability densities of continuous random variables taking on values
in the interval (0, 1) is the family of beta distributions,
Definition 14 Beta distribution If a random variable X has a density

given by
/x(x) = /x(x; a, b) B(~, b) x"-t (I - x)'-t 1(0. t)(x), (34)
where a > °and b > 0, then X is defined to have a beta distribution. IIII
The function B(a, b) = xa-1(1 - n X)b-l dx, caned the beta junction, is
mentioned briefly in Appendix A.
Remark The beta distribution reduces to the uniform distribution over

(0, 1) if a = b = 1. IIII
Remark The cumulative distribution function of a beta-distributed

random variable is
x 1
F x(x; a, b) = 1(0, l)(X) I
0 B(a, b) u
a
-
1
(l - U)b-l du + 1[1, ooix); (35)
it is often called the incomplete beta and has been extensively tabulated.
IIII
2.0
I.S
I a = 1
~~~~ __+-______~~~~lb=l
.5
FIGURE 12 ~~~----~ __ ~ ____ ~ __ ~L-_x
Beta densities. .4 .6 .8
The moment generating function for the beta distribution does not have a
simple form; however the moments are readily found by using their definition.
Theorem 19 If X is a beta-distributed random variable, then

ab
~[X] = a and VM[X]= .
a+b (a +b+ l)(a + b)2
PROOF
~[Xk] = 1
B(a, b)
II 0
y!+a-l(1 - X)b-l dx
B(k + a, b) r(k + a)r(b) . r(a + b)

B(a, b) r(k + a + b) r(a)r(b)
_ r(k + a)r(a + b).
- r(a)r(k + a + b)'
hence,
~[X] = r(a + l)r(a + b) = a

r(a)r(a+b+l) a+b'
and
2 2 r(a + 2)r(a + b) ( a ) 2
var [X] = ~[X ] - (~[X]) = r(a)r(a + b + 2) - a + b
(a + 1)a
a )2 ab (
=(a+b+l)(a+b)- a+b =(a+b+l)(a+b)2· IIII
The family of beta densities is a two-parameter family of densities that
is positive on the interval (0, 1) and can assume quite a variety of different
shapes, and, consequently, the beta distribution can be used to model an experi-
ment for which one of the shapes is appropriate.
3.5 Other Continuous Distributions

In this subsection other parametric families ofprobabiHty density functions that
will appear later in this book are briefly introduced; many other families exist.
The introductions of the three families of distributions, that go by the names of
Student's t distribution, chi-square distribution, and F distribution, are de-
ferred until Chap. VI. These three families, as we sha11 see, are very important
when sampHng from normal distributions.
Cauchy distribution A distribution which we shaH find useful for illustrative

purposes is the Cauchy, which has the density
1
(36)
fx(x; ct, {J) = n{J{1 + [(x _ ct)/{J]2} '
where - 00 < ct < 00 and {J > O.

Although the Cauchy density is symmetrical about the parameter ct, its
mean and higher moments do not exist. The cumulative distribution function
IS
F (x) - -
1 IX du
x - n -00 n{J{l + [(u - a)/{J]2}
(37)
1 1 X-ct
= - + - arc tan - -
2 n {J
Lognormal distribution Let X be a positive random variable, and let a new

random variable Y be defined as Y = loge X. If Y has a normal distribution,
then X is said to have a lognormal distribu tion. The density of a lognormal
distribu tion is given by
f .
(x, 11,
2 _ J1 exp [1
(1 ) - - 2 2 (Ioge x - 11) 2] [(0. OO)(x), (38)
x 2n(1 (1
where -00 < 11 < 00 and (1 > O.

C[X] = eJl+t a2
and (39)
for a lognormal random variable X. Also, if X has a lognormal distribution,
then C[loge (X)] = 11, and var [1oge (X)] = (12.
Double exponential or Laplace distribution A random variable X is said to

have a double exponential, or Laplace, distribution if the density function of X
is given by
fx(X) = fx(x; ct, {J) = _I exp (_ Ix - ct l) (40)

2{J {J'
where - 00 < ct < 00 and {J > O. If X has a Laplace distribution, then
C[X] = ct and (41)
Weibull distribution The density

f(x; a, b) = ab:x!'-le-aXb[(O, ooix) (42)
where a > 0 and b > 0, is ca]]ed the Weibull density, a distribution that has been
successfully used in reliability theory. For b = 1, the Weibu11 density reduces
to the exponential density. 1t has mean (1/a)llhr(l + b- 1 ) and variance
(l/a)2/h[r(1 + 2b- 1 ) - r2(l + b- 1 )).
Logistic distribution The logistic distribution is given in cumulative distribu-

tion form by
F( x . r:x fJ) ___1-,----.-:--: . (43)

" - 1 + e-(x-a)lfJ'
where - 00 < r:x < 00 and fJ > O. The mean of the logistic distribution is given
by r:x. The variance is given by fJ 2 n 2 /3. Note that F( r:x - d; r:x, fJ) =
I - F(r:x + d; r:x, fJ), and so the density of the logistic is symmetrical about cx.
This distribution has been used to model tolerance levels in bioassay problems.
Pareto distribution The Pareto distribution is given in density-function form

by
l44)
where () > 0 and Xo > O. The mean and variance respectively of the Pareto
distribution are given by
()xo ()x5 _ ( ()x o ) 2

for () > 1 and for () > 2.
()-I ()-2 ()-I
This distribution has found application in modeling problems involving distribu-

tions of incom.es when incomes exceed a certain limit Xo .
Gumbel distribution The cumulative distribution function

F(x; cx, fJ) = exp( - e -(x-a)lfJ), (45)
where - 00 < r:x < 00 and fJ> 0 is ca]]ed the Gumbel distribution. It appears
as a limiting distribu tion in the theory of extreme-value statistics.
Pearsonian system of distributions Consider a density function /x(x)

which satisfies the differential equation
I d/x(x) x+a (46)

--- = 2
/x(X) dx bo + b1x + b 2 x
for constants a, bo , b l , and b 2 • Such a density is said to belong to the Pearsonian
4 COMMENTS 119
system of density functions. Many of the probability density functions that we

have considered are special cases of the Pearsonian system. For example, if
~r r -1
I\, X e- AX
fxCx) = r(r) I[o,oo)Cx),
then
1 dfxC x) r - J x - (r - 1)/A
-- = -A + = -----
fx(x) dx x -x/).
for x > 0; so the gamma distribution is a member of the Pearsonian system with
a = -(r - 1)/)" b I = -1/)', and bo = b 2 = O.
4 COMMENTS
We conclude this chapter by making severa) comments that tie together some
of the density functions defined in Secs. 2 and 3 of this chapter.
4.1 . Approximations
Although many approximations of one distribution by another exist, we wi]]
give only three here. Others wiJ] be given along with the central-limit theorem
in Chaps. V and VI.
Binomial by Poisson We defined the binomial discrete density function,

with parameters nand p, as
for x 0, 1, ... , n.
If the parameter n approaches infinity and p approaches 0 in such a way that

np remains constant, say equal to )., then
(47)
for fixed integer x. The above fo11ows immediately from the fo]]owing con-
sideration:
n)pX(1 _ p)n-x
( x.
= (n}x (~)X(l _~)n-x
xI n n
since
A) -x
(1-;; -+1, and as n -+ 00.
Thus, for large n and small p the binomial probability(:)r(l _ p)O-X

can be approximated by the Poisson probability e-np(npr/x!. The utility of
this approximation is evident if one notes that the binomial probability in-
volves two parameters and the Poisson only one.
Binomial and Poisson by normal
Theorem 20 Let random variable X have a Poisson distribution with

parameter A; then for fixed a (b) - <1>(a) as A -+ 00. (48)
PROOF Omitted. [Eq. (48) can be proved using Stirling's formula,

which is given in Appendix A. It also fol1ows from the central-limit
theorem.] I,1/11
/1
Theorem 21 De Moivre-Laplace limit theorem Let a random variable

X have a binomial distribution with parameters nand p; then for fixed
a(b) - <1>(a) as n -+ 00. (49)
PROOF Omitted. (This is a special case of the central-limit

theorem, given in Chaps. V and VI.) II1I
Remark We approximated the binomial distribution with a Poisson

distribution in Eq. (47) for large n and small p. Theorem 21 gIVes a
normal approximation of the binomial distribution for large n. IIII
The usefulness of Theorems 20 and 21 rests in the approximations that

they give. For instance, Eq. (49) states thatP[np + aJnpq < X <np + bJnpq]
4 COMMENTS 121
is approximately equal to (b) - (a) for large n. Or if c = np + aJnpq and

d = np + bJnpq, then Eq. (49) gives that P[c < X < d] is approximately equal
to
d - np ) (C - np )
 ( Jnpq - Jnpq
for large n, and, so, an approximate value for the probability that a binomial
random variable faIJs in an interval can be obtained from the standard norma]
distribution. Note that the binomial distribution is discrete and theapproximat-
ing norma] distribution is continuous.
EXAMPLE 15 Suppose that two fair dice are tossed 600 times. Let X
denote the number of times a total of 7 occurs. Then X has a binomial
distribution with parameters n = 600 and p = ~. 8[X] = 100. Find
P[90 < X < 110].
P[90 x if (600)
-< 11 0] = j=90 j
(~)j(~)600-
6 6 '
j
a sum that is tedious to evaluate. Using the approximation given by

Eq. (49), we have
110 - 100) (90 - 100)

P[90 < X < 110] ~ ( J.i~O - Jsgo
(Ji) - <1>( -J~) ~ <1>(1.095) - <1>( -1.095) ~ .726.
IIII
4.2 Poisson and Exponential Relationship

When the Poisson distribution was introduced in Subsec. 2.4, an experiment
consisting of the counting of the number of happenings of a certain phenomenon
in time was given special consideration. We argued that under certain conditions
the count of the nmriber of happenings in a fixed time interval was Poisson dis-
tributed with parameter, the mean, proportional to the length of the interval.
Suppose now that one of these happenings has just occurred; what then is the
distribution of the length of time, say X, that one will have to wait until the
next happening? P[X> t] = P[no happenings in time interval of length t] =
e- vt , where v is the mean occurrence rate; so
Fx(t) = P[X ~ t] = 1 - P[X > t] = 1 _ e- vt for t > 0;
that is, X has an exponential distribution. On the other hand, it can be proved,
under an independence assumption, that if the happenings are occurring in
time in such a way that the distribution of the lengths of time between successive
happening~ is exponential, then the distribution of the number of happenings
in a fixed time interval is Poisson distributed. Thus the exponentia1 and Poisson
distribu tions are re1ated.
4.3 Contagious Distributions and Truncated Distributions
A brief introduction to the concept of contagious distributions is given here.

If/o('), 11('), ... ,/,.(.), ... is a sequence of density functions which are either al1
discrete density functions or all probabiHty density functions which mayor may
not depend on parameters, and Po, PI' ... , Pn , ••• is a sequence of parameters
00 00
satisfying Pi > 0 and L Pi = 1, then L Pih(x) is a density

i=O i=O
function, which is
sometimes called a contagious distribution Or a mixture. For example, if
lo(x) = 4>"0. u02 (x) (a norma1 with mean Ilo and variance O'~) and 11 (x) = 4>lll. U1 2(X),
then
where PI = P and Po = 1 - p, is a mixture of two normal densities. Equation

(50) is a1so sometimes referred to as a contaminated normal. A random variab1e
Xhas distribution given by Eq. (50) ifit is normal1y distributed with mean III and
variance O'i with probability P and normally distributed with mean Ilo and vari-
ance O'~ with probabiJity 1 - p. Contagious distributions or mixtures can be
useful mode1s for certain experiments. For instance, the mixture of two normal
distributions given in Eq. (50) has five parameters, name1y, p, Ilo, Ill' 0'0, and
0' l' If we vary these five parameters, the density can be forced to assume a
variety of different shapes, some of which are bimoda1; that is, the density has
two distinct 10cal maximums.
Physical considerations of the random experiment at hand can sometimes
persuade one to consider mode1ing the experiment with a mixture. The
experimenter may know that the phenomena that he is observing are a mixture;
for example, the radioactive partic1e emissions under observation might be a
mixture of the partic1e emissions of two, or several, different types of radioactive
materia1s.
4 COMMENTS 123
The concept of mixing can be extended. Let {f(x; 8)} be a family of

density functions parameterized or indexed by 8. Let the totaHty of values
that the parameter 8 can assume be denoted by e. If e is an interval (possibly
infinite) and g(8) is a probability density function which is for an arguments °
not in e, then
J~f(x; 8)g(8) d8 (51)
is again a density function, ca]]ed a contagious distribution or a mixture. For

example, supposef(x; 8) = e-08 x /x! for x = 0,1,2, ... andf(x; 8) = other- °
wise and
g(O) A' l l ' -1 - ;01 (ll)

r(r) u e· (O,(X) U ,
a gamma density. Then
=
A'
I (X)
O,+x-1 e -O.+ 1)0 d8
x!r(r) °
= A' . r(r + x) I(X) [(A + 1)0j'+x-1 e -().+ 1)8 d[(A + 1)8]
x!r(r) (A + 1Y+x ° r(r + x)
A )' r(r + x) 1
(
- A + 1 (x!)r(r) (A + It
= (r +xx- 1) ().+1
A )'( 1 )
A+1
x for x = 0, 1, ... ,
which is the density function of a negative binomial distribution with param-

eters rand P =A/(A + 1). We say that the derived negative binomial distri-
bution is the gamma mixture of Poissons.
is sometimes called a compound Poisson, where g(O)/(o,(X)iO) is a probability

density function.
We have sketchily i11ustrated above how new parametric families of den-
sities can be obtained from existing families by the technique of mixing. In
Subsec. 2.6 we indicated how truncation could be employed to generate new
families of discrete densities. Truncation can also be utilized to form other
families of continuous distributions. For instance, the family of beta distri-
butions provides densities that are useful in modeling an experiment for which
124 SPECIAL pARAMETRIC FAMILIES OF UNIVARIATE DISTRIBUTIONS m
it is known that the values that the random variable can assume are between 0
and 1. A truncated norma] or gamma distribution would also provide a useful
model for such an experiment. A normal distribution that is truncated at 0
on the left and at 1 on the right is defined in density form as
<Pp.,a2(x)l(o,l)(x)
ji()
x = ji( x; /1, (1 ) = (52)
ct> p.,a2( 1) - ct>ll,a2(O)
This truncated normal distribution, like the beta distribution, assumes values
between 0 and 1.
Truncation can be defined in general. If X is a random variable with
density Ix(-) and cumulative distribution Fx('), then the density of X truncated
on the left at a and on the right at b is given by
Ix(x)l(a,b)(X)
(53)
Fx(b) - Fx(a) .
PROBLEMS
1 (a) Let X be a random variable having a binomial distribution with parameters
n 25 and p = .2. Evaluate P[X < ftx 2ax].
(b) If X is a random variable with Poisson distribution satisfying P[X 0] =
P[X = 1], what is G[X]?
(c) If X is uniformly distributed over (1, 2), find z such that P[X > z·+ ftx] L
(d) If X is normally distributed with mean 2 and variance 1, find P[I X 21 < 1].
(e) Suppose X is binomial1y distributed with parameters nand p; further sup-
pose that G[X] = 5 and var [X] = 4. Find nand p.
(/) If G[X] = 10 and ax = 3, can X have a negative binomial distribution?
(g) If Xhas a negative exponential distribution with mean 2, find P[X < 11 X < 2].
(h) Name three distributions fOr which P[X < ftx] = ~..
(0 Let X be a random variable having binomial distribution with parameters
n = 100 andp =.1. Evaluate P[X < ftx - 3ax].
(j) If X has a Poisson distribution and P[X = 0] = i, what is G[X]?
(k) Suppose X has a binomial distribution with parameters n and p. For what
p is var [X] maximized if we assumed n is fixed?
(/) Suppose X has a negative exponential distribution with parameter A. If
P[X 1] = P[X > 1], what is var [X]?
(m) Suppose X is a continuous random variable with uniform distribution
having mean 1 and variance t. What is p[X < OJ?
(n) If X has a beta distribution, can G[lt X] be unity?
(0) Can X ever have the same distribution as - X? If so, when?
l
(p) If X is a random variable having moment generating function exp (e 1), -
what is 8[X]?
2 (a) Find the mode of the beta distribution.
(b) Find the mode of the gamma distribution.
PROBLEMS 125
3 'Name a parametric family of distributions which satisfies:

(a) The mean must be greater than or equal to the variance.
(b) The mean must be equal. to the variance.
(c) The mean must be less than or equal to the variance.
(d) The mean can be less than, equal to, or greater than the variance (for dif-
ferent parameter values).
4 (a) If X is norma]]y distributed with mean 2 and variance 2, express
P[ I X - i I <2] in terms of the standard normal cumulative distribution
function.
(b) If X is norma]]y distributed with mean ft > 0 and variance (12 = ft\ express
P[X < - ft I X < ft] in terms of the standard normal cumulative distribution
function.
(c) Let X be normally distributed with mean ft and variance (12. Suppose (12 is
some function of ft, say (12 = h(ft). Pick h(') sO that P[X < 0] does not
depend on J1 for ft > O.
5 Use the ahernate definition of the median as given in the remark fo]]owing Defi-
nition 18 of Chap. II. Find the median in each of the fo]]owing cases:
(a) /x(x) = ;"e-;'X[(O, a»(x),
(b) X is uniformly distributed on the interval (0 1, (J2).
(c) X has a binomial distribution with n = 4, p = .5.
(d) X has a binomial distribution with n = 5, p = .5.
(e) X has a binomial distribution with n = 2, p = .9.
*6 A contractor has found through experience that the low bid for a job (excluding
his own bid) is a random variable that is Uniformly distributed over the interval
(iC, 2C), where C is the contractor's cost estimate (no profit or loss) of the job.
If profit is defined as 0 if the contractor does not get the job (his bid is greater than
the low bid) and as the difference between his bid and his cost estimate C if he gets
the job, what should he bid (in terms of C) in order to maximize his expected
profit?
7 A merchant has found that the number of items of brand XYZ that he can sell
in a day is a Poisson random variable with mean 4.
(a) How many items of brand XYZ should the merchant stock to be 95 percent
certain that he wiJ] have enough to last for 25 days? (Give a numerical
answer.)
(b) What is the expected number of days out of 25 that the merchant will sel1
nO items of brand XYZ?
8 (a) If X is binomia]]y distributed with parameters nand p, what is the distribution
of Y=n -- X?
(b) Two dice are thrown n times. Let X denote the number of throws in which the
number on the first die exceeds the number On the second die. What is the
distribution of X?
*(c) A drunk performs a" random walk;; over pOsitions 0, ± 1, ±2, ... as fo]]ows:
He starts at O. He takes successive one-unit steps, going to the right with
probability p and to the left with probabHity 1 - p. His steps are inde-
pendent. Let X denote his position after n steps. Find the distribution of
(X + n)/2, and then find 8[X].
*(d) Let Xl (X 2) have a binomial distribution with parameters nand Pl (n and P2)'
If Pl <P2, show that P[Xl < k] > P[X2 < k] for k = 0, I, ... , n. (This
result says that the smaller the p, the more the binomial distribution is shifted
to the left.)
9 In a town with 5000 adu1ts, a sample of 100 is asked their opinion of a proposed
municipal project; 60 are found to favor it, and 40 oppose it. If, in fact, the
adults of the town were equal1y divided on the proposal, what would be the prob-
ability of obtaining a majority of 60 or more favoring it in a sample of 100?
10 A distributor of bean seeds determines from extensive tests that 5 percent of a large
batch of seeds will not germinate. He sells the seeds in packages of 200 and
gUarantees 90 percent germination. What is the probability that a given package
wi1l violate the guar3;ntee?
*11 (a) A manufacturing process is intended to produce electrical fuses with no
mOre than 1 percent defective. It is checked every hour by trying 10 fuses
selected at random from the hour's production. If 1 or more of the 10
fail, the process is halted and carefully examined. If, in fact, its prob-
ability of producing a defective fuse is .01, what is the probability that the
process will needlessly be examined in a given instance?
(b) Referring to part (a), how many fuses (instead of 10) should be tested if the
manufacturer desires that the probability be about .95 that the process wi11
be examined when it is producing 10 percent defectives?
12 An insurance company finds that .005 percent of the population die from a certain
kind of accident each year. What is the probability that the company must pay
off On more than 3 of 10,000 insured risks against such accidents in a given
year?
13 (a) If X has a Poisson distribution with P[X = 1] = P[X = 2]. what is
P[X = 1 or 2]?
(b) If X has a Poisson distribution with mean 1, show that 8[1 X-II] = 2ax/e.
*14 Recall Theorems 4 and 8. Formulate, and then prove or disprove a similar
theorem for the negative binomial distribution.
*15 Let X be normal1y distributed with mean ft and variance a 2 • Truncate the density
of X on the left at a and On the right at b, and then calculate the mean of the trun-
cated distribution. (Note that the mean of the truncated distribution should fall
between a and b. Furthermore, if a = ft - c and b = ft + c, then the mean of the
truncated distribution should equal ft.)
*16 Show that the hypergeometric distribution can be approximated by the binomial
distribution for large M and K; Le., show that
PROBLEMS 127
17 Let X be the life in hours of a radio tube. Assume that X is normally distributed
with mean 200 and variance a 2 • If a purchaser of such radio tubes requires that
at least 90 percent of the tubes have lives exceeding 150 hours, what is the largest
value a can be and still have the purchaser satisfied?
18 Assume that the number of fatal car accidents in a certain state obeys a Poisson
distribution with an average of one per day.
(a) What is the probability of more than ten such accidents in a week?
(b) What is the probability that more than 2 days will lapse between two such
accidents?
19 The distribution given by
I(x; (3) = 131

2
xe-i:(X!P)
2
[(0. a:»(x) for f3 > 0
is cal1ed the Rayleigh distribution.

(a) Show that the mean and variance exist, and find them.
(b) Does the Rayleigh distribution belong to the Pearsonian system?
4
I(x; (3) = fJ3 ..J7T' x 2e- X2 !p 2[(0. OO)(x) for f3 > 0
is ca11ed the Maxwell distribution.

(b) Does this distribution belong to the Pearsonian system?
I 2
I(x; n) = BO, [n _ 2]/2) (1- x Yn-4)!2[[_1.1l(X)
is called the r distribution.

(b) Does this distribution belong to the Pearsonian system?
22 A die is cast until a 6 appears. What is the probabi1ity that it must be cast more
than five times?
23 Red-blood-ce11 deficiency may be determined by examining a specimen of the
blood under a microscope. Suppose that a certain small fixed volume contains,
on an average, 20 red ceJIs for normal persons. What is the probability that a
specimen from a normal person will contain less than 15 red ce]]s?
24 A telephone switchboard handles 600 calls, On an average, during a rush hour.
The board can make a maximum of 20 connections per minute. Use the Poisson
distribution to evaluate the probability that the board will be overtaxed during any
given minute.
25 Suppose that a particle is equa]]y likely to release one, two, or three other particles,
and suppose that these second·generation particles are in turn each equa]]y likely
to release one, two, or three third-generation particles. What is the density of
the number of third-generation particles?
26 Find the mean of the Gumbel distribution.

27 Derive the mean and variance of the Weibul1 distribution.
* 28 Show that
P[X > k] = L ~
n
j=k
(
J
)
pjqn- j = 1
B(k, n - k + 1)
f
0
p
1I"-I( 1 - 1I)"-~ dll
for X a binomial1y distributed random variable. That is, if X is binomial1y dis-

tributed with parameters nand p and Y is beta-distributed with parameters k and
n - k+ I, then Fy(p) = 1 - Fx(k - I).
*29 Suppose that X has a binomial distribution with parameters nand p and Y has a
negative binomial distribution with parameters rand p. Show that Fx(r - 1) =
1- Fy(n-r).
*30 If U is a random variable that is uniformly distributed over the interval [0, I], then
the random variable ZA = [UA - (I - U))']/A is said to have TlIkey's symmetrical
lambda distribution. Find the first four moments of Z).. Find two different A's,
say .;\1 and ';\2, such that ZAI and Z)'2 have the same first four moments and unit
standard deviations.
IV
JOINT AND CONDITIONAL DISTRIBUTIONS,
STOCHASTIC INDEPENDENCE,
MORE EXPECTATION

The purpose of this chapter is to introduce the concepts of k-dimensional
distribution functions, conditional distributions, joint and conditional expecta-
tion, and independence of random variables. It. like Chap. II, is primarily a
" definitions~and~their-understanding" chapter.
The chapter is divided into four main sections in addition to the present
one. In Sec. 2, joint distributions, both in cumulative and density~function
form, are introduced. The important k~dimensional discrete distribution,
caned the multinomial, is included as an example. Conditional distributionsand
independence of random variables are the subject of Sec. 3. Section 4 deals
with expectation with respect to k-variate distributions. Definitions of covari-
ance, the correlation coefficient, and joint moment generating functions, a1l
of which are specia.l expectations, are given. The important concept of condi-
tional expectation is discussed in Subsec. 4.3. Results relating independence
and expectation are presented in Subsec. 4.5, and the famous Cauchy-Schwarz
inequality is proved in Subsec. 4.6. The last main section, Sec. 5, is devoted
to the important bivariate normal distribution, which gives one unified example
of many of the terms defined in the preceding sections.
130 JOINT AND CONDITIONAL DISTRIBUTIONS, STOCHASTIC INDEPENDENCE IV
This chapter is the multidimensional analog of Chap. II. It provides

definitions needed to understand distributional-theory results of Chap. V.
2 JOINT DISTRIBUTION FUNCTIONS

In the study of many random experiments, there are, or can be, more than one
random variable of interest; hence we are compe11ed to extend our definitions
of the distribution and density function of one random variable to those of
several random variables. Such definitions are the essence of this section,
which is the multivariate counterpart of Secs. 2 and 3 of Chap. II. As in the
univariate case we wi]] first define, in Subsec. 2.1, the cumulative distribution
function. Although it is not as convenient to work with as density functions,
it does exist for any set of k random variables. Density functions for jointly
discrete and jointly continuous random variables wi]] be given in Subsecs. 2.2
and 2.3, respectively.
2.1 Cumulative Distribution Function

Definition 1 Joint cumulative distribution function Let Xl' X 2 , ••• ,Xk
be k random variables a]] defined on the same probability space
.(n, .91, P[· D. The joint cumulative distribution function of Xl' "" x k ,
denoted by Fx" ... ,xJ·, ... , .), is defined as P[XI < Xl; ... ; X k '< xd for
a]] (Xl' X2' ... , Xk)' IIII
Thus a joint cumulative distribution function is a function with domain

euclidean k space and counterdomain the interval [0, I]. if k = 2, the joint
cumulative distribution function is a function of two variables, and so its
domain is just the xy plane.
EXAMPLE 1 Consider the experiment of tossing two tetrahedra (regular

four-sided polyhedron) each with sides labeled I to 4. Let X denote the
number on the downturned face of the first tetrahedron and Y the larger
of the downturned numbers. The goal is to find Fx , y(', .), the joint cu-
mulative distribution function of X and Y. Observe first that the random
variables X and Y jointly take on only the values
(1, I), (1, 2), (1, 3), (1, 4),
(2, 2), (2, 3), (2, 4),
(3, 3), (3, 4),
(4,4).
(The first component is the value of X, and the second the value of Y.)
2 JOINT DISTRIBUTION FUNGnONS 131
xz
=
8 4 •
• • •
13
"2
- 3 • • • •
~
• • • •
=2
"0
0
u
~ 1 • • • •
FIGURE 1 I I I .. Xl
Sample space for experiment of tossing 1 2 3 4
two tetrahedra. First tetrahedron
The sample space for this experiment is displayed in Fig.]. The 16

sample points are assumed to be equally likely. Our objective is to find
Fx, y(x, y) for each point (x, y). As an example let (x, y) = (2, 3), and
find Fx, y(2, 3) = P[X < 2; Y < 3]. Now the event {X < 2 and y.:s;; 3}
corresponds to the encircled sample points in Fig. I; hence Fx. y(2, 3) =
6
1 6' Similarly, Fx. y(x, y) can be found for other values of x and y.
Fx. y(x, y) is tabled in Fig. 2. IIII
We saw that the cumulative distribution function of a unidimensional

random variable had certain properties; the same is true of a joint cumulative.
We shaH list these properties for the joint cumulative distribution function of
two random variables; the generalization to k dimensions is straightforward.
TABLE OF VALUES OF Fx.r (x,y)
4<y 0 h .lL
16 -» 1
3<y<4 0 1\ Ch) -h -h
'-
2<y<3 0 -h N
4
n N
4
1<y<2 0 n -h -h ...l..
16
y<l 0 0 0 0 0
.... 1 1 x<2 2<x<3 3<x<4 4 x
FIGURE 2
Properties of bivariate cumulative distribution function F( " .)

•
(i) F( - 00, y) = lim F(x, y) = Oforal1 y, F(x, - (0) = lim F(x, y) =0
X-+ - 00 Y-+ - 00
for all x, and Jim F(x, y) = F( 00, (0) = ].

X-+ 00
Y-+oo
(ii) If Xl < and Yl < Y2, then P[Xl < X < X2; Yl < Y < Y2]
X2
= F(X2' Y2) - F(X2' Yl) - F(Xl' Y2) + F(x l , Yl) > O.
(iii) F(x, y) is right continuous in each argument; that is,
lim F(x + h, y) = Jim F(x, Y + h) = F(x, y).
o <h-+O 0 <h-+O
We wil1 not prove these properties. Property (ii) is a mono tonicity

property of sorts; it is not equivalent to F(Xl' Yl) < F(X2' Y2) for Xl < X2
and Yl < Y2 . Consider, for example, the bivariate function G(x, y) defined
as in Fig. 3. Note that G(Xl, Yl) < G(X2' Y2) for Xl < X2 and Yl < Y2'
yet G(1 +e,] +e)-G(1 +e,] -e)-G(1-e,1 +e)+G(1-e,] -e)= 1-
(l - e) - (1 - e) = 2e - I < 0 for e < t; so G(x, y) does not satisfy property
(ii) and consequently is not a bivariate cumulative distribution function.
Definition 2 Bivariate cumulative distribution function Any function

satisfying properties (i) to (iii) is defined to be a bivariate cumulative
distribution function without reference to any random variables. IIII
Definition 3 Marginal cumulative distribution function IfFx, y( " .) is

the joint cumulative distribution function of X and Y, then the cumulative
distribution functions Fx (') and F y (') are called marginal cumulative
dis tribution functions. IIII
TABLE OF G(x, y)
1 <y 0 x 1
O<y<l 0 0 y
y<O 0 0 0
x<O O<x<l l<x
FIGURE 3
2 JOINT DISTRIBUTION FUNCTIONS 133
Remark Fx(x) = Fx. y(x, (0), and Fy(y) = Fx. y(oo, y); that is, knowl~
edge of the joint cumulative distribution function of X and Y implies
knowledge of the two marginal cumulative distribution functions. fill
The converse of the above remark is not general1y true; in fact, an example
(Example 8) will be given in Subsec. 2.3 below that gives an entire family of
joint cumulative distribution functions, and each member of the family has the
same marginal distributions.
We wi11 conclude this section with a remark that gives an inequality
inv01ving the joint cumu1ative distribution and marginal distributions. The
proof is left as an exercise.
Remark Fx(x) + Fy(y) - I <Fx. y(x, y) < J Fx(x)Fy(y) for a1l x, y.

fill
2.2 J oint Density Functions for Discrete Random Variables

If Xl' X 2 , ••• , X k are random variables defined on the same probability space,
then (Xl' X 2 , ••• , X k ) is cal1ed a k~dimensional random variable.
Definition 4 Joint discrete random variables The k~dimensional ran~

dom variable (XI' X 2' ... , X,J is defined to be a k-dimensional discrete
random l'ariable if it can assume values only at a countable number of
points (XI' X2, •.. , Xk) in k~dimensional real space. We also say that
the random variables Xl' X 2 , ••• , X k are joint discrete random variables.
fill
Definition 5 Joint discrete density function If (Xl' X 2 , ••• , X k ) is
a k~dimensional discrete random variable, then the jOint discrete density
function of (Xl' X 2 , ••• , X k ), denoted byj~l,x2, .... xk(·' ., .,., .), is defined
to be
for (Xl' X2' ... , Xk), a value of (Xl' X 2 , "" X k ) and is defined to be 0
otherwise. fill
Remark I/x I, •••• xk(x l , " ' , Xk) = 1, where the summation IS over all
possib1e va1ues of (Xl' ... , X k ). fill
134 JOINT AND CONDmONAL DISTRIBUTIONS, STOCHASTIC INDEPENDENCE IV
Ix. y(x, y)
FIGURE 4 x
EXAMPLE 2 Let X denote the number on the downturned face of the first
tetrahedron and Ythelargerofthe downturned numbers in the experiment
of tossing two tetrahedra. The values that (X, Y) can take on are (I, I),
(I , 2), (I, 3), (] , 4), (2, 2), (2, 3), (2, 4), (3, 3), (3, 4), and (4, 4); hence X and
Yare jointly discrete. The joint discrete density function of X and Y
is given in Fig. 4.
In tabular form it is given as
(x, y) (l, ]) (1, 2) (I, 3) (J,4) (2, 2) (2,3) (2,4) (3, 3) (3,4) (4,4)
1 1 1 1 2 1 1 3 1 4
j ~. y(x,y) 16 16" 16 16 16 16 T6 T6" T6 16"
or in another tabular form as

1 1 1 4
4 16 16 16 1"6
1 1 3
3 16 T6 16
2 1 2
T6 16
] 1
T6
y/x 1 2 3 4 II/I
Theorem 1 If X and Yare jointly discrete random variables, then

knowledge of F x , y(', .) is equivalent to knowledge of Ix. y(', .}. Also,
the statement extends to k-dimensional discrete random variables.
2 JOINT DlsTRIBUOON FUNCTIONS 135'
PROOF Let (Xl' Yl)' (X2' Y2)' '" be the possible values of (X, Y).
If lx, y(', .) is given, then F x , y(x, y) = I/x, y(Xj, Yi)' where the summa~
tion is over an i for which Xl ::;; X and Yi < y. Conversely, if Fx. y(., .) is
given, then for (Xi, Yi), a possible value of (X, Y),
Ix,y(X;, Yi) = FX,y(Xb Yi) - limoFx,y(Xi - h, Y{)
O<h-+
- Jim FX.y(Xb Yi - h)
o <h-+O
+ Jim Fx,y(xj-h,Yi- h ). 1111
O<h-+O
Definition 6 Marginal discrete density If X and Yare jointly discrete

random variables, then Ix(') and Iy(') are cal1ed marginal discrete
density functions. More generally, Jet XiI' ... , Xim be any subset of the
jointly discrete random variables Xl' ... , X k ; then IXi I ' •••• Xim(X 11 , ••• , Xim )
is also ca]]ed a marginal density. IIII
Remark If Xl' "', X k are jointly" discrete random variables, then any
marginal discrete density can be found from the joint density, but not
conversely. F or example, if X and Yare jointly discrete with values
(Xl' YI)' (X2, Y2), ... , then
fX(X k ) = I Ix. y(Xi' Yi)

U: Xi= Xk}
and IIII
Heretofore we have indexed the values of (X, Y) with a single index,

namely i. That is, we listed values as (Xl, Yt), (X2 'Y2), ... , (Xi' Yi), ... , The
values of (X, Y) could also be indexed by using separate indices for the X and Y
values. For instance, we could let i index the possible X values, say X., ""
Xi' ... , and j index the possible Y values, say YI' ... , YJ' . • •. Then the values
of (X, Y) would be a subset of the points (Xi' Yj) for i I , 2, ... andj = 1, 2, ....
If this latter method of indexing is used, then the marginal density of X is
obtained as fo]]ows:
IX(Xk) = I/x. y(Xk' YJ)'

j
where the summation is over all Yj for the fixed Xk. The marginal density of Y
is analogously obtained. The fo]]owing example may help to c1arify these two
different methods of indexing the values of (X, Y).
EXAMPLE 3 Return to the experiment of tossing two tetrahedra, and define

X as the number on the downturned face of the first tetrahedron and Yas
the larger of the numbers on the two downturned faces. The joint
density of X and Y is given in Fig. 4. The values of (X, Y) can be Hsted

as (1, I), (I, 2), (1, 3), (l, 4), (2, 2), (2, 3), (2, 4), (3, 3), (3, 4), and (4, 4),
10 points in an. Or, if we note that X has values I, 2, 3, and 4; Y has
values I, 2, 3, and 4; and Y is greater than or equal X, the values of
(X, Y) are {(i,j): i = 1, ... ,4; j = 1, ... ,4; and i <j}. Let us use each
of these methods of indexing to evaluate Fx. y(2, 3) from the joint density.
Under the first method of indexing,
Fx, y(2, 3) = I
{i:Xi::;; 2. )Ii::;; 3}
fx. y(X;, Yi)
=fx,y(l, 1)+ lx, y(l, 2)

+ lx, y(1, 3) + Ix. y(2, 2) + Ix. y(2, 3) 6
10
Under the second method of indexing,
2 3
F x ,y(2, 3) =I Ifx,y(i,j) = 6
1 6'
i=lj=i
Similarly, an other values of Fx. y(., .) could be obtained. Also
ly(3) = I Ix, y(Xj, Yi) = lx, y(1, 3) + Ix, y(2, 3) + lx, y(3, 3)
{i:y£=3}
= l6 + rt +
3 5
16 = 1 6'
Simi1arly Iy(l) = -n" fy(2) = -hi, and ly(4) = 176' which together with
ly(3) = 156 give the marginal discrete density function of Y. 1111
EXAMPLE 4 We mentioned that marginal densities can be obtained from

the joint density, but not conversely. The following is an example of a
family of joi~t densities that a)) have the same marginals, and hence we
see that in genera] the joint density is not uniquely determined from
knowledge of the marginals. Consider altering the joint density given
in the previous examples as follows:
4
4 1
16 +e /6 - e 1
TI TI
3 /6 - e rt+e 3
1"6
1 2
2 TI 16
1
1 16
Y/x I 2 3 4
For each 0 < e < /6' the above table defines a joint density. Note that
the marginal densities are independent of e, and hence each of the joint
densities (there is a different joint density for each 0 < e < -16) has the
same marginals. IIII
We saw that the binomial distribution was associated with independent,

repeated Bernoul1i trials; we shaH see in the example below that the multinomial
distribution is associated with independent, repeated trials that generalize from
Bernoul1i trials with two outcomes to more than two outcomes.
EXAMPLE 5 Suppose that there are k + ] (distinct) possible outcomes of a

trial. Denote these 0 utcomes by J I' J2, ... , jk + I' and let Pi = P[Jd,
k+l
i = ], ... , k + I. Obviously we must have I Pi = I, just as P + q = I in
i= 1
the binomial case. Suppose that we repeat the trial n times. Let Xi
denote the number of times outcome .J i occurs in the n trials,
i = I, "', k + I. Jf the trials are repeated and independent, then the
discrete density function of the random variables Xl' ... , X k is
k+ I k
wherexj=O, ... ,nand I Xj=n. NotethatXk+1=n- I Xj'
i= I i I
To justify Eq. (I), note that the left-hand side is P[X1 = XI; X 2 = X2;
... ; X k + I Xk+ d; so, we want the probability that the n trials result in
exactly Xl outcomes ·JI' exactly X2 outcomes "2, •..• exactly Xk+ I outcomes
k+1
,Jk+l' where II Xi = n. Any specific ordering of these n outcomes has
probability p~l . p~2 ... Pk~\l by the assumption of independent trials.
and there are n!/xt! x 2 ! ... Xk+l! such orderings. IIII
Definition 7 Multinomial distribution The joint discrete density func-

tion given in Eq. (1) is called the multinomial distribution. IIII
The multinomial distribution is a (k + I) parameter family of distri-

butions, the parameters being n and PI' P2, "" Pk' Pk + 1 is, like q in the
binomial distribution, exactly determined by Pk+ t = I - PI - P2 - ... - Pk' A
.20
FIGURE 5 Xl
particular case of a multinomial distribution is 0 btained by putting, for example,

n = 3, k = 2, Pl = .2, and P2 = .3, to get
3'
!Xl,X2(Xl' X2) =f(x l , X2) = , '(3 ~ _ )' (.2Yl(.3YZ(.5)3
Xl ,X2 • Xl X2'
This density is plotted in Fig. 5.

We might observe that if Xl' X 2 , ••• , X k have the multinomial distribu-
tion given in Eq. (1), then the marginal distribution of Xi is a binomiaJ distri-
bution with parameters n and Pi' This observation can be verified by recalJing
the experiment of repeated, independent trials. Each trial can be thought of
as resuJting either in outcome 6i or not in outcome 6j, in which case the trial is
Bernoul1i, implying that Xi has a binomial distribution with parameters n and Pi'
2.3 Joint Density Functions for Continuous Random Variables
Definition 8 Joint continuous random variables and density function The

k-dimensional random variable (Xl' X 2, ... , X k ) is defined to be a
k-dimensional continuous random variable if and only if there exists a
function fXI •.... Xk(', ••• , .) ~ 0 such that
FXt. .... Xk(X l ' ""Xk)= J~koo '" J:loofxh ... ,Xk(Ut"",Uk)dul •.• duk (2)
for all (Xl' ... , Xk)' fXI ..... Xk(·' ... , .) is defined to be a joint probability
density function. IIII
As in the unidimensional case, a joint probability density function has

two properties:
(i) fx 1 , ••• ,Xk (X 1 , " ' , Xk) > O.
(ii) roo'" r /x ...... x,(x 1 , ... , x.) dx, ... dx, = 1.

A unidimensional probability density function was used to find proba-
bilities. For example, for X· a continuous random variable with probability
density Ix('), P[a < X < b] = J!/x(x) dx; that is, the area under Ix(') over the
interval (a, b) gave P[a < X < b]; and, more genera]]y, P[X E B] = JBlx(x) dx;
that is, the area under fx(') over the set B gave P[X E B]. In the two-dimen-
sional case, volume gives probabilities. For instance, let (Xl' X 2) be jointly
continuous random variables with joint probability density function
lXI, x2(x 1, x 2), and let R be some region in the X1X2 plane; then P[(X1' X 2) E R]
= JJfXI,X2(X1' x 2) dx 1 dX2; that is, the probability that (Xl' X 2) fa]]s in the
R
region R is given by the volume under IXI. X2(', .) over the region R. In particu-
lar if R = {(Xl' X2): a 1 < Xl < h1; a2 < X2 < b2}, then
Ajoint probabllitydensityfunction is defined as any nonnegative integrand

satisfying Eq. (2) and hence is not uniquely defined.
EXAMPLE 6 Consider the bivariate function

f(x, y) = K(x + y)I(o, l)(x)/(O, 1)(Y) = K(x + y)/u(x, y),
where U = {(x, y): 0 < x < ] and 0 < y < I}, a unit square. Can the
constant K be selected so that f(x, y) wi]] be a joint probability density
function? If K is positive, I(x, y) > O.
5_
00
00
f-
00
oo
Kf(x, y) dx dy = fo
1
t 1
K(x + y) dx dy
1 1
=K I I (x + y) dx dy
° °
I (t + y) dy
1
= K
°
=K(t+-D
=1
140 JOINT AND CONDITIONAL DlSTRmUTlONS, STOCHASTIC INDEPENDENCE IV
f(x, y)
(1, 1, 2)
(1,0, 1)
(0, 1, 1)
--~--~~------~--x
/ 2
I //
I /
_____ ....Y____
/ _
FIGURE 6 y
for K = 1. So lex, y) = (x + y)I(o, l)(x)I(o, 1)(Y) is a joint probability

density function. It is sketched in Fig. 6.
Probabilities of events defined in terms of the random variables can
be obtained by integrating the joint probability density function over the
indicated region; for example
t t
P[O<X<t;O<Y<:!-]= f f(x+y)dxdy
° °
= £: (1 +~) dy
_
-
1
32
+ 64
1
-2-
- 64'
which is the volume under the surface z = x + y over the region {(x, y):
0< x < t; 0 < y<!} in the xy plane. ///1
Theorem 2 If X and Yare jointly continuous random variables, then

knowledge of Fx. y(., .) is equivalent to knowledge of an/x, y(., .). The
remark extends to k~dimensional continuous random variables.
PROOF . .
For a given /x y(., .), Fx y(x, y) is obtained for any
(x, y) by
2 JOINT DlsTRmUTION FUNCTIONS 141
For given Fx,Y(', .), anfx,Y(x, y) can be obtained by
iJ2 Fx y(x, y)
fx, y(x, y) = iJ~ iJy
for x, y points, where Fx, y(x, y) is differentiable. IIII
Definition 9 Marginal probability density functions If X and Yare

jointly continuous random variables, then fx(') and fy(') are ca]]ed
marginal probability density functions. More general1y, let XiI' ... , Xim
be any subset of the jointly continuous random variables Xl' "', X k •
IXi!' "', Xim (xiI' ... , x im ) is ca]]ed a marginal density of the m-dimensional
random variable (XiI' ... , Xi m)' IIII

Remark If Xl' .. " X k are jointly continuous random variables, then
any marginal probability density function can be found. (However,
knowledge of a11 marginal densities does not, in general, imply knowledge
of the joint density, as Example 8 below shows.) If X and Yare jointly
continuous, then
fx(x) = f~ oofx, y(x, y) dy and (3)
smce
dF x(x) d [X (fOO ) ] 00
Ix(x) = dx = dx f_ 00 _ oofx, y(u, y) dy du = f _ oofx, y(x, y) dy.
I1I1
EXAMPLE 7 Consider the joint probability density
fx, y(x, y) = (x + y)l(o, l)(x)l(o, 1)(Y)'
f f (u + v) du dv
y x
Fx,Y(x, y) = l(o,l)(x)l(o,l)(Y)
°° 1 x
+ 1(0, l)(x)l[l, OO)(y) fo fo (u + v) du dv
y 1
+ 1[1, oo)(x)I(o, l)(Y) fo fo (u + v) du dv
+ l[l, oo)(x)I[1, ooly)

2
= t{(x y + xy2)1(0.1)(x)I(0, l)(Y) + (x 2 + x)l(o, l)(x)I[1, oo)(y)
+ (y + y2)1[1, oo)(x)l(o, l)(Y)} + 1[1, oo)(x)I[1, OO)(y).
fx(x) = I 00
-00
fx, y(X, y) dy
= 1(0, l)(X) f (X + y) dy
0
1
= (X + !)/(o, l)(X);
or,
aFx,y(x, oo)
fx(x) =
ax
aF x(X)
ax
a
= l(o,ll x ) ax (+x)
2
= (x + !-)/(o, l)(X). IIII
EXAMPLE 8 Let /x(x) and /y(y) be two probability density functions with
corresponding cumulative distribution functions Fx(x) and Fy(y), respec-
tively. For - I < ex < I, define
Ix, y(x, y; ex) = /x(x)/y(y}{1 + ex[ 2Fx(x) - 1][2Fy(Y) - In. (4)
We will show (i) that for each ex satisfying -I < ex < I, fx, y(x, y; ex) is a
joint probability density function and (ii) that the marginals of/x, y(x, y; ex)
are/x(x) and/y(y), respectively. Thus, {Ix, y(x, y; ex): -I < ex < I} will be
an infinite family of joint probability density functions, each having the
same two given marginals. To verify (i) we must show that/x , y(x, y; ex)
is nonnegative and, if integrated over the xy plane, integrates to I.
/x(x)fy(y}{l + ex[2F x(x) - 1][2Fy(y) - I]} >0

if I >-ex[2Fx(x) - 1][2Fy(y) - I];
but ex, 2Fx(x) - I, and 2Fy(y) - I are all between -I and I, and hence
also their product, which implies/x. y(x, y; ex) is nonnegative. Since
I = r/x(X) dx = f'oo (roo fx. y(x, y; rx) dy ) dx,

3 CONDITIONAL DISTRIBUTIONS AND STOCHASTIC INDEPENDENCE 143
it suffices to show that/x(x) and/y(y) are the marginals of/x, y(x, y; ex).
I-00
00
fx, y(x, y; ex) dy
= foo fx<x)fY(Y}{1 + ex[2F x(x) - 1][2F y(y) - I]} dy

-00
00 00
=/x(x) I-fY(Y) dy + ex/x(x)[2Fx(x) - l]f_ 00

[2F y(y) - l]fy(y) dy
00
= fx(x), noting that f_ 00 [2F y (y) - l]fy(y) dy
= fo (2u -
1
1) du = 0
by making the transformation u = Fy(y). IIII
3 CONDITIONAL DISTRIBUTIONS AND

STOCHASTIC INDEPENDENCE
] n the preceding section we defined the joint distribution and joint density
functions of several random variables; in this section we define conditional
distributions and the related concept of stochastic independence. Most defini-
tions will be given first for only two random variables and later extended to k
random variables.
3.1 Conditional Distribution Functions

for Discrete Random Variables
Definition 10 Conditional discrete density function Let X and Y be

jointly discrete random variables with joint discrete density function
Ix, y(., .). The conditional discrete density function of Y given X = x,
denoted by /Ylx(·1 x), is defined to be
f (Ix)_/x,y(x,y) (5)
YIX y - /x(x) ,
if fx(x) > 0, where fx(x) is the marginal density of X evaluated at x.

/Ylx(· Ix) is undefined for fx(x) = O. Similarly,
f (xl ) _ Ix, y(x, y) (6)

xlr Y - fy(y) ,
if fy(y) > o. IIII
Since X and Yare discrete, they have mass points, say Xl' X2, ... for X
and YI, Y2'··· for Y. If Ix(x) > 0, then X = Xi for some i, and IX(Xi)
= P[X = xJ The numerator of the right-hand side of Eq. (5) is lx, y(Xi' J'j)
= P[X = Xi; Y = Yj]; so
for Yl a mass point of Yand Xi a mass point of X; hence !Ylx(" Ix) is a condi-
tional probability as defined in Subsec. 3.6 of Chap. l. !Ylx(· Ix) is called a
conditional discrete density function and hence should possess the properties
of a discrete density function. To see that it does, consider X as some fixed
mass point of X. Then IYlx(yl x) is a function with argument Y; and to be a
discrete density function must be nonnegative and, if summed over the possible
values (mass points) of Y, must sum to 1. IYlx(yl x) is nonnegative since
!x, y(x, y) is nonnegative and Ix(x) is positive.
" "Ix. y(x, Yj) 1" fx(x)

L-fYlx(y1Ix)
1
= L-j j () =j-() 4- fx,y(x, Yj) =j-()
x x xX J xX
= 1,
where the summation is over all the mass points of Y. (We used the fact that
the marginal discrete density of X is obtained by summing the joint density of
X and Y over the possible values of Y.) So !Ylx(" Ix) is indeed a density; it
tells us how the values of Yare distributed for a given value x of X.
The conditional cumulative distribution of Y given X = x can be defined
for two jointly discrete random variables by recalling the close relationship
between discrete density functions and cumulative distribution functions.
Definition 11 Conditional discrete cumulative distribution If X and Y

are jointly discrete random variables, the conditional cumulative distribu-
tion of Y given X = x, denoted by F y1x(· Ix), is defined to be Fy1x(Y1 x) =
P[ Y < YI X = x] for Ix(x) > o. IIII
Remark Fy1x(Y1 x) = L IYlx(Yjl x). IIII

{j:Yj:;;;Y}
EXAMPLE 9 Return to the experiment of tossing two tetrahedra. Let X

denote the number on the downturned face of the first and Y the larger
of the downturned numbers. What is the density of Y given that X = 2?
Ix y(2, 2) -(6 1
ly/x(21 2) = ix(2) = T\ = 2
Ix y(2, 3) rt 1
IYlx(312) = ix(2) = T~ =4
Ix y(2,4) l6 1
ly/x(41 2) = ix(2) = n = 4'
Also,
for y = 3
for y = 4. IIII
Definition 12 Conditional discrete density function Let (Xl' ... , Xk ) be

a k-dimensional discrete random variable, and let XII"'" Xi .. and
Xii' ... , Xjs be two disjoint subsets of the random variables Xb ... , X k •
The conditional density of the r-dimensional random variable (Xi I ' ••• , XiJ
given the value (x j l ' . . . , x)J of (Xj l ' •• " X j ) is defined to be
_lxil' .... xi,.,Xh ••.. ,XJ"'(Xil' ••• , Xi,., Xii' ""Xis)
IXh .... , xJ",(xit' .. " X j,.} IIII
EXAMPLE 10 Let Xl' ... , Xs be jointly discrete random variables, Take

r = s = 2, (Xii' X i2 ) = (Xt, X 2 ), and (XiI' X h ) = (X3' Xs); then
r (I
JXl,X2IXJ,XsXbX2X3,XS-
)- lXI, X2, XJ. xs(XI, X2, Xl' Xs)
I" ( ) •
J XJ. Xs X 3 , Xs IIII
EXAMPLE II Suppose 12 cards are drawn without replacement from an

ordinary deck of playing cards. Let Xl be the number of aces drawn,
X 2 be the number of 2s, X3 be the number of 3s, and X 4 be the number
of 4s. The joint density of these four random variables is given by
where Xi = 0, I, 2, 3, or 4 and i = I, ... ,4, subject to the restriction that

LXi < 12. There-are a large number of conditional densities associated
with this density; an example is
!X2,X4Ix 1 ,xJ (X2, x 4 1x h X3)
(12 - x~~x,)
where Xi = 0, 1, .. " 4 and X2 + X 4 < 12 - Xl - X3' IIII
3.2 Conditional Distribution Functions

for Continuous Random Variables
Definition 13 Conditional probability- density function Let X and y

be jointly continuous random variables with joint probability density
function Ix. y(x, y). The conditional probability density junction of Y
given X = X, denoted by IYlx('1 x), is defined to be
I' ( 1 ) _ Ix. y(x, y) (7)
JYIX Y X - /x(x)
if Ix(x) > 0, where Ix(x) is the marginal probability density of X, and is

undefined at points when Ix(x) = O.
Similarly,
y(x, y)
IXly(xly) = if/y(Y) > 0, (8)
Iy(y)
and is undefined if /y(y) = 0 IIII
IYlx( -I x) is called a (conditional) probability density function and hence
should possess the properties of a probability density function. IYlx( -I x) is
clearly nonnegative, and
3 CONDmONAL DISTRIBUTIONS AND STOCHASTIC INDEPENDENCE 147
The density IYlx(·1 x) is a density of the random variable Y given that x

is the value of the random variable X. In the conditional density IYlx(· J x),
x is fixed and could be thought of as a parameter. Consider IYlx('1 xo), that is,
the density of Y given that X was observed to be Xo • Now Ix, y(x, y) plots as a
surface over the xy plane. A plane perpendicular to the xy plane which inter-
sects the xy plane on the line x = Xo will intersect the surface in the curve
Ix. y(xo, y). The area under this curve is
00
f i x . y(Xo, y) dy = Ix(xo)·
-00
Hence, if we divide Ix. y(Xo , y) by Ix(xo), we obtain a density which is precisely

/Ylx(yl xo)·
Again, the conditional cumulative distribution can be defined in the
natural way.
Definition 14 Conditional continuous cumulative distribution If X and

Yare jointly continuous, then the conditional cumulative distribution of Y
gi ven X = x is defined as
Fy1x(yl x) = fY IYlx(zl x) dz
-00
for all x such that/x(x) > O. IIII
EXAMPLE 12 Suppose/x. y(x, y) = (x + y)I(o, l)(x)I(o, 1)(Y)·

r (I ) - (x + y)l(o, Il x )I(o, 1)(Y) _ x +Y 1 ()
JYIX Y x - (+
X
1)1 ()
2: (0,1) X
-
X
+ 1:1 (0.1) Y
for 0 < x < I. Note that
Fy1x(ylx) = f~oo/Ylx(zlx) dz
yx +z 1 fY
=
fox+t dz =
x+t 0
(x + z) dz
I
--1 (xy + y2/2) for 0 < y < I. IIII
x+'!
Conditional probability density functions can be analogously defined for

k-dimensional continuous random variables. F or instance,
f Xl,X2.X 4 Ixl,XS (X l' x 2, x 4 I X3, Xs

)_fXl,X2,Xl,X4,Xs(Xl,X2,X3,X4'XS)
- f ( )
Xl,X s x3 , Xs
for /Xl' xs( X3 , x 5) > o.
3.3 More on Conditional Distribution Functions

We have defined the conditional cumulative distribution Fy1x(yl x) for either
jointly continuous or jointly discrete random variables. If X is discrete and Y
is any random variable, then Fy1x(Ylx) can be defined as PlY yl X = xl if x
is a mass point of X. We would like to define PlY ~ yl X = x] and more
generally P[A I X = x], where A is any event, for X either a discrete or continu-
ous random variable. Thus we seek to define the conditional probability of an
event A given a random variable X = x.
We start by assuming that the event A and the random variable X are
both defined on the same probability space. We want to define P[A I X = x].
If X is discrete, either x is a mass point of X, or it is not; and if x is a mass point
of X,
P[A I X = 1= P[A; X = xl
x P[X = xl '
which is well defined; on the other hand, if x is not a mass point of X, we are
not interested in P[A I X = xl. Now if X is continuous, P[A I X = xl cannot be
analogously defined since P[X = xl = 0; however, if x is such that the events
{x - h < X < x + h} have positive probabllity for every h > 0, then P[A I X = x]
could be defined as
P[A I X = xl = lim P[A Ix - h < X < x + hl (9)

o <h-+O
provided that the limit exists. We will take Eq. (9) as our definition of
P[A I X = xl if the indicated limit exists, and leave P[A I X = xl undefined other-
wise. (It is, in fact, possible to give P[A I X = xl meaning even'if P[X = x] = 0,
and such is done in advanced probability theory.)
We will seldom be interested in P[A I X = x] per se, but will be interested
in using it to calculate certain probabilities. We note the following formulas:
00
I (i) P[A] =I P[A I X = xdfx(Xi) (10)

i= 1
if X is discrete with mass points Xl' X2, ••••
00
(ii) P[A] = r
·-00
P[A I X = x]fx(x) dx (11)
if X is continuous.
(iii) P[A; X E B] = I P[A I X = xdfx(Xi) (12)

{i:Xf eB}
if X is discrete with mass points Xl' X2, .••.
(iv) P[A; X E B] = f P[A I X

B
= x]fx(x) dx (13)
if X is continuous.
Ai' .;gh we will nQt prove the above formulas, we note that Eq. (10) is
just tl:-:' iheorem of total probabilities given in Subsec. 3.6 of Chap. I and the
other& are generalizations of the same. Some problems are of such a nature
that it is easy to find P[A I X = x] and difficult to find P[A]. If, however, /x( .)
is known, then PtA] can be easily obtained using the appropriate one of the
above formulas.
Remark Fx. y(x, y) = S: ooFy1x(yl x')fx(x') dx' results from Eq. (13) by
taking A = {Y < y} and B = (- 00, x]; and Fy(y) = J~ooFYlx(yl x)/x(x) dx
is obtained from Eq. (II) by taking A = {Y < y}. IIII
We add one other formula, whose proof is also omitted. Suppose

A = {h( X, Y) < z}, where h( " .) is some function of two variables; then
(v) P[A I X = x] = P[h(X, Y) zl X = x] = P[h(x, Y} < zl X = x].

(14)
The following is a classical example that uses Eq. (II); another example
utilizing Eqs. (14) and (II) appears at the end of the next subsection.
EXAMPLE 13 Three points are selected randomly on the circumference of

a circle. What is the probability that there will be a semicircle on which
all three points will lie? By selecting a point" randomly," we mean that
the point is equally likely to he any point on the circumference of the
circle; that is, the point is uniformly distributed over the circumference
of the circle. Let us use the first point to orient the circle; for example,
orient the circle (assumed centered at the origin) sO that the first point
falls on the positive x axis. Let X denote the position of the second point,
and let A denote the event that all three points lie on the same half circle.
X is uniformly distributed over the interva1 (0, 2n). According to Eq. (II),
P[A] = SP[A I X = x]/x(x) dx. Note that for 0 < x < n, P[A I X = xl =
(n - x + n)/2n since, given X = x, event A occurs if and only if the
third point falls between x - nand n. Similarly, P[A I X = x] =
(x + n - n)/2n for n x < 2n. Hence P[A] = J~1t P[A I X = x](l/2n) dx =
(l/2n){jo[(2n - x)/2n] dx + J;1t(x/2n) dx}=,f. II1I
3.4 .Independence
When we defined the conditional probability of two events in Chap. I, we also
defined independence of events. We have now defined the conditional distri-
bution of random variables; so we should define independence of random
variables as well.
Definition 15 Stochastic independence Let (Xl' X 2, ... , X k ) be a

k-dimensional random variable. X., X 2 , ••• , X k are defined to be
stochastically independent if and only if
k
F Xl ..... X,,(XI' ..• , Xk) = n F XlXI)
1:= I
(15)
Definition 16 Stochastic independence Let (Xl, X 2, ..• , Xk ) be a

k-dimensional discrete random variable with joint discrete density func-
tion fx 1•••.• x,,( " ... , '). Xl" .. , X k are stochastically independent if and
only if
k
fXl ..... X"(X., .. " Xk) = nfxlxt) (16)
j= 1
Definition 17 Stochastic independence Let (Xl' "', X k ) be a k-dimen-

sional continuous random variable with joint probability density function
fx 1.... , x,,( " ... , '). Xl" .. , X k are stochastically independent if and only
if
k
fXI ..... X"(X., ... , Xk) = nfxlxi) (17)
i= 1
forallxt"",xk' IIII
Remark Often the word "stochastically" will be omitted. IIII
We saw that independence of events was closely related to conditional

probability; likewise independence of random variables is closely related to
conditional distributions of random variables. For example, suppose X and
Y are two independent random variables; thenfx, y(x, y) =/x{x)fy(y) by defini-
tion of independence; however, fx, y(x, y) = fy[x(yl x)fX<x) by definition of
conditional density, which implies thatfYlx(Y1 x) = fy(y); that is, the conditional
3 CONDmONAL DISTRIBUTIONS AND STOCHASTIC INDEPENDENCE 151
density of Y given x is the unconditional density of Y. So to show that two

random variables are not independent, it suffices to show thatfYjx(yl x) depends
on x.
EXAMPLE 14 Let X be the number on the downturned face of the first

tetrahedron and Y the larger of the two downturned numbers in the ex-
periment of tossing two tetrahedra. Are X and Y independent? Ob-
3
viouslynot, since/Y lx(21 3) = P[ Y = 21 X = 3] = 0 =I: fy(2) = P[ Y = 2] = 1 6'
IIII
EXAMPLE J 5 Let Ix. y(x, y) = (x + y)I(O.l)(x)I(o.l)(Y)' Are therefore X

and Y independent? No, since fYjx(Y 1 x) = [(x + y)/(x + !)]I(o.l)(Y) for
0< x < l,fYlx(ylx) depends on x and hence cannot equal/y(y). IIII
EXAMPLE 16 Let Ix. y(x, y) = e-(x+Y)I(o. oo)(x)I(o. oo)(Y). X and Yare

independent since
fx. y( x, y) = [e-xI(o. oo)(x)][e- YI(o. OO)(Y)] = fx(x)fy(y)
for all (x, y). IIII
It can be proved that if XI. ... , X k are jointly continuous random variables,
then Definitions IS and 17 are equivalent. Similarly, for jointly discrete
random variables, Definitions 15 and 16 are equivalent. It can also be proved.
that Eq. (15) is equivalent to P[XI E B l ; ... ; X k E Bd = n P[X,
k
i= 1
E B i ] for sets
Bl , ••• , The following important result is easily derived using the above
Bk •
equivalent notions of independence.
Theorem 3 If Xl' ... , X k are independent random variables and

gl('), "', gk(') are k functions such that Y j = giXj)' .i = I, ... , k are
random variables, then Yl , ... , Yk are independent.
Note that if gjl(Bj ) = {z: gj(z) E B), then the events
PROOF
{ Y j E B j } and {Xj E gjl(Bj)} are equivalent; consequently, P[ Yl E B l ; ... ;
Y E Bd = P[X E gl l (B ); . . .
k I X E g,;-l(Bk)] =
1 k
k
P[Xj E gjl(B)] n
j= 1
= n" P[Yj
j=l
E Bj ]. IIII
152 JelNT AND CONDITIONAL DISTRIBUTIONS, STOCHASTIC INDEPENDENCE IV
For k = 2, the above theorem states that if two random variables, say
X and Y, are independent, then a function of X is independent of a function of
Y. Such a result is certainly intuitively plausible.
We will return to independence of random variables in S ubsec. 4.5.
Equation (14) of the previous subsection states that P[h(X, Y) < zl X = x]
= P[h(x, y) < z I X = x]. Now if X and Y are assumed to be independent,
then P[h(x, y) < zl X = x] = P[h(x, y) < z], which is a probability that may
be easy to calculate for certain problems.
EXAMPLE 17 Let a random variable Y represent the diameter of a shaft

and a random variable X represent the inside diameter of the housing
that is intended to support the shaft. By design the shaft is to have
diameter 99.5 units and the housing inside diameter 100 units. If the
manufacturing process of each of the items is imperfect, so that in fact Y
is uniformly distributed over the interval (98.5, 100.5) and X is uniformly
distributed over (99, 101), what is the probability that a particular shaft
can be successfully paired with a particular housing, when" successfully
paired" is taken to mean that X - h < Y < X for some small positive
quantity h? Assume that X and Yare independent; then
P[X - h < Y < X] = foo P[X - h < Y < Xl X = x]/x(x) dx

-00
= f99
101
P[x - h < Y < x]tdx.
Suppose now that h = 1; then

x - 98.5
for 99 < x < 99.5
2
I
P[x - 1 < Y < x] = - for 99.5 < x < 100.5
2
100.5 - (x - 1)
for 100.5 < x < 101.
2
Hence,
P[X - 1 < Y < Xl = f 99

101
P[x - 1 < Y < x]t dx
99.5
=f 1(X - 98.5)! dx
99
f f
100.5 101
+ t(t) dx + (t)(100.5 - x + l)t dx = 7
1 6.
99.5 100.5 •
IIII
4 EXPECTATION 153
4 EXPECT A TION
When we introduced the concept of expectation for univariate random variables
in Sec. 4 of Chap. II, we first defined the mean and variance as particular expec-
tations and then defined the expectation of a general function of a random vari-
able. Here, we will commence, in Subsec. 4.1, with the definition of the
expectation of a general function of a k-dimensional random variable. The
definition will be given for only those k-dimensional random variables which
ha ve densities.
4.1 Definition
Definition 18 Expectation Let (Xl, ... , X k ) be a k-dimensional ran-

dom variable with density IXI •.... Xk(', ... , '). The expected value of a
function g(', ... , .) of the k-dimensional random variable, denoted by
8[g(X1> "', X,,)], is defined to be
8[g(X1> ... , Xk)] = L g(Xb ... , xk)/xl . .... X (Xl' .. " k

x,,) (18)
if the random variable (Xl' ... , Xk ) is discrete where the summation is

over all possible values of (Xh ... , X k ), and
(Xl (Xl (Xl

= f_(Xl f_(Xl ... f_(Xl g(Xl' ... , x l )!XI .... Xk (X 1 ' .•. , xk ) dX 1 ••• dXk (19)
if the random variable (Xl' ... , X,,) is continuous. ///1
In order for the above to be defined, it is understood that the sum and
mUltiple integral, respectively, exist.
Theorem 4 In particular, if g(Xl' ... , Xk) = Xi' then
8[g(Xl' "" X,,)] = 8[Xd = /lx,. (20)
PROOF Assume that (Xl, ... , X k ) is continuous. [The proof for

(Xl' ... , Xk) discrete is similar.]
(Xl (Xl (Xl
8[g(Xl' ... , X k )] = J-(Xl f_(Xl '" f_(Xl xi!xt, ... ,Xk(Xh ... , Xk) dXl ... dx"
= J(Xl
-(Xl xdx.(xt) dXr = 8[X,]
using the fact that the marginal density /XiXi) is obtained from the joint
density by
f ···f
00 00
-00 -00
fXl, ... ,Xk(X1,···,Xk)dx1···dxi-1·dxi+1··.dxk· IIII
Similarly, the following theorem can be proved.
Theorem 5 If g(X1, ... , Xk) = (Xi - 8[Xd)2, then

8[g(Xb ... , X k)] = 8[(Xi - 8[Xi])2] = var [Xi]. IIII
We might note that the" expectation" in the notation 8[Xi ] of Eq. (20)
has two different interpretations; one is that the expectation is taken over the
joint distribution of Xl' ... , X k , and the other is that the expectation is taken
over the marginal distribution of Xl. What Theorem 4 really says is that
these two expectations are equivalent, and hence we are justified in using the
same notation for both.
EXA MPLE 18 Consider the experiment of tossing two tetrahedra. Let X

be the number on the first and Y the larger of the two numbers. We gave
the joint discrete density function of X and Y in Example 2.
8[XY] = I xy/x, y(X, y)

= 1 . 1(/6) + 1 ·2(/6) + 1 . 3(ft) + 1 ·4(/6)
+ 2·2(/6) + 2 . 3(/6) + 2 . 4(/6) + 3 . 3(/6)
+ 3 . 4(-Ar) + 4 . 4(146) = \365.
8[X + y] = (1 + 1)/6 + (1 + 2)/6 + (1 + 3)/6 + (l + 4)-Ar
+ (2 + 2)l6 + (2 + 3)-Ar + (2 + 4)-Ar + (3 + 3)-h
+ (3 + 4) /6 + (4 + 4) 146 = i ~ .
8[X] = i, and 8[y] = ~~; hence 8[X + y] = 8[X] + 8[Y]. IIII
EXAMPLE 19 Suppose lx, y(x, y) = (x + y)I(o,1)(x)I(o, l)(Y)·
8[XY] = f° f
1
0
1
xy(x + y) dx dy =-1-.
8[X + Y] = tt
1 1
(x + y)( x + y) dx d y = ~.
7
8[X] = 8[Y] = 1 2. IIII
4 EXPECTATION 155
EXAMPLE 20 Let the three-dimensional random variable (Xl' X 2 , X 3) have

the density
fXl.X2,xixh X2, X3) = 8XIX2 X3 I(o.l)(xd I (o, 1)(x2)I(o,l)(X3)'
Suppose we want to find (i) 8[3XI + 2X2 + 6X3], (ii) 8[XI X 2X 3], and
(iii) 8[XI X 2]. For (i) we have g(Xh X2, X3) = 3XI + 2X2 + 6X3 and
obtain
8[g(XI , X 2 , X 3)] = 8[3XI + 2X2 + 6X3]
= nnn (3XI + 2x2 + 6X3)8xIX2 X3 dXI dX2 dX3 = \2.
For (ii), we get

8[XI X 2 X 3 ] nJ~ J5 8X/X2 x/ dXI dX2 dX3 =
2 8
2 "
and for (iii) we get 8[XI X 2 ] =l IIII
The following remark, the proof of which is left to the reader, displays a
property of joint expectation. It is a generalization of (ii) in Theorem 3 of
Chap. II.
Remark 8[~ CiUi(Xl • . . . . XJ] = ~ ci 8[u,(X" ... , X.)l for constants

CI 'C 2 '···,C m • 1/11
4.2 Covariance and Correlation Coefficient
Definition 19 . Covariance Let X and Y be any two random variables

defined on the same probability space. The covariance of X and Y,
denoted by cov [X, Y] or (Ix, y, is defined as
COv [X, Y] = 8[(X - ttx)(Y-tty)] (21)
provided that the indicated expectation exists. III/
Definition 20 Correlation coefficient The correlation coefficient, de-

noted by p[X, Y] or Px, y, of random variables X and Y is defined to be
COv [X, Y]
px , y = (Ix (Iy
(22)
provided that cov [X~ y], (Ix, and (Iy exist, and (Ix > 0 and (Iy > O. III I
156 JOINT AND CQNDmONAL DISTRIBUTIONS, STOCHASTIC INDEPENDENCE IV
Both the covariance and the correlation coefficient of random variables

X and Yare measures of a linear relationship of X and Y in the following sense:
cov [X, Y] will be positive when X - J1x and Y - J1y tend to have the same sign
with high probability, and cov [X, Y] will be negative when X - Jlx and Y - Jly
tend to have opposite signs with high probability. cov [X, Y] tends to measure
the linear relationship of X and Y; however, its actual magnitude does not have
much meaning since it depends on the variability of X and Y. The correlation
coefficient removes, in a sense, the individual variability of each X and Y by
dividing the covariance by the product of the standard deviations, and thus the
correlation coefficient is a better measure of the linear relationship of X and Y
than is the covariance. Also, the correlation coefficient is unitless and, as we
shall see in Subsec. 4.6 below, satisfies - I Px , y :s: I.
Remark cov lX, Y] = 4[(X - J1x)(Y - J1y)] = 4[XY] - J1xJ1Y·
PROOF 4[(X - J1x)( Y - J1y)] = 4[XY - J1x Y - J1y X + J1x J1y]

= 4[XY] - J1x4[ Y] - J1y 4[X] + J1xJly
= 4[XY] - J1x J1y. 1111
EXAMPLE 21 Find Px , y for X, the number on the first, and Y, the larger of
the two numbers, in the experiment of tossing two tetrahedra. We would
expect that Px, y is positive since when X is large, Y tends to be large too.
We calculated 4[XY], 4[X], and 4[ Y] in Example 18 and obtained
4[XY] = \365, 4[X] = 4-, and 4[ Y] = i ~. Thus cov [X, YJ = _\365 - t· i ~
= t%. Now 4[X2] = 34° and 4[ y2] = \76°; hence var [X] = t and
var [y] = ~~. So,
t~ .' 2
1111
Px, y = J~Jll =
4 64
EXAMPLE 22 Find Px, y for X and Y iflx, y(x, y) = (x + y)I(o, 1)(x)I(o, 1)(Y)'
We saw that 4[XY] = ~ and 4[X] = 4[ y] = 1\ in Example 19.. Now
4[X2] = 4[ y2] = 152; hence var [X] = var [y] = 1\14< Finally
1. _ J:.2... 1
3 144
Px, y = 11 = - - .
T44 II
Does a negative correlation coefficient seem right? 1111
4 EXPECTATION 157
4.3 Conditional Expectations

In the following chapters we shall have occasion to find the expected value
of random variables in conditional distributions, or the expected value of one
random variable given the value of another.
Definition 21 Conditional expectation Let (X, Y) be a two-dimensional

random variable and g( . , . ), a function of two variables. The conditional
expectation of g(X, Y) given X = x, denoted by C[g(X, Y)I X = x], is
defined to be
C[g(X, Y) I X = xl = f~oo g(x, Y)/YlxCYl x) dy (23)
if (X, Y) are jointly continuous, and
C[g(X, Y) I X = x] = L g(x, Y)/Ylx(yil x) (24)
if (X, Y) are jointly discrete, where the summation is over all. possible
~~~y @
In particular, if g(x, y)=y, we have defined C[YIX=x]=C[Ylx].
C[Ylx] and C[g(X, Y)lx] are functions of x. Note that this definition can be
generalized to more than two dimensions. For example, let (Xl' ... , X k ,
Yl , ... , Ym ) be a (k + m)-dimensional continuous random variable with density
/XI, ... , Xk , YI, ... , Ym(Xl' . " , Xk' Yl' ., ., Ym); then
C[g(Xb ... , X k , Yb ... , Ym)IXl, ... , Xk]
.= foo ... foo

-00 -00
g(x b ... , Xk' Yl' ... , Ym)
EXAMPLE 23 In the experiment of tossing two tetrahedra with X, the

number on the first, and Y, the larger of the two numbers, we found that
! for Y =2
/Ylx(yI2) = J.! for Y = 3
{
4 for Y = 4
in Example 9. Hence C[ YI X = 2] = LY/YI x(y I X = 2) = 2·1- + 3·1 + 4· t
_ 11
T'
. . . - .- -. /f //
EXAMPLE 24 For/x, y(x, y) = (x + y)I(o.l)(x)I(o, lly), we found that

x+y
/Ylx(Y Ix) = + 1 I(o.I)(Y)
X "2
for 0 < x < I in Example 12. Hence
X+ y d 1 (X- + -1)
8[ Y I X = x] = I° y x+!
I
Y=
x+! 2 3
for 0 < x < I. //11
As we stated above, 8[g( Y) Ix] is, in general, a function of x, Let us

denote it by hex); that is, hex) = 8[g(Y)lx]. Now we can evaluate the expecta-
tion of heX), a function of X, and will have 8[h(X)] = 8[8[g(Y)1 Xl].
This gi ves us
8[8[g(Y) IX]] = 8[h(X)] = J: oo

h(x)fx(x) dx
= J-
00
oo
8[g(Y)lx]fx(x) dx
= roo [J~ooU(Y)/Y1X(YI x) dY]fx(X) dx
= Joo Joo g(Y)/Ylx(Ylx)/x(x)dydx

-00 -00
= J~oo J~oog(Y)fx.Y(X' y) dy dx
= 8[g(Y)].
Thus we have proved for jointly continuous random variables X and Y
(the proof for X and Y jointly discrete is similar) the following simple yet very
useful theorem.
Theorem 6 Let (X, Y) be a two-dimensional random variable; then

8[g( Y)] = 8[8[g( Y) IX]], (25)
and in particular
8[ Y] = 8[8[ YI X]]. (26)
IIII
Definition 22 Regression curve 8[ YI X = x] is called the regression
curve of Yon x. It is also denoted by ,LlYIX=x=P.Ylx· IIII
4 EXPECTATION 159
Definition 23 Conditional variance The variance of Y given X = x is

defined by var [YI X = x] = <9'[ y21 X = x] - (<9'[ YI X = X])2. IIII
Theorem 7 var [Y] = <9'[var [YI X]] + var [<9'[ YI X]].

PROOF
<9'[var [YI X]] = <9'[ <9'[ y21 X]] - <9'[(<9'[ YI X])2]

= <9'[ y2] - (<9'[ Y])2 - <9'[(<9'[ YI X])2] + (<9'[ y])2
= var [ Y] - <9'[(<9'[ YI X])2] + (<9'[<9'[ YI X]])2
= var [Y] - var [<9'[ YI X]]. IIII
Let us note in words what the two theorems say. Equation (26) states
that the mean of Y is the mean or expectation of the conditional mean of Y,
and Theorem 7 states that the variance of Y is the mean or expectation of the
conditional variance of Y, plus the variance of the conditional mean of Y.
We will conclude this subsection with one further theorem. The proof
can be routinely obtained from Definition 21 and is left as an exercise. Also,
the theorem can be generalized to more than two dimensions.
Theorem 8 Let (X, Y) be a two-dimensional random variable and

gl ( .) and g2(') functions of one variable. Then
(i) <9'[gl( y) + g2( Y) I X = x] = <9'[gl (Y) I X = x] + <9'[g2( Y) I X = x].

(ii) <9'[g1( y)g2(X) I X = x] = g2(X)<9'[gl( Y) I X = x]. IIII
4.4 Joint Moment Generating Function and Moments
We will use our definition of the expectation of a function of several variables

to define joint moments and the joint moment generating function.
Definition 24 Joint moments The joint raw moments of Xl' ... , X k

are defined by <9'[X1 X~2 ... X~k], where the r/s are 0 or any positive
1
integer; the joint moments about the means are defined by

<9'[(Xl - /lxJ'l ... (Xk - /lxJ'k]. IIII
Remark If ri = rj = 1 and all other rm's are 0, then that particular joint
moment ~bout the means becomes <9'[(Xi - /lx,)(Xj - /lx)], which is just
the covanance between XI and Xi' IIII
\ ~
Definition 25 Joint moment generating function The joint moment

generating function of (Xl, ... , X k) is defined by
mx" ... ,x.(t., ... , I.) = of [expJ.r; xl (27)
if the expectation exists for all values of t l , ... , tk such that -h < tj < h
for some h > O,j = 1, ... , k. IIII
The rth moment of Xj may be obtained from m Xt • ... , Xk(tl, ••. , t k ) by
differentiating it r times with respect to tj and then taking the limit as all the t's
approach 0. Also 8[X~ Xj] can be obtained by differentiating the joint moment
generating function r times with respect to ti and s times with respect to tj and
then taking the limit as all the t's approach O. Similarly other joint raw
moments can be generated.
Remark mx(td = mx. y(tl' 0) = limmx. y(tl' t 2), andm y(t2) = mx. yeO, t 2)
t2-+0
= limmx y(tl' t 2 ); that is, the marginal moment generating functions can
tl-+0 •
be obtained from the joint moment generating function. IIII

An example of a joint moment generating function will appear in Sec. 5
of this chapter.
4.5 Independence and Expectation

We have already defined independence and expectation; in this section we will
relate the two concepts.
Theorem 9 If X and Yare independent and gl (.) and g2(') are two
functions, each of a single argument, then
8[gl(X)g2( Y)] = 8[gl(X)]' 8[g2( Y)].
PROOF We will give the proof for jointly continuous random
variables.
t9'[gI(X)g2(Y)] = fOO fOO gl(X)g2(y)fx,y(X, y) dx dy

-00 -00
= fOO fOO gl(X)g2(y)fx(x)fy(y) dx dy

-00 -00
= f~oogl(X)fx(X) dx . f~oog2(Y)fY(Y) dy
= 8[g leX)] . t9'[g2(Y)]' IIII
4 EXPECTATION 161
Corollary If X and Yare independent, then cov [X, Y] = O.
PROOF Take gl(X) = x - Jlx and g2(Y) = Y - Jly; by Theorem 9,
cov [X, Y] = <9'[(X - Jlx)( Y - Jly)] = <9'[gl(X)g2( Y)]

= <9'[gl(X)]<9'[g2(Y)]
= <9'[X - Jlx] . <9'[ Y - Jly] = 0 since ~[X - Jlx] = o. IIII
Definition 26 Uncorrelated random variables Random variables X and

Yare defined to be uncorrelated if and only if cov [X, Y] = o. IIII
Remark The converse of the above corollary is not always true; that is,
cov [X, Y] = 0 does not always imply that X and Yare independent, as
the following example shows. IIII
EXAM PLE 25 Let U be a random variable which is uniformly distributed

over the interval (0, 1). Define X = sin 2nU and Y = cos 2nU. X and
Yare clearly not independent since if a value of X is known, then U is
one of two values, and so Y is also one of two values; hence the conditional
distribution of Y is not the same as the marginal distribution. ~[ Y] =
g
Sb cos 2nu du = 0, and ~[X] = sin 2nu du = 0; so cov [X, Y]= ~[XY] =
S6 sin 2nu cos 2nu du = 1- S6 sin 4nu du = o. IIII
Theorem 10 Two jointly distributed random variables X and Yare

independent if and only if mx, y(tb t 2 ) = m X(tl)my(t 2) for all t1 , t2 for
which -h < tj < h, i = I, 2, for some h > O.
PROOF [Recall that mX(tl) is the moment generating function of X.

Also note that mX(tl) = mx, y(tb 0).] X and Y independent imply that
the joint moment generating function factors into the product of the
marginal moment generating functions by Theorem 9 by taking gl (x) = et1x
and g2(Y) = e2Y• The proof in the other direction will be omitted.
IIII
Remark Both Theorems 9 and 10 can be generalized from two random

variables to k random variables. IIII
162 JOINT,&.ND CONDmONAL DISTRIBUTIONS, STOCHASTIC INDEPENDENCE IV
4.6 Cauchy-Schwarz Inequality
Theorem 11 Cauchy-Schwarz inequality Let X and Y have finite

second moments; then (4[XY])2 = 14[XY] 12 < 4[X2)4[ y2), with equality
if and only if P[ Y = cX] = I for some constant c.
PROOF The existence of expectations 4[X), 4[ Y), and 4[XY)

follows from the existence of expectations 4[X2) and 4[ y2). Define
o <h(t) = 8[(tX - Y)2) = 4[X2)t 2 - 24[XY]t + 4[ y2). Now h(t} is a
quadratic function in t which is greater than or equal to o. If h(t} > 0,
then the roots of h(t) are not real; so 4(4[XY]}2 - 44[X 2)4[ y2) < 0,
or (8[XY])2 < 8[X2)4[ y2). If h(t} = 0 for some t, say to, then
8[(t o X - Y)2] = 0, which implies P[to X = Y] = 1. IIII
Corollary IPx, yl < 1, with equality if and only if one random variable
is a linear function of the other with probability 1.
PROOF Rewrite the Cauchy-Schwarz inequality as 14[UV)1 <

J 4[U2)4[ V 2), and set U = X - Jlx and V = Y- Jly. IIII
5 BIVARIATE NORMAL DISTRIBUTION

One of the important multivariate densities is the multivariate normal
density, which is a generalization of the normal distribution for a unidimensional
random variable. In this section we shall discuss a special case, the case of the
bivariate normal. In our discussion we will include the joint density, marginal
densities, conditional densities, conditional means and variances, covariance,
and the moment generating function. This section, then, will give an example
of many of the concepts defined in the preceding sections of this chapter.
5.1 Density Function
Definition 27 Bivariate normal distribution Let the two-dimensional

random variable (X, Y) have the joint probability density function
1
fx,y (x, y) =f(x, y)= 2 J
1taXay 1 - P
2
xexp ( - 1 2 [(X - JlX) 2- 2p X- Jlx Y- Jly + (Y - Jly) 2]) (28)

2(1 - P ) ax ax ay ay
5 BIVARIATE NORMAL DISTRIBUTION 163
z
z = f(x, y) for z > k
FIGURE 7 k
for - 00 < x < 00, - 00 < y < 00, where u y , u x , llx, lly, and p are con-
stants such that -I < p < I, 0 < Uy , 0 < Ux, -00 < llx < 00, and
- 00 < lly < 00. Then the random variable (X, Y) is defined to have a
bivariate normal distribution. fill
The density in Eq. (28) may be represented by a bell-shaped surface

z = I(x, y) as in Fig. 7. Any plane parallel to the xy plane which cuts the
surface will intersect it in an elliptic curve, while any plane perpendicular to the
xy plane will cut the surface in a curve of the normal form. The probability
that a point (X, Y) will lie in any region R of the xy plane is obtained by
integrating the density over that region;
P[(X, Y) is in R] = III(x, y) dy dx. (29)

R
The density might, for example, represent the distribution of hits on a vertical
target, where x and y represent the horizontal and vertical deviations from the
central lines. And in fact the distribution closely approximates the distribution
of this as well as many other bivariate populations encountered in practice.
We must first show that the function actually represents a density by
showing that its integral over the whole plane is I; that is,
I I
00 00
-00 -00
f(x, y) dy dx = 1. (30)
The density is, of course, positive. To simplify the integral, we shall substitute
x-ll y - lly
u = x and v=--- (31)
U x Uy
so that it becomes
I I
00
oo
1 .-
e -U-/(l_p2)](U2- 2pUV+V2) d v d u.
- 00 - 00 2n J 1 _ p2
On completing the square on u in the exponent, we have
and if we substitute
u - pv du
w = --;-.===- and dw = ,
JI -p2 JI- p2
the integral may be written as the product of two simple integrals
e -w /2 d w I J 1 e -v /2 d v,
oo oo
I 1 -- 2 2
(32)
- 00 "
J 2n - 2n 00
both of which are I, as we have seen in studying the univariate normal distri-
bution. Equation (30) is thus verified.
Remark The cumulative bivariate normal distribution
may be reduced to a form involving only the parameter p by making the

substitution in Eq. (31). fill
5.2 Moment Generating Function and Moments

To obtain the moments of X and Y, we shall find their joint moment generating
function, which is given by
I I
. 00 00
t2
rnX,y(tl' t 2 ) = m(tl' t 2 ) = C[elX+t2Y] = i 1X
+ Yf(x, y) dy dx.
-00 -00
Theorem 12 The moment generating function of the bivariate normal

distribution is
BIVARIATE NORMAL DISTRIBUTION 165
PROOF Let us again substitute for x and y in terms of u and v to

obtain
m(tl, t2)
The combined exponents in the integrand may be written
2(1 ~ p2) [u 2 - 2puv + v2 - 2(1 p2)t,a x u - 2(1- p2)t 2 a y vI,
and on completing the square first on u and then on v, we find this expres-
sion becomes
- 2(1 ~ p2) {[u - pv - (I - p2)t,ax J2 + (1 - p2)(V - pt,ax - t2ay)2
-(1-p2)(tfui +2pt1t2UXUy +tiu~)},

which, if we substitute
and
becomes
_!w 2 -lz2 + !(trui + 2pt1t2UXUy + tiui),
and the integral in Eq. (34) may be written
m(th t 2) = e'lf.l x+ t1 J4y exp[!(trui + 2Ptlt2 UXUy + t~u:)J
x f-00
00 foo
- 00
-2n1 e-wl/2-z1/2 dw dz
= exp[t1Jlx + t2 Jly + !(tiui + 2Ptl t2 UXUy + ti ui)]
since the double integral is equal to unity. II1I
Theorem 13 If (X, Y) has bivariate normal distribution, then
8[X] = /lx,
8[ Y] = /ly,
var [X] = ui,
var [Y] = u:,
cov [X, Y] = pux uy ,
166 JOINI' AND CONDmONAL DISTRIBUTIONS, STOCHASTIC INDEPENDENCE IV
and
Px , y = p.
PROOF The moments may be obtained by evaluating the appro-

priate derivative of m(tl' t 2) at tl = 0, t2 = O. Thus,
=..J1x2 + Gx2 .
Hence the variance of X is
Similarly, on differentiating with respect to t 2 , one finds the mean and

variance of Y to be Jly and G~. We can also obtain joint moments
by differentiating m{t., t 2 ) r times with respect to tl and s times with respect

to t2 and then putting tl and t2 equal to O. The covariance of X and Y is
4[(X - Jlx)(Y ~ Jly)] = 8[XY - XJly - YJlx + JlxJly]

= 8[XY] - Jlx Jly
82
- a 8 m(t., t 2 ) - Jlx Jly
t1 t2 tl = t2 = 0
= pGxGy .
Hence, the parameter p is the correlation coefficient of X and Y. III1
Theorem 14 If (X, Y) has a bivariate normal distribution, then X and Y

are independent if and only if X and Yare uncorrelated.
PROOF X and Yare uncorrelated if and only if cov [X, Y] = 0 or,

equivalently, if and only if Px, Y = P = O. It can be observed that if
p = 0, the joint density f(x, y) becomes the product of two univariate
normal distributions; so that p = 0 implies X and Yare independent.
We know that, in general, independence of X and Y implies that X and Y
are uncorrelated. 1I1I
5 BIVARIATE NORMAL DISI1UBUTION 167
5.3 Marginal and Conditional Densities
Theorem 15 If (X, Y) has a bivariate normal distri bution, then the mar-
ginal distributions of X and Yare univariate normal distributions; tbat is,
X is normally distributed with mean JJ.x and variance ai, and Y is nor-
mally distributed with mean JJ.y and variance ai.
PROOF The marginal density of one of the variables X, for example,
is by definition
f
QO
/x(x) = -QO
f(x, y) dy;
and again substituting
y ~ JJ.y
V=--
ay
and completing the square on v, one finds that
00 1
/x(x)=f -QO 2nax
J11 - p
2
x exp [ - -1 (X - JJ.X)2 - 1 2 (V - P X - JJ.X)2] dv.

2 ax 2(1 - p ) ax
Then the substitutions
and
show at once that
fx(X) = ~i exp [ _ ~ (X ;;x) 2],

the univariate normal density. Similarly the marginal density of Y may
be found to be
fy(y) == I 1 exp[1
,,2nay
2 - - (y - p,y) 2] .
2 ay 1111
Theorem 16 If (X, Y) has a bivariate normal distribution , then the
conditional distribution of X given Y = y is normal with mean
JJ.x + (pax/ay)(y ~ JJ.y) and variance ai(I - p2). Also, the conditional
distribution of Y given X = x is normal with mean JJ.y + (payjax)(x - JJ.x)
and variance ai(l _ p2).
PROOF The conditional distributions are obtained from the joint

and marginal distributions. Thus, the conditional density of X for fixed
values of Y is
f(x, y)
fXly(xly) = fy(y) ,
and, after substituting, the expression may be put in the form
!XIY(x Iy)
= J21!UX~1 _ pi exp {- 2uj(/- p2) [x - I1x -7.X (y - I1r)]'}, (35)
which is a univariate normal density with mean f1x + (paxlay)(y - f1y)

and with variance ai(1 - p2). The conditional distribution of Y may be
obtained by interchanging x and y throughout Eq. (35) to get
!Ylx(Y Ix)
= J21!
urJi _ p2 ex { - 2ui(11_ p2) [Y -
p I1r - p:: (x - I1X)] T (36)
IIII
As we already noted, the mean value of a random variable in a conditional

distribution is called a regression curve when regarded as a function of the fixed
variable in the conditional distribution. Thus the regression for X on Y = Y in
Eq. (35) is f1x + (paxlay)(y - f1y), which is a linear function of y in the present
case. For bivariate distributions in general, the mean of X in the conditional
density of X given Y = Y will be some function of y, say g(.), and the equation
x =g(y)
when plotted in the xy plane gives the regression curve for x. It is simply a
curve which gives the location of the mean of X for various values of Y in the
conditional density of X given Y = y.
For the bivariate normal distribution, the regression curve is the straight
line obtained by plotting
pax
x = f1x + ~ (y - f1y),
ay
as shown in Fig. 8. The conditional density of X given Y = y, fXIY(x Iy), is

also plotted in Fig. 8 for two particular values Yo and Yl of Y.
PROBLEMS 169
FIGURE 8
PROBLEMS
(a) If P[X> Y] = 1, then S[X] > SlY].
(b) If S[X] > SlY], then P[X> y] = 1.
(e) If S[X] > S[ Y], then P[X> y] > O.
(a) If Fx(z) > Fy(z) for alJ z, then SlY] > S[X].
(b) If S[y] > C[X], then Fx(z) > Fy(z) for an z.
(e) If Cry] > S[X], then Fx(z) > Fy(z) for some z.
(d) If Fx(z) Fy(z) for all z, then P[X = Y] = 1.
(e) If Fx(z) > F y(z) for all z, then P{X < Y] > O.
(I) If Y = Xl I, then Fx(z) Fy(z + 1) for all z.
3 If X I and X 2 are independent random variables with distribution given by
P[X, 1] =P[X~ = 1] ! for i = 1, 2, then are Xl and X I X 2 independent?
4 A penny and dime are tossed. Let X denote the number of heads up_ Then
the penny is tossed again. Let Y denote the number of heads up on the dime
(from the first toss) and the penny from the second toss.
(a) Find the conditional distribution of Y given X = I.
(b) Find the covariance of X and Y.
5 If X and Y have joint distribution given by
lx, y(x, y) U(O,y)(x)I(o. J)(Y).
(a) Find COy [X, Y].
(b) Find the conditional distribution of Y given X = x.
6 Consider a sample of size 2 drawn without replacement from an urn containing
three ba1ls, numbered 1, 2, and 3. Let X be the number on the first ball drawn
and Y the Jarger of the two numbers drawn.
(a) Find the joint discrete density function of X and Y.
(b) Find P[X = 11 Y 3].
(e) Find cov [X. Yl.
170 JOINT AND CONDITIONAL DISTRIBUTIONS. STOCHAS'l1C INDEPENDENCE IV
7 Consider two random variables X and Y having a joint probability density

function
Ix. r(X, y) = lxyl(o.Jr)(y)I(o. 2)(X).
(a) Find the marginal distributions of X and Y.
(b) Are X and Y independent?
8 If X has a Bemoul1i distribution with parameter p (that is, p[X = 1] = p = 1
~ p[X = OD, 8[ YI X = 0] = 1, and 8[YI X = 1] = 2, what is 8[ Y]?
9 Consider a sample of size 2 drawn without replacement from an urn containing
three balls, numbered 1, 2, and 3. Let X be the smaller of the two numbers
drawn and Y the larger.
(a) Find the joint discrete density function of X and Y.
(b) Find the conditional distribution of Y given X = 1.
(c) Find cov [X, Y].
10 Let X and Y be independent random variables, each having the same geometric
distribution. Find P[X = Y].
11 If F(·) is a cumulative distribution function:
(a) Is F(x, y) = F(x) + F(y) a joint cumulative distribution function "1
(b) Is F(x, y) = F(x)F(y) a joint cumu1ative distrIbution function?
(c) Is F(x, y) = max [F(x), F(y)] a joint cumulative distribution function?
(d) Is F(x, y) = min [F(x), F(y)] a joint cumulative distribution function "1
12 Prove
F,,(x) + Fr(y) ~ I <F". y{x, y) <VF,,(x)Fr(Y) for al1 x, y.

13 Three farr coins are tossed. Let X denote the number of heads on the first two
coins, and let Y denote the number of tails on the last two coins.
(a) Find the joint distribution of X and Y.
(b) Find the conditional distribution of Y given that X = 1.
(c) Find cov [X, Y].
14 Let random variable X have a density function 1(·), cumulative distribution
function F( .), mean p, and variance (12. Define Y = or; + fJx, where or; and fJ are
constants satisfying ~ 00 < or; < 00 and {J > O.
(a) Select or; and fJ so that Y has mean 0 and variance 1.
(b) What is the correlation coefficient between X and Y?
(c) Find the cumulative distrIbution function of Y in terms of or:. (J, and F( .).
(d) If X is symmetrically distrIbuted about p, is Y necessarily symmetrically
distributed about its mean "1 (HINT: Z is symmetrically distributed about
constant C if Z ~ C and ~(Z ~ C) have the same distribution.)
15 Suppose that random variable X is uniformly distributed over the interval (0, 1);
that is, I,,(x) = 1(0. 1)(x). Assume that the conditional distribution of Y given
X = x has a binomia1 distribution with parameters n and p = x; i.e.,
p[Y = yl X = x] = C) x'(1 - x),,-' for y = 0, 1, .•• , n.

PROBLEMS 171
(a)Find "[y].
(b)Find the distribution of Y.
16 Suppose that the joint probabiJity density function of (X, Y) is given by
Ix. y(x, y) = [1 - IX(l - 2x)(l - 2y)]I(o.1)(x)I(o. 1)(Y)'
where the parameter IX satisfies -1 < IX < 1.

(a) Prove or disprove: X and Yare independent if and only if X and Yare un-
correlated.
An isosceles triangle is formed as indicated in the sketch.

(b) If (X. Y) has the joint density given above, pick IX to maximize the expected
area of the triangle.
(c) What is the probability that the triangle falls within the unit square with
corners at (0, 0), (1, 0), (1, 1), and (0, 1) 7
*(d) Find the expected length of the perimeter of the triangle.
17 Consider tossing two tetrahedra with sides numbered 1 to 4. Let Y 1 denote the
smalJer of the two downturned numbers and Y 2 the larger.
(a) Find the joint density function of Y1 and Y 2 •
(b) Find pry! ~ 2, Y 2 2].
(e) Find the mean and variance of Y 1 and Y 2 •
(d) Find the conditional distribution of Y 2 given Y 1 for each of the possible
values of Y 1 •
(e) Find the correlation coefficient of Y1 and Y 2 •
18 Let/x. rex, y) = e-(x+7)I (0. <X!)(x)I(o. <X!)(y)
(a) Find P[X> lJ. (b) Find p[1 < X + Y < 2].
(e) Find P[X < YI X < 2 Yl. (d) Find m such that p[X + Y < m] = i.
(e) Find p[0 < X < 11 Y = 2]. (f) Find the correlation coefficient of X and Y.
*19 Let /x. rex, y) = e -"(1 - e-x)l(o. "J(x)l[O, <X!)(y) + e-x(l - e-")I(o. x)(y)I[o. <X!)(x).
(a) Show that/x. y(', .) is a probat?ility density function.
(b) Find the marginal distributions of X and Y.
(e) Find S[YI X = x] for 0 < x.
(d) Find P[X 2, Y 2].
(e) Find the correlation coefficient of X and Y.
(n Find another joint probability density function having the same marginals.
172 JoINT AND CONDITIONAL DISTRIBUTIONS, STOCH~C INDEPENDENCE IV
"'20 Suppose X and Y are independent and identically distributed random variables
with probability density function f(·} that is symmetrical about o.
(a) Prove that Pfl X + YI 21 XI] > 1_
(b) Select some symmetrical probability density function f('}, and evaluate
P[IX YI <2IXI]·
*21 Prove or disprove: If 8[YI X] = X,8[XI Y] = Y, and both 8[X2] and 8[y2] are
finite, then P[X = Y] = 1. (Possible HINT: P[X = Y] = 1 if var [X Y] = 0.)
22 A multivariate Chebyshev inequaJity: Let (Xl, .. _, Xm) be jointly distributed with
8 [Xl] ILl and var [Xl] = u~ for j = 1, "', m. Define Al = {I Xl ILl I < Vmtul}'
m
Show that P[ () A l ] > 1 - r2. for t>O.
1=1
23 Let fx( .) be a probability density function with corresponding cumulative dis-
tribution function F x(·). In terms of fx(') and/or F x('}:
(a) Find p[X> XO + .::lxl X>xo].
(b) Find P[xo < X <Xo + .::lxl X> xo].
(c) Find the limit of the above divided by .::lx as .::lx goes to O.
(d) Evaluate the quantities in parts (a) to (c) for fx(x} = 'Ae-;,xI(o. OC)(x).
24 Let N equal the number of times a certain device may be used before it breaks.
The probability is p that it will break on anyone try given that it did not break
on any of the previous tries.
(a) Express this in terms of conditional probabiJities.
(b) Express it in terms of a density function, and find the density function.
25 Player A tosses a coin with sides numbered 1 and 2. B spins a spinner evenly
graduated from 0 to 3. B's spinner is fair, but A's coin is not; it comes up 1
with a probability p, not necessarily equal to 1. The payoff X of this game is the
difference in their numbers (A's number minus B's). Find the cumulative dis-
tribution function of X.
26 An urn contains four balls; two of the balls are numbered with aI, and the other
two are numbered with a 2. Two bal1s are drawn from the um without replace-
ment. Let X denote the smaller of the numbers on the drawn balls and Y the
larger.
(a) Find the joint density of X and ·Y.
(b) Find the marginal distribution of Y.
(c) Find the cov [X, Y].
27 The joint probability density function of X and Y is given by
fx. rex, y} = 3(x + y)I(o. 1)(x y)l(o. l)(x}I(o. 1)(y}.
(Note the symmetry in x and y.)

(a) Find the marginal density of X.
(b) Find P[X + Y < .5].
(c) Find 8[ YI X = x].
(d) Find cov [X, Y].
PROBLEMS 173
28 The discrete density of X is given by Ix(x) = xl3 for x = 1, 2, and IYI x(Ylx) is
binomia1 with parameters x and !; that is,
I" x(Ylx} ~P[Y~YI X ~xl ~ (;) U).
for Y = 0, ... , x and x = 1, 2.
(a) Find 8[X] and var [X).
(b) Find 8[ Y).
(c) Find the joint distribution of X and Y.
29 Let the joint density function of X and Y be given by Ix. y(x, y) = 8xy for 0 < x
< y < 1 and be 0 e1sewhere.
(a) Find 8[ YI X = x).
(b) Find 8[XYIX=x].
(c) Find var [YI X = x).
30 Let Y be a random variab1e having a Poisson distribution with parameter "-.
Assume that the conditiona1 distribution of X given Y = y is binomia11y distrib-
uted with parameters y and p. Find the distribution of X, if X = 0 when Y = O.
31 Assume that X and Yare independent random variab1es and X ( y) has binomia1
distribution with parameters 3 and t (2 and i). Find P[X = Y).
32 Let X and Y have bivariate norma1 distribution with parameters p'x = 5, p. Y = 10,
ai = 1, and a~ = 25.
(a) If p > 0, find p when P[4 < Y < 161 X = 5] = .954.
*(b) If p = 0, find P[X + Y < 16).
33 Two dice are cast 10 times. Let X be the number of times no Is appear, and 1et
Y be the number of times two Is appear.
(a) What is the probabiJity that X and Y wi11 each be 1ess than 3?
(b) What is the probabiHty that X + Y wi11 be 4?
34 Three coins are tossed n times.
(a) Find the joint density of X, the number of times no heads appear; Y, the num-
ber of times one head appears; and Z, the number of times two heads appear.
(b) Find the conditiona1 density of X and Z given Y.
35 Six cards are drawn without rep1acement from an ordinary deck.
(a) Find the joint density of the number of aces X and the number of kings Y.
(b) Find the conditiona1 density of X given Y.
36 Let the two-dimensiona1 random variab1e (X, Y) have the joint density
lx, y(x, y) = 1(6 - ,x - y)I(O, 2)(x)I(2.4ly).
(a)Find 8[YI X = x], (b) Find8[Y2IX=x].
Find var [YI X = x],
(c) (d) Show that 8[ y] = 8[8[ YI X]].
(e) Find 8[XYI X = x).
37 The trinomia1 distribution (muhinomia1 'with k + 1 = 3) of two random variab1es
X and Y is given by
n'
Ix. y(x, y) = ' pXqY(1 -/J _ q)"-X-Y
x!y!(n - x - y)!
for x, y = 0, 1, ... , n and x + y < n, where 0 <p, 0 .-:::;: q, and p + q < 1.
174 JOINT AND CONDmONAL DISTRIBUTIONS, STOCHASTIC iNDEPENDENCE IV
,.
Find the marginal distribution of Y.
(a)
(b) Find the conditional distribution of X given Y, and obtain its expected
value.
(e) Find p[X, Y].
38 Let (X, Y) have probability density function/x. y(x, y), and let u(X) and v(Y) be
functions of X and Y, respectively. Show that
8[u(X)v(Y) I X = x] = u(x)8[v(Y) I X = x].
39 If X and Yare two random variables and 8[YI X = x] = ft, where ft does not
depend on x, show that var [Y] = 8[var [YI Xl],
40 If X and Yare two independent random variables, does 8[ YI X = x] depend
onx?
41 If the joint moment generating function of (X, Y) is given by mx. y(II' ( 2 ) =
exp[l(lf+ In] what is the distribution of Y?
42 Define the moment generating function of YI X = x. Does my(l) = 8[mYI x(I)]?
43 Toss three coins. Let X denote the number of heads on the first two and Y
denote the number of heads on the last two.
(a) Find the joint distribution of X and Y.
(b) Find 8[YI X = 1].
(e) Find px. y.
(d) Give a joint distribution that is not the joint distribution given in part (a)
yet has the same marginal distributions as the joint distribution given in
part (a).
44 Suppose that X and Y are jointly continuous random variables, /Ylx(ylx) =
I(x. X+I)(Y), and/x(x) = 1(0. O(x).
(a) Find 8[Y]. (b) Find cov [X, Y].
(e) Find P[X + Y <]]. (d) Find/xI y(xly)·
45 Let (X, Y) have a joint discrete density function
Ix. y(x, y)
pf(1 - PI)I-Xp~(l - P2)I-Y[l + «(x -PI)(Y -P2)]/(0.1)(x)/(0. l)(Y),

where 0 <PI < 1, 0 <P2 < 1, and -1 ~« < 1. Prove or disprove: X and Y
are independent if and only if they are uncorrelated.
*46 Let (X, Y) be jointly discrete random variables such that each X and Y have at
most two mass points. Prove or disprove: X and Yare independent if and only
if they are uncorrelated.
v
DISTRIBUTIONS OF FUNCTIONS OF
RANDOM VARIABLES

As the title of this chapter indicates, we are interested in finding the distribu-
tions of functions of random variables. More precisely, for given random
variables, say Xl' X 2 , ••. , X n , and given functions of the n given random variables,
say gl(·, ...•. ), g2(', ... , .), ... , gle(-, ... , .), we want, in general, to find the
joint distribution of Y 1, Y 2 , ... , Y Ie , where Yj = giX1, ..• , Xn),j = 1,2, ... , k.
If the joint density of the random variables Xl' X 2 , ••• , Xn is given, then theo-
retically at least, we can find the joint distribution of Ylo Y2 , ••• , Yle . This
follows since the joint cumulative distribution function of Y1 , ••• , Yle satisfies
the following:
FYI, ... , Yk(Ylo ... , YIe) = P[ Y1 < Y1; .•. ; Yk < YIe]
= P[gl(X1"", XII) < Yl;··· ;gle(X1,··., Xn)<YIe]
for fixed Y1' ••• , Yle' which is the probability of an event described in terms of
Xl' ..• , X n• and theoretically such a probability can be determined by integrat-
ing or summing the joint density over the region corresponding to the event.
The problem is that in general one cannot easily evaluate the desired probability
176 DISTRIBUTIONS OF FUNCTIONS OF RANDOM VARIABLES v
for each Yl' ... , Yk' One of the important problems of statistical inference, the
estimation of parameters, provides us with an example of a problem in which
it is useful to be able to find the distribution of a function of joint random
variables.
In this chapter three techniques for finding the distribution offunctions of
random variables will be presented. These three techniques are called 0) the
cumulative-distribution-function technique, alluded to above and discussed in
Sec. 3, (ii) the moment-generating-function technique, considered in Sec. 4, and
(iii) the transformation technique, considered in Secs. 5 and 6. A number of
important examples are given, including the distribution of sums of independent
random variables (in Subsec. 4.2) and the distribution of the minimum and
maximum (in Subsec. 3.2). Presentation of other important derived distributions
is deferred until later chapters. For instance, the distributions of chi·square,
Student's t, and F, all derived from sampling from a normal distribution, are
given in Sec. 4 of the next chapter.
Preceding the presentation of the techniques for finding the distribution
of functions of random variables is a discussion, given in Sec, 2, of expectations
of functions of random variables. As one might suspect, an expectation, for
example, the mean or the variance, of a function of given random variables can
sometimes be expressed in tenns of expectations of the given random variables.
If such is the case and one is only interested in certain expectations, then it is not
necessary to solve the problem of finding the distribution of the function of the
given random variables. One important function of given random variables
is their sum, and in Subsec. 2.2 the mean and variance of a sum of given random
variables are derived,
We have remarked several times in past chapters that our intermediate
objective was the understanding of distribution theory. This chapter provides
us with a presentation of distribution theory at a level that is deemed adequate
for the understanding of the statistical concepts that are given in the remainder
of this book.
2 EXPECTATIONS OF FUNCTIONS
OF RANDOM VARIABLES
2.1 Expectation Two Ways

An expectation of a function of a set of random variables can be obtained two
different ways. To illustrate, consider a function of just one random variable,
say X. Let g(') be the function, and set Y = g( X). Since Y is a random
2 EXPECT ATIONS OF FUNCTIONS OF RANDOM VARlABLES 177
variable, 8[ Y] is defined (if it exists), and 8[g( X)] is defined (if it exists). For
instance, if X and Y = g( X) are continuous random variables, then by definition
00
8[Y] = f_ooyfy(y) dy, (I)
and
8[g(X)] = f~oog(X)fx(X) dx; (2)
but Y = g(X), so it seems reasonable that 8[ Y] = 8[g(X)]. This can, in fact,

be proved; although we will not bother to do it. Thus we have two ways of
calculating the expectation of Y = g(X); one is to average Y with respect to the
density of Y, and the other is to average g(X) with respect to the density of X.
In general, for given random variables X h •.. , X,., let Y = g(Xl' .•• , XII);
then 8[ Y] = 8[g(Xh ..• , X,.)], where (for jointly continuous random variables)
8[ Y] = foo yfy(y) dy (3)

-00
and
oo .. 00
8 [g(Xl , ... , XII)] f
-00
••. J-00
g(x 1 ,···, x lI )fxl ..... XJXl , ... , XII) dXl ... dx,..
(4)
In practice, one would naturally select that method which makes the
calculations easier. One might suspect that Eq. (3) gives the better method of
the two since it involves only a single integral whereas Eq. (4) involves a multiple
integral. On the other hand, Eq. (3) involves the density of Y, a density that
may have to be obtained before integration can proceed.
EXAMPLE 1 Let X be a standard normal random variable, and Ietg(x) = x 2 •

For Y =g(X) = X 2 ,
00
8[Y] = f yfy(y) dy,

-00
and
8[g(X)] = 8[X2] = f 00
-00
x:tx(x) dx.
Now
178 DISTRIBtmONS OF FUNCTIONS OF RANDOM VARIABLES V
and
using the fact that Y has a gamma distribution with parameters r = ! and
A. = t. (See Example 2 in Subsec. 3. I below.) JI II
2.2 Sums of Random Variables

A simple, yet important, function of several random variables is their sum.
Theorem 1 For random variables Xl' ... , X"
.c[~ x,] = ~ .c[X,), (5)
and
var[f
1
Xl] =f
1
var[Xi ] + 2 L L cov[X
i<j
j , Xj]' (6)
PROOF That.c [ ~ X, = J t .c [X,] follows from a property of expec-

tation (see the last Remark in Subsec. 4.1 of Chap. IV).
var[~ x,] =.c[ (~x,- e[~ x.]) 2J =.c[ (~(XI - .c[X,])n

= eLt, J.(X,- .c[XaXX j - .c[Xj ])]
= L" L" 8[(Xj -

i=l j;.l
8[Xi ))(Xj - 8[XJ)]
= L" var[Xi) + 2 I I
i=l i<j
COV[Xi' Xj]' JIJI
Corollary If Xl' ... , X" are uncorrelated random variables, then
var [~ XI] = ~ var[X.J. /1/1
The following theorem gives a result that is somewhat related to the above
theorem inasmuch as its proof, which is left as an exercise, is similar.
EXPEcrATIONS OF FUNCTIONS OF RANDOM VARIABLES 179
Theorem 2 Let Xl' ... , XII and Y1, ••• , Ym be two sets of random vari-
ables, and let al, ••• , all and b1, ••• , bm be two sets of constants; then
Corollary If Xl, ... , X II are random variables and al' ... , a" are
cons tants, then
(8)
II
= L af var[X,] + L L a,aj cov[X j , X j ].
i=l i:f:.j
In particular, if Xl' ... , XII are independent and identically distributed

II
random variables with mean 11 x and variance 0'; and if X II = (I/n) LXi'
1
then
2
- O'x
8[XII ] = Ilx, and var [XII] = -. (9)
n
PROOF Let m n, Y, = Xi' and b i = ai' i = I, ... , n in the above

theorem; then
var[~ aiXi] = COy [t a,X" ~ b j Yl

and Eq. (8) follows from Eq. (7). To obtain the variance part of Eq. (9)
from Eq. (8), set a, = I/n and 0'; = var [Xd. The mean part of Eq. (9)
is routinely derived as
/11/
Corollary If Xl and X 2 are two random variables, then

var [Xl + X 2 ] = var [Xd + var [X 2 ] + 2 cov [XI' X 2 ]. (10)
JIll
Equation (10) gives the variance of the sum or the difference of two ran-
dom variables. Clearly
180 DISTRIBUTIONS OF FUNCTIONS OF RANDOM VARIABLES V
2.3 Product and Quotient

In the above subsection the mean and variance of the sum and difference of two
random variables were obtained. It was found that the mean and variance of
the sum or difference of random variables X and Y could be expressed in terms
of the means, variances, and covariance of X and Y. We consider now the
problem of finding the first two moments of the product and quotient of X and Y.
Theorem 3 Let X and Y be two random variables for which var [X YJ

exists; then
G[X YJ = P,x p,y + cov [X, YJ, (12)
and
var [X YJ
= p,~ var [X] + p,i var [YJ + 2P,x p,y cov [X, YJ
- (cov [X, YJ? + G[(X - p,X)2(Y - .uy}2] + (13)
2.u yG[(X - P,x)2 (Y - p,y)] + 2P,x G[(X - P,x)( Y - p,y)2].
PROOF
XY = P,x p,y + (X - P,x)p,y + (Y - p'y}p'x + (X - P,x)( Y - p'y).

Calculate G[X YJ and G[(Xy)2] to get the desired results. /11/
Corollary If X and Yare independent, G[X YJ = P,x P,y, and var [X YJ =
p'~ var [X] + p,i var [YJ + var [X] var [Y].
PROOF If X and Yare independent,
G[(X P,x)2( Y - p,y)2] = G[(X - .ux)2]G[( Y - p,y)2]
= var [Xl var [Y],
G[(X - p,X)2( Y - p,y)] = G[(X - p,X)2]G[Y - p,y] = 0,
and
Note that the mean of the product can be expressed in terms of the means
and covariance of X and Y but the variance of the product requires higher-order
moments.
In general, there are no simple exact formulas for the mean and variance
of the quotient of two random variables in terms of moments of the two random
variables; however, there are approximate formulas which are sometimes useful.
3 CUMULATIVE-DISTRIBUTION-FUNCTION TECHNIQUE 181
Theorem 4
8' [ -X] RP,x

::-- 1 cov[X, Yl+3"var
Jlx [YJ , (14)
2
Y Ily Ily Ily
and
var - [X]
Y
R::
~x)2(var[Xl
-
y Jlx
2 + var[Yl
2
_ 2cov[X,
p,y Jlx Jly
Yl) • (15)
PROOF To find the approximate formula for 8'[XI y], consider the
Taylor series expansion of xly expanded about (Ilx, Jly); drop all terms of
order higher than 2, and then take the expectation of both sides. The
approximate formula for var [XI Y] is similarly obtained by expanding in
a Taylor series and retaining only second-order terms. IIII
Two comments are in order: First, it is not unusual that the mean and
variance of the quotient XI Y do not exist even though the moments of X and Y
do exist. (See Examples 5, 23, and 24.) Second, the method of proof of
Theorem 4 can be used to find approximate formulas for the mean and variance
of functions of X and Yother than the quotient. For example,
1 var[X] ;z g(x, y)
8'[g(X, Y)] R:: g{J1x, Jly) + -2
I a
vx
2
IPX,py
1 a2 a2
+ - var[Y] ~ 2 g(x, y) + cov[X, Y] ~y ;lX g(x, y) , (16)
2 vy px.Py v v px.py
and
var[g(X, Y)] R:: var[X](: g(x, y)

X PX,py
}2 + var[Y](~
oy
g(x, y)
PX,py
)2
+ 2 cov[X, Yl(:x g(x, Yli . ~ g(x, y) ). (17)

Px.py ay px.py
3 CUMULATIVE-DISTRIBUTION-FUNCTION
TECHNIQUE
3.1 Description of Technique

If the joint distribution of random variables X h . . . , XII is given, then, theoreti-
caJIy, the joint distribution of random variables of Yt , ... , Yk can be determined,
where Yj=giXt, ... , XII),j= 1, ... , k for given functionsgt(', .. " ,), ... ,
182 DISTRIBUTIONS OF FUNcn:ONS OF RANDOM VARIABLES V
gk( " .,., . ). By definition, the joint cumulative distribution function of

Y1 , . , . , Yk is FYi' ... , Yk(Yl' ... , Yk) = P[ Y1 Yl; ... ; Yk < yd. But for each
Yl' "" Yk the event {Y1 <Yl; ... ; Yk <Yk} {Ul(X1 , ... , X,,) <Yl; ... ;
Uk(Xl , ••. , X,,) < Yk}' This latter event is an event described in terms of the
given functions Ul ( " ... , .), ... , gk(', •.. , .) and the given random variables
Xl' ... , X". Since the joint distribution of Xl' ... , X" is assumed given, presum-
ably the probability of event {Ul (Xl' ... , X,,) Yl;' .. ; Uk( Xl, ... , X,,) < Yk}
can be calculated and consequently Fyl. ... , Yk(', ... , .) determined. The above
described technique for deriving the joint distribution of Yl , ... , Yk will be
called the cumulative-distrihution-/unction technique.
An important special case arises if k I; then there is only one function,
say g( Xl' ... , X,,), of the given random variables for which One needs to derive
the distribution.
EXAMPLE 2 Let there be only one given random variable, say X, which has a
standard normal distribution. Suppose the distribution of Y = g(X) = X2
is desired.
FI'(Y)
= pry < y] = P[X 2 < y] P[ -JY ~ X <JY] = tP(Jy) - tP( -Jy)

";y ";y 1
= 2
o
f t/J(u) du = 2
0
f J-
2n
e-
tu2
du
.2 f)/ 1 tz f)/ 1 1 tz
= J2n 0 2Jz e- dz = 0 r(!-) J2z e- dz, for y > 0,
which can be recognized as the cumulative distribution function of a

gamma distribution with parameters r = ! and A. = !. ////
Other applications of the cumulative-distribution-funct~on technique

expounded above are given in the following three subsections. i,
3.2 Distribution of Minimum and Maximum

Let Xl' ... , X" be n given random variables. Define Yl = min [Xl' ... , X,,]
and Y" = max [Xh .,., X,,]. To be certain to understand the meaning of
Yn = max [Xl' .,., X,,], recall that each Xi. is a function with domain n, the
sample space of a random experiment. For each 00 E n, Xl(oo) is some real
number. Now Y" is to be a random variable; that is, for each 00, YnCoo) is to be
3 CUMULATIVE-DISTRIBUTION-FUNCTION TECHNIQUE 183
some real number. As defined, Y,,(w) = max [X1(w), ... , X,,(w)]; that is, for
a given w, Y,,(w) is the largest of the real numbers Xl (w), ... , X,,(w).
The distributions of YI and Y" are desired. F y" (y) = P[ Yn y] =
P[X I < Y; .•. ; X" < y] since the largest of the X,'s is less than or equal to y if
and only if all the X,'s are less than or equal to y. Now, if the X;'s are assumed
independent, then
" n
P[X I <y; ... ; Xn<Y]= [lP[Xt<y]= [lFxlY);
i= 1 I I
so the distribution of Yn = max [X I ' ... , Xn] can be expressed in terms of the
marginal distributions of Xl' ... , X n • If in addition it is assumed that all the
Xl' ... , X" have the same cumulative distribution, say F x('), then
" FX,(y) = [Fx(y)]n.

[I
1= 1
We have proved Theorem 5.
...,Theorem 5 If X I' ... , Xn are independent random variables and Yn =

max [Xl' ... , X,,], then
Fy"(Y) " Fx.(Y)·

= [I (18)
1= I
If Xl' ... , Xn are independent and identically distributed with common

cumulative distribution function Fx( .), then
Fy"(y) = [Fx(Y)]". (19)

1I11
Corollary If X., ... , X" are independent identically distributed con-
tinuous random variables with common probability density function/xC .)
and cumulative distribution function Fx( .), then
PROOF
Similarly,
P[YI~y]=1 P[Y1 > y] = I - P[XI > y; ,. .. ,. Xn> y]
184 DISTRlBUfiONS OF FUNCTIONS OF RANDOM VARIABLES V
since. YI is greater than y if and only if every Xi > y. And if Xl, ... , X" are
independent, then
1 - P[X I > Y; .•. ; X" > y] n

i= I
n
= 1 - " P[X i > Y] = 1 - " [1 - F xlY)].
i= 1
If further it is assumed that Xl, ..• , X" are identically distributed with common
cumulative distribution function ..F x ( .), then
I - n" [] - Fxj(Y)] = J -
1= I
[J - Fx(Y)]",
and we have proved Theorem 6.
Theorem 6 If Xl, ... , X" are independent random variables and Y1 =

min [Xl' ... , X,,], then
. (2])
And if X I, ..• , X" are independent and identically distributed with com-
mon cumulative distribution function Fx( .), then
Fy1(Y) 1 - [I - Fx(Y)]". (22)

IIII
Corollary If Xl, ••. , X" are independent identically distributed con-
tinuous random variables with common probability density ix(') and
cumulative distribution F x ('), then
PROOF
d
-Fy1 (Y) = n[1 II II
dy
EXAMPLE 3 Suppose that the life of a: certain light bulb is exponentially
distributed with mean 100 hours. If 10 such light bulbs are installed
simultaneously, what is the distribution of the life of the light bulb that
fails first, and what is its expected life? Let X i denote the life of the ith
light bulb; then Y 1 = min [Xl' .•• , X IO ] is the life of the light bulb that
fails first. Assume that the X i'S are independent.
3 CUMULATiVE-DISTRIBUTION-FUNCTION TECHNIQUE t 85
Now fx'(x) = 1 ~Oe-.tfoXI(O. oo)(x), and

Fx,(x) (1 e-1~oX)I(o, oo)(X);
so
!Yl(Y) = JO(e-l~o")10-1(ntoe-1~o")I(o. oolY)
= /o~-looo)ll(o. (0)(Y),
which is an exponential distribution with parameter A. = -10; hence 8[Y1 ] =

1/,1. = 10. 1111
3.3 Distribution of Sum and Difference of Two Random Variables
Theorem 7 Let X and Y be jointly distributed continuous random

variables with density fx. y(x, y), and let Z = X + Y and V = X Y.
.Then,
00
Iz(z) = f-
oo
Ix. y(x, z - x) dx = f_ Ix. y(z -
00
00 y, y) dy, (24)
and
00
Iy(v) = fix. y(x, x

-00
v) dx = foo Ix. y(v + y, y) dy.
-00
(25)
•
PROOFWe will prove only the first part of Eq. (24); the others are
proved in an analogous manner.
Fz(z) = P[Z < z] = P[X -f:. Y z] = II Ix. y(x, y) dx dy

= C, [()x, y(x, d dx y) y ]
= ( , [f!x,y(X, u- x) dU] dx
by making the substitution y u- x.
Now
!iZ)=d;;Z) = ~ {(o [f!x,y(X,U-X)dX] dU}

= foo Ix. y(x, z -
-00
x) dx. 1/1/
186 DISTRIBUTIONS OF FUNCTIONS OF RANPOM VARIABLES V
Coronary If X and Yare independent continuous random variables

and Z = X + Y, then
J_!Y(Z
00 00
fz1.z) =fx+y(z) = x)fx(x) dx = r

~-oo
fx(z - y)fy(y) dy. (26)
PROOF Equation (26) follows immediately from independence and

Eq. (24); however, we will give a direct proof using a conditional distribu-
tion formula. [See Eq. (II) of Chap. IV.]
00
P[Z <z] =P[X + Y < z] J -00

P[X + Y <zlX = x]fx(x)dx
00
= J P[x + Y z]fx(x) dx
-00
,00
= J -00 Fy(z x)fx(x) dx.

Hence,
fz(z) = dF z(z) = -d [J"'OO Fy(z x)fx(x) dx ]
dz dz 00
= J"'OO dF y(z - x) fx(x) dx

00 dz
= J~!Y(Z - x)fx(x) dx. //1/
Remark The formula given in Eq. (26) is often called the convolution
formula. In mathematical analysis, the function fz(') is caned the
convolution of the functions fy( .) and f x( . ). IIII
EXAMPLE 4 Suppose that X and Yare independent and identically dis-

tributed with density fx(x) = fy(x) = 1(0, 1)(x). Note that since both X
and Yassume values between 0 and I, Z = X + Y assumes values between
o and 2.
fz(z) J~!y(z - x)fx(x) dx = f~oo 1(0, l)(Z - x)/(o, l)(x) dx
foo
-00
{I (0, z)( x)/(O, 1)(z) + I (z-l, 1)(x)I[1, 2)(Z)} dx
:z; 1
/(0, 1)(Z) Jo dx + 1[1, 2)(Z) {-1 dx

= zl(o, l)(Z) + (2 - Z)/[l, 2)(Z). 1I11
3 CUMULATlVE-DISTRIBUTlON-FUNCTlON TECHNIQUE 187
FIGURE 1
3.4 Distribution of Product and Quotient
Theorem 8 Let X and Y be jointly distributed continuous random vari-

ables with density Ix, y(x, y), and let Z = XY and U = xl Y; then
iz(z) = roo Th Ix. Y( x, ~) dx = roo ~ Ix. y(~, y) dy, (27)
and
lu(u) = I
-00
00
Iyl/x. y(UY, y) dy. (28)
PROOF Again, only the first part of Eq. (27) will be proved. (See
Fig. I for z > 0.)
Fz(z) = P[Z < z] = II Ix. y(x, y) dx dy

xy:s;Z
= f: 00
[fOOfx. y(x, y) d Y] dx + foo [IZ,xlx. y(x, y) d Y] dx,
x/x 0 - 00
which on making the substitution u = xy
[J-oo/x,y(x,~) dU] dx + roo [f fx.Y(x,~) dU] dx

Z
= fO
-00 Z x x "0 -00 x x
= (ro [(, IJx,y(x,~) dX] du + (ro [( ~fx,+,~) dX] du
= L. [fro I~I fx,y(x,~) dX] du;
hence
liz) = dFz(z)
dz
= C, ~ fx,.( x,~) dx. IIII
EXAMPLE 5 Suppose X and Yare independent random variables, each

uniformly distributed over the interval (0, I). Let Z = X Y and U = X I Y.
Iz(z) = roo _111 Ix. y(x, ~) dx

"'-00 x x
00
lrAu) = r
~ - 00
IY IIx. y(uy, y) dy
(see Fig. 2)
= f~ Iy I{/(O, o(u)1 (o.l)(Y) + 1[1, oolu)/(o. '/llly)} dy

00
= 1(0, l)(u) fo y dy + 1[1. OO)(u) f

1
0
1{1l
y dy
1
=2 1(0,
1(1)2
l)(u) + 2 ;; 1[1,oo)(u).
Note that rS[XI Y] = 8[U] = 1 J~ u du + 1- J~(llu) du = 00, quite dif-

ferent from 8[X]18[ Y] = I. IIII
4 MOMENT-GENERATING-FUNCTION TECHNIQUE 189
~=--_U
FIGURE 2 1
4 MOMENT-GENERATING-FUNCTION TECHNIQUE
4.1 Description of Technique

There is another method of determining the distribution of functions of random
variables which we shall find to be particularly useful in certain instances. This
method is built around the concept of the moment generating function and will
be called the moment-generating-function technique.
The statement of the problem remains the same. For given random vari-
ables X I ' •.. , X,. with given density fXI' ... , x,,(Xt, ... , X,.) and given functions
gd " ... , .), ... , gk( . , .. " .), find the joint distribution of Y1 = g1 (X h ... , X,.),
... , Yk = gk(X" •.. , X,.), Now the joint moment generating function of
Y 1 , ••• , Y k , if it exists, is
m Yl • .... Yk(t...
l'
t) = @[elYl+···+tkYk]
'k
Ix!, ... , X,,(X t , •.• , X,.) n dXi'

n
X (29)
i= 1
If after the integration of Eq. (29) is performed, the resulting function of
t 1 , ••• , tk can be recognized as the joint moment generating function of some
known joint distribution, it will follow that Y1, ''', Yk has that joint distribu-
tion by virtue of the fact that a moment generating function, when it exists, is
unique and uniquely determines its distribution function.
For k > 1, this method will be of limited use to us because we can recog-
nize only a few joint moment generating functions. For k = 1, the moment
generating function is a function of a single argument, and we should have a
better chance of recognizing the resulting moment generating function.
190 DISTRIBUTIONS OF FUNCITONS OF RANDOM VARIABLES v
This method is quite powerful in connection with certain techniques of

advanced mathematics (the theory of transforms) which, in many instances,
enable one to determine the distribution associated with the derived moment
generating function.
The most useful application of the moment-generating-function technique
will be given in Subsec. 4.2. There it will be used to find the distribution of
sums of independent random variables.
EXAMPLE 6 Suppose X has a normal distribution with mean 0 and variance I.

Let Y = X 2 , and find the distribution of Y.
1
for t< 2.'
which we recognize as the moment generating function of a gamma with
parameters r = t and ,t = t. (It is also called a chi-square distribution
with one degree of freedom. See Subsec. 4.3 of Chap. VI.) 1///
EXAMPLE 7 Let Xl and X 2 be two independent standard normal random

variables. Let Y1 = Ul(X 1 , X 2) = Xl + X 2 and Y2 = U2(X h X 2 ) =
X 2 - Xl' Find the joint distribution of Y1 and Y 2 •
m
Yl.Y2
(t l' t)
2
= .!.P[e Y 1t1 + Y2t2]
(0
= ~"'[e(X + X2)tt + (X2 - x d 2]

1 f
= 8[eX1(t1 -(2)+X2(t1 +t2)]

= cC[eXt(ft-t2)]cC[eX2(tl +( 2 )]
= mXl(t 1 - t 2)mxi t l + t 2)
(11 - t2)2 (11 + t2)2
= exp 2 exp 2
2ti 2ti
= exp(t; + t~) = exp 2 exp 2:
= mY1(tl)mY2(t2)'
We note that YI and Y2 are independent random variables (by Theorem 10

of Chap. IV) and each has a normal distribution with mean 0 and variance
2. ., II1I
In the above example we were able to manipulate expectations and avoid
performing an integration to find the desired joint moment generating function.
In the following example the integration will have to be performed.

variables. Let Y = (X 2 - X I )2/2, and find the distribution of Y.
my(/) = of[exp YI] = of [exp (X 2 - ; X I )2 t]
= f OO foo 1
-2 exp
[(X2 - XI)2
2 t-
xi +2 X~] dXI dX2
-00 -00 TC
= roo roo 2~ exp{- ~ [x:(1 - I) + 2X,X2 I + x~(1 - t)]} dx, dX2

= foo ~ exp [- ~ x~(1- t)]
- 00 2TC 2
x {f_ OO
00
[1 - t (2 +
exp - -2- Xl
2XtX2
1_ t
t)] dXI }dX 2
1 [X~(1 - t)] exp X2 2 t

= f OO
-00
~
~2TC
exp -
2
---....:2::...-_
2(1 - t)
1
x J1 _ t {roo JJ~ exp [ _1 2 I (x, + t2/y] dx,) dX2
1
= J1 _ I i
roo exp [ - ~ (1 - I - 1,2 }i] dX2
_ 1 . J 1- t Jt=U 1 foo ( 1 1 - 2t )
- Ji - t Jl - 21 . J 1 - t J2n -00 exp -:2 1 _ t x~ dX2
= (1- 2/)-! = (! ! I)!' for t < 1/2,

which is the moment generating function of a gamma distribution with
parameters r = ! and A. = ~.; hence,
4.2 Distribution of Sums of Independent Random Variables

In this subsection we employ the moment-generating-function technique to find
the distribution of the sum of independent random variables.
Theorem 9 If X 1 ••••• X are independent random variables and the

II
moment generating function of each exists for all -h < t < h for some
,.
h > 0, let Y = I Xi; then
1
,.
my(t) = t9'[ exp LXi t] = 0 mx.(t) for -h<t<h.
j= 1
PROOF
II II
=0 G[ell =0 mxlt)
1= l j= 1
using Theorem 9 of Chap. IV. 1/1/

The power and utility of Theorem 9 becomes apparent if we recall Theorem
7 of Chap. II~ which says that a moment generating function, when it exists,
n
determines the distribution function. Thus, if we can recognize 0 mxlt) as
i"" 1
the moment generating function corresponding to a particular distribution,
,.
then we have found the distribution of I Xi' In the following examples, we
1
will be able to do just that.
EXAMPLE 9 Suppose that Xl, ... , X II are independent Bernoulli random

variables; that is, P[X i = I] = p, and P[X f = 0] = I - p. Now
mxlt) = pe + q ..
II
So
mr xlt) = .0 mxlt) = (pe' + q)lI,
1=1
the moment generating function of a binomial random variable; hence

n
LXi has a binomial distribution with parameters nand p. lIlt
1
EXAMPLE 10 Suppose that Xl, ... , XII are independent Poisson distributed
random variables, Xi having parameter A. i • Then
mx.(t) = C[e tXi ] = exp Ale' - 1),
and
, hence
mr. xlt) = nmxlt) nexp Ai(e

II
1=1
=
II
i=l
t
- 1) = exp L Ai(e t
- 1),
which is again the moment generating function of a Poisson distributed

random variable having param,eter L Ai . So the distribution of a sum of
independent Poisson distributed random variables is again a Poisson
distributed random variable with a parameter equal to the sum of the
individual parameters. I111
EXAMPLE 11 Assume that Xl, ... , XII are independent and identically dis-
tributed exponential random variables; then
So
mI:x,(t) = ,n
II
mxlt) =
( A )11
A. - t '
which is the moment generating function of a gamma distribution with

parameters n and A; hence,
An
f I: x.() -
X - r(n) X
n-l-lxj
e
()
(0, (0) X ,
the density of a gamma distribution with parameters n and A.. 1II1
EXAMPLE 12 Assume that Xl' ... , Xn are independent random variables

and
then
and
Hence
n
m'f,o;x.(t) = nmOtxlt)
i=l
= exp[(I aif,J.i)t + t(I afuf)t 2] ,
which is the moment generating function of a normal random variable; so
The above says that any linear combination (that is, L ai Xi) of inde-
pendent normal random variables is itself a normally distributed random
variable. (Actually, any linear combination of jointly normally distri-
buted random variables is normally distributed. Independence is not
required.) In particular, if
x '" N(px, oJ),
and X and Yare independent, then
X + Y", N(Jlx + Jly, uk + u~),
and
x- Y '" N(p.x - Jly, u; + u~).
If Xl' ... , Xn are independent and identically distributed random vari-
ables distributed N(Jl, ( 2 ), then
X.=!I: X I - N (I'. :2);

that is!, the sample mean has a (not approximate) normal distribution. /1/1
In the above examples we found the exact distribution of the sums of

certain independent random variables. Other examples, including the impor-
tant result that the sum of independent identically distributed geometric random
variables has a negative binomial distribution, are given in the Problems. One
n
is often more interested in the average, that is, (lIn) LXi' than in the sum.
1
Note, however, that if the distribution of the sum is known, then the distribution
of the average is readily derivable since
F(1Jn)'£X,(z) = p[;; I: X, < z] = P[L XI < nz] = Fr,x,(nz). (30)
In Examples 9 to 12 above, where we derived the distribution of a sum, we have

in essence also derived the distribution of the corresponding average. One of
the most important theorems of all probability theory, thecentral-limittheorem,

gives an approximate distribution of an average. We will state this theorem
next and then again in our discussion of sampling in Chap. VI, where we will
outline its proof.
Theorem 10 Central-limit theorem If for each positive integer n,

Xl, ... , X,. are independent and identically distributed random variables
with mean Jlx and variance ui-, then for each z
FzJz) converges to CI>(z) as n approaches 00, (31)
where
We have made use ofEq. (9), which stated that ~[XII] = Jlx and var [XII] =
uiln. Equation (31) states that for each fixed argument z the value of the
cumulative distribution function of ZII' for n = I, 2, ... , converges to the value
CI>(z). [Recall that CI>(.) is the cumulative distribution function of the standard
normal distribution.]
Note what the central-limit theorem says: If you have independent random
variables X I, ... , XII , ... , each with the same distribution which has a mean and
variance, then XII = (lIn) LXi" standardized" by subtracting its mean and
then dividing by its standard deviation has a distribution that approaches a
standard normal distribution. The key thing to note is that it does not make
any difference what common distribution the Xl' ... , XII' ... have, as long as
they have a mean and variance. A number of useful approximations can be
garnered from the central~limit theorem, and they are listed as a corollary.
Corollary If Xl' ... , XII are independent and identically distributed

random variables with common mean Jlx and variance ui, then
p [a < X.u /
x
-;X
n
< b] ~ (b) - CI>(a) , (32)
P[~ < XII < d] ~ CI> (~- Jl~) - CI>(C - Jlx) , (33)
uxlJn ux/Jn
or
IIII
Equations (32) to (34) give approximate values for the probabilities of

certain events described in terms of averages or sums. The practical utility
of the central-limit theorem is inherent in these approximations.
At this stage we can conveniently discuss and contrast two terms that are a
vital part of a statistician's vocabulary. These two terms are limiting distribution
and asymptotic distribution. A distribution is called a limiting distribution
function if it is the limit distribution function of a sequence of distribution
functions. Equation (31) provides us with an example; <J>(z) is the limiting
distribution function of the sequence of distribution functions F zt ('), Fzi'), ... ,
F zn{'), .••• Also <J>(z) is called the limiting distribution of the sequence of random
variables Z1, Z 2, •.. , Z,., .. . . On the other hand, an asymptotic distribution
of a random variable, say Y,., in a sequence of random variables Y 1 , Y 2 , ••• Y,., ...
is any distribution that is approximately equal to the actual distribution of Y,.
for large n. As an example [see Eq. (33)], we say that X,. has an asymptotic
distribution that is a normal distribution with mean J-lx and variance ui/n. Note
that an asymptotic distribution may depend on n whereas a limiting distribution
does not (for a limiting distribution the dependence on n was removed in taking
the limit). Yet the two terms are closely related since it was precisely the fact
that the sequence ZI' Z2, ... , Z,., ... had limiting standard normal distribu-
tion that allowed us to say that X,. had an asymptotic normal distribution with mean
J.lx and variance ui/n. The idea is that if the distribution of Z,. is converging
to <J>(z), then for large n the distribution of Z,. must be approximately distributed
N(O, I). But if Z,. = (X,. ~ J-lx)/(ux/Jn) is approximately distributed N(O, I),
then X,. is approximately distributed N(;tx, ui/n).
In concluding this section we give two further examples concerning sums.
The first shows how expressing one random variable as a sum of other simpler
random variables is often a useful ploy. The second shows how the distribution
of a sum can be obtained even though the number of terms in the sum is also a
random variable, something that occasionally occurs in practice.
EXAMPLE 13 Consider n repeated independent trials, each of which has

possible outcomes {) 1, ••• , Ok + l ' Let Pj denote the probability of outcome
o.J on a particular trial, and let X·J denote the number of the n trials resulting
in outcome {}j' j = 1, ... , k + l. We saw that (XH ... , X k ) had a multi-
nomial distribution. N ow let
if ath trial results in outcome {}j
otherwise;
4 MO~NT-GENERATING-FUNCTION TECHNIQUE 197
n
then Xj = I Zjl%' Now suppose we want to find coY [Xi' Xj ], In-
a=l
tuitively, we might suspect that such covariance is negative since when one
of the random variables is large another tends to be small.
by Theorem 2. Now if IX i" p, then ZiP and Zja are independent since they
correspond to different trials, which are independent. Hence
n ,. n
I I
p=la=l
COY [Zip, Z ja] = I
a=l
COY [Zil%' Z jl%]'
But COY [Zia, Zja] = C[ZiaZja] C[Zia]C[Zja], and C[ZiI%Zja] = 0 since

at least one of Zia and Zja must be O. Now C[Zia] = Pi' and C[Zja] = Pj;
socov[Xi,Xj]=-nPiPj' IIII
EXAMPLE 14 Let Xl' ... , Xn , ... be a sequence of independent and identi-

cally distributed random variables with mean Jlx and variance uj. Let N
N
be an integer-valued random variable, and define SN = L Xi; that is, SN
i= 1
is the sum of the first N X/s, where N is a random variable as are the
X/so Thus SN is a sum of a random number of random variables. Let us
assume that N is independent of the X/so Then C[SN] = C[C[SNIN]]
by Eq. (26) of Chap. IV. But C[SNI N = n] = C[Xl + ... + Xn] = nJlx;
so 8[SNIN] =NJlx, and C[C[SNI N]] = C[NJlx] = JlxC[N] = JlNJlX' Sim-
ilarly, using Theorem 7 of Chap. IV,
var [SN] = C[var [SN IN]] + var [C[SN IN]]

= C[Nui] + var [NJlx]
= uiC[N] + Jli var [N]
= JlN . ui + u~ . Jli.
Suppose now that N has a geometric distribution [see Eq. (14) of Chap. III]
with parameter p, Xi has an exponential distribution with parameter 1,
and we are interested in the distribution of S N • Further assume inde-
pendence of N and the X/so Now, for z> 0,
ao
P[SN < z] = L P[SN < zlN = n]P[N = n]
n=1
(by using the fact that a sum of independent and identically distributed
exponential random variables has a gamma distribution)
= f ).pe-
z
o
lU e(l- p )lll du = ).p f e-
z
0
lpu du =1- e- lpz•
That is, SN has an exponential distribution with parameter pl. Recall

that [see Eq. (14) of Chap. ITI] $[N] = lip and var [N] = (I p)lp2; also,
$[X] = II)., and var [Xl = IJ).2. SO, as a check of the formulas for the
mean and variance derived above, note that
1 1 1
IlN Ilx = - . -
p ). p).
and
2 22 11 1-p1 1
IlN(JX + (JNllx = p).2 + 7 ).2 = (p).) 2 ,
which are the mean and variance, respectively, of an exponential distribu-
tion with parameter pl. JIJI
5 THE TRANSFORMATION Y=g(X)

The last of our three techniques for finding the distribution of functions of
given random variables is the transformation technique. It is discussed in this
section for the special case of finding the distribution of a function of a uni-
dimensional random variable. That is, for a given random variable X we seek
the distribution of Y = g(X) for some function g( . ). Discussion of the general
case is deferred until Sec. 6 below. Both the notation Y = g(X) and the notation
y = g(x) will appear in the ensuing paragraphs; y = g(x) is the usual notation
for the function or transformation specified by g('), and Y = g(X) defines the
random variable Y as the function g(') of the random variable X.
5.1 Distribution of Y = g(X)

A random variable X may be transformed by some function g(.) to define a
new random variable Y. The density of Y, fy(y), will be determined by the
transformation g(.) together with the density fx(x) of X.
5 THE TRANSFORMATION Y g(X) 199
First, if X is a discrete random variable with mass points Xl' X 2, ••• , .

then the distribution of Y = g(X) is determined directly by the laws of prob-
ability. If X takes on the values Xl' X 2 ' ••• with probabilities/x(xl ),/x(x2)' ... ,
then the possible values of Yare determined by substituting the successive
values of X in g('). It may be that several values of X give rise to the same
value of y, The probability that Y takes on a given value, say Yj' is
/y(y) = L
(i!g(xt)= YJ)
fx(Xi)' (35)
EXAMPLE 15 Suppose X takes on the values 0, I, 2, 3, 4, 5 with prob-

abilities/x(0),/x(l),/x(2),/x(3),/x(4), and/x(5). If Y = g(X) = (X - 2)2,
note that Y can take on values 0, I, 4, and 9; then /y(O) = /x(2),
/y(I) = /x(l) + /x(3),/y(4) = /x(O) + /x(4) , and/y(9) = fx(5). IIII
Second, if X is a continuous random variable, then the cumulative dis-

tribution function of Y = g(X) can be found by integrating/x(x) over the appro-
priate region; that is,
Fy(Y) = P[Y:::;;; y] P[g(X) < y] = f

{x:g(X)5)'}
/x(x) dx. (36)
This is just the cumulative-distribution-function technique.
EXAMPLE 16 Let X be a random variable with uniform distribution over the

interval (0, 1) and let Y = g(X) = X 2 , The density of Y is desired.
Now
f f )' dx = Jy
..;~
Fy(Y) = pry < y] =

2
P[X < y] = fx(X) dx =
(x: x 2 s)'} 0
for °< y < I; so
and therefore
I I
/Y(Y)=
2Jy I(o.1)(Y)' IIII
Application of the cumulative-distribution-function technique to find the
density of Y = g(X), as in the above example, produces the transformation
technique, the result of which is given in the following theorem. 1
Theorem 11 Suppose X is a continuous random variable with prob-

ability density function f x(·). Set X = {x: f x(x) > O}. Assume that:
0) y = g(x) defines a one-to-one transformation of X onto 'D.
(ii) The derivative of x = g-l(y) with respect to y is continuous and
nonzero for y E 'D, where g-l(y) is the inverse function of g(x); that is,
g-l(y) is that x for which g(x) = y.
Then Y = g(X) is a continuous random variable with density
PROOF The above is a standard theorem from calculus on the change

of variable in a definite integral; so we wil1 only sketch the proof. Consider
the case when I is an interval. Let us suppose that g(x) is a monotone
increasing function over X; that is, g'(x) > 0, which is true if and only if
(dldy)g-l(y) > 0 Qver 'D. For y E 'D, Fy(y) = P[g(X) < y] = P[X ~
g-l(y)] =FX(g-l(y)), and hence fy(y) = (dldy)Fy(y) = [(dldy)g-l(y)]
fX(g-l(y» by chain rule of differentiation. On the other hand, if g(x)
is a monotone decreasing function over I, so that g'(x) < 0 and
(dldy)g-l(y) < 0, then Fy(y) = P[g(X) < yl = P[X > g-l(y)] = I - Fx
(g-l(y»,andthereforefy(y) = - [(dldy)g-l(y)]fx(g-l(y» = I(dldy)g-\y) I
fX(g-l(y» for y E 'D. IIII
EXAMPLE 17 Suppose X has a beta distribution. What is the distribution of

Y = -IogeX? I = {x:fx(x) > O} = {x:O < x < I}. y =g(x) = ~Iogex
defines a one-to-one transformation of X onto 'D = {y: y > O}. x =
g-l(y) = e-"', so (dldy)g-\y) = -e- Y, which is continuous and nonzero
for y E ID. By Theorem II,
d
fy(y) = dyg-l(y) fX(g-l(y»I!}(y)
= - Y I ( Y)Q -1 (I _ e - Y\b - 11 ()
e B(a, b) e ) (0, (0) Y
I
- B(a, b) e- aY(I - e Y)b 11(0, OO)(y),
In particular, if b = I, then B(a, b) = I/a; so fy(y) = ae-ay/(O, oo)(Y), an

exponential distribution with parameter a. IIII
5 THE TRANSFORMATION Y=g(X) 201
EXAMPLE 18 Suppose X has the Pareto density/x(x) = (}X- 9 - 1[[1. oo)(x) and
the distribution of Y = loge X is desired.
IY(Y) ~ ~ g-l(y) fx(g-l(y»l'f)(Y)

9y
= eY(}(e'j-9-1 1[1. OO)(e Y ) = (}e- [[o. OO)(Y)· IIII
The condition that g(x) be a one-to-one transformation of X onto ~ is
unnecessarily restrictive. For the transformation y = g(x), each point in X
will correspond to just one point in ~ ; but to a point in ~ there may correspond
more than one point in X, which says that the transformation is not one-to-one,
and consequently Theorem II is not directly applicable. If, however, X can be
decomposed into a finite (or even countable) number of disjoint sets, say Xl, ... ,
Xm , so that y = g(x) defines a one-to-one transformation of each Xi into ~,
then the joint density of Y = g(X) can be found. Let x = gi l(y) denote the
inverse of y = g(x) for x E Xi. Then the density of Y = g(X) is given by
where the summation is over those values of i for which g(x) = y for some value
of x in Xi.
EXAMPLE 19 Let X be a continuous random variable with density /x(·),

and let Y = g(X) = X2. Note that if X is an interval containing both
negative and positive points, then y = g(x) = x 2 is not one-to-one. How-
ever, if X is decomposed into Xl = {x: x E X, X < O} and X2 = {x: x EX,
X >O}, then y = g(x) d~nes a one-to-one transformation on each Xi.
Note that g11(y) = - J

y and g2 l(y) = JY.
By Eq. (37),
In particular, if
/x(x) = H)e- 1x1 ,
then
r 1 I -'Or
lY(Y) = 2 JY e Y[(o.ooly);
or, if
then
IIII
5.2 Probability Integral Transform

If X is a random variable with cumulative distribution F x ('), then F x (') is a
candidate for g(') in the transformation Y = g(X). The following theorem
gives the distribution of Y = Fx(X) if F x (') is continuous. Since F x (') is a
nondecreasing function, the inverse function Fi \.) may be defined for any
°
value of y between and I as; Fi ley) is the smallest x satisfying Fx(x) :2::: y.
Theorem 12 If X is a random variable with continuous cumulative dis-

tribution function Fx(x), then U = FxCX) is uniformly distributed Over the
interval (0, I). Conversely, if U is uniformly distributed over the interval
(0, I), then X = Fi l(U) has cumulative distribution function Fx (·).
PROOF P[U < u] = P[Fx(X) u] = P[X < Fi 1 (u)] = Fx(F-x 1 (u» =

u for 0 < u < I. Conversely, P[X x] = P[Fil(U) < x] = P[U < Fx(x)]
= Fx(x). IIII
In various statistical applications, particularly in simulation studies, it is

often desired to generate values of some random variable X. To generate a value
of a random variable X having continuous cumulative distribution function
F x ('), it suffices to generate a value of a random variable U that is uniformly dis-
tributed over the interval (0, I). This fonows from Theorem 12 since if U is a
random variable with a uniform distribution over the interval (0, I), then X =
Fi 1 (U) is a random variable having distribution F x( . ). So to get a value,
say x, of a random variable X, obtain a value, say u, of a random variable U,
compute Fx l( u), and set it equal to x. A value u ofa random variable U is called
a random number. Many computer-oriented random-number generators are
available.
6 TRANSFORMATIONS 203
EXAMPLE 20 Fx(x) = (I - e- lx )l(o, oo)(x), Fx 1(y) = -(l/A.) loge (I - y); so

-O/l) loge (l - U) is a random variable having distribution (1 - e-;'X)
1(0, OO)(X) if U is a random variable uniformly distributed over the interval
(0, I). / // /
The transformation Y = Fx(X) is called the probability integral trans-

formation. It plays an important role in the theory of distribution-free statistics
and goodness-of-fit tests.
6 TRANSFORMATIONS
In Sec. 5 we considered the problem of obtaining the distribution of a function
of a given random variable. It is natural to consider next the problem of
obtaining the joint distribution of several random variables which are functions
of a given set of random variables.
6.1 Discrete Random Variables

Suppose that the discrete density function fXh . •. , x n (x 1 , ••• , xn) of the n-
dimensional discrete random variable (Xl' ... , Xn) is given. Let l: denote
the mass points of (Xl' ... , X n ); that is,
Suppose that the joint density of Yl = gl(Xl , ... , X n), •.. , Yk = gk(Xl , ... , Xn)
is desired. It can be observed that Yl , ••• , Yk are jointly discrete and
P[Yl =Yl; ... ; Yk = Yk] =!Yl, ... ,Yk(Yl'''',Yk) = I!xt .... ,xn(Xl , ... ,xn),where
the summation is over those (Xh •.. , xn) belonging to I for which (Yl, ... 'Yk) =
(gl(Xl' ... , XII)' ••. , gk(Xh ... , XII»'
EXAMPLE 21 Let (Xl' X 2 , X 3 ) have a jOint discrete density function given

by
(0,0,0) (0,0,1) (0,1,1) (1,0,1) (1,1,0) (1,1,1)

Find the joint density of Y1 = 91(X 1, X 2 , X 3) = Xl + X 2 + X3 and

Y 2 = 92(X1, X 2 , X 3) = IX 3 - X21·
X = {(O, 0, 0), (0, 0, 1), (0, 1, 1), (1, 0, 1), (1, 1,0), (1, 1, I)}.
!Yt. Y2(0, 0) =!X •• X2.X3(0, 0, 0) = L
!Y to Y2(1, 1) =!Xt, X2. X3(0, 0,1) = i,
!Yt,Y2(2, 0) =!XI,X2.X3(0, 1,1) = ~,
!Y to Y2(2, 1) =!x lo x2.xi 1, 0,1) + !x •• x 2.x3(1, 1,0) = i,
and
IIII
6.2 Continuous Random Variables

Suppose now that we are given the JOInt probability density function
fXI •...• Xn (X1' ... , XII) of the n-dimensional continuous random variable
(Xl' ... , XII)' Let
X = {(Xl' ... , XII) :fxt. ...• Xn (X1, ... , XII) > O}. (38)
Again assume that the joint density of the random variables Y1 = 91 (X 1, ... , XII)'
... , Yk = 9k( Xl' ... , XII) is desired, where k is some integer satisfying I <k < n.
If k < n, we will introduce additional, new random variables Yk + 1 =
9k+1(X 1, ... , X,,), ... , Y II = 911 (X 1 , ••• , XII) for judiciously selected functions
9k+ l' .•. , 911; then we will find the joint distribution of Yh •.. , YII , and finally
we will find the desired marginal distribution of Y1 , ••• , Yk from the joint dis-
tribution of Y1 , ••• , YII . This use of possibly introducing additional random
variables makes the transformation Y1 = 91(Xb ... , x,,), ... , YII = 911(X1, ... , XII)
a transformation from an n-dimensional space to an n-dimensional space.
Henceforth we will assume that we are seeking the joint distribution of Y1 =
91(X1, ... , XII)' ... , YII = 911(X1, ... , XII) (rather than the joint distribution of
Y1, ... , Yk) when we have given the joint probability density of Xl, ... , XII'
We will state our results first for n = 2 and later generalize to n > 2.
LetfX1>XiX1' X2) be given. Set X ={(Xb x 2):fxl.xi x 1' X2) > O}. We want to
find the joint distribution of Y1 = 91(X 1, X 2) and Y2 = 92(X b X 2) for known
functions 91(', .) and 92(', '). Now suppose that Y1 = 91(X1' X2) and Y2 =
92(X1' X2) defines a one-to-one transformation which maps X onto, say, ~.
Xl and X2 can be expressed in terms of Y1 and Y2; so we can write, say, Xl =
911(Y1' Y2) and X2 = 921(Yb Y2)' Note that X is a subset of the X1X2 plane and
~ is a subset of the Y1Y2 plane. The determinant
aXl aXl
aYl aY2 (39)
aX2 aX2
aYl aY2
will be called the Jacobian of the transformation and will be denoted by J.

The above discussion permits us to state Theorem 13.
Theorem 13 Let Xl and X 2 be jointly continuous random variables

with density function!xl~xixl' X2). Set X = {(Xh X2):!X1.Xz(x l , X2) > O}.
Assume that:
Yl = gl(Xlo X2) and Y2 = g2(Xl' x 2) defines a one-to-one transformation
(i)
of X onto ~.
(ii) The first partial derivatives of Xl = gll(Yh Y2) and X2 = g2"l(Yl' Y2)
are continuous over ~.
(iii) The Jacobian of the transformation is nonzero for (Yl' Y2) E ~.
Then the joint density of Yl = gl(X l , X 2) and Y2 = g2(X l , X 2) is given
by
fYI, YZ(Yl' Y2) = IJ Ifxt.xz(gll(Yb Y2)' g2"l(Yl' Y2»IIl)(Yl' Y2)· (40)
PROOF We omit the proof; it is essentially the same as the derivation

of the formulas for transforming variables in double integrals, which may
be found in many advanced calculus textbooks. ~ is that subset of the
YlY2 plane consisting of points (Yl, Y2) for which there exists a (Xl' X2) EX
such that (Yl' Y2) = (gl(Xl, X2)' g2(Xl' X2».
IIl)(Yl' >'2) = I~(gll(Yl' Y2), g2"l(Yl' Y2». IIII
EXAMPLE 22 Suppose that Xl and X 2 are independent random variables,

each uniformly distributed over the interval (0, 1). Then!xI,xz(xl, x 2 ) =
I(o.1)(x l )I(o,1)(x2)· X = {(Xl' X2): O<Xl < I and 0<X2 < I}. Let
Yl = gl(Xh X2) = Xl + X2 and Y2 = g2(Xl , X2) = X2 - Xl; then Xl =
-!(Yl - Y2) = g1 (Yl' Y2)' and x 2 = t(Yl + Y2) = g2"l(Yh Y2).
1
aXl aX
- - l -!
aYl aY2 1
J= =-
aX2 aX2 2
- -
aYl aY2
206" DISTRIBUTIONS OF FUNCTIONS OF RANDOM VARIABLES V
FIGURE 3
X and ID are sketched in Fig. 3. Note that the boundary Xl = 0 of X goes

into the boundary !(Yl - Y2) = 0 ofID, the boundary x 2 = 0 of X goes into
the boundary !(Yl + Y2) = 0 ofID, the boundary Xl = I of X goes into the
boundary t(Yl - Y2) = I of ID, and the boundary X2 = I of X goes into
the boundarY-!(Yl + Y2) = 10fID. Now the transformation is one-to-one,
the first partial derivatives of g~ 1 and g2 1 are continuous, and the Jacobian
is nonzero; so
= !1(0, t) ( Yl -2 Y2) 1(0,1) (Yl +

2 Y2)
= {~ for (Yl' Y2)EID

otherwise. IIII
•

variables. Let Yl = Xl + X 2 and Y2 = X l lx 2 • Then
-l( ) YIY2
Xl=91 Yl'Y2 =1+Y2 and X2 = 92-l( Yl' Y2 ) = 1 Yl .
+ Y2
Y2 Yl
-
1 +Y2 (1 + Y2)2 Yl(Y2 + 1) Yl
J= = - = -
(1 + Y2)2·
1 (1 + Y2)3
Yl
1 + Y2 (1 + Y2)2
To find the marginal distribution of, say, Y 2 , we must integrate out YI ;

that is
00
!r (Y2) = f_!rt. r2(vl' Y2) dYl

2
_ ~ 1 foo ex [_ ~ (1 + Y~)Yi] d
- 2n (I + Y 2)2 _ 00 IY I I p 2 (1 + Y 2)2 Y 1-
Let
then
(I + Y~)
du =
(I + Y2) 2 Yl dYt
and so
a Cauchy density. That is, the ratio of two independent standard normal
random variables has a Cauchy distribution. 1/1/
EXAMPLE 24 Let Xi have a gamma density with parameters ni and A. for

i = I, 2. Assume that Xl and X 2 are independent. Again, we seek the
jOint distribution of Y1 = Xl + X 2 and Y2 = X l /X 2 -
-l( ) YIY2 and -1( ) Yl_

Xl = 91 Ylt Y2 = 1 + Y2 X2 = 92 Yl' Y2 = 1 + Y2 '
208 DISTRIBUTIONS OF FUNCrIONS OF RANDOM VARIABLES v
hence
We see thatfYl. Y2(Yt, Y2) = f Yl(Yt)fY/Y2); so Yt and Y 2 are independent.

Also, we see that the distribution of Yt = X. + X 2 is a gamma distribu-
tion with parameters nt + n z and A. If nl = nz = I, then Y2 is the ratio
of two independent exponentially distributed random variables and has
density
a density which has an infinite mean. 1/1/
EXAMPLE 25 Let Xi have a gamma distribution with parameters ni

and A for i = l~ 2~ and assume X t and X 2 are independent. Suppose now
that the distribution of Yt = X./(X t + X 2 ) is desired. We have only
the one function Yt = gt(x., X2) = x.l(xt + X2); so we have to select the
other to use the transformation technique. Since Xl and x 2 occur in the
exponent of their joint density as their sum, x. + x 2 is a good choice.
Let Y2 t : Xl + X2; then X t = YtY2, X2 = Y2 - YIY2' and
J=I -YzYz Yt
) -Yl
I Y2 .
6 TRANSFORMAT]ONS 209
Hence
fy t • Y2(Yl, Y2)
= Y2 1 1 ,tilt +1I 2(YIY2t 1 - 1(Y2 - YIY2t 2- 1 e- AYl!ID(Yl, Y2)

r(nl) r(n 2)
1I
x [ ;"lIt+ 2 ylIl+1I2-le-AYl] (y)].
r(nl + n2 ) 2 (0,00) 2
It turns out that Yl and Y2 are independent and Y1 has a beta distribu-
tion with parameters n1 and n2 • ////
Of the three conditions that are imposed on the transformation Yl =

gl(Xb X2) and Y2 = g2(Xl, X2), the sometimes restrictive condition that the
transformation be one-to-one Can be relaxed. For the transformation Yl =
gl(Xl, X2) and Y2 = g2(X 1 , x 2), each point in X will correspond to just one point
in ~ ; but to a point in ~ there may correspond more than one point in X, which
says that the transformation is not one-to-one and consequently Theorem 13 as
stated is not applicable. If, however, X can be decomposed into a finite number
of disjoint sets, say Xl' ... , Xm, so that Yl = gl(X 1 , x 2) and Y2 = g2(Xl, X2)
define a one-to-one transformation of each Xi onto ~ then the joint density of
Y1 = gl(X 1 , X 2) and Y2 = g2(X 1 , X 2) can be found. Let Xl = g;/(Yb Y2)
and X2 = g;}(Yl' Y2) denote the inverse transformation of ~ onto Xi for j = I,
... , m, and set
a -1 a -1
-9ti- -9ti
-
aYl aY2
Ji = agu-1 ag2i
-1
-- --
aYl aY2
Theorem 14 Let Xl and X 2 be two jointly continuous random variables
with density function!xl,X2(Xl, X2)' Assume that X Can be decomposed
into sets Xl' "', Xm such that the transformation Yl = gl (Xl' X2) and
Y2 = g2(Xl, x 2) is one-to-one from Xi onto~. Let Xl = gl/(Yl' Y2) and
X2 = g2"/(Yl' Y2) denote the inverse transformation of ~ onto Xi' i =
1, ... , m. Assume that all first partial derivatives of g~/ and g:;/ are
continuous on ID and that J i does not vanish on~, i = I, ... , m. Then
210 DISTRIBUTIONS OF FUNCITONS OF RANDOM VARJABLES V
m
fYI. yiYl' YZ) = L IJilfX.,X2(gl/(Yl' Yl), U2/(Yl' YZ»I\~,(YI' Yl)'
i=l
(41)
1/11
We illustrate this theorem with Example 26.
EXAMPLE 26 Assume that X I and X z are independent standard normal

random variables. Consider the transformation Yl = xi + and xi
Y 2 = Xl' which implies Xl = + J Yl - y~ and Xz = Yl so that the trans-
formation is not one-to-one. Here X = {(Xl' Xz): - 00 < Xl < 00,
- 00 < Xl < oo}, and ID ={(YH Yl): 0 < Yl < 00, -,}Y. < Y2 < J YI}'
If X is decomposed into Xl and Xz, where Xl = {(Xl' X2): 0 Xl < 00,
-00 < X2 < oo} and Xl = {(Xl' Xl): -00 < Xl < 0, -00 < X2 < oo} (in
the terminology of Theorem 14, m = 2), then our transformation is one-
to-one for Xi onto ID, i = 1,2. gll(YI' Y2) = J Yl - yi, andg 2l(Yl' Yl) =
Y2; gI21 (y., Yl) = J Yl - yi, and g2i(Yh Y2) = Yz; so
1 iJull
yn- t ,
.1.
f(Yl - Yl)-~
J1 = iJY2 = t(Yl -
o I
and
= - -1 ( Yl - Yl2)-t •
2
o
Hence,
fyhyiYh Y2) = [IJllfxl. x2(uIl(Yh Y2), UZ/(Yh Yl»
+ IJzlfxt.x/uli(Yl' Y2)' u2i(Yl' Y2»]IID(Yh Y2)
1 1 _tyl
- . -. e
JYI - y~ 2n
for YI 0 and -JYl < Y2 < JYl' Now
= ~ e- t ),lJarc sin Yl ,,/1.)

2n \ JYl -../)'1
= ;" e-ty.(~ +~) = ~ e- b • for Y, > 0,

an exponential distribution. /111
Theorems 13 and 14 can be generalized from n = 2 to n > 2. We wilI

state the generalization of Theorem 14. (Theorem 13 is a special case of Theorem
14.)
Theorem 15 Let Xl' X 2 , ••• , X rJ be jointly continuous random variables

with density function IXh .... Xn(Xt' ••• , xrJ)' Let X = {(xv· .. , xrJ):
Ix •..... x(x t "" xrJ) > O}. Assume that X can be decomposed into sets
Xl' .•. , Xm such that .
Yl gl(Xt, .'" x,,), Y2 =giXl' , .. , x,,), ... , Y" =
grJ(Xl, ... , xrJ) is a one-ta-one transformation of Xi onto ID, i = I, ... , m.
.
Let Xl = g1/(Yb ... , y,,), ... , XrJ = g;' 1 (Yl' ...• Y,,) denote the inverse
transformaTIOn of ID onto Xi' i = I, "', m. Define
1 o -1
°9ti
- - og1/
-OY2 ... ... -9ti-
OYI oY"
o9u-1
ogu t 1
- - - - ... . .. -
0921
-
Ji = oYt OY2 oY"
••••• to • '" ................. to ••••
~ -1
" . . . .. U9",
oYrJ
for i = 1, .. " m.
Assume that all the partial derivatives in J i are continuous over
ID and the determinant J i is nonzero, i = 1, ... , m. Then
m
= L IJdfxl ..... x (9i/(Yt, ... , Y,,), ... , 9rJi \Yl' ... , Yn»
i=1
n (42)
for (Yl, ... , Y,,) in ID· IIII
EXAMPLE 27 Let Xl, X 2 , and X 3 be independent standard normal random

variables, Yl = Xl' Y2 = (Xl + x2)/2, and Y3 = (Xl + X2 + x3)/3. Then
Xl = YI, X2 = 2Y2 - Yl' and x3 = 3Y3 - 2Y2; so the transformation is
one-to-one. (m = I in Theorem 15.)
1 0 o
J= -1 2 o = 6.
o -2 3
-
. 212 DISfRIBUTIONS OF FUNCTIONS OF RANPOM VARIABLES V
fYI. Y2, Y3(Yl' Y2' Y3)

= IJlfXt.x2,X3(X 1 , X2, X3)
= 6(/~;;) 3 exp{ -l[yi + (2y, - y,)' + (3Y3 - 2y,)'])
= 6()2J exp[ -!(2yi - 4y,y, + 8y~ -

3 12Y'Y3 + 9y~)].
The marginal distributions can be obtained from the joint distribution;

for instance,
f f fYI, Y2. Y3(Ytt Y2' Y3) dYl dY2

00 00
=
-00 -00
= 6()2J rooexp[ -!(6y~ - 12y, Y3 + 9y~)]

3
x (rooeXP [-!(2yi - 4y,y, + 2yi y ,) dy, )]d
= .)2 ()2,,)' r", exp [-!( 6y~ - 12y, Y3 + 6y~)]exp[ -!(3 y~)] dy,
= (J3/J2rr) exp[ - !y~];

that is, Y3 is normally distributed with mean 0 and variance t· IIII
PROBLEMS
1 (a)Let X h X 2 , and X3 be uncorrelated random variables with common variance
a 2 • Find the correlation coefficient between Xl + X 2 and X 2 + X 3 .
(b) Let Xl and X 2 be uncorrelated random variables. Find the correlation
coefficient between Xl + X 2 and X 2 - Xl in terms of var [Xl] and var [X2].
(c) Let Xl, X 2 , and X3 be independently distributed random variables with
common mean p. and common variance a 2 • Find the correlation coefficient
between X 2 - Xl and X3 - Xl'
2 Prove Theorem 2.
3 Let X have c.d.f. F x (') = F(·). What in terms of F(') is the distribution of
XI[o. CXJ)(X) = max [0, Xl?
4 Consider drawing bal1s, one at a time, without replacement, from an urn containing
M bal1s, K of which are defective. Let the random variable X( Y) denote the
number of the draw On which the first defective (nondefective) ba11 is obtained.
Let Z denote the number of the draw On which the rth defective baH is obtained.
(a) Find the distribution of X.
PROBLEMS 213
(b) Find the distribution of Z. (Such distribution is often cal1ed the negative
hypergeometric distribution.)
(c) Set M = 5 and K = 2. Find the jOint distribution of X and Y.
5 Let Xl, ••• , Xn be independent and identically distributed with common density
/x(x) = x- 2 I u , CXl)(X).
Set Y = min [Xl, ••. , X n ]. Does 8[X.] exist? If so, find it. Does 8[ Y] exist?
If so, find it.
6 Let X and Y be two random variables having finite means.
(a) Prove or disprove: 8[max [X, Yl] > max [8[X], 8[ Y]].
(b) Prove or disprove: 8[max [X, Yl + min [X, Yl] = 8[X] + 8[ Y].
7 The area of a rectangle is obtained by first measuring the length and width and
then multiplying the two measurements together. Let X denote the measured
length, Y the measured width. Assume that the measurements X and Y are
random variables with jOint probability density function given by Ix. l'(x, y) =
klr.9L.1.ILJ(x)hsw.I.2WJ(Y), where Land Ware parameters satisfying L > W> 0
and k is a constant which may depend on Land W.
(a) Find 8[XYl and var [XYl.
(b) Find the distribution of XY.
8 If X and Yare independent random variables with (negative) exponential dis-
tributions having respective parameters Al and A2 , find 8[max [X, Yl].
9 Projectiles are fired at the origin of an xy coordinate system. Assume that the
point which is hit, say (X, y), consists of a pair of independent standard normal
random variables. For two projectiles fired independently of one another, let
(XI, Yl ) and (X2 , Y 2 ) represent the points which are hit, and let Z be the distance
between them. Find the distribution of Z2. HINT: What is the distribution of
(X2 - X I )2? Of (Y2 - y 1 )2? Is (X2 - X I )2 independent of (Y2 - y.)2?
10 A certain explOsive device wiJ] detonate if anyone of n short-lived fuses lasts
longer than .8 seconds. Let Xl represent the life of the ith fuse": It can be as-
sumed that each Xi is uniformly distributed over the interval 0 to I second. Fur-
thermor~ it can be assumed that the XI'S are independent.
(a) How many fuses are needed (i.e., how large should n be) if one wants to be
95 percent certain that the device wiH detonate?
(b) If the device has nine fuses, what is the average life of the fuse that lasts the
longest?
11 Suppose that random variable Xn has a c.d.f. given by [(n - l)/n] <l>(x) + (l/n)Fn(x),
where <l> (.) is the c.d.f. of a standard normal and for each n Fn(·) is a c.d.f. What
is the limiting distribution of Xn?
12 Let X and Y be independent random variables each having a geometric distribu-
tion.
*(a) Find the distribution of X/(X + Y). [Define X/(X + Y) to be zero if
X+ y=o.]
(b) Find the joint moment generating function of X and X + Y.
214 DISTRIBUTIONS OF FUNCTIONs OF RANPOM vARIABLES v
13 Let Xl and X 2 be independent standard norma1 random variables. Let Y l = Xl

+ X 2 and Y 2 = Xl + Xl.
(a) Show that the joint moment generating function of Y l and Y 2 is
exp [tfl(1 - 2t2)]
for - ex) < t1 < ex) and - ex) < t2 < !.
Find the correlation coefficient of Y l and Y 2 •
(b)
14 Let X and Y be independent standard normal random variables. Find the m.g.f.
of XY.
15 Suppose that Xl and X 2 are independent random variable~ each having a standard
normal distribution.
(a) Find the joint distribution of (Xl + X 2)/,v'2 and (X2 ~ Xl)/vi.
(b) Argue that 2XI X 2 and Xl- Xl have the same distribution. HINT:
2 2 _ Xl+ X 2 X 2 - Xl
X 2 ~ Xl - 2 Vi Vi
16 A dry-bean supplier fi]ls bean bags with a machine that does not work very wen,
and he advertises that each bag contains 1 pound of beans. In fact, the weight
of the beans that the machine puts into a bag is a random variable with mean
16 ounces and standard deviation 1 Ounce. If a box contains 16 bags of beans:
(a) Find the mean and variance of the weight of the beans in a box.
(b) Find approximately the probabiHty that the weight of the beans in a box
exceeds 250 ounces.
(c) Find the probability that two or fewer underweight (less than 16 ounce) bags
are in the box if the weight of beans in a bag is assumed to be normally
distributed.
17 Numbers are selected at random from the interval (0, 1).
(a) If 10 numbers are selected, what is the probability that exactly 5 are less than
!?
If 10 numbers are selected, on the average how many are less than 1?
(b)
(c) If 100 numbers are selected, what is the probability that the average of the
numbers is less than 1?
18 Let Xl denote the number of meteors that collide with a test satellite during the
n
ith orbit. Let Sn = L Xl; that is, Sn is the total number of meteors that collides
1=1
with the satellite during n orbits. Assume that the XI'S are independent and
identically dist~ibuted Poisson random variables having mean A.
(a) Find I[Sn] and var [Sn].
(b) If n = 100 and A = 4, find approximately pISIOO > 440].
19 How many light bulbs should you buy if you want to be 95 percent certain that
you will have 1000 hours of light if each of the bulbs is known to have a lifetime
that is (negative) exponentially distributed with an average life of 100 hours?
(a) Assume that an the bulbs are burning simultaneously.
(b) Assume that one bulb is used until it burns out andlhen it is replaced. etc.
PROBLEMS 215
20 (a) If Xl, ... , XII are independent and identical1y distributed gamma random
variables, what is the distribution of Xl + ... XII?
(b) If X., •. " XII are independent gamma random variables and if X, has param-
eters ri and ", i 1, ... , n~ what is the distribution of Xl + ... + XII 1
21 (a) If Xl, •.• , XII are independent identically distributed geometric random
variables, what is the distribution of Xl + . . . Xn 1
(b) If Xl" •.• , XII are independent identically distributed geometric random
variables with density 0(1- (J)Jf- 1 l(l.z ....•(x), what is the distribution of
Xl + ... + XII?
(c) If Xl, .•. , XII are independent identically distributed negative binomial
random variables, what is the distribution of Xl ••• + XII ?
(d) If Xl, .•• , Xn are independent negative binomial random variables and if Xi
has parameters ri and p, what is the distribution of Xl + ... + XII?
*22 Kitty Oil Co. has decided to drill for oil in 10 different locations; the cost of
drilling at each location is $10,000. (Total cost is then $100,000.) The prob-
ability of finding oil in a given location is only "1, but if oil is found at a given
location, then the amount of money the company will get selling oil (excludi ng the
initial $10,000 deming cost) from that location is an exponential random variable
with mean $50,000. Let Y be the random variable that denotes the number of
locations where oil is found, and let Z denote the total amount of money received
from selling oj] from an the locations.
(a) Find 8[Z].
(b) Find P[Z> 100,0001 Y = 1] and P[Z > 100,0001 Y = 2].
(c) How would you find p[Z > 100,000]1 Is P[Z> 100,000] > i?
23 If Xl, ..• , X" are independent Poisson distributed random variables, show that
the conditional distribution of Xl, given Xl + ... + X", is binomial.
*24 Assume that Xl, ... , Xk+l are independent Poisson distributed random variables
with respective parameters "1, ... ,
",,+1' Show that the conditional distribution
of XI, ... , X" given that Xl ... + X"+1 = n has a multinomial distribution
with parameters n, "d", "" ""I", "1
where" = + ... +",,+1' .
25 If X has a uniform distribution over the interval ( -7T/2, 7T/2), find the distribution
of Y tan X.
26 If X has a normal distribution with mean f.L and variance (12, find the distribution,
mean, and variance of Y = eX.
27 Suppose X has c.d.f. Fx(x) exp [-e-(Jf-<l:)/P]. What is the distribution of
Y exp [-(X - (X)/{1]?
28 Let X have density
1 x lO
-
fx(x; a, b) B(a, b) (l + Xf+b 1(0, ~)(x),

where a > 0 and b > O. (This density is often called a beta distribution of the
second kind.) Find the distribution of Y = 1/(1 + X).
29 If X has a uniform distribution on the interval (0, 1), find the distribution of 11 X.
Does 8[11 XJ exist? If so, find it.
30 (a) Give an example of a distribution of a random variable X for which 8[1/ Xl

is not finite.
(b) Give an example of a distribution of a random variable X for whiCh 8[1/ Xl is
finite, and evaluate 8[1/ Xl.
31 If Ix(x) = 2xe- X2 / to . oo)(x), find the density of Y = X2.
32 If X has a beta distribution, what is the distribution of 1 - X?
33 If Ix(x) = e-x/(o. oo)(x), find the distribution of X/(1 + X).
34 If Ix(x) = 1/7T(1 + x 2), find the distribution of I/X.
35 If Ix (x) = 0 for x <0, find the density of Y = aX2 + b in terms of Ix( .) for a > o.
36 If X has the Weibu]] distribution as given in Eq. (42) of Chap. III, what is the
distribution of Y = aXb ?
37 (a) Let Y = X2 and Ix(x) = (1/8)/(0. 8)(X), 8> o. Find the c.dJ. of X and Y.
Find the density of Y.
(b) Let Y = X2 and Ix(x) = (!8)/(_8. 8)(X), 8> O. Find the c.dJ. and density
of Y.
38 If X and Yare independent random variables, each having the same geometric
distribution, find the distribution of Y - X.
39 If X and Yare independent random variables, each having the same negative
exponential distribution, find the distribution of Y - X.
40 If X, Y, and Z are independent random variables, each uniformly distributed over
(0, 1), what is the distribution of XY/Z?
41 Assume that X and Y are independent random variables, where X has a p.dJ.
given by Ix(x) = 2xl(o, l)(x) and Y has a p.dJ. given by Iy(y) = 2(1 - y)/(o, 1)(Y).
Find the distribution of X + Y.
*42 Let X and Y be independent Poisson distributed random variables. Find the
distribution of Y - X.
43 If Ix(x) = 1(0. l)(x), find the density of Y = 3X + 1.
*44 Let X and Y be two independent beta-distributed random variables. Is XYalways
beta-distributed? If not, find conditions on the parameters of X and Y that wi1l
imply that XY is beta-distributed.
45 If Ix. y(x, y) = e-(XH)/(o. oo)(x)/(o. OO)(y), find the density of Z = (X + Y)/2.
46 If lx, y(x, y) =4xye-(.%2 U 2)/(0. oo)(x)/(o. oo)(Y), find the density of V X2 + 1'2.
47 If Ix. y(x, y) = 4xyl(o, 1)(x)/(o, l)(y)' find the joint density of X2 and y2.
48 If Ix. y(x, y) = 3x/(0.x)(y)/(o, 1)(X), find the density of Z = X - Y.
49 If Ix(x) = [(1 + x)/2l/(-I,1)(X), find the density of Y = X2.
50 If Ix. y(x, y) = 1(0. 1)(x)/(o, l)(Y), find the density of Z, where
Z = (X + Y)/(-oo. 1](X+ y) + (X + Y - 1)/(1. OO)(X + y).
51 If lx, y(x, y) = e-(x+)')/(o, oo)(x)/(o. oo)(y), find the joint density of X and X + Y.
52 If Ix. y. z(x, y, z) = e-(:K+),H)/(o, oo)(x)/(o. oo)(y)/(o, oo)(z), find the density of their
average (X + Y + Z)/3.
53 If Xl and X 2 are independent and each has probability density given by
Ae-Ax/(o. oo)(x), find the joint distribution of YI = X I /X2 and Y 2 = Xl + X 2 and
the marginal distributions of YI and Y 2 •
PROBLEMS 217
*54 Let Xl and X 2 be independent random variables, each normally distributed with
parameters p. = 0 and (72 = 1. Find the joint distribution of Y1 = Xl + Xl and
(
Y 2 = X 1 !X2 • Find the marginal distribution of Y1 and of Y 2 • Are Y1 and Y 2
independent?
55 If the joint distribution of X and Y is given by
Ix. rex, y) = 2e- (.x'+.v) I[o.),](x)I[o. C;C)(Y},
find the jOint distribution of X and X + Y. Find the marginal distributions of
X and X Y.
56 Let Ix. rex, y) K(x + y)l(o. l)(x)/(o. l)(y)I(o. l)(x y).
(a) Find Ix( ').
(b) Find the jOint and marginal distributions of X + Yand Y - X.
57 Suppose !c.x. ]I) I z(x, y Iz) = [z + (I - z}(x+- y»)I(o. l)(x)l(o. l)(Y) for 0 <z < 2, and
Iz(z} = !i[o. 2](Z}.
(a) Find 8[X YJ.
(b) Are X and Y independent? Verify.
(e) Are Xand Z independent? Verify.
(d) Find the joint distribution of X and X + Y.
(e) Find the distribution of max [X, Y]IZ = z.
(I) Find the distribution of (X + Y) I Z = z.
58 A system will function as long as at least one of three components functions.
When all three components are functioning, the distribution of the life of each is
exponential with parameter iA, When only two are functioning, the distribution
of the I ife of each of the two is exponential with parameter !A; and when only One
is functioning, the distribution of its 1ife is exponential with parameter A.
(a) What is the distribution of the lifetime of the system?
(b) Suppose nOw that only one component (of the three components) is used at a
time and it is replaced when it fails. What is the distribution of the lifetime
of such a system?
59 The system in the sketch will function as long as component C 1 and at least One
of the components C 2 and C 3 functions. Let Xi be the random variable denoting
the lifetime of component C" i I, 2, and 3. Let Y = max [X2 • X 3 ] and Z =
min [Xl, n. Assume that the Xs's are independent (negative) exponential
random variables with mean I.
(a) Find 8[Z] and var [Z].
(b) Find the distribution of the lifetime of the system.
r--
- C2 :-
CI
-
c3 I-
218 DISTRIBUTIONS OF FUNCfIONS OF RANDOM VARIABLES v
60 A system, which is composed of two components~ will function as 19n9 as at least

one of the two components functions. When both components are operating~
the lifetime distribution of each is exponential with mean 1. However~ the dis-
tribution of the remaining lifetime of the good component~ after ~ne fails, is
exponential with mean t. (The idea is that after one component fails the other
component carries twice the load and hence has only half the expected lifetime.)
Find the lifetime distribution of the system.
*61 Suppose that (X, Y) has a bivariate normal distribution. Find the jOint distribu-
tion of aX + bY and eX + dY for constants a, b, e, and d satisfying ad-be#O.
Find the distribution of aX + bY. HINT: Use the moment-generating-function
technique and see Example 7.
62 Let Xl and X 2 be independent standard normal random variables. Let U be
independent of Xl and X 2 , and assume that U is uniformly distributed over(O, I}.
DefineZ = UXI + (1 - U}X2 •
(a) Find the conditional distribution of Z given U = u.
(b) Find cB'[Z] and var [Z].
* (e) Find the distribution of Z.
VI
SAMPLING AND SAMPLING DISTRIBUTIONS

The purpose of this chapter is to introduce the concept of sampling and to present
some distribution theoretical results that are engendered by sampling. It is a
connecting chapter-it merges the distribution theory of the first five chapters
into the statistical theory of the last five chapters. The intent is to present
here in one location some of the laborious derivations of distributions that are
associated with sampling and that will be necessary in our future study of the
theory of statistics, especially estimation and testing hypotheses. Our thinking
is that by deriving these results now, our later presentation of the statistical
theory will not have to be interrupted by their derivations. The nature of the
material to be given here is such that it is not easily motivated.
Section 2 begins with a discussion of populations and samples. It ends
with the definitions of statistic and of sample moments. Sample moments are
important and useful statistics. Section 3 is devoted to the consideration of
various results associated with the sample mean. The law of large numbers
and the central-limit theorem are given, and then the exact distribution of the
sample means from several of the different parametric families of distributions
220 SAMPLING AND SAMPLING DISTRIBUTIONS VI
introduced in Chap. III is given. Sampling from the normal distribution is con-
sidered in Sec. 4!J where the chi-square, F, and t distributions are defined. Order
statistics are discussed in the final section; theY!J like sample moments, are impor-
tant and useful statistics.
2 SAMPLING
2 .. 1 Inductive Inference
Up to now we have been concerned with certain aspects of the theory of prob-
ability, including distribution theory. Now the subject of sampling brings us
to the theory of statistics proper, and here we shall consider briefly one important
area of the theory of statistics and its relation to sampling.
Progress in science is often ascribed to experimentation. The research
worker performs an experiment and obtains some data. On the basis of the .
data, certain conclusions are drawn. The conclusions usually go beyond the
materials and operations of the particular experiment. In other words, the
scientist may generalize from a particular experiment to the class of all similar
experiments. This sort of extension from the particular to the general is called
inductive inference. It is one way in which new knowledge is found.
Inductive inference is well known to be a hazardous process. In fact, it
is a theorem of logic that in inductive inference uncertainty is present. One
simply cannot make absolutely certain generalizations. However, uncertain
inferences can be made, and the degree of uncertainty can be measured if the
experiment has been performed in accordance with certain principles. One
function of statistics is the provision of techniques for making inductive in-
ferences and for measuring the degree of uncertainty of such inferences'. Un-
certainty is measured in terms of probability, and that is the reason we have
devoted so much time to the theory of probability.
Before proceeding further we shall say a few words about another kind of
inference-deductive inference. While conclusions which are reached by induc-
tive inference are only prObable, those reached by deductive inference are con-
clusive. To illustrate deductive inference, consider the following two statements:
(i) One of the interior angles of each right triangle equals 90°,
(ii) Triangle A is a right triangle.
If we accept these two statements, then we are forced to the conclusion:
(iii) One of the angles of triangle A eq uals 90°.
2 SAMPLING 221
FIGURE 1 West Point graduates
This is an example of deductive inference, which can be described as a

method of deriving information [statement (iii)] from accepted facts [statements
(i) and (ii)]. Statement (i) is called the major premise, statement (ii) the minor
premise, and statement (iii) the conclusion. For another example, consider the
following:
(i) Major premise: All West Point graduates are over 18 years of age.
(ii) Minor premise: John is a West Point graduate.
(iii) Conclusion: John is over 18 years of age.
West Point graduates is a subset of all persons over 18 years old, and John
is an element in the subset of West Point graduates; hence John is also an element
in the set of persons who are over 18 years old.
While deductive inference is extremely important, much of the new knowl-
edge in the real world comes about by the process of inductive inference. In the
science of mathematics, for example, deductive inference is used to prove the-
'orems, while in the empirical sciences inductive inference is used to find new
knowledge.
Let us illustrate inductive inference by a simple example. Suppose that
we have a storage bin which contains (let us say) 10 million flower seeds which
we know will each produce either white or red flowers. The information which
we want is: How many (or what percent) of these 10 million seeds will produce
white flowers? Now the only way in which we can be sure that this question
is answered correctly is to plant every seed and observe the number producing
white flowers. However, this is not feasible since we want to sell the seeds.
Even if we did not want to sell the seeds, we would prefer to obtain an answer
without expending so much effort. Of course, without planting each seed and
observing the color of flower that each produces we cannot be certain of the
number of seeds producing white flowers. Another thought which occurs is:
Can we plant a· few of the seeds and, on the basis of the colors of these few
flowers, make a statement as to' how many of the 10 mi11ion seeds will produce
white flowers? The answer is that we cannot make an exact prediction as to

how many white flowers the seeds will produce but we can make a probabilistic
statement if we select the few seeds in a certain fashion. This is inductive
inference. We select a few of the 10 million seeds, plant them, observe the
number whiCh produce white flowers, and on the basis of these few we make a
prediction as to how many of the 10 million will produce white flowers; from a
knowledge of the color of a few we generalize to the whole 10 million. We
cannot be certain of our answer, but we can have confidence in it in a frequency-
ratio-probability sense.
2.2 Populations and Samples

We have seen in the previous subsection that a central problem in discovering
new knowledge in the real world consists of observing a few of the elements
under discussion, and on the basis of these few we make a statement about the
totality of elements. We shall now investigate this procedure in more detail.
Definition 1 Target population The totality of elements which are under

discussion and about which information is desired will be called the target
population. // / /
In the example in the previous subsection the 10 million seeds in the stor-
age bin form the target population. The target population may be all the dairy
cattle in Wisconsin on a certain date, the prices of bread in New York City on a
certain date, the hypothetical sequence of heads and tails obtained by tossing a
certain coin an infinite number of times, the hypothetical set of an infinite
number of measurements of the velocity of light, and so forth. The important
thing is that the target population must be capable of being quite well defined;
it may be real or hypothetical. .
The problem of inductive inference is regarded as follows from the point
of view of statistics: The object of an investigation is to find out something about
a certain target population. It is generally impossible or impractical to examine
the entire population, but one may examine a part of it (a sample from it) and,
on the basis of this limited investigation, make inferences regarding the entire
target population.
The problem immediately arises as to how the sample of the population
should be selected. We stated in the previous section that we could make prob-
abilistic statements about the population if the sample is selected in a certain
fashion. Of particular importance is the case of a simple random sample,
usually called a random sample, which is defined in Definition 2 below for any
2 SAMPLING 223
population which has a density. That is, we assume that each element in our
population has some numerical value associated with it and the distribution of
these numerical values is given by a density. For such a population we define
a random sample.
Definition 2 Random sample Let the random variables Xl' X 2, •.. , Xn

have a joint density fXJ' .__ • Xn( " ••• , .) that factors as follows:
fXJ.Xl •.... Xn(Xh X2, "', xn) = f(Xl)f(X2) ..... f(xn),
where f( .) is the (common) density of each Xi' Then Xl' X 2, •••• Xn is

defined to be a random sample of size n from a population with density
f(·). ///1
In the example in the previous subsection the 10 million seeds in the stor-
age bin formed the population from which we propose to sample. Each seed is
an element of the population and will produce a white or red flower; so, strictly
speaking, there is not a numerical value associated with each element of the
population. However, if we, say, associate the number 1 with white and the
num ber 0 with red, then there is a numerical value associated with each element
of the population, and we can discuss whether or not a particular sample is
random. The random variable Xi is then 1 or 0 depending on whether the ith
seed sampled produces a white or red flower, i 1, ... , n. Now if the sampling
of seeds is performed in such a way that the random variables Xl' ... , Xn are
independent and have the same density. then, according to Definition 2, the
sample is caned random.
An important part of the definition of a random sample is the meaning of
the random variables Xl, .. " X n • The random variable Xi is a representation
for the numerical value that the ith item (or element) sampled will assume. After
the sample is observed, the actual values of Xl' ... , Xn are known, and as usual,
we denote these observed values by X h ••. , X n • Sometimes the observations
Xl' ••. , Xn are called a random sample if Xl, ... , Xn are the values of Xl, .. " X n•
where Xl' "', Xn is a random sample.
Often it is not possible to select a random sample from the target popula-
tion, but a random sample can be selected from some related population. To
distinguish the two populations, we define sampled population.
Definition 3 Sampled population Let Xl' X 2, ... , Xn be a random

samp1e from a population with density f( . ); then this population is called
the sampled population.' IIII
Valid probability statements can be made about sampled populations on

the basis of random samples, but statements about the target populations are
not valid in a relative-frequency-probability sense unless the target population
is also the sampled population. We shaH give some examples to bring out the
distinction between the sampled population and the target population.
EXAMPLE I Suppose that a sociologist desires to study the religious habits

of 20-year-old males in the United States. He draws a sample from the
20-year-old males of a large city to make his study. In this case the target
population is the 20-year-old males in the United States, and the sampled
population is the 20-year-old males in the city which he sampled. He can
draw valid relative-frequency-probabilistic conc1usions about his sampled
population, but he must use his personal judgment to extrapolate to the
target population, and the reliability of the extrapolation cannot be
measured in relative-frequency-probability terms. 111/
EXAMPLE 2 A wheat researcher is studying the yield of a certain variety of
wheat in the state of Colorado. He has at his disposal five farms scat-
tered throughout the state on which he can plant the wheat and observe
the yield. The sampled population consists of the yields on these five
farms, whereas the target population consists of the yields of wheat on
every farm in the state. IIII
This book will be concerned with the problem of selecting (drawing) a
sample from a sampled population with density f( .), and on the basis of these
sample observations probability statements will be made about f( .), or infer-
ences about f( . ) will be made.
Remark We shall sometimes use the statement "population f(·)" to

mean" a population with density f( . )." When we use the word" popula-
tion" without an adjective" sampled" or "target," we shall always mean
sampled population. IIII
2.3 Distribution of Sample
Definition 4 Distribution of sample Let Xl' X 2 , ••• , Xn denote a sample
of size n. The distribution of the sample Xl' ... , Xn is defined to be the
joint distribution of X h ... , X n • II/I
Suppose that a random variable X has a density 1(') in some population,
and suppose a sample of two values of X, say Xl and X2' is drawn at random.
2 SAMPLING 225
Xl is called the first observation, and X2 the second observation. The pair of
numbers (Xl' x 2 ) determines a point in a plane, and the collection of all such
pairs of numbers that might have been drawn forms a bivariate population.
We are interested in the distribution (bivariate) of this bivariate population in
terms of the original density f( . ). The pair of numbers (Xl' X2) is a value of the
joint random variable (Xl' X 2), and Xl' X 2 is a random sample (of size 2)
from I( ,). By definition of random sample, the joint distribution of Xl and
X 2 , which we call the distribution of our random sample of size 2, is given by
fxt, X2(X l , X2) = f(x l )f(x 2)·
As a simple example, suppose that X can have only two values, 0 and 1,
with probabilities q = 1 - p and p, respectively. That is, X is a discrete ran-
dom variable which has the Bernoulli distribution
f(x) = pXql -x/{O.l}(X), (1)
The joint density for a random sample of two values fromf( .) is
IXI. X2(X l , X2) = f(x l )f(x 2) = pXI +X2q2- XI-X2f{O. l}(xl)I{o. 1}(x 2). (2)
It is to be observed that this (bivariate) density is not what we obtain as the

distribution of the number of successes, say Y, in drawing two elements from a
Bernoulli population. The density of Y is given by
frey) = G) p'q 2-, for y = 0, 1,2.
The single random variable Yequals Xl + X 2 .

It should be noted thatfxl. X2(X t , X2) gives us the distribution of the sample
in the order drawn. For instance, in Eq. (2), f X Io X 2(O, 1) = pq refers to the
probability of drawing first a 0 and then a 1.
Our comments for a random sample of size 2 generalize to a random sam-
ple of size n, and we·have the following remark.
Remark If Xl' X 2 , ... , Xn is a random sample of size n. from f( ,), then

the distribution of the random sample Xl' .. " X n , defined as the joint dis-
tribution of X h . . . , X n, is given by fxt. .... x,.(x l , ... , xn) = f(x l )'" 'f(xn)'
1//1
Note that again this gives the distribution of the sample in the order drawn.
Also, note that if X h ... , Xn is a random sample, then Xl' ... , Xn are stochasti-
cally independent.
We might further note that ou~ definition of random sampling has auto-
matical1y ru]ed out sampling from a finite population without replacement s"ince,
then, the results of the drawings are not independent.
•
2.4 Statistic and Sample Moments

One of the central problems in statistics is the following: It is desired to study a
population which has a density I( . ; 0), where. the form of the density is known
but it contains an unknown parameter 0 (if 0 is known. then the density func·
tion is completely specified). The procedure is to take a random sample
Xl' X 2 , •••• XII of size n from this density and let the value of some function,
say t(XI' X2 , ••. , XII)' represent or estimate the unknown parameter O. The
pr<?blem is to determine whiCh function will be the best one to estimate O. This
problem will be formulated in more detail in the next chapter. In this section
we shall examine certain functions, namely, the sample moments, of a random
sample. First, however, we shall define what we shall m.ean by a statistic.
Definition 5 Statistic A statistic is a function of observable random

variables, which is itself an observable random variable, which does not
contain any unknown parameters. 1/11
The qualification imposed by the word" observable" is required because
of the way yve intend to use a statistic. (" Observable" means that we can observe
the values of the random variables.) We intend to use a statistic to make in-
ferences about the density of the random variables. ~nd if the random variables
were not observable, they would be of no use in making inferences.
For example, if the observable random variable X has the density ¢Jl.<F2 (x),
where J1 and a 2 are unknown, then X - J1 is not a statistic; neither is X/a (since
they are not functions of the observable random variable X only-they contain
unknown parameters), but X, X + 3, and X2 + log X 2 are statistics.
In the formulation above, one of the central problems in statistics is to
find a suitable statistic (function of the random variables Xl' X 2 , ••• , Xn) to
represent O.
EXAMPLE 3 If X B •.. , Xn is a random sample from a density I( . ; 0), then

(provided Xl' ... , XII are observable)
_ 1 n
xn = - I x,
.. n i= I
is a statistic, and
!{min [Xh ... , XII] + max [Xl' ... , XII]}
is also a statistic. If I(x; 0) ¢o, 1 (x) and 0 is unknown, Xn 0 is not a
statistic since it depends on 0, which is unknown. ////
2 SAMPLING 227
Next we shall define and discuss some important statistics, the sample
moments.
Definition 6 Sample moments Let Xl' X 2 , ••• , Xn be a randonl sample

from the density f( . ). Then the rth sample moment about 0, denoted ~y
M;, is defined to be
1 n
M;=-
n
LXi·
i= I
(3)
In particular, if r = 1, we gf!t the sample mean, which is usually denoted by

X or Xn; that is,
Also, the rth sample moment about X n , denoted by M,., is defined to be
M,. = -1 ~ _)'
~ (Xi - Xn . (5)
n i= 1
IIII
Remark Note that sample moments are examples of statistics. IIII

We will consider in detail some properties of the sample mean in Sec. 3
below.
In Chap. II we defined the rth moment of a random variable X, or the rth
moment of its corresponding density fx(·), to be @[X'] = /1;. We could say
that 6"[XY] is the rth population moment of the population with density f(x) =
fx(x). We_ shall now show that the sample moments reflect the population
moments in the sense that the expected value of a sample moment (about 0)
---
equa}s the corresponding population moment. Also, the variance of a sample
moment wiJI be shown to be (lIn) ti~es some function of population moments.
The implication is that for a given population the values that the sample moment
assume will tend to be more concentrated about the corresponding population
moment for large sample size n than for small sample size.' Thus a sample
moment can be used to estimate its corresponding population moment (provided
the population 'moment exists).
Theorem 1 Let Xl' X 2 • Xn be a random sample from a population

••• ,
with a density f(·)· The expected value of the rth sample moment (about
0) is equal to the rth population moment; that is,
(if J.l; exists). (6)

Also,
(if /1;, exists). (7)
PROOF
var[M;] = var [~
ni=1
t X~]
= (~)2 var [.tx~]
n ,In
= (!)2 .t var[X~]
,=1
In particular, if r = 1, we get the following coro]]ary. 1111

Corollary Let Xl' X 2 • . •.• Xn be a random sample from a density I( . ),
I "
and let Xn = - L Xi be the sample mean; then
n i== 1
and - ] = -1 (J 2 ,
var [ Xn (8)
n
where It and (J2 are, respective1y, the mean and variance of I( .). 1111
As we mentioned earlier, properties of the sample mean wi11 be studied

in detail in the next section.
Theorem 1 gives the mean and variance in terms of population moments of
the rth sample moment about 0; a s!milar, though more complicated, result
can be derived for the mean and variance of the rth sample moment about the
sample mean. We wiJ] be content with looking only at the particular case
n
r = 2, that is, at M2 = (lIn) L (Xi - X)2. M2 is sometimes cal1ed the sample
i= 1
v!!~iaI}.9..e..; although we will take Definition 7 as our definition of the sample
varIance.
2 SAMPLING 229
Definition 7 Sample variance Let Xl) X 2 , ••• , Xn be a random sample

from a 9-ensity f( . ); then
for n> 1 (9)
is defined to be the sample variance. /III
The reason for taking 8 2 rather than M 2 as our definition of the sample
variance (both measure dispersion in the sample) is that the expected value of
8 2 equals the population variance.
The proof of the following remark is left as an exercise.
I nn
2
Remark 8 n2 = 8 = 2 ( 1) '"
L, '"L, (Xi - Xj) 2 . 1111
nn- i= 1 j= 1
Theorem 2 Let Xh X 2 , ... , Xn be a random sample from a density

f( .), and let
Then
and var [8 2 ] for n > 1. (10)
PROOF (Only the first part wil1 be proved.) Recall that (J2
= 8[(X - 11)2] and 11,. 8[(X 11)"]. We commence by noting and prov-
ing an identity that is quite useful.
n n
L (Xi -
i= 1
11)2 =
i
L1(Xi - x)2 + n(X - 11)2 (11)
sInce
L (Xi - 11)2 = L (Xi - X + X - 11)2 L [(Xi - X) + (X - 11)]2

= L [(Xi - X)2 + 2(X, - X)(X -·11) + (X - 11)2]
= L (Xi - X)2 + 2( X - 11) L (Xi - X) + n(X - Il)2
= L (Xi - X)2 + n(X - tt)2.
230 SAMPUNG AND SAMPLING DISTIUBUTIONS VI
Using the identity of Eq. (11), we obtain
S[8 2 ] = S[ 1
n-1
L (Xi - X)2]
=
n - 1 L=l
±
1 sf (Xi - p)2 - n(X - p)2]
= n ~ 1 {.t1 8 [(X, - p)2] - n8[(X - p)2]}

= n~ 1 ttu nvar[X]}2
-
= 1 ( n(12 - n (12)
- = (12.
n -1 n
Although the derivation of the formula for the variance of 8 2 can
be accomplished by utilizing the above identity [Eq. (11)] and
_ 1,. 1
X- p = - L. Xi - - np
n n
111 I
= -
n
L Xi - n- L p = -n L (Xi - p),
such derivation is lengthy and is omitted here only to be relegated to

the Problems. / / //
Sample moments are examples of statistics that can be used to estimate

their population counterparts; for example, M; estimates It;, X estimates It,
and 8 2 estimates (12. In each case, we are taking some function of the sample,
which we can observe, and using the value of that function of the sample to
estimate the unknown population parameter.
3 SAMPLE MEAN
The first sample moment is the sample mean defined to be
I "
X = X" = -
ni=l
L Xi'
where Xh X 2 , ••• , X,. is a random sample from a density I( .). X is a function
of the random variables Xl' ... , X,., and hence theoretically the distribution of
X can be found. In general, one would suspect that the distribution of X
3 SAMPLE MEAN 231
depends on the density f(·) from which the random sample was selected, and
indeed it does. Two characteristics of the distribution of X, its mean and
variance, do not depend on the density f( . ) per se but depend on]y on two charac-
teristics of the density f(·). This idea is reviewed in the following subsection,
while succeeding subsections consider other results involving the samp]e mean.
The exact distribution of X will be given for certain specific densities f( . ).
It might be helpful in reading this section to think of the sample mean X
as an estimate of the mean J1 of the density f(·) from which the samp]e was
selected. We might think that one purpose in taking the sample is to estimate
J1 with X.
3.1 Mean and Variance
Theorem 3 Let Xl' X 2 , ••• , Xn be a random sample from a density f(·),

n
which has mean J1 and finite variance (12, and let X = (lIn) I Xi. Then
i=l
and (12)
IIII
Theorem 3 is just a restatement of the corol1ary of Theorem 1. In light

of using a va]ue of X to estimate J1, let us note what Theorem 3 says. @[ X] = J1
says that on the average X is equal to the parameter J1 being estimated or that
the distribution of X is centered about J1. var [Xl = (1ln)(12 says that the
spread of the va]ues of X about J1 is sma]] for a ]arge samp]e size as compared to
a small sample size. For instance, the variance of the distribution of X for a
sample of'size-W is one-ha]f the variance of the distribution of X for a sample of
size 10. So for a large sample size the values of X (which are used to estimate
J1) tend to be more concentrated about J1 than for a smal1 sample size. This
notion is further exemplified by the law of large numbers considered in the next
subsection.
3.2 Law of Large Numbers

Letf(· ; (J) be the density of a random variable X. We have discussed the fact
that one way to get some information about the density function f(· ; (J) is to
observe a random sample and make an inference from the sample to the popu]a-
tion. If (J were known, the density functions would be completely specified,
--
232~AMPLING
"'" -...-' J
AND SAMPLING DISTRIBUTIONS VI
and no inference from the sample to the popUlation would be necessary. There-
fore, it seemS that we would like to have the random sample tell us something
about the unknown parameter 0. This problem will be discussed in detail in
the next chapter. In this subsection we shall discuss a related particular
problem.
Let S[X] be denoted by J-l in the density f(·). The problem is to estimate
J-l. In a loose sense, 8[X] is the average of an infinite number of values of the
random variable X. In any real-world problem we can observe only a finite
number of values of the random variable X. A very crucial question then is:
Using only a finite number of values of X (a random sample of size n, say), can
any reliable inferences be made about S[X], the average of an infinite number of
values of X? The answer is "yes"; reliable inferences about S[X] can be made
by using only a finite sample, and we shaH demonstrate this by proving what is
called the weak law of large numbers. In words, the law states the following: A
positive integer n can be determined such that if a random sample of size n or
larger is taken from a population with the density 1(') (with 8[X] = 11), the
probability can be made to be as close to I as desired that the sample mean X
will deviate from J-l by less than any arbitrarily specified small quantity. More
precisely, the weak law of large numbers states that for any two chosen small
numbers e and c5, where e > 0 and 0 < c5 < 1, there exists an integer n such that
if a random sample of size n or larger is obtained from f(·) and the sample
mean, denoted by Xn , computed, then the probabiJity is greater than I - c5
(i.e., as close to I as desired) that Xn deviates from 11 by less than e (i.e., is ar-
bitrarily close to J-l). In symbols this is written: For any e> 0 and 0 n
The weak law of large numbers is proved using the Chebyshev inequality given
in Chap. II.
Theorem 4 Weak law of large numbers Let I( . ) be a density with mean

J-l and finite variance (J2, and let xn be the sample mean of a random
sample of size n from f(')' Let e and c5 be any two specified numbers
satisfying e > 0 and 0 I - b. (13)
PROOF Theorem 5 in Subsec. 4.4 of Chap. II stated that P[g( X) k]
< cB'[g(X)]/k for every k> 0, random variable X, and nonnegative func-
tion g(.). Equivalently, P[g(X) < k] > 1 - 8[g(X)]/k.
3 SAMPLE MEAN 233
2
Letg(X) = (Xn - J.l)2 and k = 8 ; then
P[ -8 < Xn - J.l < 8] = P[I Xn - J.l1 < 8]
- 2 2 1 S[(Xn - J.l?]
= P[ I Xn - J.l1 < 8 ] >-
8
2
IIII
Below are two examples to illustrate how the weak law of large numbers
can be used.
EXAMPLE 4 Suppose that some distribution with an unknown mean has

variance equal to 1. How large a sample must be taken in order that the
probability wi1l be at least .95 that the sample mean Xn will lie within .5
of the population mean? We have a 2 = 1, 8 = .5, and () = .05; therefore
a2 1
n > {)8 2 = .05(.5)2 = 80. IIII
EXAMPLE 5 How large a sample must be taken in order that you are 99
percent certain that Xn is within .5a of J.l? We have 8 = .5a and () = .01.
Thus
IIII
We have shown that by use of a random sample inductive inferences to

populations can be made and the reliability of the inferences can be measured in
terms of probability. For instance, in Example 4 above, the probability that
the sample mean will be within one-half unit of the unknown population mean
is at least .95 if a sample of size greater than 80 is taken.
3.3 Central-limit Theorem

Although we have already stated the central-limit theorem in our study of
distribution theory in Chap. V, we will repeat it here in our study of the sample
mean X because it gives the asymptotic distribution of X. At the outset of this
~~~PLING AND SAMPLING DISTRIBUTIONS VI
section we indicated that we were interested in the distribution of x. The

central-limit theorem, which is one of the most important theorems in all of
probability and statistics, tells us approximately how X is distributed.
Theorem 5 Central-limit theorem Let/(') be a density with mean J.! and

finite variance a 2 • Let X" be the sample mean of a random sample of
size n from/(·). Let the random variable Z" be defined by
Z = X" - 8[X,,] = X" - J.! • (14)

" Jvar [X,,] alJn
Then, the distribution of Z" approaches the standard normal distribution
as n approaches infinity. IIII
Theorem 5 tells us that the limiting distribution of Z" (which is X" stand-
ardized) is a standard normal distribution, or it tells us that X" itself is ap-
proximately, or asymptotically, distributed as a norma] distribution with mean
Jl and variance a2 In.
The astonishing thing about Theorem 5 is the fact that nothing is said
about the form of the original density function. Whatever the distribution
function, provided only that it has a finite variance, the sample mean wiJ] have
approximately the normal distribution for large samples. The condition that
the variance be finite is not a critical restriction so far as applied statistics is
concerned because in almost any practical situation the range of the random
variable will be finite, in which case the variance must necessarily be finite.
The importance of Theorem 5, as far as practical applications are con-
cerned, is the fact that the mean X" of a random sample from any distribution
with finite variance a 2 and mean J.! is approximately distributed as a normal
random variable with mean J.! and variance a2 /n.
We shall not be able to prove Theorem 5 because it requires rather ad-
vanced mathematical techniques. However, in order to make the theorem
plausible, we shall outline a proof for the more restricted situation in which
the distribution has a moment generating function. The argument will be
essentially a matter of showing that the moment generating function for the
sample mean approaches the moment generating function for the normal
distribution.
Recall that the moment generating function of a standard normal dis··
t2 t2
tribution is given by e-t • (See Subsec. 3.2 of Chap. III.) Let m(t) = e1: •
Let mzJt) denote the moment generating function of Z" . It is our purpose to
show that mz.,(t) must approach met) when n, the sample size, becomes large.
3 SAMPLE MEAN 235
Now
mz.(t) = .'le'z"] = 8[exp tZ.] = c[exp (t ~/J:)] .r[exp (~L :~J:)] =
= c9'[Ii exp (! .Xi - 11)] = .11 [ex p (Jt _.Xi - 11)]

&
i=1 n ujJn n a
1=1
using the independence of Xl' ... , X n • Now if we let Y j = (Xi - II}jO', then
mylt), the moment generating function of Y j , is independent of i since a II Yt
have the same distribution. Let my(t) denote my.(t); then
Hence,
mz.(t) = [my C~) r· (15)
The rth derivative of my(tj J;;) evaluated at t = 0 gives uS the rth moment about
the mean of the densityf(·) divided by (O'J~y, so we may write
t ) III t 1 112 ( t ) 2 1 113 ( t ) 3

my ( In =1+~ In + 2! 0'2 In + 3! 0'3 J~ + "', (16)
and since III = 0 and 112 = 0'2, this may be written
my ( J-t n) = 1 + n1- (1-2 t 2

+ 1 113
J 3 t
3! nO'
3 I 114 4
+ 4,4 )
t +... .
.n a
(17)
Now lim (1 + ujnt = ett2 , where u represents the expression within the paren-
theses in Eq. (17). We have lim m ZI1(t) = ett \ so that in the Jimit, Zn has the
same moment generating function as a standard normal and, by a theorem
similar to Theorem 7 in Chap. II, has the same distribution.
The degree of approximation depends, of course, on the sample size and
on the particular density f(·). The approach to normality is iI1ustrated in
Fig. 2 for the particular function defined by f(x) = e- X l(o,oo)(x). The solid
curves give the actual distributions, while the dashed curves give the normal
approximations. Figure 2a gives the original distribution which corresponds to
samples of 1 ; Fig. 2b shows the distribution of sample means for n = 3; Fig. 2c
236 ')AMPLING AND SAMPLING DISTRIBUTIONS VI
"---~
1.5 n =1 1.5 n = 3 1.5 n = 10
1.0 1.0
-'--------'------"-- x
o 1 2 x o 1 2 x o 1
(a) (b) (c)
FIGURE 2
gives the distribution of sample means for n = 10. The curves rather exaggerate
the approach to norma1ity because they cannot show what happens on the tails
of the distribution. Ordinarily distributions of samp1e means approach
normality fair1y rapid1y with the sample size in the region of the mean, but more
s10wly at points distant from the mean; usual1y the greater the distance of a
point from the mean, the more slowly the normal approximation approaches the
actual distribution.
In the fol1owing subsections we wi]] give the exact distribution of the
sample mean for some specific densities f(')'
3.4 Bernoulli and Poisson Distributions

If Xl' X 2 , ••• , Xn is a random sample from a Bernoulli distribution, we can
find the exact distribution of X n. CVVe know that Xn is approximately normally
distributed.) The density from which we are sampling is
n
We know (see Example 9 of Chap. V) that L Xi has a binomial distribution;
1
that is,
hence, the distribution of Xn is given by
p [ X. = :] = (~)p'r' for k = 0, I, ... , n. (18)

3 SAMPLE MEAN 237
So Xn , the sample mean of a random sample from a Bernoulli density, takes on

the values 0, lin, 21n, ... , 1 with respective binomial probabilities
(n)2 p q 2 n-2
, ... ,
If Xl' ... , Xn is a random sample from a Poisson distribution with mean

A, then L: Xi also has a Poisson distribution with parameter nA (see Example 10
of Chap. V), and hence
_ k] [ n ] e - nA( nA)k
[
P X n =- =P L:Xj=k
n ;=1
=------
k!
for k = 0, 1, 2, _.. , (19)
which gives the exact distribution of the sample mean for a sample from a
Poisson density.
3.S Exponential Distribution

Let X h X 2 , •• _, Xn be a random sample from the exponential density
I(x) = ee- 8xI(o. co)(x).
According to Example II of Chap_ V, L" Xi. has
1
a gamma distribution with
parameters nand e; that is,
1
f :EXt () -
Z - r(n) Z
,,-lene - 8z1(0, co) (Z,)
or
for y > 0,
and so
for y > 0.
Or,
nx 1
P[X" < xl = f(j r(n) z" lW'e- 9z dz
x 1
=
fo r(n)
- - (nu)"-le ne- n8u n du'
,
that is, X" has a gamma distribution with parameters n and nO.
3.6 Uniform Distribution

Let Xh ... , Xn be a random sample from a uniform distribution on the interval
(0, I]. The exact density of Xn is given by
Ix.(x) = :~: (n ~ I)! [(nxr- I

(7)(nx - 1)" + G)(nx - 2r- ' - .. ,
- I
+ ( -It(:)(nx- k)n-I] [(kin, )/nj(X).

(HI (20)
The derivation of the above (using mathematical induction and the convolution
formula) is rather tedious and is omitted. Instead let us look at the particular
cases n = I, 2, 3.
lx/x) = IX{x) = l(o,l](x),

Ix 2 (x) = 2(2x)l(o,i](x) + 2[2x - 2(2x - I) ]lu.1](x)
J4x for 0 < x ~ t
- \4(1 - x) for t < x ~ I,
and
for 0 < x <t

for t < x <i
for i < x ~ I.
lx/x), fxl x), and lx/x) are sketched in Fig. 3, and an approach to normality
can be observed. (In fact, the inflection points of I X 3(X) and of the normal
approximation occur at the same points!)
We have given the distribution of the sample mean from a uniform distri-
bution on the interval (0, I]; the distribution of the sample mean from a uniform
distribution over an arbitrary interval (a, b] can be found by transformation.
3.7 Cauchy Distribution

Let Xl' ... , Xn be a random sample from the Cauchy density
I
f(x) = n{J{1 + [(x - a)/Pf} ;
4 SAMpLiNG FROM THE NORMAL DISTRIBUTIONS 239
l~--~~--------~----
~----~-------L----~~~x
FIGURE 3
then X" has this same Cauchy distribution for any n. That is, the sample mean
has the same distribution as one of its components. We are unable to easily
verify this result. The moment-generating-function technique fails us since the
moment generating function of a Cauchy distribution does not exist. Mathe-
matical induction in conjunction with the convolution formula produces
integrations that are apt to be difficult for a nonadvanced calculus student to
perform. The result, however, is easily obtained using complex-variable
analysis. In fact, if we had defined the characteristic function of a random
variable, which is a generalization of a moment generating function, then the
above result would follow immediately from the fact that the product of the
characteristic functions of independent and identically distributed random
variables is the characteristic function of their Sum. A major advantage of
..
characteristic functions over moment generating functions is that they always
exist.
r
4 SAMPLING FROM THE NORMAL DISTRIBUTIONS
4.1 Role of the Normal Distribution in Statistics

It wi1l be found in the ensuing chapters that the normal distribution plays a very
predominant role in statistics. Of course, the central-limit theorem alone
ensures that this win be the case, but there are other almost equally important
reasons.
In the first place, many populations encountered in the course of research

in many fields seem to have a normal distribution to a good degree of approxima-
tion. It has often been argued that this phenomenon is quite reasonable in
view of the central-limit theorem. We may consider the firing of a shot at a
target as an i11ustration. The course of the projectile is affected by a great
many factors, a11 admittedly with small effect. The net deviation is the net
effect of all these factors. Suppose that the effect of each factor is an observa-
tion from some population; then the total effect is essential1y the mean of a set
of observations from a set of populations. Being of the nature of means, the
actual observed deviations might therefore be expected to be approximately
normally distributed. We do not intend to imply here that most distributions
encountered in practice are normal, for such is not the case at all, but nearly
normal distributions are encountered quite frequently.
Another consideration which favors the normal distribution is the fact
that sampling distributions based on a parent normal distribution are fairly
manageable analytically. In making inferences about populations from
samples, it is necessary to have the distributions for various functions of the
sample observations. The mathematical problem of obtaining these distribu-
tions is often easier for samples from a normal population than from any other,
and the remaining subsections of this section will be devoted to the problem of
fmding the distributions of several different functions of a random sample from
a normally distributed population.
In applying statistical methods based on the normal distribution, the
experimenter must know, at least approximately, the general form of the distri-
bution function which his data follow. If it is normal, he may use the methods
directly; if it is not, he may sometimes transform his data so that the transformed
observations follow a normal distribution. When the experimenter does not
know the form of his population distribution, then he may use other more
general but usual1y less powerful methods of analysis called nonparametric
methods. Some of these methods wi11 be presented in the final chapter of this
book.
4.2 Sample Mean
One of the simplest of all the possible functions of a random sample is the
sample mean, and for a random sample from a normal distribution the dis-
tribution (exact) of the sample mean is also normal. This result first appeared
as a special case of Example 12 in Chap. V. It is repeated here.
4 SAMPLING FROM THE NORMAL DISTRIBUTlONS 241
Theorem 6 Let Xn denote the sample mean of a random sample of size n

from a normal distribution with mean J.1 and variance 0'2. Then Xn has
2
a normal distribution with mean it and variance a /n.
PROOF To prove this theorem we shall use the moment-generating-
function technique.
= tG [ .n exp tXt]
n
,:1
-
n
=.n tXt]
tG exp-
n
n
,~\
[
= n n
l== I
mXi
( t) =
-
n
nexp [J.1t-n + 2-1 (at)
n
i= I
- 2]
n
;(0'1)2]
= exp [ J.1t + n '
which is the moment generating function of a normal distribution with

mean J.1 and variance a2 /n. 1//I
Since we have the exact distribution of Xn , in considering estimating J.1

with Xn , we will be able to calculate, for instance, the (exact) probability that
our .. estimator)) Xn is within any fixed amount of the unknown parameter J.1.
4.3 The Chi..Square Distribution

The normal distribution has two unknown parameters J.1 and 0'2. In the previous
subsection we found the distribution of X n , which" estimates" the unknown J.1.
In this subsection, we seek the distribution of
which estimates" the unknown 0'2. A density function which plays a central
H
role in the derivation of the distribution of 8 2 is the chi-square distribution.
Definition 8 Chi..square distribution If X is a random variable with

density
(21)
then X is defined to have a chi-square distribution with k degrees offreedom;

or the density given in Eq. (21) is called a chi-square density with k degrees
offreedom, where the parameter k, called the degrees offreedom, is a posi-
tive integer. IIII
Remark We note that a chi-square density is a particular case of a
gamma density with gamma parameters r and A equal, respectively, to
kl2 and t. Hence, if a random variable X has a chi-square distribution,
tS'[X] = kl2 = k,
t
X] - kl2 - 2k (22)
var [ - (1/2)2 - ,
and
_[_t]k12 _ [1
~2
1 ]k12 t < 1/2. (23)
mX<t) - - t - - 2t '
1111
Theorem 7 If the random variables Xi' i = 1,2, ... , k, are normal1y and
independently distributed with means f.11 and variances af, then
U= Ik (X.' _11.)2 Fl
j=l (Ii
has a chi-square distribution with k degrees of freedom.

PROOF Write Zi = (Xl - f.1i)l(Ii; then Zi has a standard normal
distribution. Now
mu(t) = tS'[exp tV] = tS'[exp(t I zf)]
= ,,[;v.exp tZ;] = lI/[ex ptZn

But
oo
-1e -t(1-2t)z 2 d z
=
f -00 J2n
=
1 foo /1-2t
e -!(1-2t)Z2 d z
)1 - 2t -00 J21C
1 1
for t< 2.'
/1 - 2t
4 SAMPLING FROM THE NORMAL DISTRIBUTIONS 243
the ]a tter integral being unity since it represents the area under a norma]
curve with variance 1/(1 - 2t). Hence,
k 1 = (1
n0"[exp tZlJ = n J 1-2t k
1 2t
)k12
for t<
1
2:'
i=l -
i=l
the moment generating function of a chi-square distribution with k

degrees of freed om. fill
Corollary If Xl, ... , Xn .is a random samp]e from a norma] distribution
n
with mean J.1 and variance (12, then U = L (Xi - J.1)2/(12 has a chi-square
i= 1
distribution with n degrees of freedom. fill
We might note that if either J.1 or (12 is unknown, the U in the above
coro]]ary is not'a statistic. On the other hand, if J.1 is known and (12 is unknown,
we could estimate ,,2 with (I fn) it. (X, - li)2 (note that .c [(I In),t, (X, - JI)2] =
(lln),t. C[(X, - JI)2] = (1/n),t, ,,2 = ,,2), and find the distribution of
n
(1/n) L (Xi - J.1)2 by using the coro]]ary.
i= 1
Remark In words, Theorem 7 says, "the sum of the squares of inde-

pendent standard normal random variables has a chi-square distribution
with degrees of freedom equa] to the number of terms in the sum." fill
Theorem 8 If Zl' Z2' ... , Zn is a random samp]e from a standard

normal distribution, then:
(i) Z has a norma] distribution with mean Qand variance lin.
n
(i~) Z and L (Zi - 2)2 are independent.
i= 1
n
(iii) L (Zi - Z)2 has a chi-square distribution with n - I degrees
i= 1
of freedom.
PROOF (Our proof wi]] be incomp]ete.) (i) is a special case of
"-
Theorem 6. We wi11 prove (ii) for the case n = 2. If n = 2,
- Zl + Z2
Z= 2
and
L(Z, - 2)2 (Z. _Z. ;Z2)' + (Z2 _Z. ; z2f

=
(ZI - Z2)2 (Z2 - ZI)2

= 4 + 4
(Z2 - ZI)2 .
-
2 '
so 2 is a function of ZI + Z2, and L (Zi - 2)2 is a function of Z2 - ZI;

so to prove 2 and L (Zj - Z)2 are independent, it suffices to show that
ZI + Z2 and Z2 - ZI are independent. Now
and, similarly,
mz Z -Z 1(t 2 ) = exp t~.
Also,
mZ +zz, ZZ-Zl(t It t 2 )
1
= G[e (Zl +zz)+tz(Zz-Zt)]
1
= G[e(h -tz) Zl e(tl +tz)ZZ] = G[e(tt-t z )zJ]G[e(t l +t 1)Z2]
= et(t1-tZ)Zetch +tZ)2 = exp tf exp t~

= mZ 1 +Z z(tl)mZ z -Z t (t2);
and since the joint moment generating function factors into the produc~
of the marginal moment generating functions, ZI + Z2 and Z2 - ZI are
independent.
n
To prove (iii), we accept the independence of Z and L (Zj -
1
2}2 for
arbitrary n. Let us note that L zf = L (Z, - 2 + Zf = L (Zl - 2)2 +
22 L(Zj - Z) + L Z2 = L(Zj - 2)2 + nZ2; also L(Zi - zf and nZ2
are independent; hence
So,
t < 1/2
4 SAMPLING FROM THE NOR.MAL DISTRIBUTIONS 245
noting that J~Z has a standard normal distribution implying that nZ/
has a chi-square distribution with one degree of freedom. We have
shown that the moment generating function of L (Zi - Z)2 is that of a
chi-square distribution with n - 1 degrees of freedom, which completes
~~~ M
Theorem 8 was stated for a random sample from a standard normal dis-
tribution, whereas if we wish to make inferences about 11 and a 2 , our sample
is from a normal distribution with mean p and variance a 2 • Let Xl' ... , Xn
denote the sample from the norma] distribution with mean Il and variance a 2 ;
then the Zi of Theorem 8 could be taken equal to (Xi - II)la.
(i) of Theorem 8 becomes:
(i') Z = (lIn) L (X, - Il)la = (X - tt)/a has a normal distribution with
mean 0 and variance 1In.
(H) of Theorem 8 becomes:
(ii') Z = (X - p)la and L(Zi - Z)2 = L [(Xi - p)la - (X - p)laf =
L [(Xi - %)2/a2] are independent, which implies X and L (Xi - %)2 are
independent.
(iii) of Theorem 8 becomes:
(iii') L (Zi - Z)2 = L [(Xi - X)2/a 2 ] has a chi-square distribution with
n - 1 degrees of freedom.
n
Corollary If 8 2 = [l/(n - 1)] I (Xi - X)2 is the sample variance of a
i= I
random sample from a normal distribution with mean Il and variance (r2,
then
",
(24)
has a chi-square distribution with n - 1 degrees of freedom.

PROOF This is just (iii'). IIII
Remark Since 8 2 is a linear function of U in Eq. (24), the density of 8 2
can b~ obtained from the density of U. It is
1
f~p(y) = ( n - 2 1)(n-1)/2 . y(n-3)/2 e -(n-l)y/2a j
2 (y) (25)
2a r[(n _ 1)/2] (0. cO) •
IIII
246 sAMPLING AND SAMPLING DISTRIBUTIONS VI
Remark The phrase "degrees of freedom" can refer to the number of

independent squares in the sum. For example, the Sum of Theorem 7 has
k independent squares, but the sum in (iii) of Theorem 8 has only n - I
independent terms since the relation L (Zi - Z) = 0 enables one to com-
pute anyone of the deviations Zj - Z, given the other n - I of them. IIII
All the results of this section apply only to normal populations. In fact,
it can be proved that for no other distributions (i) are the sample mean and
sample variance independently distributed or (ii) is the sample mean exactly
normally distributed.
4.4 The F Distribution

A distribution, the F distribution, which we shall later find to be of considerable
practical interest, is the distribution of the ratio of two independent chi-square
random variables divided by their respective degrees of freedom. We suppose
that U and V are independently distributed with chi-square distributions with
m and n degrees of freedom, respectively. Their joint density is then [see
Eq. (21)]
_ 1 (m-2)/2 (n-2)/2 -!(u+v)

Iv. v(u, v) - r(m/2)r(n/2)2(m+n)/2 u v e 1(0, oo)(u)/(o. oo)(v).
(26)
We shall find the distribution of the quantity
X- Ulm (27)
- Vln'
which is sometimes referred to as the variance ratio. To find the dIstribution
of X, we make the transformation X = (Ulm)/(Vln) and Y = V, obtain the
joint distribution of X and Y, and then get the marginal distribution of X by
integrating out the y variable. The Jacobian of the transformation is (mln)y; so
r ( ) _ m 1 (m x )(m-2)/2 (n-2)/2 e -H(m/n)xy+y]

JX, y x, Y - n y r(m/2)r(n/2)2(m+n)/2 n y y ,
and
00
fx(x) = fix, y(x, y) dy

o
_ 1 (m)m/2 x(m-2)/2 feCI (m+n-2)/2 e -H(m/n)x+ IJy dy
- r(m/2)r(n/2)2(m+n)/2 n 0 y
r[(m + n)/2] (m)m/2 x(m-2)/2

= r(m/2)r(n/2) ~ [1 + (m/n)x](m+n)/2 1(0,. eCllx ). (28)
Definition 9 F distribution If X is a random variable having density

given by Eq. (28), then X is defined to be an F-distributed random variable
with degrees offreedom m and n. jjjj
The order in which the degrees of freedom are given is important since the
density of the F distribution is not symmetrical in m and n. The number of
degrees of freedom of the numerator of the ratio mjn that appears in Eq. (28)
is always quoted first. Or if the F-distributed random variable is a ratio of two
independent chi-square-distributed random variables divided by their respective
degrees of freedom, as in the derivation above, then the degrees of freedom of the
chi-square random variable that appears in the numerator are always quoted
first.
We have proved the following theorem.
Theorem 9 Let U be a chi-square random variable with m degrees of

freedom; let V be a chi-square random variable with n degrees of freedom,
and let U and V be independent. Then the random variable
X= Ujm
Vjn
is distributed as an F distribution with m and n degrees of freedom. The

density of X is given in Eq. (28). jjjj
The following corollary shows how the result of Theorem 9 can be useful
in sampling.
Corollary If Xl' ... , X m + 1 is a random sample of size m + I from a

normal population with mean I1x and variance a 2 , if Y1 , •.. , Yn + 1 is a
random sample of size n + I from a normal population with mean Jiy and
variance a 2 , and if the two samples are independent, then it follows
m+l
that (lj( 2
) L (Xi - X)2 is chi-square distributed with m degrees of
1
n+l
freedom, and (1 j( 2) L (Yj - y)2 is chi-square-distributed with n degrees
1
of freedom; so that the statistic
L (Xi - X)2jm
L (Yj - y)2jn
has an F distribution with m and n degrees of freedom. jjjj

248 sAMPLING AND SAMPLING DISTRlBunoNS VI
We close this subsection with several further remarks about the F dis-
tribution.
Remark If X is an F-distributed random variable with m and n degrees

of freedom, then
$[X] = n for n > 2

n-2
and
2n2(m + n - 2)
var [X] = ---::---..;..- fer n > 4. (29)
men - 2)2(n - 4)
PROOF At first it might be surprising that the mean depends only

on the degrees of freedom of the denominator. Write X as in Eq. (27);
that is,
X = Ulm.
Vln'
then
$[X] = $ [-Ulm]
Vln
n
=-$[U]$
m
- [1]
V
.
But $[U] = m by Eq. (22), and
$ [~] = 1 (!)TI/2 J"oo ! . v(TI-2)/2e-tv dv

V r(n/2) 2 0 v·
= 1 (~)TlI2Joov(n-4)/2 e-tv dv
r(n/2) 2 0
= r[(n -
2
2)/2] (!)nI2(!) -(,,-2)/ = 1 ;
r(n/2) 2 2 n- 2
and so
$[X] = ( mn) $[U]$ [1] =

V
n m
m n - 2 = n - 2.
n
The variance formula is similarly derived. 1II1
Remark If X has an F distribution with m and n degrees of freedom,

then 11 X has an F distribution with nand m degrees of freedom. This
result allows one to table the F distribution for the
,
upper tail only. For
example, if the quantile '.95 is given for an F distribution with m and n

degrees of freedom, then the quantile '~05 for an F distribution with nand
m degrees of freedom is given by 1/'.95' In general, if X has an F dis-
tribution with m and n degrees of freedom and Y has an F distribution
with nand m degrees of freedom, then the pth quantile point of X, 'p,
is the reciprocal of the (1 - p)th quantile point of Y, ,'l-P' as the following
shows: .
but
1 - p = P[ Y < l;~ - p] ;
so
I
I
'l-P = 'p' 1/11
Remark If X is an F-distributed random variable with m and n degrees

of freedom, then
W= mXln
1 + mXln
has a beta density with parameters a = m/2 and b = n/2. //1/
4.5 Student's t Distribution

Another distribution of considerable practical importance is that of the ratio of
a standard normally distributed random variable to the square root of an in-
dependently distributed chi.square random variable divided by its degrees of
freedom. That is, if Z has a standard normal distribution, if U has a chi-
square distribution with k degrees of freedom, and if Z and U are independent,
we seek the distribution of
Z
X= .....
jUlk
The joint density of Z and U is given by
(30)
,." ,
If we make the transformation X = zljuii and Y = U) the Jacobian is

jYlk, and so
_
Ix. y(x, y) - If 1
~ __ 1 (l)kf2
_
'k j2n r(k/2) 2 y
00
(kf2)-1 -'t.v -txly/k
e e /(o.oo)(Y)
Ix(x) = fIx. y(X, y) dy

-00
= fon r(!/2) (~r/2 foro y'12-1+te-t(1+X'I')Y dy
r[(k + 1)/2] 1 1
- (31)
r(k/2) jkn (1 + x 2 /k)(k+ 1)/2'
Definition 10 Student's t distribution If X is a random variable having

density given by Eq. (31), then X is defined to have a Student's t distribu-
tion, or the density given in Eq. (31) is called a Student's t distribution
with k degrees offreedom. IIII
We have derived the following result.
Theorem 10 If Z has a standard normal distribution, if U has a chi-

square distribution with k degrees of freedom, and if Z and U are inde-
pendent, then Z/j Ulk has a Student's t distribution with k degrees of
freedom. IIII
The following corollary shows how the result of Theorem lOis applicable
to sampling from a normal popUlation.
Corollary If Xl' ... , Xn is a random sample from a norm~l distribution

with mean fl and variance 0'2, then Z = (X - fl)/(O'IJn) has a standard
t
normal distribution and U = (1/0'2) (Xi - xf has a chi-square distribu-
tion with n - 1 degrees of freedom. Furthermore, Z and U are inde-
pendent.(see Theorem 8); hence
(X - fl)/(O'/j~) _ jn(n - I)(X - fl)

j(1/O' 2 ) I (Xi - Xf/(n - I) j I (Xi - 1)2
has a Student's t distribution with n - 1 degrees of freedom. / /11
5 ORDER STATISTICS 251
We might note that for one degree of freedom the Student's t distribution
reduces to a Cauchy distribution; and as the number of degrees of freedom
increases, the Student's t distribution approaches the standard normal distribu-
tion. Also, the square of a Student's t-distributed random variable with k
degrees of freedom has an F distribution with I and k degrees of freedom.
Remark If X is a random variable having a Student's t distribution with

k degrees of freedom, then
C[X] =0 if k > 1 and var [X] '= kl(k - 2) if k > 2. (32)
PROOF The first two moments of X can be found by writing

X = ZIJUlk as in Theorem 10 and using the independence of Z and U.
The actual derivation is left as an exercise. IIII
This completes Sec. 4 on sampling from the normal distribution. Note

that we considered the distribution of functions of only two different statistics,
namely, the sample mean and sample variance. In the next chapter we will find
that these two statistics are the only ones of interest in sampling from a normal
distribution; they will turn out to be sufficient statistics.
5 ORDER STATISTICS
5.1 Definition and Distributions

In Subsec. 2.4 we defined what we meant by statistic and then gave the sample
moments as examples of easy-to-understand statistics. In this section the
concept of order statistics will be defined, and some of their properties will be
investigated. Order statistics, like sample. moments, play an important role
in statistical inference. Order statistics are to population quantiles as sample
moments are to population moments.
Definition 11 Order statistics Let Xl' X 2 , ••• , Xn denote a random

sample of size n from a cumulative distribution function F(')' Then
Yt < Y2 < ... < Yn , where Y i are the Xi arranged in order of increasing
magnitudes and are defined to be the order statistics corresponding to
the random sample Xl' .. " Xn • IIII
We note that the Yj are statistics (they are functions of the random sample
Xh X 2 , ••• , Xn) and are in order. Unlike the random sample itself, the order
statistics are clearly not independent, for if Y j > y, then Yj + l > y.
We seek the distribution, both marginal and joint, of the order statistics.
We have already found the marginal distributions of YI = min [Xl' ... , Xn] and
Yn = max [Xl' ... , Xn] in Chap. V. Now we will find the marginal cumulative
distribution of an arbitrary order statistic.
Theorem 11 Let YI < Y 2 < · .. ~ Yn represent the order statistics from

a cumulative distribution function F( . ). The marginal cumulative distri-
bution function of Y a , a = 1, 2, ... , n, is given by
F Yr.,(y) = ±(~)[F(y)]j[l
)=tZ J
- F(y)r- j (33)
PROOF For fixed y, let
then
n
i
2:= 1Zi = the number of Xi < y.
n
Note that 2: Zj has a binomial distribution with parameters nand F(y).
i= I
Now
.
FyJy) = P[Ya < y] = P[2: Zi > a] = f
}=tZ
(~)[F(y)]j[l -
J
F(y)r- i .
The key step in the proof is the equivalence of the two events {Ya < y}
and {2: Zj > a}. If the ath order statistic is less than or equal to y, then
surely the number of Xi less than or equal to y is greater than or equal to
a, and conversely. 1111
Corollary Fy"(y) = .f (n.)J [F(y)J1[l -

J=n
j
F(y)r- = [F(Y)r,
and
Fy,(Y) = J. C)[F(y)]i[l - F(Y)r- i = 1 - [1 - F(y)]". /III

Theorem 11 gives the marginal distribution of an individual order statistic

in terms of the cumulative distribution function F(')' For the remainder of
this subsection, we will assume that our random sample Xl' ... , Xn came from a
probability density function f('); that is, we assume that the random variables
Xi are continuous. We seek the density of Y a , which, of course, could be
obtained from Eq. (33) by differentiation of Fy"'(y). Note that
fy,/y)
= lim Fy,.{y + Ay) - Fy",(Y) = lim P[y < Ya y + Ay]
A,.-O Ay A,.-O Ay
= lim P[(a: -l)ofthe Xi ~ y; one Xi in(y,y + Ay];(n - a:) of the Xi> y + Ay]
A,.-O Ay
= lim f n! [F(y)r-l[F(y + Ay) - F(y)][1 - F(y + Ay)]n-a}
A,.-O \( a: - 1) U !(n - a:)! Ay
n'
- (a: _ l)!('n _ a:)! [F(y)]a-l[1 - F(y)]n-af(y)·
We have made sensible use of the multinomial distribution. Similarly, we can

derive the joint density of Ya and Yp for 1 < a: Y + Ay]
nl
~----~-----------------
(a: - I)! 1 t(P - a: - I)! l!(n - {I)!
x [F(x)]a-l[F(y) - F(x + Ax)]fl-a-l[1 - F(y + Ay)]n-Pf(x) Axf(y) Ay;
hence
fy",. Y,(X' y) =
(a: - 1)1 (P
n'
_ a: '_ I)! (n _ P)![F(x)]a-l[F(Y) - F(x)]p-a-l[l - F(y)]n-Pf(x)f(y)
for x < y, and
fYt%. yiX, y) = 0 for x >y.

In general, fy 10''', y,,(y l' . . . ,YJ

1
= lim P[YI < Y 1 < Yl + AYl; ... ; Yn < Y n < Yn + AYn]
Ayr-+O
i= 1
nAYi
n
1
= lim P[one Xi in (y., Yl + AYl]; . " ; one Xi in (Yn, Yn + AYn]]
AYt-+
O
i =1
n
n
AYI
n!
lim [F(y 1 + Ay 1) - F(y 1)]· .... [F(Yn + AYn) - F(Yn)]
=
AYt-+O
i
n=
n
1
AYI
for Yl < Y2 < ... < Yn'

andfYl, ... , Yn(Yl"'" Yn) = 0, otherwise.
We have derived the following theorem.
Theorem 12 Let Xl' X 2 , ••• , Xn be a random sample from the prob-

ability density function f(') with cumulative distribution function F(·).
Let Y1 < Y2 < ... < Yn denote the corresponding order statistics; then
fyJy) = (a _ 1)! en
n' _ a)! [F(y)]'~-I[1 - F(y)]n-'i(Y); (34)
n!
fy"" YfJ(x, y) = '(a - l)! (P - a - 1)! (n - P) !
x [F(x)]a-l[F(y) - F(x)]p-a-l
x [1 - F(y)]n-Pf(x)f(y)I(x. oo)(Y); (35)
fYI •.... Yn(Yl, ... , Yn)
= {~!fly,) ..... f(y.) for Yl < Y2 < . . . < Yn (36)
otherwise. IIII
Any set of marginal densities can be obtained from the joint density
fyIo .... Yn(Yl' ... , Yn) by simply integrating out the unwanted variables.
5.2 Distribution of Functions of Order Statistics

In the previous subsection we derived the joint and marginal distributions of the
order statistics themselves. In this subsection we will find the distribution of
certain functions of the order statistics. One possible function of the order
statistics is their arithmetic mean, equal to
1 n
- L Yj •
n j= 1
II II
Note, however, th~t (lIn) L Yj = (lin) L Xi' the sample mean, which was the
j=l" i=l
subject of Sec. 3 of this chapter, We define now some other functions of the
order statistics,
Definition 12 Sample median, sample range, sample midrange Let

Yl < ... ~ Yn denote the order statistics of a random sample Xl' " " Xn
from a density f( . ), The sample median is defined to be the middle order
statistic if n is odd and the average of the middle two order statistics if n
is eye..!!, The sample range is defined to be Yn - Yl , and the sample mid-
range is defined to be (Y} + Yn)/2, IIII
If the sample SIze . IS,~t'3
odd,t., then t h e d'Istn'b'
I
ution 0 f t h Id
e samp e me 'Ian IS
,
given by Eq. (34); for example, if n = 2k + I, where k is some positive integer,
then" Yk + l is the sample median whose distribution is given by Eq, (34). If the
sample size is even, say n = 2k, then the sample median is (Yk + Yk + l )/2, the
distribution of which can be obtained by a transformation starting with the
joint density of Yk and Yk + l , which is given by Eq. (35),
We derive now the joint distribution of the sample range and midrange,
from which the marginals can be obtained.
By Eq, (35), we have
fYl. y •.(x, y) = n(n - 1)[F(y) - F(x)]n-2f(x)f(y) for x < y, (37)
Make the transformation R = YII - Yl and T = (Yl + Yn )/2 or r = y - x and
t = (x + Y)/2. Now x = t - r12, and y = t + r12; hence
ax ax
-
J=
ar at
-
-! 1
= -1,
! 1
ay ay
-
ar at
and we obtain Theorem 13.
Theorem 13 If R is the sample range and T the sample midrange from

a probability density function, then their joint distribution is given by
fR, r(r, t) =
n(n - 1)[F(t + rI2)- F(t - rI2)]n-2f(t - rI2)f(t
,
+ r12) for r > 0, (38)
and the marginal distributions are given by
and fT(t) = foo fR, T(r, t) dr. (39)

o
IIII
EXAMPLE 6 Let Xb ... , Xn be a random sample from a uniform distribu-

tion on (Jl - j3a, Jl + j3a). Here Jl is the mean, and a2 is the variance
of the sampled population.
1
I(x) = 2)30' 1(p._..,I3(1, p.+..,I3(1)(X),
J
(40)
JR(r) = ffR. T(r, t) dt = n(J- 1)

(2 3ay
,.-2 (2J3U - r)l(o 2,130)(r).
,
(41)
We note that IR(r) is independent of the parametero Jl.

fT(t) = ffR. r(r, t) dr
n(n - 1)
= (2j30't f0
min[2t-2(p.-..,I30'),2(p.+..,I3(1)-2t]
rn - 2 dr . I ~ - (t)
(p.-..,I30'.p.+..,I3(1) ,
which simplifies to
Jr(t) = 2:/Ju CJ 3: + 1r'I(-I.o{/3:) +

2j3U (1 - tJ3 1[0'I)C3:)' :r' (42)
From Eq. (41), we can derive 8[R] = 2j3a(n - l)/(n + 1). /1//
Certain functions of the order statistics are again statistics and may be used
to make statistical inferences. For example, both the sample median and the
midrange can be used to estimate Jl, the mean of the population. For the uni-
form density given in the above example, the variances of the sample mean, the
sample median, and the sample midrange are compared in Problem 33.
5.3 Asymptotic Distributions

In Subsec. 3.3, we discussed the asymptotic distribution of the sample mean
X n. We saw that Xn was asymptotically normally distributed with mean Jl and
variance a2 /n. We now consider the question: Is there an asymptotic dis-
tribution for the sample median? We will state (without proof) a more general
result.
s ORDER STATISTICS 257
Since for asymptotic results the sample size n increases, we let yin) <
Y!IJ) < ... < y!IJ) denote the order statistics for a sample of size n. The super-
script denotes the sample size. We will give the asymptotic distribution of that
order statistic which is approximately the (np)th order statistic for a sample of
size n for any 0 <p < I. We say "approximately" the (np)th order statistic
because np may not be an integer. Define Pn to be such that nPn is an integer
and Pn is approximately equal to P; then Y!;~ is the (nPn)th order statistic for a
sample of size n. (If Xl' ... , Xn are independent for each positive integer n, we
will say Xl' ... , X n , ... are independent.)
Theorem 14 Let Xl' ... , Xn , ••• be independent identically distributed

random variables with common probability density f(') and cumulative
distribution function F(·). Assume that F(x) is strictly monotone for
e
0< F(x) < 1. Let p be the unique solution in x of F(x) = P for some
0< P < 1. ~ep is the pth quantile.) Let Pn be such that nPn is an integer
and nlPn - pi is bounded. Finally, let Y;;~ denote the (npn)th order
statistic for a random sample of size n. Then Yn~~ is asymptotically
distributed as a normal distribution with mean e,p and vanance
p(I - p)ln[f(ep)p. IIII
EXAMPLE 7 Let P = -!; then ep is the population median, and Theorem 14

states that the sample median is asymptotically distributed as a normal
distribution with mean the popUlation median and variance 1/4n[f(e,1/2)]2.
In particular, if f( .) is a normal density with mean fl and variance a 2 ,
then the sample median is asymptotically normally distributed with
mean fl and variance 1/4n[f(fl)f = TCa 2 /2n. Recall that the sample mean
is normally distributed with mean fl and variance (J2In. 11/1
In Theorem 14 above we considered a certain kind of asymptotic distribu-

tion of order statistics. We will now consider yet another kind. In the above
we looked at the asymptotic distribution of that order statistic which was
approximately the (np)th order statistic for a sample of size n. Such an order
statistic had (approximately) lOOp percent of the n observations to its left.
That is, its relative position remained unchanged as n, the sample size, in-
creased; it always had (approximately) the same percentage of the n observations
to its left. We will now consider the asymptotic distribution of that order
statistic whose absolute position remains unchanged. That is, we con-
sider the asymptotic distribution of, say, Yin) for fixed k and increasing n.
yt) is the kth smallest order statistic for a sample size n > k, and k remains
fixed. In order to make the presentation somewhat simpler, we will take k = 1,
.258 SAMPUNO AND SANPUNO DISTRIBUTIONS VI
in which case Y1 n ) is the smallest of the n observations. We note that we could

just as well consider the kth largest order statistic, namely y~=:?k+" which for
k = 1 specializes to y~n), the largest order statistic for a sample of size n. Either
y~n) or y~n) is often referred to as an extreme-value statistic.
Practical applications of extreme-value statistics are many. The old
adage that a chain is no stronger than its weakest link provides a simple example.
If Xi denotes the" strength" of the ith link of a chain with n similar links, then
y~n) = min [Xb "', X,.] is the" strength" of the chain. Also, in measuring the
results of certain physical phenomena such as floods, droughts, earthquakes,
winds, temperatures, etc., it can be seen that under certain circumstances one is
more interested in extreme values than in average values. For instance, it is
the extreme earthquake or flood, and not the average earthquake or flood, that
is more damaging. We can see that results, whether exact or asymptotic, for
extreme-value statistics Can be just as important as results for averages.
For the most part we will concentrate on finding the asymptotic distribu-
tion of y~,.). One might wonder why we should be interested in an asymptotic
distribution of y~n) when the exact distribution, which is given by Fyn(n)(Y) =
[F(y)]n, where F(') is the c.d.f. sampled from, is known. The hope is that we
will find an asymptotic distribution which does not depend on the sampled
c.d.f. F(')' We recall that the central-limit theorem gave an asymptotic dis-
tribution for Xn which did not depend on the sampled distribution even though
the exact distribution of X,. could be found.
In searchi~g for the asymptotic distribution of y!n), let us pattern our
development after what was done in deriving the asymptotic distribution of Xn •
According to the law oflarge numbers, Xn has a degenerate limiting distribution;
that is, the limiting c.d.f. of Xn is the cumulative distribution that assigns all its
maSS to the point /1. Such a limiting distribution is not useful if one intends to
use the .limiting distribution to approximate probabilities of events since it
assigns each event a probability of either 0 or 1. To circumvent such difficulty,
we first" centered" the values of Xn by subtracting /1, and then we " inflated"
the values of Xn - /1 by mUltiplying them by J~IO', and, consequently, we were
able to get a non degenerate limiting distribution; that is, accordingtothecentral-
limit theorem, J~( Xn - /1)/0' had a standard normal distribution as its limiting
distribution. A general procedure, when one is looking for a limiting distribu-
tion of, say, Z,. , is to first'" center" the Zn by subtracting a constant, sayan, and
to then" scale" Zn - an by dividing by another constant, say bn . In the case of
=
the central-limit theorem, Zn = Xn, an /1, and bn = O'/J~. In the case of
= J
Theorem 14 above, Zn = Y~;~, an e p, and bn = p(l - p)/n[f(ep)f. Por both
of these two Cases the sequence of constants {an} did not depend on n. In the
Case at hand, namely when Zn = y!n), the sequence of constants {on} is likely to
depend on n since y!n) tends to increase with n. Let us look at a couple of

examples.
EXAMPLE 8 Consider sampling from the logistic distribution; that is,

F(x) = (1 + e-~-l. Find the limiting distribution of (y~n) - an)/bn .
There are two problems: First, what should we take the sequences of
constants {an} and {bn} to be? And, second, what is the limiting distribu~
tion of (y~n) - an)/bn for the selected constants {an} and {bn}? It seems
reasonable that the "centering" constants {an} should be close to c[y~n)];
so we seek an approximation to C[ y~n)]. Now F(Xd, ... , F(Xn) is a
random sample from the uniform distribution over (0, 1); hence F( y~n»)
is the largest of a sample of size n from a uniform distribution over (0, 1).
ThatC[F(y~n»)] = n/(n + 1) can then be routinely derived. Now F(c[y~n)]) ~
C[F( y~n»)] or
F(c[y~n)]) = + exp (_c[y~n)])}-1
{l
= 1 - {l + exp (C[ y~n)])}-l
~ 1 -en + 1)-1
n
---
n +1
= C[F( y~n»)],
which implies that
n ~ exp (C[ y~n»)]
or that
C[ y~n)] ~ loge n.
Finally, since C[ y~n)] ~ log n (from here on we use log n for loge n), a
reasonable choice for the sequence of "centering" constants {an} seems
to be the sequence {log n}. Weare seeking
limp[y~n) - an < Y] = 1imp[y~n) -log n < Y]

n-+oo bn n-+oo bn
= 1imP[Y~") < bny + log n]
n-+oo
= lim [F(bny + log n)]n

n-+oo
= lim(l + e-bnY-logn)-n
n-+oo
= lim(} + (l/n)e-bny)-n
n-+oo
Hence, if {an} and {bn} are selected so that {an} = {log n} and {bn} = {l},
respectively, then the limiting distribution of (y~n) - an)lbn = y~n) - log n
is exp ( - e - Y). IIII
EXAMPLE 9 Consider sampling from the exponential distribution so that

F(x) = (l - e-).x)I(o, oo)(x). Again, let us find the limiting distribution of
(y~n) _ an)lbn . As in Example 8, tC[F( y~n»] = nl(n + 1) = 1 -l/(n + 1).
Now
and
so
1
n + I ~ exp { - AtC[ y~n)]},
or
] 1
tC[ y!n)] ~ J log (n + 1) ~ J log n.
Hence, it seems reasonable to use an = (IIA) log n.
Y] = limp[Yn - !A log n < b Y] n
r
n ..... oo
= !~n;, [F(b n Y+ ~ log n)

= lim(l - e-;'bny-logn)"
= lim
' ( 1 - -1 e-;'b
,. ..... 00 n
l1 y )n
Hence the limiting distribution of(y!n) - an)lbn = [y~n) - (1IA) log n]/O/A)
is exp (-e- Y ). We note that we obtained the same limiting distribution
here as in Example 8. Here we were sampling from an exponential
distribution, and there we were sampling from a logistic distribution.
IIII
In each of the above two examples we were able to obtain the limiting
distribution of (y~n) - Qn)lbn by using the exact distribution of y!n) and ordinary
algebraic manipulation. There are some rather powerful theoretical results
concerning extreme-value statistics that tell us, among other things, what limiting
distributions we can expect. We can only sketch such results here. The
interested reader is referred to Refs. 13, 30, and 35.
Theorem 15 Let X., ... , X n , ••• be independent and identically dis-

tributed random variables with c.dJ. F('). If (y!n) - an)lbn has a limiting
distribution, then that limiting distribution must be one of the following
three types:
Gt(y; y) = e- Y - Y /(0, oo)(Y), where y > O.

G2 (y; y) = e-1Y1Y/(-OO,O)(Y) + /[0, oo)(Y), where y > O.
G3(Y) = ex P ( - e - Y). 1111
Theorem 15 states what types of limiting distributions can be expected.

The following theorem gives conditions on the sampled F(') that enable us to
determine which of the three types of limiting distributions correspond to the
sampled F(·).
Theorem 16 Let X., ... , Xn , ••• be independent and identically dis-

tributed random variables with c.d.f. F(·). Assume that (y~n) - an)lbn
has a limiting distribution. The limiting distribution is:
(i) G1(. ; y) if and only if

. 1 - F(x)
11m = rY for every r > O.
x-+oo 1 - F( rx)
(ii) G2 (·; y) if and only if there exists an Xo such that
and F(xo - E) < 1 for every E > O.

and
lim
1 - F(xo - rx) = rY
for every r > O.
o<x-+o I - F(xo - x)
(iii) G 3 (·) if and only if
lim n[I - F(Pnx + an)] = e- X

for each x,
n-+oo
where
n - 1 }
Ci" = inf {z: n < F(z)
and
Pn = inf {z: 1 - (ne)-l < F(a" + z)}. fill
Note that if F(') is strictly monotone and continuous, then an is given by

F(an) = (n - 1)/n, or an = F-1({n - 1}/n); and P" is given by F(a" + P,,) = 1 -
(ne)-I, or P" = F- 1(1 - {ne} -1) - a" = F- 1(1 - {ne} -1) - F- 1({n - I}/n),
EXAMPLE 10 Take F(x) = (1 - e-1j/(o. oo)(x) as in Example 9. a" is such

that F(a,,) = (n - 1)/n or 1 - e- lrzn = 1 - l/n, which implies that a" =
(1/l) log n. p" is such that F(Ci" + P,,) = I - (ne) -lor 1 _ e-1(1/1) log n-1Pn
= 1 - (ne)-l, or p" = I/l.
for each x,
so, as we saw in Example 9, the exponential distribution has G 3 (') as its

corresponding limiting extreme-value distribution, fill
EXAMPLE 11 Take F(x) = F(x; y) = [1 - (l - x)yl/(o. 1 )(x) + /[1. oolx). Note

that for Xo = I, F(xo) = 1 and F(xo - e) < 1 for every e > O. Also,
, 1 - F(xo - 1'x) I' (1 - Xo + 1'x)Y "

I1m = 1m =1">
o<x"'o 1 F(xo - x) O<X"'O (1- Xo + xY ,
so F('; y) has G 2(·; y) as its limiting extreme-value distribution. fill
EXAMPLE 12 Take F(x) = F(x; y) the c.d.f. of a t distribution with y degrees

of freedom.
. 1 - F(x) _ I' f(x) . [1 + (1'X)2/y](Y+ 0/2 _ y.

hm - 1m = 11m 2 (1'+ 1)/ 2 - l' ,
x"'oo 1 - F(rx) X"'oo 1'f(1'x) x"'oo 1'(1 + x /y)
so the t distribution with y degrees of freedom has GI (' ; y) as its limiting

extreme-value distribution. fill
Theorem 16 gives conditions on the sampled c.d.f. F(') that enable us to

determine the proper limiting extreme-value distribution for (y~n) - an)jbn · The
theorem does not tell us what the constants {an} and {b n } should be. If, how-
ever, the conditions for the third type are satisfied, then we have
n[l - F(Pnx + an)] ~ e- x , as n ~ 00,
and
p [Y!'~: a. < x] --+ exp (-e-,,).
Now, p[(y~n) - a,.)jbn <x] = [F(bnx + an))"; hence

[F(bnx + an)]n ~ exp (-e X),
or
or
n[l - F(bnx + a,.)] -+ e- x ;
and we see that a,. can be taken equal to ctn and bn = Pn . Thus, for the third
type the constants {an} and {bn} are actually determined by the condition for that
type. We shall see below that for certain practical applications it is possible to
estimate {an} and {bn}.
Since the types G1(. ; y) and G 2 (' ; y) both contain a parameter, it can be
surmised that the third type G 3(') is more convenient than the other two in
applications. Also, G 3 (y) = exp (-e-)') is the correct limiting extreme-value
distribution for a number of families of distributions. We saw that it was
correct for the logistic and exponential distributions in Examples 8 and 9; it is
also correct for the gamma and normal distributions. What is often done in
practice is to assume that the sampled distribution F(') is such that exp (-e- Y)
is the proper limiting extreme-value distribution; one can do this without assum-
ing exactly which parametric family the sampled distribution F(') belongs to.
One then knows that p[(y~n) - an)jbn ~ y] ~ exp (-e-y) for every y as n ~ 00.
Hence,
for large fixed n. Or
or
It is true that and bn are given in terms of the (1 - 1/n)th quantile and the
On
(1 - 1/ne)th quantile of the sampled distribution; however, for certain applica-

tions they can be estimated, in which case we would have an approximate distri-
bution for y!n), a distribution that is valid for a variety of different distributions
that could be sampled from. (One might note that in applications of the central-
limit theorem, which states that Xn is approximately distributed as N(Jl, (J2/n),
often Jl and (J2 are unknown and consequently they also have to be estimated.)
The preceding indicates how powerful the asymptotic extreme-value theory can
be. We have merely introduced the subject. For instance, we stated some re-
sults for the asymptotic distribution of y~n); one could state similar results
for y~n), y1 n), or y~~k+l' The interested reader is referred to Refs. 13, 30,
and 35.
5.4 Sample Cumulative Distribution Function

We have repeatedly stated in this chapter that our purpose in sampling from some
distribution was to make inferences about the sampled distribution, or popula-
tion, which was assumed to be at least partly unknown. One question that
might be posed is: Why not estimate the unknown distribution itself? The
answer is that we can estimate the unknown cumulative distribution function
using the sample, or empirical, cumulative distribution function, which is a func-
tion of the order statistics.
Definition 13 Sample cumulative distribution function Let Xl, X 2 , ••• ,

Xn denote a random sample from a cumulative distribution function F('),
and let y. < Y2 < ... < Yn denote the corresponding order statistics.
The sample cumulative distribution function, denoted by Fn(x), is defined by
Fix) = (lIn) x (number of Yj less than or equal to x) or, equivalently,
by Fn(x) = (lIn) x (number of Xi less than or equal to x). IIII
For fixed x, Fn(x) is a statistic since it is a function of the sample. (The

dependence of Fn(x) on the sample may not be clear from the notation itself.)
We shall see that Fix) has the same distribution as that of the sample mean of a
Bernoulli distribution.
Theorem 17 Let FnCx) denote the sanlple cumulative distribution function

of a random sample of size n from F( . ); then
P [F.(X) =~] =(:) [F(x)l'[l - F(x)r', k = 0, 1, ... , n. (43)

PROBLEMS 265
PROOF Let Zi = Ic- rCJ,X](Xi ); then Zi has a Bernoulli distribution

n
with parameter F(x). Hence, L1 Zi' which is the number of Xi less than
or equal to x, has a binomial distribution with parameters nand F(x). But
n
Fn(x) = (lIn) L Zi . The result follows. IIII
1
Much more could be said about the sample cumulative distribution func-
tion, but we will wait until Chap. Xl on nonparametric statistics to do so.
PROBLEMS
1 (a) Give an example where the target population and the sampled population
are the same.
(b) Give an example where the target population and the sampled population
are not the same.
2 (a) A company manufactures transistors in three different plants A, B, and C
whose manufacturing methods are very similar. It is decided to inspect thOse
transistors that are manufactured in plant A since plant A is the largest plant
9
and statisticians are available there. In order to inspect a week s produc-
tion, 100 transistors will be selected at random and tested for defects. Define
the sampled population and target population.
(b) In part (a) above, it is decided to use the results in plant A to draw conclu-
sions about plants Band C. Define the target population.
3 (a) What is the probability that the two observations of a random sample of two
from a population with a rectangular distribution over the unit interval will
not differ by more than I?
(b) What is the probability that the mean of a sample of two observations from a
rectangular distribution over the unit interval will be between! and.£?
4 (a) Balls are drawn with replacement from an urn containing one white and two
black balls. Let X = 0 for a white ball and X = 1 for a black ball. For
samples XI, X 2 , " ' , X9 of size 9, what is the joint distribution of the observa-
tions? The distribution of the sum of the observations?
(b) Referring to part (a) above, find the expected values of the sample mean and
sample variance.
5 Let XI, •. " Xn be a random sample from a distribution which has a finite fourth
moment Define IL = @"[XI ),U 2 = var [Xd, 1L3 = @"[{XI - 1L)3], 1L4 = @"[{X1 - 1L)4].
n n
X = (lIn) 2: Xi. and 8
I
2
= [l/{n -1)] L (X,_X)2.
I
n n
(a) Does 8 2 = [l/2n(n - 1)] L L {Xi - X J )2?
'''' I J'" I
266 SAMPLING AND SAMPLING DISTRIBUTIONS vI
*(b) Find var [~)2].

*(c) Find cov [X, ~)2], and note that cov [X, ~)2] = 0 if f-t3 = o.
Possible HINT: ~(X, - f-t)2 = ~(Xi - X)2 + (1In)~~(Xj - f-t)(Xj - f-t).
6 *(a) For a random sample of size 2 from a population with a finite (2r)th moment,
n
find tf[Mr] and var [M,], where Mr = (lIn) L (Xi - XnY.
i= 1
(b) For a random sample of size n from a population with mean f-t and rth
central moment f-t" show that
7 (a) Use the Chebyshev inequality to find how many times a coin must be tossed
in order that the probability will be at least .90 that X will lie between .4
and .6. (Assume that the coin is true.)
(b) How could one determine the number of tosses required in part (a) more
accurately, i.e., make the probability very nearly equal to .90? What is the
number of tosses?
8 If a population has u = 2 and X is the mean of samples of size 100, find limits
between which X - f-t will lie with probability .90. Use both. the Chebyshev
inequality and the central-limit theorem. Why do the two results differ?
9 Suppose that Xl and X 2 are means of two samples of size n from a population
with variance u 2 • Determine n so that the probability will be about .01 that the
two sample means will differ by more than u. (Consider Y = Xl - X 2 .)
10 Suppose that light bulbs made by a standard process have an average life of 2000
hours with a standard deviation of 250 hours, and suppose that it is considered
worthwhile to replace the process if the mean life can be increased by at least
10 percent. An engineer wishes to test a proposed new process, and he is willing
to assume that the standard deviation of the distribution of lives is about the
same as for the standard process. How large a sample should he examine if he
wishes the probability to be about .01 that he will fail to adopt the new process if
in fact it produces bulbs with a mean life of 2250 hours?
11 A research worker wishes to estimate the mean of a popUlation using a sample
large enough that the probability will be .95 that the sample mean will not differ
from the population mean by more than 25 percent of the standard deviation.
How large a sample should he take?
12 A polling agency wishes to take a sample of voters in a given state large enough
that the probability is only .01 that they will find the proportion favoring a certain
candidate to be less than 50 percent when in fact it is 52 percent. How large a
sample should be taken?
13 A standard drug is known to be effective in about 80 percent of the cases in which
it is used to treat infections. A new drug has been found effective in 85 of the
first 100 cases tried. Is the superiority of the new drug well established? (If
PROBLEMS 267
the new drug were as equally effective as the old, what would be the probability
of obtaining 85 or more successes in a sample of 1001)
14 Find the third moment about the mean of the sample mean for samples of size n
from a Bernoulli population. Show that it approaches 0 as n becomes large
(as it must if the normal approximation is to be valid).
15 (a) A bowl contains five chips numbered from 1 to 5. A sample of two drawn
without replacement from this finite population is said to be random if all
possible pairs of the five chips have an equal chance to be drawn. What is
the expected value of the sample mean 1 What is the variance of the sample
mean 1
(b) Suppose that the two chips of part (a) were drawn with replacement; what
would be the variance of the sample mean 1 Why might one guess that this
variance would be larger than the one obtained before?
*(c) Generalize part (a) by considering N chips and samples of size n. Show that
the variance of the sample mean is
u2 N n
n N-l'
where u 2 is the population variance; that is
u2 = ~ ~
N 1=1
(i __N_+_l)
2
2
16 If Xl, X 2 , X3 are independent random variables and each has a uniform distribu-
tion over (0, I), derive the distribution of (Xl X 2 )/2 and (Xl + X 2 X 3)/3.
2
17 If XI, ... , XII is a random sample from N(p., u ), find the mean and variance of
8 =J~(XI- X)2.
n-l
18 On the F distribution:
(a) Derive the variance of the F distribution. [See part (d).]
(b) If X has an F distribution with m and n degrees of freedom, argue that 1/ X
has an F distribution with nand m degrees of freedom.
(c) If X has an F distribution with m and n degrees of freedom, show that
mX/n
w
1 + mX/n
has a beta distribution.

(d) Use the result of part (c) and the beta function to find the mean and variance
of the F distribution. [Find the first two moments of mX/n = W/(l - W)].
268 SAMPLING AND SAMPLING DlSl'R.IBUTlONS VI
19 On the I distribution:
(a) Find the mean and variance of Student's t distribution. (Be careful about
existence.)
(b) Show that the density of a t distributed random variable approache-s the
standard nonnal density as the degrees of freedom increase. (Assume that
the" constant" part of the density does what it has to do.)
(c) If X is t-distributed, show that X2 is F-distributed.
(d) If X is I-distributed with k degrees of freedom, show that 1/(1 + X2!k) has
a beta distribution.
20 Let Xl, X 2 be a random sample from N(O, O. Using the results of Sec. 4 of
Chap. VI, answer the following:
(a) What is the distribution of {X 2 X l )tV2?
(b) What is the distribution of (Xl + X 2)2/(X2 X I )2?
(c) What is the distribution of (X2 + Xl)!V (Xl - X 2)2?
(d) What is the distribution of lIZ if Z = Xf! Xl?
21 Let Xl, ..... XII be a random sample from N(O, 1). Define
1 k 1 II
-k 2: Xi
1
and XII - k =
n-
k 2:
k+l
X,.
Using the results of Sec. 4, answer the following:

(a) What is the distribution of l(Xk + X II - k)?
(b) What is the distribution of kXl + (n - k)XII2_k ?
(c) What is the distribution of Xf! X~?
(d) What is the distribution of XII XII?
22 Let X h ••• , XII be a random sample from N{jJ., (12), Define
1 II
X n - k = n- k 2: X"
k+l
1 n
X=- 2: X,.
n 1
I n
S:-k = n- k - '" (Xi
1 L. X-n)
- k2
k+l
and
.. PROBLEMS 269
Using the results of Sec. 4, answer the following:

(a) What is the distribution of a-Z[(k 1)81 + (n - k - 1)8~_,,]?
(b) What is the distribution of O)(X" XII-a)?
(c) What is the distribution of a-z(X, - p,r~?
(d) What is the distribution of 8U8~-" ?
l
(e) What is the distribution of (X p,)/(8lv r;)?
23 Let Zlo Z Z be a random sample of size 2 from N(O, 1) and Xl, X z a random sample
of size 2 from N(1, 1). Suppose the Z,'S are independent of the X/so Use the
results of Sec. 4 to answer the following:
(a) What is the distribution of X Z?
(b) What is the distribution of {Z. + Zz)/V[{X z - X.)i + {Zz Zlr~]/2?
(c) What is the distribution of [(X. X z)Z {Z. - Zz)Z + {Z. + Zz)Z1/2?
(d) What is the distribution of (Xz + XI 2)z/(Xz - XI)Z?
24 Let XI be a random variable distributed N(;, P), ; = 1, 2, 3. Assume that the
random variables Xl, X z, and X 3 are independent. Using only the three random
variables Xl, X z , and X3:
(a) Give an example of a statistic that has a chi-square distribution with three
degrees of freedom.
(b) Give an example of a statistic that has an F distribution with one and two
degrees of freedom.
(c) Give an example of a statistic that has t distribution with two degrees of
freedom.
25 Let Xl, Xz be a random sample of size 2 from the density
J(x) = le-·hI(o. al)(x).
Use results on the chi-square and F distributions to give the distribution of

XI/XZ '
26 Let VI, Vz be a random sample of size 2 from a uniform distribution over the
interval (O, 1). Let Y. and Y z be the corresponding order statistics.
(a) For 0 < yz < 1, what is JJ'. I J'z=yz(Yllyz), the conditional density of YI given
Y z =yz?
(b) What is the distribution of Y z - Y I ?
27 If XI, X z, ... XII are independently and normally distributed with the same
'I
mean but different variances af, a~., ... , a: and assuming that V = L(x,/u1)/'L.(1/uJ)
and V = 'L.{X, - V)z/uf are independently distributed, show that V is normal and
V has the chi-square distribution with n - 1 degrees of freedom.
28 For three samples from normal populations {with variances af. a~, and an, the
sample sizes being nl, n z , and n3, find the joint density of
Sf
V=- and
Sl
where the 81, 8~, and Sj are the sample variances. (Assume that the samples
are independent.)
270 SAMPLING AND SAMPLING DISI'RIBUTIONS VI
29 Let a sample of size nl from a normal population (with variance an

have sample
variance Sf, and let a second sample of size n z from a second normal population
(with mean p.-z and variance a~) have mean X and sample variance S~. Find the
joint density of
and v
(Assume that the samples are independent.)
30 For a random sample of size 2 from a normal density with mean 0 and variance t,
find the distribution of the range.
31 (a) What is the probability that the larger of two random observations from any
continuous distribution will exceed the median?
(b) Genera1ize the result of part (a) to samples of size n.
32 ConSidering random samples of size n from a population with density /(x), what
is the expected value of the area under /(x) to the left of the smallest sample
observation?
·33 Consider a random sample X h ••• ., XII from the uniform distribution over the
interval (p V3a, I' + V3a). Let Y 1 < ... <YII denote the corresponding
order statistics.
(a) Find the mean and variance of Y II - Y 1•
(b) Find the mean and variance of (Y1 + Y II )/2.
(c) Find the mean and variance of Yt+l if n 2k + 1, k = 0., 1, ....
(d) Compare the variances of XII, Yk+h (Y1 + YII)/2.
HINT: It might be easier to solve the problem for VI, •.. , VII., a random sample
from the uniform distribution over either (0, 1) or (-1, I), and then make an
appropriate transformation.
34 Let Xl. ... ., XII be a random sample from the density
I
/(x;a,fJ) = 2P exp [-I(x- a)/PI],
where - 00 < (X < 00 and f3 > O. Compare ~he asymptotic distributions of the
sample mean and the sample median. In p~rticular, compare the asymptotic
variances.
• 35 Let XI, ... , XII be a random sample from the cumulative distribution function
F(x) = {I - exp [-x/O - x)]}I(o. l){X) + 1[1. q;)(x). What is the limiting distri-
bution of{y~lI) - all)/bll , where all log n/(1 + log n)and b;l = Oog n) (1 + logn)?
What is the asymptotic distribution of Y~")?
36 Let Xl, ••• , XII be a random sample from/(x; 8) = ()e-bl(o. q;)(x), () >0.
(a) Compare the asymptotic distribution of XII with the asymptotic distribution
of the sample median.
(b) For your choice of {all} and {bll }, find a limiting distribution of (Y!") - all)/bll'
(c) For your choice of {all} and {bll}' find a limiting distribution of (YIII> - all)/bll'
VII
PARAMETRIC POINT ESTIMATION

Chapter VI commenced with some general comments about inference. There,
it was indicated that a sample from the distribution of a population is useful
in making inferences about the population. Two important problems in
statistical inference are estimation and tests of hypotheses. One type of estima-
tion, namely pOint estimation, is to be the subject of this chapter.
The problem of estimation, as it shall be considered herein, is loosely
defined as follows: Assume that some characteristic of the elements in a popula-
tion can be represented by a random variable X whose density is fx( .; 0)
f(' ; 0), where the form of the density is assumed known except that it contains
an unknown parameter 0 (if 0 were known, the density function would be com-
pletely specified, and there would be no need to make inferences about it).
Further assume that the values Xl, X2, .•• , Xn of a random sample Xl' X 2 , ••• , Xn
from f(·; 0) can be observed. On the basis of the observed sample values
X h X 2 ' ••• , XII it is desired to estimate the value of the unknown parameter 0 or
the value of some function, say r(8), of the unknown parameter. This estima-
tion can be made in two ways. The first, called point estimation, is to let the
272 PARAMETRIC POINT ESTIMATION vn
value of some statistic, say t(Xb ... , X n), represent, or estimate, the unknown
reO); such a statistic t(Xb ... , Xn) is called a point estimator. The second,
called interval estimation, is to define two statistics, say t 1(Xh •.. , Xn) and
tiXb ···, X n), where t 1 (Xh ···, Xn) < '2(X1 , ••• , X n), so that (tl(XI , ••• , XII)'
t 2 (Xb ..• , Xn)) constitutes an interval for which the probability can be deter-
mined that it contains the unknown reO). For example, if f( . ; 0) is the normal
density, that is,
f(x; 0) = f(x; tl, u) = 4>/l,u 2(x) = 1

r::L exp [2! (X u It) 2] ,
y 21tu
where the parameter 0 is (p, u), and if it is desired to estimate the mean, that is,
n
reO) = tl, then the statistic X (ljn) L Xi is a possible point estimator of
I
r(O) = tl, and (X - 2 j8 2 /n, X +2 j8 2 /n) is a possible interval estimator

n
of reO) = tl. {Recall that 8 2 [lj(n - 1)] L (Xi - X)2.} Point estimation
1
will be discussed in this chapter and interval estimation in the next.
Point estimation admits two problems: the first, to devise some meanS
of obtaining a statistic to use as an estimator; the second, to select criteria and
techniques to define and find a best" estimator anl0ng many possible estima-
u;
tors. Several methods of finding point estimators are introduced in Sec. 2.

One of these, and probably the most important, is the method of maximum
likelihood. In Sec. 3 several "optimum" properties that an estimator or sequence
ofestimators may possess are defined. These include closeness, bias and variance,
efficiency, and consistency. The loss and risk functions, essential elements
in decision theory, are defined as possible tools in assessing the goodness of
estimators.
Section 4 is devoted to sufficiency, an important and useful concept in the
study of mathematical statistics that will also be utilized in succeeding chapters.
Unbiased estimation is considered in Sec. 5. The Cramer-Rao lower bound
for the variance of unbiased estimators is given, as well as the Rao-Blackwell
theorem concerning sufficient statistics. A brief look at invariant estimators
is presented in Sec. 6. Bayes estimation is considered in Sec. 7. A Bayes
estimator is given as the mean of the posterior or from the decision-theoretical
viewpoint as an estimator having smallest average risk. Some results in the
simultaneous estimation of several parameters are given in Sec. 8. Included is
the notion of ellipsoid of concentration of a vector of point estimators and the
Lehmann-Scheffe theorem. Section 9 is devoted to a brief discussion of some
optimum properties of maximum-likelihood estimators.
2 METHODS OF FINDING ESTIMATORS 273
Frequent use, of some of the distribution-theoretical results f{)r stat~stics~

which Were derived in earlier chapters~ especially Chaps. V and VI, wIll be
noted throughout this chapter. After all, estimators .are.sta~isti~s, and to study
properties of estimators, it is desirable to look at then dIstnbutIons.
2 METHODS OF FINDING ESTIMATORS

Assume that Xl' ... , Xn is a random sample from a density f(' ; e), where the form
of the density is known but the parameter e is unknown. Further assume that
e is a vector of real numbers, say e = (e h .. " ek ). (Often k will be unity.)
We sometimes say that el , ... , ek are k parameters. We will let 9, called the
parameter space, denote the set of possible values that the parameter e can
assume. The object is to find statistics, functions of the observations Xl' ... , X n ,
to be used as estimators of the ej,j = I, ... , k. Or, more generally, our object
is to find statistics to be used as estimators of certain functions, say 1'1 (e), ... , 1'r(e).
of e = (elt •.. , ek ). A variety of methods of finding such estimators has been
proposed on more or less intuitive grounds. Several such methods will be
presented. along with examples, in this section. Another method, that of the
method of least squares will be discussed in Chap. X.
An estimator can be defined as in Definition I.
Definition 1 Estimator Any statistic (known function of observable

random variables that is itself a random variable) whose values are used
to estimate 1'(e), where 1'(') is sonle function of the parameter e, is defined
to be an estimator of 1'(e). IIII
An estimator is always a statistic which is both a random variable and a

function. For instance, suppose Xb "', Xn is a random sample from a density
f(· ; e) and it is desired to estimate 1'(e), where 1'(') is some function of e. Let
I(X1' ... , Xn) be an estimator of 1'(e). The estimator I( Xl, ... , Xn) can be
thQught of in two related ways: first, as the random variable , say T , where
T = I(X}, ... , X n), and, second, as the function 1(-, ... , .). Naturally, one
needs to specify the function 1(', ... , .) before the random variable T =
I(X}) ... , Xn) is defined. In all we have three types of tees : the capital Latin T,
which represents the random variable I(X}) ... , X n), the small script I, which
represents the function 1(', ... , .), and the small Latin t, which represents a
value of T; that is, t = I(Xl' ... , x n ). Let us adopt the convention of calling
the statistic (or random variable) that is used as an estimator an Uestimator"
and calling a value that the statistic takes on an "estimate." Thus the word
, ,
'274 PARAMETRIC POINT ESTIMATION VII
"estimator" stands for the function, and the word" estimate" stands for a
value of that function; for example, Xn =.!.. f

ni=l
Xi is an estimator of a mean J.l,
and xn is an estimate of J.l. Here Tis X n , t is Xn , and 1(' , ... , .) is the function
defined by summing the arguments and then dividing by n.
Notation in estimation that has widespread usage is the following: 8 is
used to denote an estimate of (J, and, more generally, (81 , " ' , Ok) is a vector
that estimates the v~ctor «(Jl' ... , (Jk), where OJ estimates (Jj' j = 1, .,., k. If
8 is an estimate of (J, then 0 is the corresponding estimator of (J; and if the
discussion requires that the function that defines both 8 and 0 be specified, then
it can be denoted by a small script theta, that is, 0 = 9( Xl' ... , Xn).
When we speak of estimating (J, we are speaking of estimating the fixed
yet unknown value that (J has. That is, we assume that the random sample
Xl, ... , Xn came from the density f(' ; (J), where (J is unknown but fixed. Our
object is, after looking at the values of the random sample, to estimate the fixed
unknown (J. And when we speak of estimating -r(0), we are speaking of estimat-
ing the value -r«(J) that the known function -r(') assumes for the' unknown but
fixed (J.
2.1 Methods of Moments

Letf(' ; (Jl' ... , (Jk) be a density of a random variable X which has k parameters
(Jl' ... , (Jk' As before let It; denote the rth moment about 0; that is, J.l; = c[xr].
In general J.l; will be a known function of the k parameters (Jl' .. " (Jk' Denote
this by writing J.l; = J.l;«(Jl, "', (Jk)' Let Xl' •.. , Xn be a random sample
from the density f('; (Jl' ... , (Jk), and, as before, let Mj be the Jth sample
moment; that is,
1 n
Mj =- LXi.
n i= 1
Form the k equations

j = 1, ... , k, (1)
in the k variables (Jl, ... , (Jk, and let 9 1, ... , 9 k be their solution (we assume
that there is a unique solution). We say that the estimator (9 b ... : 9 k ),
where 8j estimates (Jj' is the estimator of «(Jl, .. " (Jk) obtained by the method of
moments. The estimators were obtained by replacing population moments by
sample moments. Some examples follow.
2 MElHODS of FINDING ESTIMATORS 275
EXAMPLE 1 Let Xl' ..• , XII be a random sample from a normal distribution
with mean Il and variance u 2 • Let (Ov ( 2 ) = (p, u). Estimate the param-
2
eters Il and G b.y the method of moments. Recall that u = J12 - (p~)2
and Il = Ill. The method-of-moments equations become
M~ = J11 = J1~(J1, u) = J1
2
M2 J12 Jl1,(J1, u) u + J12,
and their solution is. the following: The method-of-moments estimator
of Jt is M~ X, and the method-of-moments estimator of u is'
X I
JM 2 - 2 = J(lln) xl- X2 = JI (Xi -X)2In. Note that the
method-of-moments estimator of u given above is not jSi. IIII
EXAMPLE 2 Let Xl' ... , Xn be a random sample from a Poisson distribu-

tion with parameter 1. Estimate A. There is only one parameter, hence
only one equation, which is
M~ = J1~ = J1~(A) = A.
Hence the method-of-moments estimator of A is M~ = X, which says
estimate the population mean 1 with the sample mean X. IIII
EXAMPLE 3 Let Xl' ... , Xn be a random sample from the negative expo-
nential density f(x; 0) Oe- 9xI(o,oolx). Estimate 0. The method-of-
moments equation is
M~ = J1~ = J1~(0) = !;
o
hence the method-of-moments estimator of 0 is I/M~ = I/X. IIII
EXAMPLE 4 Let Xl' •.. , Xn be a random sample from a uniform distribu-

tion on (p. - j3u, J1 + j3u). Here the unknown parameters are two,
namely II and u, which are the population mean and standard deviation.
The method-of-moments equations are
and
· '276' PARAMETRIC POINT ESTIMATION VB
hence the method-of-moments estimators are X for II and
for (J'.
We shall see later that there are better estimators of p and (J' for this
distribution. III!
Method-of-moments estimators are not uniquely defined. The method-

of-moments equations given in Eq. (1) are obtained by using the first k raw
moments. Central moments (rather than raw moments) could also be used to
obtain equations whose solution would also produce estimators that would be
labeled method-of-moments estimators. Also, moments other than the first
k could be used to obtain estimators that would be labeled method-of-moments
estimators.
If, instead of estimating Uh, ... , Ok), method-of-moments estimators of,
say, 'l'l(Ol' ... , Ok), ... , 'l',.«(Jv .•• , Ok) are desired, they can be obtained in several
ways. One way would be to first find method-of-moments estimates, say
Ol' ... , Ok, ofO l , ... , Ok and then use riOb "', Ok) as an estimate of 7:j«(Jh ... , (Jk)
for j = I, "', r. Another way would be to form the equations
j I, ... ,r
and solve them for 7: 1 , ••• , 7:,.. Estimators obtained using either way are called
method-of-moments estimators and may not be the same in both cases.
2.2 Maximum Likelihood

To introduce the method of maximum likelihood, consider a very simple estima-
tion problem. Suppose that an urn contains a number of black and a number
of white balls, and suppose that it is known that the ratio of the numbers is
3/1 but that it is not known whether the black or the white balls are more
numerous. That is, the probability of drawing a black ball is either or i.
If n balls are drawn with replacement from the urn, the distribution of x, the
*
number of black balls, is given by the binomial distribution
f(x; p) (;) p'q" • for x = 0, 1, 2, ... , n,
where q = 1 - p and p is the probability of drawing a black ball. Here p = 1,

or p = i.
We shall draw a sample of three balls, that is, n = 3, with replacement and
2 METHODS' OF FINDING ESTIMATORS 277
attempt to estimate the unknown parameter p of the distribution. The estima-

tion problem is particularly simple in this case because we have only to choose
between the two numbers .25 and .75. Let us anticipate the results of the
drawing of the sample. The possible outcomes and their probabilities are given
below:
Outcome: x 0 1 2 3
9 1.1 1.7.
f(x; 1) «1 6-4" 64 64
1.1 1
f(x; -1)
117
"8. 64 n9 '6-4"
In the present example, if we found x = 0 in a sample of 3, the estimate .25 for

p would be preferred over .75 because the probability it is greater than 6 4'
1
. i.e., because a sample with x = 0 is more likely (in the sense of having larger
probability) to arise from a population with p ! than from one with p i.
And in general we should estimate p by .25 when x = 0 or 1 and by .75 when
x = 2 or 3. The estimator may be defined as
p= p(x) = {.25
for x = 0, 1
.75 for x = 2, 3.
The estimator thus selects for every possible x the value of p, say /1, such that
f(x; /1) > f(x; p'),
where p' is the alternative value of p.
More generally, if several alternative values of p were possible, we might
reasonably proceed in the same manner. Thus if we found x = 6 in a sample
of 25 from a binomial population, we should substitute all possible values of p
in the expression
for 0 <p< I (2)
anti choose as our estimate that value of p which maximizedf(6; p). For the
given possible values of p we should find our estimate to be 265- The position
of its maximum value can be found by putting the derivative of the function
defined in Eq. (2) with respect to p equal to 0 and solving the resulting equation
for p. Thus,
~f(6")
dp ,p = (25)
6 P5 (l - p) 18 [6(1 - p) 19p],
278 PA~C POINT ESTIMATIoN vn
and on putting this equal to 0 and solving for p, we find that p 0, I, 265 are
the roots. The first two roots give a minimum, and so our estimate is therefore
p 265' This estimate has the property that
f(6; P) > f(6; p'),
where pi is any other value of p in the interval 0 p ~ 1.
In order to define maximum-likelihood estimators, we shall first define the
likelihood function.
Definition 2 Likelihood function The likelihood function of n random

variables XI' X 2 , " ' , Xn is defined to be the joint density of the n random
variables, say fXI, ... ,x.lxl , ••• , Xn; 0), which is considered to be a function
of 8. In particular, if XI, ... , Xn is a random sample from the density
f(x; 0), then the likelihood functionisf(xl; 8)f(X2; 8)· '" 'f(xn; 8). IIII
Notation To remind ourselves to think of the likelihood function as a

function of 8, we shall use the notation L(O; Xv ••• , xn) or L(. ; Xh ..• , xn)
for the likelihood function. IIII
The likelihood function L(O; XI' ••• , xn) gives the likelihood that the
random variables assume a particular value Xl, X2, ... , Xn . The likelihood is
the value of a density function; so for discrete random variables it is a proba-
bility. Suppose for a moment that 8 is known; denote the value by 00 , The
particular value of the random variables which is "most likely to occur" is that
value x~, Xl' ... , x~ such that fXI, ... , xn(Xf, ... , Xn; ( 0) is a maximum. For
example, for simplicity let us assume that n = 1 and Xl has the normal density
with mean 6 and variance 1. Then the value of the random variable which is
most likely to occur is Xl = 6. By" most likely to occur" we mean the value
x~ of Xl such that cP6,I(xD>cP6,I(XI)' Now let us suppose that the joint
density of n random variables is fx!, ...• xn(xI , ••• , Xn; 8), where 8 is unknown.
Let the particular values which are observed be represented by Xl, X2' ., ., x~.
We want to know from which density is this particular set of values most likely
to have come. We want to know from which density (what value of 0) is the
likelihood largest that the set x~, ... , x~ was obtained. In other words, we
want to find the value of 0 in 9, denoted by {J, which maximizes the likelih~~d
function L(O; x~, ... , x~. The value {J which maximizes the likelihood function
is, in general, a function of Xf, ••• , Xn, say (J = 9(Xh X2,"'" xJ. When this is
e
the case, the random variable = 8(Xf, X 2 , ••• , Xn) is called the maximum-
likelihood estimator of O. (We are assuming throughout that the maximum
of the likelihood function exists.) We shall now formalize the definition of a
maximum-likelihood estimator.
Definition 3 Maximum-likelihood estimator Let

L(O) =L(O; Xl' ... , Xli)
be the likelihood function for the random variables Xl' X 2 , "', Xli' If
lJ [where (J = 9(Xl' X2, ... , Xli) is a function of the ~bservations Xl, ... , Xli]
is the value of 0 in e which maximizes L(O), then e = 8(Xl' X 2 , ••• , Xli)
is the maximum-likelihood estimator of O. {} = 8(xI' ... , Xli) is the
maximum-likelihood estimate of 0 for the sample Xl' .. " Xli' 1//1
The most important cases which we shall consider are those in which
Xb X 2 , .,., Xli is a random sample from some density f(x; 0), so that the
likelihood function is
L(O) = f(Xl ; O)f(X2; 0) ... f(x lI ; 0).
Many likelihood functions satisfy regularity conditions; so the maximum-

likelihood estimator is the solution of the equation
dL(O)
dO = O.
Also L(O) and log L(O) have their maxima at the same value of 0, and it is some-
times easier to find the maximum of the logarithm of the likelihood.
If the likelihood function contains k parameters, that is, if
II
L(Ol' O2 , . .. , Ok) = nf(Xj; Oh O2 , ••• , Ok),

i= 1
then the maximum-likelihood estimators of the parameters 01 , O2 , ... , Ok are

the random variables 0 1 = 9t (Xb " ' , Xn), O2 = 82 (Xl , ... , XJ, ... , 0 k =
8lXb ... , Xli)' where {}l' (}2, "', 6k are the values in e which maximize
L(Ob Oz, ... , Ok)'
If certain regularity conditions are satisfied, the.point where the likelihood
is a maximum is a solution of the k equations
aLeOb ... , Ok)

aO = O.
k
In this case it ~ay also be easier to work with the logarithm of the likelihood.
We shall Illustrate these definitions with some examples.
280. pAItAMETlUC POINT ESTIMATION VII
EXAMPLE 5 Suppose that a random sample of size n is drawn from the

Bernoulli distribution
o ~ p .s; 1 and q = 1 - p.
The sample values Xh X2, " ' , XII will be a sequence of Os and Is, and the
likelihood function is
L(p) = n
II
i= 1
ylql-XI = pf.Xiqn-f.xi,
and if we let
we obtain
*log L(P) = y log p + (n - y) log q
and
dlogL(p) y n- y
=- - ,
dp p q
remembering that q = 1 - p. On putting this last expression equal to 0

and solving for p, we find the estimate
y 1 _
P= - = -
n n
2. Xi = x, (3)
which is intuitively what the estimate for this parameter should be. It is
also a method-of-moments estimate. For n = 3, let us sketch the likeli-
hood function. Note that the likelihood function depends on the x/s
only through 2. Xi; thus the likelihood function can be represented by the
following four curves:
Lo = L(P; 2. Xi = 0) = (1 - p)3
Ll L(P; 2. Xi = 1) = P(1 - p)2
L2 = L(p; 2. Xi = 2) = p2(1 - p)
L3 = L(P; 2. Xl = 3) = p3,
which are sketched in Fig. 1. Note that the point where the maximum of
each of the curves takes place for 0 p 1 is the same as that given in
Eq. (3) when n = 3. 11I1
,.. Recall that log x means loge x.
L(p)
FIGURE 1
EXAMPLE 6 A random sample of size n from the normal distribution has

the density
Ii J2nu
i=l
1 e-(1/2a 2
)(xi-Jt)2 = (~)n/2 exp [- ~ I (Xi -
2nu 2u
tl)2] .
The-logarithm of the likelihood function is
L lie = - n-log 2n - n-log u 2 - - I2 I (x - tl) 2

2 2 2u i ,
where u > 0 and - 00 < tl < 00.

To find the location of its maximum, we compute
and
aLate n 1 1 2
au 2 = - 2. u 2 + 2u4 I (Xi - tl) ,
and on putting these derivatives equal to 0 and solving the resulting

equations for tl and u 2 , we find the estimates
(4)
"'2
U = -n1 "L. (Xi - 2,
- X) (5)
which turn out to be the sample moments corresponding to tl and u 2 •

I111
282 PAR.AMETRIC pOINT ESTIMATION VII
EXAMPLE 7 Let the random variable X have a uniform density given by

f(x; 0) = I[o-t,o+-u(x),
where - 00 < 0 < 00; that is~ e = real line. The likelihood function
for a sam pIe of size n is
n
nI
II
L(O; X h ••• , x,,) = nf(x i ; 0) = lo - t , O+t1(Xj)

1=1 i=l
= I h 'n-t,)ll+tl.°), (6)
where Yl is the smallest of the observations and Y1I is the largest. The
last equality in Eq. (6) follows since n." IlO-t,o+u(xi) is unity if and on1y
i= 1
if all Xl' ••• , X" are in the interval [0 - t, 0 + t], which is true if and
only if 0 - t < YI and Y" s 0 + t, which is true if and only if
y" - t 0 <YI + t. We see that the likelihood function is either 1
(for Yn - t < 0 < YI + i) or 0 (otherwise); hence any statistic with value
iJ satisfying Y1I - i < {j YI + t is a maximum-likelihood estimate.
Examples are Yn - i, Yl + i, and llil + Y1I)' This latter is the midpoint
between Yn -!- and Yt + i, or the midpoint between Yl and Y1I' the
smallest and largest observations. / / //
EXAMPLE 8 Let the random variable X have a uniform distribution with

density given by
1
f(x; 0) = f(x; Jl, 0') = 2J3a hll-./3o-,Il+,/3o-1(x),
where - 00 O. (Recall Example 4.) Here the likeli-
hood function for a sample of size n is
L(p, 0'; Xl"'" Xn) = (~30')"fIJ11l - 43 o- 1l+.J3o-lx ,)

i=l •
= (J 3 0') \ . -'/3 ., ,.J(Y ,)I[j>., • +,/•• J<y,)
= (d3.,.) '1[(.-,.)/,/3,00)(0')1[<1.-.)/,/3, oo)(O')/[j>" OO)(Y,).
where YI is the smallest of the observations and Yn is the largest. The

likelihood function is (2J3a)-n in the shaded area of Fig. 2 and 0 else-
where. (2J3a) -n within the shaded area is clearly a maximum when 0'
FIGURE 2
is smallest, which is at the intersection of the lines J.! - )3 (J = Yl and

J1 + J3(J = Yn. Hence the maximum-likelihood estimates of J.! and (J are
1
.a = '2 (y 1 + Y n) (7)
and
1
lJ = J3
2 3
(yn - Yl), (8)
which are quite different from the method-of-moments estimates given in

Example 4. IIII
The above four examples are sufficient to illustrate the application of the
method of maximum likelihood. The last two show that one must not always
rely on the differentiation process to locate the maximum.
The function L(O) may, for example, be represented by the curve in Fig. 3,
where the actual maximum is at fJ, but the derivative set equal to 0 would locate
0' as the maximum. One must also remember that the equation 8LI80 = 0
locates minima as well as maxima, and hence one must avoid using a root of
the equation which actually locates a minimum.
We shall see in later sections (especially Sec. 9 of this chapter) that the
maximum-likelihood estimator has some desirable optimum properties other
than the intuitively appealing property that it maximizes the likelihood function.
In addition, the maximum-likelihood estimators possess a property which is
sometimes called the invariance property of maximum-likelihood estimators. A
little reflection on the meaning of a single-valued inverse will convince one of
the validity of the following theorem.
284 PARAMETRIC POINT ESTIMATION VII
L(O)
~--~--------~--------------~O
FIGURE 3 0'
Theorem 1 Invariance property of maximum-likelihood estimators Let

e = 8(Xl' X 2 , ••• , Xn) be the maximum-likelihood estimator of 0 in the
density I(x; 0), where 0 is assumed unidimensional. If r( .) is a function
with a single-valued inverse, then the maximum-likelihood estimator of
r(O) is r(e). 1111
For example, in the normal density with 110 known the maximum-likelihood
estimator of (J2 is
By the invariance property of maximum-likelihood estimators, the maximum-

likelihood estimator of (J is
Similarly, the maximum-likelihood estimator of, say, log (J2 is
log [~.tl (Xi - 1'0)2].

The invariance property of maximum-likelihood estimators that is exhib-
ited in Theorem 1 above can and should be extended. Following Zehna [43],
we extend in two directions: First. 0 will be taken as k-dimensional rather
than unidimensional, and, second, the assumption that r(·) has a single-valued
inverse will be removed. It can be noted that such extension is necessary by
considering two simple examples. As a first example, suppose an estimate of
the variance, namely O(l - 0), of a Bernoulli distribution is desired. Example
5 gives the maximum-likelihood estimate of 0 to be X, but since O(l - 0) is not
a one-to;'one function of 0, Theorem 1 does not give the maximum-likelihood
2 MEmoDS OF f'INDING ESTIMATORS 285
estimator of 8(1 - 8). Theorem 2 below will give such an estimate, and it will
be x(1 - x). As a second example, consider sampling from a normal distribu-
tion whete both It and a2 are unknown, and suppose an estimate of C[X2] =
It'}. + (12 is desired. Example 6 gives the maximumMlikelihood estimates of J1
and a2 , but It'}. + (12 is not a one-to-one function of J.i and 0'2, and so the
maximum-likelihood estimate of It'}. + 0'2 is not known. Such an estimate will
be obtainable from Theorem 2 below. It will be x2 + (lIn) L (Xl' - X)2.
Let 0 = (01 , ••• , OJ be a k-dimensional parameter, and, as before, let e
denote the parameter space. Suppose that the maximum-likelihood estimate
of t(O) = (tl (0), ... , 1'r(O», where 1 < r < k, is desired. Let T denote the range
space of the transformation t(·) = (tl (.), ... , t r (· ».
T is an rMdimensional
space. Define M( t; Xl' ... , Xn) = sup L(O; Xl' .•• , Xn)· M(· ; Xh .•• , Xn)
{6:.(6)=.}
is called the likelihood function induced by t(· ).* When estimating 0 we max-
imized the likelihood function L(O; X h •.• , xn) as a function of 0 for fixed
Xb " ' , x,,; when estimating t = 1'(0) we will maximize the likelihood function
induced by t(·), namely M(1'; Xl' .•• , X n), as a function of t for fixed Xh ••• , X n •
Thus, the maximum-likelihood estimate of t = t(O), denoted by t, is any
value that maximizes the induced likelihood function for fixed Xl' ..• , x"; that
is, t is such that M(t; Xl' •.. , xn) M(t; Xl' ..• , X,,) for all l' E T. The invari-
ance property of maximum-likelihood estimation is given in the following
theorem.
Theorem 2 Let 8 = (81 , ... , 8 k ), where OJ = 8/Xl , ••• , Xn)' be a

maximum-likelihood estimator of 0 = (01 , ••• , Ok) in the density
f(· ; 01, ... , Ok)' If t(O) = (1'1(0), ... , tr(O» for 1 r k is a transforma-
tion of the parameter space ~, then a maximum-likelihood estimator of
t(O) = (1'1(0), ... , tr(O» is 1'(0) = (t l (8), ... , t,.(0». [Note that tiO) =
ti~l' ... , O}.); so the maximum-likelihood estimator of tj(Ol' ... , Ok) is
t i 0 1 , ••• , 0 k), j = 1, .. ., r.]
PROOF Let e= (e 1 , ••• , eJ

be a maximum-likelihood estimate of
0= (Oh ... , Ok)· It suffices to show that M(1'(fJ); Xl' •.. , Xn) >
M( t; Xl' •.• , xn) for any l' E T, which follows immediately from the in-
equality M(t; Xl'·'·' Xn) = sup L(O; Xl"'" X) sup L(O° X .•• X)
{8: T(8)=t} " / f e e ' 1, 'n
= L(fJ; x[, ... , Xn) = sup L(O; x[, ... , Xn) = M(1'(tJ)· X ••• X). IIII
(8; T(8)=T(tJ)} '1, ,n
*The notation
. "sup" is used here. and elsewhere I'n th'IS b 00 k ,as 1't IS
' usua II y use d In
.
mathematIcs.
. I t 'f" .. •For those
I readers Who are not acq a' t d 'th
u In e WI th' t t'
IS no a lon, not mue
h
IS OS I sup IS rep aced by "max" h ' . . .
• were max IS an abbreVIatIon for maxImum. .
286 P.AIlANETRIC POINT ESTIMATION VII
It is precisely this property of invariance enjoyed by maximum-likelihood

estimators that allowed us in our discussion of maximum-likelihood estimation
to consider estimating (Ott ... , Ok) rather than the more general T 1 (OI' ... , Ok),
... , 1:l'(Oh .. " Ok)'
°
EXAMPLE 9 In the normal density, let = (Oh ( 2 ) = (j1., (i2). Suppose
T(O) = J1 + Zq u, where Zq is given by tfJ(Zq) = q. T(O) is the qth quantile.
According to Theorem 2, the maximum-likelihood estimator of T(O) is
X + zqj(l/n) I (X~ - X)2. 1I11
2.3 Other Methods

There are several other methods of obtaining point estimators of param-
eters. Among these are (i) the method of least squares, to be discussed in
Chap. X, (ii) the Bayes method, to be discussed later in. this chapter, (iii) the
minimum-chi-square method, and (iv) the minimum-distance method. In this
subsection we will briefly consider the last two. Neither will be used again in
this beok.
Minimum-chi-square method Let Xl' ... , Xn be a random sample from a

density given by I x(X; 6), and let fJ\, "', fI' k be a partition of the range of X.
The probability that an observation falls in cell fI' j ' j = I, "', k, denoted by
Pj(O), can be found. For instance, if Ix(x; 0) is the density function of a con-
tinuous random variable, then Pj(O) = P[X falls in cell fl'j] = SYJ Ix(x; 6) dx.
k •
Note that I p/6) = 1. Let the random variable N j denote the number of
j=1
k
X/s in the sample which falls in cell fl'j' j = I, ... , k; then I Nj = n, the
j=1
sample size. Form the following summation:
where nj is a value of N j • The numerator ofthejth term in the sum is the square
of the difference between the observed and the expected number of observations
falling in cell fj' j • The minimum-chi-square estimate of 6 is that tJ which
minimizes x2 • It is that 0 among all possible O's which makes the expected
number of observation in cell fl'j "nearest" the observed number. The minimum-
chi-square estimator depends on the partition fl'1' ••• , fj' k selected.
2 ME1HODS Oil FINDING ESTIMATORS 287
EXAMPLE 10 Let Xl' .. ., XII be a random sample from a Bernoulli distri-

bution; that is, fx(x; 0) = OX(l - O)l-X for x 0, 1. Take N, = the
numberoo f bservations equal to j for j = 0, 1. .Here
.
the range of the
f h b
observation X is partitioned into the two sets conslstmg 0 t e num ers
°
and 1 respectively.
I [n. nptO)] 2 [no - n(l - 0)]2 (n I - nO)2
x -
2
- "
jf:o J
nptO) =
n(l - 0) + n0
[n - n 1 n(l - 0)]2
- +~I - nO)2 = (n 1 - nO)2 1 .
- n(l - 0) nO n 0(1 - 0)
The minimum of X2 as a function of (} can be found by inspection by

noting that x2 = 0 for 0 = n1/n. Hence (j = ntln. For this example
there was only one choice for the partition [/.' ., ., [/ k' The estimator
found is the same as what would be obtained by either the method of
moments or maximum likelihood. IIII
Often it is difficult to locate that () which minimizes X2; hence, the denomi-
nator np/O) is sometimes changed to nj (if nj = 0, unity is used) forming a
k
modified X2 = I ([n) - np/O)]2Inj}. The modified minimum-chi-square estimate
j= I
of 0 is then that () which minimizes the nlodified X2 •
Minimum-distance method Let XI' ... , Xn be a random sample from the

distribution given by the cumulative distribution function Fx(x~ 0) = F(x~ 0),
and let d(F, G) be a distance function that measures how "far apart" two cumula-
tive distribution functions F and G are. An example of a distance function is
d(F, G) = sup IF(x) -"G(x) I, which is the largest vertical distance between F
x
and G. See Fig. 4.
The minimum-distance estimate of 8 is that (j among all possible
which d(F(x; 8), F ,ix» is mil~Jmized, where F n(x) is the sample cumulative
for °
distribution function. Thus, 8 is chosen so that F(x; (j) will be "closest" to
F,,(x) , which is desirable since we saw in Subsec. 5.4 of Chap. VI that for a
fixed argument x the sample cumulative distribution function has the same
distribution as the mean of a binomial distribution; hence, by the law of large
numbers F "(x) "converges" to F(x). The minimum-distance estimator might be
intuitively appealing, but it is almost always difficult to find since locating
{} which minimizes d(F(x; 8), F,,(x») is seldom easy. The following example is
an exception.
d(F. G)
~--~--------------------------------~X
FIGURE 4
EXAMPLE 11 Again let Xl' ... , Xn be a random sample from a Bernoulli

distribution; then
F(x; B) = (I - B)/[o, 1)(x) + 1[1, OCllx),-
where 0 < B :s:: 1. Let nj = the number of observations equal to};} = 0, 1.
Then
no
F ,.(x) = -;;: 1[0, l/X) + 1£1, OCl)(X).
Now if the distance function d(F, G) = sup IF(x) - G(x) I is used, then
x
d(F(x; 8), Fix» is minimized if 1 - 0 is taken equal to noln or 0 = ndn =
Lxi/no Hence &= x. IIII
For a more thorough discussion of the minimum-chi-square method, see
Cramer [11] or Rao [17]. The minimum-distance method is discussed in
Wolfowitz [42].
3 PROPERTIES OF POINT ESTIMATORS

We presented several methods of obtaining point estimators in the preceding
section. All the methods were arrived at on a more or less intuitive basis. The
question that now arises is: Are some of many possible estimators better, in
some sense, than others? In this section we will define certain properties. which
an estimator mayor may not possess, that will help us in deciding whether one
estimator is better than another.
3.1 Closeness
If we have a random sample Xv ... , Xn from a density, say f(x; 8), which is
known except for 0, then a point estimator of 1'(8) is a statistic, say t(XI' ... , XII)'
whose value is used as an estimate of 1'(8). We will assume here that 1'(0) is a
3 PRoPERnES OF POINt ESTIMATORS 289
real-valued (not a vector) function of the unknown parameter 8. [Often 1'(8)

will be 6 itself.) 'ldeallY7 we.. would like the value of !(XI , •.• , Xn) to be the
unknown 1'(0), but this is not possible except in trivial cases, one of which follows.
EXAMPLE 12 Assume that one can sample from a density given by

f(x; e) = I(8--hIH-!)(X),
where e is known to be an integer. That is, S, the parameter space,
consists of all integers. Consider estimating 8 on the basis of a single
observation Xl' If !(x I ) is assigned as its value the integer nearest Xl' then
the statistic or estimator t(XI ) will always correctly estimate 8, In a
sense, the problem posed in this example is really not statistical since one
knows the value of () after taking one observation. //1/
Not being able to achieve the ultimate of always correctly estimating the
unknown 't'«(), we _look for an estimator t( Xl' ... , X,,) that is "close" to 1'(e).
There are several ways of defining" close," T = !(XI' ... , Xn) is a statistic
and hence has a distribution, or rather a family of distributions, depending on
what () is. The distribution of Ttells how the values t of Tare distributed, and
we would like to have the values of T distributed near 1'(0); that is, we would like
to select t(·, ,.,' .) so that the values of T = t(XI' ... , Xn) are concentrated
near 't'«()). We saw that the mean and variance of a distribution were, respec-
tively, measures of location and spread. So what we might require of an
estimator is that it have its mean near or equal to 1'(0) and have small variance.
These two notions are explored in Subsec. 3.2 below and then again in Sec. 5.
Rather than resorting to characteristics of a distribution, such as its mean
and variance, one can define what" concentration" might mean in terms of the
distribution itself. Two such definitions follow.
Definition 4 More concentrated and most concentrated Let T =

t(Xh ... , Xn) and T' = t'(XI' ... , Xn) be two estimators of 1"(8). T' is
called a more concentrated estimator of 1"(8) than T if and only if
P8[1"(e) -). < T' < 1"(0) +).] > P8 [1'(O) -). < T < 1"(8) +).] for all ). > 0
and for each 0 in S. An estimator T* = t*(XI , " ' , X") is called most
concentrated if it is more concentrated than any other estimator, /11/
Remark The subscript 8 on the probability symbol P e['] is there to

emphasize that, in general, such probability depends on e. For instance,
290 pAIlAMETRIC poINT ESTIMATION vn
IIIPe[1"(O) - ..l < T < 1"(0) + ..l], the event {r(O) - ..l < T:5: 1"(0) +..l} is
described in terms of the random variable T, and, in general, the distri-
bution of T is indexed by O. IIII
We see from the definition that the property of most concentrated is

highly desirable (Pitman [41], in defense of his calling a 'most concentrated
estimator best, stated that such an estimator is undeniably best"); unfortu-
H
nately, most concentrated estimators seldom exist. There are just too many
possible estimators for anyone of them to be most concentrated. What is then
sometimes done is to restrict the totality of possible estimators under con-
sideration by requiring that each estimator possess some other desirable property
and to look for a best or most concentrated estimator in this restricted class.
We will not pursue the problem of finding most concentrated estimators, even
within some restricted class, in this book.
Another criterion for comparing estimators is the following one.
Definition 5 Pitman-closer and Pitman-closest Let T = I(XI' ... , X,,)

and T' = 1'( Xl' ... , X n) be two estimators of 1'(8). T' is called a Pitman-
closer estimator of 1"(8) than T if and only if
for each 8 in e.
An estimator T* is called Pitman-closest if it is Pitman-closer than any
other estimator. IIII
The property of Pitman-closest is, like the property of most concentrated,
desirable, yet rarely will there exist a Pitman-closest estimator. Both Pitman-closer
and more concentrated ,are intuitively attractive properties to be used to com-
pare estimators, yet they are not always useful. Given two estimators T and
T', one does not have to be more concentrated or Pitman-closer than the other.
What often happens is that one, say T, is Pitman-closer or more concentrated
for some 0 in 8, and the other T' is Pitman-closer or more concentrated for
other 0 in e; and since 8 is unknown, we cannot say which estimator is preferred.
Since Pitman-closest estimators rarely exist for applied problems, we will not
devote further study to the notion in this book; instead, we will consider other
ways of measuring the closeness of an estimator to 1"(0).
Competing estimators can be compared by defining a measure of the close-
ness of an estimate to the unknown 1"(8). An estimator T' = I'(XI' ... , X,,)
of 1"(8) will be judged better than an estimator T = t( Xl' ... , X,,) if the measure
of the closeness of T' to 1"(0) indicates that T' is closer to 1"(0) than T. Such
concepts of closeness will be discussed in Subsecs. 3.2 and 3.4.
3 PROPERTIES OF POINT ESTIMATORS 291
In the above we were assuming that n, the sample size, was fixed. Still
another meaning can be affixed to" closeness" if one thinks in terms of increasing
sample size. It seems that a good estimator should do better when it is based
on a large sample than when it is based on a small sample. Consistency and
asymptotic efficiency are two properties that are defined in terms of increasing
sample size; they are considered in Subsec. 3.3. Properties of point estimators
that are defined for a fixed sample size are sometimes referred to as small-sample
properties, whereas properties that are defined for increasing sample size are
sometimes referred to as large-sample properties.
3.2 Mean-squared Error

A useful, though perhaps crude, measure of goodness or closeness of an esti-
mator t(XI' ... , Xn) of 'l{O) is what is called the mean-squared error of the
estimator.
Definition 6 Mean-squared error Let T = t(Xl , " ' , Xn) be an

estimator of 1'(8). Go[[T - 1'(8)]2] is defined to be the mean-squared error
of the estimator T = t(XI' ... , Xn)· IIII
Notation Let MSE,(O) denote the mean-squared error of the estimator
T = t(XI , .•• , Xn) of 1'(8). IIII
Remark The subscript 8 on the expectation symbol G(J indicates from
which density in the family under consideration the sample came. That is,
G6 [[T - 1'(8)]2]
= Ge[[t(Xh "', Xn) - 1'(O)j2]
= J... J[t(XI' •.. , xn) - 1'(O)]~(XI; 0) ••• f(x,,; 8) dxl ••• dx n ,
where f(x; 8) is the probability density function from which the random
sample was selected. IIII
The name" mean-squared error" can be justified if one first thinks of
the difference t - 1'(0), where t is a value of Tused to estimate 1'(8), as the error
made in estimating 1'(8), and then interprets the U mean" in "mean-squared
error" as expected or average. To support the contention that the mean-
squared error of an estimator is a measure of goodness, one merely notes that
ife[[T - 1'(O)Fl is a measure of the spread of T values about 1'(0), just as the'
variance of a random variable is a measure of its spread about its mean. If we
VII
________-1----_. 8
- --MSE'l (6)
.,.--- ...........
MSE, (6)
FIGURE 5
were to compare estimators by looking at their respective mean-squared errors,

naturally we would prefer one with small or smallest mean-squared error. We
could define as best that estimator with smallest mean-squared error, but such
estimators rarely exist. In general, the mean-squared error of an estimator
depends on O.
For any two estimators Tl = 11(X1, ... , Xn) and T2 = 1 2(Xb ... , Xn) of
r(O), their respective mean-squared errors MSEt}(O) and MSEtiO) as functions
of 0 are likely to cross; so for some 0, 11 has smaller MSE, and for others 12 has
smaller MSE. We would then have no basis for preferring one of the estimators
over the other. See Fig. 5.
The following example shows that except in very rare cases an estimator
with smallest mean-squared error will not exist.
EXAMPLE 13 Let Xl' ... , Xn be a random sample from the density f(x; 0),
where 0 is a real number, and consider estimating 0 itself; that is, r(O) = O.
We seek an estimator, say T* = 1*(Xb ... , X n), such that MSEt*(O) <
MSEA.O) for every 0 and for any other estimator T = I(Xb ... , Xn) of O.
Consider the family of estimators T90 = 190(X1, ... , Xn) 00 indexed by =
00 for 00 E e. For each 00 belonging to 8, the estimator 160 ignores the
observations and estimates 0 to be 00 , Note that
MSEtl1 /O) = 6"9[[1 90(X1, ... , Xn) - 0]2]

= $9[(0 0 - 0)2] = (0 0 - 0)2;
so MSEtO()(Oo) = 0; that is, the mean-squared error of 180 evaluated at

0= 00 is O. Hence, if there is to exist an estimator T* = 1*(X1, ... , Xn)
satisfying MSEt*(O) < MSEt(O) for every 0 and for any estimator
I, MSEt*(O) - O. [For any 0 0 , MSEt*(Oo) = 0 since MSEt*(Oo) <
MSE t 00 (0 0 ) = 0.] In order for an estimator 1* to -have its mean-
squared error identically 0, it must always estimate 0 correctly, which
means that from the sample you must be able to identify the true parameter
value. IIII
3 PROPERTIES OF POINT ESTIMATORS 293
One reason for being unable to find an estimator with uniformly smallest
mean-squared error is that the class of all possible estimators is too large-
it includes some estimators that are extremely prejudiced in favor of particular O.
For instance, in the example above t(Jo(X1 , ••• , XII) is highly partial to Bo since
it always estimates 0 to be 00 • One could restrict the totality of estimators by
considering only estimators that satisfy some other property. One such
property is that of unhiasedness.
Definition 7 Unbiased An estimator T = t(Xl' ... , XII) is defined to

be an unbiased estimator of 'r(B) if and only if
for all 0 E e. /11/
An estimator is unbiased if the mean of its distribution equals -,;(B) , the
function of the parameter being estimated. Consider again the estimator
t(Jo(X" "', XII) = eo of the above example; tf(J[t60(X1 , ... , XII)] = tf(J[Bo] =
60 :# 6; so t(Jo(X1 , ••• , XII) is not an unbiased estimator of B. Ifwe restricted the
totality of estimators under consideration by considering only unbiased estima-
tors, we 'Could hope to find an estimator with uniformly smallest mean-squared
error within the restricted class, that is, within the class of unbiased estimators.
The problem of finding an unbiased estimator with uniformly smallest mean-
squared error among all unbiased estimators is dealt with in Sec. 5 below.
Remark
MSEt (6) = var [T] + {'r(B) - tfe[T]}2. (9)
So if Tis an unbiased estimator of 'r(B), then MSEAB) = var [T].
~
PROOF
MSEAO) = 8 e[[T - -,;(B)]2] = tfe[«T - 8 e[T])- {'r(e) - tfe[T]})2]
= 8 e[(T - 8 e[TD2] - 2{'r(O) - 8 e[T]}cf(J[T - cfe[T]]
+ tfe[{-,;(B) - 8 e[T]}2] = var [T] + {-,;(B) - 8(J{T]}2. IIII
The term -,;(B) - tfe[T] is called the bias of the estimator T and can be
either positive, negative, or zero. The remark shows that the mean-squared
error is the sum of two nonnegative quantities; it also shows how the mean-
squared error, variance, and bias of an estimator are related.
EXAMPLE 14 Let Xl~ ... , XII be a random sample from/(x; B) = ¢p.,al(x).

Recall that the maximum~likelihood estimators of p and (12 are, respec-
tively, X and (lin) L (Xi - X)2. (See Example 6.) Now 8(J[X] = p; so
X is an unbiased estimator of 11, and hence the mean-squared error of

X = ct8 [(X - 11)2] = var [X] = a 2ln. We know that ct 9 [8 2] = a 2 ; so
ct 9 [(1/n) L (Xi - X)2] = [(n - 1)ln]ct9 [[1/(n - 1)] L (Xi - X)2]
= [(n - 1)ln]ct9 [8 2 ] = [(n - 1)ln]a2 •
Hence the maximum-likelihood estimator of a 2 is not unbiased. The
mean-squared error of (lin) L (Xi - X)2 is
ct [[(lln) L (Xi - X)2 - a2]2]
9
= var [(lIn) L (Xi - X)2] + {a 2 - ct 9 [(1ln) L (Xi - X)2]}2
=
(n - 1)2 var [8 2
] + (n -1 )2
a2 _ - - a2
n2 n
=
(n - 1)2 1(
2 - 114 -
n- 3
a
4) + a-, 4
n n n- 1 n2
using Eq. (10) of Theorem 2 in Chap. VI. IIII
Remark For the most part, in the remainder of this book we will take
the mean-squared error of an estimator as our standard in assessing the
goodness of an estimator. IIII
3.3 Consistency and BAN

In the previous subsection we defined the mean-squared error of an estimator
and the property of unbiasedness. Both concepts were defined for a fixed
sample size. In this subsection we will define two concepts that are defined
for increasing sample size. In our notation for an estimator of ,(8), let us use
Tn = liX1, ... , Xn), where the subscript n of I indicates sample size. Actually
we will be considering a sequence of estimators, say Tl = 11(X1), T2 = 1 2(X1, X 2)
T3 = 1 3(X1, X 2 , X 3), ... , Tn = I iX1, ... , X n), .... An obvious example is
n
Tn = I iX1, ... , Xn) = X n = (lin) L Xi' Ordinarily the functions In in the
i= 1
sequence will be the same kind of function for each n.
When conSidering a sequence of estimators, it seems that a good sequence
of estimators should be one for which the values of the estimators tend to get
closer to the quantity being estimated as the sample size increases. The following
definitions formalize this intuitively desirable notion of limiting closeness.
Definition 8 Mean..squared-error consistency Let T1, T2 , ... , Tn ...

be a sequence of estimators of ,(8), where Tn = liXl' ... , Xn) is based
on a sample of size n. This sequence of estimators is defined to be a
3 PROPERTIES OF pOINT ESTIMATORS 295
mean-squared-error consistent sequence of estimators of 1:(8), if and only if

lim 8 e[[T,. - ,;(8)]2] = 0 for all 8 in 9. IIII
Remark Mean..squared-error consistency implies that both the bias

and the variance of Tn approach 0 since 8 9 [[Tn - ,(8)]2] = var [Tn]
+ {,(O) - 8 e[Tn]}2. IIII
EXAMPLE 15 In sampling from any density having mean J1 and variance

,.
(12~ let Xn = (lIn) L Xi be a sequence of estimators of J.l and S; =
i= 1
n
[l/(n - 1)] L (Xi - X n) be a sequence of estimators of (12. 8 [(X n - J.l)2] =
i= 1
var [X n] = (121n -40 as n -4 00 ; hence the sequence {X n} is a mean-squared-
error consistent sequence of estimators of J1.
0"[(8; - 0-
2)2] = var [8;] = ~ (P4 -: ~ 0-4) .... 0
as n -+ 00, using Eq. (10) of Chap. VI; hence the sequence {S;} is a
mean-squared-error consistent sequence of estimators of (12. Note that if
Tn = (lIn) I (Xi - X)2, then the sequence {Tn} is also a mean-squared-
error consistent sequence of estimators of (12. IIII
There is another weaker notion of consistency given in the following

definition.
Definition 9 Simple consistency Let T1 , T 2 , ••• , Tn, ... be a sequence

of estimators of 1:"(0), where Tn = t n(Xb ... , Xn). The sequence {Tn} is
defined to be a simple (or weakly) consistent sequence of estimators of ,(8)
if for every 8 > 0 the following is satisfied:
lil!l Pe[,;(8) - 8 < Tn < 1:"(8) + 8] =1 for every 0 in El. I1II
,.-+ 00
Remark If an estimator is a mean-squared-error consistent estimator,

it is also a simple consistent estimator, but not necessarily vice versa.
PROOF
Pe[1:"(O) - 8 < Tn < 1:(8) + e] = P[I Tn - ,(0) I < 8]
= Pe[[Tn - 1:"(8)]2 < 8 2] > I _ tfe[[Tn ~ 1:(0)j2]

e
by the Chebyshev inequality. As n approaches infinity, $8[[Tn - 1"(0)]2]

approaches 0; hence lim Po[1"(e) - e < Tn < 1"(0) + e] = 1. fill
n-oo
We close this subsection with one further large-sample definition.
Definition 10 Best asymptotically normal estimators (BAN estimators)

A sequence of estimators Ti, ... , T:, ... of 1"(8) is defined to be best
asymptotically normal (BAN) if and only if the following four conditions
are satisfied:
(i) The distribution of In[T: -

1"(8)] approaches the normal
distribution with mean 0 and variance U*2(0) as n approaches infinity.
(ii) For every e > 0,
lim po[1
n-oo
T: - 1"(0)1 > e] = 0 for each 0 in e.
(iii) Let {Tn} be any other sequence of simple consistent estimators
for which the distribution of j~[Tn - 1"(8)] approaches the normal dis-
tribution with mean 0 and variance u 2(0).
(iv) u 2 (0) is not less than u*\8) for all 0 in any open interval. fill
Remark The abbreviation BAN is sometimes replaced by CANE,

standing for consistent asymptotically normal efficient. fill
The usefulness of this definition derives partially from tp.eorems proving

the existence of BAN estimators and from the fact that ordinarily reasonable
estimators are asymptotically normally distributed.
It can be shown that for samples drawn from a normal density with
n
mean /l and variance u 2 the sequence T: = (I/n) L Xi = X n for n = I, 2, ...
i:::: 1
is a BAN estimator of /l. In fact, the limiting distribution of J~(X n - /l) is

normal with mean 0 and variance u 2 , and no other estimator can have smaller
limiting variance in any interval of /l values. However, there are many other
estimators for this problem which are also BAN estimators of /l, that is, esti·
mators with the same normal distribution in the limit. For example,
n = 1, 2, ... ,
is a BAN estimator of /l. BAN estimators are necessarily weakly consistent

by (ii) of the definition.
3 PROPERTIES OF pOINT ESTIMATORS 297
3.4 Loss and Risk Functions

In Subsec. 3.2 we used mean-squared error of an estimator as a measure of the
closeness of the estimator to 1"(8). Other measures are possible, for example,
cK(l[1 T - 1"(0)1],
called the mean absolute deviation. In order to exhibit and consider still other
measures of closeness, we will borrow and rely on the language of decision
theory. On the basis of an observed random sample from some density
function, the statistician has to decide what to estimate 1"(8) to be. One might
then caB the val ue of some estimator T = t( Xl' ... , X n) a decision and call the
estimator itself a decision function since it tells us what decision to make. Now
the estimate t of 1"(0) might be in error; if so, some measure of the severity of
the error seems appropriate. The word" loss '" is used in place of "error,"
and" loss function" is used as a measure of the" error." A formal definition
follows.
Definition 11 Loss function Consider estimating 1"(0). Let t denote

an estimate of 1"(0). The loss function, denoted by t(t; 0), is defined to
be a real-valued function satisfying (i) t(t; 8) > 0 for all possible estimates
t and all 8 in e and (ii) t(t; 0). = 0 for t = 1"(0). t(t; 0) equals the loss
incurred if one estimates 1"(8) to be t when 0 is the true parameter value. IIII
In a given estimation problem one would have to define an appropriate

10ssJutlcfion for the particular problem under study. It is a measure of the
error and presumably would be greater for large error than for small error. We
would want the loss to be small; or, stated another way, we want the error in
estimation to be small, or we want the estimate to be close to what it is estimating.
EXAMPLE 16 Several possible loss functions are:

(i) tl(t; 8) = [t - 1"(0)]2.
(ii) t 2 (t; 8) = 1t - 1"(8) I.
(iii) t (t. 8) = fA if 1t - 1"(8) I > e
3, \0 if 1 t - 1"(0) I ~ e, where A > o.
(iv) tit; 0) = p(O) 1t - 1"(0)1' for P(O) > 0 and r > O.
tl is called the squared-error loss function, and t 2 is called the absolute-
error loss function. Note that both t l and t 2 increase as the error t - 1"(8)
increases in magnitude. t 3 says that you lose nothing if the estimate t
is within e units of 1"(0) and otherwise you lose amount A. t 4 is a general
loss function that includes both t l and t 2 as special cases. IIII
We assume now that an appropriate loss function has been defined for our
estimation problem, and we think of the loss function as a measure of error
or loss. Our object is to select an estimator T = t(Xl , •.. , Xn) that makes
this error or loss small. (Admittedly, we are not considering a very important,
substantive problem by assuming that a suitable loss function is given. In
general, selection of an appropriate loss function is not trivial.) The loss
function in its first argument depends on the estimate t, and t is a value of the
estimator T; that is, t = t(Xb " ., xn)· Thus, our loss depends on the sample
Xl' ... , X n • We cannot hope to make the loss small for every possible sample,
but we can try to make the loss small on the average. Hence, if we alter our
objective of picking that estimator that makes the loss small to picking that
estimator that makes the average loss small, we can remove the dependence of
the loss on the sample Xl' "" X n • This notion is embodied in the following
definition.
Definition 12 Risk function For a given loss function t(·; .), the risk
function, denoted by [!llO), of an estimator T = t(Xl , ..• , Xn) is defined
to be
[!liO) = <if(J[t(T; 0)]. (10)
fIll
The risk function is the average loss. The expectation in Eq. (10) can be
taken in two ways. For example, if the density f(x; 0) from which we sampled
is a probability density function, then
C(J[t(T; 0)] = C(J[t(t(Xh ... , X n); 0)]
= f. .. Jt(t(Xb"" ,x n); O)l~ f(Xi; 0) dXjo

Or we can consider the random variable T and the density of T. We get
Co[t(T; 0)] = J
t( ; O)fT(t) dt,
where fT(t) is the density of the estimator T. In either case, the expectation
averages out the values of Xl' . " , X n •
EXAMPLE 17 Consider the same loss functions given in Example 16. The
corresponding risks are given by:
(i) C(J[[T - T(O)j2], our familiar mean~squared error.
(ii) <ifo[! T T(O) 1], the mean absolute error.
(iii) A'Po[!T-T(O)! >8].
(iv) p(O)Co[! T - T(O)I''j. 1/11
4 SUFFICIENCY 299
Our object now is to select an estimator that makes the average loss (risk)
small and ideally select an estimator that has the smallest risk. To help meet
this objective, we use the concept of admissible estimators.
Definition 13 Admissible estimator For two estimators ~ =

t 1(X1, ... , Xn) and T2 = t 2(X1, ... , X n), estimator t1 is defined to be a
better estimator than t 2 if and only if
for all (J in e
and
for at least one (J in e.
An estimator T = t(X1' ... , Xn) is defined to be admissible if and only if
there is no better estimator. IIII
In general, given two estimators t1 and t 2 neither is better than the other;
that is, their respective risk functions as functions of (J, cross. We observed
this same phenomenon when we studied the mean-squared error. Here, as
there, there will not, in general, exist an estimator with uniformly smallest risk.
The problem is the dependence of the risk function on (J. What we might do
is average out (J, just as we average out the dependence on x., ... , Xn when
going from the loss function to the risk function. The question then is: Just
how should (J be averaged out? We will consider just this problem in Sec. 7
on the Bayes estimators. Another way of removing the dependence of the risk
function on (J is to replace the risk function by its maximum value and compare
estimators by looking at their respective maximum risks, naturally preferring
that estimator with smallest maximum risk. Such an estimator is said to be
minimax.
Definition 14 Minimax An estimator t* is defined to be a minimax

estimator if and only if sup 9ft *«(J) < sup 9ft<(J) for every estimator t. 1III
() ()
Minimax estimators will be discussed in Sec. 7.
4 SUFFICIENCY
Prior to continuing our pursuit of finding best estimators, we introduce the
concept of sufficiency of statistics. In many of the estimation problems that
we will encounter, we will be able to summarize the information in the sample
'-------
Xl' ... , Xn· That is, we will be able to find some function of the sample that
tells us just as much about () as the sample itself. Such a function would be
sufficient for estimation purposes and accordingly is called a sufficient statistic.
Sufficient statistics are of interest in themselves, as well as being useful in
statistical inference problems such as estimation or testing of hypotheses.
Because the concept of sufficiency is widely applicable, possibly the notion
should have been isolated in a chapter by itself rather than buried in this chapter
on estimation.
4.1 Sufficient Statistics

Let Xl, .. ' , Xn be a random sample from some density, say f(' ; (}). We defined
a statistic to be a function of the sample; that is, a statistic is a function with
domain the range of values that (Xl' "', Xn) can take on and counterdomain
the real numbers. A statistic T = t(Xl , ... , Xn) is also a random variable; it
condenses the n random variables Xl, X 2 , ... , Xn into a single random variable.
Such condensing is appealing since we would rather work with unidimensional
quantities than n-dimensional quantities. We shall be interested in seeing if we
lost any "information" by this condensing process. The condensing can also be
viewed another way. Let X denote the range of values that (Xl, ... , Xn) can
assume. For example, if we sample from a Bernoulli distribution, then X is a
collection of all n-dimensional vectors with components either 0 or 1; or if we
sample from a normal distribution, then X is an n-dimensional euclidean space.
Now a statistic induces or defines a partition of X. (Recall that a partition of X
is a collection of mutually disjoint subsets of X whose union is X.) Let
t(·, ... , .) be the function corresponding to the statistic T = t(Xl , ... , Xn).
The partition induced by t(·, ... , .) is brought about as follows: Let to denote
any value of the function t(·, ... , .); that subset of X consisting of all those
points (Xl' ... , xn) for which t(Xl, ... , xn) = to is one subset in the collection
of subsets which the partition comprises; the other subsets are similarly formed
by considering other values of t(·, ... , '). For example, if a sample of size 3
is selected from a Bernoulli distribution, then X consists of eight points
(0,0,0), (0,0,1), (0, 1,0), (1, 0,0), (0, 1, 1), (1, 0, 1), (1, 1,0), (1,1, 1). Let
t(Xb X2, X3) = x] + x 2 + x 3 ; then t(·, " .) takes on the values 0,1,2, and 3.
The partition of X induced by t(·, ., .), consists of the four subsets {CO, 0, O)},
{CO, 0,1), (0, 1,0), (1, 0, O)}, {CO, 1, 1), (1, 0, 1), (1, 1, O)}, and {(l, 1, 1)} corre-
sponding, respectively, to the four values 0, 1, 2, and 3 of t(·, " '). A statistic
then is really a condensation of X. In the above example, if we use the statistic
t(·, " .), we have only four different values to worry about instead of the eight
different points of X.
4 SUFFICIENCY 301
Several different statistics can induce the same partition. In fact, if

t(., ... , .) is a statistic, then any one~to-one function of t has the same partition
as t. In the example above t'(xl , Xl, X3) = 6(Xl + Xl + X3)2, Or even
t"(x ,X , X ) = xi + X~ + x~, induces the same partition as t(Xl' X2' X3) =
X +1 X ~ X 3. One of the reasons for using statistics is that they do condense
l and ff suc~ is our only reason for using a statistic, then any two statistics with
the same partition are of the same utility. The important aspect of a statistic
x
is the partition of that it induces, not the values that it assumes.
A sufficient statistic is a particular kind of statistic. It is a statistic that
condenses ~ in such a way that no "" information about 0" is lost. The only
information about the parameter 0 in the density f( . ; 0) from which we sampled
is contained in the sample Xl' ... , Xn; so, when we say that a statistic loses no
information, we mean that it contains all the information about 0 that is con-
tained in the sample. We enlphasize that the type of information of which we
are speaking is that information about 0 contained in the sample given that we
know the form of the density; that is, we know the function f('; .) in f(' ; 0),
and the parameter 0 is the only unknown. We are not speaking of information
in the sample that might be useful in checking the validity of our assumption
that the density does indeed have formf('; ').
Now we shall formalize the definition of a sufficient statistic; in fact, we
shall give two definitions, namely, Definitions 15 and 16. It can be argued that
the two definitions are equivalent, but we will not do it.
Definition 15 Sufficient statistic Let Xl' ... , Xn be a random sample

from the density f('; 0), where 0 may be a vector. A statistic S =
O(Xb ... , Xn) is defined to be a sufficient statistic if and only if the con~
ditional distribution of Xl' ... , Xn given S = s does not depend on 0 for
any value s of s. IIII
Note that we use S = o(Xl' ... , X n), instead of T = t(Xl' ... , X n), to
denote a sufficient statistic. Some care is required in interpreting the condi-
tional distribution of Xl' ... , Xn given S = s, as Example 19 and the paragraph
preceding it demonstrate.
The definition says that a statistic S = o(Xl' ... , Xn) is sufficient if the
conditional distribution of the sample given the value of the statistic does not
depend on O. The idea is that if you know the val ue of the sufficient statistic,
then the sample values themselves are not needed and can tell you nothing more
about 0, and this is true since the distribution of the sample given the sufficient
statistic does not depend on O. One cannot hope to learn anything about 0 by
sampling from a distribution that does not depend on O.
302 PARAMETRIC POIN'l' ESTIMATION VB
EXAMPLE 18 Let Xb X 2 , X3 be a sample of size 3 from the Bernoulli

distribution. Consider the two statistics S = o(X., X 2 , X 3) = Xl
+ X 2 + X3 and T = t(XI , X 2 , X 3) = X I X 2 + XJ . We will show that
0(', " .) is sufficient and t(·, " .) is not. This first column of Fig. 6 is X.
Values of Values of
S T j X l.X2,X3JS j X l,x2,X3I T
1-p
(0,0,0) 0 0 1
l+p
1 I-p
(0,0, 1) 1 1
3 1 +2p
1 p
(0, 1,0) 1 0
3 l+p
1 p
(l, 0, 0) 1 0
3 l+p
1 p
(0,1,1) 2 1
3 1 + 2p
1 P
(1,0,1) 2 1
3 1 +2p
1 p
(l, 1,0) 2 1
3 1 +2p
(1,1,1) 3 2 1 1
FIGURE 6
The conditional densities given in the last two columns are routinely
calculated. For instance,
P[X I = 0; X 2 = 1; X 3 = 0; S = 1]
P[S = 1]
(1 - p)P(1 - p) 1.
= G)p(l - p)2 = 3'
4 sumClENCY 303
and
o _P[X I =0;X z =I;X 3 =0;T=0]
IXI. X2.X3IT==o(0, 1,01 )- P[T = 0]
(1 _ p)zp p p
-
- (1 - p)3 + 2(1 - p)zp 1- p + 2p 1+P
The conditional distribution of the sample given the values of S is inde-

pendent of p; so Sis a sufficient statistic; however, the conditional distribu-
tion of the sample given the values of Tdepends on p; so Tis not sufficient.
We might note that the statistic T provides a greater condensation of X
than does S. A question that might be asked is: Is there a statistic which
provides greater condensation of I than does S which is sufficient as well ?
The answer is "no" and can be verified by trying all possible partitions of
X consisting of three or fewer subsets. IIII
In the case of sampling from a probability density function, the meaning

of the term "the conditional distribution of Xl' ... , Xn given S = s that H
appears in Definition 15 may not be obvious since then P[S = s] = O. We can

give two interpretations'. The first deals with the joint cumulative distribution
function and uses Eq. (9) of Subsec. 3.3 in Chap. IV; that is, to show that
S = o(XI' ... , Xn) is sufficient, one shows that P[XI Xl;"'; X" ~ xnl S = s]
is independent of 0, where P[XI < Xl; ... ; X" < Xn 1S = s] is defined as in
Eq. (9) of Chap. IV. The second interpretation is obtained if a one-to-one
transformation of Xl' X z , ... , Xn to, say, S, Yz , ... , Yn is made, and then it is
demonstrated that the density of Yz , ... , Yn given S = s is independent of O. If
the distribution of Yz , ... , Yn given S = s is independent of 0, then the distribu-
tion of S, Yz , ... , Yn given S = s is independent of 0, and hence the distribution
of Xl' X z , •.. , Xn given S = s is independent of O. These two interpretations
are illustrated in the following example.
EXAMPLE 19 Let Xl' "" X,. be a random sample from f(-; 0) = ¢(J.I(·);
that is, X h ... , Xn is a random sample from a normal distribution with
mean 0 and variance unity. In order to expedite calculations, we take
n = 2. Let us argue that S = o(Xl' Xz) = Xl + X z is sufficient using
the second interpretation above. The transformation of (Xl' Xz) to
(S, Y z), where S = Xl + X z and Yz = X z - Xh is one-to-one; so it
suffices to show that fY2 IS(Yz Is) is independent of O. Now
I Y2Is(Yzl s) = fr2.s(Yz,S) = fY2(yZ)fS<s) = f ( )

fs(s) Is(s) Y2 Yz
304 PARAMETRIC POINT ESTIMATIoN vn
(using the independence of Xl + X 2 and X2 - Xl that was proved in

Theorem 8 of Chap. VI), but
since
Y2 ,..,., N(O, 2),
which is independent of O.
The necessary calculations for the first interpretation above are
less simple. We must show that P[XI < Xl; X 2 x21 S = s] is inde-
pendent of O. According to Eq. (9) of Chap. IV,
P[XI < Xl; X 2 < x 2 1 S = s] = lim P[XI < Xl; X 2 < x21s - h < S < s + h].
h ...... o
Without loss of generality, assume that Xl < X2 . We have the following

three cases to consider: (i) s < X., (ii) Xl < s < X2 , and (iii) X2 < s.
P[XI < Xl; X 2 ::::;; Xli S = s] is clearly 0 (and hence independent of 0) for
case (iii). Let us consider (i). [Case Oi) is similar.]
P[XI < Xl; X 2 < Xli S = s]

= lim P[XI Xl; X 2 < x 2 1s - h < S < s + h]
h ...... o
. 1
hm 2h P[XI Xl; Xl < Xl;S - h < S < s + h]
h-+o
- ----------------------------------
lim Zl P[s - h < S < S + h]
h ...... o h
. 1
hm Zh P[XI < Xl; X 2 X2; s - h < S < s + h]
h ...... O
------------------------------------
fs(s)
Note that (see Fig. 7)
1
lim -
h ...... O Zh
J f
Xl
S-h-X2
s+h-Il
s-h-Il
!Xl(U)!X2(V) dv du,
4 SUFFICIENCY 305
v
___ X2 _ _ _ _ _ _ _ _ _ _ _ + __ _
I
u+
-41---_u
FIGURE 7
and hence
Finally, then,
P[XI < Xl; X 2 < x21S = s]

J:~X2 !Xt(U)!X2(S - U) du
-
!S(S)
2
J:~x2(1/2n)e -![u - 2uO+6 2+(s-u)2_ 2(s-u)6+6 2] du
(1/ J 4n)e -![(s- 26)/"'2]2

(I/J~) J:~X2e-![U2+(S-U)2] du
(1 IJ'Z)e- !(s2/2)
which is independent of O. 11//
306 pARAMETRIC POINT ESTIMATIoN VII
Definition 15 of a sufficient statistic is not very workable. First, it does

not tell us which statistic is likely to be sufficient, and, second, it requires us to
derive a conditional distribution which may not be easy, especially for con-
tinuous random variables. In Subsec. 4.2 below, we will present a criterion
that may aid us in finding sufficient statistics.
Although we will not so argue, the following definition is equivalent to
Definition 15.
Definition 16 Sufficient statistic Let Xb ... , Xn be a random sample

from the density f(·; e). A statistic S = d(Xl' ... , Xn) is defined to be a
sufficient statistic if and only if the conditional distribution of T given S
does not depend on e for any statistic T = t(Xb ... , Xn). IIII
Definition 16 is particularly useful in showing that a particular statistic

is not sufficient. For instance, to prove that a statistic T' = t'(Xl , ... , Xn) is
not sufficient, one needs only to find another statistic T = t(Xb ... , Xn) for
which the conditional distribution of T given T' depends on e.
For some problems, no single sufficient statistic exists. However, there
will always exist jointly sufficient statistics.
Definition 17 Jointly sufficient statistics Let Xl' "', Xn be a random

sample from the density f(' ; e). The statistics S1, ... , Sr are defined to be
jointly sufficient if and only if the conditional distribution of Xb ... , Xn
given Sl = Sl' ... , Sr = Sr does not depend on e. IIII
The sample Xl, ... , Xn itself is always jointly sufficient since the condi-
tional distribution of the sampJe given the sample does not depend on e. Also,
the order statistics Yl , ... , Yn are jointly sufficient for random sampling. If the
order statistics are given, say, by (Yl = Yl, ... , Yn = Yn), then the only values
that can be taken on by (Xl' ... , Xn) are the permutations of Yl, ... , Yn' Since
the sampling is random, each of the n! permutations is equally likely. So, given
the values of the order statistics the probability that the sample equals a partic-
ular permutation of these given values of the order statistics is lIn!, which is
independent of e. (Sufficiency of the order statistics also follows from Theorem
5 belOW.)
If we recall that the important aspect of a statistic or set of statistics is the
partition of x that it induces, and not the values that it takes on, then the
validity of the following theorem is evident.
4 SUFFICIENCY 307
Theorem 3 If SI = 01(X1, ... , Xli)' " ., S,. = o,,(Xb ... , Xli) is a set of
jointly sufficient statistics, then any set of one-to-one functions, or trans-
formations, of 8 1 , •.. , S,. is also jointly sufficient. I111
For example, if L Xi and L X; are jointly sufficient, then X and

L (Xi - X)2 = L xl -" nX2 are also jointly sufficient. Note, however, that
X 2 and L (Xi - X)2 may not be jointly sufficient since they are not one~to-one
functions of L Xi and L X;.
e
We note again that the parameter that appears in any of the above three
definitions of sufficient statistics can be a vector.
4.2 Factorization Criterion

The concept of sufficiency of statistics was defined in Definitions 15 to 17 above.
In many cases, a relatively easy criterion for examining a statistic or set of
statistics for sufficiency has been developed. This is given in the next two
theorems, the proofs of which are omitted.
Theorem 4 Factorization theorem (single sufficient statistic) Let X h

X 2 ' ... , Xli be a random sample of size n from the density f(·; e), where
the parameter e may be a vector. A statistic S = o(Xl , ... , Xli) is sufficient
II
ifand only if the joint density of X b ... , Xli' which is nf(x i ; e), factors as
i 1
fXt, ''', Xn(X h ... , Xli; e) = g(O(Xl' ... , Xli); e)h(Xb ... , Xli)
= g(S; e)h(Xl' ... , Xli)' (II)
where the function h(Xl' ... , Xli) is nonnegative and does not involve the
parameter e and the function g(O(Xh ... , Xli); e) is nonnegative and
depends on Xl, ... , Xli only through the function 0(', ... , '). IIII
Theorem 5 Factorization theorem (jointly sufficient statistics) Let Xb

X 2 , ... , Xli be a random sample of size n from the density f(· ; (J), where
the parameter e may be a vector. A set of statistics Sl = 0l(Xl , ... , Xn),
•.. , S,. = o,,(Xh ... , Xli) is jointly sufficient if and only if the joint density
of X h •.. , Xli can be factored as
fxt, ... ,Xn(Xl' ... , Xli; 0)
= g(Ol(Xb ... , Xli)' .•. , o,,(Xb ... , xJ; (J)h(Xh ., ., Xli) (12)
= g(Sb ... , s,.; (})h(Xl' ... , Xli)'
where the function h(Xl' ... , xn) is nonnegative and does not involve the
parameter fJ and the function g(Sb ... , sr; fJ) is nonnegative and depends
on Xl' ... , Xn only through the functions .11(', ... , .), .. " dr(", .. " .). II/I
Note that, according to Theorem 3, there are many possible sets of suffi-
cient statistics. The above two theorems give us a relatively easy method for
judging whether a certain statistic is sufficient or a set of statistics is jointly
sufficient. However, the method is not the complete answer since a particular
statistic may be sufficient yet the user may not be clever enough to factor the
joint density as in Eq. (11) or (12). The theorems may also be useful in discover-
ing sufficient statistics.
Actually, the result of either of the above factorization theorems is in-
tuitively evident if one notes the following: If the joint density factors as
indicated in, say, Eq. (12), then the likelihood function is proportional to
g( Sl' ... , Sr; fJ), which depends on the observations Xb .. " Xn only through
.11' .. " dr [the likelihood function is viewed as a function of fJ, So h(xb ... , xn)
is just a proportionality constant], which means that the information about
fJ that the likelihood function contains is embodied in the statistics .11(', ... , .),
••• , d r(', ••• , ').
Before giving severa~ examples, we remark that the function h(·, "', .)
appearing in either Eq. (11) or (12) may be constant.
EXAMPLE 20 Let Xb ... , Xn be a random sample from the Bernoul1i density

with parameter fJ; that is,
and
Then
n n
.n !(Xi; fJ) = n fJxi(I -
,=1 i=l
e)l-xiJ {O, l}(Xi)
= fJl:Xi(1 - fJt-l: Xi nl{O,l}(Xi)'

n
i=l
n
If we take fJl:Xi(I - fJt -l:x, as g(d(Xl' "', x n); fJ) and n
i= 1
I{O,l}(Xj) as
h(Xl' .. " xn) and set d(Xl' ... , xn) = L
Xi' then the joint density of
Xb ... , Xn factors as in Eq. (11), indicating that S = d( Xl' ... , Xn) = Xi L
is a sufficient statistic. IIII
4 SUFFICIENCY 309
EXAMPLE 21 Let Xl' ... , XII be a random sample from the normal density
with mean Jl and variance unity. Here the parameter is denoted by Jl
instead of 8. The joint density is given by
n
Ix 10 •••• X n( Xl' •. " x.. ; Jl) =

,.
n ¢p.l(Xj)
i= 1
= Ii J2n
i=l
1 exp [- ~ (Xi - Jl)2]
2
1
= (2n)II/2 exp
[1
- 2i~l
II
(Xi - Jl)
2]
= (2,,~nI2 exp [ - ~ (L: xl - 2Jl L: x, + nJl2)]
= (2nt/2
1 exp (II" X. -
r i...J ,
~2 Jl2) exp (- 2~ i...J
'\~ x~).
I
Ifwetakeh(xl"'" XII) = [l/(2n)II/2] exp (-lIxf) andg(O(xl' ... , XII); Jl) =

exp £J.L I Xi - (nI2)Jl2], then the joint density has been factored as in
Eq. (11) with O(Xl' ... , XII) = I Xi; hence I Xl is a sufficient statistic.
(Recall that X II is also sufficient since any one-to-one function of a suffi-
cient statistic is also sufficient.) II/I
EXAMPLE 22 Let Xl' ••. , Xn be a random sample from the normal density
¢p, a2( . ). Here the parameter () is a vector of two components; that is,
8 = (j.t, u). The joint density of Xl, ... , XII is given by
iQ <P
II
p ,a2(X i ) = I\ J2nu
n 1 [1 (X' - Jl) 2]
exp - 2 I U
= (2:),,12 ,,-0 exp [ _ ~ L: (XI ~ Jl) 2]

= (2,,~/2 ,,-0 exp [ - 2!2 (L: xl - 2jl L: XI + nJl2)] ;
so the joint density itself depends on the observations Xl' ... , XII only
through the statistics 01(Xh ... , XII) = L Xi and 02(X I , ... , XII) = I x;;
that is, the joint density is factored as in Eq. (12) with h(Xb ... , xn) = 1.
Hence, L Xi and I Xl are jointly sufficient. It can be shown that
X It and S2 = [l/(n - 1)] I (Xi - X)2 are one-to-one functions of I Xi
and L xl; so XII and S2 are also jointly sufficient. 11//
EXAMPLE 23 Let Xl' ... , X,. be a random sample from a uniform distribu-
tion over the interval [01 , O2 ], The joint density of Xh •.. , X,. is given by
,. 1
IXI ..... X/Xl' ••. , x,.; Oh ( 2 ) = n0
i= 1 2 -
0 I[fJt.fJ2](Xi)
1
1 n
= (0
2 -
0 ),.
1
n l[fJ
i:;l
I • fJ2](Xj)
1
= (0 _ ( )" l[fJt.1n](Yl)1[;Pl,fJ2](Y")'
2 1
where
and y,. = max [Xb ... , X,.].
The joint density itself depends on Xh " ' , x,. only through Y1 and y,.;
hence it factors as in Eq. (12) with h(X1' ... , X,.) = 1. The statistics Y1
and Y,. are jointly sufficient. Note that if we take 01 = 0 and O2 = 0 + 1,
then Y1 and Y,. are still jointly sufficient. However, if we take 01 = 0 and
O2 = 0, then our factorization can be expressed as
1
= 0" 1[0, fJ](Y,.)I [0, 1n](Y 1)'
Taking g(a(xb ... , X,.); 0) = (1/8")1[0. fJ](yn) and h(Xb ... , xn) = 1[0. Yn](Yl),
we see that Y,. alone is sufficient. IIII
The factorization criterion of Eqs. (11) and (12) is primarily useful in

showing that a statistic or set of statistics is sufficient. It is not useful in
proving that a statistic or set of statistics is not sufficient. The fact that we
cannot factor the joint density does not mean that it cannot be factored; it could
be that we are just not able to find a correct factorization.
If we go back and look through our examples on maximum-likelihood
estimators (see Examples 5 to 8), we will see that all the maximum-likelihood
estimators that appear there depend on the sample Xl' ... , X,. through sufficient
statistics. This is not something that is characteristic of the relatively simple
examples we had given but something that is true in general.
Theorem 6 A maximum-likelihood estimator or set of maximum-

likelihood estimators depends on the sample through any set of jointly
sufficient statistics.
4 SUFFICIENCY 311
PROOF If Sl = 01(Xb ... , X n), "" Sk = 0k(Xb ... , Xn) are jointly
sufficient, then the likelihood function can be written as
= nn
£=1
!(Xi; 0)
= g(ol(XI , ••• , x n), ••• , 0k(XI , .,., x n); O)h(XI' ... , x n)·
As a function of 0, L(O; Xl' ... , xn) will have its maximum at the same
place that g(SI' ..• , Sk; 0) has its maximum, but the place where 9 attains
its maximum can depend on Xl' ... , Xn only through Sb ... , Sk since 9
~~. HH
We might note that method-of-moment estimators may not be functions
of sufficient statistics. See Examples 4 and 23.
4.3 Minimal Sufficient Statistics

When we introduced the concept of sufficiency, we said that our objective was to
condense the data without losing any information about the parameter. We
have seen that there is more than one set of sufficient statistics. For example,
in sampling from a normal distribution with both the mean and variance un-
known, we have noted three sets of jointly sufficient statistics, namely, the sample
Xl' ... , Xn itself, the order statistics YI , .•. , Yn , and X and 8 2 . We naturally
prefer the joi ntly sufficient set X and 8 2 since they condense the data more than
either of the other two. (Note that the order statistics do condense the data.)
The question that we might ask is: Does there exist a set of sufficient statistics
that condenses the data more than X and 8 2 ? The answer is that there does
not, but we will not develop the necessary tools to establish this answer. The
notion that we are alluding to is that of a minimum set of sufficient statistics,
which we label minimal sufficient statistics.
We noted earlier that corresponding to any statistic is the partition of ~
that it induces. The same is true of a set of statistics; a set of statistics induces
a partition of ~. Loosely speaking, the condensation of the data that a statistic
or set of statistics exhibits can be measured by the number of subsets in the
partition induced by that statistic or set of statistics. If a set of statistics has
fewer subsets in its induced partition than does the induced partition of another
set of statistics, then we say that the first statistic condenses the data more than
the latter. Still loosely speaking, a minimal sufficient set of statistics is then a
sufficient set of statistics that has fewer subsets in its partition than the induced
312 pARAMETRIC POINT ESTIMATION VII
partition of any other set of sufficient statistics. So a set of sufficient statistics

is minimal if no other set of sufficient statistics condenses the data more. A
formal definition is the following.
Definition 18 Minimal sufficient statistic A set of jointly sufficient

statistics is defined to be minimal sufficient if and only if it is a function of
every other set of sufficient statistics. IIII
Like many definitions, Definition 18 is of little use in finding minimal

sufficient statistics. A technique for finding minimal sufficient statistics has
been devised by Lehmann and Scheffe [19], but we will not present it. If the
joint density is properly factored, the factorization criterion will give us minimal
sufficient statistics. All the sets of sufficient statistics found in Examples 20
to 23 are minimal.
4.4 Exponential Family

Many of the parametric families of densities that we have considered are
members of what is called the exponential class, or exponentialfamity, not to be
confused with the negative exponential family of densities which is a special case.
Definition 19 Exponential family of densities A one-parameter family

(lJ is unidimensional) of densities f(· ; lJ) that can be expressed as
f(x; lJ) = a(lJ)b(x) exp [c(lJ)d(x)] (13)
for - 00 < x < 00, for all lJ E e, and for a suitable choice of functions
a(')' b('), c(·), and d(·) is defined to belong to the exponential family or
exponential class. IIII
EXAMPLE 24 If f(x; lJ) = 8e- ex/(0. ooix), then f(x; lJ) belongs to the expo-
nential family for a(fJ) = lJ, b(x) = /(0. tt:l)(x), c(lJ) = - 8, and d(x) = x in
Eq. (13). IIJI
EXAMPLE 25 Iff(x; lJ) =f(x; A) is the Poisson density, then

e-A.1x
f(x; A) = ,/{o.l ... i x )
x.
4 SUFFICIENCY 313
A
In Eq. (13), we can take a(A) = e- , b(x) = (1Ix!)I{o.l •... }(x), C(A) = log A,
and d(x) = x; so f(x; A) belongs to the exponential family. /III
Remark If f(x; 0) = a(O)b(x) exp [c«(J)d(x)], then
,U f (x,; 6) = a'(6{Jj, b(X,)] exp [c(6),t,d(x,)].

and hence by the factorization criterion L d(Xi) is a sufficient statistic. IIII
. The above remark shows that, under random sampling, if a density belongs
to the one-parameter exponential family, then there is a sufficient statistic. In
fact, it can be shown that the sufficient statistic so obtained is minimal.
The one-parameter exponential family can be generalized to the k-param-
eter exponential family.
Definition 20 k-parameter exponential family A family of densities

f( . ; (h, ... , Ok) that can be expressed as
k
f(x; (}l' ... , (}k) = a({}l, ... , (}k)b(x) exp L Ci{}l' ... , (}k)d/x) (14)
j=l
for a suitable choice of functions a(·, ... , .), b(·), ci·, ... , .), and dj (·),
j = 1, ... , k, is defined to belong to the exponential family. IIII
In Definition 20, note that the number of terms in the sum of the exponent
is k, which is also the dimension of the parameter.
EXAMPLE 26 Iff(x; {}l, (}2) = cP". 2 (x) , where (Ol' (}2)

(I = (Ji, a), thenf(x; {}1I (}2)
belongs to the exponential family.
1 exp (1
= fo
2na
- -2 . Ji
a
2
2
) exp (1
-
2a
-2 x2 + ..!!..-2 x
a
)
.
Take a(p, a) = (1/J2na) exp (--!- . Ji 2I( 2), b(x) = 1, Cl(P, a) = -1/2a2,
2
C2(P, a) = Jila , dl(x) = x 2, and d2(x) = x to show that cP",(l2(X) can be
expressed as in Eq. (14). IIII
EXAMPLE 27 If
then
J(X; °
1, (
1
2) = B(Ol' 0l) 1(0,1) (x) exp [(Ot - 1) log x + (0 2 - I) log (1 - x)];
so J(x; °1, ( 2) belongs to the exponential family with a(Ol' ( 2 ) =

I/B(Ol, ( 2 ), b(x) = 1(0, l)(X), Ct (0) = 01 - 1, ciO) = O2 - I, d1(x) =
log x, and d2(x) = log (1 - x). IIII
k
Remark Iff(x; Of, ... , Ok) = a(Ol, ... , 0k)b(X) exp I
j=l
C/Oh ... , Ok)d/x),
then, under random sampling,
n f (Xi; ° °k)
II
1, . . .,
i= 1
II n
and so by the factorization criterion I d1(Xi ), "', I dk(Xi ) is a set of
£:=1 i=1
jointly sufficient statistics. I d1(X,), ... , I dk(Xi) are in fact minimal
sufficient statistics. 1I11
II II
EXAMPLE 28 From Example 27, we see that I log Xi and I log (1 - Xl)
i=l 1==1
are jointly minimal sufficient when sampling from a beta density. 1111
Our main use of the exponential family will not be in finding sufficient
statistics, but it will be in showing that the sufficient statistics are complete, a
concept that is useful in obtaining" best" estimators. This concept will be
defined in Sec. 5.
Lest one get the impression that all parametric families belong to the
exponential family, we remark that a family of uniform densities does not
belong to the exponential family. In fact, any family of densities for which the
range of the values where the density is nonnegative depends on the parameter
does not belong to the exponential class.
°
s UNBIASED ESTIMATIoN 315
5 UNBIASED ESTIMATION
Since estimators with uniformly minimum mean-squared error rarely exist, a
reasonable procedure is to restrict the class of estimating functions and look
for estimators with uniformly minimum mean-squared error within the restricted
class. One way of restricting the class of estimating functions would be to
consider only unbiased estimators and then among the class of unbiased esti-
mators search for an estimator with minimum mean-squared error. Con-
sideration of unbiased estimators and the problem of finding one with uniformly
minimum mean-squared error are to be the subjects of this section.
According to Eq. (9) the mean-squared error of an estimator T of 1'(6)
can be written as
t!8[[T - 1'(6)]2] = var6 [T] + {1'(0) - t!8[T]}2,
and if T is an unbiased estimator of 1'(0), then t!8[T] = 1'(8), and so
t!6[[T - 1'(0)]2] = varo [T]. Hence, seeking an estimator with uniformly
minimum mean-squared error among unbiased estimators is tantamount to
seeking an estimator with uniformly minimum variance among unbiased
estimators.
Definition 21 Uniformly minimum-variance unbiased estimator (UMVUE)

Let Xl' ... , Xn be a random sample from f(·; 0). An estimator
T* = t*(XI' "', XJ of 1'(0) is defined to be a uniformly minimum-variance
unbiased estimator of 1'(0) if and only if (i) t!8[T*] = 1'(0), that is, T* is
unbiased, and (ii) var8 [T*] < var6 [T] for any other estimator T =
t(XI ,·.·, XII) of 1'(0) which satisfies t!o[T] = 1'(0). IIII
In Subsec. 5.1 below we will derive a lower bound for the variance of
unbiased estimators and show how it can sometimes be useful in finding an
UMVUE. In SUbsec. 5.2 we will introduce the concept of completeness and
show how it in conjunction with sufficiency can sometimes be used to find an
UMVUE.
5.1 Lower Bound for Variance

Let X b ··., Xn be a random sample from/(·; (J), where 6 belongs to e. Assume
that e is a subset of the real line. Let T = t(Xh ... , Xn) be an unbiased
estimator of 1'(0). We will consider the case where f(·; 0) is a probability
density function; the development for discrete density functions is analogous.
We make the following assumptions, caned regularity conditions:
(i) aoa log f(x;O) exists for all x and all O.
316 PAltAMETRlC POINT ESTIMATIoN vn
(ii) ! r··f ,a !(x,; 0) dx, ... dx.
= r. f:O,a!(X,;9)dX, "·dx•.
(]'I']') 8 f . ..
ae ft{ Xl' ••• , XII )IT" f( Xi; e) dXl , .• dXn
i= 1
= r. f t(x" ... , x.) ! ,D/(X,; IJ) dx, ... dx•.
(iv) 0 < S. [[:0 logj(X; 0)],] < 00 for all 9 in e.
Theorem 7 Cramer-Rao inequality Under assumptions (i) to (iv)

above
vare [T] (15)
n8'[[:9 10gj(X; /l)] ']
where T = t{Xl' ... , Xn) is an unbiased estimator of ,(e). Equality
prevails in Eq, (15) if and only if there exists a function, say K(e, n),
such that
f 8e8 log!(xi; e) = K(e, n)[t(xh .. "
n
Xn) - ,(e)], (16)
Equation (15) is called the Cramer-Rao inequality, and the right-hand side
is called the Cramer-Rao lower bound for the variance of unbiased estimators
of ,(e).
PROOF
8 ,(e) = 8e
,'(e) = 8e 8 f . . . Jt{Xl" ,IJ" f(Xj; e) dX1 , •. dXII
•. , XII)
1
= J..... f t {X1' •.. , xn) 80

~ [rlf{Xi; e)] dX1 ... dXII
i=1
- T(O) :0 f ... f,U [f(X,; 0) dx;l

= r. f t(X" .... x.) :9 [.D/(X,; 9)] dx, ." dx.
- «9) r. f :9 [,:0, !(x,; 9)] dx, ... dx.
UNBIASED ESTIMAnON 317
5
= r·. I[t(x" ... , x.,} - 1(/1)] :0 Lr"l fix,; 0] <lx, ... <lx,
= I··' I[t(x" •.. , x,) - -r(6)] [:6 IOg.U j(x,; 0)]
= 4'. [[t(X " •.. , X.) - -r(/I)] [:6 log tV(X,; 0)]].
Now by the Cauchy-Schwarz i'nequality
[1'(0)]2 < S.[[t(X " ... , X,) - 1(0)]2]S. [[:0 log ,:0. f(X,; 6)] 'l
or
but
8'[[a~ log,:o. f(X,; 6)rJ = s.[[,t, a~ log/(X,; 6)n

= ~ ~ 8.[[:6 logf(X,; 6)] [:6 10gf(X); 0)]]
= n.r.[[:6 10gf(X; 0)] 2],
using the independence of Xi and Xj and noting that
S.[:o!Og/(X; /I)] = I [! log/(x; /I)]i(x; 0) dx

0 a a
=
f aof(x; (J) dx = 00 ff(x; 0) dx = ao (1) = O.
The inequality in the Cauchy-Schwarz inequality becomes an
equality if and only if one function is proportional to the other; in our case
this requires that :0 log ,:0. /(x,; 0) be proportional to I(x" ... , x.) - 1(0)
or that there exists K = K(O, n) such that
;o!Og Ux /(x,; 0)] = K(O, n)[t(x" ... , xJ - 1(0)]. 1111

318 pARAMETRIC POINT ESTIMATION vn
The regularity conditions, which were stated for probability density

functions, can be modified for discrete density functions, leaving the statement
of the theorem unchanged.
The theorem has two uses: First, it gives a lower bound for the variance
of unbiased estimators. An experimenter using an unbiased estimator whose
variance was close to the Cramer-Rao lower bound would know that he was
using a good unbiased estimator. Second, if an unbiased estimator whose
variance coincides with the Cramer-Rao lower bound can be found, then this
estimator is an UMVUE. Equation (16) aids in finding an estimator whose
variance coincides with the Cramer-Rao lower bound. In fact, if there exists a
T* = t*(Xl' ... , X,.) such that
,. 0
~ ae logf(xi; e) = K(e, n)[t*(xl' ... , xJ - -r*(e)]
for some functions K(e, n) and -r*(e), then T* is an UMVUE of -r*(e).
EXAMPLE 29 Let Xl' ... , X,. be a random sample from f(x; e) =

ee- 9x/(o.OO)(x). Take -r(e) = e. It can be shown that the regularity con-
ditions are satisfied. -r' (e) = 1 ; hence
o 0
Note that oe logf(x; e) = ae (log e - ex) = lIe - x, and so
Hence, the Cramer-Rao lower bound for the variance of unbiased esti-
mators of e is given by
Similarly the Cramer-Rao lower bound for the variance of un-

biased estimators of -r(e) = lIe is given by
s UNBIASED ESTIMATION 319
The left..hand side of Eq. (16) is
t1
!-.log/(x,;
00
e) = t
1
!..(lOg e -
00
eXi) = f.1 (-e - Xi) = -n(xn - ~),
1
By taking K(O, n) = -n and utilizing the result of Eq. (16), we see that
X n is an UMVUE of lIe since its variance coincides with the Cramer-Rao
lower bound. IIII
EXAMPLE 30 Let Xl' ... , Xn be a random sample fromJ(x; 0) =J(x; A) =

e- AA:1I: Ix! for x = 0, I, 2, ....
a a e-.lAx
A). 10gJ(x; A) = a). log x!
a X
= OA (-A + x log A -log x!) = -I + l'
Therefore
... [[~ logf(X; l)n =8 [(~ -1)'] = ;, 8[(X -l)']

1 1 1
= ;.:z var [X] = ).2 A= ~ ,
and so the denominator of the Cramer-Rao lower bound is n/A. Now,

if 1'(A) = e- A = P[X = 0], then the Cramer-Rao lower bound for the
vanance of unbiased estimators of 1'(.1) = e- A is given by var [T] >
n
k- /n. Note that T = (l/n) I I{o)( Xi) is an unbiased estimator of 1'(.1)
2A
i= 1
n n
= e- A since 8[T] = (lIn) 'L 8[I{o}(Xj)] = (l/n) Ie-A = e- A. I{olX;) = 1
i=l 1==1
if Xi = 0, and I(oj(Xi) = 0 otherwise; so 8[I{o}(Xi )] = 1 ' P[Xi = 0]
+ 0 . P[Xj #- 0] = e- A• T is the proportion of observations in the sample
that are equal to O. var [T] (l/n)e-A(l - e- A), as compared to the
Cramer-Rao lower bound, which is (1/n)k-2A. Note that
(l/n)e-A(I - e- A) > (l/n)J.e- 2A,
as it should be. An UMVUE of 1'(A) = e- A is found in Example 34.

We note that I (O/OA) logJ(x,; A) = I (-1 + XdA) = (njA)(x - A);
hence, X is the UMVUE of A by Eq. (16). ///1
In general, the Cramer-Rao lower bound is not an attainable lower bound;

that is, there often exists a lower bound for variance that is greater than the
Cramer-Rao lower bound. We will see several such examples in Subsec. 5.2
below. We win see that an UMVUE can exist whose variance does not coincide
with the Cramer-Rao lower bound.
We conclude this subsection with several remarks, the statements of which
are not necessarily mathematically precise. All the same, the remarks are
important and do relate some earlier concepts to the Cramer-Rao lower bound.
Remark' Under certain assumptions involving the existence of second

derivatives and the validity of interchanging the order of certain differen-
tiations and integrations,
0".[ [:0 iogf( X; O)rJ = -of. [::2 iogf( X; IJ) l IIII
This remark is computationally useful if the first expectation is more

difficult to obtain than the second. The proof is left as an exercise.
Remark If the maximum-likelihood estimate ofB, say (j = S(Xl' "', x n),

is given by a solution to the equation
a a
aB log L(B; Xb ... , Xn) = aB log Jl !(X
n
j ; B) = 0,
and if T* = t*( Xl, ... , Xn) is an unbiased estimator of T*(B) whose
variance coincides with the Cramer-Rao lower bound, then t*(Xb' .. , xn) =
T*(S( Xl, ... , Xn».
PROOF
a n
o= aB log iQ !(x j ; B) (1-9 = K(B, n)[t*(xl' ... , Xn) - T*(B)] (lo:=(j
by Eq. (16) and the definition of e. IIII

This remark tells us that under the conditions of the remark a maximum-
likelihood estimator is an UMVUE!
Remark If T* t*{Xl , ... , Xn) is an unbiased estimator of some

=
T*(B) whose variance coincides with the Cramer-Rao lower bound, then
f(·; B) is a member of the exponential class; and, conversely, if f(· ; B)
is a member of the exponential class, then there exists an unbiased esti-
mator, say T*, of some function, say T*(B), whose variance coincides with
the Cramer-Rao lower bound. 1///
We win omit the proof of this remark. It relates the Cramer-Rao lower
bound to the exponential family; in fact, it tells us that we will be able to find
an estimator whose variance coincides with the Cramer-Rao lower bound if
and only if the density from which we are sampling is a member of the expo-
nential class. Although the remark does not explicitly so state, the following is
true: There is essentially only one function (one function and then any linear
function of the one function) of the parameter for which there exists an unbiased
estimator whose variance coincides with the Cramer-Rao lower bound. So,
what this remark and the comments following it really tell is: The Cramer-
Rao lower bound is of limited use in finding UMVUEs. It is useful only if we
sample from a member of the one-parameter exponential family, and even then
it is useful in finding the UMVUE of only one function of the parameter.
Hence, it behooves us to search for other techniques for finding UMVUEs, and
that is what we do in the next subsection.
5.2 Sufficiency and Completeness

In this subsection we will continue our search for UMVUEs. Our first result
will show how sufficiency aids in this search. Loosely speaking, an unbiased
estimator which is a function of sufficient statistics has smaller variance than an
unbiased estimator which is not based on sufficient statistics. In fact, let
f( . ; 0) be the density from which we can sample, and suppose that we want to
estimate t(O). Let us assume that T = t(XI' " ., Xn) is an unbiased estimator
of t(O) and that S = Q(XI , •.. , Xn) is a sufficient statistic. It can be shown that
another unbiased estimator, denoted by T', can be derived from T such that
(i) T' is a function of the sufficient statistic Sand (ii) T' is an unbiased estimator
of t(8) with variance less than or equal to the variance of T. Therefore, in our
search for UMVUEs we need to consider only unbiased estimators that are
functions of sufficient statistics. We shall formalize these ideas in the following
theorem.
Theorem 8 Rao-Blackwell Let Xh ... , Xn be a random sample from

the density f(· ; 8), and let SI = Ql(X1 , ... , X)
n" •• , S k = Vk (X1,· .. , X)
A
n
be a set of jointly sufficient statistics. Let the statistic T = t(Xb ... , X,,)
be an unbiased estimator of t(8). Define T' by T' = 8[TIS1' ... , Sk]'
Then,
(i) T' is a statistic, and it is a function of the sufficient statistics
Sb ... , Sk' Write T' = t'(Sb ... , Sk).
(ii) 8 o[T'] = t(8); that is, T' is an unbiased estimator of t(8).
(iii) varo [.T'] < varo [T] for every 8, and varo [T'] < vare [T] for
some 8 unless T IS equal to T' with probability 1.
PROOF (i) SI' .•• , Sk are sufficient statistics; so the conditional

distribution of any statistic, in particular the statistic T, given Sf, ... , Sk
is independent of B; hence T' = G[TI Sf, ... , Sk] is independent of B, and
so T' is a statistic which is obviously a function of Sf, ... , Sk'
(ii) G9[T'] = G9[G[TISl' ... , Sd] = G9[T] = r(B) [using Eq. (26) of
Chap. IV]. (iii) We can write
var9[T] = G9[tT-G o[T'])2] =G6[(T- T' +T' - 8 6[T,])2]
= tf6[(T - T,)2] + 2G6[(T - T')(T' - tf6[T'])] + var6 [Tl
But
8 6[(T - T')(T' - 8 6[T'])] = 8 6[tf[(T - T' )(T' - tfll[T'])! Sf, .. " Sk]]'
and
G[(T - T')(T' - G6[T']) ISI = Sl; ... ; Sk = sd
= {t'(Sl' .. " Sk) - 8 6[T']}G[(T - T')I SI = SI; ... ; Sk = sd
= {t'(sp ... , Sk) - 8 o[T']}(t9'[T lSI = SI; ... ; Sk = SkJ
- 8[T' lSI = SI; ... ; Sk = Sk])
= {t'(S1' ... , Sk) - tB'6[T']}[t'(Sj, ... , Sk) - t'(S1, ... , Sk)]
=0,
and therefore
var9 [11 = Go[(T - T')2] + var6 [Tf] > varo [T'].
Note that varo [T] > varo [T'] unless Tequals T' with probability l. II/I
For many applications (particularly where the density involved has only
one unknown parameter) there will exist a single sufficient statistic, say
S = O(Xl' ... , X n), which would then be used in place of the jointly sufficient
set of statistics SI' ... , Sk' What the theorem says is that, given an unbiased
estimator, another unbiased estimator that is a function of sufficient statistics
can be derived and it will not have larger variance. To find the derived statistic,
the calculation of a conditional expectation, which mayor may not be easy, is
required.
EXAMPLE 31 Let Xl' ... , X,. be a random sample from the Bernoulli density
I(x; B) = (]x(I - B)1-;;c for x = 0 or 1. Xl is an unbiased estimator of
r(B) = B. We use Xl as T = t(Xl , ... , XII) in the above theorem. Xi L
is a sufficient statistic; so we use S = I Xi as our set (of one element) of
UNBIASED ESTIMATION 323
sufficient statistics. According to the above theorem T' = t9'[T 1 S1 =

8[X 1L X.] is an unbiased estimator of 0 with no larger variance than
T =1 Xl- Let us evaluate t9'[T 1S1. We first find the conditional distribu-
tion of Xl given L Xi = S. Xl takes on at most the two values 0 and 1.
P[x, =O;.tXi= $]
p[X 1 = OIL Xi = s] = [n p==l ]
P LX' =s
1 i=
- P[tXi
t= 1
=s] -
P[X 1 = 1]· p[ f Xi= -1]

i=2
S O· (n - 11)8
S-
s - 1(1 _ 0t- 1 - s+ 1
S
P[it Xi s] - --(-:)-es-(1-0-r---
s - - ~
We note in passing that the conditional distribution of Xl given L Xl = S

is independent of 8, as it should be. Also, we could have derived the
conditional distribution with much less effort by asking: Given that you
have observed s successes in n trials, what is the probability that the
first trial resulted in a success? This probability is sIn. (See Example 28
in Chap. I.)
324 PAltAMETRlC POINT ESTIMATION vu
hence,
II
L Xi
T' = i=l
n
The variance of Xl is 9(1 - 8), and the variance of T' is 9{ I - 8)ln; so

for n > 1 the variance of T' is actually smaller than the variance of
T = Xl' //1/
Before leaving Theorem 8, two comments are appropriate: First, if the

unbiased estimator T is already a function of only Sl' ... , S'k, then the derived
statistic T' will be identical to T, and hence no improvement in variance can be
expected. Second, although the set of jointly sufficient statistics is an arbitrary
set, in practice one would naturally use a minimal set of jointly sufficient statistics
if such were available.
Theorem 8 tells us how to improve on an unbiased estimator by con-
ditioning on sufficient statistics. For some estimation problems this unbiased
estimator, obtained by conditioning on sufficient statistics, will be an UMVUE.
To aid in identifying those estimation problems for which a derived estimator
is an UMVUE, the concept of completeness of a family of densities is useful.
Definition 22 Complete family of densities Let Xl' ... , Xn denote a

random sample from the density f(·; 9) with parameter space 8, and let
T = t(XI , ... , XII) be a statistic. The family of densities of T is defined
=
to be complete if and only if 8 6[~(T)] 0 for all 9 e e implies that
P6[~(T) = 0] = I for all 9 e 8, where ~(T) is a statistic. Also, the statistic
T is said to be complete if and only if its family of densities is complete.
111/
Another way of stating that a statistic T is complete is the following:

Tis complete if and only if the only unbiased estimator of 0 that is a function of
T is the statistic that is identically 0 with probability 1.
EXAMPLE 32 Let Xl' ... , Xn be a random sample from the Bernoulli

density. The statistic T = Xl - X 2 is not complete since 8 o[Xt - X 2 ] = 0
II
and Xl - X 2 is not 0 with probability I. Consider the statistic T = I

1
Xi'
Let ~(T) be any statistic that is a function of T for which 8 8[~(T)] == 0
5 UNBIASED ESTIMATION 325
for all (J e E'l, that is, for 0 < (} < 1. To argue that Tis complete, we must
show that A:(t) = 0 for t = 0, 1, ... , n. Now
8 9 [4T)] = I A:(t)(n)t 8'(1 -

t=O
6)n-t = (l - fJ)n
t-O
t ~(t)(~) [0/(1 - O)J;
hence, 8 9 [A:(T)] = °
for all 0 < 6 < 1 implies that
or
for all a, where a = 81(1 - 8). Now in order for a polynomial in a to be

identically 0, each coefficient of at, t = 0, ... , n, must be 0; that is,
x( t) (;) = 0 for t = 0, ... , n, but (~) # 0; so x( t) = 0 for t = 0, ... ,n. 11/1
EXAMPLE 33 Let Xh ... , Xn be a random sample from the uniform distri-

bution over the intet:Val (0, 8), where E'l = {6: 0 > O}. Show that the
statistic Yn is conlplete. We must show that if 8 9 [A:(Yn)] == 0 for all 8 > 0,
then Po[A\Yn) = 0] == 1 for all 8 > O.
8 o[A\Yn)] = J4.y)frn(Y) dy = J 4y)0-nnyn-l dy,

(J
and 8o[A:(Yn)] = 0 for all 0 > 0 when and only when

n 9
lln r A:(y)yn-t dy = 0 for all 0> 0
u "'0
or
JA:(y)yn-l dy =0
o
(J
for all 8 > O.
Differentiating both sides of this last identity with respect to 0 produces

A:(8)on-l = 0, which in turn implies that A:(fJ) 0 for 8> o. =//1/
In general, demonstrating completeness can require tricky analysis. The

two above examples are exceptions. We state now, without proof, a theorem
that gives us completeness of a statistic. It will be our main tool for arguing
completeness.
Theorem 9 Let Xl' ... , Xn be a random sample fromf(· ; 8), 8 E 8, where

e is an interval (possibly infinite). If f(x; 6) = a(8)h(x) exp [c(B)d(x)],
that is, f( .; 8) is a member of the one-parameter exponential family,
then L d(Xi ) is a complete minimal sufficient statistic. IIII
Theorem 9 shows once again the importance of the exponential family or
exponential class. We are finally adequately prepared to state the theorem that
is useful in finding UMVUEs.
Theorem 10 Lehmann-Scheffe Let Xb ... , Xn be a random sample

from a density f(·; 8). If S = O'(Xb ... , Xn) is a complete sufficient
statistic and if T* = t*(S), a function of S, is an unbiased estimator of
-r(8), then T* is an UMVUE of -r(8).
PROOF Let T' be any unbiased estimator of -r(B) which is a function

of S; that is, T' = t'(S). Then ~(I[T* - T'] == 0 for a1l 8 E 6, and
T* - T' is a function of S; so by completeness of S, P(I[t*(S) = t'(S)] = 1
for all 8 E 6. Hence there is only one unbiased estimator of -r(8) that
is a function of S. Now let T be any unbiased estimator of -r(8). T*
must be equal to ~[T IS] since ~[T IS] is an unbiased estimator of -r(8)
depending on S. By Theorem 8, var(l [T*] < var(l [T] for all 8 E 6; so
T* is an UMVUE. IIII
Let us review what this important theorem says: First, if a complete
sufficient statistic S exists and if there is an unbiased estimator for -r(8), then
there is an UMVUE for -r(8); second, the UMVUE is the unique unbiased
estimator of -r(B) which is a function of S.
To actually find that unbiased estimator of -r(B) which is a function of S,
we have several ways of proceeding. First, simply guess the correct form of the
function of S that defines the desired estimator. Second, guess or find any
unbiased estimator of -r(8), and then calculate the conditional expectation of
the unbiased estimator given the sufficient statistic. Third, solve for t*(·)
in the equation 8(1[t*(S)] = -r(B). Such an equation becomes the integral
J
equation t*(s)fs(s) ds = -r(8) in the case of a continuous random variable S
and becomes the summation L t*(s)fs(s) == -r(8) for S a discrete random variable.
We will employ two of these methods in the following examples.
EXAMPLE 34 Let Xl' ... , Xn be a random sample from the Poisson density
e-.ll x
f(x; l) = , for x = 0, 1, ....
x.
5 UNBIASED ESTIMATION 327
We saw in Example 25 that/ex; 2) belongs to the exponential family with

d(x) = x. By Theorem 9, the statistic L Xi is complete and sufficient.
To :find the UMVUE of A itself, it suffices to guess a function of L Xi
whose expectation is 2. Noting that 2 is the population mean, (lIn) L Xi
is the obvious choice; so (lIn) L Xi is the UMVUE of A.
Consider now estimating rCA) = e- A = P[Xi = 0]. (Recall Ex-
ample 30.) Let us derive the UMVUE of e- A by calculating the condi-
tional expectation of some unbiased estimator given the sufficient statistic.
Any unbiased estimator will do as the preliminary estimator whose
conditional expectation ~eeds to be calculated; so we may as well choose
one that would make the calculations easy. I{o}( Xl) is an unbiased estimator
of e- A and is relatively simple since it can assume only the two values 0
and 1. By Theorem 10, tf[I{o}(X1) IL Xi] is the UMVUE of e- A• To find
the desired conditional expectation, we first find the conditional distribu-
ti on of Xl given L Xi'
Therefore,
hence
t9'[l{o)(X,)IL Xi =s] = PIX, =OIL Xi =s] = r ~ It
. h A
IS t e UMVUE of e- for n > 1. For n = 1, I{o}(X1) is an unbiased
estimator which is a function of the complete sufficient statistic X, and
hence I(o}(X1) itselfis the UMVUE of e- A• The reader may want to ~erive
the mean and variance of
and compare them with the mean and variance of the estimator
II
(lin) I I{o}(X,) given in Example 30. IIII

i=l
EXAMPLE 35 Let Xl' ... , XII be a random sample from I(x; 0) =

Oe- 8xI(o.oo)(X). Our object is to find the UMVUE of each of the following
functions of the parameter 0: 0, 1/0, and e- K8 = P[X > K] for given K.
Since Oe- 8xI(o. oolx) is a member of the exponential class (see Example 24),
n
the statistic S = I Xi is complete and sufficient.
1
II
X n = (lIn) I Xi'
i I
which is a function of the complete sufficient
II
statistic S = I Xi, is an unbiased estimator of 110; hence by Theorem 10,

i= 1
X II is the UMVUE of 110.
To find the UMVUE of 0, one might suspect that the estimator is
n
of the form elI Xi' where e is a constant which may depend on n. Now
1
for n > 1. So ~9[eII Xi] = 0 when e = n - 1; hence (n - l)/I Xi is the

UMVUE of 0 for n> 1. The variance of (n - l)/I Xi is given by
02j(n - 2) for n > 2.
Although one might be able to guess which function of S = I Xi
is an unbiased estimator for e- K8 , let us derive the desired estimator by
starting with the following simple unbiased estimator of e- K9 : I(K.oo)(X1),
Note that C8[I(K,oolX1 )] = 0 . P[XI < K] + 1 . P[XI > K] = P[XI > K] =
e- K9 ; so I(K, oo)(X1) is indeed an unbiased estimator of e- K8 , and therefore
by Theorems 8 and 10 C 9[I(K, oo)(X1) IS] is the UMVUE of e- K9. Now,
C8 [I(K , oo)(XI ) IS = s] = P[I(K. oo)(X1) = 11 S = s] = P[X1 > K IS = s). In
order to obtain P[XI > K IS = s], we will first find the conditional distribu..
tion of Xl given S = s.
IX1IS • .(Xll s) ~XI

IXbS<XI' s) Axl ~s
- lS<s) As
P[Xl <XI <X1 + Ax l ; S

~~~------~~---.~~------~
<*X, < S+ AS]
(l/r(n)]8"s" Ie 6s ~s
P[XI <XI <XI +Ax1; S- X1 <tXI < S -XI +A5]

~~--------~~~~~~~~----------~
[1/r( n ) ]8"s" Ie 6s ~s
P[x I < XI < XI + Ax I]P [s - XI < t X, < S- XI + AS]

- [l/r(n)]8"s" Ie 68 ~s
for x I < s and n > L
cS'6[I(K,oo)(X 1 )!S=S]=P[X l >K!S=s]= fKoofxdS=.,cXlls)dX1
_ JS r(n) (s - Xt)"-2
- Kr(n-l) S,,-1 dXt
=
n -1
,,-1
fO y"-2( -dy)
s s-K
n- 1 y,,-t s-K
- S,,-l n - 1 0
for s > K and n > 1, where the substitution y =s- Xl was made. Hence,
is the UMVUE for e-K(J for n > 1. (Actually the estimator is applicable
for n = 1 as well.) It may be of interest and would serve as a check to
verify directly that
is unbiased.
where the substitution u = s - K was made. //1/
In closing this section on unbiased estimation, we make several remarks.
Remark F or some functions of the parameter there is no unbiased

estimator. For example, in a sample of size 1 from a binomial density
there is no unbiased estimator for 1/8. Suppose there were; let T = t(X)
denote it. Then 8(J[T] = ±
x=O
t(x) (n)x 9"(1 -
= 1/8, which says that
8)n-x
an nth-degree polynomial in 8 is identical to 1/8, which cannot be. ///1
Remark We mentioned in Subsec. 5.1 that the Cramer-Rao lower bound
is not necessarily the best lower bound. For example, the Cramer-Rao
lower bound for the variance of unbiased estimators of 8 in sampling
from the negative exponential distribution is given by 82 /n (see Example 29),
and the variance of the UMVUE of 8 is given by 82 /(n - 2) (see Ex-
ample 35). 82 /(n - 2) is necessarily the best lower bound. //11
Remark For some estimation problems there is an unbiased estimator
but no UMVUE. Consider the following example. 1/1/
EXAMPLE 36 Let Xb ... , Xn be a random sample from the uniform density

over the interval (8, 8 + 1]. We want to estimate 8. X n -! and
(Y1 + Yn)/2 -! are unbiased estimators of 8, yet there is no UMVUE
6 LOCATION OR SCALE JNVARlANCE 331
of (J. For fixed 0 <p < 1, consider the estimator g(XI - p) + p, where
the function g(y) is defined to be the greatest integer less than y. Now
8[g(Xl- p) + p]
=1 9+1
9
g(Xl - p) dXl +p = f 6+1-p
6-p
g(y) dy + p.
For fixed and p, there exists an integer, say N
(J = N(B, p), satisfying
() - p < N < (J + 1 - p. Hence,
S[g(Xl - p) + p]
N f6 + I - p
= f
6-p
8+ 1- p
g(y) dy +p= f6-p
(N - 1) d y +
N
N dy +p= (}.
So g(XI - p) + p is an unbiased estimator of B. Moreover, if (} + 1 - P

is an integer, say J, then g(XI - p) = J - 1 for all Xl satisfying
(} - p < Xl - P < B + 1 - p; so g(Xl - p) + p = J - 1 + p = B + 1 - P
- 1 + p = B for all B < Xl < B + 1; that is, g(XI - p) + p estimates B with
no error, and hence has zero variance for B + 1 - P equal to any integer.
So, we have an estimator, namely g(Xl - p) + p, which has zero variance
for (} = any integer - 1 + p. But 0 < P < 1 is arbitrary; so for any fixed
B, say Bo , we can find an unbiased estiumtor of B which has zero variance
at (}o' Hence, in order for an estimator to be the UMVUE of B, it must
have zero variance for all (}; that is, it must always estinlate B without
error. Qearly, no such estimator exists. {The reader may wish to show
that var [g(Xl - p) + p] = [N - (B - p))[(B + 1 - p) - N], where N =
N«(}, p) is an integer satisfying B - p < N B + 1 - p.} ////
Remark It· is sometimes possible to find an UMVUE even when a

minimal sufficient statistic is not complete. See Prob. 11, p. 313, in
Rao [17]. fIll
6 LOCATION OR SCALE INVARIANCE

In the last section we employed the property of unbiasedness as a means of
restricting the class of estimators with the hope of finding an estimator having
minimum mean-squared error within the restricted class. In this section we will
indicate how an alternative property, the property of invariance, can be used to
restrict the class of estimators. Our discussion will be limited to only two types
of invariance, namely, location invariance and scale invariance; a fuller discus-
sion, which is beyond the scope of this book, can be found in Refs. [12] and [19].
6.1 Location Invariance

If the observations Xl' •.. , Xn represented measurements of some sort and the
parameter being estimated was also measured in the same units, one might
reasonably require that an estimator t(·, ... , .) satisfy the property
t(XI + c, X2 + c, "" xn + c) = t(XI' "', xn) + c for every constant c. The
idea is that if a constant c is added to each of the measurements Xl' "', Xn'
then the estimator evaluated at the adjusted measurements Xl + c, ... , Xn + c
ought to adjust the estimated values t(XI' •.. , xn) by adding the same constant
to it. For example, suppose that it is desired to estimate the average weight
of a group of pigs when the only method available for weighing is for a person
to stand on a scale holding a pig; so both the pig and person are weighed. If
one person were to hold the pigs, the measurements (weights) Xl + c, ... ,
Xn + c would be obtained, where Xi is the weight of the ith pig and c is the person's
weight. If, on the other hand, someone else were to hold the pigs, the measure~
ments Xl + c', ... , xn + c' would be obtained, where c' is the other person's
weight. It seems reasonable that the estimate of the average weight of the group
of pigs obtained should not depend on which person held the pigs; that is, the
estimate should not vary with c, the weight of the pig holder. We define a
location-invariant estimator accordingly.
Definition 23 Location invariant An estimator T = t( Xl' ... , Xn) is

defined to be location-invariant if and only if t(XI + c, ... , xn + c) =
t(Xb "', xn) + c for all values Xb •.. , xn and all c. IIII
A number of the estimators that we have encountered are location-
invariant, for example, X nand (Yi + Yn)/2, as the following shows:
L (Xi + c) LXi
t(XI + c, " ., Xn + c) =
n
= --
n
+ c = t(XI' ... , xn) + c
for t(XI' ... , xn) = xn; and

t(XI + c, ... , Xn + c)
min [Xl + c, ... , Xn + c] + max [Xl + c, ... , Xn + c]
-
2
min [Xl' ... , Xn] + c + max [X I, ... , Xn] +C
-
2
6 LOCATION OR SCALE INVARlANC£ 333
for t(X., •.• , XII) = (YI + y,J/2. On the other hand, quite a number of
2
estimators are not location-invariant; for example, 8 and YII - Y1, as the fol-
lowing shows: Take T = t(XI , ... , XII) = 8 = L (Xi - XJ /(n - 1); then
2 2
t(XI + C, ••• , XII + c) = L [Xi + c - L (Xi + c)ln]2/(n - 1) = t(X., ... , XII)' in-
stead of t(XI' •.• , xJ + c. Now take T = t(XI' ... , X,J = Y,. - Yi; then
t(XI + c, •.. , X,. + c) = max [Xl + c, ... , X,. + c] - min [Xl + c, ... , Xn + c] =
max [Xl' ... , xn] + c - min [Xl' ... , X,.] - C = t(X., ... , x,J, instead of
t(XI' .•. , XII) + c.
Our use of location invariance will be similar to our use of unbiasedness.
We will restrict ourselves to looking at loc~tion-invariant estimators and seek
an estimator within the class of location-invariant estimators that has uniformly
sma1lest mean-squared error. The property of location invariance is intuitively
appealing and turns out also to be practically appealing if the parameter we are
estimating represents location.
Definition 24 Location parameter Let {f(· ; e), e E 9} be a family

of densities indexed by a parameter 8, where 9 is the real line. The param-
eter e is defined to be a location parameter if and only if the density
f(x; 8) can be written as a function of X - e; that is, f(x; 8) = h(x - 8)
for some function h( . ). Equivalently, e is a location parameter for the
density fx(X; e) of a random variable X if and only if the distribution of
X - e does not depend on e. 111/
We note that if e is a location parameter for the family of densities
{f( . ; 8), e E 8}, then the function h( . ) of the definition is a density function
given by h( . ) = f(· ; 0).
EXAMPLE 37 We will give examples of several different location parameters.

Iff(x; e) = <P(J. I (X), then e is a location parameter since
tf>•• ,(x) = J2tt

1 [12(x - 8)2] = <Po. ,(x - 8).
exp -
Or if X is distributed normally with mean e and variance I, then X - 8

has a standard normal distribution; hence the distribution of X - 8 is
independent of e.
AIf f(x; e) = 1(6 - i. 6 + t lx), then e is a location parameter since
f(x; (J) = 1(6 -to 9 +t)(X) = 1( -t. t)(X - e), a function of X - e.
If f(x; a) = 1/11:[1 + (x - a)2], then a is a location parameter since
f(x; a) is a function of X-a. IIII
334 PARAMEI'RlC POINT ESTIMATION vn
We will now state, without proof, a theorem that gives within the class
of location-invariant estimators the uniformly smallest mean-squared error
estimator of a location parameter. The theorem is from Pitman [41].
Theorem 11 Let X., ... , Xn denote a random sample from the density
e
I( . ; e), where e is a location parameter and is the real line. The
estimator
t(X 1 ,· .. ,X )=
f etl !(X
n
t; e) de
(17)
n
f}]l!(X i ; e) de
is the estimator of e which has uniformly smallest mean-squared error
within the class of location-invariant estimators. ////
Definition 25 Pitman estimator for location The estimator given in

Eq. (17) is defined to be the Pitman estimator for location. IIII
According to the formula given in Eq. (17), determining the Pitman

estimator requires evaluating the integrals given in the numerator and denomi-
nator; such evaluation may not be easy. Note that the integration is with
respect to the parameter; SO the resulting ratio will be a function of Xl, ... , X n.
EXAMPLE 38 Let Xl' ... , Xn be a random sample from a normal distribution

with mean e and variance unity. We saw in Example 37 that e is a loca-
tion parameter. Our object is to find the Pitman estimator of e, which is
given by Eq. (17). In the following series of equalities one should be
forewarned that cancellations and insertions are being made simultaneously
in the numerator and denominator.
Je(I/J~)n exp [-t I (Xi - e)2] de
f(I/J2n)n exp [-t L (Xi - 8)2] de
-----------------------
6 LOCATION OR SCALE INV ARlANCE 335
I8 exp .[ -(nI2)(8 - XIIi] d8
I exp [- (nI2)(8 - X ,,)2] d8
I 8[1/-J'~(1I-J';;)]exp{ -t[(8 - X n)/{lj-J';;)]2}dO

-
I [l/J2n(lJJn)] exp{ -t[(8 Xn)/(I/-J';;)f} d8
=X n
by noting that the last denominator is just the integral of a normal density
with mean Xn and variance lin and hence is unity, and the last numerator
is the mean of this same normal density and hence is X".
We note that, for this example, the Pitman estimator of 8, which is
uniformly minimum mean-squared error among location-invariant esti-
mators, is identical to the UMVUE of 8; that is, the estimator that
is best among location-invariant estimators is also best among unbiased
estimators. IIII
EXAMPLE 39 Let Xb ... , X" be a random sample from a uniform distribu-

tion over the interval (8 - t, 8 + t). According to Example 37, 8 is a
location parameter. The Pitman estimator of 8 is
f 8 n" I(O-t. 8+t)(X

i=1
i) d8
"
I8.n l(x,-t. Xi + t)(8) d8
p=1
- ----------
J Dl(8-t,0+tlXi)d8
i=1
JfI
i=1
I(X,-t,Xt+tl 8 )d8
J~!~t8 d8 _1 . (Y1 + t)2 - (Y" - t)2 Y1 + Y"

J~!~t d8 - 2 (Y1 + t) - (Y" - t) = 2
Recall that for this example there is no UMVUE of 8. (See Example 36.)
Jill
Remark A Pitman estimator for location is a function of sufficient

statistics.
PROOF If 8 1 = 0l(Xh ••. , Xn), ••. , 8.,. = 0k(X1 , •.• , Xn) is a set of
n
sufficient statistics, then by the factorization criterion I1 !(Xl; 8) =
i= 1
g(Sb ••• , SJi;; 8)h(Xb •.• , x,,); so the Pitman estimator can be written as
f 8J} I(Xi;8) d8 f 8g(Sb ... , Sk; fJ)h(Xl' ... , Xn) d8

n
I in I(X 8) d8 fg(Sl' . '" Sk; 8)h(Xl' ... , Xn) d8

i;
f 8 g(S S k; 8) d8
l' ... ,
Ig(Sl' ... , Sk; 8) d8

which is a function of Sl' ... , Sk' IIII
6.2 Scale Invariance

For those experiments in which measurements can be made in different units,
such as length being measured in either inches or centimeters, weight being
measured in either pounds or kilograms, or volume being measured in either
quarts or liters, one might reasonably require that his statistical procedure be
independent of the measurement units employed. If the statistical procedure
is that of point estimation, then one might require that the estimator that is to
be used satisfy the property of scale invarianee defined below. The idea is that
an estimator will be scale-invariant if the estimator does not depend on the sca1e
of the measurement.
Definition 26 Scale-invariant An estimator T = t(Xl' ... , XJ is

defined to be scale-invariant if and only if t(exl' " . , eXn) = et(xl' .. , , xn)
for all values Xl' ... , xn and all e > O. IIII
A number of the estimators that we have considered are scale-invariant,

including X n , JS2, (Yl + Yn)/2, and Yn - Yl . Our discussion of scale-
invariant estimators will be limited to problems concerning estimation of scale
parameters defined below.
Definition 27 Scale parameter Let {f( . ; 8), 8 > O} be a family of

densities indexed by a real parameter 8. The parameter 8 is defined to be a
scale parameter if and only if the density I(x; 8) can be written as (1/8)h(xI8)
for some density h(·). Equivalently, 8 is a scale parameter for the
density Ix(x; 8) of a random variable X if and only if the distribution of
XI8 is independent of 8. / / //
Note that if 8 is a scale parameter for the family of densities {/(' ; 8),8 > O},
then the density h( . ) of the definition is given by hex) = I(x; 1).
LOCATION OR SCALE INVARIANCE 337
6
EXAMPLE 40 We give several examples of scale parameters. If f(x; 1) =

(l/l)e-x/ Al 0, t;1)ix ), then 1 is a scale parameter since e~Y1(0, oo)(Y) is a
density. ~ote that this parameterization of the negative exponential
distribution is not the parameterization that we have used previously.
If
f(x; 6) = "'0, .,(x) = J;1tlT exp [ - H~) 1

then (Jis a scale parameter since (I/J2ic) exp (-ty2) is a density.
If f(x; 0) = (1/ 8)1(0. 9) (x) = (l/8)I(o.1)(xl8), then 0 is a scale param-
eter since I(O.l)(Y) is a density.
If f(x; 8) = (lIO)I(9. 29)(X) = (1/8)1(1, 2lx /8), then 8 is a scale param-
eter since 1(1. 2)(Y) is a density. II/I
Our sole result for scale invariance, a result that is comparable to the
result of Theorem lIon location invariance, requires a slightly different frame··
work. Instead of measuring error with squared-error loss function we measure
it with the loss function t(t; 8) = (t - 6)216 2 = (tiO - 1)2. If It - OJ represents
error, then 1001 t - 01/6 can be thought of as percent error, and then (t - 8)2/62
is proportional to percent error squared. We state the following theorem, also
from Pitman [41], without proof.
Theorem 12 Let X b · · . , Xn bearandomsamplefromthedensityf(' ;8),

where 0> 0 is a scale parameter. Assume that f(x; 0) = 0 for x < 0;
that is, the random variables Xi assume only positive values. Within
the class of scale-invariant estimators, the estimator
has uniformly smallest risk for the loss function t(t; 0) = (t - 8)2/02. /1//
...
Definition 28 PItman estimator for scale The estimator given in Eq.
(18) is defined to be the Pitman estimator for scale. /111
Remark The Pitman estimator for scale is a function of sufficient

statistics. IIII
338 PAR.AMEI'RlC POINT ESTIMATION VII
EXAMPLE 41 Let Xl' ... , Xn be a random sample from a density I(x; lJ) =
(1/0)/(0,6)(x). The Pitman estimator for the scale parameter B is
frO (110 2) fI (1/0)1(0, 8)(X ,) dO

o 1=1 _
f o-n-
00 2
_ Y..:. :;.."_ __
dO
n
lU1(1/0)/(0. 6)(X,) dO
00 00
3
fo (1/0 ) f " o-n- 3 dO
y
{1/[(n + 2) _1]}yn-(n+2)+1 n +2
= {1/[(n + 3) - 1]}yn-<n+3)+ 1 = n + 1 Yn'
We know that Yn is a complete sufficient statistic and 8[Yn ] =[nl(n + 1)]0;
so by the Lehmann-Scheffe theorem [en + 1)/n]Yn is the UMVUE of o.
IIII
EXAMPLE 42 Let X., ... , Xn be a random sample from the density f(x; A) =
(IIA) exp (-xIA)/(o. oolx). The Pitman estimator for the scale param-
eter A is
00
2 n
fo (1/A )IIJ !(X , ; A) dA t 00
(1IAn+2)exp( - I: XdA)dA
r (I/A
1
3
) D,f(X,; A)dA
-
('(I/An + 3)eXp ( - L XJA)dA
t (<x/I:
00
x,)n+2 e-O:(I: X,/<x 2) d<x

-
fo
00
(<x/I: Xi) n+3 e -«(I: X,J<x2) d<x
= I: x. r(n + 1) = I: X, .
'r(n + 2) n +1
(It can be shown that the UMVUE of A is Xi/n.) I:
Note that I:
X,Jn is a scale-invariant estimator, and, hence, since
I:Xil(n + 1) is the sca1e-invariant estimator having uniformly smallest
risk for the loss function (t - Of 102 , the risk of Xi/en + 1) is uniformly I:
I: 2
smaller than the risk of Xi/n. Also, since here risk equals 1/0 times the
MSE, the MSE of I:
Xil(n + 1) is uniformly smaller than the MSE of
I: Xil n . 1/1/
BAYES ESTIMATORS 339
7
7 BAYES ~STIMATORS
In our considerations of point-estimation problems in the previous sections .of
this chapter, we have assumed that our random sample came from some densIty
f( . ; 8), where the function/( . ; .) was assumed known. Moreover, we have
assumed that 8 was some fixed, though unknown to us, point. In some real-
world situations which the density f( . ; 8) represents, there is otten additional
infonnation about 6 (the only assumption which we heretofore have made about
8 is that it can take on values in 8). For example, the experimenter may have
evidence that 6 itself acts as a random variable for which he may be able to
postulate a realistic density function. For instance, suppose that a machine
which stamps out parts for automobiles is to be examined to see what fraction
o of defectives is being made. On a certain day, 10 pieces of the machine's
output are examined, with the observations denoted by Xl' X 2 , ••• , X 10 , where
Xi = 1 if the ith piece is defective and Xl = 0 if it is nondefective. These can
be viewed as a random sample of size 10 from the Bernoulli density
for 0 <0< 1,
which indicates that the probability that a given part is defective is equal to the
unknown number O. The joint density of the 10 random variables Xl' X 2 , ••• ,
XlO is
for 0 :::; 0 < 1.
The maximum-likelihood estimator of 0, as explained in previous sections, is

e = X. The method of moments gives the same estimator. Suppose, how-
ever, that the experimenter has some additional information about 8; suppose
that he has observed that on various days the value of 0 changes and it appears
that the change can be represented as a random variable with the density
ge(O) = 60(1 - 0)/[0,1](0).

An important question is: How can this additional information about 0 be
used to estimate 00 , where 60 is the value that ® was equal to on the day the
sample was drawn?
To examine this problem, we will assume, in addition to the assumption
that our random sample came from a density 1(' ; 8), that the unknown param-
eter 8 is the value of some random variable, say ®. We wUI still be interested
in estimating some function of 0, say 1'(0). If ® is a random variable, it has
a distribution. We let G( .) = Ge (') denote the cumulative distribution function
of ® and g(.) = ge(') denote the density function of ® and we assume these
. . '
functIons contaIn no unknown parameters. In order to emphasize that the
distribution of E> is over the parameter space, we have departed from our custom
of using F( .) and f( .) to represent a cumulative distribution function and
density function, respectively, and have used G( . ) and g( . ) instead.
If we assume that the distribution of E> is known, we have additional
information. So an important question is: How can this additional information
be used in estimation? It is this question that we will address ourselves to in
the foll<?wing two subsections. In many problems it may be unrealistic to
assume that B is the value of a random variable; in other problems, even though
it seems reasonable to assume that B is the value of a random variable E> the
distribution of E> may not be known, or even ifit is known, it may contain other
unknown parameters. However, in some problems the assumption that the
distribution of E> is known is realistic, and we shall examine this situation.
7.1 Posterior Distrib ution

Heretofore we have used the notationf(x; B) to indicate the density of a random
variable X for each B in 8. Whenever we want to indicate that the parameter
B is the value of a random variable E>, we shall write the density of X as f(x 1 B)
instead of f(x; 8). We should note that f(x 1B) is a conditional density; it is
the density of X given E> = B. A more complete notation for f(x 18) would be
fXIS=O(xl B).
Let X h . . . , Xn be a random sample of size n from the density f( ·18),
where B is the value of a random variable E>. Assume that the density of E>,
gs(·), is known and contains no unknown parameters, and suppose that we
want to estimate L(B). How do we incorporate the additional information of
known gs( . ) into our estimation procedures? In the past, we thought of the
likelihood function as a single expression that contained all our information; the
likelihood function included the observed sample Xh .•. , Xn as well as the form
of the density f(x; B) we sampled from in its expression. Now we need an
expression that contains all the information that the likelihood function con-
tains plus the added information of the known density gs(·)' gs(·) is called the
prior distribution of E>. It summarizes what we kn.ow about B prior to taking
a random sample. What we seek is an expression that summarizes what we
know about B after we take a random sample. We seek the posterior distribu-
tion of E> given Xl = X h ... , Xn = x n·
Definition 29 Prior and posterior distributions The density gs(') is

called the prior distribution of E>. The conditional density of E> given
Xl = Xl' .•• , Xn = x n , denoted by fSlxl =Xl .....Xn =xn(B 1 Xl' ••• , xJ, is
called the posterior distribution of E>. IIII
7 BAYES ESTIMATORS 341
Remark
[,D, I(x" 6)]U,,(6) (19)
= f La I(X,16)]g,,(8) d8
for random sampling. [Recall that fYlx=iylx) =fx,y(x, y)lfx<.x) =
fXIY=:V<x I y)fy(y)Lfx(x).] 1///
The posterior distribution replaces the likelihood function as an expression
that incorporates all information. If we want to estimate 6 and parallel the
development of the maximum-likelihood estimator of 6, we could take as our
estinlator of 6 that 6 which maximizes the posterior distribution, that is, estimate
6 with the mode of the posterior distribution. However, unlike the likelihood
function (as a function of 6), the posterior distribution is a distribution function;
so we could just as well estimate 6 with the median or mean of the posterior
distribution. We will use the mean of the posterior distribution as our estimate
of 6, and in general we could estimate 1'( 6) as the mean of 1'( 9) given Xl = Xl, ••• ,
X" = x,,; that is, take 8[1'(9) I Xl = Xh ... , X" = x,,] as our 'estimate of 1'(6).
Definition 30 Posterior Bayes estimator Let Xh .•. , X" be a random

sample from a density f(x 16), where 6 is a value of the random variable
9 with known density ge(·). The posterior Bayes estimator of 1'(6) with
respect to the prior ge(') is defined to be
8[ 1'(9) I X h ... , X,,]. (20)
1///
Remark
8[1'(9)1 Xl = Xl' ••• ' X" = X,,] = J1'(6)!elxl=Xl •.... Xn:;x.J6Ixh ... ' x,,) d6
(21)
f r(8>ta I (x, 8)] g,,(8) d8
1
-
f La I(X.l8)] g.,(8) d8 .
/111
One might note the Similarity between the posterior Bayes estimator of
1'(6) = 6 and the Pitman estimator of a location parameter [see Eq. (17)].
EXAMPLE 43 Let Xl' ... , Xn denote a random sample from the Bernoulli
density f(x 18) = £F(1 - e)l-x for x = 0, 1. Assume that the prior distri-
bution of 9 is given by gs(8) = /(0,1) (8); that is, 9 is uniformly distribu-
ted over the interval (0, 1). Consider estimating 8 and T(e) = 8( 1 - 8).
Now
so the posterior Bayes estimator of 8 with respect to the uniform prior

distribution is given by
8[9 IXl = Xh ..• , Xn = xJ

= f 8!SIXl=Xlo ...,Xn=xn(Olxh·'·' Xn) d8
_ J~88rX'(1 - 8)n-rXi d8 _ B(L Xi + 2, n - LXi + 1)
- U 8r XI( 1 - 8t - r Xi d8 - B(L Xi + 1, n - LXi + 1)
_ r(L Xi + 2)r(n - LXi + 1) . r(n + 2)
- r(n + 3) r(L Xi + l)r(n - L Xi + 1)
LXi + 1
- n +2 .
Hence the posterior Bayes estimator of 8 with respect to the uniform prior
distribution is given by (L Xi + 1)/(n + 2). Contrast this to the
maximum-likelihood estimator of 8, which is Xt!n. L L
Xt!n is unbiased
and an UMVUE, whereas the posterior Bayes estimator is not unbiased.
To obtain the posterior Bayes estimator of, say T(8) = 8(1 - 8), we
calculate
= f 8(1- 8)!slx =xI, ...,xn=xn(8I x

1 l,··" xn) d8
U8(1 - 8)8rXi (1 - 8t- rxi d8

-
J~8r XI(1 - 8t- r Xi d8
r(L Xi + 2)r(n - L Xi + 2) r(n + 2)

= r(n+4) 'r(L x i+ 1)r(n-L x i+ 1)
(L Xi + l)(n - LXi + 1)
-
(n + 3)(n + 2)
So the posterior Bayes estimator of 8(1 - 8) with respect to a uniform
prior distribution is (L Xi + 1)(n - L Xi + l)/(n + 3)(n + 2). IIII
We noted in the above example that the posterior Bayes estimator that
we obtained was not unbiased. The following remark states that in general a
posterior Bayes estimator is not unbiased.
Remark Let T~ = t~(XI' .,., XJ denote the posterior Bayes estimator

of 't(6) with respect to a prior distribution G( ,). If both T~ and 1'(9)
have finite variance, then either var [T~ 18] = 0, or T~ is not an unbiased
estimator of 't(6). That is, either T~ estimates 1'(8) correctly with proba-
bility 1, or T~ is not an unbiased estimator of 1'(8),
PROOF Let us suppose that T~ is an unbiased estimator of 1'(8);
that is, 8[T~ 18] = 1'(8), By definition we have T~ = t~( X., •.. , Xn) =
8[1'(9)1 X., •.. , Xn]. Now
var [T~] = 8[var [T~I 9]] + var [8[T~1 9]]
= 8[var [T~I 9]] + var [1'(9)],
and
var ['t(9)] = 8[var [1'(9)1 X., ••. , Xn]] + var [8[1'(9)1 Xl"'" Xn]]
= 8[var [1'(9) 1X., •.. , Xn]] + var [T~];
hence, 8[var [T~I 9]] + 8[var [1'(9)1 Xl"'" Xn]] = O. Since both
8[var [T~ 19]] and 8[var [1'(9) 1X., •. " Xn]] are nonnegative and their
sum is 0, both are O. In particular, 8[var [n
19]] = 0, and since
var [T~I 9] is non-negative and has zero expectation, var [T~18] = O. Jill
7.2 Loss-function Approach

In Subsec. 3.4 we introduced the concepts of loss and risk. These two concepts
were used to assess goodness of estimators. In this section we discuss how the
additional information of knowledge of a prior distribution of 9 can be used
in conjunction with loss and risk to define or select an optimum estimator.
We commence with a review of the problem we hope to solve. Let
X., ... , Xn be a random sample from a density f(xl8), 8 belonging to e, where
the function f( • IfJ) is assumed known except for 8. We assume that the
unknown 8 is the value of some random variable 9 and that the distribution of
9 is known and contains no unknown parameters. On the basis of the random
sample X., •.. , Xn we hope to estimate 1'(8), some function of 8. In addition,
we assume that a loss function I(t; 8) has been specified, where I(t; 8) represents
the loss incurred if we estimate 1'(8) to be t when 8 is the parameter of the
density from which we sampled, For any estimator T = t( X., ... , X J, we
344 pA.RAMBTlUC POINT ESTIMATION vn
noted in Subsec. 3.4 that Co[t(T; B)] represented the average loss of that esti-
mator, and we defined this average loss to be the risk, denoted by {!ttCB), of the
estimator t(·, ... , '). We further noted that two estimators, say 11 =
tl(XI , ... , Xn) and Tz = t 2 (XH ••• , X n), could be compared by looking at their
respective risks {!t'l(B) and {!t,lB) , preference being given to that estimator with
smaller risk. In general, the risk functions as functions of B of two estimators
may cross, one risk function being smaller for some B and the other smaller
for other B. Then, since B is unknown, it is difficult to make a choice between
the two estimators. The difficulty is caused by the dependence of the risk
function on B. Now, since we have assumed that Bis the value of some random
variable 9, the distribution of which is also assumed known, we have a natural
way of removing the dependence of the risk function on B, namely, by averaging
out the B, using the density of 9 as our weight function.
Definition 31 Bayes risk Let Xl, ... , Xn be a random sample from a

density f(x IB), where B is the val ue of a random variable 9 with cumula-
tive distribution function G(') = Ge ( .) and corresponding density
g( . ) = ge( '). In estimating ,(0), let t(t; B) be the loss function. The
risk of estimator T = t(Xt , ••• , XII) is denoted by {!tiB). The Bayes risk
of estimator T = t(XH ... , Xn) with respect to the loss function t( .; .)
and prior cumulative distribution G(')' denoted by ~(t) = ~(., G( I), is
defined to be
-l( I) f
= -l/, G( I) = ~ {!tiB)g(B) dB. (22)
fIll
The Bayes risk of an estimator is an average risk, the averaging being over
the parameter space ~ with respect to the prior density g(.). For given
loss function t( .; .) and prior density g( . ) the Bayes risk of an estimator is a
real number; so now two competing estimators can be readily compared by
comparing their respective Bayes risks, still preferring that estimator with smaller
Bayes risk. In fact, we can now define the best" estimator of ,(B) to be that
U
estimator with smallest Bayes risk.
Definition 32 Bayes estimator The Bayes estimator of ,(B), denoted by

T,*, G = ti, G(X1 , .•• , XJ, with respect to the loss function t( .; .) and
prior cumulative distribution G('), is defined to be that estimator with
smallest Bayes risk. Or the Bayes estimator of ,(B) is that estimator
t1. G satisfying
-l/, G( t*) = ~/, G( ti, G) .s;; -l/, G( I)
for every other estimator T = t(XI' ... , X,.) of ,(B). 11II

The posterior Bayes estimator of T(8), defined in Definition 30, was defined
without regard to a loss function, whereas the definition given above requires
specification of a loss function.
The definition leaves the problem of actually finding the Bayes estimator,
which may not be easy for an arbitrary loss function, unsolved, However, for
squared-error loss, finding the Bayes estimator is relatively easy, We seek that
estimator, say 1*( • , "', .), which minimizes the expression J~ BitCe)g(e) de =
Iii C8 [[/(Xb ... , x,.) - 1:(e)]2]g(e) de as a function over possible estimators
I( • , •. " .). Now,
J~ Co[[/(X 1, , •• , Xn) - T(e)]2]g(e) de
= r f r [/(XH ... , Xn) -

JiilJ~
T(6)]2IxI, ...• xJX., •.. , xnl 8) Ii dXi}g(e) de
£=1
= f~(it ~ [«0) - I(x" ... , x,,) l' Ix .. .... x.(x" ... , x.1 O)g(O) dO}
Ix 10 •••• x..(X h •.. , Xn)
n
• IXI ••••• X ..(Xl' ••• , Xn) n dx,
1= 1
·IXl, •••• X ..(x., ... ,Xn) n

i= 1
dXi'
and since the integrand is nonnegative, the double integral can be minimized
if the expression within the braces is minimized for each x., ... , X n • But the
expression within the braces is the conditional expectation of [T(e) - I(XH
... , xn)]2 with respect to the posterior distribution of e given Xl = Xl' ... t
Xn = X n , which is minimized as a function of I(x., ... , xn) for I*(x., ... , xn)
equal to the conditional expectation of T(e) with respect to the posterior distribu-
tion of e given Xl = Xl' ..• , Xn = x n • {Recall that C[(Z - a)2] is minimized
as a function of a for a*"= C[Z].} Hence the Bayes estimator of T(e) with re-
spect to the squared-error loss function is given by
(23)
which is identical to the estimator given in Eq. (21).

346 pARAMETRIC POINT ESTIMATION vu
For a general loss function, we seek that estimator which minimizes

J~ 9flO) g(O) dB. Again,
f- 9fJ.B)g(B) dB
~
=fft [J/(t(x" ... , x,); IJ)fXh ...• xJx" ... , x, I IJ) V, dx,]O(IJ) dIJ
= J. [J/( I(x" ... , x,); IJ)felx, =x" .... x.=x.(IJI x" ... , x,) dIJ]
n
• fXl •.•.• Xn(Xl' .•• , xn) n dx
i= 1
i,
and minimizing the double integral is equivalent to minimizing the expression

within the brackets, which is sometimes called the Rosterior risk. So, in general,
the Bayes estimator of r(B) with respect to the loss function t( .; .) and prior
density g( • ) is that estimator which minimizes the posterior risk, which is the
expected loss with respect to the posterior distribution of 9 given the observa-
tions Xl, ... , x n • We have the following theorem and corollaries.
Theorem 13 Let Xl, .•. , Xn be a random sample from the density

f(x IB), and let g(B) be the density of 9. Further let t(t; B) be the loss
function for estimating r(B). The Bayes estimator of r(O) is that estimator
1*( . , ... , .) which minimizes
as a function of Ie , ... , .). IIII

Corollary Under the' assumptions of Theorem 13, the Bayes estimator
of r(B) is given by
(25)
for a squared-error loss function. IIII

Corollary Under the assumptions of Theorem 13, the Bayes estimator
of B is given by the median of the posterior distribution of 9 for a loss
function equal to the absolute deviation. IIII
The proofs of the theorem and first corollary preceded the statement of the
theorem. The second corollary follows from the observation that
f._I (J - t(Xh"" x,,)lfs1xl=xl ..... XI'l=xl'l(BI Xl , ••• , X,,) dB

~
is minimized as a function of I( . , ... , .) for 1*( . , ... , .) equal to the median

of the posterior distribution of 8. {Recall that &[ IZ - a I] is minimized as a
function of a for a* = nledian of Z.}
EXAMPLE 44 Let Xl, ... , X" be a random sample from the normal density
with mean B and variance 1. Consider estimating B with a squared-error
loss function. Assume that 8 has a normal density with mean 110 and
variance 1. Write 110 = Xo when convenient. According to Eq. (25) the
Bayes estimator is given as the mean of the posterior distribution of 8.
fSlx 1 =X1. •••• Xn=xl'l(B IXb .•• , X,,)
= Ix" ...• x.llx!> ... , x.IO)g(O) _ tV/(X;! 0)] g(O)

Ix, ..... X.(x .. ... , x.) f [.V/(X,IO)]g(O) dO
(11 J21t)" exp [ - t t (x, - 0)2] (1/J21t) exp [ - t(0 - /lO)2]
r., (1/.fiic)· exp [ - t t

--=--------~~~------~---------------------
(X, - 6)' ] (1/.fiic) exp [ - t(O - /lo)2] dO
_ exp [ -t,t
o
(X,- 0)2]
f., [-t,t.
exp (x, - 0)2] dO
_ exp {-t[(n + 1)02- 21!,to x, + It/f)}

f., exp {-t[(n + 1)62 - 20,t/, + txl]} dO
_ exp ( -[en + 1)/2l{02 ~ 20 t xrl(n + 1) + [t xr/(n + !)n)
roo exp (-[(n + 1)/21{02 - 26 t xrt(n + I) + [t xl/(n + on) dO
_ [lIJ2n/(n + 1)] exp {-[(n + 1)/2l[0 - t xr/(n + l)rJ
roo [IIJ2n!(n + 1)] exp (-[(n + 1)/2][0 - L xt!(n + 1)] 2} dO
= 1 e
J2n/(n +1) xp
(_n+l["
2
]21
~xd(n + 1) J; (J -
the denominator is unity since it is the integral of a density. We have

shown that the posterior distribution of e is normal with mean
L~ xi/(n + 1) and variance 1/(n + 1); hence the Bayes estimator of 8 with
respect to squared-error loss is
ft n
Xo + L Xi Jlo + L Xi
1 1
--.....-.- - --- -
n+1 n+1
Since the posterior,distribution of e is normal, its mean and median are

the same; hence
n
Jlo+LXi
1
n+1
is also the Bayes estimator with respect to a loss function equal to the
absolute deviation. II/I
EXAMPLE 45 Let Xl' ... , X" be a random sample from the density I(x 18) =
(l/fJ)/(0.8)(x), Estimate 8 with the loss function let; fJ) = (t - 8)2/82.
Assume that e has a density given by g(8) = 1(0.1)(8). Let y" denote
max [Xl, ... ,xft ]. Find the posterior distribution of 9.
n I(0.8)(x;)I(0. 1)(8)
ft
(l/fJ)ft
f8I X l=xlo •••• Xn=Xn
t
( ll
V Xl"'"
)
X" = -l~--"-----
i= I
fo (1/8)"Xl I(0.8/X i) d8
(1/ 8 )"I(Yn.l)(8)
-
fYn(1/8)"I(Yn.n(8) d8
I
(1 18)nI (Yn. 1)(fJ)

= [l/(n - l)](l/y:-l - 1)'
We seek that estimator which minimizes Eq. (24), or we seek that esti-
mator 1(') which minimizes
J{[/(Yn) - 8fI82}(1/~)/(Yn.1)(8) d8
[l/(n - l)](lly: 1 - 1)
or that estimator which minimizes

1 1
I1n
[t(yJ - 0]2 (1'+ 2 dO
= [t(y,.)]2 f"'n on+21 dO - 2t(Yn) I1" '1ndO"+' nI1 on1 dO.

1
(1'+ 1 (26)
Equation (26) is a quadratic equation in t(·); this quadratic equation

assumes its minimum for
/III
We note that the Bayes estimators derived in Examples 43 to 45 are

functions of sufficient statistics. It can be shown that this is generally true;
that is, a Bayes estimator is always a function of minimal sufficient statistics. In
fact, under quite general conditions it can be shown that the Bayes estimator
corresponding to an arbitrary prior probability density function, which is
positive for all 0 belonging to e, is consistent and BAN. So, even if you do not
know the correct prior distribution, a Bayes estimator has some desirable
optimum properties. And if you do know the correct prior distribution and
accept the criterion that a best estimator is one that minimizes average loss, then
the Bayes estimator corresponding to the known prior distribution is optimum.
Even in those problems when the prior distribution is unknown, the
concept of Bayes estimation can benefit us. It provides us with a technique of
determining many estimators that we might not have otherwise considered.
Each possible prior distribution has a corresponding estimator, whose merits
can be judged by using our standard methods of comparison. Thus, we have
yet another method of finding estimators to append to the methods given in
Sec. 2.
Bayes estimation can also sometimes be useful as a tool in obtaining an
estimator possessing some desirable property that does not depend on prior
distribution information. The property of minimax is such a property, and
in the next subsection we will see how Bayes estimation can sometimes be used
to find a minimax estimator. Another such property is given below. Our
objective has been to minimize risk, but since risk depended on the parameter,
we were unable to find one estimator that had smaller risk than all others for
all parameter values. Minimax circumvented such difficulty by replacing the
risk function by its maximum value and then seeking that estimator which
minimized such maximum value. Another way of getting around the difficulty
arising from attempting to uniformly minimize risk is to replace the risk function
VII
by the area under the risk function and to seek that estimator which has the
least area under its risk function. We note that if the parameter space e
is
an interval, the estimator having the least area under its risk function is the
Bayes estimator corresponding to a uniform prior distribution over the interval e.
This is true because for a uniform prior distribution the Bayes risk is propor-
tional to the area under the risk function, and hence minimizing the Bayes risk is
equivalent to minimizing area.
7.3 Minimax Estimator

We defined a minimax estimator at the end of Subsec. 3.4 as an estimator whose
maximum risk is less than or equal to the maximum risk of any other estimator.
Such an estimator might be considered "conservative" since it protects against
the worst that can happen; it seeks to minimize the maximum risk. The follow-
ing theorem is sometimes useful in finding a minimax estimator.
Theorem 14 If T* = 1*( Xl' ... , Xn) is a Bayes estimator having con-

stant risk, that is, Btt-(O) = constant, then T* is a minimax estimator.
PROOF Let g*( . ) be the prior density corresponding to the Bayes
estimator 1*( . , ... , .).
sup Btt*(O) = constant = Bt,*(O)
6e~
= f Btr<.O)g*(O) dO < J~r BttCO)g*(O) dO ~ supe~ BtiO)

fi 6
for any other estimator I( . , ... , '). IIII
EXAMPLE 46 Find the minimax estimator of 0 in sampling from the Bernoulli

distribution using a squared-error loss function. We seek a Bayes esti-
mator with constant risk. The family of beta distributions is a family of
possible prior distributions. We hope that for one of the beta prior
distributions the corresponding Bayes estimator will have constant risk.
A Bayes estimator is given by
J~OOE X'(l - Ot- EX'[l IB(a, b)]oa-l(l - O)b-l dO
J~OEX'(l - o)n-Ex'[l/B(a,b)]O a-l(l - O)b-l dO
JAO Ex,+Q(1 - O)"-EXI+b-l dO
= JAOExl+a-l(l - o)n-Ex,+b-l dO
_ BeL Xi + a + 1, n - 2: Xi + b) _ 2: Xi + a
- B(2: + a, n - 2: + b) - n + a + b .
Xi Xi
8 VECTOR OF PARAMETERS 351
So -the- Bayes estimator with respect to a beta prior distribution having

parameters -a and b is given by
We now ev8Juate the risk of (2: Xl + a)/(n + a + b) with the hope that we
will be able to select a and b so that the risk will be constant. Write
t': B(Xl, .•. , x.) = A 2: Xi + B = (2: Xl + a)/(n + a + b); then t!it*A,B(O)
+
= '4 [(A 2: Xi B - O?] = 8[[A(2: Xi - nO) + B - 0 + nAO]2] =
A 24[(2: Xi - nO)2] + (B - 0 + nAO)2 = nA 20(l - 0) + (B - 0 + nA0)2 =
02[(nA - 1)2 _ nA2] + O[nA2 + 2(nA - I)B] + B2, which is constant if
(nA - 1)2 - nA2 = 0 and nA2 + 2(nA - I)B = O. Now (nA - 1)2 - nA2
=0 if A=1/jn(jn+1), and nA2+2(nA-l)B=0 if B=
-nA 2/2(nA - 1), which is 1/2(Jn + 1) for A = l/j;t(J~ + 1). On
solving for a and b, we obtain a = b = j~/2; so (2: Xl + ~/2)/( n + In)
is a Bayes estimator with constant risk and, hence, minimax. // //
8 VECTOR OF PARAMETERS
In this section we present a brief introduction to the problem of simultaneous
point estimation of several functions of a vector parameter. We will assume
that a random sample Xl' ... , Xn of size n from the density I(x; 01, ... , Ok) is
available, where the parameter 0 = (01, ... , Ok) and parameter space e are
k-dimensional. We want to simultaneously estimate 't'l (0), ... , 't'r(O) , where
't'iO) , j = I, ... , r, is some function of 0 = (01, ... , Ok)' Often k = r, but this
need not be the case. An important special case is the estimation of 0 =
(81, ... , e k) itself; then r = k, and 't'l (0) = 01, ... , 't' k(O) = Ok' Another important
special case is the estimation of 't'(0); then r = 1. A point estimator of
('t'l(O), ... , 't',.(O» is a vector of statistics , say(Tl" .. , T,.), where Tj = tiXl" .. , Xn)
and Tj is an estimator of 't'j(0).
Our presentation of the method of moments and maximum-likelihood
method as techniques for finding estimators included the possibility that the
parameter be vector-valued. So we already have methods of determining esti-
mators. What we need are some criteria for assessing the goodness of an esti-
mator, say (Th ... , T,.), and for comparing two estimators, say (Tl' ... , T,.)
and (T{, ... , T;). As was the case in estimating a real-valued function 't'(e) ,
where we wanted the values of our estimator to be close to 't'(0), we now want
the values of the estimator (1;., ... , Tr) to be close to ('t'l (0), ... , 't'r(O». We
want the dis~ribution of (Tit ... , T,.) to be concentrated around ('t'1(0), ... , 't'r(O».
352 pAR.AM£TlUC POINT ESTIMATION VII
There are a number of ways of measuring the closeness of an estimator. For

instance, in comparing two estimators the definitions of " more concentrated"
and" closer," given in Subsec. 3.1, can be generalized to r dimensions. We will,
however, restrict ourselves to consideration of unbiased estimators and define
several ways of measuring the closeness of an unbiased estimator. No attempt
will be made in this book to generalize to r dimensions the notions of loss and/or
risk, invariance, Bayes estimation, and minimax. As far as optimum estima-
tion is concerned, we will be content to consider only unbiased estimators and
look for a best estimator within the restricted class of unbiased estimators.
Definition 33 Unbiased An estimator (T1 ) ••• , T,.), where T j =

t lX1 , ••• , X,.}, j = 1, ... , r, is defined to be an unbiased estimator of
(r1 (0), ... , 't,.(0) if and only if 4 e[1jl = 't/8) for j = 1, ... , r and for all
Oee. /1//
In Sec. 5, where we considered unbiased estimation of a real-valued func-

tion 'teO), we employed the variance of an estimator as a measure of its closeness
to't(O). Here we seek a generalization of the notion of variance to r dimensions.
Several such generalizations have been proposed; we will consider four of them,
called (i) vector of variances, (ii) linear combination (with nonnegative coeffi-
cients) of variances, (iii) ellipsoid of concentration, and (iv) Wilks' generalized
variance. The last two require some knowledge of matrices.
Possibly the simplest way of generalizing the concept of variance to
r dimensions is to use the vector of variances of the unbiased estimators 11, ... , T,..
That is, let the vector (vare [Td, .,., varo [T,,]) be a measure of the closeness of
the estimator (T1' ... , 1;.) to (-r1 (0), ... , t,,(O». The disadvantage of such a
definition is that our measure is vector-valued and consequently sometimes
difficult to work with. One way of circumventing this disadvantage is to use a
linear combination of variances, that is, measure the closeness of the estimator
(T1 , ••• , T,.) to ('t1(0), ... , 't,.(0» with IJ=1 a j varo [1)] for suitably chosen aj > O.
Both of these generalizations of variance embody only the variances of the
Tj , j = I, ... , r. The Tj are likely to be correlated; so one might justifiably
think that our measure of closeness of (T1' ... , 1;.) to ('t1 (8), ... , 't,.(0») should
incorporate the covariances of the T/s.
Notation If (T1' .,., 1;.) is an unbiased estimator of ('t1(0), ... , 't,,(0»,

let uiiO) = cove [Tf, 1)]. The matrix whose ijth element is Ujj(O) is called
the covariance matrix of the estimator (T1' ... , 1;.). Let uU(O) denote the
ijth element of the inverse of the covariance matrix. / // /
8 VECTOR OF pARAMETERS 353
FIGURE 8 ~----------------------~tl
Definition 34 Ellipsoid of concentration Let (T1' ... , T,.) be an un·

biased estimator of (r1 (0), ... , 'r,.(0)). Let uii(O) be the ijth element of the
inverse of the covariance matrix of (T1' ... , T,.), where the ijth element of
the covariance matrix is Ui/(/) = cOVo [T., Tj ]. The ellipsoid of concen-
tration of (T1' ••. , T,.) is defined as the interior and boundary of the ellipsoid
,. ,
L J=L uii(O)[ti -
i= 1 1
1:iCO)][tj - 'riO)] = r+ 2. (27)
See Fig. 8 for r = 2. 1111
Loosely speaking, the ellipsoid of concentration measures how concen-

trated the distribution of (T1' ... , T,.) is about (1:1(0), ... , 1:,(0». [In fact, if one
considers the vector random variable, say (U1 , .•• , U,), unifofI!1ly distributed
over the ellipsoid of concentration, it can be proved that (U1 , ••• , U,) and
('11, ... , T,.) have the same first- and second-order moments.] The distribution of
an estimator (T1' ... , T,.) whose ellipsoid of concentration is contained within
the ellipsoid of concentration of another estimator (T{, ... , T:) is more highly
concentrated about ('r1 (0), ... , 1:,(0) than is the distribution of (T{, ... , T;).
It is known that the determinant of the covariance matrix of an estima-
tor is proportional to the square of the volume of the corresponding ellipsoid of
concentration; hence another generalization of variance is as in Definition 35.
Definition 35 Wilks' generalized variance Let ('11, .,., T,.) be an

unbiased estimator of ('l'1 (0), ... , 1:,.(0). Wilks' generalized variance
of (T1' ... , T,.) is defined to be the determinant of the covariance matrix
of (T1""" T,.). 1111
354 pARAMETRIC POINT ESTIMA110N VII
Theorem 8, which showed how sufficiency could be used to improve on an

arbitrary unbiased estimator, generalizes to r dimensions. The generalization
is stated without proof.
Theorem 15 Let Xl' ... , Xn be a random sample from the density

I(x; Ob ... , Ok), and let Sl = d l (Xl , ... , Xn), ... , Sm = d m(Xl , .,., X,J be
a set of jointly sufficient statistics. Let (1";., ••. , T,.) be an unbiased esti-
mator of ('rl (0), ... , 'rlO». Define Tj = 8[Tj ISl' ... , Sm], j = 1, ... , r.
Then,
(i) (T{, ... , T;) is a statistic and an unbiased estimator of

('r l (O), ... , 'tr(O)), and T; = tj(Sb ... , S,J; that is, Tj is a function of the
sufficient statistics Sl' ... , Sm,j = 1, ... , r.
(ii) varo [Ti] < varo [Tj ] for every 0 E i:J, j = 1, ... , r.
(iii) The ellipsoid of concentration of (T{, ... , T;) is contained in
the ellipsoid of concentration of (1";., ..• , T,.), for every (J E e. ////
We might note that (ii) implies that
r r
L
j= 1
aj varo [Tn < L aj varo [Tj ]
)= 1
and (iii) implies that Wilks' generalized variance of (T{, ... , T;) is smaller than
Wilks' generalized variance of (Th ... , T,.).
Theorem 10 of Sec. 5 can also be generalized to r dimensions, but first the
concept of completeness has to be generalized.
Definition 36 Joint completeness For Xl' ... , X n, a random sample

from the density I(x; 01 , ••• , Ok), let (1";., ••• , Tm) be a set of statis-
tics. Tl , . .. , Tm are defined to be jointly complete if and only if
8 0 [,-,;::(1";., ••• , Tm)] = 0 for all 0 E i:J implies that Po[.-.;::(1";., .•• , Tm) = 0] == 1
for all 0 E i:J, where .-.;::(1";., •.. , TJ is a statistic. // / /
EXAMPLE 47 Let Xl' ... , Xn be a random sample from
where 81 < O2 , Write 0 = (81 , 82 ), Let Yl = min [Xh ... , X,.] and
Y,. = max [Xl' ... , Xn]. We want to show that Yi and Yn are jointly
complete. (We know that they are jointly sufficient.) Let ~(Yl' Y,,) be
an unbiased estimator of 0, that is,
<9' 8[~(Yl' Y,,)] == 0 for all 0 E ~.
Now
which is identically 0 If and only if

8 fYn~(yl' y")(y,, -
2
,,-2_
f81 81
Yl) dYl dy,. = 0
Differentiate both sides with respect to O2 , and obtain
now differentiate both sides of the resulting identity with respect to

Oh and obtain - ~(01' ( 2 )(0 2 - 01ya-2 == 0 for all 01 < O2 , and hence
~(01' ( 2 ) = 0 for 01 < O2 ; that is, ~(Yl' YII) = 0 for Yl < Y,., where Yl
and y,. are the possible values of Y1 and Y,.. We have shown that Y1 and
Y,. are jointly complete. / // /
If the density /(x; Oh ... , Ok) is a member of the k-parameter exponential

family, a'set of jointly complete and sufficient statistics can be found using the
following theorem. It is a k-dimensional analog of Theorem 9 and is stated
without proof. The following theorem is not precisely stated; certain regularity
conditions are omitted [16].
Theorem 16 Let Xl' ... , X,. be a random sample from/(x; 01 , ••• , Ok)'
k
If/(x; 01 , ... , Ok) =a(OI, ... , 0k)b(x)exp [I
j=1
c/0 1 , .," Ok) d.(x)j, that
J
. is,/(x; 01 , Ok) is a member of the k-parameter exponential family, then
.t.
•• "
Ct. d. (X,) • ... • d.( Xi) ) is a minimal set of jointly complete and sufficient
statistics. // //
3S6 pARAMETRIC POINT ES'I1MATION VII
EXAMPLE 48 Let Xl' ... , Xn be a random sample from
Now
n n
so L Xi and L X; are jointly complete and sufficient statistics by
i=l i=l
Theorem 16. IIII
We will state without proof the vector analog of Theorem 10. In the
same sense that an UMVUE was optimum, this following theorem gives an
optimum estimator for a vector of functions of the parameter.
Theorem 17 Let Xl' ... , Xn be a random sample from/ex; 01, ... , Ok)'
Write 0 = (0 1, ... , Ok)' If Sl = 01(X1, ... , X n), ... , Sm = 0m(X1, ... , Xn)
is a set of jointly complete sufficient statistics and if there exists an un-
biased estimator of (r 1 (0), ... , rr(O», then there exists a unique unbiased
estimator of (r1(0), ... , rr(O») , say Tt = tt(Sl' ... , Sm), ... , T: =
t:(Sl' ... , Sm), where each tj is a function of Sl' ... , Sm, which satisfies:
var9 [Tt] ~ var 9 [1)] for every () E E:j, j = 1, ... , r, for any
(i)
unbiased estimator (11, ... , T,) of (r 1 ((}), •.. , r,((}».
(ii) The ellipsoid of concentration of (Tt, ... , T:) is contained in
the ellipsoid of concentration of (11, ... , 7;.), where (T1' ... , T,) is any
unbiased estimator of (r1 ((}), ••• , rr«(}»' II/I
There are four different maximal subscripts, all of which are intended.
n denotes the sample size, k denotes the dimension of the parameter 0, m is the
number of real-valued statistics in our jointly complete and sufficient set, and r
is the dimension of the vector of functions of the parameter that we are trying
to estimate. In practice, it will turn out that usually' k = m. The estimator
(Tt, ... , T:) is optimal in the sense that among unbiased estimators it is the
best estimator using any of the four generalizations of variance that have been
proposed.
Just as was the case in using Theorem 10, we have two ways of finding
(Tt, ... , r:). The first is to guess the correct form of the functions tt, ... , t:,
which are functions of Sl' ... , Sm' that will make them unbiased estimators of
't1 (0), ... , 'f,(l1). The second is to find any set of unbiased estimators of
'fl(O), ••• , 'f,i.l1) and then calculate the conditional expectation of these unbiased
estimators given the set of jointly complete and sufficient statistics. We employ
only the first method in the following examples.
EXAMPLE 49 Let Xi) ... , Xn be a random sample from the density

f(x; 01 , ( 2 ) = [1/(0 2 - ( 1 )]/(81. (2)(X)' Suppose we want to jointly estimate
the range and midrange, that is, 't1 (8) = 82 - 01 and 't2(0) = (0 1 + ( 2 )/2.
We know that Yi = min [Xl' .• " Xn] and Yn = max [Xi) , .. , Xn] are
jointly sufficient (see Example 23); also, they are jointly complete (see
Example 47). Hence, to find the unbiased estimator (T!, Tn which has
uniformly smallest variance for each component among all unbiased
estimators, it suffices to find the unbiased estimator that is a function of
the jointly complete sufficient statistics. Since 4[Y1]= 01+ (0 2- (1)/(n+ 1)
and 4[Yn] = O2 - (82 - (1)/(n + 1), ([(n + 1}/(n - 1}](Yn- Y1 ), (Yl + Yn)/2)
is the unbiased estimator of (0 2 - 01 , (01 + ( 2 )/2) that we are seeking. 111I
EXAMPLE 50 Let Xi) ... , Xn be a random sample from the normal den-
sity f(x; 01, ( 2 ) = tP/l.fl2(X). By Examples 22 and 48, L Xi and L Xl are
jointly complete and sufficient statistics. Hence, by Theorem 17,
(L X,ln, L (X, - X)2/(n - 1)) is an unbiased estimator of (Il, 0'2) whose
corresponding ellipsoid of concentration is contained in the ellipsoid of
concentration of any other unbiased estimator. [NoTE: L (X, - X)2 =
L Xf - nX2; so the estimator L (Xi - X)2/(n - 1) is a function of the
L
jointly complete and sufficient statistics Xi and X;,] L
For this same example, suppose we want to estimate that function
of 8 = (p., 0'2) satisfying the following integral equation:
for (X fixed and known. 'teO) is that point which satisfies P[Xi > 'teO)] = (X;
that is, it is that point which has 100(X percent of the mass of the population
density to its right, or r(O) is the (1 - (X)th quantile point. We have
1 - (X = <b(['t(0) - 1l]10'); so 'teO) = Jl + ZI-«O', where Zl-« is given by
«)(Zl -J = 1 - (x, Since (X is known, Zl _« can be obtained from a table
of the standard normal distribution. To find the LTMVUE of reO), it
suffices to ~nd the unbiased estimator of Jl + Zl _« 0' which is a function of
3S8 PARAMETRIC POINT ESTIMATION vu
L Xi and L xl· We know that X is the UMVUE of /1, and it can be

verified that
r[(n - 1)/2]
r(nI2)j2 J'L (Xi - X)2 = T*
say, is the UMVUE of (1; hence X + Zl-a T* is the UMVUE of T(O). We

have employed Theorem 17 for r = 1; our vector of functions of the
parameter that we wanted to estimate was unidimensional. fill
9 OPTIMUM PROPER TIES OF

MAXIMUM-LIKELIHOOD ESTIMATION
Several methods of finding point estimators were presented in Sec. 2 of this
chapter. There, and in succeeding sections, we have particularly emphasized
the method of maximum likelihood. In this section we will partially justify
such emphasis by considering some optimum properties of maximum-likelihood
estimators.
For simplicity of presentation, let us consider the maximum-likelihood
estimation of the parameter 0, which is to be estimated on the basis of a random
sample from a density f(· ; 0), where 0 is assumed to be a real number. That is,
let us consider the unidimensional-parameter case and estimate 0 itself. Recall
that for the observed sample Xl' ... , Xn the maximum-likelihood estimate of 0
is that value, say 0, of 0 which maximizes the likelihood function L(O; Xl' ... , Xn)
= nf(Xi;
n
i= 1
0). Let an = 8iXl , ... , Xn) denote the maximum-likelihood
estimator of 0 based on a sample of size n. We defined and discussed in Sec. 3
of this chapter a number of properties that an estimator mayor may not possess.
Recall that some of these properties, such as unbiasedness and uniformly
minimum variance, are referred to as small-sample properties, and others of
these properties, such as consistency and best asymptotically normal, are referred
to as large-sample properties. The use of the word" small" in "small-sample"
is somewhat misleading since a small-sample property is really a property that is
defined for a fixed sample size, which may be fixed to be either small or large.
By a large-sample property, we mean a property that is defined in terms of the
sample size increasing to infinity. Our main result of this section will be con-
tained in Theorem 18 below and will concern optimum large-sample properties
of maximum-likelihood estimation.
9 OPTIMUM PROPERTIES OF MAXIMUM-LIKELIHOOD ESTIMATION 359
We have already observed some small-sample properties of maximum-

likelihood estimation. For instance, we have noted two things: first, that some
maximum-likelihood estimators are unbiased and others are not and, second,
that some maximum-likelihood estimators are uniformly minimum-variance
unbiased and others are not. For example, in the density f(x; 8) = 4>6, I (x) the
maximum-likelihood estimator of 8 is X, which is the uniformly minimum-
variance unbiased estimator of 8, whereas in the density f(x; 8) = (1/8)/[0. 8](X)
the maximum-likelihood estimator of 8 is Y,. = max [Xl' ... , X,.], which is
biased. [We might note here that the Yn in this last example can be corrected
for bias by multiplying Yn by (n + l)/n and that the estimator that is thus
obtained is uniformly minimum variance unbiased.]
One property that it seems reasonable to expect of a sequence of esti-
mators is that of consistency. Theorem 18 will show, in particular, that generally
a sequence of maximum-likelihood estimators is consistent.
Theorem 18 If the density f(x; 8) satisfies certain regularity conditions

and if An = 8n(Xh ••. , Xn) is the maximum-likelihood estimator of 8 for
a random sample of size n from f(x; 8), then:
(i) An is asymptotically normally distributed with mean 8 and
variance 1/1149[[:0 logf(X; 6)]1
(ii) The sequence of maximum-likelihood estimators AI' ... , An'
... is best asymptotically normal (BAN). / ///
We will not be able to prove Theorem 18. In fact, we have not precisely
stated it, inasmuch as we have not delineated the regularity conditions. We
do, however, want to emphasize what the theorem says. Loosely speaking, it
says that for large sample size the maximum-likelihood estimator of 8 is as
good an estimator as there is. (Other estimators might be just as good but not
better.)
We might point out one feature of the theorem, namely, that the asymptotic
normal distribution of the maximum-likelihood estimator is not given in terms of
the distribution of the maximum-likelihood estimator. It is given in terms of
f( • ; B), the density sampled. Also, the variance of the asymptotic normal
distribution given in the theorem is the Cramer-Rao lower bound.
EXAMPLE 51 Let Xl"'" Xn be a random sample from the negative

exponential distribution f(x; 8) = Be-Ox/[o. a:»(x). It can be routinely
demonstrated that the maximum-likelihood estimator of B is given by
360 PARAMETRIC POINT ESTIMA nON VB
Fa
nIL Xi = II X Fa. According to Theorem 18 above, the maximum-likeli-

1
hood estimator has an asymptotic normal distribution with mean 0 and
variance equal to
1
-----------:~ = -.
n4.[[! 10gf(X; 0)] 2] n
(See Example 29.) IIII
We have ordinarily considered estimation of T(O) some function of 8,

rather than estimation of 8 itself. For maximum-likelihood estimation, we
noted (see Theorem 2) that the maximum-likelihood estimator of -r(8) was given
by T(a), where a was the maximum-likelihood estimator of O. If we assume
that T( . ) is differentiable, then it can be shown that T(a) has an asymptotic
normal distribution with mean T(O) and variance
[T'(8)]2
a
"88 [[oologf(X; Ii)] ]
2 '
which is the Cramer-Rao lower bound. (See Theorem 7.)

Maximum-likelihood estimators possess similar optimum large-sample
properties in the case of a k-dimensional parameter. For instance, it can be
proved (again under regularity conditions) that the joint distribution of the
maximum-likelihood estimators is asymptotically distributed as a multivariate
normal distribution. Let us illustrate for the case when k = 2; that is,
8 = (0 1 , 82 ). Recall that the bivariate normal distribution is specified by the
five parameters /1h /12, ui, ui, and p. (See Sec. 5 of Chap. IV.) It tUrns out
that under certain regularity conditions the joint distribution of the maximum-
a
likelihood estimators a1 and z is asymptotically distributed as a bivariate
normal distribution with parameters /11 = 81, /12 = 82 ,
-.t.[::~IOgf(X; 0)]
(J2
1- - ,
n~
2 -
u2
-.t'[::i logf (X; 0)]
-
nA
9 oPTIMUM PROPERTIES OF MAXIMUM-LIKELlHOOD ESTIMATION 361
and
PU 1U 2 = ,
nA
where
A = .,.[::~ log/(X; 9)]".[::~ log/(X; 9)] - (8.[00::9, log/(X; 9)])'.

EXAMPLE 52 Let Xl' •.. , X,. be a random sample from the density
f( x', ll)
v
= f(x', 0H 8)
2
= '+'01.Oz (x) = J2n8
,I,. 1 e-(1/ 20 z)(x-0 1 )Z
•
2
We have already derived, in Example 6, the maximum-likelihood estimators

of 01 and 82 ; they are, respectively,
1 II
G1 = - ~ X· and
nf '
According to the above, the asymptotic large-sample joint distribution of
G1 and G2 is a bivariate normal distribution with means 01 and O2 •
Since 10gf(X; 0) = -t log 2n - t log 8 2 - (l/20 2 )(X - ( 1)2, the required
deri vatives are
82 I
80 2 logf( X; 0) = - ll'
1 u2
82 X - 01
80 80 10gf(X; 0) = - 02 '
2 1 2
and
and because
and
-&.[::~ logf(X; II)] = ~r

which gives A = 1/20~. Finally, then, ui = 021n, u~ = 20iln, and p = o.
IIII
PROBLEMS
1 An urn contains black and white balls. A sample of size n is drawn with replace-
ment. What is the maximum-likelihood estimator of the ratio R of black to
white balls in the urn? Suppose that one draws balls one by one with replacement
until a black ball appears. Let X be the number of draws required (not counting
the last draw). This operation is repeated n times to obtain a sample X., X 2 , ••• ,
X n • What is the maximum-likelihood estimator of R on the basis of this sample '/
2 Suppose that n cylindrical shafts made by a machine are selected at random from
the production of the machine and their diameters and lengths measured. It is
found that Nil have both measurements within the tolerance limits, Nl2 have
satisfactory lengths but unsatisfactory diameters, N21 have satisfactory diameters
but unsatisfactory lengths, and N22 are unsatisfactory as to both measurements.
2: N'J = n. Each shaft may be regarded as a drawing from a multinomial popUla-
tion with density
~ll
PH P12 P21
~12 ~21(l
-PH - Pl2 - P21
)~22
for X'J = 0, l; 2: Xu = 1
having three parameters. What are the maximum-likelihood estimates of the
parameters if Nu = 90, N12 = 6, N21 = 3, and N22 = 1 '/
3 Referring to Prob. 2. suppose that there is no reason to believe that defective
diameters can in any way be related to defective lengths. Then the distribution
of the X,Jcan be set up in terms of two parameters: PI, the probability of a satis-
factory length, and q., the probability of a satisfactory diameter. The density of
the Xu is then
(Plqlt ll [Pl(1 - ql)t 12 [(1 - PI)qlf 2I [(l - Pl)(l - ql)t
22
for XIJ =0, l; 2: X'J = 1.
What are the maximum-likelihood estimates for these parameters'/ Are the prob-
abilities for the four classes different under this model from those obtained in the
above problem '/
4 A sample of size nl is to be drawn from a normal population with mean P,l and
variance 0':.
A second sample of size n2 is to be drawn from a normal population
with mean P,2 and variance O'~. What is the maximum-likelihood es~timator of
• • Jl~_ h
() = P,l - P,2 '/ If we assume that the total sample SIze n = nl + n2 IS xed, Ow
should the n observations be divided between the two populations in order to
minimize the variance of the maximum-likelihood estimator of () '/
PROBLEMS 363
5 A sample of size n is drawn from each of four normal populations. all of which
have the same variance a 2 • The means of the four populations are a + b + c,
a + b - c, a - b + c, and a - b - c. What are the maximum-likelihood estima-
tors of a, b, c, and a 2 ? (The sample observations may be denoted by X'J, i = 1,
2, 3,4 andj = 1,2, ... , n.)
6 Observations X., X 2 , ••• , Xn are drawn from normal populations with the same
mean ft but with different variances a~, a~, ... ,a;. Is it possible to estimate all
the parameters? If we assume that the a~ are known, what is the maximum-
likelihood estimator of ft ?
7 The radius of a circle is measured with an error of measurement which is dis-
tributed N(O, ( 2 ), a 2 unknown. Given n independent measurements of the
radius, find an unbiased estimator of the area of the circle.
8 Let X be a single observation from the Bernoulli density f(x; (J) =
(Jx(1- (J)l-xI{o.l)(X), where 0 < (J < 1. Let tl(X) = X and t 2(X) = l.
(a) Are both tl(X) and t 2(X) unbiased? Is either?
(b) Compare the mean-squared error of tl(X) with that of t 2(X).
9 Let Xl, X 2 be a random sample of size 2 from the Cauchy density
.(J _ 1 (J
f(x, ) - 17[1 + (x _ 8)2]' - 00 < < 00.
Argue that (Xl + X 2 )/2 is a Pitman closer estimator of (J than Xl is. [Note that
(Xl + X 2 )/2 is not more concentrated than Xl since they have identical distribu-
tions.]
10 Let (J denote some physical quantity, and let Xl, ••• , Xn denote n measurements of
the physical quantity. If (J is estimated by 0, then the residual of the ith measure-
ment is defined by X, - 0, i = 1, "', n. Show that there is only one estimator
with the property that the residuals sum is 0, and find that estimator. Also,
find that estimator which minimizes the sum of squared residuals.
II Let Xl, ••• , Xn be a random sample from some density which has mean ft and
variance a 2 •
n
(a) Show that .2 a, X, is an unbiased estimator of ft for any set of known constants
1
n
al, •.. , an satisfying .2 a, =
1
1.
n n
(b) If L a, =
1
1, show that var [L a, X,] is minimized for a, =
1
lIn, ; = 1, ... , n.
n n
[HINT: Prove that

-
L a: = L (a, -
1 1
I/n)2 + lIn when
f~ a, = 1.]
12 Let X., ••• , Xn be a random sample from the discrete density function f(x; 8) =
(Jx(1 - (J)l-xI{o. l)(X)~ where 0 :::;;: (J < l. Note that e = {(J: 0 < (J < n.
(a) Find a method-of-moments estimator (J, and then find the mean and mean-
squared error of your estimator.
(b) Find a maximum-likelihood estimator of (), and then find the mean and
mean-squared error of your estimator.
364 pAR.AMJ!TR.IC poINT ESTIMATIoN VB
13 Let Xl, X 2 be a random sample of size 2 from a normal distribution with mean 8
and variance 1. Consider the following three estimators of 8:
T, = t,(Xl, X 2 ) = IX, + tX2
T2 = t 2 (X" X 2 ) = IX, + iX2
T3 = t 3 (X" X 2 ) = IX, + IX2 •
For the loss function t(t; 8):::;;: 38 2 (t - 8)2, find £1t,,(8) for i = 1, 2, 3, and
(a)
sketch it.
(b) Show that T, is unbiased for i = 1, 2, 3.
14 Let X" ... , XII: .•. be independent and identically distributed random variables
from some distribution for which the first four central moments exist. We know
that 8[8 2 ] 0'2 and
var [8 2 ] = -1 ( P,4 - -
n--3 0'4) ,
n n-l
where
Is 8 2 a mean-squared-error consistent estimator of 0'2 ?

15 In genetic investigations one frequently samples from a binomial distribution
f(x) = (7) p'q'" -. except that observations of x = 0 are impossible; so, in fact, the
sampling is from the conditional (truncated) distribution

m)
X
yqm-x
( qm I{l.
_ 2 •••• • m}(x).
Find the maximum-likelihood estimator of p in the case m = 2 for samples of

size n. Is the estimator unbiased ?
16 Let X be a single observation from N(O, 8). (8 = 0'2.)
(a) Is X a sufficient statistic?
(b) Is I XI a sufficient statistic?
(c) Is X2 an unbiased estimator of 8?
(d) What is a maximum-likelihood estimator of V 8?
(e) What is a method-of-moments estimator of V 8?
17 Let X have the density f(x;8) = (8/2)lxl(1- 8)'-l x II{_,. o.,}(x), o:s;: 8 <1.
Define t(x) = 2/{l)(x).
(a) Is X a sufficient statistic? A complete statistic?
(b) Is IXI a sufficient statistic? A comp~e statistic?
(c) What is a maximum-likelihood estimator of 8?
(d) Is T= t(X) an unbiased estimator of 8?
(e) Does f(x; 8) belong to an exponential class ?
(f) Find an estimator with uniformly smaller mean-squared error than that of
t(X), if such exists.
PROBLEMS 365
18 Let Xh X 2 , ••• , Xn be a random sample from the density

I(x; 8) = ()x- 2 /(6. CIJ)(x)
where 8>0.
(a) Find a maximum-likelihood estimator of ().
(b) Is YI = min [X., ••. , Xn] a sufficient statistic?
19 Let Xl, ••• , XII be a random sample from I(x; 8) = Ie -Ix - 61, - 00 < () < 00.
(a) Discuss sufficiency for this density.
(b) Obtain a method-of-moments estimator of ().
(e) Find a maximum-likelihood estimator of ().
(d) Does I(x; 8) belong to an exponential class?
20 Find a maximum-likelihood estimator for oc in the density I(x; oc) =
(2/oc 2 )(oc-- x)/eo • )(x) for samples of size 2 Is it a sufficient statistic? Estimate oc
II
by the method of moments. What is the maximum-likelihood estimator of the

population mean?
21 Let Xl, ••. , Xn be a random sample from I(x; 8) = (1/8)/(0. 61(X), where () > O.
Define Yn = max [X., •.. , Xn] and Yl = min [X., ... , X n]·
(a) Estimate () by the method of moments. Call the estimator T l • Find its
mean and mean-squared error.
(b) Find the maximum-likelihood estimator of (). Call the estimator T 2 • Find
its mean and mean-squared error.
(e) Among all estimators of the form a Yn , where a is a constant which may depend
on n, find that estimator which has uniformly smallest mean-squared error.
Call it Ts. Find its mean and mean-squared error.
(d) Find the UMVUE of (). Call it T 4 • Obtain its mean and mean-squared
error.
(e)Let Ts = Yl + Yn • Find the mean and mean-squared error of Ts.
(I) What estimator of () would you use and why?
(g) Find the maximum-likelihood estimator of the variance of the population.
22 Let Xl, . .. , Xn be a random sample from the Bernoulli distribution, say P[X = 1] =
() = 1 - P[X = 0].
(a) Find the Cramer-Rao lower bound for the variance of unbiased estimators of
()(1 - 8).
(b) Find the UMVUE of ()(1 - 0) if such exists.
23 Assuming r known, find the maximum-likelihood estimator for" for a random
sample of size n from a gamma distribution. Find a sufficient statistic if one
exists,,\ Is your maximum-likelihood estimator unbiased? Is there an UMVUE
of ,,?
24 Let X., ... , Xn be a random sample from ()x 6 -1leo. 1)(x), where () > O.
(a) Find the maximum-likelihood estimator of t-t = ()/(1 + ()).
(b) Find a sufficient statistic, and check completeness. Is L: X, a sufficient
statistic?
(e) Is there a function of () for which there exists an unbiased estimator whose
variance coincides with the Cramer-Rao lower bound?
*(d) Find the UMVUE of each of the following: (), 1/(), t-t = ()/(l + 8).
25 Let X., •.• , X,. be a random sample from the binomial distribution
(:)P"(I- p)m-., x ~ 0,1, ••. , m, where m is known and 0 <p< 1.
Estimatep by the method of moments and the method of maximum likelihood.
(a)
Is there an UMVUE of p? If so, find it.
(b)
*26 Let Xl, ... , Xn be a random sample from the discrete density function
f(x; 8) = (1/8)lu . 2 ••••• 9) (X) ,

e
where () = 1, 2, .... That is, = {(): () = 1, 2, ... } = the set of positive integers.
(a) Find a method-of-moments estimator of (). Find its mean and mean-squared
error.
(b) Find a maximum-likelihood estimator of (). Find its mean and mean-squared
error.
(c) Find a complete sufficient statistic.
(d) Let T = Yn , the largest order statistic. Show that the UMVUE of () is
[Tn+1 - (T-1)n+1]/[T" - (T- 1)n].
27 Let X be a single observation from the density [l/B«(), 8)]X 9- 1(1- x)9- 11(0. 1)(X).
Is X a sufficient statistic? Is X complete?
28 An experimenter knows that the distribution of the lifetime of a certain component
is negative exponentially distributed with mean 1/(). On the basis of a random
sample of size n of lifetimes he wants to estimate the median lifetime. Find both
the maximum-likelihood and uniformly minimum-variance unbiased estimator of
the median.
29 Let Xl, •.• , Xn be a random sample from N«(), 1).
(a) Find the Cramer-Rao lower bound for the variance of unbiased estimators of
(), ()2, and p[X > 0].
(b) Is there an unbiased estimator of ()2 for n = ]? If so, find it.
(c) Is there an unbiased estimator of P[X > O]? If so, find it.
(d) What is the maximum-likelihood estimator of P[X > O]?
(e) Is there an UMVUE of ()2? If so, find it.
(f) Is there an UMVUE of P[X > O]? If so, find it.
30 For a random sample from the Poisson distribution, find an unbiased estimator of
7(A) = (1 + A)e- A• Find a maximum-likelihood estimator of T(A). Find the
UMVUE of T(A).
31 Let X., .•. , Xn be a random sample from the density
2x
f(x; 8) = 7f2 1(0.9)(x)
where () > O.
(a) Find a maximum-likelihood estimator of ().
v-
(b) Is Yn = max [Xl, ••• , Xn] a sufficient statistic? Is Yn complete?
(c) Is there an UMVUE of ()? If so, find it.
PROBLEMS 367
32 Let XI, ••• , X" be a random sample from the density

f(x; U) = 0(1 + x)-U + 9)/(0. C()(x) for 0 >0.
(a) Estimate 0 by the method of moments assuming 0 > 1.
(b) Find the maximum-likelihood estimator of I/O.
(c) Find a complete and sufficient statistic if one exists.
(d) Find the Cramer-Rao lower bound for unbiased estimators of I/O.
(e) Find the UMVUE of I/O if such exists.
(f) Find the UMVUE of 0 if such exists.
33 Let X h ••• , X" be a random sample from
f(x; 0) for 0 o.
(a)Find a maximum-likelihood estimator of O.
(b)Suppose n 1, so that you have only one observation, say X = Xl' Clearly
X is a sufficient statistic. Is X a minimal sufficient statistic? Is X complete?
34 Let Xl, ••. , X" be a random sample from the negative exponential density
f(x; 0) = Oe-h/[o. C()(x).
(a) Find the uniformly minimum-variance unbiased estimator of var [Xl] if such
exists.
(b) Find an unbiased estimator of I/O based omy on Y~") = min [Xl, ... , XJ.
Is your sequence of estimators mean-SQuared-error consistent?
35 Let Xl, ~ .• , X" be a random sample from the density
log 0
f(x; 0) = 0 _ 1 OX/(o.l)(x), 0>1.
(a)Find a complete sufficient statistic if there is one.

(b)Find a functiOn of 0 fQr which there exists an unbiased estimator whose
variance coincides with the Cramer-Rao lower bound if such exists.
n
• 36 Show that
.g. [{ :0 log/(X; 6) = - 8. [ :;. log/(X; 6) l

37 Let Xl, "', X" be a random sample from the density
f(x; U) = e-(x-B) exp (_e-(X-B»,
where - 00 < 0 < 00.
(a) Find a method-of-moments estimator of O.
(b) Find a maximum-likelihood estimator of O.
(c) Find a complete sufficient statistic.
(d) Find the Cramer-Rao lower bound for unbiased estimators of O.
(e) Is there a function of 0 for which there exists an unbiased estimator, the
variance of which coincides with theCramer-Rao lower bound? If so, find it.
*(f) Show that r'(n)/r(n) - log (2: e- Xt ) is the UMVUE of O.
368 pARAMBTRIC POINT ESTIMATION vn
38 Let X., ••• , Xn denote a random sample from
I(x; fJ) = 19(x) = fJ/l(x) (1 - fJ)/o(X),
where 0 fJ < 1 andh(') and/o(') are known densities.

(a) Estimate fJ by the method of moments.
(b) For n = 2, find a maximum-likelihood estimator of fJ.
(c) Find the Cramer-Rao lower bound for the variance of unbiased estimators
of fJ.
39 Suppose that a(') and b(') are two nonnegative functions such that I(x; fJ) =
a(fJ)b(x)/(o.9)(X) is a probability density function for each fJ > O.
(a) What is a maximum-likelihood estimator of fJ?
(b) Is there a complete sufficient statistic? If so, find it.
(e) Is there an UMVUE of fJ? If so, find it.
40 Let Xl, ••• , Xn be a random sample from N(fJ, fJ), fJ > O.
(a) Find a complete sufficient statistic if such exists.
(b) Argue that Xis not an UMVUE of fJ.
(e) Is fJ either a location or scale parameter?
41 Let Xl, ..• , Xn be a random sample from N(fJ, fJ2), - 00 < fJ < 00.
(a) Is there a unidimensional sufficient statistic?
(b) Find a two-dimensional sufficient statistic.
(e) Is Xan UMVUE of fJ? {HINT; Find an unbiased estimator of fJ based on 8 2 ;
call it T*. Find a constant a to minimize var [aX + (1 - a)T*].}
(d) Is fJ either a location Or scale parameter?
42 Let Xl, ... , Xn be a random sample of size n from the density
1
I(x; fJ) = -0 1[9. 29](X), fJ > O.
Find a maximum-likelihood estimator of fJ.

(a)
We know that Y1 and Yn are jointly sufficient. Are they jointly complete?
(b)
Find the Pitman estimator for the scale parameter fJ.
(e)
(d) For a and b constant (they may depend on n), find an unbiased estimator of
fJ of the form a YI b Y n satisfying P[ Y n/2 < a YI b Yn < Yd = 1 if such
exists. Why is P[Yn/2 < aYI bY" YI ] = 1 desirable?
43 Let Z 1, ••• , Z" be a random sample from N(O, fJ2), fJ > O. Define X, = IZ, I" and
consider estimation of fJ and fJ2 on the basis of the random sample Xl, ... , X".
(a) Find the UMVUE of fJ2 if such exists.
(b) Find an estimator of fJ2 that has uniformly smaller mean-squared error than
the estimator that you found in part (a).
(c) Find the UMVUE of fJ if such exists.
(d) Find the Pitman estimator for the scale parameter fJ.
(e) Does the estimator that you found in part (d) have uniformly smaller mean-
squared error than the estimator that you found in part (e)?
PROBLEMS 369
44 Let XI, ••• , XII be a random sample from

f(x; 0) = e-(X-6)I[6. o:»(x) for - 00 < () < 00.
(a) Find a sufficient statistic.

(b) Find a maximum-likelihood estimator of 0.
(c) Find a method-of-moments estimator of 0.
(d) Is there a complete sufficient statistic? If so, find it.
(e) Find the UMVUE of () if one exists.
(f)Find the Pitman estimator for the location parameter ().
(g)Using the prior density g(O) = e- 6 I(o. 0:»(0), find the posterior Bayes estimator
of 0.
45 Let Xl, ••• , X" be a random sample from f(xI8) 8x 6 - 1 I(o.1)(x), where e > o.
Assume that the prior distribution of e is given by
where r and A are known.

(a) What is the posterior distribution of e?
(b) Find the Bayes estimator of () with respect to the given gamma prior distribu-
tion using a squared-error loss function.
46 Let X be a single observation from the density f(xl 0) = (2x/()2)I(o. ,)(x), where
() > O. Assume that e has a uniform prior distribution over the interval (0, 1).
For the loss function (t; 0) ()2(t - 0)2, find the Bayes estimator of 0.
47 Let Xl, X 2 , ••• , XII be a random sample of size n from the follOWing discrete
density:
where () > O.
(a) Is there a unidimensional sufficient statistic? If so, is it complete?
(b) Find a maximum-likelihood estimator of 0 2 = P[XI = 2]. Is it unbiased?
(c) Find an unbiased estimator of () whose variance coincides with the correspond-
ing Cramer-Rao lower bOund if such exists. If such an estimate does not
exist, prove that it does not.
(d) Find a uniformly minimum-variance unbiased estimator of ()2 if such exists.
(e) Using the squared-error loss function find a Bayes estimator of 0 with respect
to the beta prior distribution
(f) Using the squared-error loss function, find a minimax estimator of ().
(g) Find a mean-squared error consistent estimator of 02 •
48 Let Xl, ... , Xn be a random sample from a Poisson density

e-fJ(Jx
- , - 1,0. 1 • ... }(X),
X.
where () > O. For a squared-error loss function find the Bayes estimator of () for
a gamma prior distribution. Find the posterior distribution of 0. Find the
posterior Bayes estimator of T( () = P[Xf = 0].
49 Let Xl, ... , Xn be a random sample from f(xl () = (1/()/(o.fJ)(x), where () > o.
For the loss function (t - 8)2/()2 and a prior distribution proportional to
()-IZ/(1. CXI)«() find the Bayes estimator of 8.
50 Let Xl, ... , Xn be a random sample from the Bernoulli distribution. Using the
squared-error loss function, find that estimator of () which has minimum area
under its risk function.
51 Let Xl, ... , Xn be a random sample from the geometric density
f(x; 8) = ()(1 - ()"I,o.l . ... }(x)

where 0 < () < 1.
(a) Find a method-of-moments estimator of ().
(b) Find a maximum-likelihood estimator of ().
(c) Find a maximum-likelihood estimator of the mean.
(d) Find the Cramer-Rao lower bound for the variance of unbiased estimators of
1 - ().
(e) Is there a function of () for which there exists an unbiased estimator the
variance of which coincides with the Cramer-Rao lower bound? If so,
find it.
(f) Find the UMVUE of (1- ()/() if such exists.
(g) Find the UMVUE of () if such exists.
(h) Assume a uniform prior distribution and find the posterior distribution of 0.
For a squared-error loss function, find the Bayes estimator of () with respect
to a uniform prior distribution.
52 Let () be the true I.Q. of a certain student. To measure his I.Q., the student takes
a test, and it is known that his test scores are normally distributed with mean fL and
standard deviation 5.
(a) The student takes the I.Q. test and gets a score of 130. What is the maximum-
likelihood estimate of ()?
(b) Suppose that it is known that I.Q.'s of students of a certain age are distri-
buted normally with mean 100 and variance 225; that is, 0 ,.... N(100, 225).
Let X denote a student's test score [X is distributed N«(), 25)]. Find the pos-
terior distribution of 0 given X = x. What is the posterior Bayes estimate
of the student's I.Q. if X = 130.
*53 Let Xl, ... , Xn be a random sample from the density
PROBLEMS 371
where - 00 < a < 00 and f3 o. Show that Yl and 2: Xi are jOintly sufficient.
It can be shown that Yl and 2: (Xi - Yl ) are jOintly complete and independent of
each other. Using such results, find the estimator of (a, f3) that has an ellipsoid of
concentration that is contained in the ellipsoid of concentration of any other un-
biased estimator of (a, (J). (Yl = min [Xl"'.' X n1.)
54 Let X., ... , X. be a random sample from the density
f(x; a, 8) = (1- 8)(JX-«I(<<, «+1, ... )(x),
where - 00 < a < 00 and 0 < (J < 1.
(a) Find a two-dimensional set of sufficient statistics.
• (b) Find the maximum-likelihood estimator of <a,
(J).
55 Let Xl, ... , Xn be a random sample from the density
f<x; 8) = e2x l(o.9}(x) + 2(1-

1_
x)
1(9.1](X),
(J
where 0 < (J < 1.

(a) Estimate (J by the method of moments.
(b) Find the maximum-likelihood estimator of (J for n = 1 and n = 2.
(c) For n 1 find a complete sufficient statistic if such exists. Find a UMVUE
of (J for n = 1 if such exists.
*(d) Find the maximum-likelihood estimator of (J.
VIII
PARAMETRIC INTERVAL ESTIMATION

Chapter VII dealt with the point estimation of a parameter, or more precisely,
point estimation of a value of a function of a parameter. Such point estimates
are quite useful, yet they leave something to be desired. In all those cases when
the point estimator under consideration had a probability density function, the
probability that the estimator actually equaled the value of the parameter being
estimated was O. (The probability that a continuous random variable equals
anyone value is 0.) Hence, it seems desirable that a point estimate should be
accompanied by some measure of the possible error of the estimate. For in-
stance, a point estimate might be accompanied by some interval about the point
estimate together with some measure of assurance that the true value of the
parameter lies within the interval. Instead of making the inference of estimating
the true value of the parameter to be a point, we might make the inference of
estimating that the true value of the parameter is contained in some interval.
We then speak of interval estimation, which is to be the subject of this chapter.
Like point estimation, the problem ofinterval estimation is twofold. First,
there is the problem offinding interval estimators, and, second, there is the prob-
2 CONFIDENCE INTERVALS 373
lern of determining good, or optimum, interval estimators. The considerations

of these two problems that will appear in this chapter will be incomplete.
Further considerations will be presented at the end of the next chapter on testing
hypotheses. The mathematics of interval estimation and hypotheses testing are
closely related. Either concept could be used to introduce the other. In this
book, we have decided to introduce interval estimation first, right after our
presentation of point estimation, then introduce hypotheses testing, and finally
point out the close mathematical relationship between the two.
The introduction to interval estimation that appears in this chapter will
not be as thorough as was our discussion of point estimation in the last chapter.
One should not infer from this that interval estimation is less important since in
practice the opposite is usually true. It is just easier to present the basic theory
of point estimation. No concerted effort will be given to the problem of finding
optimum interval estimators. The chapter will be divided into six main sections,
the first being this introductory section. Section 2 will be devoted to confidence
intervals, where the notion is introduced and defined. One method of finding
confidence intervals will also be given as well as some idea as to what an optimum
confidence interval might be. Section 3 will consider several examples of con-
fidence intervals that are associated with sampling from the normal distribution.
Such discussion will hinge on the results of Sec. 4 of Chap. VI. Several general
methods of finding confidence intervals are given in Sec. 4; another method,
which utilizes the theory of hypotheses testing, will be given at the end of Chap.
IX. A brief discussion of large-sample confidence intervals appears in Sec. 5,
and Sec. 6 presents another type of interval estimation, namely, Bayesian interval
estimation.
2 CONFIDENCE INTERVALS
2.1 An Introduction to Confidence Intervals
In practice, estimates are often given in the form of the estimate plus or minus a
certain amount. For instance, an electric charge may be estimated to be
(4.770 + .005)10- 10 electrostatic unit with the idea that the first factor is very
unlikely to be outside the range 4.765 to 4.775. A cost accountant for a publish-
ing company in trying to allow for all factors which enter into the cost of produc-
ing a certain book (actual production costs, proportion of plant o,("erhead, pro-
portion of executive salaries, etc.) may estimate the cost to be 83 + 4.5 cents per
volume with the implication that the correct cost very probably lies between
374 PARAMETRIC INTERVAL ESTIMA nON vm
78.5 and 87.5 cents per volume. The Bureau of Labor Statistics may estimate
the number of unemployed in a certain area to be 2.4 + .3 million at a given
time, feeling rather sure that the actual number is between 2.1 and 2.7 million.
What we are saying is that in practice One is quite accustomed to seeing estimates
in the form of intervals.
In order to give precision to these ideas, we shall consider a particular
example. Suppose that a random sample (1.2, 3.4, .6, 5.6) of four observations
is drawn from a normal population with an unknown mean JI and a known
standard deviation 3. The maximum-likelihood estimate of JI is the mean of the
sample observations:
x = 2.7.
We wish to determine upper and lower limits which are rather certain to contain
the true unknown parameter value between them.
In general, for samples of size 4 from the given distribution the quantity
z=X
__ JI
~
will be normally distributed with mean 0 and unit variance. X is the sample
mean, and; is (J/J~. Thus the quantity Z has a density
1
fz(z) = ¢(z) = ;_ e -i z2 ,
y2n
which is independent of the true value of the unknown parameter; so we can

compute the probability that Z will be between any two arbitrarily chosen
numbers. Thus, for example,
1.96
P[ -1.96 < Z < 1.96] =
f-1.96
¢(z) dz = .95. (1)
In this relation the inequality - 1.96 < Z, or

X - JI
-1.96 < ; ,
is equivalent to the inequality

JI <X + ;(1.96) = X + 2.94,
and the inequality
Z < 1.96
is equivalent to
J1 > X - 2.94.
We may therefore rewrite Eq. (1) in the form
P[X - 2.94 < J1 < X + 2.94] = .95, (2)
and substituting 2.7 for X we obtain the interval

( - .24, 5.64).
It is at this point that a certain abuse of language takes place since the
random interval (X - 2.94, X + 2.94) and the interval (- .24, 5.64) are each
called a confidence interval, or more precisely a 95 percent confidence interval.
[The interval (- .24,5.64) is the value of the random interval (X - 2.94, X + 2.94)
when X = 2.7.] The meaning of Eq. (2) is the following: The probability that
the random interval (X - 2.94, X + 2.94) covers the unknown true mean J1 is .95.
That is, if samples of size 4 were repeatedly drawn from the normal population
and if the random interval (X - 2.94, X + 2.94) were computed for each sample,
then the relative frequency of those intervals that contain the true unknown mean
J1 would approach 95 percent. We therefore have considerable confidence that
the observed interval, here (- .24, 5.64), covers the true mean. The measure of
our confidence is .95 because before the sample was drawn .95 was the prob-
ability that the interval that we were going to construct would cover the true
,\
mean. .95 is called the confidence coefficient.

Similarly, intervals with any desired degree of confidence between 0 and 1
can be obtained. Thus, since
P[-2.58 <Z < 2.58] = .99,
a 99 percent confidence interval for the true mean is obtained by converting the
inequalities as before to get
P[X - 3.87 < J1 <X + 3.87] = .99

and then substituting 2.7 for X to get the interval (- 1.17, 6.57).
It is to be observed that there are, in fact, many possible intervals with the
same probability (with the same confidence coefficient). Thus, for example,
since
P[ -1.68 < Z < 2.70] = .95,
another 95 percent confidence interval for J1 is given by the interval (-1.35, 5.22).
This interval is inferior to the One obtained before because its length 6.57 is
greater than the length 5.88 of the interval (- .24, 5.64); it gives less precise
information about the location of J1. Any two numbers a and b such that
376 pARAMETRIC INTERVAL ESTIMATION VIII
FIGURE 1
95 percent of the area under ¢(z) lies between a and b will determine a 95 percent
confidence interval. Ordinarily one would want the confidence interval to be
as short as possible, and it is made so by making a and b as close together as
possible because the relation P[a < Z < b] = .95 gives rise to a confidence in-
terval of length ((J/Jn)(b - a). The distance b - a will be minimized for a
fixed area when ¢(a) = ¢(b), as is evident on referring to Fig. 1. If the point b
is moved a short distance to the left, the point a will need to be moved a lesser
distance to the left in order to keep the area the same; this operation decreases
the length of the interval and will continue to do so as long as ¢(b) < ¢(a).
Since ¢(z) is symmetrical about z = 0 in the present example, the minimum value
of b - a for a fixed area occurs when b = - a. Thus for x = 2.7, (- .24, 5.64)
gives the shortest 95 percent confidence interval, and (-1.17, 6.57) gives the
shortest 99 percent confidence interval for 1'.
In most problems it is not possible to construct confidence intervals which
are shortest for a given confidence coefficient. In these cases one may wish to
find a confidence interval which has the shortest expected length or is such that
the probability that the confidence interval covers a value 1'* is minimized, where
1'* =f:. 1'.
The method of finding a confidence interval that has been illustrated in the
example above is a general method. The method entails finding, if possible, a
function (the quantity Z above) of the sample and the parameter to be estimated
which has a distribution independent of the parameter and any other parameters.
Then any probability statement of the form P[a < Z < b] =}' for known a and
b, where Z is the function, will give rise to a probability statement about the
parameter that we hope can be rewritten to give a confidence interval. This
method, Or technique, is fully described in Subsec. 2.3 below. This technique
is applicable in many important problems, but in others it is not because in these
others it is either impossible to find functions of the desired fOrm or it is impos-
sible to rewrite the derived probability statements. These latter problems can
be dealt with by a more general technique to be described in Sec. 4.
The idea of interval estimation can be extended to include simultaneous
FIGURE 2
estimation of several parameters. Thus the two parameters of the normal distri-
bution may be estimated by some plane region R in the so-called parameter
space, that is, the space of all possible combinations of values of Jt and (12. A
95 percent confidence region is a region constructible from the sample such that
if samples were repeatedly drawn and a region constructed for each sample,
95 percent of those regions in a long-term relative-frequency sense would include
the true parameter point (Po, (l5)(see Fig. 2).
Confidence intervals and regions provide good illustrations of uncertain
inferences. In Eq. (2) the inference is made that the interval - .24 to 5.64 covers
the true parameter value, but that statement is not made categorically. A
measure, .05, of the uncertainty of the inference is an essential part of the
statement.
2.2 Definition of Confidence Interval

In the previous subsection we attempted to give some feel for the concept of a
confidence interval by discussing a simple example. In this subsection we define,
in general, what a confidence interval is and in the next subsection describe one
method of finding confidence intervals.
We assume that we have a random sample Xl' ... , Xn from a density
f( . ; 0) parameterized by O. Previously, in Chap. VII, we considered point
estimates of say 1:(0), some real function of O. Now we look for an interval
estimate of 1:(0).
Definition 1 Confidence interval Let Xl' ... , Xn be a random sample

from the density f(· ; 8). Let Tl = 11 (Xl' ... , X,.) and T2 = 1 2 (Xb ... , Xn)
be two statistics satisfying Tl < T2 for which P8[T1 < 1:(8) < T 2 ] = y, where
y does not depend on 8; then the random interval (Tl' T 2 ) is called a 100y
percent conjidence interval for 1:(0); Y is called the conjidence coefficient; and
Tl and T2 are called the lower and upper confidence limits, respectively,
for r(O). A value (tl' t 2 ) of the random interval (Tl' T 2 ) is also called a
100y percent confidence interval for r(O). IIII
378 PAIt.AMBTRlC INTERVAL ESTIMATION vm
We note that one or the other, but not both, of the two statistics
tl(Xb " ' , Xn) and t 2 (Xl , .,., Xn) may be constant; that is, One of the two end
points of the random interval (Tl , T 2 ) may be constant.
Definition 2 One..sided confidence interval Let Xl' ... , X,. be a random

sample from the density f( . ; 0). Let Tl = tl(Xl , ... , XJ be a statistic
for which PO[Tl < 1:(0)] = y; then Tl is called a one-sided lower corifidence
limit for 1:(0). Similarly, let T2 = t iXl , ... , Xn) be a statistic for which
Poh(O) < T2] == y; then T2 is called a one-sided upper confidence limit for
1:(0). (y does not depend On 0.) IIII
EXAMPLE 1 Let Xl' ... , Xn be a random sample from f(x; 0) = q,O.9(X).

Set Tl = tl(Xl , , .. , XJ = X - 6/~ and T2 = t 2(Xl , .. ,' XJ = X +
61 ~; then (Tl' T2 ) constitutes a random interval and is a confidence
interval for 1:(6) = Owithconfidencecoefficienty = Po[X - 61J-;i < 0 < X +
61j;i]=Po[-2 < (X - 6)/(3/~) <2]= <1>(2) - <1>(-2) =.9772 -.0228 =
.9544. Also, if a random sample of 25 observations has a sample mean
of, say, 17.5, then the interval (17.5 - 61J25, 17.5 + 61J25) is also
called a 95.44 percent confidence interval for O. '1111
Remark If a confidence interval for () has been determined, then, in

essence, a whole family of confidence intervals has been determined.
More specifically, for a given lOGy percent confidence interval estimator of
o a 100y percent confidence-interval estimator of 1:(0) can be obtained,
where 1:(') is any strictly monotone function. For example if 1:(')
is a monotone, increasing function and (Tl = tl(Xl , ... , XJ, T2 =
t 2 (Xl , ••• , Xn» is a 100y percent confidence interval for 0, then
(1:(Tl ), 1:(T2» is a 100y percent confidence interval for 1:(0) since
As was the case in point estimation, our problem is twofold: First, we need
methods of finding a confidence interval, and, second, we need criteria fOl~
comparing competing confidence intervals or for assessing the goodness of a
confidence interval. In the next subsection, we will describe One method of
finding confidence intervals and call it the pivotal-quantity method.
2 CONFIDENCE INTER VALS 379
FIGURE 3
2.3 Pivotal Quantity

As before, we assume a random sample Xl' •.. , X,. from some density f('; 0)
parameterized by O. Our object is to find a confidence~interval estimate of'r(O),
a real-valued function of O. 0 itself may be vector-valued.
Definition 3 Pivotal quantity Let Xl, ... , X,. be a random sample from
the density f(· ; 0). Let Q = ?(Xl , ... , X,.; 0); that is, let Q be a function
of Xl' ..• , X,. and O. If Q has a distribution that does not depend on 0,
then Q is defined to be a pivotal quantity. IIII
EXAMPLE 2 Let Xv ... , X,. be a random sample from f(x; 0) = <P8.9(X),

X - 0 is a pivotal quantity since X - 0 is normally distributed with mean 0
and variance 91n. Also (X - O)/(3/~) has a standard normal distribution
and, hence, is a pivotal quantity. On the other hand, XIO is not a pivotal
quantity since X10 is normally distributed with mean unity and variance
910 2 n, which depends on O. IIII
Our hope is to utilize a pivotal quantity to obtain a confidence interval.
Pivotal-quantity method If Q = ?(Xv ... , X,.; 0) is a pivotal quantity and

has a probability density function, then for any fixed 0 < y < 1 there will exist ql
and q2 depending on y such that P[ql < Q < q2] = y. Now, if for each possible
sample value (Xh .•• , X,.), ql < ?(x l , ... ,X,.; 0) < q 2 if and only if tl (Xl' ... ,xJ <
'reO) < t 2(Xl' ... , X,.) for functions tl and t 2 (not depending on 0), then (Tl' T 2)
is a 100y percent confidence interval for 'reO), where Ti = ti(Xb ••• , X,.), i = 1, 2.
Before illustrating the pivotal-quantity method with a simple example we

make several comments. First, ql and q 2 are independent of 0 since the distribu-
tion of Q is. Second, for any fixed y there are many possible pairs of numbers
ql and q2 that can be selected so that P[ql < Q < q2] = y. See Fig. 3. Different
pairs of ql and q2 will produce different tl and t 2' We should want to select
that pair of ql and q2 that will make t 1 and t 2 close together in some sense. For
instance, if t 2(Xl , ... , X,.) - tl (Xl! ... , X,.), which is the length of the confidence
380 PAR.AMETRIC INTEllVAL ESTIMATION vm
FIGURE 4
interval, is not random, then we might select that pair of ql and q2 that makes the
length of the interval smallest; or if the length of the confidence interval is
random, then we might select that pair of ql and q2 that makes the average
length of the interval smallest.
As a third and final comment, note that the essential feature of the pivotal-
quantity method is that the inequality {ql < 9(Xl, ... , Xn; 8) < q2} can be re-
written, or inverted or" pivoted," as {tl(X l , ... , xn) < -r(8) < tix l , ... , xn)} for
any possible sample value Xl' ... , x n • [This last comment indicates that
"pivotal quantity" may be a misnomer since according to our definition Q =
9(Xl , ... , Xn; 8) may be a pivotal quantity, yet it may be impossible to "pivot"
it.]
EXAMPLE 3 Let Xl' ... , Xn be a random sample from ¢o, l(X), Consider
estimating -r(8) = 8. Q = 9(Xl , ... , Xn; 8) = (X - 8)/(JlTn) has a
standard normal distribution and, hence, is a pivotal quantity. fQ(q) =
¢(q). For given I' there exist ql and q2 such that P[ql < Q < q2] = I' (in
fact, there exist many such ql and q2)' See Fig. 4.
Now {ql < (x - e)IJ1/n < q2} if and only if {x - q2Jl/n < e <
x -qlJl/n}; so (X - q2JlTn, X .:.. qlJl/n) is a 1001' percent confidence
interval for e. The length of the confidence interval is given by
(X - qlJ lin) - (X - q2#) = (q2 - ql)JlTn; so the length will be
made smallest by selecting ql and q2 so that q2 - ql is a minimum under the
restriction that I' = P[ql < Q < q2] = (q2) - (ql), and q2 - ql will be a
minimum if ql = -q2' as can be seen from Fig. 4. IIII
The steps in the pivotal-quantity method of finding a confidence interval

are two: First, find a pivotal quantity, and, second, invert it. We will comment
further on techniques for finding pivotal quantities in Sec. 4. The method is
thoroughly exploited in the next section.
3 SAMPLING FROM THE NORMAL DISTRIBUTION 381
3 SAMPLING FROM THE NORMAL DISTRIBUTION

Let Xl' ... , X It be a random sample from the normal distribution with mean Il
and variance 0'2. The first three subsections of this section are generated by.
the cases (i) confidence interval for Il only, (ii) confidence interval for 0'2 only,
and (iii) simultaneous confidence interval for Il and 0'. The fourth subsection
considers a confidence interval for the difference between two means.
3.1 Confidence Interval for the Mean

There are really two cases to consider depending on whether or not 0'2 is known.
We leave the case 0'2 known as an exercise. (The technique is given in Example 3.)
We want a confidence interval for Il when 0'2 is unknown. In our general dis-
cussion in Sec. 2 above OUr parameter was denoted by O. Here 0 = (Il, 0'), and
-r(0) = Il. We need a pivotal quantity. (X -Il)/(O'/,/;;) has a standard normal
distribution; so it is a pivotal quantity, but {ql < (x -Il)/(O'/Jn) < q2} cannot be
inverted to give {t 1 (Xl' ... , x,.) < Il < t 2(Xl' ... , x,.)} for any statistics tl and t 2 .
The problem with (X - Il)/(O'j);;) seems to be the presence of 0'. We look for a
pivotal quantity involving only Il. We know that
has a t distribution with n - I degrees of freedom. [Recall that 8 2 =

L (Xi - X)2j(n - 1).] So (X - ll)j(8j);;) has a density that is independent of
Il and 0'2; hence it is a pivotal quantity. Now one has {ql < (x - Il)/(<J/Jn) < q2}
if and only if {x - q2(<JjJ;;) < Il < x - ql(<J/);;)}, where ql and q2 are such that
P[ql < (X - 1l)J(8/J;;) < q2] = 1'; therefore, (X - q2(8/.jn), X - ql(8/Jn» is
a lOOy percent confidence interval for Il. The length of this confidence interval
is (q2 - ql)(8/Jn), which is random~ For any given sample the length will be
minimized if ql and q2 are selected so that q2 - ql is a minimum. A little
reflection will convince One that ql and q2 should be symmetrically selected about
0, Or the following argument can be advanced. We seek to minimize
subject to
(3)
382 pARAMETRIC INTERVAL ESTIMATION vm
where fr(t) is the density of the t distribution with n - 1 degrees of freedom.

Equation (3) gives q2 as a function of ql' and differentiating Eq. (3) with respect
to ql yields
To minimize L, we set dLjdql = 0; that is,
but
8 (d q 2 ) 8 (frlql) )
J~ dql - 1 = In frlq2) - 1 =0
if and only if fT(ql) = fT(q2)' which implies that ql = q2 [in which case
S:!fT(t) dt =f:. y] or ql = -q2' ql = -q2 is the desired solution, and such ql
and q2 can be readily obtained from a table of the t distribution.
3.2 Confidence Interval for the Variance

Again there are two cases depending On whether or not Il is assumed known, and
again we leave the case Il known as an exercise. We want a confidence interval
for (1'2 when Il is unknown. We need a pivotal quantity that can be inverted.
We know that
Q = L (Xl - X)2 = (n - 1)82
(1'2 (1'2
has a chi-square distribution with n - 1 degrees offreedom; hence Q is a pivotal

quantity. Also, one has
if and only if
so
2 2
(n - 1)8 , (n - 1)8 )
( q2 ql
3 SAMPLING FROM THE NORMAL DISTRIBUTION 383
Chi-square density with

n - 1 degrees of freedom
FIGURE 5
is a looy percent confidence interval for (12, where ql and q2 are given by
P[ql < Q < q2] = y. See Fig. 5.
ql and q2 are often selected so that P[Q < ql] = P[Q > q2] = (1 - y)/2.
Such a confidence interval is sometimes referred to as the equal-tails confidence
interval for (12. ql and q2 can be obtained from a table of the chi-square
distribution. Again, we might be interested in selecting ql and q2 so as to
minimize the length, say L, of the confidence interval.
(1ql q2I).
L = (n - 1)82 - - -
Letfa(q) be a chi-square density with n - 1 degrees of freedom; then differen-

tiating
with respect to ql yields
and so
dL .= (n _ 1)8 2(- 12 + q
12 d 2) = (n _ 1)8 2(- 12 + ~fa(ql)) = 0,
dql ql q2 dql ql q2fa(q2)
which implies that q;fa(ql) = q~fa(q2)' The length of the confidence interval
will be minimized if ql and q2 are selected so that
subject to
A solution for ql and q2 can be obtained by trial and error or numerical integra-
tion.
384 pAJt.AMETRIc INTBllVAL ESTIMATION vm
FIGURE 6 j - t. 97s J111/n j + t'97S J {]l/n
We might note that for any ql and q2 satisfying
is a 10Dy percent confidence interval for (1.
3.3 Simultaneous Confidence Region for the Mean and Variance

In constructing a region for the joint estimation of the mean /1 and variance (12
of a normal distribution, one might at first be inclined to use Subsec. 3.1 and 3.2
above. That is, for example, one might construct a confidence region as in
Fig. 6 by using the two relations
and
p [
en - 2
1)8
2
2
< (1 < 2
2
(n - 1)8 ]
= .95, (5)
X.975 X.025
where t.975 is the .975th quantile point of the t distribution with n - 1 degrees
of freedom and X~25 and X.~75 are the .025th quantile point and .975th quantile
point, respectively, of the chi-square distribution with n - 1 degrees of freedom.
The region displayed in Fig. 6 does indeed give a confidence region for (/1, (12),
but we do not know what its corresponding confidence coefficient is. [It is not
.95 2 since the two events given in Eqs. (4) and (5) are not independent.]
A confidence region, whose corresponding confidence coefficient can be
3 SAMPLING FROM 11IE NORMAL DISTRmunoN 385
2 (n - l)a2
---(1 = "
I 2 ql
(Il - X)I = ql(1
n
L-------~~~---------+Il
FIGURE 7 X
readily evaluated, may be set up, however, using the independence of X and S2.
Since
and
are each pivotal quantities, we may find numbers ql' q2' and qi such that
p [ -q, < ~/ln <q,] = y, (6)
and
(7)
Also, since Ql and Q2 are independent, we have the joint probability

2
X- It (n - 1)S ]
(1/ J n
P [ -ql < < ql; q2 < 2 < q2 = 1'11'2' (8)
(1
The four inequalities in Eq. (8) determine a region in the parameter space, which
is easily found by plotting its boundaries. One merely replaces the inequality
signs by equality signs and plots each of the four resulting equations as functions
of It and (12. A region such as the shaded area in Fig, 7 will result.
We might note that a confidence regiOn for (It, (1) could be obtained in
exactly the same way; the equations would be plotted as fUnctions of (1 instead of
(12, and the parabola in Fig. 7 would become a pair of straight lines given by
p. = x + ql(1/J;; intersecting at x On the It axis.

The region that we have constructed does not have minimum area for
fixed 1'1 and 1'2' Its advantage is that it is easily constructible from existing
tables and it will differ but little from the region of minimum area unless the
sample size is smalL The region of minimum area is roughly elliptic in shape
and difficult to construct.
386 pARAMETRIC INTERVAL ESTIMATION vm
3.4 Confidence Interval for Difference in Means

Let Xl' ... , X m be a random sample of size m from a normal distribution with
mean 1'1 and variance (f2, and let Yb •.• , Yn be a random sample of size n from
a normal distribution with mean 112 and variance (f2. Assume that the two
samples are independent of each other. We want a confidence interval for
112 - Ill' Y - X is normally distributed with mean 112 - 1'1 and variance
(f2jn + (fIjm. L (Xi - X)2j(f2 is chi-sQuare-distributed with m - 1 degrees of
freedom, and L (Yi - Y)2j(f2 is chi-sQuare-distributed with n - 1 degrees of
freedom; hence
L (Xi - X)2 + L (Yi - Y)2
(12 (12
is chi-sQuare-distributed with m + n - 2 degrees of freedom. Finally,
Q= [( y - X) - (1'2 - IlI)]jJ~(12-jm + (12jn

J[L (Xi - X)2 +L (Yi - Y)2]j(12(m + n - 2)
( Y - X) - (1l2 - Ill)
J(ljm + 1jn) [2: (Xi - X)2 +L (Yi - Y)2]j(m +' n - 2)
( Y - X) - (1l2 - Ill)
-
J(ljm +Ijn)(S;)
has a t distribution with m + n - 2 degrees of freedom. Thus it follows that
y = P[ - t(1 +y)/2 < Q < t(1 +Y)/2], where t(l +y)/2 is the [(1 + y)/2]th quantile point
of the t distribution with m + n - 2 degrees of freedom. S;
is an unbiased
estimator of the common variance (f2. (The subscript p can be thought of as an
abbreviation for "pooled"; S;is a pooled estimator of (12, the two samples
being pooled together.) Now
(y - x) - (1l2 - Ill)
-t(1+Y)/2< J (l/m + l/n)o; <t(1+y)/2
if and only if
(,V - X) - t (1+1){2 (: +~) d~ < /12 - /11 < (y - x) + +rH.J (~ +~)d!;

t(1
hence
(- -
(Y-X)-t(1+y)/2 J(1 1) 2 - -
m+n Sp,(Y-X)+t(1+Y)/2 J(l 1)S2)
m+~ p (9)
is a lOOy percent confidence interval for 112 - Ill'

4 METHODS OF FINDING CONFIDENCE INTERVALS 387
We assumed above that we had two independent random samples. Now

assume that (Xl' Yl ),···, (X,., Y,.) is a random sample from the bivariate normal
distribution with parameters given by III = 4'[X], Ilz = 4'[ Y], O'I = var[X),
O'~ = var[ y], and p = cov [X, Y]/ 0'10' 2 • The 0 bject is to find a confidence-
interval estimate of J.lz - Ill' Let Die = Y t - Xb i = 1, ... , n; then Db ... , D,.
are independent and identically distributed random variables with Common
normal distribution having mean J.lD = Ilz - J.ll and variance O'~ = O'i + O'i -
2pO'I 0' 2 • Pretending that Db ... , D,. is our random sample and proceeding as
in Subsec. 3.1, we obtain the following 100y percent confidence interval for
112 - Ill:
(-
D- t(1+'1)/2
Jl: (D, - D)2 D-
n(n - 1)
,
r
+ t(I+'1)/2 (D, - D)2)
n(n - 1)
(10)
where t(1 +'1)1 2 is the [(1 + y)/2]th quantile point of the t distribution with n - 1
degrees of freedom. The above obtained confidence interval for 112 - III is
often referred to as the confidence interval for the difference in means for paired
observations. The ith X observation is paired with the ith Yobservation.
4 METHODS OF FINDING CONFIDENCE INTERVALS

In this section we will discuss two methods of obtaining a confidence interval.
A third method will be presented in Chap. IX .
./ 4.1 Pivotal-quantity Method

We described the pivotal-quantity method of finding confidence intervals in
Sub sec. 2.3, but we left unanswered the question of whether or not a pivotal
quantity would actually exist for a given problem. The following remark gives
a partial answer to this question.
Remark If Xb ... , Xn is a random sample from f(· ; 8), for which the
corresponding cumulative distribution fUnction F(x; 8) is continuous
in x, then, by the probability integral transform, F(Xi; 8) has a uniform
distribution over the interval (0, 1). Hence -log F(Xi ; 8) has the den-
sity e-ul(o,oo)(u) since P[-logF(Xi; 8);:::u] =P[logF(Xj ; 8)< -u]=
P[F(X,; 8) se- ] = e- for u > O. Finally -
U U
L
log F(X.· 8) has a gamma
distribution with parameters n and I; that is, "
388 pARAMETRIC INTERVAL ESTIMATION VIII
p[ -log q2 < - ,t,iogF(X,; 9) < -log q,]

-IOgq. 1
=
I -logqz
__
r(n)
zit le- z dz
= p[q, < lJ.F(X,; 9) < q2] (11)
So
It n
TI F(Xi ; 8), or - L log F(X j ; 8),
i= 1 i"'" 1
is a pivotal quantity. II/I

The remark shows that any tim~ that we sample from a population having
a continuous cumulative distribution function, a pivotal quantity exists. Note
that as of yet we have no assurance that the pivotal quantity exhibited by the
remark can be utilized to find a confidence interval. If, however, F(x; 8) is
n
monotone in 8 for each x, then TI F(Xi; 8) is also monotone in 8 for each
i"" 1
Xl' ... , Xn , and such monotonicity allows one to find a confidence interval for 8.
n
We see from Fig. 8 that ql < TI F(xj; 8) < q2 if and only if 11 (X 1 , "', xn) < 8 <
i 1
1 2(Xl, ... , x n), where II and 12 are defined as indicated.
EXAMPLE 4 Let Xl, ... , Xn be a random sample from the density r(x; 8) =
8xO- 1I(o.1)(x); then F(x; 8) = XOI(O.llx) + I[t,oo)(x), If ql and q2 are
selected [see Eq. (11)] so that
l' = p[q, < laF(X,; 9) < q2]

= p[q, < ,Ij,xr < q2]
= P[IOg q, < 9 log Iv'XI < log q2]
= P[-log q2 < -810g IT Xi< -log ql]
i= 1
log q 2 log q 1 ]
=P It <8< It '
[ 10glU Xi 10gJ:\ Xi
1~----------------------~--
FIGURE 8
then
is a 100)' percent confidence interval for 8. IIII
We conclude this subsection with two further comments regarding the

existence of pivotal quantities. First, if 8 is a location parameter, then Xi - 8
has a distribution independent of 8 by definition and, hence, is a pivotal quantity
as are a variety of other random quantities, including L Xi - n8, Yj - 8, Yl +
Yn - 28, etc. Second, if 8 is a scale parameter, then by definition Xd8 is
distributed independently of 8 and, hence, is a pivotal quantity as are L Xd8,
YJl8, etc.
4.2 Statistical Method

t
As usual, we assume that we have a random sample Xl' ... , Xn from the density
f(' ; ( 0), We further assume that the parameter 80 is real and that the param-
eter space e is some interval. (In this SUbsection, we will let 80 denote the
true parameter value.) We seek an interval estimate of 80 itself. Let T =
t(Xb ... , Xn) be some statistic. The statistic T can be select~d in several ways.
For instance, if a sufficient statistic (unidimensional) exists, then T could be
taken to be a sufficient statistic; or if no sufficient statistic exists, Tcould be taken
to be a point estimator, possibly the maximum-likelihood estimator, of 80 ,
The actual choice of T might depend on the ease with which the operations that
need to be performed to obtain the confidence interval can be performed. One
of those operations will be the determination of the density of T.
390 pAJlAMBTRlc INTERVAL ESTIMATION VBI
~ h1 (6)
Area = Pl
t
FIGURE 9
Let fT(t; 0) denote the density of T. We will proceed as though T is a

continuous random variable; although the technique will also work for T as a
discrete random variable. We can define two fUnctions) say hl(O) and h2 (0)) as
follows:
and (12)
where Pl and P2 are two fixed numbers satisfying 0 < Pl) 0 < P2 ) and Pl + P2 < 1.
See Fig. 9.
hl(O) and h2 (0) can be plotted as functions ofO. We will assume that both
hl ( .) and h2 (·) are strictly monotone) and for our sketch we will assume that
they are monotone) increasing functions. We know that hl (0) < h2 (0). See
Fig. 10.
Let to denote an observed value of T; that is, to = t(Xl, ... , xn) for an
observed random sample Xb ..• , Xn • Plot the value of to on the vertical axis in
Fig. 10, and then find Vl and V 2 as indicated. For any possible value of to, a
corresponding Vl and V2 can be obtained, so Vl and V2 are functions of to; denote
these by Vl = Vl(tO) and V2 = V2(t O)' The interval (Vb V 2) will turn out to be a
100(1 - Pl - P2) percent confidence interval for 00 , To argue that this is so,
let us repeat Fig. 10 as Fig. 11 and add to it. (Figure 10 indicates the method of
finding the confidence interval.)
to ---------
FIGURE 10
FIGURE 11
We see from Fig. 11 that hi (0 0) < to = t(X I , ... ~ Xn) < h 2(00) if and only if
VI = VI(Xb ••• ~ xn) < 00 < V2 = V2(XI~ ... ~ xn) for any possible observed sample
(Xj ~ ... ~ x n). But by definition of hi (.) and h2(')~
P8o[h l (0 0) < t(Xb ... ~ Xn) < h 2(00)] = 1 - PI - P2;
so
P80 [VI(Xh ' ••• ~ Xn) < 00 < V2(Xh ... ~ Xn)] = 1 - PI - P2;
that is~ as stated~ (VI~ V 2) is a 100(1 - PI - P2) percent confidence interval for
00 ~ where Vi = Vl(X b ... , Xn) for i = 1, 2.
We might note that the above procedure would work even if hi (.) and
h2 ( .) were not monotone fUnctions, only then we would obtain a confidence
region (often in the form of a set of intervals) instead of a confidence interval.
EXAMPLE 5 Let Xl' ... , Xn be a random sample from the density f(x; ( 0) =
(1/(Jo)I(0.8o)(x), We want a confidence interval for 00 , Y n =
max [Xl, ... ~ Xn] is known to be a sufficient statistic; it is also the maximum-
likelihood estimator of(Jo' We will use Yn as our statistic Tthat appears
in the above discussion; then
fT(t; 8) = n (~r-l ~ I(0.6)(t).

For given PI and P2 , find hi (0) and h2(0). PI = S~1(8) ntn-1o- n dt implies
that Ji1 (8) t n- l dt = onpl/n, which in tum implies [hl(o)]n/n = onpl/n~ or
finally hi CO) = Opt/no Similarly, P2 = S:1(8) ntn-Io- n dt implies that
8" - [h 2(o)]n = 8"P2 or h 2(0) = 0(1 - P2)I/n. See Fig. 12, which is Fig. 10
for the example at hand.
392 PAIlAMETlUC INTERVAL ESTIMATION VID
-=~----------------~----6
FIGURE 12 Vl
For observed to = max [Xl' ••• ,Xn ], VI is such that h 2 (Vl) = to;
that is, hlv 1) = Vl(1 - P2)I/n = to or VI = to(l - P2)-lln. Similarly,
V 2 = top-lIn. So a 100(1 - PI - pz) percent confidence interval for 00 is
given by (Y,.(1 - P2)-I/n, Y nPl l !n). We could worry about selecting PI
and Pl so that the confidence interval is shortest subject to the restriction
that 1 - PI - Pl = y. The length of the confidence interval is
L = Y~[pilin - (1 - P2) -lIn]
and so the length will be shortest if PI and Pl are picked so as to minimize

pill,. - (1 - P2)-1/1I subject to 1 - PI - P2 = Y and 0 <PI + Pl < 1,
which is accomplished by picking P2 = 0 and PI = 1 - y.
We might note that Yn/O is a pivotal quantity and a confidence
interval for 0 can be obtained more easily using the pivotal-quantity
method. / // /
We observe in the example above, and in general for that matter, that
hI (0) and hiO) are really not needed. For a given observed value to =
I(Xb ••• , xn) of the statistic T, we need to find VI = 'Vt(Xb ... , xJ) and V 1 =
'Vl(Xb ••• , x 1I ). V2 can be found by solving for 0 in the equation \
\\
Pt = J
hl(tJ)=to
-00
/,.(t; 8) dt;
\
(13)
V1 is the solution; VI can be found by solving for 0 in the equation '
Pl = f 00
h;z(6) =to
fT(t; 8) dt; (14)
VI is the solution.
We mentioned at the outset that the method would work for discrete
random variables as well as for continuous random variables. Then the inte-
grals in Eqs. (12) to (14) would need to be replaced by summations. Two
5 LARGE-SAMPLE CONFIDENCE INTERVALS 393
popular discrete density functions are the Bernoulli and Poisson. One could be
interested in confidence-interval estimates of the parameters in each. In
Example 6 to follow we will consider the Bernoulli density function; the Poisson
case is left as an exerCIse.
EXAMPLE 6 Let Xh •.• , Xn be a random sample from the Bernoulli density;

that is, P[X = 1] = 00 = 1 - P[X = 0]. We know that T = L Xi is a
sufficient statistic; furthermore, T has a binomial distribution; that is,
P[T = t1 = (~)8W - (}o)"-' for t= 0, 1, ... , n. We want a confidence-
interval estimate for 00 , Suppose we observe T = to (necessarily an inte-
ger). According to Eqs. (13) and (14) we need to solve for 0 in each of
the equations
and
To actually solve these equations, a table of the binomial distribution is

useful. If Pt .0509, P2 = .0159, n = 20, and if T = 4 is observed then
the 93.33 percent confidence interval (.05, .40) for 00 is obtained. 11I1
S LARGE-SAMPLE CONFIDENCE INTERVALS

We have seen in our studies of point estimation that it is sometimes possible to
find a sequence of estimators, say Tn = tn(Xt , ..• , Xn), of 0 in a density f(' ; 0)
that are asymptotically normally distributed about 0; that is, Tn is approximately
normally distributed with mean 0 and variance, say, 0';(0), where 0';(0) indicates
that the variance is a function of 0 (since it will ordinarily depend on 0) and the
sample size n. In particular, we have seen in Sec. 9 of Chap. VII that for large
samples the maximum-likelihood estimator, say Gn = 8n(Xt , ..• , X n ), for a
parameter 0 in a density f(·; 0) is approximately normally distributed about 0
under rather general conditions. The large-sample variance of the maximum-
likelihood estimator was seen to be, say,
0' 2(0) - 1 _ -1
.. 2 2 - o-g-f-(x-;0:::-:-::)]'
n10 [{(0/00) log f (X; O)} ] - -nS'":""o-:":[(--:a2;-1a::-::O-::')-1 (15)
394 p.A.llAMln"l';UC INTBIlVAL ESTIMATION VIII
When such a sequence of asymptotically normally distributed estimators {Tn} of

o exists, it is sometimes possible to obtain approximate confidence intervals
quite easily. (Tn - 6)/0',.(0) can be treated as an approximate pivotal quantity,
and, therefore, for large sample size a confidence interval with an approximate
confidence coefficient y may be determined by converting the inequalities
T -0
-z < ;n(O) < z, (16)
where z = Z(1 +1)/2 is defined by fl)(z(l +y)/2) = (1 + y)/2 or fl)(z) - fl)( -z) = y.
The above described method will always work to find a large-sample confidence
interval provided the inequality -z < (Tn - O)IO'n(O) < z can be inverted.
EXAMPLE 7 Let Xb ... , Xn be a random sample from the density f(x; 0) =

Oe-ox/(o. «)(x). We know (see Example 51 in Chap. VII) that the maxi-
mum-likelihood estimator of 0, which is 1/Xn , has an asymptotic normal
distribution with mean 0 and variance equal to
1
----------::- - -.
m9'9[{(81 80) 10gf(X; n
Therefore,
y,;:;;p [ -z<
1/Xn - 0
fii2i:. <z
]
V 021n
-ZO 1 ZO]
=p [ v1n<Xn- O <vIn
=p [ l/Xn < II
u < l(Xn ] ,
1 + zlvln 1 - ~/Jn
\
\
and hence
is a large-sample confidence interval for 0 with an approximate confidence

coefficient y, where z is given by fl)(z) - fl)( -z) = y. 1111
EXAMPLE 8 Consider sampling from the Bernoulli distribution with param-

eter 0 = P[X = 1] = 1 - P[X = 0]. The maximum-likelihood estimator
5 LARGE-8AMPLE CONFIDENCE INTERVALS 395
of 0 is Q = X n, and it has variance u;(O) = 0(1 - O)ln. An approximate

100y percent confidence interval for 0 is obtained by converting the in-
equalities in
p [ -z< 0-0 <z] :::::::y

J 0(1 - O)ln
to get
p[2nQ + Z2 - zJ4nQ + Z2 - 4nQ2

2(n + Z2)
2nQ + Z2 + z)4nQ + Z2 - 4nQ2] ""
< (J < 2(n + Z2) "" Y . (17)
These expressions for the limits may be simplified since in deriving the
large-sample distribution certain terms containing the factor l/.j;i are
neglected; that is, the asymptotic normal distribution is correct only to
within error terms of size a constant times II);;. We may therefore
neglect terms of this order in the limits in Eq. (17) without appreciably
affecting the accuracy of the approximation. This means simply that we
may omit all the Z2 terms in Eq. (17) because they always occur added to a
term with factor n and will be negligible relative to n when n is large to
within the degree of approximation that we are assuming. Thus Eq. (17)
may be rewritten as
p [e- - )(;)(1 n
Z
-Q)
< 0 < e~ + Z
J0(l -(;»]
n
::::::: y. (18)
In particular,
[-
P E> - 1.96~
1~(1 n-Q) < (J < °+ 1.96
JQ(l - Q)]
n : : : : .95
gives an approximate 95 percent confidence interval for 0 for large

samples. IIII
We may observe that Eq. (18) is just the expression that would have been
obtained had Q been substituted for 0 in u;(O). The substitution would imply
that
A- 0
)~(l - (;)/n
396 PAIlAMETRIC INTERVAL ESTIMATION VIII
is approximately normally distributed with mean 0 and unit variance. It is,

in fact, true in general that in the asymptotic normal distribution of a maximum-
likelihood estimator ~ the variance (j;(0) may be replaced by its estimator (j;(~)
without appreciably affecting the accuracy of the approximation. We shall not
prove this fact but shall use it because it greatly simplifies the conversion of the
inequalities that is required to get the large-sample confidence intervals. For
instance,
readily converts to give

P[~ - z(jn(~) < 0 < ~ + z(jn(~)] ~ y, (19)
where ~ is the asymptotically normally distributed maximum-likelihood esti-

mator of 0, (jn(~) in this expression is the maximum-likelihood estimator of
(jn(O) (which is the large-sample standard deviation of ~), and z is given by
(z) - <1>( - z) = y.
We noted in Sec. 9 of Chap. VII that under regularity conditions the joint
distribution of the maximum-likelihood estimators of the components of a
k-dimensional parameter is asymptotically normally distributed. Although we
will not so argue, such a result could be used to obtain a large-sample confidence
regIOn.
The large-sample confidence intervals presented in this section have an
optimum property which we shall point out but not prove. Recall that in the
earlier sections, particularly Sec. 3, of this chapter we were concerned with finding
the shortest interval for a given probability. Loosely speaking, an ana10gous
optimum property of large-sample confidence intervals based on maximum-
likelihood estimators is this: Large-sample confidence intervals based on the
maximum-likelihood estimator will be shorter, on the average, than intervals deter-
mined by any other estimator.
6 BAYESIAN INTERVAL ESTIMATES

In Sec. 7 of Chap. VII we examined what is called Bayes estimation. There
we assumed that a random sample, say Xl' ... , X n , from some density f(' ; 0) =
f('IO) was available, where the form of the functionf( '1') was known and the
fixed value of 0 was unknown. We further assumed that the unknown fixed
value of 0 was the value of a random variable e with known density, denoted by
ge(') and called the prior density of e. We then used this additional knowledge
6 BAYESIAN INTERVAL ESTIMATES 397
of a known prior density to define the posterior distribution of ®, and from this
posterior distribution we defined the posterior Bayes point estimator of O. In
this section we use this same posterior distribution of ® to arrive at an inter1(ai
estimator of O.
If 1(' 10) is the density sampled from and ge(') is the prior density of ®,
then the posterior density of® given (Xl' ... , Xn) = (Xl'" . , xn) is [recall Eq. (19)
of Chap. VII]
[Qf(x. 10) ]9,,(11)

f",x, =xo ...•• x.= ••elll x" ... , xJ = J[Qfexd 0)]9,,(0) dO: (20)
For fixed /', any interval,. say (tl' t 2), satisfying
is defined to be a l00/, percent Bayesian interval estimate of O. In practice, one

would naturally pick those tl and t2 satisfying Eq. (21) for which t2 - tl is smallest.
Note that ti = ti~Xb .•. , X,.); that is, t{ is some function of the observations
Xi) ••• , X n •
EXAMPLE 9 Let Xl' ... , Xn be a random sample from the normal density
with mean 0 and variance 1. Assume that ® has a normal density with
mean Xo and variance 1. Consider estimating O. We saw in Example 44
of Chap. vn that the posterior distribution of ® is normal with mean
n
Lo xi/en + 1) and variance lj(n + 1). We seek tl and t2 satisfying
f fejx .... xI. ... ,xn ... xJO I

t2-
/' = Xh ., ., xn) dO
t1
= «»(t2 - 'tXi/(n + 1») _«»(t l - 'tXi/(n + 1»). (22)

Jl/(n + 1) Jl/(n + 1)
If z is such that «»(z) - «»( - z) = /" then
n
t2 =
~ Xi
+z
J--1---- and
n+l n+l
398 PARAMETRIC INTERVAL ESTIMATION VIII
gives the shortest 1001' percent Bayesian interval estimate of Note that e.
the corresponding 1001' percent confidence-interva1 estimate of is given e
by (t xJn - zJlIn, t xJn + zJi/n). The only difference in the results
of the two methods for this example is that the sample size seems to
increase by I and the apparent" additional observation" is the mean of
the assumed prior normal distribution. / // /
PROBLEMS
1 Let X be a single observation from the density
I(x; 6) = (JxfJ-1I(o.1)(X),
where (J > O.
(a) Find a pivotal quantity, and Use it to find a confidence-interval estimator of (J.
(b) Show that (Y/2, Y) is a confidence interval for (J. Find its confidence coef-
ficient. Also, find a better confidence interval for (J. Define Y = -l/log X.
2 Let Xl, "', Xn be a random sample from N«(J, 8), (J > O. Give an example of a
pivotal quantity, and Use it to obtain a confidence-interval estimator of (J.
3 Suppose that Tl is a 100y percent lower confidence limit for T«(J) and T2 is a 100y
percent upper confidence limit for T(6). Further assume that PfJ(Tl < T 2] = 1.
Find a 100(2y - 1) percent confidence interval for T(8). (Assume y > ~.)
4 Let Xl, ...• Xn denote a random sample from I(x; 6) = l(fJ-~. fJ+~)(x). Let
Y1 < ... < Y,. be the corresponding ordered sample. Show that (Yl~ Yn) is a
confidence interval for (J. Find its confidence coefficient.
5 Let Xl, "', X,. be a random sample from/(x; 8) = (Je-fJxl(o. OO)(x).
(a) Find a 100y percent con'fidence interval for the mean of the population.
(b) Do the same for the variance of the population.
(e) What is the probability that these intervals COver the true mean and true
variance, simultaneously?
(d) Find a confidence-interval estimator of e - fJ = P(X > I].
(e) Find a pivotal quantity based only on YJ, and use it to find a confidence-
interval estimator of (J. (Yl = min (Xl, . ", X,.].)
6 X is a single observation from (Je-fJxl(o. OO)(x), where (J > O.
(a) (X, 2X) is a confidence interval for l/(J. What is its confidence coefficient?
(b) Find another confidence interval for 1/(J that has the same coefficient but
smaller expected length.
7 Let Xl, X2 denote a random sample of size 2 from N«(J, 1). Let Yl < Y 2 be the
corresponding ordered sample.
(a) Determine yinP(Yl < (J < Y 2 ] = y. Find the expected length oftheinterval
(Y1 , Y 2 ).
(b) Find that confidence-interval estimator for 8 using X - (J as a pivotal quantity
that has a confidence coefficient y, and compare the length with the expected
length in part (a).
PROBLEMS 399
8 Consider random sampling from a normal distribution with mean p. and variance
a2•
(a) Derive a confidence interval estimator of p. when a 2 is known.
(b) Derive a confidence interva1 estimator of a 2 when p. is known.
9 Find a 90 percent confidence interval for the mean of a normal distribution with
a = 3 given the sample (3.3, -.3, -.6, -.9). What would be the confidence
interva1 if a were unknown?
10 The breaking strengths in pounds of five specimens of manila rope of diameter
3
1 6 inch were found to be 660, 460, 540, 580, and 550.
(a) Estimate the mean breaking strength by a 95 percent confidence interval

assuming normaJity.
(b) Estimate the point at which only 5 percent of such specimens would be
expected to break.
(c) Estimate a 1 by a 90 percent confidence interval; a1so a.
(d) Plot an 81 percent confidence region for thejoint estimation of p. and a 1 ; for
p. and a.
11 A sample was drawn from each of five populations assumed to be normal with the
same variance. The va1ues of (n - 1)~)2 = L (X, - %)2 and n, the sample size,
were
82 : 40 m W 42 ~
n: 6 4 3 7 8
Find 98 percent confidence limits for the common variance.
12 Develop a method for estimating the ratio of variances of two norma1 populations
by a confidence interval (1"'-
13 What is the probability that the length of a t confidence interva1 for p. when
sampling from a norma1 distribution win be less than a for samples of size 20?
14 In sampling from a norma1 population compare the average length of the two
confidence intervals for the mean p. when (a) a is known and (b) a is unknown.
15 Show that the length and the variance of the length of the t confidence interval
for p. when sampling from a normal population approach 0 with increasing sample
size.
16 In sampling from a normal population with both p. and a unknown, how large a
sample must be drawn-to make the probability .95 that a 90 percent confidence
interval for p. will have length less than a15?
17 Show that the length of the confidence interval for a of a norma1 population
approaches 0 with increasing sample size.
18 To test two promising new lines of hybrid com under norma1 farming conditions,
a seed company selected eight farms at random in Iowa and planted both lines in
experimental plots on each farm. The yields (converted to bushels per acre) for
the eight locations were
Line A: 86 87 56 93 84 93 75 79
Line B: 80 79 58 91 77 82 74 66
Assuming that the two yieldS are jointly normally distributed, estimate the difference
between the mean yields by a 95 percent confidence interval.
400 PARAMETRIC INTERVAL ESTIMATION VIU
19 X., ... , X,. is a random sample from (l/6)X(1-6)/6/ro . J](x), where 0 O. Find a
l00y percent confidence interval for O. Find its expected length. Find the
limiting expected length of your confidence interval. Find n such that
P[length (50] p for fixed (5 and p. (You may use the central-limit theorem.)
20 Develop a method for estimating the parameter of the Poisson distribution by a
confidence interval.
21 Find a good l00y percent confidence interval for 0 when sampling fromf(x; (J) =
1(6-1:.6+1:)(x).
22 Find a gOod tOOy percent confidence interval for 0 when sampling fromf(x; 0) =
(2X/02)/(O.6)(X), where 0 > O.
23 One head and two tails resulted when a coin was tossed three times. Find a
90 percent confidence interval for the probability of a head.
24 Suppose that 175 heads and 225 tails resulted from 400 tosses of a coin. Find a
90 percent confidence interval for the probability of a head. Find a 99 percent
confidence interval. Does this appear to be a true coin?
25 Let Xl, ... , X" be a random sample fromf(x; 6) = f(x; p., 0") = cPp,.02(X). Define
T«(J) by f:6) cPp,.a 2 (x) dx = a(a is fixed). Recall what the UMVUE of T(O) is. Find
a l00y percent confidence interva1 for T(O). (If you cannot find an exact lOOy per-
cent confidence interval, find an approximate one).
26 Let Xl, ... , Xn be a random sample from f(x; (J) = cP6.I(X). Assume that the
prior distribution of@ is N(lLo,~),lLoandO"~known. Find a lOOypercentBayesian
interval estimator of 0, and compare it with the corresponding confidence interval.
27 Let Xl, "', XII be a random sample from f(x 16) = OX'-I 1(0. H(X), where 0> O.
Assume that the prior distribution of @ is given by
1
De(O) = r(r) ){Or-I e -A6[(o. CC)«(J),
where r and A are known. Find a 95 percent Bayesian interval estimator of O.

*28 Let X denote the life in hours ofa radioactive particle. Suppose Xhas a density
f(x; 0) = Oe-'X[{o. oolx).
A random sample of n particles is put under observation, but the experiment is to

stop when the kth particle has expired; i.e., it is intended not to wait until all the
particles have ceased activity but only until k of them (k fixed in advance) have
done so. The data consist of the k measurements Yt , •• " Yk and n - k measure-
ments known only to exceed Y", where Y is the lifetime of the ith particle to
j
expire. Find the maximum-likelihood estimator of the mean lifetime I/O. Also
find a confidence-interval estimator of I/O.
IX
TESTS OF HYPOTHESES

There are two major areas of statistical inference: the estimation of parameters
and the testing of hypotheses. We shall study the second of these two areas in
this chapter. Our aim will be to develop general methods for testing hypotheses
and to apply those methods to some common problems. The methods developed
will be of further use in later chapters.
In experimental research, the object is sometimes merely to estimate param-
eters. Thus one may wish to estimate the yield of a new hybrid line of corn.
But more often the ultimate purpose will involve some use of the estimate.
One may wish, for example, to compare the yield of the new line with that of a
standard line and perhaps recommend that the new line replace the standard
line if it appears superior. This is a common situation in research. One may
wish to determine whether a new method of sealing light bulbs will increase
the life of the bulbs, whether a new germicide is more effective in treating a
certain infection than a standard germicide, whether one method of preserving
foods is better than another insofar as retention of vitamins is concerned,
and so on.
402 TESTS OF HYPOTHESES IX
Using the light-bulb example as an illustration, let us suppose that the

average life of bulbs made under a standard manufacturing procedure is 1400
hours. It is desired to test a new procedure for manufacturing the bulbs. The
statistical model here is this: We are dealing with two populations of light bul bs:
those made by the standard process and those made by the proposed process.
We know (from numerous past investigations) that the mean of the first popula-
tion is about 1400. The question is whether the mean of the second population
is greater than or less than 1400. Traditionally, to answer this type of question,
we set up the hypothesis that one mean is greater than the other mean. Then,
on the basis of a sample from the population of the proposed process we shall
either accept or reject the hypothesis.
F or our example, we formulate the hypothesis that the proposed process
is no better than the standard process. Generally we hope that the hypothesis
will be rejected. To test the hypothesis, a number of bulbs are made by the new
process and their lives measured. Suppose that the mean of this sample of
observations is 1550 hours. The indication is that the new process is better, but
suppose that the estimate of the standard deviation of the mean rtl In is 125
(n being the sample size). Then a 95 percent confidence interval for the mean of
the second population (assuming normality) is roughly 1300 to 1800 hours. The
sample mean 1550 could very easily have come from a popUlation with mean 1400.
We have no strong grounds for rejecting the hypothesis. If, on the other hand,
rtl In were 25, then we could very confidently reject the hypothesis and pronounce
the proposed manufacturing process to be superior.
The testing of hypotheses is seen to be closely related to the problem of
estimation. It will be instructive, however, to develop the theory of testing
independently of the theory of estimation, at least in the beginning.
In order to conveniently talk a bout testing of hypotheses, we need to intro-
duce some language and notation and give some definitions. As was the case
when we studied estimation, we will assume that we can obtain a random sample
Xl' ... , Xn from some density f(· ; lJ). A statistical hypothesis will be a hypoth-
esis about the distribution of the population.
Definition 1 Statistical hypothesis A statistical hypothesis is an asser-

tion or conjecture about the distribution of one or more random variables.
If the statistical hypothesis completely specifies the distribution, then it is
called simple; otherwise, it is called composite. IIII
Notation To denote a statistical hypothesis, we will use a script capital
J'f followed by a colon that in turn is followed by the assertion that
specifies the hypothesis. IIII
1 INTRODUCTION AND SUMMARY 403
EXAMPLE 1 Let Xl' ... , Xn be a random sample from I(x; fJ) = CPo, 2S(X),
The statistical hypothesis that the mean of this normal population is less
than or equal to 17 is denoted as follows: Yf: fJ < 17. Such a hypothesis
is composite; it does not completely specify the distribution. On the
other hand, the hypothesis Yf: fJ = 17 is simple since it completely specifies
the distribution. IIII
Definition 2 Test of a statistical hypothesis . A test of a statistical hy-

pothesis Yf is a rule or procedure for deciding whether to reject Yf. IIII
Notation Let us use a capital upsilon 1 to denote a test. IIII
EXAMPLE 2 Let Xl' ... , Xn be a random sample from I(x; fJ) = CPo, 2S(X),
Consider Yf: fJ ,:5;; 17. One possible test 1 is as follows: Reject Yf if and
only if X > 17 + 51Jn. /III
A test can be either randomized or nonrandomized. The test 1 given in

Example 2 above is an example of a nonrandomized test. Another possible
test, say 1', of Yf in Example 2 is the following: Toss a coin, and reject Yf if
and only if a head appears. Such is an example of a randomized test. Although
we will make little use of randomized tests in this book, we do include their
definition. Definitions of both nonrandomized and randomized tests follow.
Notation As in previous chapters, we let X denote the sample space of

observations, or the potential data set; that is, X = {(Xl,··., X n ): (Xl"'" Xn)
is a possible value of (Xl' ... , X n)}· IIII
Definition 3 Nonrandomized test and critical region Let a test 1 of a

statistical hypothesis Yf be defined as follows: Reject Yf if and only if
(Xl, ... , xn) E Cy , where Cy is a subset of X; then 1 is called a nonrandom-
ized test, and Cy is called the critical region of the test 1. IIII
EXAMPLE 3 Let Xl, "" Xn be a random sample from I(x; fJ) = CPo, 25(X),
X is euclidean n space. Consider Yf: fJ < 17 and the test 1: Reject Yf if
and on)y if x > 17 + 51 J~. Then 1 is nonrandomized, and Cy =
{(Xl' ... , X n ): x > 17 + 5IJ~ }. IIII
Remark A nonrandomized test Y of Jf is a decomposition of l: into C r

and Cr such that if(x1, ... , xn) E C r , Jf is rejected. So a nonrandomized
test is specified by its corresponding critical region. fill
Definition 4 Randomized test A test Y of a hypothesis Jf is defined to
be a randomized test if Y is defined by the function t/Jy(X1' "', xn) =
P[Jf is rejected I (Xl' " ' , xn) is observed]. The function t/Jy(', "', .) is
called the critical/unction of the test Y. //11
The actual performing of a nonrandomized test Y of Jf is straightforward;
one observes a random sample, say Xl' ••• , Xn ,checks to see whether the observed
sample falls in the critical region, and rejects Jf when it does. On the other
hand, to perform a randomized test Y of Jf, one first observes the random sample,
say Xl' ••• , X n , then evaluates t/Jy(x b •. " x n), and finally observes the result of
some auxiliary Bernoulli trial that has t/Jy(x 1, .. " xn) as its probability of success,
and if the Bernoulli trial resu.lts in a success, then Jf is rejected. Since the
performance of the auxiliary Bernoulli trial is extraneous to the actual testing
problem, that is, it does not depend on the data Xl' •• " Xn of the experiment, one
might reasonably wonder why its result should be the deciding factor in accepting
or rejecting the hypothesis. It is for this reason that r rriized tests are not
often employed in practice; when they are, usuall e sample space l: is decom-
posed into three sets, one where the hypothesis is accepted, another where the
hypothesis is rejected, and the third where" randomization" takes place. This
third region is often the boundary between the acceptance and rejection region
and/or a region where it is not easy to decide whether to accept or reject. The
following example may help in understanding randomized tests.
EXAMPLE 4 Let Xb "', X 10 be a random sample of size 10 fromf(x; lJ) =

lJx(I - lJ)l- x for X = 0 or 1. Suppose we want to test the hypothesis
10
Jf: lJ < t. A possible test Y is the following: Reject Jf if I Xi > 5, accept
1
10
Jf if I Xi < 5, and decide between rejecting and accepting by tossing a
1
10
fair coin if L Xi = 5. (The tossing of the coin is the auxiliary Bernoulli
1
trial.) Such a test Y partitions l: into three regions, say A, B, and C,
where
A = {(X" "', X,o): ~ X; < 5},

B = {(X" ... , X,o): ~ XI = 5},
and
c {(x" ... , XlO): ~ Xi> 5}.

The critical function of test T is given by
1 if (Xb ... , XI0) E C
t/lr(Xl' .•• , XI0) = 1/2 if (Xb ••• , X 10) E B
o IIII
Remark We saw that a nonrandomized test was specified by its critical
region. Likewise, a randomized test is specified by its critical function.
In fact any function t{t(., .,., .) with domain l! and counterdomain the
interval [0, I] is a possible critical function and defines a randomized
~ M
The following remark shows that a nonrandomized test is a particular case
of a randomized test.
Remark If test T has a critical function defined by
1/IY{X" ... , xJ = {~ if (Xl' ••• , x n)'

.
otherwlse
E C r) -- 1
C
r
(
Xl"'" Xn ,
)
then T is a nonrandomized test with a critical region Cr. IIII

As we mentioned earlier, we will not make extensive use of randomized
tests. Theorem I below requires their use; other than that, their only use will
be in obtaining tests of exact size (see Definition 7), and then only for sampling
from discrete distributions.
In many hypotheses-testing problems two hypotheses are discussed: The
first, the hypothesis being tested, is called the null hypothesis, denoted by :7t 0 ,
and the second is called the alternative hypothesis, denoted by :7t l' The think-
ing is that if the null hypothesis is false, then the alternative hypothesis is true,
and vice versa. We often say that :7t 0 is tested against, or versus, :7tI' If the
null hypothesis :7t 0 is not rejected, we say that :7t0 is accepted, With this kind
of thinking, two types of errors can be made.
Definition 5 Types of error and size of error Rejection of :7t0 when it is

true is called a Type I error, and acceptance of :7t0 when it is false is called
a Type II error. The size of a Type I error is defined to be the probability
that a Type I error is made, and similarly the size of a Type II error is the
probability that a Type II error is made. IIII
If the distribution from which the sample was obtained is parameterized by

0, where 0 E f), then associated with any test is a power function, defined as in
Definition 6.
Definition 6 Power function Let Y be a test of the null hypothesis .7t 0 +

The power function of the test Y, denoted by 1ty(O), is defined to be the
probability that .7t 0 is rejected when the distribution from which the
sample was obtained was parameterized by O. IIII
The power function will play the same role in hypothesis testing that mean-
squared error played in estimation. It will usually be our standard in assessing
the goodness of a test or in comparing two competing tests. An ideal power
function, of course, is a function that is 0 for those 0 corresponding to the null
hypothesis and is unity for those 0 corresponding to the alternative hypothesis.
The idea is that you do not want to reject .7t 0 if .7t 0 is true and you do want to
reject .7t 0 when .7t 0 is false.
Remark 1ty(O) = P8[reject .7t 0]' where 0 is the true value of the param-
eter. If Y is a nonrandomized test, then 1ty(O) = P8 [(X1 , ••• , Xn) E Cy],
where Cy is the critical region associated with test Y. If Y is a randomized
test with critical function t/ly (. , ... , +), then
1ty(O) = P 8[reject.7t 0]
= f··· f P[reject .7t 0 I Xl' + •• , X n]!Xl • .... Xn(X 1 , ••• , Xn ; 0) Ii

1= 1
dXi
= cf8 [t/ly(X I, ++., X n)] + ,
The argument is similar for discrete ran~/variables+ IIII
EXAMPLE 5 Let Xb +", Xn be a random sample fromf(x; 0) = cP8.2S(X),
Consider .7t0: 0< 17 and the test Y: Reject if and only if X > 17 + 51 J~.
-
1ty(0) = P 8 [ X > 17 + J~ =
5 ]
I - 
(17 + 51J~ -
51J n
0) .
For n = 25, 1ty(O) is sketched in Fig. 1.
1.0
.5
~----~----~----~~~~-----8
FIGURE 1 0
The power function is useful in telling how good a particular test is.
In this example, if 0 is greater than about 20, the test Y is almost certain
to reject :J'f 0' as it should. And if 0 is less than about 16, the test Y is
almost certain not to reject :J'f 0, as it should. On the other hand, if
17 < 0 < 18 (so :J'f 0 is false), the test Y has less than half a chance of
rejecting :J'f 0 • II!!
Definition 7 Size of test Let Y be a test of the hypothesis :J'f0: 0 E eo,
where eo c e; that is, eo is a subset of the parameter space S. The size
of the test Y of :J'f0 is defined to be sup [nr(O)]. The size of the test for
8e~o
a nonrandomized test is also referred to as the size of the critical region.
IIII
Remark Many writers use the terms "significance level" and "size of
test" interchangeably. We, however, will avoid use of the term" signi-
ficance level/' intending to reserve its use for tests of significance, a type of
statistical inference that is closely related to hypothesis testing. Tests of
significance will not be considered in this book; the interested reader is
referred to Ref. [37]. IIII
EXAMPLE 6 Let Xl' "" Xn be a random sample from f(x; 0) = 4>0. 2S(x).
Consider the :J'f 0 = 0 < 17 and the test Y: Reject :J'f0 if X > 17 + 51Jn.
~o = {O: 0 < 17} and the size of the test Y is sup [nr(O)]
_ Oe~o
= sup
Osl7
[I - (J)(17 +515~nn -6)] = 1 - (J)(l) ~ .159. IIII
In our study of point estimation, we found that for certain considerations

we could restrict attention to estimators that were functions of sufficient statistics
only. The same is true for testing hypotheses when the power function is used
as a basis of comparing tests, as the following theorem shows.
Theorem 1 If Xl' ... , X 1I is a random sample from/ex; e), where e E 9,

and Sl = 61(X1, ... , X1I ) , • • • , S,. = 6,.(X1, ... , X1I ) is a set of sufficient sta-
tistics, then for any test i with critical function 0/1' there exists a test, say
if, and corresponding critical function, say 0/1', depending only on the set
of sufficient statistics which satisfies 1t1 (e) = 1t1'(e) for all e E 8.
PROOF Define o/l'(Sb "" S,.) = 8[0/1(X1, ... , X1I ) lSI = Sl, ... ,
S,. = S,.]; then 0/1' is a critical function. Furthermore, 1t1'(e) =
4 9 [o/y,(Sl' ... , S,.)] = 8 9 [8[0/1(X1, ... , X1I ) ISb . ", S,.]] = 8 9 [I/II(X1, ... ,
X 1I) ] = 1ty(e). IIII
The theorem shows that given any test, another test which depends only on
a set of sufficient statistics can be found, and this new test has a power function
identical to the power function of the original test. So, in our search for good
tests we need only look among tests that depend on sufficient statistics.
We have introduced some of the language of testing in the above. The
problem of testing is like estimation in the sense that it is twofold: First, a
method of finding a test is needed, and, second, some criteria for comparing
competing tests are desirable. Although we will be interested in both aspects
of the problem, we will not discuss them in that order. First we will consider,
in Sec. 2, the problem of testing a simple null hypothesis against a simple alter-
native. Two approaches will be assumed. The first will use the power function
as a basis for setting goodness criteria for tests, and the second will use a loss
function. The Neyman-Pearson lemma is stated and proved. It will turn out
that all those tests, which are best in some sense, will be of the form of a simple
likelihood-ratio, which is defined.
Tests of composite hypotheses will be discussed in Sec. 3. The section
will commence, in Subsec. 3.1, with a discussion of the generalized likelihood-
ratio principle and the generalized likelihood-ratio test. ~. his principle plays a
central role in testing, just as maximum likelihood played central role in es-
timation. It is a technique for arriving at a test that in gen ral will be a good
test, just as maximum likelihood led to an estimator that in general was quite a
good estimator. For a book of the level of this book, it is probably the most
important concept in testing. The notion of uniformly most powerful tests
will be introduced in Subsec. 3.2, and several methods that are sometimes use-
ful in finding such tests will be presented. Unbiasedness and invariance in es-
timation are two methods of restricting the class of estimators wi th the hope
of finding a best estimator within the restricted class. These two concepts play
essentially the same role in testing; they are methods of restricting the totality of
2 SIMPLE HYPOTHESIS VERSUS SIMPLE ALTERNATIVE 409
possible tests with the hope offinding a best test within the restricted class. We
will discuss only unbiasedness, and it only briefly in Subsec. 3.3. Subsection
3.4 will summarize several methods of finding tests of composite hypotheses.
Section 4 will be devoted to consideration of various hypotheses and
tests that arise in sampling from a normal distribution. Section 5 will consider
tests that fall within a category of tests generally labeled chi-square tests.
Included will be the asymptotic distribution of the generalized likelihood-ratio,
goodness-of-fit tests, tests of the equality of two or more distributions, and tests
of independence in contingency tables. Section 6 will give the promised dis-
cussion of the connection between tests of hypotheses and interval estimation.
The chapter will end with an introduction to sequential tests of hypotheses in
Sec. 7.
The reader will note that our discussion of tests of hypotheses is not as
thorough as that of estimation. Both testing and estimation will be used in
later chapters, especially in Chap. X. Also, a number of the nonparametric
techniques that will be presented in Chap. XI will be tests of hypotheses.
We stated at the beginning of this section that testing of hypotheses is one
major area of statistical inference. A type of statistical inference that is closely
related (in fact so closely related that many writers do not make a distinction) to
hypothesis testing is that of significance testing. The concept of significance
testing has important use in applied problems; however, we will not consider it
in this book. The interested reader is referred to Ref. [37].
2 SIMPLE HYPOTHESIS VERSUS

SIMPLE ALTERN A TIVE
2.1 Introduction
In this section we consider testing a simple null hypothesis against a simple
alternative hypothesis. This case is actually not very useful in applied statistics,
but it will serve the purpose of introducing us to the theory of testing hypotheses.
We assume that we have a sample that came from one of two completely
specified distributions. Our object is to determine which one. More precisely,
assume that a random sample Xl"'" Xn came from the density fo(x) or fl (x)
and we want to test :Yf 0 : Xi distributed asJo('), abbreviated Xi '" fo{'), versus
:Yf1: Xi '" It ( .). If we had only one observation Xl and foe .) and It (.) were
as in Fig. 2, one might quite rationally decide that the observation came from
fo(') if fo(x 1) > Ji(Xl) and, conversely, decide that the observation came from
It (.) if fi (Xl) > fo(xl)' This simple intuitive method of obtaining a test can be
expanded into a family of tests that, as we shall see, will contain some good tests.
FIGURE 2
Definition 8 Simple likelihood-ratio test Let Xl' ... , Xn be a random

sample from either fo{') or It (.). A test Y of :Yf 0: Xi""" fo{ .) versus
:Yf1 : Xi""" fi(') is defined to be a simple likelihoodMratio test ifY is defined
by
Reject :Yfo if A < k,
Accept :Yf 0 if A > k,
Either accept :Yf 0, reject :Yf 0, or randomize if A = k, (1)
where
and k is a nonnegative constant. [L j = Lixl' ... , xn) is the likelihood

function for sampling from the density Ii' ).] 1111
For each different k we have a different test. For a fixed k the test says to
reject :Yf 0 if the ratio of likelihoods is small; that is, reject :Yf0 if it is more likely
(Ll is large compared to Lo) that the sample came from fi ( .) than from lo{ . ).
Such a test certainly has intuitive appeal. In fact, one might suspect that an
optimum test will have to be the form of a simple likeliho?dMr.ati~.~
Optimality of a test of a simple hypothesis versus asimpLe-atrernative can
be approached in two ways. One way, using the power of the test to set goodness
criteria, is discussed in SUbsec. 2.2, and the other way, using a loss function and a
decisionMtheoretical approach, is considered in Subsec. 2.3.
2.2 Most Powerful Test

Let Xl' ..• , Xn be a random sample from the density fo{') or the density 11 (').
Let us write lo(x) = f(x; ( 0 ) and It (x) = f(x; ( 1 ); then Xl' ... , Xn is a random
sample from one or the other member of the parametric family {/(x; 0): = 00
orO = Ol}' e = {Oo, 0l} is a parameter space with only two points in it. 00 and
°
(}l are known. We want to test :tt 0 : (} = (}o versus :ttl : (} = (}l' Corresponding
to any test 1 of :tt 0 versus :ttl is its power function 1'ly(8). A good test is a test
for which 1ty«(}o) = P[reject :tt 0 l:tt 0 is true] is small (ideally 0) and 1'ly(81) =
P[reject -*'ol:tfo is false] is large (ideally unity). One might reasonably use the
two values 1ty«(Jo) and 1'lY«(}I) to set up criteria for defining a best test. 1'ly«(}o) =
size of Type I error, and I - 1'lY«(}I) = P[accept:tt 0 I:tt 0 is false] = size of Type II
error; so our goodness criterion might concern making the two error sizes small.
F or example, one might define as best that test which has the smallest sum of the
error sizes. Another method of defining a best test, made precise in the following
definition, is to fix the size of the Type I error and to minimize the size of the
Type II error.
Definition 9 Most powerful test A test 1* of :tt 0: (} = (}o versus

:ttl: (} = (}l is defined to be a most powerful test of size (1 (0 < (1 < 1) if and
only if:
(i) 1'ly-«(}o) = (1. (2)
(ii) 1'lY-«(}I) > 1'lY«(}I) for any other test 1 for which 1'ly«(}o) < (1. (3)
1/11
A test 1* is most powerful of size (1 if it has size (1 and if among all other
tests of size (1 or less it has the largest power. Or a test 1* is most powerful of
size (1 if it has the size of its Type I error equal to (1 and has smallest size Type II
error among all other tests with size of Type I error (1 or less.
The justification for fixing the size of the Type I error to be (1 (usually
small and often taken as ,.05 or .01) seems to arise from those testing situations
where the two hypotheses are formulated in such a way that one type of error is
more serious than the other. The hypotheses are stated so that the Type I error
is the more serious, and hence one wants to be certain that it is small.
The following theorem is useful in finding a most powerful test of size (1.
The statement of the theorem as given here, as well as the proof, considers only
nonrandomized tests. We might note that the statement and proof of the
theorem can be altered to include all randomized tests.
Theorem 2 Neyman-Pearson lemma Let Xl, ... , Xn be a random

sample from f(x; (}), where (} is one of the two known values (}o or (}l'
and let 0 < (1 < 1 be fixed.
Let k* be a positive constant and C* be a subset of ~ which satisfy:
(i) P8o[(XI, ... , Xn) E C*] = (1. (4)
..) ~ _ L(Oo; Xl' ... , XJ -

_ LO
(11 / .. -
°
L( 1; Xb ... , XJ Ll
- k* (5)
and )" > k* if (Xb ... , xn) E C*.
Then, the test y* corresponding to the critical region C* is a most powerful

test of size ct of Jf 0: 0=°
0 versus Jf 1 : °°
= 1 , [Recall that L) =
n f(Xi; OJ) for j = 0 or 1 and C* is the complement of
n
L(OJ; Xl' ... , Xn) =
i= 1
C*; that is, C* = l - C*.]
PROOF Suppose that k* and C* satisfying conditions (i) and (ii)

exist. If there is no other test of size ct or less, then y* is automatically
most powerful. Let Y be another test of size ct or less, and let C be its
corresponding critical region. We have P8o[(X1, ... , Xn) E C] < ct. We
must show that 1ly.(Ol) > 1ly(Ol) to complete the proof. {For any subset
R of X, let us abbreviate J.~. J (,V, I(x,; 8J) dx,] as JR L J for j = 0, 1.
Our notation indicates that fo( .) and It ( . ) are probability density func-
tions. The same proof holds for discrete density functions.} Showing
that 1ly.(01 ) > 1ly(Ol) is equivalent to showing that Ic. Ll > Ic L 1 • See
Fig. 3.
Now Ic. L l - Ic L l = Ic.c Ll - Icc. L l > (llk*) Ic*cLo- (llk*) Icc*L o
sinceLl > Lo/k* on C* (hence also on C*C) and Ll < Lo/k*, or -Ll
-Lo/k*, on C* (hence also on CC*). But (l/k*) (Jc*c Lo - Icc. Lo)
= (llk*){Jc.cLo + Ic*cLo - Ic.cLo - Icc.Lo) = (llk*)(Ic*L o - IcLo) =
(l/k*)(ct - size of test Y) > 0; so Ic. Ll - Ic Ll > 0, as was to be shown.
IIII
We comment that k* and C* satisfying conditions (i) and (ii) do not always
exist and then the theorem, as stated, would not give a most powerful size-ct
test. However, whenever fo( .) and It ( .) are probability density functions, a k*
and C* will exist. Although the theorem does not explicitly say how to find k*
and C*, implicitly it does since the form of the test, that is, the critical region, is
given by Eq. (5). In practice, even though k* and C* do exist, often it is not
necessary to find them. Instead the inequality A S; k* for (Xh ... , xn) E C* is
manipulated into an equivalent inequality that is easier to work with, and the
actual test is then expressed in terms of the new inequality. The following
examples should help clarify the above.
FIGURE 3
8X
EXAMPLE 7 Let Xl"'" Xli bea random sample from/ex; (J) = (Je- /(0. oo)(x),
where (J = 80 or 8 = 81 , (Jo and (JI are known fixed numbers, and for
concreteness we assume that (JI > (Jo' We want to test .Yf0: (J = (Jo versus
.Yfl : (J = (JI' Now Lo = 03 exp (-(Jo LXi), L1 = 0i exp (-(JI I Xi), and
according to the NeymanftPearson lemma the most powerful test will have
the form: Reject .Yf0 if )" < k* or if «(JO/(JI)'J exp [ - «(Jo - (JI) I < k*, xa
which is equivalent to .
LXi < 0, ~ 00 log. [ (::) nk*] = k' (say),
where k' is just a constant. The inequality A =:;; k* has been simplified and
expressed as the equivalent inequality L Xi k'. Condition (i) is ex =
PfJo[reject .Yfol = PfJo[ I Xi =:;; k']. We know that Xi has a gamma L
distribution with parameters n and 0; hence
" X·l -< k'] =

p fJo [ f..J f -r(n)I
0
k'
(In
0
X'- l e- x8o dx = ex ,
an equation in k', from which k' can be determined; and the most powerful
test of size ex of .Yfo: (J = (Jo versus .Yfl : (J = (JI' (JI > (Jo is this: Reject .Yfo
if LXi k', where k' is the exth quantile point of the gamma distribution
with parameters nand (Jo· 11//
EXAMPLE 8 Let Xl' ... , Xn be a random sample from /(x; (J) =
(Jx(1 - (J)l-X I{O.I}(X), where (J = (Jo or (J = (J1' We want to test .Yf0: (J = 00
versus .Yf1 : 0 = (Jl' where, say, (Jo < 01 , Lo = (J~Xi(1 - (Jo)n-r.xi, Ll =
(Jr Xt (1 Odn-r.xt, and so A < k* if and only if
(J~Xi(1 - (JO)'J-r.Xi/(J r
Xi(l - (Jl)n-r. Xi < k*, ('-
if and only if
or if and only if I Xi > k', where k' is a constant. {Note that

loge [0 0 0 - ( 1 )/0 - ( 0 )01 ] < O.} So a most powerful test would be of the
form: Reject :tf 0 if I Xi is large. For definiteness, let us take 0 0 = !,
01 = i, and n = 10. We must find k' so that
(X=P80=1/4[reject:tfO]=P80=1/4[IXi:C:k'] = I
10 (10)
. (1)
- i (3) 10 - i
- .
i =k' 1 4 4
If (X = .0197, then k' = 6, and if (X = .0781, then k' = 5. For (X = .05, there
is no critical region C* and constant k* of the form given in the Neyman-
Pearson lemma. In this example our random variables are discrete, and
for discrete random variables it is not possible to find a k* and C* satisfying
conditions (i) and (ii) for an arbitrary fixed 0 < (X < 1. In practice, one is
usually content to change the size of the test to some (X for which a test
given by the Neyman-Pearson lemma can be found. We might note,
however, that a most powerful test of size (X does exist. The test would be
a randomized test. For the example at hand, if we take (X = .05, the
randomized test with critical function
1
.0584
o
is the most powerful test of size (X = .05. IIII
In closing this subsection, we note that a most powerful test of size (x,
given by the Neyman-Pearson lemma, is necessarily a simple likelihood-ratio
test.
2.3 Loss Function

As in the last subsection, we assume that we have a random sample Xl, ... , Xn
from one or the other of the two completely known densities fo( .) = f( . ; ( 0 ) and
it ( .) = f(· ; ( 1 ), On the basis of the observed sample we have to decide from
which of the two densities the sample came; that is, we test :tf 0: 0 = 00 versus
:tf1: 0 = 01, We can make one of two decisions, say do or d 1, where d j is the
decision that fi') is the density from which the sample came, j = 0, 1. We
assume that a loss function is available.
Definition 10 Loss function In testing :tf 0: 0 = 0 0 versus :tf1 : 0 = Ob

define t(d i ; OJ) = loss incurred when decision d i is made and OJ is the true
SIMPLE HYPOTHESIS VERSUS SIMPLE ALTERNATIVE 415
parameter value for i = 0, 1 andj = 0, 1, where dE is the decision of deciding

that hypothesis :Yf i is correct. We will adopt the convention that
ted;; (Ji) = 0 for i = 0, 1 and t(di ; OJ) > 0 for i #=j. IIII
The values of the loss function are the amounts that are lost if decision dE
is made when (J.J was correct. With our convention, nothing is lost if the right
decision is made, and a positive amount is lost if a wrong decision is made. If
we think in terms of a (nonrandomized) test Y having a critical region Cr , then
decision d1 is made if the observed sample (Xl' ... , XII) belongs to C r, and
decision do is made if (Xl' •.. , xn) belongs to Cr· A test can be thought of as a
decisionfunction since for a given observed sample the test tells us which decision
to make. We do not consider the problem of selecting an appropriate loss
function, and hence we will always assume that an appropriate loss function has
been prescribed. Note that if decision d1 is made when 00 is correct, then a
Type I error is made, and if decision do is made when 01 is correct, then a Type IT
error is made.
In comparing two tests we naturally prefer that test which has smaller loss,
and among all tests we would prefer that test which has smallest loss. However,
seldom will there exist one test that has smallest loss for both possible decisions
and for both 00 and 01 , This motivates the defining of average loss, and by
continuing to borrow language from decision theory we define the risk function
of a test.
Definition 11 Risk function For a random sample X h " ' , Xn from

f(' ; 00 ) or f(· ; 01), let Y be a test of :Yf 0: 0 = 00 versus :Yf 1: (J = 01 having
a critical region Cr. For a given loss function t(·; .), the risk function
of 1, denoted by !!lr(O), is defined to be the expected loss; that is
Remark
aly(O) = t(d1 ; 0)Pe[(X1 , XJ E Cr ] + t(do ; 0)Pe[(X1 ,
••• , ••• , Xn) E (;r]
= t(d1 ; O)nr(O) + t(do; 0)[1 - nr(O)]; (6)
that is, the risk function is a linear function of the power function; the
coefficients in the linear function are determined by the values of the loss
function. Since 0 assumes only two values, !!lr(O) can take on only two
values, which are
and al y{Ol) = t(do ; 01)[1 - n 1 (01)]' (7)
IIII
Our object is to select that test which has smallest risk, but, unfortunately,
such a test will seldom exist. The difficulty is that the risk takes on two values,
and a test that minimizes both of these values simultaneously over all possible
tests does not exist except in rare situations. (See Prob. 3.) Not being able
to find a test with smallest risk, we resort to another less desirable criterion, that
of minimizing the largest value of the risk function.
Definition 12 Minimax test A test 1m of ff 0: fJ = (}o versus ff 1: (} = fJ1

is defined to be minimax if and only if
max [&frm(fJO), &f rm (OI)] < max [&fr(fJo), &fr «(}I)]
for any other test 1. IIII
The following theorem is sometimes useful in finding a minimax test. (As
with Theorem 2, we state the theorem and proof in terms of nonrandomized
tests.)
Theorem 3 For a random sample Xl' ... , X" from f(·; ( 0 ) orf(· ; ( 1 ),
°°
consider testing ff 0: fJ = 00 versus ff 1 : = 1 , If a test Ymhas a critical
region given by Cm = {(Xl' ... , xJ: A <km }, where k m is a positive constant
such that &frm(fJ O) = &f rm«(}I), then 1m is minimax. Recall that
A = LolLI = " f(XI; (}o)]/[ I1
[I1 " f(Xi; ( 1)],
1= I i I
PROOF We will assume that f(·; fJ o) and f(·; ( 1 ) are probability

density functions. The proof for discrete density functions is similar.
Let 1 be any other test with a critical region Cr which satisfies
&fr(Oo) < &frm(Oo). [Note that jf &fr«(}o) > &frm«(}o), then Y would not
even be a candidate for minimax.] We have &fr(fJ o) < &frm«(}o),
t(d1 ; 0o)1tr«(}o) < t(dl ; Oo)1tr m(fJ O) , or 1tr«(}o) 1trm(fJO); that is, 1 has size
less than or equal to that of 1m' But, by the Neyman-Pearson lemma,
Ym is the most powerful test of size 1tl'm(OO); hence 1tr(OI) < 1trm(OI), 1 -
1tr(OI) > 1 - 1trm«(}I), t(do ; 0d[l - 1tr(fJd] > t(do ; fJ1)[l - 1trm(fJI)], or
&fr(OI) > &frm(OI); so, we have max [,!{lrm«(}o), &frm «(}I)] = &f rm(fJ1 ) <
,!{lY{(}I) < max [&fr«(}o), &fr(fJd]; that is, 1m is minimax. IIII
EXAMPLE9 Let Xl, ... , X" be a random sample fromf(x; fJ) = fJe- 8xI(o, oo)(x),
° °°
For 01 > 0 , test ffo: 0= 00 versus ff l : = 1 , In Example 7, we found
the most powerful size-tX test. We seek now to find the minimax test for
a loss function given by t(d1 ; fJ o) = a and t(do ; fJ1) = h. According to
Theorem 3, the minimax test 1m is given by Cm = {(Xl' ... , x,,): A k m }
Before leaving minimax tests we make two comments: First, if 10(') and
11 ( .) are discrete density functions, then there may not exist a k m such that
r71 ym(()o) = r71 Ym (()l) unless randomized tests are allowed; and, second, a minimax
test as given in Theorem 3 was a simple likelihood-ratio test.
In the above we assumed that Xl, ... , Xn was a random sample from
1(' ; (J), where () = ()o or ()l' and for each (), 1(' ; ()) is completely known. We
also assumed that we had an appropriate loss function. Now, we further
assume that ()o and ()l are the possible values of a random variable e and that
we know the distribution of e, which is called the prior distribution, just as in
our considerations of Bayes estimation. e is discrete, taking on only two
values ()o and ()l; so the prior distribution of e is completely given by, say, g,
where 9 = p[e = ()l] = 1 - p[e = ()o]. We mentioned above that, in general,
a test with smallest risk function for both arguments does not exist. Now that
we have a prior distribution for the two arguments of the risk function, we can
define an average risk and seek that test with smallest average risk.
Definition 13 Bayes test A test Y9 of .Yf0: () = ()o versus .Yf1 : () = ()l is

defined to be a Bayes test with respect to the prior distribution given by
9 = pre = ()l] if and only if
(1 - g)PAYg(()o) + gPAYg(()l) < (l - g)PAy(()o) + gPAY(()l) (8)

for any other test Y. fill
To find a Bayes test we seek a critical region Cg that minimizes
(1 - g)PAy(()o) + gPAY(()l) = (1 - g)t(dl ; ()o)1Ly(()o) + gt(do ; ()l)[l - 1Lr(()l)]
= (l - g)t(dl ; ()o) $e Lo + gt(do ; ()l) $e Ll
as a function of the region C. Now
(1 - g)t(dl ; (Jo) Ie Lo + gt(do ; ()l) $e Ll
= gt(do ; (Jl) + $d(l - g)t(dl ; ()o)Lo - gt(do ; ()l)Ll ], (9)
which is minimized if C is defined to be all (Xl' ... , xn) for which the last inte N
grand of Eq. (9) is negative; that is
Cg = {(Xb .. " x,,): (l - g)t(d1 ; Oo)Lo - gt(do ; (1)L 1 < O}. (10)
We have proved the following theorem.
Theorem 4 The Bayes test Yg of .Yf'0: 0 = 00 versus .Yf'1: 0 = 01 with

respect to a prior distribution. given by g = p[e = 01 ] has a critical region
defined by
(II)
IIII
We note that once again a good test, in this case a Bayes test, turns out to
be a simple likelihood-ratio test. The exact form of the Bayes test is given by
Eq. (11).
EXAMPLE 10 Let Xl, "', X" be a random sample from f(x; 0) =

Oe -8x1(0, colx). Test.Yf' 0: 0 = 00 versus .Yf'1: (J = (Jl' The critical region
of a Bayes test is given by
_ ( . 00 exp ( - 00 I Xi) gt(do ; (J1) }

C -l(Xl' .. " xn)· (Ji exp (- 01 I
g Xi) < (I - g)t(d ; (Jo)
1
f I g(Ji t(do ; ( 1 ) }
= 1(Xh ... , Xn): I Xi < (Jl - (Jo loge (1 - g)Oo t(d ;( )
1 0
for 01 >0 0 , IIII
3 COMPOSITE HYPOTHESES
In Sec. 2 above we considered testing a simple hypothesis against a simple
alternative. We return now to the more general hypotheses-testing problem,
that of testing composite hypotheses. We assume that we have a random sample
fromf(x; 0), (J E 9, and we want to test .Yf'o: 0 E 9 0 versus.Yf'l: (J E 9f, where
eo c e, e 1 c 8, and eo and e1 are disjoint. Usually 9 1 = 9 - 9 0 , We
begin by djscussing a general method of constructing a test.
3 COMPOSITE HYPOTHESES 419
3.1 Generalized Likelihood-ratio Test

For a random sample Xl' .," X" from a densityf(x; 0), 0 E 9, we seek a test of
eo
Jff o : (J E versus Jff1 : 0 E 9 1 = 9 - 9 0 ,
Definition 14 Generalized likelihood-ratio Let L(O; x h ••. , x,,)

be the likelihood function for a sample Xl' ... , Xn having joint density
fx." ... Xn (Xl' ... , x,,; 0) where 0 E 9. The generalized likelihood-ratio,
denoted by 2 or An , is defined to be
(12)
1111
Note that 2 is a function of Xl' ... , x,,, namely 2(Xl' ... , x,,). When the
observations are replaced by their corresponding random variables Xl, ... , X"'
then we write A for 2; that is, A = 2(Xl' ... , X,,). A is a function of the random
variables Xl, ... , X" and is itself a random variable. In fact, A is a statistic
since it does not depend on unknown parameters.
Several further notes follow: (i) Although we used the same symbol ), to
denote the simple likelihood-ratio, the generalized likelihood-ratio does not
e
reduce to the simple likelihood-ratio for = {Oo, Ol}' (ii) 2 given by Eq. (12)
necessarily satisfies 0 < 2 < I; 2 >0 since we have a ratio of nonnegative
quantities, and )" < I since the supremum taken in the denominator is over a
larger set of parameter values than that in the numerator; hence the denominator
cannot be smaller than the numerator. (iii) The parameter 0 can be vector-
valued. (iv) The denominator of A is the likelihood function evaluated at the
maximum-likelihood estimator. (v) In our considerations of the generalized
likelihood*ratio, often the sample X h . . . , X" will be a random sample from a
density f(x; 0) where (J E 9.
The values 2 of the statistic A are used to formulate a test of Jff 0 = 0 E 9 0
versus Jff1 : 0 E 9 - 9 0 by employing the generalized likelihood-ratio test prin-
ciple, which states that Jff 0 is to be rejected if and only if 2 < 20 , where 20 is
some fixed constant satisfying 0 < 20 < I. (The constant 20 is often specified
by fixing the size of the test.) A is the test statistic. The generalized likelihood-
ratio test makes good intuitive sense since 2 will tend to be small when Jff 0 is
not true, since then the denominator of 2 tends to be larger than the numerator.
In general, a generalized likelihood-ratio test will be a good test; although there
are examples where the generalized likelihood-ratio test makes a poor showing
420 TSSTS OF HYPOTHESES IX
compared to other tests. One possible drawback of the test is that it iSlIOtn:ez
times difficult to find sup L(O; Xb •.. , x n); another is that it can be difficult to'
find the distribution of A which is required to evaluate the power of the test.
EXAMPLE 11 Let Xl' ... , Xn be a random sample from f(x; 0) =

Oe -8x/(0. oolx), where t:i = {o; 0 > O}. Test JIP 0: 0 < 00 versus JlPI : > 00 • °
sUEL[(O;xH""Xn)] = sup [On exp (-OIxi)] = ( "n )n e- n,
8eft 8>0 ~Xi
and
sup (L(O; Xl' ••. , Xn)] = sup (on exp ( - 0 L Xl)]

(I e.eo 0<8.s(lo
n
(Inx)"e-. if --::;;;, 00
LXi
-
n
~ exp (-0 0 LXi) if
r>Oo'
Xi
Hence
n
1 if - - < 00
LXi -
A= (13)
O~ exp ( - 00 LXi) n
(niI xi)ne-n
if
r >°
Xi
0 ,
If 0 < Ao < 1, then a generalized likelihood-ratio test is given by the follow-

ing: Reject JIP 0 if A < Ao , or
reject JIP 0 if "'\'n > 00 and 00 In Xl)n exp ( -0 Ix; + n) < Ao,
~Xi
( 0
(14)
or reject JIP 0 if Oox < 1 and (0 0 x)n e- n(8ox-l) < Ao. Write y = 00 x, and
note that yne-n(.v-I) has a maximum for y = 1. Hence y < 1, and
yne n(y-l) < Ao if and only if y < k, where k is a constant satisfying
o < k < 1. See Fig. 4.
We see that a generalized likelihood-ratio test reduces to the follow-
ing:
Reject JlPo if and only if Box < k, where 0 < k < 1; (15)
O~~--+---~l----------~==~-Y
FIGURE 4
that is, reject :Yf 0 if x is less than some fraction of 1/00 , If that gen-
eralized likelihood-ratio test having size a is desired, k is obtained as the
solution to the equation
"k 1
f
a = P9o[OoX < k] = P9o [Oo L Xi < nk] = 0 r(n) u"- e- duo
l U
(Note that P9[OoX < k] P9o [OoX < k] for 0 <0 0 ,) III/
We note that in the above example the first form of the test, as given in
Eq. (14), is rather messy, yet after some manipulation the test reduces to a very
simple form as given in Eq. (15). Such a pattern often appears in dealing with
generalized likelihood-ratio tests-their first form is often foreboding, yet the
tests often simplify into some nice form. We will observe this again in Sec. 4
below when we consider tests concerning sampling from the normal distribution.
We might note, by considering the factorization criterion, that a generalized
likelihood-ratio test must necessarily depend only on minimal sufficient statistics.
In Sec. 5 below, a large-sample distribution of the generalized likelihood-
ratio is given. This will provide us with a method of obtaining tests with
approximate size a.
3.2 Uniformly Most Powerful Tests

In Subsec. 3.1 above we exhibited a method of obtaining a family of tests of
:Yf 0: () E E)o versus :lf l : () E E) - E)o. We now define one optimum property
that such a test may possess. It is defined in terms of the power function 1'Cy(O)
and the size of the test.
Definition 15 Uniformly most powerful test A test y* of :Yf0: 0 E 8 0

versus :Ye l : 0 E e - E)o is defined to be a uniformly most powerful size-a
test if and only if:
(i) sup 1'Cy.(O) = a.
gei!o
(ii) 1'Cy.(ll) > 1'Cy(O) for all () E E) - 8 0 and for any test Y with size
less than or equal to a. /11/
A test y* is uniformly most powerful of size (X if it has size (X and if among

all tests of size less than or equal to (X it has the largest power for all alternative
values of 8. The adverb" uniformly" refers to " all" alternative 8 values. A
uniformly most powerful test does not exist for all testing problems, but when
one does exist, we can see that it is quite a nice test since among all tests of size (X
or less it has the greatest chance of rejecting .Yf0 whenever it should.
EXAMPLE 12 Let Xl, ... , Xn be a random sample from I(x; 8) =

8e- 9x/(o.ooi x ), where 8 = {8: 8 >8 o}. Find a uniformly mOrtowerful
test of .lito: 8 = 80 versus .Yf1 : 8 > 80 , For fixed 81 > 80 , we termined
in Example 7 that the most powerful test of .Yf0: 8 = 80 versus 1: 8 = 81
was given by the following: Reject .lito if I Xi < k', where k' was given as
a solution to the equation
k' 1
(X = fo r(n) 8~ ~-1 e- 90x dx.
Such a test was given by the Neyman-Pearson lemma. Note that the test
in no way depends on 81 except that 81 > 80 ; hence, we would get the same
most powerful test for any 81 > 80 , and thus the test is actually uniformly
most powerful! /1/ /
The above example provides us with an example of a situation where a

uniformly most powerful test can be obtained using the Neyman-Pearson lemma.
That same technique can be used to find uniformly most powerful tests in more
general situations, such as those given in Theorems 5 and 6 below, which are
given without proof.
Theorem 5 Let Xl' ... , Xn be a random sample from the density I(x; 8),
8 E 8, where (3 IS some interval. Assume that I(x; 8) =
n
a(8)b(x) exp [c(8)d(x)], and set t(Xb ... , xn) = L d(xJ
1 ..
(i) If c(8) is a monotone, increasing function in 8 and if there exists

k* such that P9o[t(X1, ... , Xn) > k*] = (x, then the test y* with a critical
region C* = {(Xb ... , xn): t(Xl' . ", xn) > k*} is a uniformly most power-
ful size-a test of .Yf0: 8 <8 0 versus .Yf1 : 8 > 80 or of .Yf0: 8 = 80
versus .Yf1: 8 > 80 •
Oi) If c(8) is a monotone, decreasing function in 8 and if there
exists k* such that P 9o [t(X1 , ••• , Xn) < k*] = (x, then the test y* with a
critical region C* = {(Xl' ' , " x n): t(x l , " " Xn) < k*} is a uniformly most
powerful size-a test of .Yt0: 0 00 versus J'l'l: 0 > 00 or of J'l' 0: 0 = 00
versus J'l'l: 0 > 00 , IIII
EXAMPLE 13 Let Xl' , .. , Xn be a random sample from f(x; 0) =

Oe- 8xI(0, oolx), where e = {O: 0 > O}. Test Jr 0: 0 < 00 '1ersus Jrl : 0 > 00 ,
f(x; 0) = 01(0. oo)(x) exp (-Ox) = a(O)b(x) exp [c(O)d(x)]; so t(xI , .• , ' xn)
n
=I Xb and c(O) = -0. c(O) is a monotone, decreasing function; so
I
by (ii) of Theorem 5 a uniformly most powerful test is given by the follow-
ing: Reject Jr0 if and only if I Xi < k*, where k* is given as a solution to
a = P 80 ['\'
'-'
X I < k*] = f -r(n) on un- e- 8oU du
k·
0
1
0
I
•
IIII
Definition 16 Monotone likelihood-ratio A family of densities {f(x; 0):

oE 9, 9 an interval} is said to have a monotone likelihood-ratio
if there exists a statistic, say T = t(XI , ... , Xn), such that the ratio
L(O'; Xl, ... , xn)/L(O"; X., .. ,' Xn) is either a nonincreasing function of
t(x l , ... , xn) for every 0' < Olf or a nondecreasingfunction of t(x l , ••. , xn)
for every Of < 0", IIII
Note that in the term "monotone likelihood-ratio" the likelihood-ratio is not
a generalized likelihood-ratio; it is a ratio of two likelihood functions.
EXAMPLE 14 If {f(x; 0): 0 E 8} = {Oe- 8xI(0. oo)(x): 0 > O}, then
L(O'; x., ... , xn) = (O')n exp (- 0' I X,) = (o,)n exp [ _ (0' _ 0") I x.]
L(O"; x., ... , xn) (Ollr exp ( - 0" I Xi) 0" I
which is a monotone, increasing function in I Xi' IIII
EXAMPLE 15 If{f(x; 0): OE9} = {(110)/(0.8)(x): 0 >O}, then

n
/
L(O'; x., ... , xn) (1/8')"iU 1(0,8 )(xi) =(l/0't1(0,8/)(Yn)
L(O"; Xl, .•. , Xn) (l/O"yn 1(0,8") (Xi) (l/0,,)nl(0,8")(Yn)
i= I
for 0 < Yn < 0'
for 0'::;; Yn < 0",

424 TBSTS OF HYPOTHESES IX
which is a monotone, nonincreasing function in Yn = max [x h ... , xnl.

[Note that YlI cannot fall outside of the interval (0, 0") when 0 is either 0'
or 0".] IIII
Theorem 6 Let Xl"", Xli be a random sample fromf(x; 0), where e is

some interval. Assume that the family of densities {f(x; 0): 0 E t)} has a
monotone likelihood-ratio in the statistic t(Xl , .. " Xn):
(i) If the monotone likelihood-ratio is nondecreasing in t(Xl,
· .. , xn) and if k* is such that P6o [t(Xl , ... , Xli) < k*] = Ci, then
the test corresponding to the critical region C* = {(x h " ' , xn):
t(Xl' ... , xn) < k*} is a uniformly most powerful test of size ct of
J'l' 0: 0 < 00 versus J'l'1: 0 > 00 ,
(ii) If the monotone likelihood-ratio is noruncreasing in t(XI'
· .. , xn) and if k* is such that P"o[t(XI , ... , Xn) > k*] = a, then the
test corresponding to the critical region C*={(xl , ... , xn): t(XI'
· . " Xli) > k*} is a uniformly most powerful test of size ct of J'l' 0: 0 < 00
versus J'l'l: 0 > 00 , IIII
EXAMPLE 16 Let Xl, ... , Xn be a random sample from f(x; 0) =

(110)1(0. 6)(X), where 0 > O. Test J'l'0: 0 00 versus J'l'1: 0 > 00 . We
saw in Example 15 that the family of densities has a monotone, nonincreas-
ing likelihood-ratio in t(Xh ... , xn) = YlI = max [Xh ... ,Xli]' According
to Oi) of Theorem 6, a uniformly most powerful size-a test is given by the
following: Reject J'l' 0 if YlI > k*, where k* is given as the solution to
ct = P6o [Yll > k*] = fk n 0Y ) e;;dY

60
*
(
0
n- I
=
1
O~ [O~ - (k*)lI] = 1-
(k*) 11
0 '
0
which implies that k* = 00 ViI - ct. IIII
Several comments are in order. First, the null hypothesis was stated as
o < 00 in both Theorems 5 and 6; if it had been stated as 0 > 00 , tIte two theorems
would remain valid provided the inequalities that define the critical regions were
reversed. Second, Theorem 5 is a consequence of Theorem 6. Third, the
theorems consider only one-sided hypotheses.
This completes our brief study of uniformly most powerful tests. We have
seen that a uniformly most powerful test exists for one-sided hypotheses if the
density sampled from has a monotone likelihood-ratio in some statistic. There
are many hypothesis-testing problems for which no uniformly most powerful
test exists. One method of restricting the class of tests, with the hope and
intention of finding an optimum test within the restricted class, is to consider
unbiasedness of tests, to be defined in the next subsection.
3.3 Unbiased Tests

There are many hypotheses-testing problems for which a uniformly most powerful
test does not exist. In these cases it may be possible to restrict the class of tests
and find a uniformly most powerful test in the restricted class. One such class
that has some merit is the class of unbiased tests.
Definition 17 Unbiased tests A test Y of the null hypothesis :Yf 0: (J E 8 0

against the alternative hypothesis :Yf1 : (J E e 1 is an unbiased test if and only
if
sup ny«(J) < inf ny«(J). I111
Be ~o Be ~1
Consequently in an unbiased test the probability of rejecting :Yt 0 when it is

false is at least as large as the probability of rejecting :Yf 0 when it is true. In
many respects this seems to be a reasonable restriction to place on a test. If
within this restricted class a test exists that is uniformly most powerful, then we
have a uniformly most powerfUl unbiased test. An elaborate theory has been
developed for finding uniformly most powerful unbiased tests, but we will not
study it. See [16].
3.4 Methods of Finding Tests

We presented in Subsec. 3.1 the generalized likelihood-ratio principle; it provides
us with one method of obtaining tests of hypotheses. In Subsec. 3.2 we gave a
method of finding a uniformly most powerful test for certain testing problems.
There are still other methods of finding tests. One method is sketched on pages
456 to 459 of Subsec. 5.4 of this chapter. Another method, which might be
called the confidence-interval method, is to use a confidence interval to obtain a
test. F or instance, if it is desired to test :Yf0: (J = (J 0 versus :Yt1: (J =1= (J 0, then
we might compute a confidence-interval estimate of (J from the data, and if the
interval contains (Jo, accept :Yf0, and otherwise, reject it. If the confidence
interval had a confidence coefficient y, then the resulting test would have size
1 - y. We will say more about how confidence intervals might be used to
obtain tests in Sec. 6 below.
A useful and intuitive technique for obtaining tests is the following:
Discover some statistic which behaves differently under the two hypotheses, and
utilize the different behavior to design a test. As an illustration, consider testing

.Yf0: e ~ eo versus Jr 1 : e > eo, where the sample Xl, ... , Xn is selected from
the density I(x; e) = 4>8,1 (x). The statistic X has a normal distribution with
mean e and variance 1In; hence the statistic X will tend to be smaller when .Yf0
is true than when .Yf0 is false. The statistic X behaves differently under the two
hypotheses. A reasonable test, then, would be to reject .Yf0 for X large; that is,
reject .Yf 0 if X > k, where k is determined by, say, fixing the size of the test.
(We know from Subsec. 3.2 above that such a test is uniformly most powerful.)
To employ this technique, a statistic has to be discovered which will behave
differently under the two hypotheses. There are various ways of approaching
the task of discovering such a statistic. For instance, if a sufficient statistic
exists, then it is a natural candidate to try; or a good estimator, such as a maxi-
mum-likelihood estimator, of the parameter or parameters that are used to
specify the hypotheses is another possibility for the needed statistic. In the
above simple illustration X was all these since X is the maximum-likelihood
estimator of e as well as being a sufficient statistic. We make frequent use of
this intuitive technique for obtaining tests in the remaining sections.
EXAMPLE 17 Let Xl, ... , Xn be a random sample from a Poisson distribution

with mean e. Suppose that it is desired to test that the mean is a fixed
value, say eo; that is, test .Yf 0: e = eo versus .Yf1 : e =/; eo. We know that
X is the maximum-likelihood estimator of e and that X will tend to be
distributed about eo if .Yf0 is true. Consequently, the following test seems
reasonable: Accept.Yf 0 if C1 < x < C2' and otherwise reject it, where C1 and
C2 are selected so that the test will have a desired size. To be specific, let
n = 10 and eo = I. The test given by "Accept .Yf 0 if and only if.4 < x <
1.6" has size given by
I - P8=1[.4 < X < 1.6] = I - P8=1[4 < L Xi < 16]

15 e- 10 loi
=I-
5
L j!
~ .078~ IIII
A test that has been quite extensively applied in various fields of science is
.Yf0: e = eo against .Yf1 : e =/; eo. For example, let e be the mean difference of
yields between two varieties of wheat. It is often suggested that it is desirable
to test the hypothesis .Yf0: e = 0 against .Yf1 ; e =/; 0, that is, to test if the two
varieties are different in their mean yields. However, in this situation, and
many others where e can vary continuously in some interval, it is inconceivable
that e is exactly equal to 0 (that the varieties are identical in their mean yields).
Yet this is what the test is stating: Are the two mean yields identical (to one
ten-billionth of a bushel, etc.)? In many cases it seems more realistic for an

experimenter to select an interval about 00 , say 01 < 00 < O2 , and test :Yf0: 01
o< O2 against the alternative :Yf1 : 0 < 01 or 0 > O2 • For example, it may be
feasible to set 01 = -j- and O2 = t in the above illustration and test if the
difference of the mean yields of the two varieties is between -t bushel and +1
bushel against the alternative that it is not in this interval. A test that is
uniformly most powerful for the above hypothesis may be difficult or impossible
to devise, but if/ex; 0) is a density with a single parameter, then the maximum·
likelihood estimator (3 may sometimes be used to construct a test and the power
of this test compared with the ideal power function for a test of size ct. A test
of the following form may be used for some densities: Reject lito if () is not in
some interval, say (Ch C2), and accept :Yf 0 if t} is in the interval, where C1 and C2
are chosen so that the test has size ct. Often Cl and C2 can be chosen so that
where/~(t}; 0) is the density of (3 when 0 is the parameter. The power function

of this test is
C2
nCO) = 1-
fCl
f ~(t); 0) dt} for 0 in 9.
This power function can be compared with the ideal power function, and if it
does not deviate further from the ideal than the experimenter can tolerate, the
test may be useful even though it may not be a uniformly most powerful test.
Let us illustrate the above with a simple example.
EXAMPLE 18 Let Xl' ... , Xn be a random sample from 4>(J,l(X). Test

:Yf 0: 1 <0 < 2 versus :Yf1 : () 2. X is the maximum-likelihood
estimator of 0; it has a normal distribution with mean 0 and variance lin.
According to the above we would like to select Cl and C2 so that
We have
~(~) - ~(J~) = ~(JD-t) - ~(~) = 1- ~,

and we can see from Fig. 5 that C1 = i - d and C2 = i + d, where d is given
by, say,
<ll ( d
Jl/n
+ t) - (~- ~
Nn --I-ct.
2
.5
.05 ~_ _----'..::::_"""~L...-...__--'---_---+ e
FIGURE 6 o 1 2 3
For example, if ex = .05 and n = 16, then d -:: : ; .911; so Cl -:::::; .589, and
C2 -:::::; 2.411. The power function is given by
nCO) = 1 - Pe[cl < X < c2l = 1 - Pe[.589 < X < 2.411]
and is sketched in Fig. 6. IIII
4 TESTS OF HYPOTHESES-SAMPLING
FROM THE NORMAL DISTRIBUTION
A number of the foregoing ideas are well illustrated by common practical testing
problems-those problems of testing hypotheses concerning the parameters of
normal distributions. The section is subdivided into four subsections, the first
two dealing with just one normal population and the last two dealing with several
normal populations.
4.1 Tests on the Mean

We shall assume that we have a random sample of n observations Xl' ... , Xn
from a normal population with mean J.l and variance (J2, and we will be interested
in testing hypotheses about J.l. There is quite a variety of hypotheses about the
mean J.l that can be formulated; we begin by considering one-sided hypotheses.
4 TESTS OF HYPOTHESES-SAMPLING FROM THE NORMAL DISTRIBUTION 429
:Yt 0: Jl < Jlo versus :Yt I : Jl > Jlo In testing :Yt0 : Jl S Jlo versus :Yt I : Jl > Jlo there
are two cases to consider depending on whether or not (]2 is assumed known,
If (]2 is assumed known, our parameter space is the real line, and we are testing a
one-sided hypothesis; so we have hope of finding a uniformly most powerful
test, Since (]2 is assumed known, it is a known constant; hence
= 1 e-HIl/tI)2e-!(x/tI)2e(lt/tI)X,
J2n(]
which is a member of the exponential family with
_ 1 -Hllltl)2
and d(x) = x.
( )
aJl- k e ,
y2n(]
The conditions for Theorem 5 are satisfied; so the uniformly most powerful size·
n
ct test is given by the following: Reject:Yt 0 if t(XI' .,., Xn) = I Xi > k~', where k*
I
is given as a solution to Pllo[l: Xi > k*] = ct. Now ct = Plto [I Xi > k*] =
I - q,«k* - nJlo)/jn(J); so (k* - nJlo)/jn(] = ZI-a' where Zl-a is the (1 - ct)th
quantile of the standard normal distribution, The test becomes the following:
Reject :Yt 0 if I Xi > nJlo + In(]ZI-a, or reject :Yt 0 if x >- Jlo + «(J/jn)zl_a'
If (J2 is assumed unknown, then testing :Yt0: Jl ~ Jlo versus :Yt1 : Jl > Jlo is
equivalent to testing :Yt0: (J E 9 0 versus :Yt1 : (J ~ eo,
where (J = (Jl, (]2), 9 =
{(Jl, (J2): - 00 < Jl < 00; (J2 > O}, and 9 0 = {(Jl, (]2): Jl ~ Jlo; (J2 > O}. Toobtain
a test, we could use the generalized likelihood-ratio principle, or we could find
some statistic that behaves differently under the two hypotheses and base our
test on it. Such a statistic is T = (X - Jlo)/(Sljn), where X is the sample mean
and S2 is the sample variance, Since Twould tend to be larger for Jl > Jlo than
for Jl < Jlo , a test based on T is given by the following: Reject :Yt 0 if T is large;
that is, reject :Yt0 if T> k, If Jl = Jlo, then T has a t distribution with n - I
degrees of freedom; so k can be determined by setting ct = PIl =Ilo[T > k], which
implies that k = tl-a:(n - I), the (1 - ct)th quantile of a t distribution with n - I
degrees of freedom, It can be shown that the test derived here is a generalized
likelihood·ratio test having size ct.
ft' 0: J1. = J1.o versus :YtI : Jl =1= Jlo Again, we have two cases to consider depend-
ing on whether or not (Jz is assumed known. For (Jz known, we know that
(X - z(l +"f)/2«(JI.j~), X + Z(l +"f)/z«(Jljn)) is a lOOy percent confidence interval
430 TESTS OF ID'PQTHESES IX
for Jl, where Z(l + y)/2 is the [(1 + y)/2]th quantile of the standard normal distribu-
tion. A possible test is given by the following: Reject .Yt0 if the confidence
interval does not contain Jlo . Such a test has size I - y since
_ (J - (J ]
Pp.=p.o [ X - Z(1+y)/2 J;';' < Jlo < X + Z(1+y)/2 yin = y.
If (12 is assumed unknown, we could obtain a test, similar to the one above,
using the lOOy percent confidence interval
Instead, let us find a generalized likelihood-ratio test.
L(P. (12; Xl>"" X.) = (~J exp [ - ~ Ie' P)1 ~

9 0 = {(Jl, (J2): Jl = Jlo; (12 > O}, and 8 = {(Jl, (J2): - 00 < Jl < 00; (12 > O}. We
have already seen that the values of Jl and (12 which maximize L(Jl, (12; Xl' ... , xn)
in 8 are fl = x and fJ2 = (lIn) (Xi - x)2; soL
nl2
2. _ n -n12
sup L(Jl, (1 ,Xb ···, Xn) - [ 2 ~ _2 e .
~ n ~ (Xi - x) ]
To maximize Lover 8 0 , we put Jl = Jlo, and the only remaining parameter is

(12; the value of (12 which then maximizes L is readily found to be (12 =
(lIn) L (Xi - Jlo)2, which gives
2. _ [ n ] nl2 _ nl2
sup L ( Jl, (1 ,Xl"'" Xn) - ~ 2 e .
~o 2n ~ (Xi - Jlo)
The generalized likelihood-ratio is then
We note now that A is a monotonic function of

4 TESTS OF HYPOTHESES-SAMPLING FROM tHE NORMAL DISTRIBUTION 431
and so a critical region of the form A < Ao is equivalent to a critical region of the
fonn t2(Xl' ... , x n ) > k 2. A generalized likelihood-ratio test is then given by
the following: Reject :Yf 0 if and only if
T2= [~/ln°r >k2,

or accept :1'f 0 if and only if - k < T < k. Since T has a t distribution with n - I
degrees of freedom when J-l = J-lo' if k is selected SO that J~k fT(t; n - 1) dt =
1 - a, then our test will have size a. k is given by tl -a/2(n - 1), the (l - a/2)th
quantile of a t distribution with n - 1 degrees of freedom. We might note that
this size-a test obtained by using the generalized likelihood-ratio principle is the
same size-a test that we obtained above using the confidence-interval method
of obtaining tests with the confidence interval
_ s _ s)
(X - t(1+I')/2(n - 1) In' X + t(1+I')/in - 1) In '
where y = I - a. Although we will not prove it, the test that we have obtained
is uniformly most powerful unbiased.
We have found tests on the mean of a normal distribution for both one-
sided and two-sided hypotheses. One might note that the one-sided null hy-
pothesis J-l < J-lo could be reversed and comparable results obtained. There are
other hypotheses about the mean that could be formulated, such as :Yf 0: J-ll <
J-l < J-l2 versus :Yf1: J-l < J-ll or J-l > J-l2 .
4.2 Tests on the Variance

As in the last subsection, we shall assume that we have a random sample of size n
from a normal population with mean J-l and variance a 2. We will be interested
in testing hypotheses about a 2 •
:Yf 0: a < a~ versus :Yf 1 : a 2 > a~ There are two cases to consider depending on
2
whether or not J-l is assumed known. If J-l is known, then our parameter space
is an interval, and our hypothesis is one-sided; so we have a chance of finding a
uniformly most powerful size-a test.
f(x; 8) = f(x; ( 2 ) = fi1 2

e-(1/2a )(x-Il)2,
2na
which is a member of the exponential family with a(a 2 ) = (2m'r 2 )-t, b(x) = 1,
2
c(a ) = - 1/2a2, and d(x) = (x - J-l)2. [J-l is known; so d(x) is a function of x
2
only] c(a ) is a monotone, increasing function in a 2 ; so, by Theorem 5, the
test with critical region = {(Xl' ... , Xn): L (Xi - 1l)2 > k*} is uniformly most
powerful of size ct, where k* is given by Pa2=ao2[L (Xi - 1l)2 > k*] = ct, which
implies that k* = (}"~ xi -in), where xi -cz(n) is the (l - ct)th quantile point of the
chi-square distribution with n degrees of freedom.
If Il is unknown, a test can be found using the statistic V = L (Xi - X)2/(}"~.
V will tend to be larger for (}"2 > (}"~ than for (}"2 < (}"~; so a reasonable test would
be to reject :Yf 0 for V large. If (}"2 = (}"~ , then V has a chi-square distribution
with n - I degrees of freedom, and Pa2=ao2[V> xi-in - I)] = ct, where
xi -cz(n - I) is the (l - ct)th quantile of a chi-square distribution with n - I
degrees of freedom. It can be shown that the test given by the following:
Reject :Yfo if and only if L (Xi - X)2/(}"~ > xi-cz(n - I) is a generalized likeli-
hood-ratio test of size ct.
:Yf 0: (}"2 = (}"~

versus :Yfl : (}"2 =1= (}"~ We leave the case Il assumed known as an
exercise. For Il unknown, so that 8 0 = {(Il, (}"): - 00 < Il < 00; (}"2 = (}"~}, we
can find a size-ct test using the confidence-interval method. In Subsec. 3.2 of
Chap. VIII, we found the following lOOy percent confidence interval for (}"2:
(n - I)S2, (n - l)S2),
( q2 ql
where ql and q2 are quantile points of a chi-square distribution with n - I degrees

of freedom, say fQ(q; n - I), satisfying -.
A size-(ct = I - y) test is given by the following: Accept :Yf 0 if and only if (}"5 is
contained in the above confidence interval. It is left as an exercise to show that
for a particular pair of ql and q2 the test of size ct derived by the confidence-
interval method is in fact the generalized likelihood-ratio test of size ct.
4.3 Tests on Several Means

In this subsection we will consider testing hypotheses regarding the means of two
or more normal populations. We begin with a test of the equality of two means.
Equality of two means In many situations it is necessary to compare two

means when neither is known. If, for example, one wished to compare two
proposed new processes for manufacturing light bulbs, one would have to base
the comparison on estimates of both process means. In comparing the yield of
4 TESTS OF HYPOTHESE8-SAMPLING FROM THE NORMAL DISTRIBUTION 433
a new line of hybrid com with that of a standard line, one would also have to use
estimates of both mean yields because it is impossible to state the mean yield of
the standard line for the given weather conditions under which the new line
would be grown. I t is necessary to compare the two lines by planting them in
the same season and on the same soil type and thereby obtain estimates of the
mean yields for both lines under similar conditions. Of course the comparison
is thus specialized; a complete comparison of the two lines would require tests
over a period of years on a variety of soil types.
The general problem is this: We have two normal populations-one with
a random variable Xl' which has a mean III and variance CTi, and the other with
a random variable X 2 , which has a mean 112 and variance CTi. On the basis of
two samples, one from each population, we wish to test the null hypothesis
versus .Yt 1: III # 112 , CTi > 0, CTi > O.
The parameter space 9 here is four-dimensional; a joint distribution of Xl and

X 2 is specified when values are assigned to the four quantities (Ill' 1l2, CTi, CTD.
The subspace 9 0 is three-dimensional because values for only three quantities
(Il, CTi, CT~) need be specified in order to specify completely the joint distribution
under the hypothesis that III = 112 = Il, say.
We shall suppose that there are nl observations (X11' X 12 , ... , X 1n1 ) in
the sample from the first population and n2 observations (X21' X 22 , ... , X 2n )
from the second. The likelihood function is
_ ( - -12)nl/2 exp
-
2nCTl
[l I
-"!
nl (Xli-lll)2]( - -12 )n2/2 exp [ -t In2 (X2J-1l2)2]
1 CTl 2nCT2 1 CT2
'
and its maximum in 9 is readily seen to be
Ifwe put III and 112 equal to Il, say, and try to maximize L with respect to Il, CTi,
and CT~, it will be found that the estimate of Il is given as the root of a cubic
equation and will be a very complex function of the observations. The resulting
generalized likelihood-ratio A will therefore be a complicated function, and to
find its distribution is a tedious task indeed and involves the ratio of the two
variances. This makes it impossible to determine a critical region 0 < A < k
434 TISTS OF BYPOTHBSES IX
for a given probability of a Type I error because the ratio of the population
variances is assumed unknown. A number of special devices can be employed
in an attempt to circumvent this difficulty, but we shall not pursue the problem
further here. For large samples the following criterion may be used: The root
of the cubic equation can be computed in any instance by numerical methods, and
A can then be calculated; furthermore, as we shall see in Sec. 5 below, the
quantity - 2 log A has approximately the chi·square distribution with one
degree of freedom, and hence a test that would reject for - 2 log )" large could
be devised.
When it can be assumed that the two populations have the same variance,
the problem becomes relatively simple. The parameter space 9 is then three ..
dimensional with coordinates (Ilh Ilz, (1z), while 9 0 for the null hypothesis
III = Ilz = Il is two-dimensional with coordinates {Jl, (1z). In 9 we find that
the maximum-likelihood estimates of Ill' Ilz, and (1z are, respectively, Xf, Xz,
and
so
for Il
and
which gives
supL
~o
nl + n2 ](n +n )/Z
1 2
[ 2n [I (Xli - XI)Z + I (XZi - x )Z +

2 nlnz (Xl - Xz)z]
nl + n2
Finally,
it = (1 + [n l n2/(n l ~ n2)](XI - X2)2 2) 2
-(n1 +n )/2.
I (Xli - Xl) + I(X2j - x2 ) (17)

This last expression is very similar to the corresponding one obtained in
Subsec. 4.1, and it turns out that this test can also be performed in terms of a
quantity which has the t distribution. We know that Xl and X 2 are independ-
ently normally distributed with means III and 112 and with variances (i /n l and
u 2 /n2' Also it is readily seen that Xl - X 2 is normally distributed with mean
III - 112 and variance u 2(1/nl + I/nz)· Under the null hypothesis the mean of
Xl - X z will be O. The quantities I (XIt - XI)2ju 2 and I (XZ) - Xz)Z/u 2 are
independently distributed as chi-square distributions with nl - I and n z - I
degrees of freedom, respectively; hence their sum has the chi-square distribution
with nl + nz - 2 degrees of freedom. Since under the null hypothesis
Z= ~l-XZ
uJI/nl + l/nz
is normally distributed with mean 0 and unit variance, the quantity
T= JnlnZ/(nl + n2)(XI - X 2) (18)

J[I (Xli - XI)z + I (XZi - X 2)Z]/(nl + nz - 2)
has the t distribution with nl + nz - 2 degrees of freedom. [Note that we do

have independence of the numerator and denominator in Eq. (18).] The gener-
alized likelihood-ratio is
(19)
and its distribution is determined by the t distribution. The test would, of

course, be done in terms of T rather than 2. A 5 percent critical region for Tis
Z
T > [t.97S(nl + nz - 2)]2, where l.97S(nl + nz - 2) is the .975th quantile of the
t distribution with nl + n2 - 2 degrees of freedom.
If we want to test :Yf 0: III = 112 versus :Yf I : III > Ilz or :Yf 0: III ::;;; Ilz versus
:Yf I: III > Ilz, a size-IX test is given by the following: Reject :Yf 0 if and only if
T> tl-lnl + n z - 2), where Tis defined in Eq. (18) and tl-a:(n l + n2 - 2) is the
(1 - IX)th quantile of the t distribution with nl + nz - 2 degrees of freedom.
Equality of several means The test presented above can be extended from
just two normal popUlations to k normal populations. We assume that we have
available k random samples, one from each of k normal populations; that is,
436 TBSTS OF HYPOTHESES IX
let Xii' •.. , XJnJ be a random sample of size nj from the jth normal population,
j = I, ... , k. Assume that the jth population has mean Ii) and variance (12.
Further assume that the k random samples are independent. Our object is to
test the null hypothesis that all the population means are the same versus the
alternative that not all the means are equal. We seek a generalized likelihood·
ratio test. The likelihood function is given by
k
where n = L nj.
j= I
The parameter space 9 is (k + 1)-dimensional with coordinates
(iii' ... , lik' (12), and 9 0 , the collection of points in the parameter space corre-
sponding to the null hypothesis, is two-dimensional with coordinates (Ii, (12), where
Ii = iii = ... = lik' In 9, the maximum-likelihood estimates of iii' ... , lik, (12
are given by
I nJ
A
II
r j =X.
-
J. =- I x",
JI j = 1, ... , k,
n)l= I
and
1
"'2 = -
(1-
~
Lk I( x .. - x·- )2.
HJ
)1 J.' (20)
n j=l i=l
hence,
In 9 0 , the maximum-likelihood estimates of Ii and (12 are
and
and so
21l ~ ~ (Xji _X)2] -n12
sup L = )~ e- nI2 •
§o
[ n
- 2
- ~~ (Xji - X J,)
= 1+
k - 1 ~ nix), - x)2/(k - 1)
J 2
l-n'2
[ n - k ~~(Xjl- x),) I(n - k)
A generalized likelihood-ratio test is given by the following: Reject :Yf 0 if and

only if A Ao ' But A < Ao if and only if
k
I niXj. - x)2j(k - 1)
j= 1
r= " " 2 > some constant, say c. (21)
~~ (Xji - XJ.) I(n - k)
J i
The ratio r is sometimes called the variance ratio, or F ratio. The constant c is
determined so that the test will have size (X; that is, c is selected so that
P[R > c 1:Yf0] = (x. Note that Xj. is independent of I (Xji - X l )2 and,hence, the
i
numerator of Eq. (21) is independent of the denominator. Also, under :Yf 0'
note that the numerator divided by (12 has a chi-square distribution with k - I
degrees of freedom, and the denominator divided by (12 has a chi-square dis-
tribution with n - k degrees of freedom. Consequently, if :Yf 0 is true, R has an
F distribution with k - land n - k degrees of freedom; so the constant c is the
(1 - (X)th quantile of the F distribution with k - 1 and n - k degrees of freedom.
The testing problem considered above is often referred to as a one-way
analysis of variance. In some experimental situations, an experimenter is
interested in determining whether or not various possible treatments affect the
yield. F or example, one might be interested in finding out whether various
types of fertilizer applications affect the yield of a certain crop. The different
treatments correspond to the different populations, and when we test that there
is no population difference, we are testing that there is no treatment" effect.
U
The term " analysis of variance" is explained if we note that the denominator of
the ratio in Eq. (21) is an estimate of the variation within populations and the
· ..
, ,.-
numerator is an estimate of the variation between populations when means are

equal. We are analyzing variance to test equality of means.
4.4 Tests on Several Variances
Two variances Given random samples from each of two normal populations
with means and variances (/11' uI) and (/12' ui), we may test hypotheses about
the two variances. We will consider testing:
(i) .Yf0: ui < ui versus .Yf1: ui > ui
(ii) .Yf0: ui > ui versus .Yf1: ui < ui
(iii) .Yf0: ui = ui versus .Yf1: ui =1= ui
If X 11 , .•. , X l711 is a random sample from a normal density with mean /11
and variance ui, if X 21 , ..• , X 2712 is a random sample from a normal density
with mean /12 and variance ui, and if the two samples are independent, then we
know that
has the F distribution with n1 - 1 and n2 - 1 degrees of freedom, and, in partic-

ular, the statistic
has the F distribution with nl - 1 and n2 - 1 degrees of freedom when ui = ui .

Note that the statistic R tends to be large when ui > ui and small when ui < ui,
and so we can capitalize on this different behavior to formulate tests for the
hypotheses (i) to (iii). F or instance, in testing .Yf0: ui ui,
we would reject .Yf0 for large R, or a size-a test is given by the following: Reject
.Yf 0 if and only if R exceeds F1 - rz (n 1 - 1, n 2 - 1), the (1 - a)th quantile of the
F distribution with nl - 1 and n 2 - 1 degrees of freedom. Similarly, a test of
.Yf0: ui > ui versus .Yf1: ui < ui is given by the following: Reject .Yf0 if and
only if R is less than Finl - 1, n2 - 1), the ath quantile of the F distribution
with n 1 - 1 and n2 - 1 degrees of freedom. A test of .Yf0: ui = ui versus
.Yf1 : ui =1= ui should be two-tailed; that is, .Yf0 should be rejected for small or
large R. In other words, a test is given by the following: Accept .Yf0 if and only
if kl < R < k2' where kl and k2 are selected so that the test will have size a.
It is customary to make the two tails have equal areas of al2 (although this
is not quite the best test); then kl = Frz / 2(n 1 - 1, n2 - 1), and k2 =
F1 - rz / 2 (n 1 - 1, n 2 - 1).
4 TEST OF HYPOTHESES--SAMPUNG FROM THE NORMAL DISTRIBUTION 439
We might mention that the above defined tests can all be derived using the
generalized likelihood-ratio principle.
Equality of several variances Let X jI , •.. , Xj"j be a random sample of size nj

from a norma] population with mean Jl j and variance a}, j = 1, ... ,k. Assume
that the k samples are independent. Our object is to test the null hypothesis
:If0: (1'I = (1'i = ... = (1'~ against the alternative that not all variances are equal.
The likelihood function
- n n .1 k"j
e- U (XJI-Jl.J)/O'jJ2
j= 1 i~ 1 J21l (1'j
and the maximum-likelihood estimates of Jlj' (1'1, j = 1, ... , k are given by
and
1 nj
hJ = - I (Xji - XjJ z.
nJ i= 1
The null hypothesis states that all (1'] are equal. Let (1'2 denote their common
value; then So = {(Jll' ... , Jlk' (1'2): - 00 < Jlj < 00; (1'2 > O}, and the maximum·
likelihood estimates of JlI" .. , Jlk' (1'2 over So are given by
j = 1, ... , k,
and
Therefore,
) _ s~f L (fzt'J12 exp (-2»/2)

,- s~p L - Ii (~)"j/2 exp (- I
j= 1 (})
nj/2)
_
n (hJ)"i/
j=1
k
Z
- (I n) hJ/I n)Y'''j/z'
A generalized likelihood-ratio test is given by the following: Reject :If 0 if and
only if A :S; Ao . We would like to determine the size of the test for any constant
AO or find Ao so that the test has size a, but, unfortunately, the distribution of the
generalized likelihood-ratio is intractable. An approximate size-a test can be
obtained for large nJ since it can be proved that - 2 log A is approximately
distributed as a chi-square distribution with k - 1 degrees of freedom. Accord~
ing to the generalized likelihood~ratio principle .Yf0 is to be rejected for small
A; hence .Yf 0 should be rejected here for large - 2 log A; that is, the critical region
of the approximate test should be the right tail. So the approximate size-a test
is the following: Reject .Yf 0 if and only if - 2 log A > Xl -a.Ck - I), the (l - a)th
quantile of the chi~square distribution with k - I degrees of freedom. (Several
other approximations to the distribution of the likelihood~ratio statistic have
been given, and some exact tests are also available.)
5 CHI-SQU ARE TESTS
In this section we present a number of tests of hypotheses that one way or another
involve the chi-square distribution. Included will be the asymptotic distribution
of the generalized likelihood-ratio, goodness-of-fit tests, and tests concerning
contingency tables. The material in this section will be presented with an aim
of merely finding tests of certain hypotheses, and it will not be presented in such
a way that concern is given to the optimality of the test. Thus, the power
functions of the derived tests will not be discussed.
5.1 Asymptotic Distribution of Generalized Likelihood-ratio

On two occasions in Sec. 4 we found that the distribution of the generalized
likelihood-ratio was intractable, and both times we indicated that an approximate
test could be obtained by using an asymptotic distribution of the generalized
likelihood-ratio. The following theorem, which we shall not be able to prove
because of the advanced character of its proof, gives the asymptotic distribution
of the generalized likelihood~ratio.
Theorem 7 Let X b ••• , Xn be a sample with joint density Ix.,· .. , Xn

(', ... , . ; 0), where 0 = (0 1 , •.• , Ok)' that is assumed to satisfy quite general
regularity conditions. Suppose that the parameter space (3 is k-dimen~
sional. In testing the hypothesis
5 em-SQUARE TESTS 441
where O~, •.. , O~ are known and Or+l' ..• , Ok are left unspecified, -2 log An
is approximately distributed as a chi-square distribution with r degrees of
freedom when :If 0 is true and the sample size n is large. //11
We have assumed that 1 <r < k in the above theorem. If r = k, then all
parameters are specified and none is left unspecified. The parameter space e is
k-dimensional, and since :If 0 specifies the value of r of the components of
(01 , •.• , Ok)' the dimension of So is k - r. Thus, the degrees of freedom of the
asymptotic chi-square distribution in Theorem 7 can be thought of in two ways:
first, as the number of parameters specified by :If 0 and, second, as the difference
e
in the dimensions of and 9 0 ,
Recall that An is the random variable which has values
which in tum is the generalized likelihood-ratio for a sample of size n. 9 0 is

that subset of 9 that is specified by :If o. The generalized likelihood~ratioprin~
ciple dictates that :If 0 is to be rejected for A" small, but since - 2 log An increases
as An decreases, a test that is equivalent to a generalized likelihood-ratio test is
one that rejects for -2 log A" large. Now, since the theorem gives an approx-
imate distribution for the values - 2 log An when :If 0 is true, a test with approx-
imate size a; is given by the following:
Reject :If 0 if and only if - 2 log An > xf _«(r),

where XI _«(r) is the (l - a)th quantile of the chi-square distribution with r
degrees of freedom. Note that the degrees of freedom r is the number of compo-
nents of the parameter space that are specified by the null hypothesis.
Because of the specific form of the null hypothesis in the theorem, it may
appear that the result is not too widely applicable. The null hypothesis of the
theorem specifies the values of a subset of the k components of the k-dimensional
parameter space, and not many null hypotheses are of that form. However,
often the density can be reparameterized so that the null hypothesis is of the
form given in the theorem. We illustrate with two examples.
EXAMPLE 19 Recall that in Subsec. 4.3 we discussed testing :If 0: III 1l2'
O'f > 0, O'i > 0 versus :lfl : III =1= 1l2' O'i > 0, O'i > 0, where III and 0';
are
the mean and variance of one normal population and 112 and O'i are
the mean and variance of another. Here the parameter space is four-
dimensional, and although :If0 does not appear to be of the form given
442 TESTS OF HYPO'IHESES IX
in Theorem 7, we can reparameterize to make it of that form. Let

01 = Jl1 - 112, O2 = Jl2' 03 = O'r, and 04 = O'~ • In terms of the reparam-
eterization, :Yf 0 becomes :Yf 0: 01 = O~ = 0, O2 , 03 , 04 ; that is, the com-
ponent 01 is specified to be 0, and the remaining three components are
unspecified. The theorem is now applicable for the reparameterization;
that is, the asymptotic distribution of - 2 log A' is known (and is the chi-
square distribution with one degree of freedom) for :Yf 0 true, where A' is
the generalized likelihood-ratio obtained under the reparameterization.
However, because of the invariance property of maximum-likelihood
estimators, the generalized likelihood-ratio A' obtained under the reparam-
eterization is the same as the generalized likelihood-ratio A obtained
before reparameterization. jill
EXAMPLE 20 In Subsec. 4.4 we tested :Yf 0: O'r = ... = 0';, 111' ... , 11m, where
I1j and 0'] were, respectively, the mean and variance of the jth normal
population, j = 1, ... , m. (In Subsec. 4.4, k was used instead of m.) If
we make the following reparameterization, :Yf 0 will have the desired form
of Theorem 7:
Now:Yf o becomes :Yfo: 01 = 1, ... , 0m-l = 1, Om, Om+l'"'' 02m; that is,
the first m - 1 components are specified to be 1 and the remaining are
unspecified. Theorem 7 is now applicable, and, again, because of the
invariance property of maximum-likelihood estimates, the generalized
likelihood-ratio obtained before and after reparameterization are the
same; hence the asymptotic distribution of - 2 log A, as claimed in
Subsec. 4.4, is the chi-square distribution with m - 1 degrees of freedom
when :Yf 0 is true. // /I
5.2 Chi-square Goodness-of-fit Test

We commence this section with an example of a testing problem that involves
the specification of the parameters of a multinomial distribution. It is hoped
that this example will help motivate the presentation of the goodness.of-fit test.
If a population has a multinomial density
"+1
f(x 1 , ••• , x" ; Ph "', p,,) = n pji,
J=l
(23)
5 CHI-sQUARE nsTS 443
1+1
where xi = 0 or 1, j = 1, ... , k + 1; 0 < Pj < 1, j = 1, "', k + 1; I Xj = 1;
j= 1
k+l
and I Pj = 1 (as would be the case in sampling with replacement from a
j=l
population of individuals who could be classified into k + I classes or categories),
a common problem is that of testing whether the probabilities Pj have specified
numerical values. Thus, for instance, the result of casting a die may be classified
into one of six classes, and on the basis of a sample of observations we may wish
to test whether the die is true, that is, whether Pj = t for j = 1, ... , 6. One can
also think in terms of independent, repeated trials, where each trial can result in
anyone of k + 1 outcomes, called classes or categories. The density in Eq. (23)
then gives the density for the outcome of one trial. The result of one trial can
be represented by the multivariate random variable (Xl' ••. , Xk), where Xj is
unity if the trial results in category j and is 0 otherwise. Pj is the probability
that a trial results in category j. Now if we independently repeat the trial n
times, we have n observations of the multivariate random variable (Xl' ... , X k);
we can display them as
If we let N j = I" Xl}' then the random variable N J is the number of the n trials
t= 1
resulting in category j. We know that (Nl' ... , N k ) has a multinomial distribu..
tion. (See Example 5 in Subsec. 2.2 of Chap. IV.)
To test the null hypothesis .1fo: Pj = pJ, j = 1, ... , k + 1, where pJ are
given probabilities summing to unity, we hope to employ the generalized
likelihood-ratio principle. The likelihood function is given by
L = L(Pl' .•. , Pk; X 1b ••• , Xu, ••. , X"h ..• , X"k) = n" k+n pjli.
i= 1 }= 1
1
(24)
. The parameter space e has k dimensions (given k of the k + 1 P/s, the remaining
one is determined by I Pj = 1), while 9 0 is a point. It is readily found thatL is
maximized in 9 when
where nj is a value of the random variable N j • Hence,

1 k+1
sup L
~
= n
n J=1
n njJ.
444 TESTS OF HYI'01"HBSES IX
The maximum of Lover n (P1 )'tJ, and so the generalized

eo is its only value j=l
1;+1
likelihood-ratio is
A = n" n
k+l (pO)"J
1'=1
-1
nj
.
A generalized likelihood-ratio test is given by the following: Reject JIP 0 if and

only if A < )'0 , where the constant Ao is chosen to give the desired probability of
a Type I error. For small n, the distribution of the generalized likelihood-ratio
may be tabulated directly in order to determine Ao ; for large values of n, we may
use Theorem 7, which states that -2 log A has approximately the chi-square
distribution with k degrees of freedom. The chi-square approximation is sur~
prisingly good even if n is small provided that k > 2.
Another test which is still commonly used for testing JIP 0 was proposed (by
Karl Pearson) before the general theory of testing hypotheses was developed.
This test uses the statistic
o k+l (N j - npJ)2
Qk = I 0' (25)
j=1 np1'
which tends to be small when JIP 0 is true and large when JIP 0 is false. Note that
N j is the observed number of trial outcomes resulting in category j and nPJ is the
expected number when JIP 0 is true. It can be easily shown (see Prob. 39)
that
k+l I
S[Q~] = I -0 [np/l - Pj) + n2 (pj - pJ)2], (26)
J= 1 np J
where the Pj are the true parameters. If JIP 0 is true, then S[Q~] = I (1 - pJ) =
k + 1 - 1 = k. The following theorem gives a limiting distribution for Q~ when
the null hypothesis .;t(? 0 is true.
Theorem 8 Let the possible outcomes of a certain random experiment

be decomposed into k + 1 mutually exclusive sets, say Ab "', A k + 1 •
Define PJ = P[A j], j = I, ... , k + 1. In n independent repetitions of the
random experiment, let N j denote the number of outcomes belonging to
k+l
set A j , j = 1, ... , k + 1, so that I Nj = n. Then
j=1
k+l (N. - np.)2

Qk = I J J (27)
j=1 npj
has as a limiting distribution, as n approaches infinity, the chi-square

distribution with k degrees of freedom. 1//1
5 CHI-SQUARE TESfS 445
We will not prove the above theorem, but we will indicate its proof for
k = 1. What needs to be demonstrated is that for each argument X~ FQk(x)
converges to Fx2(k/x) as n --'). 00, where FQk(') is the cumulative distribution
function of the random quantity Qk and F X2(k)(') is the cumulative distribution
function of a chi-square random variable having k degrees of freedom. (Note
that k + 1, the number of groups, is held fixed, and n, the sample size, is increas-
ing.) If k = 1, then
We know that NI has a binomial distribution with parameters n and PI and that
Y,. = (Nl - npI)/JnpI(l - PI) has a limiting standard normal distribution;
hence, since the square of a standard normal random variable has a chi-square
distribution with one degree of freedom, we suspect that Y; = Q1 has a limiting
chi-square distribution with one degree offreedom, and such can be easily shown
to be the case, which would give a proof of Theorem 8 for k = 1.
Theorem 8 gives the limiting distribution for the statistic
o _ k+1
~
(N.J _ npO)2
j
Qk-.i..J
}=1 npj °
when the null hypothesis Yf o : p) = pJ,j = 1, ... , k + 1, is true. Thus a test of
Yf 0: p} = pJ, j = 1, ... , k + 1, which has approximate size a, is given by the
following:
Reject Yf ° if and only if Q~ > xr -a.(k),

the (1 - a)th quantile of the chi-square distribution with k degrees of freedom.
We now have two large-sample tests of the null hypothesis Yf o : Pj = pJ,
i = 1, ... , k + I, the one just defined, which uses Theorem 8, and the other given
in terms of the generalized likelihood-ratio, which uses Theorem 7. It can, in
fact, be shown that the two tests are equivalent for large samples.
EXAMPLE 21 Mendelian theory indicates that the shape and color of a

certain variety of pea ought to be grouped into four groups, "round and
yellow," "round and green," .. angular and yellow;' and angular and
U
green," according to the ratios 9/3/3/1. For n = 556 peas, the following
were observed (the last column gives the expected number):
Round and yellow 315 312.75
Round and green 108 104.25
Angular and yellow 101 104.25
Angular and green 32 34.75
9 3 3
A size-.05 test of the null hypothesis .Yf0: PI = 1 6' P2 = 1 6' P3 = 1 6'
and P4 = /6 is given by the following:
4 (N. _ npO)2
Reject.Yf ° if and only if Q~ = L J ° j exceeds XT-alk) = X~95(3) =
1 npj
7.81.
The observed Q~ is
(315 - 312.75)2 (l08 - 104.25)2 (101 - 104.25)2 (32 - 34.75)2
312.75 + 104.25 + 104.25 + 34.75
~ .470,
and so there is good agreement with the null hypothesis; that is, there is a
good fit of the data to the model. IIII
Theorem 8 can be generalized to the case where the probabilities Pj may

depend on unknown parameters. The generalization is given in the next
theorem.
Theorem 9 Let the possible outcomes of a certain random experiment

be decomposed into k + 1 mutually exclusive sets, say AI' ... , A k + 1 •
Define Pj = P[Ajl, j = 1, ... , k + 1, and assume that Pj depends on r
unknown parameters 01 , ••• , On SO that Pj = /tiOl' ... , Or), j = 1, ... ,
k + 1. In n independent repetitions of the random experiment, let N j
denote the number of outcomes belonging to set A j ' j = 1, ... , k + 1, so
k+l
that L Nj = n. Let 9 1 , ••• , 9 r be BAN estimators (e.g., maximum-
j= 1
likelihood estimators) of 01 , ••• , Or based on N 1 , ••• , Nk • Then, under
certain general regularity conditions on the /t / s,
' _ k~l (N j - nPj )2
Qk - L - . . . . (28)
j=l nPj
has a limiting distribution that is the chi-square distribution with k - r
degrees of freedom, where Pj = /ti91' ... , 9 r ), j = 1, ... , k + 1. IIII
5 CHI-SQUARE TESTS 447
The proof of Theorem 9 is beyond the scope of this book. The limiting
distribution given in Theorem 9 differs from the limiting distribution given in
Theorem 8 only in the number of degrees of freedom. In Theorem 8 there are k
degrees of freedom, and in Theorem 9 there are k - r degrees of freedom; the
number of degrees of freedom has been reduced by one for each parameter that
is estimated from the data.
No mention of hypothesis testing is made in the statement of Theorem 9.
However, we will show now how the results of the theorem can be used to obtain
a goodness-of~fit test. Suppose that it is desired to test that a random sample
Xt, ... , Xn came from a density I(x; 81, ... , 8,.), where 81, ... , 8,. are unknown
parameters but the function I is known. The null hypothesis is the composite
hypothesis .Yf0: Xi has density I(x; 81, ... , 8,.) for some 81, ... , 8,.. The null
hypothesis states that the random sample came from the parametric family of
densities that is specified by I( . ; 81 , ••• , 8,.). If the range of the random variable
Xi is decomposed into k + 1 subsets, say AI' ... , A k + l , if Pj = P[X, E A j ], and if
N j = number of X/s falling in A j' then, according to Theorem 9,
k+l (N - np.)2
Q" = L1 j
nPj
J
is approximately distributed as the chi-square distribution with k - r degrees of

freedom if n is large and .Yf 0 is true, where Pj = /tiel, ... , e,.) and i is a e
maximum-likelihood estimator of 8i , i = 1, ... , r, obtained from the statistics
N l , ••• , N k • {Note that /ti81' ... ,8,.) =P[XiEA j ], which for a continuous
random variable Xi equals SA.jl(x; 81 , ••• , 8,.) dx.} Hence, a test of .Yf0 can be
obtained by rejecting .Yf0 if and only if the statistic Qfc is large; that is, reject
.Yf0 if and only if Q" exceeds xi -a(k - r), where xi -a(k - r) is the (1 - ct)th
quantile of the chi-square distribution with k - r degrees of freedom. Such a
test is called a goodness-of-fit test since it tests whether or not the observations
Xl, ••. , Xn fit, or are consistent with, the assumption that they are observations
from the density I(x; 81 , ••• , 8..).
In the above, the 8" for i = 1, ... , r, were estimated by using the statistics
N l , •.• , Nk rather than Xl' ••• , X n • The statistics N l , . , . , Nk give the number
of X observations falling in each of the A j subsets or groups. In practice, often
the values of the X/s are not recorded, and then the group totals N l , ••. , Nk
constitute the available information. If, however, the observations Xl' ... , Xn
were available, then one could estimate 8i , i= 1, ... , r, more efficiently by using,
say, maximum-likelihood estimators based on Xl ... , X n • When such esti-
mators are used, the limiting distribution of Q~ is no longer a chi-square distribu~
tion with k - r degrees of freedom; instead, the limiting distribution of Qi is
bounded between a chi-square distribution with k - r degrees' of freedom and a
chi-square distribution with k degrees of freedom. In a sense, some of the

"lost" r degrees of freedom are recouped by efficiently estimating (h, ... , Or.
For a proof of Theorem 9 and further discussion of the above, the reader is
referred to Kendall and Stuart [14].
EXAMPLE 22 Suppose it is desired to test the hypothesis that an observed

random sample XI, ••• , Xn has been drawn from some normal population.
Let the n sample values Xl' .•• , Xn be grouped into k + 1 classes. For
example, thejth class could be taken as all those observations falling in the
interval (Zj_I' Zj], j = 1, ... , k + 1, for some Zo < Zl < Z2 < ... <Zk
< zk+l,where Zo = - 00 and Zk+1 = + 00. Then
Let {l and ff be the maximum-likelihood estimates of Jl and (1 based on

nl, ... , nk, where nj is the number of observations falling in thejth interval.
Then,
can be determined from the sample, and so the value

k+ I (n - nftJ)2
qk = I J "
j= I nPJ
of Q~ can also be obtained from the sample. The hypothesis that the
sample came from a normal population would be rejected at the ex lever if
q/c > xI -rlk - 2). If, on the other hand, Jl and (1 were obtained from
maximum..likelihood estimators based on X., ... , X n , then the asymptotic
distribution of Q~ would fall between a chi-square distribution with k - 2
degrees of freedom and a chi.. square distribution with kdegrees of freedom.
The hypothesis would be rejected if q~ > C, where C falls between
xi _rlk - 2) and XI -rJ.(k). Note that for k large there is little difference
between XI-rlk - 2) and xf-rlk). IIII
5.3 Test of the Equality of Two Multinomial

Distributions and Generalizations
A problem that is of great practical importance is that of testing whether several
random samples can be considered as drawn from the same population. For
instance, in Subsec. 4.3 we tested whether several assumed normal populations
could be considered the same normal population. In this subsection we first
indicate a test of the hypothesis that two multinomial populations can be con-
sidered the same and then indicate some generalizations. Suppose that there are
k + 1 groups associated with each of the two multinomial populations. Let the
first popula ti on have associated probabiIi ties Pll' P12' ... , PI k' PI, k+ 1 and the
second P21' P22' ... , P2k , P2, k+l· It is desired to test .Yt0: P1l = P2j (= Pj' say),
j = 1, ... , k + 1. For a sample of size nl from the first population, let N 1j
denote the number of outcomes in group j, j = 1, ... , k + 1. Similarly, let N 2j
denote the number of outcomes in group j of a sample of size n2 from the second
population. (Here we are assuming that the sample sizes nl and n2 are known.)
We know that
I
k+l (N ij - niPij )2
j= 1 nj Pij
has a limiting chi-square distribution with k degrees of freedom for i = 1 and 2;

hence
(N
I2 k+I 1
ij - niPij
)2
i=1 j=1 niPij
has a limiting chi-square distribution with 2k degrees of freedom if the two

random samples are independent. If.Yt 0 is true, then
Q2A; = I2 k+l(N
I ij - n p)2
i j (29)
i=1 j=1 niPj
has a limiting chi-square distribution with 2k degrees of freedom. If.Yt 0
specifies the values Pj' then Q2k is a statistic and can be used as a test statistic.
On die other hand, if the Pj defined by .Yt0 are unknown, then they have to be
estimated. If.Yt 0 is true, the two samples can be considered as one random
sample of size n 1 + n 2 from a multinomial population with probabilities PI, ... ,
Pk+l. Maximum-likelihood estimators of the Pj are then (N 1j + N 2j )/(nl + n2)'
j = 1, ... , k, and if the Pj in Eq. (29) are replaced by their maximum-likelihood
estimators, we then obtain
Q;k = ± kII [Nij - ni(N Ij + N 2j)/(nl + n2)F .

i= 1 j= 1 ni(N 1l + N 2j)/(n 1 + n2)
(30)
It can be shown that Q;k has a limiting chi-square distribution with 2k - k = k

degrees of freedom. (This result is not a direct corollary of Theorem 9; it
would, however, be a corollary of a generalization of Theorem 9 from one to
two populations.) Again the degrees of freedom of the limiting distribution of
Q2k have been reduced by unity for each parameter estimated.
Another test of the homogeneity of two multinomial populations can be

derived by finding the generalized likelihood-ratio A and employing Theorem 7
to obtain the limiting distribution of - 210g A. (Reparameterization is required
before Theorem 7 can be employed directly.) The details of finding such a test
are left as an exercise.
EXAMPLE 23 In an opinion survey regarding a certain political issue there

was some question as to whether or not the eligible voters under 25 years
of age might view the issue differently from those over25. Fifteenhundred
individuals of those over 25 were interviewed, and 1000 of those under 25
were interviewed with the following results (the data are obviously arti-
ficial to facilitate calculations):
Opposed Undecided Favor Total
Under 25 400 100 500 1000
Over 25 600 400 500 1500
Total 1000 500 1000 2500
Test the null hypothesis that there is no evidence of difference of opinion

due to the different age grouping; that is, test .Yf0: PI} = P2j = p} , j = 1,
2, 3. PI and P2 need to be estimated. We can calculate the value of the
statistic given in Eq. (30) as follows:
(400 - 1000 . 1000/2500)2 (100 - 1000· 500/2500)2
1000· 1000/2500 + 1000 . 500/2500
(500 - 1000 . 1000/2500)2 (600 - 1500 . 1000/2500)2
+ 1000· 1000/2500 + 1500·1000/2500
(400 - 1500 . 500/2500)2 (500 - 1500· 10oo/2500?
+ 1500· 500/2500
. + 1500' 1000/2500 = 125 .
The 99 percent quantile point for the chi-square distribution with two
degrees of freedom is only 9.21; so there is strong evidence thatthetwoage
groups have different opinions on the political issue. fill
The technique presented in this subsection can be generalized in two direc-

tions. First, a test of the homogeneity of several, rather than just two, multi-
nomial populations can be obtained, and, second, a test of the hypothesis that
several given samples are drawn from the same population of a specified type
(such as the Poisson, the gamma, etc.) can be obtained using a procedure similar
to that above. We illustrate with an example.
EXAMPLE 24 One hundred observations were drawn from each of two Poisson
populations with the following results:
0 1 2 3 4 5 6 7 8 9 or more Total
Population 1 11 25 28 20 9 3 3 0 1 0 100
Population 2 13 27 28 17 11 1 2 1 0 0 100
Total 24 52 56 37 20 11 200
Is there strong evidence in the data to support the contention that the two
Poisson populations are different? That is, test the hypothesis that the
two populations are the same. This hypothesis can be tested in a variety
of ways. We first use the chi-square technique mentioned above. We
group the data into six groups, the last including all digits greater than 4, as
indicated in the above table. If the two populations are the same, we have
to estimate one parameter, namely, the mean of the common Poisson
distribution. The maximum-likelihood estimate is the sample mean,
which is
0(24) + 1(52) + 2(56) + 3(37) + 4(20) + 5(4) + 6(5) + 7(1) + 8(1)
200
420
= 200 = 2.1.
The expected number in each group of each population is given by
0 1 2 3 4 5 or more
12.25 25.72 27.00 18.90 9.92 6.21
The value of the statistic in Eq. (29), where niPj is replaced by the estimates
given in the above table, can be calculated. It is approximately 1.68.
The degrees of freedom should be 2k - 1 (one parameter is estimated),
which is 9. The test indicates that there is no reason to suspect that the
two assumed Poisson populations are different Poisson populations. IIII
We mentioned earlier that there are several methods of testing the null
hypothesis considered here. For example the generalized likelihood-ratio prin-
ciple and employment of Theorem 7 yield a test that the student may find in-
structive to find for himself.
5.4 Tests of Independence in Contingency Tables

A contingency table is a multiple classification; for example, in a public opinion
survey the individuals interviewed may be classified according to their attitude
on a political proposal and according to sex to obtain a table of the form
Favor Oppose Undecided
Men 1154 475 243

Women 1083 442 362
This is a 2 x 3 contingency table. The individuals are classified by two criteria,

one having two categories and the other three categories. The six distinct
classifications are called cells. A three-way contingency table would have been
obtained had the individuals been further classified according to a third criterion,
say, according to an annual-income group. If there were five income groups
set up (such as under $2000, $2000 to $4000, ... ), the contingency table would be
called a 2 x 3 x 5 table and would have 30 cells into which a person might be
put. It is often quite convenient to think of the cells as cubes in a block two
units wide, three units long, and five units deep. If the individuals were still
further classified into eight geographic locations, one would have a four-way
(2 x 3 x 5 x 8) contingency table with 240 cells in a four-dimensional block
with edges two, three, five, and eight units long. A contingency table provides
a convenient display of the data for ultimately investigating suspected relation-
ships. Thus one may suspect that men and women will react differently to a
certain political proposal, in which case one would construct such a table as the
one above and test the null hypothesis that their attitudes were independent of
their sex. To consider another example, a geneticist may suspect that suscep-
tibilitytoacertain disease is heritable. He would classify a sample of individuals
according to (i) whether 01 not they ever had the disease, (ii) whether or not their
fathers had the disease, and (iii) whether or not their mothers had the disease.
In the resulting 2 x 2 x 2 contingency table he would test the null hypothesis
that classification (i) was independent of (ii) and (iii). Again a medical research
worker might suspect a certain environmental condition favored a given disease
and classify individuals according to (i) whether or not they ever had the disease,
5 CHI-SQUARE TESfS 453
(ii) whether or not they were subject to the condition. An industrial engineer
could use a contingency table to discover whether or not two kinds of defects in
a manufactured product were due to the same underlying cause or to different
causes. It is apparent that the technique can be a very useful tool in any field
of research.
Two-way contingency tables We shall suppose that n individuals or items are

classified according to two criteria A and B, that there are r classifications
AI, A 2 , ••. , A,. in A and s classifications B 1 , B 2 , •• Bs in B, and that the number
0 ,
of individuals belonging to Ai and B} is Nij. We have then an r x scontingency

table with cell frequencies N ij and I N i} = n:
(31)
As a further notation we shall denote the row totals by N i. and the column totals
by N .}; that is,
N·I. =".t.- N··

}
'J and N·
oj =".t.- N .. ,
i
I}
Of course,
It N i. = 2: N. j = n.
J
We shall now set up a probability model for the problem with which we
wish to deal. The n individuals will be regarded as a sample of size n from a
multinomial population with probabilities Pi} (i = 1, 2, ... , r; i = 1, 2, .. " s).
The probability density function for a single observation is
where
or 1 and L
t, }
xi} = L
We wish to test the null hypothesis that the A and B classifications are independ~
ent, i.e., that the probability that an individual falls in B j is not affected by the
A class to which the individual happens to belong. Using the symbolism of
Chap. I, we would write
and
or
P[A i n B}l = P[AilP[Bjl.
If we denote the marginal probabilities P[Ail by Pi. (i = 1, 2, ... , r) and the
marginal probabilities P[Bjl by P.j (j = I, 2, ... , s), the null hypothesis is simply
I Pi. = 1, I P.j = 1. (33)

When the null hypothesis is not true, there is said to be interaction between the
two criteria of classification.
The complete parameter space 8 for the distribution of N l l , •.• , N rs has
rs - 1 dimensions (having specified all but one of the Pi)' the remaining one is
fixed by I Pi} = 1), while under :Yf 0 we have a parameter space 8 0 with
i,j
r - 1 + s - 1 dimensions. (The null hypothesis is specified by Pi., i = I, ... , r,
and P.), j = 1, ... , s, but there are only r - 1 + s - I dimensions because
I Pi. = I and I P.j = 1.) The likelihood for a sample of size n is
L=np?y
i, }
and its maximum in (3 occurs when
In 8 0 ,
L = n(Pi. p.)nu = (0 P?:·) (n P~/),

I,) I } (34)
and its maximum occurs at
n.}
and P.} = - . (35)
n
The generalized likelihood-ratio is therefore
(36)
The distribution of A under the null hypothesis is not unique because the hy-
pothesis is composite and the exact distribution of A does involve the unknown
parameters Pi. and P.j; hence, it is very difficult to solve for 20 in sup P 8[A ~ 20] =
eo
ex. For large samples we do have a test, however, because - 2 log A is in that
case approximately distributed as a chi-square random variable with
rs - 1 - (r +s- 2) = (r - 1)(s - 1)
degrees of freedom and on the basis of this distribution a unique critical region
for 1 may be determined. The degrees of freedom rs - 1 - (r + s - 2) is
obtained by subtracting r + s - 2, which is the dimension of 9 0 , from ra - 1,
which is the dimension of 9. Also, (r - 1)(a - 1) is the number of parameters
specified by :Yt0 • (See Theorem 7 and the comment following it.) Actually,
the null hypothesis :Yto: Pij = Pi.P.j is not of the form required by Theorem 7;
so it might be instructive to consider the necessary reparameterization. For
convenience, let us take r = s = 2. Now (3 = {(el , O2 , ( 3 ) = (PH' Po, P21):
Pu > 0; PI2 > 0; P21 ~ 0; and P11 + PI2 + P21 :::;;; I}. Let 8 with points
1
(e~, 0; , ( 3) denote the reparameterized space, where O~ = PI I - Pl. P.I' 0; = PI.,

and OJ = P.I' It can be easily demonstrated that 9' is a one.. to.. one transforma-
tion of 9. Also, the null hypothesis :Yt0: Pu = P1. P.I, Pu = P1.(1 - P.I), and
P21 = (1 - PI )P.I in the original parameter space 9 becomes :Yt~: O~ = 0 and
0; and 0; unspecified in the reparameterized space 9'. [Note that PI2 =
P1.(l - P.I) is equivalent to PI. - PH = PI.(1 - P.I)' which is equivalent to P11 -
P1.P.I = O. Similarly for PZI = (1 - P1.)P.I·] :Yt~ is of the form required by
Theorem 7. In general, a point in the (ra - l)-·dimensional parameter space 9
can be conveniently displayed as
PH PI2 PI,s-1 Ph
P21 PZZ P2, s-1 PZs
P,.-I.I P,.-I,2 Pr-I.s-l Pr-l, s

Prl Pr2 Pr,s-l
and a point in the reparameterized space 9' can be displayed as
PH - P1.P.l P12 - P1.P.2 Pl,8-1 - P1.P., s-1 PI.

P21 - PZ.P.l P22 - PZ.P.2 Pz, s-1 - P2.P., s-1 P2.
P,.-I, 1 - P,.-I,.P.l P,.- I, 2 - p,.- 1 ,.P.2 P,.-I,8-1 - P,.-I,.P.,s-1 P,.-I •.

P.l P.2 P.,s-1
In casting about for a test which may be used when the sample is not large,
we may inquire how it is that a test criterion comes to have a unique distribution
for large samples when the distribution actually depends on unknown param-
eters which may have any values in certain ranges. The answer is that the
parameters are not really unknown; they can be estimated, and their estimates
approach their true values as the sample size increases. In the limit as n becomes
infinite, the parameters are known exactly, and it is at that point that the dis-
tribution of A actually becomes unique. It is unique because a particular point
in 8 0 is selected as the true parameter point, so that the N ij are given a unique
distribution, and the distribution of A is then determined by this distribution.
It would appear reasonable to employ a similar procedure to set up a test
for small samples, i.e., to define a distribution for A by using the estimates for
the unknown parameters. In the present problem, since the estimates of the
Pi. and P.j are given by Eq. (35), we might just substitute those values in the
distribution function of the N lj and use the distribution to obtain a distribution
for A. However, we should still be in trouble; the critical region would depend
on the marginal totals Nt. and N. j ; hence the probability of a Type I error would
vary from sample to sample for any fixed critical region 0 < A < Ao.
There is a way out of this difficulty, which is well worth investigation
because of its own interest and because the problem is important in applied
statistics. Let us denote the joint density of all the N ij briefly by f(nt), the
marginal density of all the Nt. and N. j by g(ni" n.}), and the conditional density
of the N i}, given the marginal totals, by
!(ni})
I
f( ni}ni.,n.j)= ( )
g nt., n.}
Under the null hypothesis, this conditional distribution happens to be inde-
pendent of the unknown parameters (as we shall show presently); the estimators
N i./n and N ,}In form a sufficient set of statistics for the Pt, and P.j' This fact
will enable us to construct a test.
The joint density of the N ij is simply the multinomial distribution
fen/i) =f(nl1, n12, ... , n,,) = It

i. J
nij
! n]l;'Y
t, }
(37)
in e, and in 8 0 (we are interested in the distribution of A under :if0) this becomes
f(n", nt., .•. , n,.) = nn! ,

nij'
(0 P~.i') (I1 p~l)'
~ }
(38)
t, j
To obtain the desired conditional distribution, we must first find the distribution
of the N i. and N.}, and this is accomplished by summing Eq. (38) over all sets
of nil such that
I} = n}
"~ n·· •
and "n,.=n
~} ..
I.
(39)
i j
For fixed marginal totals, only the factor l/flnjj! in Eq. (38) is involved in
the sum; so we have, in effect, to sum that factor over all nij subject to Eq. (39).
The desired sum is given by comparing the coefficients of fl x';'. in the expression
i
(Xl + ... + xrfoJ(xl + ... + x r)n. 2
+ ... + xrt· = (Xl + ... + Xr)n.
••• (Xl s
(40)
On the right-hand side the coefficient of fl Xii. is simply
n!
(41)
fl ni.!·
i
On the left-hand side there are terms with coefficients of the form
n. l ! n.2J n' fl. n.j!

•s· _ =}=--_ (42)
fli
nil! fl ni2!
i i
fl nisI i,flj nij! '
where nij is the exponent of Xi in the jth multinomial. In this expression the
nij satisfy conditions of Eq. (39); the first condition is satisfied in view of the
multinomial theorem, while the second is satisfied because we require the ex-
ponent of Xi in these terms to be ni.. The sum of all such ,coefficients, Eq. (42),
must equal Eq. (41); hence, we may write
n!
(43)
(\I nl.! I} n.,! ) .
This is precisely the sum that we require because there is obviously one and only
one coefficient of the form of Eq. (42) on the left of Eq. (40) for every possible
contingency table, Eq. (31), with given marginal totals. The distribution of the
N i. and N ,} is, therefore,
g( n
i.,
n) -
.} -
t)2
(fl ni.(n!)(f] n.}!)
(fl Pi.ni')(f] P.j'
n. /) (44)
which shows, incidentally, that the N i • are distributed independently of the N. j

under :Yf 0; this is unexpected because N 1. and N. I , for example, have the
random variable Nll in common!
The conditional distribution of the N ij' given the marginal totals, is
obtained by dividing Eq. (38) by Eq. (44) to obtain
(flni. !)(fln.)!)
f(n ll , n12 , ... , nrs I nl.' n2., .. " n.s) = 'fl "
n. nij.
(45)
lr-------------~-----------
which, happily, does not involve the unknown parameters and shows that the
estimators are sufficient.
To see how a test may be constructed, let us consider the general situation
in which a test statistic A for some test has a distribution/A(A; 8) which involves
an unknown parameter 8. If 8 has a sufficient statistic, say T, then the joint
density of A and T may be written
fA,T(A, t; 8) =/AIT(Alt)/T(t; 8),
and the conditional density of A, given T, will not involve 8. Using the condi-
tional distribution, we may find a number, say Ao(t), for every t such that
f A.o(t)
o
/AIT(AI t) dA = .05, (46)
for example. In the At-plane the curve A = Ao(t) together with the line A = 0
will determine a region R. See Fig. 7. The probability that a sample will give
rise to a pair of values (A, t) which correspond to a point in R is exactly .05
because
co
P[(A, T) E R] = J-co J fA, T(A, t; 8) dA dt
A.o(t)
= roo [('ImO. It) d.l.]fT(t; 0) d t
= Jco .05fT(t; 8) dt
- co
=.05.
Hence we may test the hypothesis by using Tin conjunction with A. The
critical region is a plane region instead of an interval 0 < A < Ao; it is such a
region that, whatever the unknown value of 8 may be, the Type I error has a
specified probability. The test in any given situation actually amounts to a

conditional test; we observe T and then perform the test by using the interval
o < A < Ao(t) using the conditional distribution of A, given T. !tis to be observed
that this device cannot be employed unless there is a sufficient statistic for e.
The above technique is obviously applicable when e is a set of parameters
rather than a single parameter and has a set of sufficient statistics. In particular,
the technique may be employed to test the null hypothesis of a two-way con-
tingency table using Eq. (36) to define A. One merely uses the conditional
distribution of Eq. (45) and determines an interval 0 < A < Ao(ni.; n.j) which has
the desired probability of a Type I error for the observed marginal totals.
In applications of this test one is confronted with a very tedious computa-
tion in determining the distribution of A unless r, s, and the marginal totals are
quite small. It can be shown, however, that the large-sample approximation
may be used without appreciable error except when both rand s equal 2. In
the latter instance, other simplifying approximations have been developed (see,
for example, Fisher and Yates, "Tables for Statisticians and Biometricians,"
Oliver & Boyd Ltd., Edinburgh or London, 1938), but we shall not explore the
problem that far.
Another test of the :Yf 0 given in Eq. (33) is obtained if the distribution in
Eq. (45) is replaced by its multivariate normal approximation since then it can
be shown that the statistic
Q = I [Nij - n(Ni./n)(N.j!n)F
i,j n(Ni./n) (N.j!n) (47)
has approximately the chi-square distribution with rs - 1 - (r - 1 + s - 1) =

(r - 1)(s - 1) degrees offreedom. The test criterion is to reject :Yf 0 for large Q.
This is the criterion first proposed (by Karl Pearson) for testing the hypothesis,
and it differs from -2 log A by terms of order I/Jn. The two criteria are
therefore essentially equivalent unless n is small. The argument that Q is a
reasonable test statistic is entirely analogous to that used in Subsec. 5.2 above to
justify Eq. (25). The statistic Q of Eq. (47) has intuitive appeal. Nij is the
observed numberin the ijthcell, and n(N i.!n)(N.j/n) isan estimator of the expected
number in the ijth cell when :Yf 0 is true. Thus, Q will tend to be small for :Yf 0
true and large for :Yf 0 false.
Three-way contingency tables If the elements of a population can be clas-

sified according to three criteria A, B, and C with classifications Ai (i = 1, 2, ... ,
S1)' B j (j = 1, 2, ... , S2), and Ck (k = 1, 2, ... , S3), a sample of n individuals may
be classified in a three-way S1 x S2 X S3 contingency table. We shall let Pijk
represent the probabilities associated with the individual cells and Ntlk be the
numbers of sample elements in the individual cells, and, as before, marginal
totals will be indicated by replacing the summed index by a dot; thus
and (48)
There are four hypotheses that may be tested in connection with this table.
We may test whether all three criteria are mutually independent, in which case
the n utI hypothesis is
Pilk = PI .. P.l.P ..k' (49)
where Pi .. = L L Pilk'
j t
P.j. = Li L Pijk'
k
and P ..k = L L Pijk;
i 1
or we may test
whether anyone of the three criteria is independent of the other two. Thus to
test whether the B classification is independent of A and C, we set up the null
hypothesis
Pijk = Pl.k P.}., (50)
where Pi.k = L
}
Pijk'
The procedure for testing these hypotheses is entirely analogous to that

for the two-way tables. The likelihood of the sample is
L= II p7j,\
i,j. k
(51)
where
L
i. j. k
PUk = 1 and L
i.}. k
nijk = n.
In (3 the maximum of L occurs when
so that
To test the null hypothesis in Eq. (50), for example, we make the substitution
of Eq. (50) into Eq. (51) and maximize L with respect to the PUr. and P.). to find
" nt.k " n .j.

Plk=-
. n and Pj
.. =-,
n
and
sup L =
i!o n
~n (II n7:ir.
I, If.
k
) (II n~l')'
j
(53)
6 TESTS OF HYPOTHESES AND CONFIDENCE INTERV AIS 461
The generalized likelihood-ratio). is given by the quotient of Eqs. (52) and (53),
and in large samples - 2 log A has the chi-square distribution with
8 18 2 83 - 1- [(8183 - 1) + 82 - 1] = (8183 - 1)(82 - 1)
degrees of freedom. Again the large-sample distribution is quite adequate for
many purposes. (8183 -1) +(82 -1)isthedimensionofB o , and 8 18 2 83 -1 is
the dimension of t).
A test statistic analogous to that given in Eq. (47) for testing independence
in a 2 x 2 contingency table can also be derived. F or testing :If 0: A and C
classifications are independent of the B classification, such a test statistic is
Q= L L L [N ijk - n(N u /n)(N. j ./n)]2 .

i j k n(Ni.k/n)(N.J./n ) (54)
Under :lfo , Q has an asymptotic chi-square distribution with 818283 -1

- (8183 - 1) - (82 - 1) = (8183 - 1)(82 - 1) degrees of freedom. Again, the
statistic Q of Eq. (54) has intuitive appeal since N ijk is the observed number in
cell ijk and n(Ni.kln)(N.j.!n) is an estimator of the expected number when :If 0
is true.
6 TESTS OF HYPOTHESES AND

CONFIDENCE INTERVALS
In Subsec. 3.4 above we noted that a confidence interval for a unidimensional
parameter e could be used to obtain a test of :If 0: e = eo versus :If1: e =1= eo.
In this section we will further explore that concept and show that one can
reverse the operation; that is, one can use a family of tests of :If 0: e = eo versus
:lf1 : e =1= eo (the family is generated by varying eo) to obtain a confidence
interval for e. Our considerations in this section will not be very thorough; our
intent is merely to present an introduction to the usefulness of the close relation-
ship between hypothesis testing and confidence intervals.
Our discussion can be made somewhat more general if we speak in terms
of confidence 8et8 rather than confidence interval8. As usual, let X denote the
sample space, e the parameter space, and (Xh ... , x,,) the observed sample.
Definition 18 Confidence set A family of subsets of the parameter

space t) indexed by (Xl' ... , x,,) E X, denoted by 9 = {t)(Xl' ... , x,,):
S(Xl' ... , x,,) c e; (Xl"'" X,,) E X}, is defined to be afamity of confidence
sets with confidence coefficient y if and only if
P9[9(Xh ... , X,,) contains e] = y for all e E S. (55)
1//1
It should be emphasized that any member, say 8(Xl' ... , X,.), of the family
of confidence sets is a subset of 8, the parameter space. 8(X1 , • •• , Xn) is a
random subset; for any possible value, say (Xl' ... , X,.), of (Xl' ... , X,.),
8(Xl' ... , X,.) takes on the value 8(Xl' ... , X,.), a member of the family 9. To
aid in the interpretation of the probability statement in Eq. (55), note that for a
fixed (yet arbitrary) fJ "8 (Xl' ... , X,.) contains fJ" is an event [it is the event that
the random interval 8(Xl' ... , X,.) contains the fixed fJ] and the fJ that appears
as a subscript in P8 is the fJ that indexes the distribution of the X/s appearing
in 8(Xl' ... , X,.).
For instance, suppose Xl' ... , X,. is a random sample from N(fJ, 1).
8 = {fJ; - 00 < fJ < oo}. Let the subset 8(x 1 , ••• , X,.) be the interval
(x - z/Jn, x + z/J~), where z is given by cI>(z) - cI>( -z) = y; then the family of
subsets 9 = {8(Xl' ... , xJ: 8(Xh ... , X,.) = (x - z/J~, x + z/Jn)} is a family
of confidence sets with a confidence coefficient y since
P.[9(X" ... , X.) contains 0] = P.[x - J~ < 0 < X + ~]
= P8 [ -z < X -fJ]
l/Jn < z = y for all fJ E 8.
The family 9 is a family of confidence intervals for fJ having a confidence

coefficient y. In general, then, a confidence interval is an example of a .
confidence set.
Confidence sets can be constructed from tests of hypotheses, as we nOW
show. Let 180 be a size-ct test (nonrandomized) of the null hypothesis .1f0:
fJ = fJo, and let X(fJ o) be the acceptance region of the test 180 • [The acceptance
region is the set complement of the critical region; that is, if the critical region is
given by C(fJ o) , then X(fJ o) = X - C(fJ o)'] Note that X(fJ o) is a subset of X
indexed by fJ o • Since the test 180 has size ct,
P80 [(X1 , ••• , X,.) E X(fJ o)] = 1 - ct.
If we nOw vary fJo over B and for each fJo we have a test 180' then we get a
family of acceptance regions, namely, {X(fJ o): fJo E 8}. X(fJ o) is the acceptance
region of test 180' One can now define
8(Xb ... , Xli) = {fJo: (Xl' ... , X,.) E X(fJ o)}. (56)
Clearly 8(x1 , •• • , X,.) is a subset of 8. Furthermore, the family {B(Xl' ... , X,.)}
is a family of confidence sets with a confidence coefficient y = 1 - ct since One has
{8(Xl' ... , XJ contains fJo} if and only if one has {(Xb ... , X,.) E X(fJ o)}, and so
P80[8(Xl' ... , X,.) contains fJo] = P80 [(X1 , ••• , Xn) E X(fJ o)] = 1- ct.
6 TESTS OF HYPOTHESES AND CONFIDENCE INTERVALS 463
EXAMPLE 25 Let Xl' ... , X" be a random sample from N(fJ, 1), and consider
testing .1f0: fJ = fJ 0 • A test with size rl is given by the following: Reject
.1f0 if and only if Ix - fJo I z/j~, where z is defined by 4l(z) - 4l( -z) =
1 - rl. The acceptance region of this test is given by
30(110) = {(x" ... , x.); 110 - ,in < x< 110 + ,in}.
We can now define, as in Eq. (56),
S(XH ... , xn) = {fJ o: (XH ••• , xn) E X(fJo)}
= {1I0: 110 - ,in <i < 110 + ,in}

= {1I0: x - j;; < 110 < x + ,in}.
S(Xl' ... , xJ is a confidence set (in fact a confidence interval) with a

confidence coefficient y = 1 - rl. ////
The general procedure exhibited above shows how tests of hypotheses can
be used to generate or construct confidence sets. The procedure is reversible;
that is, a given family of confidence sets can be "reverted" to give a test of
hypothesis. Specifically, for a given family {e(Xl' ... , xn)} of confidence sets
with a confidence coefficient y, if we defined
then the nonrandomized test with acceptance region X(fJ o) is a test of .1f0: fJ = fJo
with size rl = 1 - y.
The usefulness of the strong relationship between tests of hypotheses and
confidence sets is exemplified not only in the fact that One can be used to construct
the other but also in the result that often an optimal property of one carries over
to the other. That is, if one can find a test that is optimal in some sense, then
the corresponding constructed confidence set is also optimal in some sense, and
conversely. We will not study the very interesting theoretical result alluded to
in the previous sentence, but we will give the following in order to give some idea
of the types of optimality that can be expected. (See the more advanced books
of Ref. 16 and Ref. 19 for a detailed discussion.) An optimum property of
confidence sets is given in the following definition.
Definition 19 Uniformly most accurate A family {9*(x 1, ... , xn)} of

confidence sets with a confidence coefficient y is defined to be a uniformly
most accurate family of confidence sets at a confidence coefficient y if for
any other family {9(X1' .,., xn)} of confidence sets with a coefficient y
P8[9*(X1, ... , Xn) contains 0'] < P8[9(X1, ... , Xn) contains 0']
for all 0 and 0'. IIII
Definition 19 is saying that 9*(Xb •.. , Xn) is less likely to contain an
incorrect 0' than is 9(X1, •.. , X n), whereas both 9*(X1, ••. , Xn) and 8(X1, . .. , Xn)
have the same probability of containing the correct O. As you may have guessed,
uniformly most accurate confidence sets rarely exist. However, uniformly most
accurate confidence sets within restricted classes of confidence sets could also be
defined, and then one could be hopeful of the existence of such optimal confidence
sets. A general type of result that derives from the close relationship between
tests of hypotheses and confidence sets is the following: If y* is a uniformly most
powerful size-a test of :Yf 0: 0 = 00 within some restricted class of tests, then the
confidence set corresponding to y* is uniformly most accurate with coefficient
y = 1 - a within some restricted class of confidence sets. With such a result
one can see how an optimality of a test can be transferred to an optimality of a
corresponding confidence set, and therein lies the real utility of the close rela-
tionship between hypotheses testing and confidence sets.
7 SEQUENTIAL TESTS OF HYPOTHESES
7.1 Introduction
Sequential analysis refers to techniques for testing hypotheses or estimating
parameters when the sample size is not fixed in advance but is determined during
the course of the experiment by criteria which depend on the observations as they
occur. In this section we propose to consider, and then only briefly, one form
of sequential analysis, namely, the sequential probability ratio test.
In Sec. 2 above we considered testing the simple null hypothesis :Yf 0: 0 = 00
versus the simple alternative hypothesis :Yf1: 0 = 01 , It was shown (Neyman-
Pearson lemma) that for samples of fixed size n, the test which minimized the
size, say p, of the Type II error for fixed size, say a, of the Type I error was a
simple likelihood-ratio test. That is, for fixed n and a, p was minimized.
Suppose now that it is desired to fix both a and P in advance and then find that
simple likelihood-ratio test having minimum sample size n and having size of
Type I error equal to a and size of Type II error p. The solution of such a
problem is illustrated in the following example.
7 SEQUENTIAL TESTS OF HYPOTHESES 46S
EXAMPLE 26 A manufacturer of a certain component, say, an oil seal, knows

the history of his current manufacturing process. He knows, for instance,
that the distribution of lifetimes of the seals now being manufactured is,
say, N(lOO, 100). A new manufacturing process is suggested; the manu-
facturer wants to continue with his present manufacturing process if the new
process is not better (longer mean lifetime), yet he also wants to be quite
certain to switch to the new process if the new process increases the mean
lifetime by, say,S percent. He proposes to take a sample of observations
of lifetimes of seals made by the new process and then from the sample
decide whether or not the process has longer mean life. He models the
experiment by assuming that the random variable X, representing the life-
time of a seal manufactured using the new process, is distributed as
NCO, 100), and he wants to test .1P0: 0 ::::; 100 versus .1P1: 0 > 100. He
fixes his error sizes and wants to determine the sample size n so that, say,
.01 = ex = P8=100[reject .1P0] and .05 = fJ = P 9 =10S[accept :Yf 0]'

That is, he seeks to determine n so that there is only a 1 percent chance of
rejecting that the new process is no better than the old when it is not, yet
there is a 95 percent chance of rejecting that the mean lifetime of the new
process is less than 100 when in fact it is 5 percent larger. It can be shown
that the simple likelihood-ratio test is equivalent to the test of rejecting
:Yf 0 for large X n • Thus he seeks to determine n and k so that
and
or
.01 = 1_ <f>(k - 100) and .05 = <f> (

k - 105)
! / •
IO/y' n IO/yn
<f>(k - 100) = .99

IOIJn
implies
k -100
r: ~ 2.326,
10/yn
and
k - 105)
<f> (
10
1)1, = .05
implies
k - 105
;: ~ -1.645;
lO/y n
which together imply that 100 + 10(2.326)IJ~ ~ 105 - 10(1.645)IJn) or

n ~ 63.08; so a sample of size 64 is needed. II I I
Referring to the above example, the following considerations make se-

quential analysis interesting both from the theoretical and practical viewpoint.
In drawing the 64 observations to test .Yf0, it is possible that among the first
few observations, say 20, 30, or 40, the evidence is quite sufficient relative to ct
and P for accepting or rejecting .Yf 0, and then observing additional observa-
tions would be a waste of time and effort. In other words, the possibility is
raised that, by constructing the test in a fashion which permits termination of
the sampling at any observation, one can test .Yf0 with fixed error sizes ct and P
and yet do so with fewer than 64 observations on an average. This is in fact
the case; although it may at first appear surprising in view of the fact that the
best test for fixed sample size requires 64 observations. The saving in observa-
tions is often quite large, sometimes as much as 50 percent! We will study such
a sequential procedure in the remaining subsections.
7.2 Definition of Sequential Probability Ratio Test
Consider testing a simple null hypothesis against a simple alternative hypothesis.

In other words, suppose a sample can be drawn from one of two distributions
(it is not known which one) and it is desired to test that the sample came from
one distribution against the possibility that it came from the other. If Xl'
X 2' ••. denotes the random variables, we want to test .Yf 0: Xi "-' lo( .) versus
.Yf1: Xi "-' 11 (. ). The simple likelihood-ratio test was of the following form:
Reject .Yf 0 if .< = ~: < k for some constant k > o.

The sequential test that we propose to consider employs the likelihood-ratios
sequentially. Define
7 SEQUENTIAL TESTS OF HYPOTHESES 467
for m = 1, 2, ... , and compute sequentially )"h A2' ... ,. For fixed ko and kl
satisfying 0 < ko < k1' adopt the following procedure: Take observation Xl and
compute A1 ; if A1 s ko, reject :Yf 0; if A1 > kh accept :Yf0; and if ko < )"1 < k1'
take observation X2, and compute A2 • If A2 s ko, reject :Yf 0; if A2 > k1'
accept :Yf 0; and if ko < A2 < kh observe X3' etc. The idea is to continue
sampling as long as ko < A) < k1 and stop as soon as Am s ko or Am ~ k1'
rejecting :Yf 0 if Am < ko and accepting :Yf 0 if Am > k 1 • The critical region of
ao
the described sequential test can be defined as C = U Cn , where
n=1
)"n(Xh ••. , xn) S k o}. (58)
A point in Cn indicates that :Yf 0 is to be rejected for a sample of size n. Sim-

ao
ilarly, the acceptance region can be defined as A = U An' where
n= 1
Definition 20 Sequential probability ratio test F or fixed 0 < ko < k1' a test
as described above is defined to be a sequential probability ratio test. IIII
When we considered the simple likelihood-ratio test for fixed sample size
n, we determined k so that the test would have preassigned size ct. We now
want to determine ko and k1 so that the sequential probability ratio test will have
preassigned ct and {J for its respective sizes of the Type I and Type II errors.
Note that
ct = P[reject :Yf 0 1:Yf 0 is true] = I

n=1
00
f Lo(n)
en
(60)
and
{J = P[accept :Yf 0 1:Yf 0 is false] = I
00
n=1
f L (n),
An
1 (61)
where, as before, Sen Lo(n) is a shortened notation for J... J

en
[Ii 1=1
!o(Xi ) dXi]'
For fixed ct and {J, Eqs. (60) and (61) are two equations in the two unknowns ko
and k 1 • (Both An and Cn are defined in terms of ko and k 1 .) A solution of
these two equations would give the sequential probability ratio test having the
desired preassigned error sizes ct and {J. As might be anticipated, the actual
determination of ko and kl from Eqs. (60) and (61) can be a major computational
project. In practice, they are seldom determined that way because a very simple
and accurate approximation is available and is given in the next subsection.
We note that the sample size of a sequential probability ratio test is a
random variable. The procedure says to continue sampling until An =
An(Xl' ... , xn) first falls outside the interval (ko, k l ). The actual sample size then
depends on which Xi'S are observed; it is a function of the random variables
Xl, X 2 , ••• and consequently is itself a random variable. Denote it by N.
Ideally we would like to know the distribution of N or at least the expectation
of N. (The procedure, as defined, seemingly allows for the sampling to continue
indefinitely, meaning that N could be infinite. Although we will not so prove,
it can be shown that N is finite with probability 1.) One way of assessing the
performance of the sequential probability ratio test would be to evaluate the
expected sample size that is required under each hypothesis. The following
theorem, given without proof (see Lehmann [16]), states that the sequential
probability ratio test is an optimal test if performance is measured using expected
sample size.
Theorem 10 The sequential probability ratio test with error sizes a and
/3 minimizes both G[NI Jt o is true] and G[NI Jt l is true] among all tests
(sequential or not) which satisfy the following: P[Jt 0 is rejected I Jt 0 is
true] < a, P[Jt 0 is accepted I Jt 0 is false] < /3, and the expected sample
size is finite. // //
Note that in particular the sequential probability ratio test requires fewer
observations on the average than does the fixed-sample-size test that has the
same error sizes. In Subsec. 7.4 we will evaluate the expected sample size for
the example given in the introduction in which 64 observations were required
for a fixed-sample-size test with preassigned a and /3.
7.3 Approximate Sequential Probability Ratio Test

We noted above that the determination of ko and kl that defines that particular
sequential probability ratio test which has error sizes a and P is in general
computationally quite difficult. The following remark gives an approximation
to ko and k l ·
Remark Let ko and kl be defined so that the sequential probability
ratio test corresponding to ko and kl has error sizes a and /3; then ko and
kl can be approximated by, say, k~ and k~, where
a 1- a
k~ = (62)
k~ = 1- /3 and
P
.
7 SEQUENTIAL TESTS OF JIYPOTHESES 469
PROOF (Assume J, P[N = n IJt';] = 1 for i = 0, 1.)

(1. = P[reject -*' 0I-*'o is true] =
00
L f Lo(n) < L f koL1(n)

n= 1 en
00
n= 1 en
L fenLl(n) =
00
= ko ko P[reject -*'0 I-*' 1 is true]

n= 1
= k o(1- P),
and hence ko > (1.1( I - P). Also
1- (1. = P[accept -*'0I-*'o is true)
= k1P[accept -*'0 I-*' 1 is true] = kIP,
and hence kl < (1 - (1.)IP. Note that the approximations k~ = (1.10 - P)

and k~ = (1 - (1.)IP satisfy
'
k 0=1 (1.
_ P< k 0 < k 1 ~ 1 -P (1. = k'
1- (63)
1/11
Remark Let (1.' and P' be the error sizes of the sequential probability
ratio test defined by leO and k't given in Eq. (62). Then (1.' + P' < (1. + p.
Let A' and C' (with corresponding A~ and C~) denote the
PROOF
acceptance and critical regions of the sequential probability ratio test
defined by ko
and k 1. Then
and
1 - (1.' = L
00 f Lo(n) > 1-(1.
I P L f , L 1(n) = 1-(1.
00
P P';
n;;:; 1 An n 1 An ~
hence (1.'(1 - P) (1.(1 - P'), and (1 - (1.)P' < (l - (1.')P, which together
implythat(1.'(1- P) + (1 - (1.)P' ~ (1.(1- P') + (1- (1.')por(1.' + P' < (1. + p.
///1
Naturally, one would prefer to use that sequential probability ratio test
having the desired preassigned error sizes a and P; however, since it is difficult
to find the ko and kl corresponding to such a sequential probability ratio test,
instead one can use that sequential probability ratio test defined by ko and kl of
Eq. (62) and be assured that the sum of the error sizes a' and P' is less than or
equal to the sum of the desired error sizes a and p.
7.4 Approximate Expected Sample Size of

Sequential Probability Ratio Test
The procedure used in performing a sequential probability ratio test is to continue
sampling as long as ko < Am < kl and stop sampling as soon as Am < ko or
Am > k 1. If Zi = loge [!O(Xi)!.fl(Xi)], an equivalent test is given by the following:
m
Continue sampling as long as loge ko < L Z i < loge kl' and stop sampling as soon
1
m m
as L Zi < loge ko (and then reject *'0) or L Zi > loge kl (and then accept *'0)'
1 1
As before, letN be the random variable denoting the sample size of the sequential
probability ratio test, and let Zi = loge [fo(Xi)/Ji(Xi)]. Equation (64), given in
the following theorem, is useful in finding an approximate expected sample size
of the sequential probability ratio test.
Theorem 11 Wald's equation Let ZI, Z2, ... , Zn, ... be independent
identically distributed random variables satisfying &[ IZi I] < 00. Let N
be an integer-valued random variable whose value n depends only on the
values of the first n Zi'S. Suppose &[N] < 00. Then
&[ZI + ... + Z N] = &[N] . &[ZJ (64)
PROOF &[ZI + ... + ZN] = &[&[ZI + ... + ZNIN]]
00
= L &[ZI +' ,,+ZnIN=n]P[N=n]

n=1
00 n
= L L
n=1 i=1
&[ZIIN = n]P[N = n]
00 00
= L Li
i= 1 n=
&[Zd N = n]P[N = n]
00
= L &[Zi IN > i]P[N > i]

i== 1
00
= L &[Zi]P[N > i]
i= 1
00
= &[Z;1 L P[N > i]

i= 1
= &[Zi]&[N].
7 SEQUENTIAL TESTS OF HYPOTHESES 471
($[Zi] = $[ZiI N > i]sincetheevent {N > i} depends only onZ1 , ""Z£-1

00
and hence is independent of Z, . Also $ [N] = L P[N;;::: i] follows from

i= 1
Eq. (6) of Chap. II.) II!!
If the sequential probability ratio test leads to rejection of J'e0' then the
random variable ZI + ... + ZN < lo~ ko, but ZI + ... + ZN is close to lo&: ko
since ZI + ... + ZN first became less than or equal to loge ko at the Nth observa-
tion; hence $[ZI + ... + ZN] ~ loge ko . Similarly, if the test leads to acceptance,
$[ZI + ... + ZN] ~ logekl ;hence$[ZI + ... + ZN] ~ plogeko + (1 - p)lo&:kl'
where p = P[ J'e 0 is rejected]. Using
_ 8[ZI + ... + ZN] I'V P lo&: ko + (l - p) loge kl

[
$ N] - $[Zi] I'V $[Zl]
we obtain
D[N I - v P ' ] I'V rx loge ko + (I - rx) loge kl

.:n 0 IS true
$[Zj I J'e0 IS true]
@ I'V •
rx loge [rx!(1 - P)] + (1

- rx) loge [(1 - rx)!P]
I'V _________________ ~--------------- _ __ _ _
(65)
'" $[Zil J'eo is true] ,
and
. (1 - P) loge ko + Ploge kl
$[N IJ'e0 IS false] ~ $[Z, I J'e0 is false]
,....,
(l - P) loge [rxl(l - P)] + P lo~ [(1 - rx)IP]
I'V
(66)
$[Zd J'e 0 is false]
EXAMPLE 27 Consider sampling from N(6, q2), where q2 is assumed known.

Test J'e 0: 6 = 60 versus J'e 1 : 6 = 61 , Now
hence
8[Zd -*' 0 is true] = - ~

2(1
[(8~ - (1) - 280 (80 - ( 1)]
1 2
= 2(12 (8 1 - ( 0) ,
and
For a; = .01, P= .05, (12 = 100,80 = 100, and 8 1 = 105 (as in Example 26),
Eq. (65) reduces to
8J[NI1I',p' ] '" .01 loge (.01/.95) + .99 loge (.99/.05) ,..., 24

({) en 0 IS true '" 25/200 '"
and
8[N I-*'o is false] ~ 34.
The average sample sizes of 24 and 34 for the sequential probability ratio
test compare to a sample size of 64 for the fixed-sample-size test. 1///
PROBLEMS
1 Let X have a Bernoulli distribution, where P[X = 1] = () = 1 - P[X = 0].
(a) For a random sample of size n = 10. test -*'0: () i. Use
the critical region {L x, 6}.
(i) Find the power function, and sketch it.
(ii) What is the size of this test?
(b) For a random sample of size n 10:
(i) Fmd the most powerful size-o: (0: .0547) test of -*'0: e= i versus
-*'1: () = 1.
(ii) FIlld the power of the most powerful test at 1. e
(c) For a random sample of size 10, test -*'0: () i versus -*'1: e= 1.
(i) Fmd the minimax test for the loss function 0 = t(do ; eo) = t(d1 ; ( 1 ),
{(do; e•.) = 1719, t(d.; eo) 2241.
(ii) Compare the maximum risk of the minimax test with the maximum risk
of the most powerful test given in part (b).
(d) Again, for a sample of size 10, test -*'0: () 1 versus -*'1: = 1. Use the e
above loss function to find the Bayes test corresponding to prior probabilities
given by
9 = (1719/2241)0)10 + 34 '
PROBLEMS 473
2 Let X have the density f(x; 8) 8x'- 1 1(o.1)(X).

(a) To test.:,y0: 8 < 1 versus Jf'1 : 8 > 1, a sample of size 2 was selected, and the
critical region C {(Xl, X2): 3/4x. < X2} was used. Find the power function
and size of this test.
(b) For a random sample of size 2, find the most powerful size-[<x = 1(1 - In 2)]
test of Jf' 0: 8 = 1 Vel\.:15 Jf'1: 8 2.
(c) Are the tests that you ob~ined in parts (a) and (b) unbiased?
(d) For a random sample of size 2, find the minimax test of Jf' 0: 80 = 1 versus
Jf'1: 8. = 2 using the loss function {(do; 80 ) = {(d1 ; 81) 0, {(do; 81) =
1 - loge2, {(d1 ; 80 ) 1 + 108e2.
(e) For a random sample of size n, find the Bayes test corresponding to prior
probabilities given by g = i of Jf' 0: 8 = 1 versus Jf'1: 8 = 2 using the loss
function {(do; 80 ) {(d1 ; 81) = 0, {(do; 81) = 1, {(d1 ; 80 ) = 2.
(j) Test Jf' 0: 8 = 1 versus Jf'.: 8 = 2 using a sample of size 2. Let a. = size of
Type I error and fJ size of Type II error. Find the test that minimizes the
largest of <X and fJ.
3 Let e = {l, 2}, and suppose you have one observation from the density
1(6-i-.6+i-)(x). Show that a test that has uniformly smallest risk among all tests
exists, and find it.
4 Let X be a single observation from the density
f(x; 8) 8x'- 1 1(o.1)(x),
where 8 >0.
(a) In testing Jf' 0: 8 <1 versus Jf'1: 8 > 1, find the power function and size of
the test given by the following: Reject Jf' 0 if and only if X> 1.
(b) Find a most powerful size-<x test of Jf' 0: 8 2 versus Jf'1: 8 = 1.
(c) For the loss function given by {(do; 2) = {(d.; 1) 0, {(do; 1) = {(d1 ; 2) 1,
find the minimax test of Jf' 0: 8 = 2 versus Jf'1: 8 = 1.
(d) Is there a uniformly most powerful size-a. test of Jf' 0: 8 > 2 versus ~1: 8 < 2?
If so, what is it?
(e) Among all possible simple likelihood-ratio tests of Jf' 0: 8 = 2 versus Jf'1:
8 = 1, find that test that minimizes <X + fJ, where <X and fJ are the respective
sizes of the Type I and Type II errors.
(f) Find the generalized likelihood-ratio test of size <X of Jf' 0: 8 = 1 versus
Jf'.:8#;1.
5 Let X be a single observation from the density f(x; 8) = (28x + 1 - 8)1[0. 11(X),
where 1 <8<1.
(a) Find the most powerful size-a. test of Jf' 0: 8 = 0 versus Jf'1: 8 = 1. (Your
test should be expressed in terms of <x.)
(b) To test Jf' 0: 8 <0 versus Jf'1: 8 0, the following procedure was used:
Reject :?e0 if X exceeds 1. Find the power and size of this test.
(c) Is there a uniformly most powerful size-a. test of Jf' 0: 8 <0 versus Jf' 1~ 8 > O?
If so, what is it?
I
(d) What is the generalized likelihood-ratio test of :¥f 0: 8 = 0 ve~us :¥f I! 8 ¢ O?
(e) Among all possible simple likelihood-ratio tests of :¥f 0: (J = 0 versus :¥f I:
8 = 1 find that test which minimizes 0'. + {3, where 0'. and (3 are the respective
(f) Given a set of observations. all of which fall between 0 and I, indicate how
you would test the hypothesis that the ob~ ,~vations came from the density
f(x; 8).
6 Let Xl, ..• , Xn denote a random sample from f(x; 8) = (1/8)1(0. 6)(X), and let
Y l , ... , Y n be the corresponding ordered sample. To test :¥f 0: 8 = 80 versus
:¥f I: 8 ¢ 80 , the following test was used: Accept :¥f 0 if 80 ( \Y;;) < Yn < 80 ;
otherwise reject.
(a) Find the power function for this test, and sketch it.
(b) Find another (nonrandomized) test that has the same size as the given test,
and show that the given test is more powerful (for all alternative 8) than the
test you found.
7 Let Xl, •.• , Xn denote a random sample from
f(x; 8) =( 1/8)x(I-6)/61(0. l)(x).
Test :¥f 0: 8 < 80 versus :¥f I: 8> 80 •

(a) For a sample of size n, find a uniformly most powerful (UMP) size-O'. test if
such exists.
(b) Take n = 2, 8 0 = 1, and 0'. = .05, and sketch the power function of the UMP
test.
8 Let Xl, .'., Xn be a random sample from the Poisson distribution
e- 6 8J<
f(x; 8) = -,-l{o.l. 2 • ••• )(x).
x.
(a) Find the UMP test of :¥f 0: 8 = 80 versus :¥f I: 8 > 80 , and sketch the power
function for 80 = 1 and n = 25. (Use the central-limit theorem. Pick
0'. = .05.)
(b) Test:¥f 0: 8 = 80 versus :¥fl: 8 ¢ 80 • Find the general form of the critical
region corresponding to the test arrived at using the generalized likelihood-
ratio principle. (The critical region should be defined in terms of L Xl.)
(c) A reasonable test of :¥f 0: 8 = 80 versus :¥f I: 8 ¢ 80 would be the following:
Reject if I X - 80 I >K. For 0'. = .05, find K so that P[reject :K 0 I:¥f 0] = .05.
(Assume that n is large enough so that the central-limit theorem can be used
to find an approximation to K.)
9 Let e = {80 , 81 }. Show that any test arrived at using the generalized likelihood-
ratio principle is equivalent to a simple likelihood-ratio test.
10 To test :¥f 0: 8 < 1 versus :¥f I : 8 > 1 on the basis of two observations, say Xl and
X 2 , from the uniform distribution on (0, 8), the following test was used:
Reject:¥f 0 if
PROBLEMS 475
(a) Find the power function of the above test, and note its size. [Recall that
Xl + X 2 has a triangular distribution on (0, 28).]
(b) Find another test that has the same size as the given test but has greater power
for some () > 1 if such exists. If such does not exist, explain why.
11 Let Xl, ••. , Xn be a random sample of size n fromf(x; 8) = ()2 xe - 61cl(0. CO)(x).
(a) In testing .1f 0: () < 1 versus .1f1 : () > 1 for n = 1 (a sample of size 1) the
following test was used: Reject .1f 0 if and only if Xl < 1. Find the power
function and size of this test.
(b) Find a most powerful size-a test of .1f 0: () = 1 versus .1f1: () = 2.
(c) Does there exist a uniformly most powerful size-a test of .1f 0: () < 1 versus
.1f1: () > I? If so, what is it?
(d) In testing .1f 0: () = 1 versus .1f 1: () = 2, among all simple likelihood-ratio
tests find that test which minimizes the sum of the sizes of the Type I and Type
II errors. You may take n = 1.
12 Let Xl, .•. , Xn be a random sample from the uniform distribution over the interval
«(), () + 1). To test .1f 0: () = 0 versus .1f1: () > 0, the following test was used:
Reject .1f 0 if and only if Yn > 1 or Y 1 > k, where k is a constant.
(a) Determine k so that the test wiIl have size a.
(b) Find the power function of the test you obtained in part (a).
(c) Prove or disprove: If k is selected so that the test has size a, then the given
test is uniformly most powerful of size a.
13 Let X .. ••. , Xm be a random sample from the density ()lxB 1 - 11(0.o(x), and let
Y 1, "', Yn be a random sample from the density ()2y fJ2 -11(0, 1)(y). Assume that
the samples are independent. Set Vf = -logeXf' i = 1, ... , m, and VJ =
-loge YJ, j = 1, •.. , n.
(a) Find the generalized likelihood-ratio for testing .1f 0: ()1 = ()2 versus
.1f 1 :()1¢()2.
(b) Show that the generalized likelihood-ratio test can be expressed in terms of
the statistic
L Vf
T - =-....;;;;..;.----==---
-LU,+LV/
(c) If.1f 0 is true, what is the distribution of T? (You do not have to derive it if
you know the answer.) Does the distribution of T depend on () = ()1 = ()2
given that .1f 0 is true?
14 Find a genera1ized likelihood-ratio test of size a: for testing .1f0: () < 1 versus
.1f 1: () > 1 on the basis of a random sample Xl, "', Xn from f(x; 8) =
()e- b l(o. CO)(x).
15 Let X be a single observation from the density f(x; 8) = (1 + 8)xBl(o. 1)(x), where
(»-l.
(a) Find the most powerful size-a: test of .1f 0: () = 0 versus .1f 1: () = 1.
(b) Is there a uniformly most powerful size-a: test of.1'f'0 : () < 0 versus .1f1 : () > O?
If so, what is it?
(c) Among all possible simple likelihood-ratio tests of :K 0: () 0 versus :K.:

() = 1 find a test which minimizes 2/X + p, where /X and p are the respective
(d) Find a generalized likelihood-ratio test of :K 0: () = 0 versus :K 1: () if; 0..
16 Let X., ••• , Xm bea random sample from ().e- 6,J(I(o. CCI)(X), and let Y., ••• , Yn be a
random sample from ()2 e- 6271(0. CCI)(y)' Assume that the samples are independent.
(a) Find the generalized likelihood-ratio for testing :K 0: ()1 = ()2 versus
:K.: ()l =f; ()2 •
(b) Show that the generalized likelihood-ratio test can be expressed in terms of
the statistic T = .2 X,f(.2 X, + .2 YJ). Argue (or show) that the distribution
of T does not depend on () = (). = ()2 when :K 0 is true.
17 Use the confidence-interval technique to derive a test of :K 0: P-l = P-2 versus
:K 1: p-. =f; P-2 in sampling from the bivariate normal distribution. Such a test is
often called a paired t test. (See the last paragraph in Subsec. 3.4 in Chap. VIII.)
18 Given the sample (- .2, - .9, - .6, .1) from a normal population with unit variance,
test whether the population mean is less than 0 at the .05 level (Le., with probability
.05 of a Type I error). That is, test :K 0: p- 0 at the .05 level versus :K. : p- > O.
19 Given the sample (-4.4, 4.0, 2.0, - 4.8) from a normal population with variance 4
and the sample (6.0, 1.0, 3.2, - .4) from a normal population with variance 5, test
at the .05 level that the means differ by no more than one unit. Plot the power
function for this test. Plot the ideal power function.
20 A metallurgist made four determinations of the melting point of manganese:
1269, 1271, 1263, and 1265 degrees centigrade. Test the hypothesis that the mean
II. of this population is within 5 degrees centigrade of the published value of 1260.
Use /X = .05. (Assume normality and (12 5.)
21 Plot the power function for a test of the null hypothesis :K 0: - 1 < p- < 1 for a
normal distribution with known variance using sample sizes 1, 4, 16, and 64. (Use
the standard deviation (J as the unit of measurement on the fL axis and .05 proba-
bility of Type I error.) Plot the ideal power function.
22 Let Xl, ... , Xn be a random sample of size n from a normal density with known
variance. What is the best critical region for testing the null hypothesis that the
mean is 6 against the alternative that the mean is 4 ?
23 Derive a test of :K 0: (J2 < 10 against :K1: (J2 > 10 for a sample of size n from a nor-
mal population with a mean of O.
24 In testing between two values 11.0 and fLl for the mean of a normal population,
show that the probabilities for both types of error can be made arbitrarily small
by taking a sufficiently large sample.
25 A cigarette manufacturer sent each of two laboratories presumably identical
samples of tobacco. Each made five determinations of the nicotine content in
milligrams as follows: (i) 24, 27, 26, 21, and 24 and (ii) 27, 28, 23, 31, and 26.
Were the two laboratories measuring the same thing? (Assume normality and a
common variance.)
PROBLEMS 477
26 The metallurgist of Prob. 20, after assessing the magnitude of the various errors
that might accrue in his experimental technique, decided that his measurements
should have a standard deviation of 2 degrees centigrade or less. Are the data
consistent with this supposition at the .05 level? (That is, test .:¥f0: a <2.)
27 Test the hypothesis that the two samples of Prob. 19 came from populations with
the same variance. Use Ct = .05.
28 The power function for a test that the means of two normal populations are equal
depends on the values of the two means ILl and 1'2 and is therefore a surface. But
the value of the function depends only on the difference (1 = 1'1 - 1'2, so that it
can be adequately represented by a curve, say {3(U). Plot {3(U) when samples of 4
are drawn from one population with variance 2 and samples of 2 are drawn from
another population with variance 3 for tests at the .01 level.
29 Given the samples (1.8, 2.9, 1.4, 1.1) and (5.0, 8.6,9.2) from normal populations,
test whether the variances are equal at the .05 level.
30 Given a sample of size 100 with X = 2.7 and 2: (X, X)2 225, test the null
hypothesis .:¥f 0: IL 3 and a 2 = 2.5 at the .01 level, assuming that the population
is normal.
31 Using the sample of Prob. 30, test the hypothesis that I' a 2 at the .01 level.
32 Using the sample of Prob. 30, test at the 0.1 level whether the .95 quantile point,
say ~ = ~.95, of the population distribution is 3 relative to alternatives ~ < 3.
Recall that ~ is such that J~ ~ f(x) dx = .95, where f(x) is the population
density; it is, of course, I' + I.645a in the present instance where the distribution is
assumed to be normal.
33 A sample of size n is drawn from each of k normal populations with the same
variance. Derive the generalized likelihood-ratio test for testing the hypothesis
that the means are all O. Show that the test is a function of a ratio which has
the F distribution.
34 Derive the generalized likelihood-ratio test for testing whether the correlation of
a bivariate normal distribution is O.
35 If XI, X 2 , ••• , XII are observations from normal populations with known variances
ai, vi, ... , a~, how would one test whether th~ir means were all equal?
36 A newspaper in a certain city observed that driving conditions were much improved
in the city because the number of fatal automobile accidents in the past year was 9
whereas the average number per year over the past several years was 15. Is it
possible that conditions were more hazardous than before? Assume that the
number of accidents in a given year has a Poisson distribution.
37 Six 1-foot specimens of insulated wire were tested at high voltage for weak spots
in the insulation. The numbers of such weak: spots were found to be 2, 0, 1, 1, 3,
and 2. The manufacturer's quality standard states that there are less than 120
such defects per 100 feet. Is the batch from which these specimens were taken
worse than the standard at the .05 level ? (Use the Poisson distribution.)
38 Consider sampling from the normal distribution with unknown mean and variance:
(a) Find a generalized likelihood-ratio test of :K 0: u 2 < u~ versus :K I: u 2 > u& •
(b) Find a generalized likelihood-ratio test of :K 0: u 2 u6 versus :K I: u 2 :f:. u~.
39 (a) Suppose (Nit 0, Nt) is multinomially distributed with parameters n,
o.
PI, ..• ,p", where NUl = n- N. ~.,. N"andp"+l = 1 "

2PJo Theorem
1
8 states that
has a limiting chi-square distribution. Find the exact mean and variance
of Q.
(b) Let (Nt, • 0 N,,) be distributed as in part (a). Define
.,
[See Eq. (25).] Find 8[Qf]. [See Eq. (26).] Is 8[Q~] for PI = p~, .•. ,
PUI = P~+l less than or equal to 8[Qf] for arbitrary Ph ... , p,,+!?
40 A psychiatrist newly employed by a medical clinic remarked at a staff meeting that
about 40 percent of all chronic headaches were of the psychosomatic variety. His
disbelieving colleagues mixed some pills of plain flour and water, giving them to
all such patients on the clinic's rolls with the story that they were a new headache
remedy and asking for comments. When the comments were all in they could be
fairly accurately classified as follows: (i) better than aspirin, 8, (ti) about the same
as aspirin, 3, (iii) slower than aspirin, 1, and (iv) worthless, 29. While the doctors
were somewhat surprised by these results, they nevertheless accused the psychiatrist
of exaggeration. Did they have good grounds?
41 A die was cast 300 times with the following results:
Occurrence: 1 2 3 4 5 6
Frequency: 43 49 56 45 66 41
Are the data consistent at the ,05 level with the hypothesis that the die is true?
42 Of 64 offspring of a certain cross between guinea pigs, 34 were red, 10 were black,
and 20 were white. According to the genetic model, these numbers should be in
the ratio 9/3/4. Are the data consistent with the model at the .05 level?
43 A prominent baseball player's batting average dropped from .313 in one year to
.280 in the following year. He was at bat 374 times during the first year and 268
times during the second. Is the hypothesis tenable at the .05 level that his hitting
ability was the same during the two years?
44 Using the data of Prob. 43, assume that one has a sample of 374 from one Bernoulli
population and 268 from another. Derive the generalized likelihood-ratio test
for testing whether the probability of a hit is the same for the two populations.
How does this test compare with the ordinarY test for a 2 x 2 contingency table?
PROBLEMS 479
45 The progeny of a certain mating were classified by a physical attribute into three
groups, the numbers being 10 53, and 46. According to a genetic model the
9
frequencies should be in theratios p 2j2p{1-p)j{l-p)2. Are the data consistent

with the model at the .05 level?
46 A thousand individuals were classified according to sex and according to whether
or not they were color-blind as follows:
Male Female
Normal 442 514

Color-blind 38 6
According to the genetic model these numbers should have relative frequencies
given by
P p2
2
-+pq
2
q q2
-
2 2
where q = 1 - p is the proportion of color-blind individuals in the population.

Are the data consistent with the model ?
47 Treating the table of Prob. 46 as a 2 x 2 contingency table, test the hypothesis that
color blindness is independent of sex.
48 Gilby classified 1725 school children according to intelligence and apparent family
economic level. A condensed classification follows:
Dull Intelligent Very capable
Very weB c10thed 81 322 233

Well c10thed 141 457 153
Poorly c10thed 127 163 48
Test for independence at the .01 level.

49 A serum supposed to have some effect in preventing colds was tested on 500 indi-
viduals, and their records for 1 year were compared with the records of 500 un-
treated individuals as follows:
More than
No colds One cold one cold
Treated 252 145 103

Untreated 224 136 140
Test at the .05 level whether the t~o trinomial populations may be regarded as the
same.
50 According to the genetic model the proportion of individuals having the four
blood types should be given by:
0: q2
A:p2+2pq
+
B: r2 2qr
AB: 2pr
where p + q + r = I. Given the sample 0, 374; A, 436; B, 132; AB, 58; how
would you test the correctness of the model?
51 Galton investigated 78 families, classifying children according to whether or not
they were light-eyed, whether or not they had a light-eyed parent, and whether or
not they had a light-eyed grandparent. The following 2 x 2 x 2 table resulted :
Grandparent
Light Not
Parent
Ught Not Light Not
'0
:aU Ught
Not
1928
303
552
395
596
225
508
501
Test for complete independence at the .01 level. Test whether the child classifica-
tion is independent of the other two classifications at the .01 level.
52 Compute the exact distribution of A for a 2 x 2 contingency table with marginal
totals Nt. = 4, N 2. = 7, N.1 = 6, N.2 = 5. What is the exact probability that
- 21o&e A exceeds 3.84, the .05 level of a chi-square distribution for one degree of
freedom?
53 In testing independence in a 2 x 2 contingency table, find the exact distribution of
the generalized likelihood-ratio for a sample of size 2. Do the same for samples
of size 3 and 4. Discuss.
54 Let Xl, •.• , Xn be a random sample from N(I-', a 2 ), where a 2 is known. Let A
denote the generalized likelihood-ratio for testing Jf' 0: I-' = 1-'0 versus :Yf 1: I-' :f:. 1-'0'
Find the exact distribution of - 2 lo&eA, and compare it with the corre-
sponding asymptotic distribution when Jf' 0 is true. HINT: L (X, - X)2 =
'"
L. (X, - I-')2 - n(X
- - 1-') 2•
55 Here is an actual sequence of outcomes for independent Bernoulli trials. Do you
think p (the probability of success) equals l?
s I I I s, sIs I I, s I I I I, I I sIs, s I I I I,
s I Is/, I I I I s, I I s I /, s I I I I, s s I I I.
PROBLEMS 481
~.
If you do not think p is 1, what do you think p is? Give a confidence-interval
estimate of p. If the above data were generated by tossing two dice, then what
would you think p is? If the data were generated by tossing two coins, then what
would you think pis? (If the data were generated by tossing two dice, assume that
the possible values of pare j/36, j 0, ... , 36. If the data were generated by
tossing two coins, assume that the possible values of p arej/4,j = 0, ... ,4.)
56 In sampling from a Bernoulli distribution, test the null hypothesis that p = t
against the alternative that p = t. Let p refer to the probability of two heads
when tossing two coins, and carry through the test by tossing two coins, using
(X = f3 .10. (The alternative was obtained by reasoning that tossing two coins
can result in the three outcomes: two heads, two tails, or one head and one tail,
and then assuming each of the three outcomes equally likely.)
57 Show that the SPRT (sequential probability ratio test) of f' = fLo versus f' = fLl for
the mean of the normal distribution with known variance may be performed by
plotting the two lines
and
II
in the ny plane and then plotting 2:

1
Xi against n as the observations are made.
The test ends when one of the lines is crossed.
58 Consider sampling from f(x; e) = (1/9)1(0. B)(X), e> o. Discuss the sequential
probability ratio test of e eo versus () = ()1 with ()o < ()1.
59 Let XI, X 2 , " ' , XII' .•. be independent random variables all having the same
Bernoulli distribution given by P[XII = l] = () = 1 - P[XII = 0]. To test -*'0: e t
versus -*'1: () = 1, the following sequential test was used: Continue sampling as
long as n/2 - 2 < 2: Xi < n/2 + 2; if and when 2: Xi is first less than or equal to
n/2 - 2, accept -*'0; and if and when 2: Xi is first greater than or equal to n/2 + 2,
accept -*'1' Is this test a SPRT 1
60 Assume that X has a Poisson distribution with mean e. Consider testing
-*' 0: () = I versus -*'1: () 2. Fix (X = f3 = .05.
(a) Find the fixed sample size necessary to achieve the prescribed error sizes.
(b) Derive the (approximate) sequential probability ratio test, and show that it
can be based on the statistics 2:" XI, n = 1, 2, ...•

i-I
(c) Find the approximate expected sample sizes for the sequential probability
ratio test.
x
LINEAR MODELS

The purpose of this chapter is to discuss a special case of the linear statistical
model. There is a large amount of material available on this subject, but we
will only discuss the special case of the simple linear model. Some authors
refer to this as the theory of straight-line regression. To study this model will
require the use of some of the theory in previous chapters, such as distribution
theory, point and interval estimation concepts, and some material from hypoth-
esis testing. This chapter will demonstrate how these concepts can be utilized
for a situation, the simple linear model, that is important in applied statistics.
In Sec. 2 two examples are given to illustrate how the simple linear model
can be used to simulate real-world problems. In Sec. 3 the simple linear model
will be rigorously defined and put into a framework that will allow us to study it
by using s.tatistical procedures from previous chapters. In the remaining
sections discussion will be centered around point estimation, interval estimation,
and testing hypotheses on the parameters in the model under two different
assumptions about the distributions of the random variables in the model.
2 EXAMPLES OF THE LINEAR MODEL 483
2 EXAMPLES OF THE LINEAR MODEL

In this section we will give two examples to illustrate how the linear model arises
in applied problems.
EXAMPLE 1 The distance s that a particle travels in time t is given by the

formula s = Po + Plt, where Pl is the average speed and Po is the
position at time t = O. If Po and Pl are unknown, then s can be observed
for two distinct values of t and the resulting two equations solved for Po
and Pl' For example, suppose that s is observed to be 2 when t = 1, and
s is 11 when t = 4. This gives 2 = Po + Pl and 11 = Po + 4Ph and the
solution is Po = -1, Pl = 3; so s = -1 + 3t. Suppose that for some
reason the distance cannot be observed accurately, but there is a measure-
ment error which is of a random nature. Therefore s cannot be observed,
but suppose that We can observe Y, where Y = s + E and E is a random
error whose mean is O. Substituting for s gives us
where Yis an observable random variable, t is an observable nonrandom

variable, Eis an unobservable random variable, and Po and Pl are unknown
parameters. We cannot solve for Po and Pl by observing two sets of
values of Yand t, as We did with sand t above, since there is no functional
relationship between Yand t. The objective in this model is to find Po
and Pl and hence evaluate s = Po + Pl t for varipus values of t. Since s
is subject to errors and cannot be observed, We cannot know Po and Pl,
but by observing various sets of Yand t values statistical methods can be
used to obtain estimates of Po, Pl' and s. This type of model is a
functional-relationship model with a measurement error. / / //
EXAMPLE 2 For another example, consider the relationship between the

height h and weight W of individuals in a certain city. Certainly there is no
functional relationship between wand h, but there does seem to be some
kind of relation. We shall consider them as random variables and shall
postulate that (W, H) has a bivariate normal distribution. Then the
expected value of H for a given value w of W is given by
8[HI W = w] = Po + P1W, (2)
where Po and Pl are functions of the parameters in a bivariate normal
density. Although there is no functional relationship between Hand W, if
484 LINEAR MODELS x
they are assumed jointly normal, there is a linear functional relationship

between the weights and the average value of the heights. Thus we can
write the following: Hand Ware jointly normal, and
8[HI W = w] = Po + PI W ;
or We can write
H w = Po + PI W + E,
where E is a normally distributed random variable denoting error. This
is a regression model, and although it came from a somewhat different
problem than the functional relationship in Example 1, they both are
special cases of a linear statistical model, which will be discussed in this
~~6. @
3 DEFINITION OF LINEAR MODEL

Let Jl( . ) be a linear function of a real variable x. This is defined by Jl(x) =
Po + PIX, where x is in a domain D. Quite often D will be the entire real line,
a half line, or a bounded interval on the real line. To model the situations
referred to in Examples 1 and 2 above, we assume that there exists a family of
c.d.f.'s (one c.d.f. for each x in D) such that the mean of the c.d.f. corresponding
to a given x (say xo) in D is Po + PIXO' Thus the means of the c.d.f.'s are on the
line defined by Jl(x) = Po + PIX., See Fig. 1. The objective is to sample some
of the c.d.f/s and on the basis of the sample to make statistical inferences about
Po, Ph etc.
The sampling is accomplished as follows:
(i) A set of n x's in D is observed and denoted by Xh X2' ... , X n.
The x's are not random variables, but they may be selected either by
some random procedure or by purposeful selection.
(ii) Each Xi determines a c.dJ. whose mean is Po + PIX! and
whose variance is a 2 • From this c.d.f. a value is selected at random and
denoted by Y,. (Y 1 is a shortened notation for YXt ' )
Thus We have a set ofn pairs of observations, which we denote by (Yh Xl),
(Y2 , X2), ... , (Yn , Xn). We have assumed that
and
3 DEFINITION OF LINEAR MODEL 485
~~----~--------~------~--~x
FIGURE 1 D
Therefore we can define random variables E l , E 2 , ••• , En by
for i = 1, 2, ... , n,
and the Ei satisfy
and
So we can write
for i = I, 2, ... , n,
where
and
and this defines a linear model. We summarize these ideas below.
Definition 1 Linear model Let the function J1('} be defined by

J1(x} = Po + PIX for all x in a set D. For each x in Diet F y",( . } be' a
c.d.f. with a mean equal to tt(x), that is, Po + PIX, and variance 0'2. Let
Xl> X2' "" xn bean observed set ofnx's from D. For Xi let Yj be a random
sample of size 1 from thec.d.f. F yxt (' ) for i = 1,2, ... , n. Then (Yl , Xl),
(Y2 , X2), .•• , (~, xn) is a set of n observations related by
486 LINEAR MODELS x
and (3)
var [Yi ] = (12, i = 1,2, ... , n.
These specifications define a linear statistical model. IIII
Note We can write Eq. (3) as
Yj = /30 + /31 X, + Ei
G[Ei ] = 0 (4)
var [Ei] = (12,
wherei = 1,2, ... , n. /1/1
Note The word" linear" in "linear statistical model " refers to the fact
that the function p(.) is linear in the unknown parameters. In the
simple example We have referred to, p( . ) is defined by p(x) = /30 + /31 x; x
in D, and this is linear in x, but this is not an essential part of the definition
of this linear model. For example, Y = p(x) + E, where p(x) = /30 + /31 ~
is a linear statistical model. IIII
Note In many situations some additional assumptions on the c.dJ.
Fyx( • ) will be made, such as normality. Also, generally the sampling
procedure will be such that the Yi will be either jointly independent or
pairwise uncorrelated. In fact We shall discuss inference procedures for
two sets of assumptions on the random variables defined in Cases A and B
below. 1/1/
Case A For this caSe We assume that the n random variables are jointly
independent and each Yi is a normal random variable. IIII
Case B For this case We assume only that the Yi are pairwise uncor-
related; that is, cov [Yj, YJ] = 0 for all i :F j = I, 2, ... , n. 1//1
For Case A We shall discuss the following:

(i) Point estimation of /30, /31, (12, and p(x) for any x in D
(ii) Confidence interval for /30' /31, (12, and p(x) for any x in D
(iii) Tests of hypotheses on /30, /31' and (12
For Case B We shall discuss the following:
(iv) Point estimation of /30' /31, (12, and p(x) for any x in D
4 POINT ESTIMATION-CASE A 487
4 POINT ESTIMATION-CASE A
For this case YI , Y2, ... , Yn are independent normal random variables with
means Po + PIXh Po + PI X2, •.• , Po + PIX,. and variances a2 • To find point
estimators, We shall use the method of maxim urn likelihood. The likelihood
function is
(5)
and
2 n n 2 1~ 2
log L(Po , PI' a ) = - -2 log 21t - -2 log a - 2 2 1... (Y t - Po - PtXi) .
a i=l
The partial derivatives of log L(Po, Ph a 2 ) with respect to Po, Ph and a 2 are
obtained and' set equal to O. We let Po, PI' 8 2 denote the solutions of the
resulting three equations. The three equations are given below (with some
minor simplifications):
(6)
The first two equations are called the normal equations for determining Po and
Pl' They are linear in Po and PI and are readily solved. We obtain
(7)
(8)
(9)
These are maximum-likelihood estimates of Ph Po, and a 2 , respectively. We

notice that the x/s must be such that L (Xi - X)2 -=J:. 0; that is, there must be at
least two distinct values for the Xi'
488 LINEAR MODELS x
Note that since
fr.(Yi; PO' P1> 0-) = (21ta'r t exp [ -! (Y' - P:- P, Xr]

= (21tC1
2
rt eXp [ - 2~2 (Po + P,xii]
1 2
X exp ( - 20'2 Y i + Po Pl )
0'2 Yi + 0'2 Xi Y i ,
!Yj(Yi; Po, Ph 0') is a member of a three-parameter exponential family; hence,

by a generalization of Theorem 16 of Chap. VII
n n
L Yi,
i= 1
"X·
L. , y.
i= ]
J (10)
is a set of minimal sufficient and jointly complete statistics. Furthermore,

since the set of statistics given in Eq. (10) is a one-to-one transformation of the
estimators (statistics) defined by Eqs. (7) to (9), the estimators are themselves
minimal sufficient and jointly complete.
To further examine the properties that the estimators possess, we shall
find the joint distribution of statistics corresponding to Po, Pb 0'2. To do this,
we shall first find the moment generating function of 0 1 , O2 , and 0 3 , which
are random quantities with values defined by
1\
ttl -
_ Po - Po
,
1\
tt2 -
_ p] - p] , (11)
0' 0'
By Definition 25 of Chap. IV the joint moment generating function of 0 h O2 ,
0 3 is defined to be
met 1, t 2 , t 3 ) = G[i 101 +t202 +130 3 ]
if the expectation exists for - h < ti < h for some h > O. We obtain
exp [ - hL (Yi - Po - p, x,i ]

x (21t0'2)"/2 dYl ... dYn'
where in the integral the quantities ~1' ~2' ~3 will be written in terms of Yi and Xi'
This integral is straightforward but tedious to evaluate, and the result is
m(t 1 ,t 2 ,t3 )= ( expt ( t 21 , , (L xNn_2+2tlt2 [-"C x -2

]
L. Xi - X) L. Xi - X)
+ t2
2 L (Xi 1_ x)2
}) x (1 - 2t
3
)-(n-2)/2 for t <to
'3
4 POINT ESTIMATION-CASE A 489
From this moment generating function we can learn a number of things:

(i) It factors into a function of tl and t2 only times a function of t3
only_ We write this result as m(tb t 2 , t 3 ) = ml(tb t2)m2(t 3 )- By using
Theorem 10 of Chap. IV we know that the random variables associated
with t I and t2 are independent of the random variable associated with
t3; that is, 0 1 and O2 are independent of0 3 , which implies that the maxi-
mum-likelihood estimators of /30 and /31 are jointly independent of the
maximum-likelihood estimator of 0'2_
(ii) Since by a generalization of Theorem 7 in Chap. II a moment
generating function uniquely determines the distribution of the random
variables involved, we shall try to recognize the form of ml(t l , t2 ) and the
form of m2(t 3 )· We note by Theorem 12 of Chap. IV that ml(t 1, t 2) is the
moment generating function of a bivariate normal distribution, and, of
course, we obtain the means, variances, and covariance. We see that
the random variables, say Do and D., associated with Po and P1 are bivari-
ate normal random variables with means (/30' /31) and covariance matrix
0'2 L xt _O'2;X
n L (Xl - X)2 L (Xi - X)2
(12)
_O'2j 0'2
L (Xi - j)2 L (Xi - X)2
Another way to state this is the following: (Do, DI ) is a bivariate no.rmal

random variable with parameters
and
-u2 x
cov [Do, Dd = L (Xl - X)
-2
(iii) We recognize that m2(t 3) is the moment generating function
of a chi-square random variable with n - 2 degrees of freedom. Hence
we have
490 UNEAR MODELS x
which is distributed as a chi-square distribution with n - 2 degrees of

freedom. (Here, and in the rest of this chapter, &2 and a2 are used to
denote the random variables with values lJ2 and ij2 respectively.) By
Eq. (22) of Chap. VI we get
2
na ]
tff [ (12 =n-2;
so we define lJ2 by
A2 n -2 1 f A 2 A
(I = (I = 2 i...J (Yi - Po - PI X,) .

n-2 n- i=l
We shall summarize these results in the following theorem.
Theorem 1 Consider Case A of the simple linear model given in

Definition 1. The maximum-likelihood estimators of PI' Po, and (12
(corrected for bias) are given by
n _I (Yi - Y)( Xi - x)
1 - I (Xi - xl '
(13)
These estimators satisfy the following:

(i) They are jointly complete sufficient statistics.
(ij) They are unbiased estimators of their respective parameters.
(iii) (no, n I) is independent of &2.
(iv) (no, BI) has a bivariate normal distribution with mean (Po, PI)
and covariance matrix given by Eq. (12).
(v) (n - 2)it2/(l2 is a chi-square random variable with n - 2
degrees of freedom. / // /
In Chap. VII, we noted that maximum-likelihood estimators possess a

number of good properties, but they, in general, are not minimum-variance un-
biased estimators. We now employ a minor generalization of Theorem 11
of Chap. VII along with the results of Theorem 1 above, to state a strong op-
timal property about the estimators no, n h &2 of Po, Ph (12.
Theorem 2 Consider the simple linear model given in Definition 1. Let

'C(Po, Ph (12) be any known function of the parameters Po, Ph and (12
for which an unbiased estimator exists. Then there exists an unbiased
5 CONFIDENCE INTERV AL8-CASE A 491
estimator of 't(PO' PI' 0'2) that is a function of no, n

h and &2. We denote
this estimator by I(n n
o , h &2), and it is the UMVUE of 't(Po, Ph 0'2).
PROOF This result follows from a generalization of Theorem 17

of Chap. VII, since no, n
l , &2 is a set of sufficient complete statistics. 11//
Corollary The UMVUE of each of the parameters Po, PI' and 0'2 is
given by no, n
h and &2, respectively, in Theorem 1. 1III
Corollary The UMVUE of Jl(x) = Po + PIX for any X in the domain
D is flex), where p(x) = no n
+ 1x. (fl(x) is the random variable with
values Jl(x) = Po + PIX.) IIII
Corollary For any two known constants c1 and c2 the UMVUE of
no
c1Po + C2 /31 is CI n
+C 2 l · 1//1
5 CONFIDENCE INTERVALS-CASE A
To obtain a y-Ievel confidence interval on al, we note by Theorem I that
_ {n - 2)&2
U ---al-=---
is distributed as a chi-square random variable with n - 2 dJ. (degrees of free-

dom). Hence U is a pivotal quantity, and we get
P[Xtl -y)/2(n - 2) ::; U ::; Xtl +y)/2(n - 2)] = y.
If we substitute for U and simplify, we get
~2
p[X(l(n+Y)/2(n
2
- 2)&2
- 2)
cr < 2
(n - 2)&2
X(I-y)/2(n - 2)
]
= 1', (14)
and this is a 1001' percent confidence interval on 0'2.

To obtain a y-Ievel confidence interval on Po, we note that by Theorem 1:
(i) Z = (no - L
Po)J'L (Xi - x)2n10'2 xf is distributed as a stand-
ard normal random variable.
(U) (n - 2)&210'2 = U is distributed as a chi-square random variable
with n - 2 d.f.
(iii) Z and U are independent.
Hence, by Theorem 10 of Chap. VI
492 LINEAR MODELS x
is distributed as Student's t distribution with n - 2 d.f. Hence T is a pivotal

quantity. We get
P[ - t(1 +'1)/2(n - 2) ~ T < t(1 +'1)/in - 2)] = y,
and if we substitute for T, we get
{ -t(l+,)/.(n - 2) < J}o; Po In Lr~f x)' < t(1+,)/.(n - 2)] = y.
Mter simplifying We get the following for a l00y percent confidence interval
on Po:
[
p no - t(I+'1)/2(n - 2)&
J n
Lxf
L (Xi -'-X) 2
< Po < no + t(1+1)/2(n - 2)&

Jn LLX; ]
(Xi - xi = y.
We note that by Theorem 1 we get
and the estimated variance of no, which we write as var [no], is given by
,... r8]
var LllO =
~2
0- "
LX;
( -)2 •
n L.. Xi - X
Then the confidence statement can be written as
p[no - t(1+'1)/2(n - 2)Jvar [no] Po < no + t(I+1)/2(n - 2)Jvar [no]] - y.

(15)
To obtain a y-Ievel confidence interval on pb we note that by Theorem 1:
(i) Z = (n 1 - pl)JL 2
(Xi - x)2ja is distributed as a standard
normal random variable.
(ii) U = (n - 2)&2ja2 is distributed as a chi-square random variable
with n - 2 d.f.
(iii) Z and U are independent.
Hence by Theorem 10 of Chap. VI
5 CONFIDENCE INTERVALS-CASE A 493
is distributed as Student's t distribution with n - 2 dJ. Hence T is a pivotal

quantity, and we get
P[ - t(1 +y)/2(n - 2) < T < t(J +y)/2(n - 2)] = 1'.
If we substitute for T, we obtain
{ - t(1+Yl/2(n - 2) ~ (n, - P,)J'I (X~-; X)2 < t(1 +Yl/z(n - 2)J = y.
After some simplification we get

2
p[fJ] - t(1 +y)/2(n - 2)J'I (Xi1t- _ 2
X)
<p, < jj,
+ t(1 +Yl/2(n - 2)J'I (X~~ J X)2 = y,
and this is a 1001' percent confidence interval on /31' We note from Theorem 1
that
(12
var [fJ d = L (Xi -

-2
X)
and that the estimated variance of fJ], which is denoted by var [8]], is given by
a2
var [fJ d = L (Xi -
-)2 .
X
If we substitute this into the confidence-interval statement, we obtain
p[fJ 1 - t(1 + y)/i n - 2)J~ar [fJd < /31 < fJ] + t(l +y)/2(n - 2)Jvar [fJ1J] = 1'.
(16)
To obtain a y-ievel confidence interval on J.l(x) for any X in the domain D,
we note that
(i) J.l(x) = /30 + /3]x.
(ii) fi(x) = fJo + fJ 1 (X).
(iii) G[fi(x)] = J.l(x).
(iv) var [fi(x)] = var [8 0 + fJ] x]
= var [fJo] + 2xcov [fJo, fJ]] + x 2 var [fJd
=
L (Xi -
(12
X)2
(L x~ _ 2xx + X2)
n
= 'I(x:~ x)2 [(X - x)' + ~ 'I (x, - X)2]
= (1
2 [1~ + L(x - (Xi
X)2 ]
- X)2 .
494 UNEAR MODELS x
(v) Z = [tl(X) - Jl(x)]/Jvar [jl(x)] is distributed as a standard

normal random variable.
(vi) U = (n - 2)&2/q2 is distributed as a chi-square random vari-
able with n - 2 d.f.
(vii) U and Z are independent.
(viii) T = tl(x) - Jl(x) =~fio + filx - Po - PIX

J var[ft(x)] J8'2[I/n + (x - x)2tf. (Xi - xi]
is distributed as Student's t distribution with n - 2 d.f. Hence T is a
pivotal quantity, and we obtain
P[ -t(1+1)/2(n - 2) ~ T ~ t(1+1)/2(n - 2)] = y.
If we substitute for T and simplify, we get
or
p[fio + fi 1x - t(1+1)/2(n - 2)Jvar [tl(x)] ~ Po + P1 X

~ Do + fi 1x + t(1 +1)/2(n - 2)Jvar [(l(x)l] = 1',
and a 1001' percent confidence interval for Po + P1X is obtained.
6 TESTS OF HYPOTHESES-CASE A
In the linear model there are many tests that could be of interest to an investi-
gator. For example, he may want to test whether the line goes through the
origin, i.e., to test if the intercept is equal to zero, or perhaps test whether the
intercept is positive (or negative). These are indicated by
Jf0: Po = 0 versus Jf1 : Po i= 0,

Jf0: Po > 0 versus Jf1 : Po < 0,
Jf0: Po < 0 versus Jf1 : Po > o.
These tests indicate that there is no interest in the slope P1 or the variance q2
6 TESTS OF HYPOTHESES-<::ASE A 495
On the other hand the interest may be in the slope rather than the intercept, and
an investigator could be interested in testing
-*' 0: PI = 0 versus-*, 1: PI -::f:. 0,

-*' 0= P1 < 0 versus -*'1: P1 > 0,
etc. Rather than testing whether the intercept (or slope) is equal to 0 an investi-
gator may be interested in testing whether it is equal to a given number. For
exam pIe he may be interested in testing
-*'0: Po = 2 versus -*'1: Po < 2,

-*'0: PI = 1 versus -*'1 : P1 -::f:. 1,
etc. We shall derive a test of the hypothesis
We could just as well derive a test of the hypothesis
or a test of the hypothesis
To test
one obvious choice for a test statistic is
Under-*,0 the random variable T is distributed as Student's t distribution with

n - 2 degrees of freedom. Thus a test procedure with size t:J. is the following:
Reject-*, 0 if and only if IT I > t 1 _ a/ 2 (n - 2).
By comparing this with Eq. (16) we notice that this test is equivalent to the
procedure of setting a 1 - t:J. confidence interval on the parameter PI and rejecting
the hypothesis if and only if the confidence interval does not contain O.
We will now show that this test is a generalized likelihood-ratio test.
Corresponding to the notation in Chap. IX we note that in testing
-*'0: PI = 0 versus -*' 1: PI -::f:. 0
496 LINEAR MODELS x
the parameter spaces 9, 9 0 , and e 1 are as given below, where 0 = (Po, PI' 0'2):
9 = {(Po, Pb ~): - 00 < Po < 00; - 00 < PI < 00; 0'2 > O}
9 0 = {(flo, Ph 0'2): - 00 < Po < 00; PI = 0; 0'2 > O}
9 1 = 9 - eo.
We must determine A, where
sup L(O; Yl' .•• , Yn)
A = (I eeo (7)
sup L(O; Yl' ... , Yn)
(le~
L(O; Yb ... , Yn) = L(f3o, pb 0'2)
= (2n:2)"12 exp [ - 2~ L (y, - Po - Ptxi} (I 8)
and the values of Po, PI' 0'2 that maximize this for 0 E 9 are the maximum-
likelihood estimates given in Eqs. (7) to (9). Thus we get
where 0'-2 = -1 '" 1J - PIX,).

'--' (y, - Po 1J 2 To find sup L(O; Yb ... , Yn), we substitute
n (I eeo
PI = 0 into Eq. (18) above and get
1 [1 '"
L(flo, 0'2) = (2n0'2t12 exp - 20'2 '--' (y i-PO)2] .
But this is the likelihood function for a random sample of size n from a normal
distribution with mean Po and variance ~. The values of Po and (12 that
maximize the likelihood function are the maximum-likelihood estimates
p~ = 57
and
Thus
sup L(O; Yl, .. " Yn) = (0'*22n)-nI2 exp [-

(I e~o
~
2(1
L (Yt - p~i]
= (0'*2 2n) -nI2e-nI2.
6 TESTS OF HYPOTHESES-CASE A 497
We obtain
,- ( -&2 )n12
A - 2
0'*
for the generalized likelihood-ratio. Instead of A we will examine the quantity
(n - 2)(A -21n - 1), which is a monotonic function of A and hence will give an
equivalent test function. We get
L(Yi - y)2 - L(Yi - Po - Pt XJ2
L (Yi - Po - Pt X i)2
Replace Po with Po = y - PtX in the numerator, and get
A-21n _ 1 = L (Yi - y)2 - L [(Yi - y) - Pt(x i - x)F.
L (Yi - Po - Pt X j)2
Hence,
2
( _ 2)(A -21n _ 1) = PI L (Xi - X)2 = PI L (Xi -
x)2/O'
n fl2 &2/0'2'
which is the ratio of the values of two independent chi-square random variables
(under .J'f0: Pt = 0) divided by their respective degrees of freedom, which are
1 for the numerator and n - 2 for the denominator. Thus en - 2)(A -21n - 1)
has an F distribution with 1 and n - 2 degrees of freedom under .J'f 0 . The
generalized likelihood-ratio test says to reject .J'f 0 if and only if A < AO, or if
and only if
en - 2)(A- 2In - 1) > en - 2)(Ao2In - 1) = A~ (say),
or if and only if
[PI L (Xi - x)2]/fl2 > A~,
where A~ is chosen for a desirable size of Type I error.
Note that (n - 2)(A -21n - 1) is the square of
Jvar [fiIl'
and recall that the square of a Student's t-distributed random variable with n - 2
degrees of freedom has an F distribution with 1 and n - 2 degrees of freedom.
Thus we have verified that if the confidence-interval statement in Eq. (16) is
used to test .J'f0: Pt = 0 versus .J'ft : Pt =P 0, it is a generalized likelihood-ratio
test.
We will generalize this result slightly in the following theorem.
498 UNEAR MODELS x
Theorem 3 In the linear model given in Definition 1 the generalized

likelihood-ratio test of size ex of .1t0: PI = b 1 (b 1 is a given constant) versus
.1t 1: P1 i= bi is given by the following: Use Eq. (16) to set a 1 - ex con-
fidence interval on Pb and reject .1t0 if and only if the confidence interval
does not include b t - IIII
We shall state a theorem concerning a test of hypothesis on Po, and the
proof will be asked for in Prob. 18.
Theorem 4 In the linear model given in Definition 1 the generalized

likelihood-ratio test of size ex of .1t0: Po = b o(b ois a given constant) versus
.1t1 : Poi=b o is given by the following: Use Eq. (15) to set a I-ex
confidence interval on Po, and reject .1t 0 if and only if the confidence
interval does not include boo IIII
There are many other tests that are of interest for the linear model and the
interested reader can consult Refs. 17, 29, 31, and 32.
7 POINT ESTIMATION-CASE B
For this case Y1 , Y2 , ••. , Yn are pairwise uncorrelated random variables with
means Po + PI Xt, Po + PI X2, ... , Po + PtXn and variances a2. Since the joint
density of the Yj is not specified, maximum-likelihood estimators of Po, Pb and
(J'2 cannot be obtained. In models when the joint density of the observable
random variables is not given, a method of estimation called least-squares can
be utilized.
Definition 2 Least-squares Let (Y" Xi), i = 1, 2, ... , n, be n pairs of

observations that satisfy the linear model given in Definition 1. The
values of Po and PI that minimize the sum of squares
are defined to be the least-squares estimators of Po and Pl' IIII

To find the least-squares estimators of Po and PI' We must find the values
that minimize
1'1
Lcpo, PI) = L (Yi - Po - PI Xt)2,

I"" 1
7 POINT ESTIMATION-CASE B 499
and clearly these are the same values that maximize the likelihood function in
Eq. (5). Hence We have the following theorem.
Theorem 5 In Case B of the simple linear model given in Definition 1 the

least-squares estimators of Po and PI are given by fio and fib where
fi - L (Yi - Y)(Xj - x)
(19)
1- ~ ( -)2 '
L. Xi - X
1111
The least-squares method gives no estimator for (J2, but an estimator of

(J2 based on the least-squares estimators of Po and P1 is
For Case A the maximum-likelihood estimators of Po, P1' and (J2 had some
desirable optimum properties. The first corollary of Theorem 2 states that
fio and fi 1are uniformly minimum-variance unbiased estimators. That is, in the
class of all unbiased estimators of Po and Pb the estimators fio and fi1 in Eq. (13)
have uniformly minimum variance. No such desirable property as this is
enjoyed by least-squares estimators for Case B. For Case A the assumptions
are much stronger than for Case B, where the distribution of the random
variables Yj is assumed to be unknown; so We should not expect as strong an
optimality in the estimators for Case B.
For Case B, we shall restrict our class of estimating functions and deter-
mine if the least-squares estimators have any optimal properties in the restricted
class. Since C[Yd = Po + P1X" We see that Po (and P1) can be given by the
expected value of linear functions of the Y i • Within this class of linear functions
We will define minimum-variance unbiased estimators.
Definition 3 Best linear unbiased estimators Let Y1 , Y2 , ••• , Yn be

observable random variables such that C[Yj] = rlO), where r;(') are
known functions that contain unknown parameters 0 (0 may be vector-
valued). To estimate any OJ in 0, consider only the class of estimators
that are linear functions of the random variables Yj • In this class consider
only the subclass of estimators that are unbiased for OJ. If in this
restricted class an estimator of OJ exists which has smaller variance than
any other estimator of OJ in this restricted class, it is defined to be the best
linear unbiased estimator of OJ ("best" refers to minimum variance). 1111
500 UNEAR MODELS x
It should be noted that there are two restrictions on the estimating func-
tions before the property of minimum variance is considered. First, the class of
estimating functions is restricted to linear functions of the Yi . Second, in
the class of linear functions of the Yi only unbiased estimators are considered.
Finally, then, consideration is given to finding a minimum-variance estimator in
the class of estimating functions that are linear and unbiased.
We will now prove an important theorem that gives optimum properties
for the point estimators of /30 and /31 derived by the method of least squares for
Case B. This theorem is often referred to as the Gauss-Markov theorem.
Theorem 6 Consider the linear model given in Definition 1, and let the
assumptions for Case B hold. Then the least-squares estimators for /31
and /30 given in Eq. (19) are the respective best linear unbiased estimators
for /31 and /30 .
PROOF We shall demonstrate the proof for /30; the proof for /31
is similar. Since We are restricting the class of estimators to be linear, We
haVe :9 0 = I aj Yj • We must determine the constant aj such that:
(i) 8[:9 0] = /30; that is, :9 0 is an unbiased estimator of /30 .

(ii) var [:9 0] is a minimum among all estimators satisfying (i).
For (i) We must haVe
This gives the two equations which must be satisfied
and (20)
Now
var [:9 0] = 8[(:9 0 - /30)2] = 8[(I aj Yj - /30)2]

= 8[[I aj(/3o + /31Xj + E j ) - /3oY]
= 8[(/30 I aj + /31 I ajx] + I ajEj - /30)2J.
By the restrictions of Eq. (20)
var[:9 o] = 8[(I ajE])2] = 8[I aJ EJ +

]
I
]
Ii ajajEjE i ].
j*i
7 POINT ESTIMATION-CASE B 501
The quantity 8[E jE J ] is 0 if i t= j since, by assumption, the E, are un-

correlated and have means O. Hence
var [fJo] = (12 L aJ.
Since (12 is a constant, to minimize var [fJo] We need to minimize L aJ.
a;
Thus constants aj must be found which minimize L subject to the restric-
tions of Eq. (20). Using the theory of Lagrange multipliers, We must
. . .
nnmnnze
L =L aJ - Al (L aJ - 1) - A2 L aJ x j '
Taking derivatives, one finds
t = 1,2, ... , n (21)
8L
-
OAI
= - L aJ + 1 = 0,
8L
OA2 = -LajXj=O.
If we sum Over the first n equations, We get (using L at = 1)

2=nAl+A2Lxj, (22)
If we multiply the jth equation in (21) by x j and add, We get
2 L x j aJ = Al L x j + A2 L xJ,
or since I ajXj = 0, this becomes
If we substitute this into (22), We get
A _ - 2 I xtln _ - 2x
2 - '\' 2 -2 - '\' ( 2
'-' Xi - nx '-' Xi - Xl
and
1
11.1 -
_ 2
'\'
L xf/n 2 •
'-' (XI - x)
Substituting Al and A2 into the tth equation in (21) and solving for at
gives
SOl LINEAR MODELS x
The best linear unbiased estimator of Po is therefore

i'\
DO -
_"
L at
y. _
t -"
YL x; - x L Y
_ 2
t Xt _
-
-
Y -
i'\-
.01 X,
L(Xi - x)
which is the one given by least squares, and so the proof is complete. A
similar proof holds for PI . 1111
PROBLEMS
1 Assume that the data below satisfy the simple linear model given in Definition 1
for Case A.
y: -6.1 -0.5 7.2 6.9 -0.2 -2.1 -3.9 3.8
x: -2.0 0.6 1.4 1.3 0.0 -1.6 -1.7 0.7
Find the maximum-likelihood estimates of /30, /31, and a 2 ,
2 In Prob. 1 find the UMVUE of /30 + 3/31'
3 In Prob. I find a 95 percent confidence interval on /30; on /31; on a 2 •
4 In Prob. I find a 90 percent confidence interval on p.(x) for x = -1.0.
5 In the simple linear model for Case A find the maximum-likelihood estimator of 8,
where () = /30 + 3/31 + 2a 2 •
6 In Prob. 5 find the UMVUE of ().
7 In the simple linear model for Case A, show that p proportion of the distribution
of Yat x = Xo is below gp, where gp /30 + /31XO + Zp a and Zp is given by (zp) = p.
8 In Prob. 7 find the UMVUE of gp •
9 Use the data in Prob. I to evaluate the UMVUE of gl' in Prob. 7.
10 The hardness Y of the shells of eggs laid by a certain breed of chickens was as-
sumed to be roughly linearly related to the amount x of a certain food supplement
put into the diet of the chickens. The model was assumed to be a simple linear
model for Case A. Data were collected and are given below:
y,: .70 .98 1.16 1.75 .76 .82 .95 1.24 1.75 1.95
Xi: .12 .21 .34 .61 .13 .17 .21 .34 .62 .71
Test the hypothesis that /31 1.00 versus the hypothesis /31 :t 1.00. Use a Type I
error probability of 5 percent.
11 In Prob. 10 test the hypothesis /31 > I versus the hypothesis /31 < I.
12 In Prob. 10 test the hypothesis p.(.50) > 1.5 versus the hypothesis p.(.50) 1.5.
Use a Type I error probability of 10 percent.
13 In Prob. 10 compute a 90 percent confidence interval on 2a.
14 In the simple linear model for Case A find the UMVUE of /3da 2 •
15 Consider the simple linear model given in Definition I except var [Y,] = a/ 1 a l ,
where at, i = I, 2, ... ., n., are known positive numbers. Find the maximum-
likelihood estimators of /30 and /31'
PROBLEMS 503
16 What are the conditions on the Xi in the simple linear model for Case A so that
.fio and .fi 1 are independent?
17 In the simple linear model for Case A show that Yand .fit are uncorrelated. Are
they independent?
18 Prove Theorem 4.
19 In Theorem 6 give the proof for the best linear unbiased estimator of f31.
20 For the simple linear model for Case B prove that the best (minimum-variance)
linear unbiased estimator of f30 + f31 is 130 + 131, where .fio and .fit are the least-
squares estimators of fJo and fJh respectively.
21 Extend Prob. 20 to cofJo + CtfJh where Co and Cl are given constants.
XI
NONPARAMETRIC METHODS

The important place ascribed to the normal distribution in statistical theory is
well justified on the basis of the central-limit theorem. However, often it is
not known whether the basic distribution is such that the central-limit theorem
applies or whether the approximation to the normal distribution is good enough
that the resulting confidence intervals and tests of hypotheses based on normal
theory are as accurate as desired. For example, if a random sample of size n
is taken from a population with a normal density and a .95 confidence interval
is set about the mean (see Sec. 3.1 of Chap. VIII) then the frequency interpreta-
tion is the following: If repeated random samples are taken from this popula-
tion and if a 95 percent confidence interval is obtained for each random sample,
in the long run 95 percent of these intervals will contain the mean of the density.
If sampling is from a density that is not normal, then, instead of 95 percent of
the intervals containing the mean, it may be 99 or 90 percent, or some other
percentage. If it is close to 95 percent, say 93 to 97 percent, usually the experi-
menter will be satisfied. However, if it deviates a large amount from the desired
1 INTRODUCTION AND SUMMARY 50S
percentage, then the experimenter will probably not be satisfied. In cases where
it is known that the conventional methods based on the assumption of a normal
density are not applicable, an alternative method is desired. If the basic distri-
bution is known (but is not necessarily normal), one may be able to derive exact
(or sufficiently accurate) tests of hypotheses and confidence intervals based on
that distribution. In many cases an experimenter does not know the form of the
basic distribution and needs statistical techniques which are applicable regardless
of the form of the density. These techniques are called nonparametric or
distribution-free methods.
The term "nonparametric" arises from considerations of testing hypoth-
eSeS (Chap. IX). In forming the generalized likelihood-ratio, for example, one
deals with a parameter space which defines a family of distributions as the para-
meters in the functional form of the distribution vary over the parameter space.
The methods to be developed in this chapter make no use of functional forms
or parameters of such forms. They apply to very wide families of distributions
rather than only to families specified by a particular functional form. The term
" distribution-free" is also often used to indicate similarly that the methods do
not depend on the functional form of distribution functions.
The nonparametric methods that will be considered will, for the most part,
be based on the order statistics. Also, although the methods to be presented are
applicable to both continuous and discrete random variables, We shall direct our
attention almost entirely to the continuous case.
Section 2 will be devoted to considerations of statistical inferences that
concern the cumulative distribution function of the population to be sampled.
The sample cumulative distribution function will be used in three types of
inference, namely, point estimation, interval estimation, and testing. Popula-
tion quantiles have been defined for any distribution function regardless of the
form of that distribution. Section 3 deals with distribution-free statistical
methods of making inferences regarding population quantiles. Section 4 studies
an important concept, that of tolerance limits. The similarities and differences
of tolerance limits and confidence limits are noted.
In Sec. 5 We return to an important problem in the application of the
theory of statistics. It is the problem of testing the homogeneity of two popula-
tions. This problem was first mentioned in Subsec. 4.3 of Chap. IX when we
tested the equality of the means of two normal populations. I t was considered
again in Subsec. 5.3 of Chap. IX when We tested the equality of two multi-
nomial populations. We indicated there that the derived test using a chi-square-
type statistic couId be used to test the equality of two arbitrary populations, and
so we had really anticipated this chapter inasmuch as We derived a distribution-
free test. Other distribution-free tests of the homogeneity of two populations
506 NONPARAMETRlC METHODS xi
will be presented in Sec. 5. Included will be the sign test, the run test, the
median test, and the rank-sum test.
In this chapter we present only a very brief introduction to nonparametric
statistical methods. This chapter is similar to the last inasmuch as it includes
use of the three basic kinds of inference that Were the focus of our attention in
Chaps. VII to IX. We shall see that much of the required distributional theory
is elementary, seldom using anything more complicated than the basic principles
of probability that Were considered in Chap. I and the binomial distribution.
2 INFERENCES CONCERNING A
CUMULATIVE DISTRIBUTION FUNCTION
2.1 Sample or Empirical Cumulative Distribution Function

In Subsec. 5.4 of Chap. VI, We defined the sample cumulative distribution
function (c.d.f.). We indicated there that it could be used to estimate thecumula-
tive distribution function from which We sampled. In this subsection some
results about the sample c.d.f. will be reviewed and used to formulate point
.
estimates. In the two following subsections the sample c.d.f. will be utilized to
test a hypothesis (in Subsec. 2.2) and to set a confidence interval (in Subsec. 2.3).
Recall that (see Definition 13 in Chap. VI) the sample c.d.f. is defined by
1
Fn(x) = - (number of Xi less than or equal to x)
n
where Xl, ... , Xn is a random sample from some c.d.f. F(·). According to
Theorem 17 of Chap. VI,
P [F.(X) =~] =(;) [F(x)]"[l - F(x)r" k = 0, 1, ... , n, (2)
where Fn( . ) is the sample c.d.f. corresponding to c.d.f. F( .). From Eq. (2),
we see that
S[Fn(x)] = ±~ (n)
k=O n k
[F(x)t[1 - F(x)]n-k = F(x) (3)
and similarly
1
var [Fn(x)] = - F(x)[1 - F(x)]. (4)
n
2 INFERENCES CONCERNING A CUMULATIVE DISTRIBUTION FUNCTION 507
In fact, since Fn(x) is the sample mean of random variables I(-ro,x](X1),

... ,I( ro, x]( Xn), we know by the central-limit theorem that F n(x) is asymptoti-
cally normally distributed with mean F(x) and variance (l/n)F(x)[1 - F(x)].
Equations (3) and (4) show that for fixed x, Fn(x) is an unbiased and
mean-squared-error consistent estimator of F(x), regardless of the form of F( . ).
If one is interested in estimating F(x) for every x (rather than for a fixed x),
then one is interested in saying something about how close Fn(x) is to F(x)
jointly Over all values x; hence the following result is of interest:
P[ sup IFn(x) - F(x)l -----+~ 0] = 1. (5)
-ro<x<ro
Equation (5), known as the Glivenko-Cantelli theorem, states that with prob-
ability one the convergence of Fn(x) to F(x) is uniform in x. We can define
Dn = sup IFix) - F(x)l. (6)
-ro<x<ro
Dn is a random quantity that measures how far F n(') deviates from F(')'
Equation (5) states that P[lim Dn = 0] = 1; so, in particular, the c.d.f. of Dn ,
say FDn( • ), converges to the discrete c.d.f. that has all its mass at O. In the next
subsection we will consider the limiting distribution of J~ Dn. Equation (5)
tells us that the estimating function Fn(x) of the c.d.f. F(x) converges to F(x)
uniformly for all x with probability one.
Instead of a point estimate of F(x) = P[X ~ x], one might be interested
in a point estimate of F(y) - F(x) = P[x < X ~ y] for fixed x < y. The follow-
ing remark is useful in showing that Fn(Y) - Fix) is an unbiased mean-squared-
error consistent estimator of F(y) - F(x).
Remark
.1
cov [Fn(x), Fn(Y)] = - F(x)[1 - F(y)] for y > x. (7)
n
PROOF
508 NONPAR.AMETRIC METHODS
1
= ;;{8[I(-oo,x](X 1)I( oo,yiXl)]
- 8[I(-oo,X](X 1)]8[I( oo,yiXl)]}

1
= - [F(x) - F(x)F(y)]
n
1
=- F(x)[l -.- F(y)]. 1111
n
Using Eq. (7), one sees immediately that
var [FiY) - Fn(x)] = var [Fn(Y)] - 2 cov [Fix), Fn(Y)] + var [Fn(x)]
1
=- [F(y) - F(x)][l - F(y) + F(x)];
n
mean-sQuared-error consistency of Fn(Y) - Fn(x) as an estimator of F(y) - F(x)

follows immediately.
Rather than estimating P[x < X <y], i.e., the probability that X falls
in some interval, one might consider estimating P[X E B], i.e., the probability
n
that X falls in some set B. It can be shown that (seethe Problems) (lfn) L IB(Xi )
1= 1
is an unbiased estimator of P[X E B], and
1 n 1
var [ - L
n i= 1
]
IB(Xj) = - P[X E B](l - P[X E BD,
n
hence, is mean-sQuared-error consistent.
2.. 2 Kolmogorov-Smirnov Goodness-of-fit Test
We noted above that Fn(x) has an asymptotic normal distribution. Equivalently,

~ [Fn(x) - F(x)] has a limiting normal distribution with mean 0 and variance
F(x)[l - F(x)]. We now state (without proof) a result that gives the limiting
distribution of
In Dn = J~ -oo<x<oo
sup IFix) - F(x)l.
Theorem 1 Let Xh ••• , Xn , • •• be independent identically distributed

random variables having common continuous c.d.f. F x( . ) = F( . ).
Define
Dn = diXb ... , Xn) = sup I Fn(x) - F(x) I,
-oo<x<oo
where F n(x) is the sample c.d.f. Then
limF",nf)n(x) = linlP[J~Dn ~ x] (8)

n-+oo n-+oo
say. IIII
The c.d.f. given in Eq. (8) does not depend on the c.d.f. from which the
sample was drawn (other than that it be continuous); that is, the limiting
distribution of J~ Dn is distribution-free. This fact allows Dn to be broadly
used as a test statistic for goodness of fit. For instance, suppose one wishes to
test that the distribution that is being sampled from is some specified continuous
distribution; that is, test .1l' 0: Xi '" F o( .), where F o( .) is some completely
specified continuous c.d.f. If.1l' 0 is true,
Kn = In(XI , •.• , Xn) = In sup
-oo<x<oo
1Fn(x) - F o(x) 1 (9)
is approximately distributed as H( . ), the c.d.f. given in Eq. (8). If.1l' 0 is false,

then Fn( . ) will tend to be near the true c.dJ. F(') and not near Fo( . ), and
consequently sup 1F n(x) - F o(x) I will tend to be large; hence a reasonable
-oo<x<oo
test criterion is to reject .1l' 0 if sup IFn(x) - Fo(x) I is large. Since
-oo<x<oo
Kn = J~ sup 1 Fn(x) - F o(x) I is approximately distributed as H(') when
-oo<x<oo
.1l' 0 is true and H(') has been tabulated, k l - ex can be determined so that
1 - H(k l - ex ) = Ct, and hence P[Kn > k l - ex ] ~ Ct. That is, the test defined by
"Reject .1l' 0 if and only if Kn > k l - ex " has approximate size Ct. Such a test is
often labeled the Kolmogorov-Smirnov goodness-of-fit test. It tests how well a
given set of observations fits some specified c.dJ. F o( '). The fit is measured
by the so-called Kolmogorov statistic sup IFix) - F o(x)l. Theorem 1 gives
-oo<x<oo
an asymptotic distribution for Dn. The exact distribution of Dn has been
tabled for various n. See Ref. 44.
EXAMPLE 1 A question of at least curious interest is the following: Are

the times of birth uniformly distributed over the hours of the day? For
37 consecutive births (actual data) the following times Were observed:
7:02 P.M., 11 :08 P.M., 3:56 A.M., 8:12 A.M., 8:40 A.M., 12:25 P.M., 1 :24 A.M.,
8 :25 A.M., 2 :02 P.M., 11 :46 P.M., 10 :07 A.M., 1 :53 P.M., 6 ~45 P.M., 9 :06 A.M.,
3:57 P.M., 7:40 A.M., 3:02 A.M., 10:45 A.M., 3:06 P.M., 6:26 A.M., 4:44 P.M.,
12:26 A.M., 2:17 P.M., 11 :45 P.M., 5:08 A.M., 5:49 A.M., 6:32 A.M., 12:40 P.M.,
1 :30 P.M., 12:55 P.M., 3:22 P.M., 4:09 P.M., 7:46 P.M., 2:28 A.M., 10:06 A.M.,
11 :19 A.M., 4:31 P.M. Both the hypothesized uniform c.d.f. and the sample
c.dJ. are sketched in Fig. 1.
510 NONPARAMETRJC METHODS XI
t
1 ------
4:44 P.M.
~--~~----~----~----~~----~----~x
8 A.M. 12 noon
o 480 720 1440 min.
FIGURE 1
One can calculate In sup

x
1Fn(x) - F(x)1 = J371~~ - }~~ril ~ .85.
The critical value for size (J.. = .10 is greater than 1.22; so, according
to the Kolmogorov-Smirnov goodness-of-fit test, the data do not indicate
that the hypothesis that times of birth are uniformly distributed through-
out the hours of the day should be rejected. IIII
The Kolmogorov-Smirnov goodness-of-fit test assumed that the null

hypothesis was simple; that is, the null hypothesis completely specified (no
unknown parameters) the distribution of the population. One might inquire
as to whether such a goodness-of-fit testing procedure can be extended to a
composite null hypothesis which states that the distribution of the population
belongs to some parametric family of distributions, say {F( . ; 8): 8 E a}. For
such null hypotheses, sup I Fn(x) - F(x; 8) 1 is no longer a statistic since it depends
x
on an unknown parameter 8. An obvious way of removing the dependence on
8 is to replace 8 by an estimator, say 0, similar to what was done in the classical
chi-square goodness-of-fit test. The test statistic then becomes sup IF n{x)
x
- F{x; 0) I. The distribution of such a test statistic is not known and, in general,
depends on the hypothesized parametric family. Although some studies (often
Monte Carlo) have been reported in the literature, much remains to be done
before a Kolmogorov-Smirnov goodness-of-fit test for composite hypotheses
becomes a practical testing tool.
2.3 Confidence Bands for Cumulative Distribution Function

Theorem 1 can also be used to set confidence bands on the c.d.f. F( .) sampled
from. Let ky be defined by H(ky) = y, where H( . ) is the c.d.f. in Eq. (8). A
brief table of y and ky is
y .99 .95 .90 .85 .80
1.63 1.36 1.22 1.14 1.07
It follows that
P[ J~ sup IF,,(x) - F(x)1 < ky] ~ y,

x
but
P[ In sup IFn(x) -
x
F(x) I < ky] = P[ sup IF,,(x) -
x
F(x) I < ky/Jn]
= P[F.(X) - j~ < F(x) < F.(x) + j~ for all x],
noting that
k
sup IF,,(x) - F(x) I < JY~
x n
if and only if
for all x. Using the fact that 0 < F(x) :::; 1, We have
that is, the band with lower boundary defined by L(x) = max [0, F,,(x) - ky/ J~]
and upper boundary defined by U(x) = min [F,,(x) + ky/ In,, 1] is an approxi-
mate 100y percent confidence band for the c.d.f. F( . ), where the meaning of the
confidence band is given in Eq. (10).
512 NONPARAMETRIC METHODS XI
3 INFERENCES CONCERNING QUANTILES
3.1 Point and Interval Estimates of a Quantile

Throughout this section, we will assume that we are sampling from a continuous
c.d.f., say F(')' Recall (see Definition 17 of Chap. II) that the qth quantile of
c.d.f. F( . ), denoted by eq , is defined by F(e q ) = q for fixed q, 0 < q < 1. In
particular, for q = t, e! is called the median. We saw in Subsec. 4.6 of Chap. II
that quantiles can be used to measure location and dispersion of a c.d.f. For
instance, e!, (e q + e.- q )/2, etc., are measures of location, and e.9 - e..,
e .75 - e.2 5' etc., are measures of dispersion.
In Subsec. 2.1, we considered estimating F(x) for fixed x; now, We con-
sider estimating eq such that F(e q ) = q for fixed q. We know that if X is a
continuous random variable with c.d.f. F( . ), then the random variable F(X)
has a uniform distribution (see Theorem 12 in Chap. V) over the interval (0, 1).
Hence F(Yj) has the same distribution as the jth order statistic from a uniform
distribution, and We know that 8[F(Yj)] = j/(n + 1). (As usual, Y., ... , Yn
are the order statistics corresponding to the random sample X•., ... , Xn .)
e
Consequently, we might estimate q with Yj if q ~ j/(n + 1). [lfj/(n + 1) < q <
e
(j + 1)/(n + 1), one could estimate q by interpolating between the order statistics
Yj and Yj + •. ]
e
A confidence-interval estimate of q can be obtained by using two order
statistics, the interval between them constituting the confidence interval. We
are interested in computing the confidence coefficient for a pair of order statistics.
P[Yj < eq < Yk ] = P[F(Yj ) 5:: F(e q) = q < F(Yk )]
= 1 - P[F(Y) > q] - P[F(Yk ) < q]
= P[F(Yj ) < q] - P[F(Yk ) < q].
Recall that
fYj(Y) = U _ l)~ ~n _ J)! [F(y)]i-'[l - F(y)]n- if(y);
hence for Z = F( lJ)

dz
-dy =f(y) ,
and so,
fz(z) = dz;dY frj(Y) = U _ l)~~n ~ i)! zi-'(\ - zr i

for 0 < z < 1.
3 INFERENCES CONCERNING QUANTILES 513
Thus,
...
P[F(Yj ) < u] = fo fz(z) dz
=
-----
B(j,
1
j + 1)
n -
f u
0 z
.
J 1(1 - z)n- j+ 1 1 dz
= IB.,(j, n - j + 1),
called the incomplete beta function, which is extensively tabulated. Hence,
which is the confidence coefficient of the interval (Yj , Yk ). In practice, of course,

we are interested in going in the other direction; that is, for fixed I' pickj and k
(and consequently order statistics YJ and Yk ) such that
IBq(j,n '-j+ 1)- IBq(k,n -k + 1)=1',
and then (Yj , Yk ) is a 1001' percent confidence interval for q . Of course, for e
arbitrary I' there will not exist a j and k so that the confidence coefficient is
exactly y.
The confidence coefficient can be obtained another way.
But
P[Yj :::;; eq] = P[jth order statistic < eq]
= P[j or more observations < e
q]
n
= I P[exactly i observations < eq]
i=j
= ,tj(;) [F(e.)]'[l - F(e,»)"-'

= .t.(~)qi(l _q)n- i;
&= J I
hence,
Note that a table of the binomial distribution can now be used to evaluate the
confidence coefficient.
EXAMPLE 2 For a sample of size 10,

what is the confidence coefficient of
the interval (Y2, Y9 ), which is a confidence-interval estimator of the pop-
ulation median? We have
P[Y 2 ~ ~t < Y9 ]
(n) (1)2:
= ~8 i n
= ~8 (10)
i
(1)2: 10
= .9784. IIII
We have presented one way, using order statistics, of obtaining point esti-
mates or confidence-interval estimates for a quantile.
Besides being extremely general in that the method requires few assump-
tions about the form of the distribution function, the method is extraordinarily
simple. No complex analysis or distribution theory was needed; the simple
binomial distribution provided the necessary equipment to determine the con-
fidence coefficient. The only inconvenience was the paucity of confidence levels
that could be attained.
3.2 Tests of Hypotheses Concerning Quantlles

Let X h . . . , Xn denote a random sample from a probability density function,
say f( '). Suppose that it is desired to test that the qth quantile of the popula-
tion sampled from is a specified value, say~. That is, it is desired to test
where f( .) is unspecified (other than being a probability density function).

The confidence-interval method of deriving a test (see Subsec. 3.4 of Chap. IX)
can be used; for instance, obtain a 1001' percent confidence interval for ~q, and
accept :leo if and only if the derived confidence interval contains~. Such a test
has size 1 - 1'.
An alternative test is the so-called one-sample sign test. It is a very simple
test based on the value of a statistic that represents the nurrlber of the n trans-
formed observations that have a positive sign. To illustrate the principle in-
volved in the sign test, consider testing :leo: ~q = ~ versus :Ie1 : ~q =1= ~ for a
random sample Xl' ... , Xn from some unspecified probability density function
f( '). Let Z denote the number of X/s that exceed~. Equivalently, Z is the
number of Xl - ~, ... , Xn - ~ that have a positive sign. If:le 0 is true, Z has a
binomial distribution with parameters nand p = 1 - q = If f(x) dx. So if
:Ie 0 is true, one would expect Z to be near np, and hence an intuitively appealing
test is to accept :leo if and only if Z is near np. Since the distribution of Z is
known, one can determine what is meant by " near" by fixing the size of the test.
4 TOLERANCE UMlTS 515
e
For example, suppose q = t so that q = e t = median; then a possible test of
.1l'o:e t =e versus .1l'l:e t =l=e is to accept.1l'o if and only if jZ-npj =
1Z - nl2j < c, where c is a constant determined by
P[ IZ - nl2l < c] = 1 - a,
where a is the desired size of the test. Now
n/2L+c (n). (l)j(l)n-

n/2 -c ) -2 -2 ;
j
P[ IZ - n /21 < c] =
so c can be determined from a binomial table. (For small sample sizes, not
many a's are possible, unless randomized tests are used.) The power function
of such a test can be readily obtained since the distribution of Z is still binomial
even when the null hypothesis is false; Z has the binomial distribution with
parameters nand p = P[X > e]. Such a power function could be sketched as a
function of p.
Note also that the sign test can be used to test one-sided hypotheses. For
e e
instance, in testing .1l' 0: q ~ e versus .1l'1: q > e, the sign test says to reject
.1l' 0 if and only if Z, defined as above, is large. Again the power function can
be easily obtained.
4 TOLERANCE LIMITS
An automatic machine in a ball-bearing factory is supposed to manufacture

bearings .25 inch in diameter. The bearings are regarded as acceptable from an
engineering standpoint if the diameter falls between the limits .249 and .251 inch.
Production is regularly checked each day by measuring the diameter of a random
sample of bearings and computing statistical tolerance limits L1 and L2 from
their samples. If Ll is above .249 and L2 is below .251, the production is
accepted. How large should the sample be so that one can be assured with
90 percent probability that the statistical tolerance limits will contain at least
80 percent of the population of bearing diameters? There is a simple non-
parametric solution to problems of this kind.
In more general terms, let f( . ) be a probability density function, and on
the basis of a sample of n values it is desirable to determine two numbers, say
Ll and L2 , such that at least .80, say, of the area under f( . ) is between Ll and
L 2 • On the basis of a sample We cannot be certain that .80 of the area under
f( . ) is between Ll and L 2 , but we can specify a probability that it is so.
516 NON PARAMETRIC METHODS XI
In other words, we want to find two functions Ll = 11(Xb . . . , Xn) and

L2 = 12(X1, ... , Xn) of the random sample Xb ... , Xn such that the probability
that
is equal to ,}" for specified'}' and 13. We summarize with a definition.
Definition 1 Tolerance limits Let Xl, ... , Xn be a random sample from

continuous c.d.f. F(·) having a density function f(·). Let Ll =
11(X1, ... , Xn) < L2 = 12(X1, ... , Xn) be two statistics which satisfy:
(i) The distribution of F(L 2) - F(Ld does not depend on F( . ).
(ii) P[F(L 2) - F(Ld > 13] = '}'.
Then Ll and L2 will be defined to be 10013 percent distribution-free tolerance
limits at probability level'}'. // //
Remark Note that the random quantity F(L2) - F(Ld represents the
area under f( . ) between Ll and L2 . / // /
For continuous random variables, order statistics Yj and Yk (j < k) form

tolerance limits. To obtain the coefficients 13 and'}' in the definition of tolerance
limits, We need the distribution of F(L 2) - F(Ll). Recall that
n1
fyjo Yk(Yj, Yk) = (j _ 1)1 (k - I - j)! (n - k)!
x [F(y)F-1[F(Yk) - F(Yj)]k-l- j [1 - F(Yk)]n-kf(Yj)f(Yk).
Make the transformation Z = F(Yk ) - F(Yj ) and Y = F(Y), find the joint distri-
bution of Yand Z, and then integrate out Y to get the marginal distribution of Z.
The following obtains:
I' ( _ n1 k- 1- j( n- k +j
(12)
JZ z) - (k _ 1 _ j)!(n _ k + j)! Z 1 - z) I(O,l)(Z),
which is a beta distribution with parameters k - j and n - k + j + 1. Now

P[Z < 13] = f fz(z) dz = IBp(k - j, n - k +
p
j + 1),
°
the incomplete beta function, which is tabled. Also, recall that
IBp(k - j, n - k + j + 1) = f. (~)f3i(1 - f3t-

k-j I
i
.
Thus for any 13, the probability level'}' can be computed.

4 TOLERANCE LIMITS 517
EXAMPLE 3 For a random sample of size 5, use (Yi, Ys ) as a tolerance

interval for 75 percent of the population; that is, p = .75. What is the
corresponding probability level y? We seek
y = P[F(Ys) - F(Y1) > .75]
= 1 - P[F(Ys) - F(Y1 ) < .75]
= 1- ±(~)(.75y(.25)5-i
1=4 I
= .3672. /III
We might note that, in general,

k j k-j
8[Z] = 8[F(Yk)] - 8[F(Yj )] = n + 1 - n + 1 = n + 1 . (13)
For Example 3 8[Z] = : = j-.
EXAMPLE 4 Suppose that it is desired to determine how large a sample must

be taken so that the probability is .90 that at least 99 percent of a future
day's output of bearings will have diameters between the largest and
smallest observations in the sample. The quantities are y = .90 and
p = .99, and we want to determine n such that
P[F(Yn ) - F(Y1 ) > P] = y,
where the density of Z = F(Yn) - F(Y1 ) is given by Eq. (12) for j = 1 and
k = n. We get
Y = P[Z > Pl = f n(n -

1
p
1)2"'-2(1 - z) dz = 1 - npn-l + (n - l)pn.
If we substitute for y and p, we get the equation

.90 = 1 - n(.99)n-l + (n - 1)(.99),
which can be solved to determine n. The solution is n ~ 388. IIII
There are similarities and differences between tolerance limits and con-
fidence limits. Tolerance limits, like confidence limits, are two statistics, one
less than the other, that together form an interval with random end points. The
user of either interval is reasonably confident (the degree of confidence being
measured by the corresponding confidence level) that the interval ob~ained
contains what it is claimed to. This is where the similarity ends. A confidence
interval is an interval thought to contain a fixed unknown parameter value. On
the other hand, a tolerance interval is an interval thought to contain a prescribed

proportion of the values of the random variable under consideration. In other
words, a confidence interval is an interval thought to contain an unknown fixed
parameter value that characterizes the distribution of population values, whereas
a tolerance interval is an interval thought to contain actual population values,
and not some characteristic of them.
5 EQUALITY OF TWO DISTRIBUTIONS
5.1 Introduction
In this section various tests of the equality of two populations will be studied.
As we mentioned in Sec. 1 above, we first studied the equality of two populations
when we tested that the means from two normal populations Were equal in
Subsec. 4.3 of Chap. IX. Then again in Subsec. 5.3 of Chap. IX, we gave a test
of homogeneity of two populations. A great many nonparametric methods
have been developed for testing whether two populations have the same distribu-
tion. We shall consider only four of them; a fifth will be briefly mentioned at
the end of this subsection.
The problem that we propose to consider is the following: Let Xl, ... , Xm
denote a random sample of size m from c.d.f. F x( . ) with a corresponding density
function fx( .), and let Y1 , ••• , Y,. denote a random sample of size n from
c.d.f. F y(') with a corresponding density function fy( '). (Note that We are
departing from our usual convention of using Y's to represent the order statistics
corresponding to the X's.) Further, assume that the observations from Fx( . )
are independent of the observations from F y( '). Test.1l' 0: F x(z) = F y(z) for
all z versus .1l'1: F x(Z) =1= F y(z) for at least one value of z. In Sec. 2 above We
pointed out that the sample c.d.f. can be used to estimate the population c.dJ.
In the case that .1l' 0 is true, that is, F x(z) = F y(z), We have two independent
estimators of the common population c.d.f., one using the sample c.d.f. of the
X's and the other using the sample c.d.f. of the Y's. Intuitively, then, one might
consider using the closeness of the two sample c.d.f.'s to each other as a test
criterion. Although we will not study it, a test, called the two-sample
Kolmogorov-Smirnov test, has been devised that USes such a criterion.
We will assume throughout that the random variables under consideration
are continuous and merely point out at this time that the methods to be presented
can be extended to include discrete random variables as well. In our pre-
sentation, we will consider testing two-sided hypotheses and will not consider
one-sided hypotheses, although the theory works equally well for one-sided
hypotheses.
5 EQUALITY OF TWO DISTRIBUTIONS 519
5.2 Two-sample Sign Test

The first test that we win consider is the two-sample sign test. We shall See
that, in a certain sense, this test is a nonparametric analog of the paired t test
(see Prob. 17 in Chap. IX). For this test We assume that the sampling situation
is such that the X and Yobservations are paired; that is, we observe (Xl' Yi),
... , (Xn' Yn)· One could think that the X observation is "untreated" (a
control) and the corresponding Y observation is "treated," the object of the
test being to determine if there is a "treatment" effect. We wish to test
-*' 0: F x( . ) = F y( .). Assume that (Xb YJ.), ... , (Xn' Yn) is a random sample
from some joint distribution Fx, y( " '). Further assume that Fx. y( " .) is
such that P[X> Y] = P[X < Y] = i when -*'0 is true. (ReCall that we are
assuming continuous random variables, and then such an assumption is satisfied
if X and Yare independent.) Consider a test based on the signs of the
differences Xi - Yj, i = 1, ... , n. For instance, define
n
then Zi has a Bernoulli distribution, and consequently Sn = L Zi has a binomial
i= 1
distribution with parameters nand p = P[Xi > Yi ]. If -*'0 is true, p = i, and
&[Sn] = n12. If the alternative hypothesis is two-sided so that p = P[Xi > Yd
can be either larger or smaller than i, then a possible test criterion is to accept
-*'0 if Sn is close to n12, that is, accept -*'0 if ISn - nl21 :s; k, where k is deter-
mined by fixing the size of the test. k is easily determined from a binomial
table, and we have a very simple test of the equality of the two populations.
One can See that avoidance of the assumption that Xi and Y, are inde-
pendent is desirable. For example, Xi might represent an observation on the
ith entity before some" treatment" and Yi the observation on the same entity
after" treatment." In such a case one is not likely to have independence of
Xi and Yj since they are observations taken on the same entity, yet one can
sometimes test that there is no " treatment" effect by testing that the" before "
and" after" popUlations are the same.
5.3 Run Test

As before let Xl, ... , Xm denote a random sample from Fx( . ) and YI , ... , Y,.
a random sample from F y( .). A rather simple test of -*'0: F x(z) = F y(z) for
all z is based on runs of values of X and values of Y. To understand the
meaning of runs, combine the m x observations with the n y observations and
then order (in ascending order of magnitude) the combined sample. For
example, if m = 4,and n = 5, one might obtain
y x x y x y y y x. (14)
A run is a sequence of letters of the same kind bounded by letters of another
kind except for the first and last position. Thus, in Eq. (14) the ordering
starts with a run of one y value, then follows a run of two x values, then a run
of one y value, and so on; six runs are exhibited in Eq. (14). It is apparent that
if the two samples are from the same population, the x's and y's will ordinarily
be well mixed, and the total number of runs will be large. If the two popula-.
tions are widely separated so that their range of values does not overlap, then
the number of runs will be only two, and, in general, differences between the
two populations will tend to reduce the number of runs. Thus the two popula-
tions may have the same mean or median, but if the x population is concentrated
while the y population is dispersed, there will be a tendency to have a long y
run on each end of the combined sample, and there will thus be a tendency
to reduce the number of runs. A test then is performed by observing the total
number of runs, say Z, in the combined sample and rejecting .1P 0 if Z is less
than or equal to some specified number Zo. Our task now is to determine the
distribution of Z under .1P 0 in order that for a given test size We may specify Zo.
If .1P 0 is true, it can be argued that the possible arrangements of the
m x values and n y values are equally likely. It is clear that there are exactly
(m ; n) such arrangements. To find P[Z = z], it is necessary now to count
all arrangements with exactly z runs. Suppose z is even, say 2k; then there
must be k runs of x values and k runs of y values. To get k runs of x values,
the m x's must be divided into k groups. We can form these k groups, or runs,
by inserting k - 1 dividers into the m - 1 spaces between the m x values with no
more than one divider per space. We can place the k - 1 dividers into the
m - 1 spaces in (~=::) ways. Similarly, We can construct the k runs of
y values in (; =: !) ways. Any particular arrangement of the k runs of x values

can be combined with any arrangement of the k runs of y values; furthermore,
the first run in the combined arrangement can be either a run of x values or a
run of y values; hence there are a total of 2 (~ =: D(; =: :) arrangements having
exactly z = 2k runs. Hence
Similarly, for z odd
(m- I) (n - I) (m - 1) (n - I)
P[Z = z] = P[Z = 2k + I] =k k- I + k- I k . (16)
(m ;n)
To test Jf 0 with size of Type I error equal to 0;, one finds the integer Zo so that
(as nearly as possible)
%0
L P[Z =
%=2
z] = 0; (17)
and rejects Jf 0 if the observed value of Z dOes not exceed zo.

The computation involved in Eq. (17) can become quite tedious unless
both m and n are small. Fortunately, the distribution of Z is approximately
normal for large samples, and in fact the approximation is usually good enough
for practical purposes when both m and n exceed 10. If Jf 0 is true, the mean
and variance of Z are
2mn
8[Z] = +I (18)
m+n
and
2mn(2mn - m - n)
Z =
vu [] . (19)
(m+n)2(m+n-l)
The asymptotic normal distribution of Z under Jf 0 has mean and variance given
in Eqs. (18) and (19). This asymptotic normal distribution can be used to
determine the critical value Zo for large samples.
The run test is sensitive to both differences in shape and differences in
location between two distributions.
5.4 Median Test

Let Xl, ... , Xm be a random sample from Fx( . ) and 11, ... , ~ be a random
sample from F y( .). As in the previous subsection, combine the two samples,
and order them. Let Zl < Z2 < ... < Zm+n be the combined ordered sample.
The median test of Jf o : Fx(u) = Fy(u) for all u consists of finding the median,
say Z, of the z values and then counting the number of x values, say mb which
exceed z and the number of y values, say nb which exceed z. If Jf 0 is true, ml
should be approximately ml2 and nl approximately n12. We can USe either the
statistic Ml or the statistic Nl to construct the test. Let us use Ml = number
of X's which exceed Z, the median of the combined sample. If m + n is even,
there are exactly (m + n)j2 of the observations (combined x's and y's) greater
than the median of the combined sample. (Since we have an even number of
continuous random variables, no two are equal, and the median is midway be-
tween the middle two.) It can be easily argued that
P[M, = mtl ('::)(m + n~2 - m,)

=
( :++n;/2)
for m + n even and .1f 0 true. A similar expression obtains for m +n odd.
Such a distribution can be used to find a constant k such that
and our test is given by the following:

m
Reject .1f 0 if and only if Ml - 2 >k.
Just as in the run test, an asymptotic normal distribution of Ml can be

derived, but we will not study it.
5.5 Rank-sum Test

A very interesting nonparametric test for two samples was described by
Wilcoxon and studied by Mann and Whitney. Given two random samples
Xl' X 2 , ••• , Xm and Yl, ... , Yn from populations with absolutely continuous
c.d.f.'s Fx( . ) and Fy( . ), respectively, one arranges the m + n observations in
ascending order and then replaces the smallest observation by 1, the next by 2,
and so on, the largest being replaced by m + n. These integers are called the
ranks of the observations. Let Tx denote the sum of the ranks of the m x values
m+n
and Ty the sum of the ranks of the n y values. Note that Tx + Ty = Ij =
j=l
(m +n+ l)(m + n)j2; so Ty is a linear function of Tx' We could base a test
on either statistic Tx or Ty . Let us use Tx. Tx is linearly related to another
statistic, which we denote by U. Set
n m
U= I II[Y;.oo)(XJ,
j= 1 i=l
(20)
the number of times an X exceeds a Y. For a given set of observations, let

rh r2' ... , rm denote the ranks of the x values, and let Xl, ... , x:n denote the
ordered x values. Clearly Xl exceeds (rl - 1) y-values, X2 exceeds (r2 - 2) y-

values, and so on, and Xm exceeds (rm - m) y-values. Hence
m. • m(m + 1)
u= I 1 (ri -I) = I
1=
ri - I' = tx -
2
,
or
To find the first two moments of Tx , we find the first two moments of U.
G[U] = G[II I[fj' 00)(X1)] = II G[I[fj' oo)(Xi )]

= II P[Xi > lj] = IIp = mnp,
where
p = P[Xi > Yj ] = f P[Y f

< x I X = x]fx(x) dx = F f(x)fx(x) dx.
If .J'f 0 is true,
1
P= f Fx(x)fx(X) dx = fo u du = t·
SimilarlY, the variance of U can be found. The derivation is somewhat more
complicated since one needs the expected value of U 2 • From the mean and
variance of U, the mean and variance of Tx can be obtained. If.J'f 0 is true,
they are given by
and
rr] mn(m + n + 1)
var [ .Lx = 12 • (23)
The exact distribution of Tx turns out to be a very troublesome problem

for large m and n. However, Mann and Whitney have calculated the distribu-
tion for small m and n, have shown that Tx is approximately normally distributed
for large m and n, and have demonstrated that the normal approximation is
quite accurate when m and n are larger than 7. Thus for samples of reasonable
size one can use the normal approximation with mean and variance given by
Eqs. (22) and (23) to find a critical region for testing .J'f 0: F x(z) = F y(z) for all z
versus .J'f1: F x(Z) =1= F y(z). The test would be the following:
Reject .J'f 0 if I Tx - G[Tx] I is large;
524 NON PARAMETRIC METHODS XI
that is,
Reject :K0 if and only if 1 Tx - 8[Tx] I ~ k,
where k is determined by fixing the size of the test and using the asymptotic
normal distribution of Tx.
EXAMPLE 5 Find the exact distribution of Tx under :K 0 for m = 3 and

n = 2. Each of the following arrangements is equally likely if :K0 is
true:
x x x y y, x x y x y, x x y y x, x y x x y, x y x y x,
x y y x x, y x x x y, y x x y x, y x y x x, y y x x x.
The corresponding Tx values are, respectively, 6, 7, 8, 8, 9, 10,9, 10, 11, 12;
so
1
P[ Tx = 6] = P[ Tx = 7] = 1 0'
P[Tx = 8] = P[Tx = 9] = P[Tx = 10] = lo'

and
P[Tx = 11] = P[Tx = 12] = /0' IIII
PROBLEMS
1 n
1 Show that T= - L IB(X,) is an unbiased estimator of P[X E B]. Find var rll,
n ,= 1
and show that T is a mean-squared-error consistent estimator of P[X E B].
1 n
2 Define Fn(BJ) = - L IBiXt) for j = 1, 2. Find cov [Fn(B 1), Fn(Bz)].
n,=l
3 Let Y1 , ••• , Yn be the order statistics corresponding to a random sample of size n
from a continuous c.d.f. F(' ).
(a) Find the density of F( Yj ).
(b) Find the joint density of F( Y t ) and F( Y j ).
(c) Find the density of [F(Yn) - F(Yz)]/[F(Yn) - F(Y1 )].
4 Let Xl, ... , Xn be independent and identically distributed random variables
having common continuous c.d.f. F( .). Let YI < ... < Yn be the corre-
SpOnding order statistics, and define F n(') to be the sample c.d.f. Set Dn =
sup 1Fn(x) - F(x) I.
-CXl<X<CXl
(a) Find the exact distribution of Dn for n = 1.

(b) Do the same forn = 2. HINT: Does Dn = max [F(YI ), 1- F(Y1),F(Yz) -I,
1 - F(Yz)]?
(c) Argue that the exact distribution of Dn will not depend on F(')'
PROBLEMS 525
5 Show that the expected value of the larger of a random sample of two observations
from a normal population with mean 0 and unit variance is VV;:;' and hence that
for the general normal population the expected value is fL + a/V;:;'.
6 If (X, Y) is an observation from a bivariate normal population with means 0, unit
variances, and correlation p, show that the expected value of the larger of Xand Y
is V(l - p)/Tr.
7 We have seen that the sample mean for a distribution with infinite variance (such
as the Cauchy distribution) is not necessarily a consistent estimator of the popula-
tion mean. Is the sample median a consistent estimator of the population median?
8 Construct a (approximate) 90 percent confidence band for the data of Example 1.
Does your band include the appropriate uniform distribution?
9 Let Yl < ... < Y s be the order statistics corresponding to a random sample from
some continuous c.d.f. Compute P[Yl < fso < Y s] and P[Yz < fso < Y 4 ].
Compute P[Yl < fzo < Y z ]. Compute P[Y3 < f7S < Ys].
10 Let Yl and Y,. be the first and last order statistics of a random ~~ple of size n
from some continuous c.d.f. F(·). Find the smallest value o'f n such that
P[F( Y,.) -F( Yl ) > .75] > .90.
11 Test as many ways as you know how at the 5 percent level that the following two
samples came from the same population:
x 1.3 1.4 1.4 1.5 1.7 1.9 1.9
y 1.6 1.8 2.0 2.1 2.1 2.2 2.3
12 Let Xl, ... , Xs denote a random sample of size 5 from the density f(x; fJ) =
1(6 - t. 6 + t)(x). Consider estimating O.
(a) Determine the confidence coefficient of the confidence interval (Yl , Ys).
(b) Find a confidence interval for 0 that has the same confidence coefficient as in
part (a) using the pivotal quantity (Yl + Y s )/2 - O.
(c) Compare the expected lengths of the confidence intervals of parts (a) and (b).
13 Find var [U] when F x (') - F·A·). See Eq. (20).
14 Equation (21) shows that Uand Txare linearly related. Find the exact distribution
of U or Tx when JIt' 0 is true for small sample sizes. For example, take m = 1,
n = 2; m = 1, n = 3; m = 2, n = 1 ; m = 3, n = 1; and m = n = 2.
15 We saw that G[U] = mnp. Is U/mn an unbiased estimator of p = P[Xt > Y j ]
whether or not JIt' 0 is true? Is U a consistent estimator of p?
16 A common measure of association for random variables X and Y is the rank
correlation, or Spearman's correlation. The X values are ranked, and the observa-
tions are replaced by their ranks; similarly the Yobservations are replaced by their
ranks. For example, for a sample of size 5 the observations
x 20.4 19.7 21.8 20.1 20.7
y 9.2 8.9 11.4 9.4 10.3

526 NONPAllAMETRIC METHODS XI
are replaced by
r(x) 3 1 5 2 4
r(y) 2 1 5 3 4
Let r(Xt ) denote the rank of XI and r( Y t ) the rank of Y t • Using these paired
ranks, the ordinary sample correlation is computed:
[r(XI) - i(X)][r( Yt) - i( Y)]
Spearman's correlation = S = -::::========;;;;:====== ,
[r(XI) - i(X)]Z [r( Yt) - i( Y)]2
where f(X) = 2: r(Xt)!n and f(Y) = 2: r(Yt)/n.

(a) Show that S = 1 - 6 2: D1/(n 3 - n), where D t = r(Xt) - r(Yt ).
(b) Compute the ordinary correlation and Spearman's correlation for the above
data.
17 ArgUe that the distribution of S in Prob. 16 is independent of the form of the
distributions of X and Y provided that X and Yare continuous and independently
distributed random variables. Hence S can be used as a test statistic in a non-
parametric test of the null hypothesis of independence.
18 Show that the mean and variance of S (in Prob. 17) under the hypothesis of inde-
pendence are 0 and 1/(n - 1), respectively.
APPENDIX A
MATHEMATICAL ADDENDUM
1 INTRODUCTION
The purpose of this appendix is to provide the reader with a ready reference to some
mathematical results that are used in the book. This appendix is divided into two
main sections: The first, Sec. 2 below, gives results that are, for the most part, com-
binatorial in nature, and the last gives results from calculus. No attempt is made to
prove these results, although sometimes a method of proof is indicated.
2 NONCALCULUS
2.1 Summation and Product Notation

A sum of terms such as n3 + n4 + ns + n6 + n7 is often designated by the symbol
7
~ ni. ~ is the capital Greek letter sigma, and in this connection it is often called the
1=3
summation sign. The letter I is called the summation index. The term fol1owing L is
called the summand. The" i = 3 " below ~ indicates that the first term of the sum is
obtained by putting i = 3 in the summand. The" 7 " above the L indicates that the
528 MATHEMATICAL ADDENDUM APPENDIX A
final term of the sum is obtained by putting i = 7 in the summand. The other terms
of the sum are obtained by giving ithe integral values between the limits 3 and 7. Thus
5
L: (-I)J-2jx1J = 2x4 - 3x6 + 4x8 - 5x10 •
J=2
An analogous notation for a product is obtained by substituting the capital

Greek letter n for L:. In this case the terms resulting from substituting the integers
for the index are multiplied instead of added. Thus
EXAMPLE 1 Some useful formulas involving summations are listed below.

They can be proved using mathematical induction.
i
l.d
i = n(n + 1) .
2
(1)
i i2 = n(n + 1)(2n + 1) . (2)

1=1 6
~ ;4 = n(n + 1)(2n + 1)(3n 2 + 3n - 1).

t~1 30 (4)
Equation (1) can be used to derive the following formula for an arithmetic series
or progression:
" [a + (j - d
L:
J=1
l)d] = na 2n(n-l). (5)
A companion series, the finite geometric series, or progression, is given by

II-I 1 -r"
L: ar} = a
J-O 1- r
. (6)
II1I
2.2 Factorial and Combinatorial Symbols and Conventions

A product of a positive integer n by all the positive integers smaller than it is usually
denoted by n! (read" n/actorial"). Thus
n! = n(n - 1)(n - 2) .... ·1= n (n -

II-I
J=O
j). (7)
O! is defined to be 1.
2 NONCALCULUS 529
A product of a positive integer n by the next k - 1 smaller positive integers is

usually denoted by (n)" ' Thus
(n)" = n(n 1) ..... (n - k + 1)
= n" (n j 1), (8)

J=1
Note that there are k terms in the product in Eq. (8).
Remark (n)" = n!/(n - k)!, and (n)n = n!/O! = nJ. The combinatorial symbol
(;) is defined as follows:
n) = (nh n!
(k (9)
k! (n k)!k!'
(;) is read .. combination of n things taking k at a time" or more briefly as
" n pick k"; it is also called a binomial coefficient. Define
if k<O or k n. (10)
11/1
Remark
(~) = (:) = 1.
(;) = (n ~k)'
(n : I) = (~) + (k ~ I) for n 1, 2, •.. and k = 0, ± 1, 2, ., ..
(11)
Equation (11) is a useful recurrent formula that is easily proved. /III
Both (n). and the combinatorial symbol (;) can be generalized from a positive integer
n to any real number t by defining ,
/) t(t - 1) ..... (t - k + 1)
(I)" = t(1 - 1) ..... (I - k 1),
(k k!
for k = 1, 2, ... , (12)
and (~) = !for k = O.

- n) = (- n)( - n - 1) ..... ( - n ~ k + 1)
Remark ( k k!
n(n 1) ..... (n +k - 1)
= (-1)~ - - - - k - ! - - - -
= (-I)' r k-1) k . II/I
2.3 Stirling's Formula

In finding numerical values of probabilities, one is often confronted with the evaluation
of long factorial expressions which can be troublesome to compute by direct multiplica-
tion. Much labor may be saved by using Stirling's formula, which gives an approximate
value of n!. Stirling's formula is
or
where 1 - 1/(12n 1) < r(n) < 1. To indicate the accuracy of Stirling's formula, 10!
was evaluated using five-place logarithms and Eq. (13), and 3,599,000 was obtained.
The actual value of 10! is 3,628,800. The percent error is less than 1 percent, and the
percent error will decrease as n increases.
2.4 The Binomial and Multinomial Theorems

The binomial theorem is often given as
(a i (~)] aJb",-J
+ b)" = J=O (15)
for n, a positive integer. The binomial theorem explains why the (;) are sometimes
called binomial coefficients. Four special cases are noted in the following remark.
Remark (16)
(1 - t)'" =
}."O
i (~)
J
(-1)JtJ, (17)
2'" =
}""O
i (~),
J
(18)
3 CALCULUS 531
and
0= i: (-1)1(~).
}=o J
(19)
1111
Expanding both sides of
(l + x)"(l + xt = (l + X)Hb
and then equating coefficients of x to the nth power gives
}=o
i (a)j (n-bj ) = (a +nb)' (20)
a formula that is particularly useful in considerations of the hypergeometric distribution.

A generalization of the binomial theorem is the multinomial theorem, which is
( .2")" IT n a;',
n'"
a} =.2,,' (21)
}=1 1",1
nd
l= 1
where the summation is over all nonnegative integers nl, nz, ... , n" which sum to n.
A special case is
Also note that
3 CALCULUS
3.1 Preliminaries
It is assumed that the reader is familiar with the concepts of limits, continuity, differenti~
ation, integration, and infinite series. A particular limit that is referred to several
times in the book is the limit expression for the number e; that is,
lim (l X)l/X = e. (24)

X"" 0
Equation (24) can be derived by taking logarithms and utilizing l'Hospital's rule,
which is reviewed below. There are a number of variations of Eq. (24)t for instance,
lim (1
x-+ a:J
+x 1)~ =e (25)
and
lim (1 + Ax)1/x = e). for constant A. (26)
X'" 0
A rule that is often useful in finding limits is the following so-called I'Hospital's
rule: If f( • ) and g( . ) are functions for which lim f(x) = lim g(x) = 0 and if
' f'(x)
IIm--
x-+a g'(x)
exists, then so does

, f(x)
II m -
x ... a g(x) ,
and
lim f(x) = lim f'(x)
x-+a g(x) x'" a g'(x) .
EXAMPLE 2 Find lim [(llx) lo~ (1 x)]. Let f(x) = lo~ (l x) and g(x) = x;
"" ... 0
then
= I = lim [~IO~ (l + X)] .

, f'(x)
IIm--
x'" 0 g'(x)
lim
x-+ 0
_1_
1+x x'" 0 X
IIII
Another rule that we use in the book is Leibniz' rule for differentiating an integral:
Let
1(1) = J lI(t)
(I(t)
f(x; t) dx,
where f(· ; . ), g( . ), and h( • ) are assumed differentiable. Then

dl
- =
dt
J- lI(t)
(I(t)
of
ot
dh
dx+ f(h(t); t)-
dt
dg
f(g(t);t) dt . (27)
Several important special cases derive from Leibniz' rule; for example, if the
integrandf(x; t) does not depend on t, then
d [ dh dg
dt {(t)
h(t)
f(x) dx
]
= f(h(t» dt f(o(t)) dt ; (28)
in particular, if g(l) is constant and h(t) = t, Eq. (28) simplifies to
~ [f [(x) dx] = [(/), (29)

3 CALCULUS 533
3.2 Taylor Series

The Taylor series for [(x) about x = a is defined as
[(2)(a)(x - a)2 [(")(a)(x - a)"
[(x) = [(a) + [(1)(a)(x - a) + 2! ---,--+R", (30)
n.
where
and a<c<x.
R" is called the remainder. f(x) is assumed to have derivatives of at least order n + l.
If the remainder is not too large, Eq. (30) gives a polynomial (of degree n) approxima w
tion, when R" is dropped, of the function [(.). The infinite series corresponding to
Eq. (30) will converge in some interval if lim Rn 0 in this interval. Several important
,,-+ <Xl
infinite Taylor series, along with their intervals of convergence, are given in the follow-
ing examples.
EXAMPLE 3 Suppose [(x) = eX and a O. Then

X2 x"
~= 1 +x+-
21 31
<Xl x)
=2-j! )=0
for - 00 <x< 00. (31)
1II1
EXAMPLE 4 Suppose [(x) = (1- x)t and a = 0; then [(1)(x) = t(1 x)t-l,
[(2)(X) = t(t l)(1 X)t-2, .•. , [U)(x) =( -l)Jt(t - l)'" ··(t - j + 1)(1 - x)t- J,
and hence
<Xl x)
[(x) (l-xr= 2
)=0
(-1)1(1»)-.
J!
= i (t.)(_X)J
)=0 ]
for -1 <x < 1. (32)
I111
There are several interesting special cases of Eq. (32). t = - n gives
i-I)
i
xJ for -1 <x < 1; (33)
t = - 1 gives the geometric series

<Xl
(1 - X)-l = L
):;0
xJ; (34)
534 MATHEMATICAL ADDENDUM APPENDlX A
t -2 gives
«l
(1 - X)-2 ~ U + l)xJ • (35)

J-O
EXAMPLE 5 Suppose f(x) l0ie (1 x) and a = 0; then

x2
lo~ (1 + x) = x - "2 for -1 <x< 1. (36)
II/!
The Tay lor series for functions of one variable given in Eq. (30) can be generalized
to the Taylor series for functions of several variables. For example, the Taylor series
for f(x, y) about x = a and y = b can be written as
f(x,y) f(a, b) + f;t;(a, b)(x - a) ~(a, b)(y - b) +
1
2! £f;t;;t;(a, b)(x - a)2 + 2fx,,(a, b)(x a)(y - b) + f.".,,(a, b)(y - b)2] + ... ,
where
and similarly for the others.
3.3 The Gamma and Beta Functions

The gamma function, denoted by r( . ), is defined by
r(t) = J xt -Ie-;t; dx
«l
for t > O. (37)
o
r(t) is nothing more than a notation for the definite integral that appears on the right-
hand side of Eq. (37). Integration by parts yields
r(t + 1) = tr(t), (38)

and, hence, if t = n (an integer),
r(n + 1) = n!. (39)
If n is an integer,
1 . 3 . 5 ..... (2n 1) ~ I-
r(n +!) = 211 V 1T, (40)
3 CALCULUS 535
and, in particular,
The beta function, denoted by B( ., .), is defined by

1
B(a, b) = 10 xa-1(l - X)b-l dx for a > 0, b > O. (42)
Again, B(a, b) is just a notation for the definite integral that appears on the right-hand
side of Eq. (42). A simple variable substitution gives B(a, b) = B(b, a). The beta
function is related to the gamma function according to the following formula =
r(a)r(b)
B(a, b) = r(a + b)· (43)
APPENDIX B
TABULAR SUMMARY OF PARAMETRIC FAMILIES
OF DISTRIBUTIONS
1 INTRODUCTION
The purpose of this appendix is to provide the reader with a convenient reference
to the parametric families of distributions that were introduced in Chap. III. Given
are two tables, one for discrete distributions and the other for continuous distribu~
tions.
538 PARAMETRIC FAMILIES OF DISTRIBUTIONS APPENDIX B
Table 1 DISCRETE DISTRIBUTIONS
Name of
parametric
family of Parameter Mean
distributions Discrete density functions f(·) Space fL = ~[X]
Discrete 1 N+1
uniform f(x) =N l{l • .... N}(X) N= 1,2, ...
2
Bernoulli f(x) = pXql-xl{o. l}(x) O<p<l p

(q = 1- p)
Binomia1 l(x) =(~)PXq"-XI{o. 1 •••.• n,{x) O<p< 1 np

n = 1,2,3, .,.
(q = 1 - p)
(~)e\!=:) K
Hypergeometric flx) = (~) l{o. 1 ..... n}(X) M=1,2, ... n-
M
K = 0, 1, ... , M
n=1,2, ... ,M
e-AAx
Poisson f(x) =-,-l{o. 1 • ••• )(x) A>O A
x.
q
Geometric f(x) = p¢l(o. 1. ... }(x) O<p<l
p
(q = 1- p)
Negative rq
binomial
f()
x -_(r+x-1)
x p '¢/(0.1 •••• ) (x) O<p<l
p
r> 0
(q = 1 - p)
t
Moment
generating
Variance Moments fL: = tf[xr] or fLr = tf[(X - fLY] function
a 2 =tf[(X - fL)2] and/or cumu]ants Kr tf[e fX ]
N2-1 , N(N + 1)2

12
fL3 = 4
(N + 1)(2N + 1)(3N + 3N - 2
1)
30
pq fL: = p for an r
npq fL3 = npq(q - p)

fL4 = 3n 2p2q2 npq(1- 6pq) +
K M-KM-n
nM M M-1
"[X(X - I) 0 0 o(X - r + I)J ~ r! (~y) not useful
Kr=A for r = 1,2, '" exp[A(e t -l)]

fL3 =A
fL4 = A + 3A2
q q+q2 P
p2 fL3 = ---pz- 1- qe t
q+ 7q2 +q3
fL4 = p4
rq r(q + q2)
p2 fL3 = p3 (1 !qetf
r[q + (3r + 4)q2 + q3]

fL4 = p4
\
S40 PARAMETRIC FAMILIES OF DISTRIBUTIONS APPENDIX B
Table 2 CONTINUOUS DISTRIBUTIONS
Name of
parametric
family of Cumulative distribution function F(o) Parameter Mean
distributions or probability density functionf(o) space p, = tB'[X]
Uniform or 1 a+b
rectangular f(x) = -b--a l(a,b](x) -00< aO
"I~"'" .i
-'
1
Exponential f(x) = >"e-A:X:I(oo~)(x) >">0
X
>..r >">0 r
Gamma f(x) = r(r)x"-le-A:X:I(oo~)(x) r>O X
1 a>O a
Beta f(x) = B(a, b)x"-l(l- x)b-ll(o.l)(X)
b>O a+b
1 -00< ex < 00 Does not

Cauchy f(x) = 1T~{1 + [(x - ex)/PJ2} ~>O exist
Lognormal f(x) = -00 < p,< 00 exp[p.+ lu2]

u>O
1
v2 exp[-(1o8eX-P,)2/2u2]I(o.~)(x)
x 1TU
Double 1 (Ix-ex') -oo<ex<oo ex

exponential f(x) =2~ exp - ~ ~> 0
2 CONTINUOUS DIsTRmunONs 541
Moments fL; = rS'[XT]

Variance or fLr = rS'[(X - fLY] Moment generating
2
a____rS'_[_(X
____fL_)2_]____a_n-d-/o-r-c-u-m-u-1-an-t-s-K-r----____________f_un_ct
__ io_n_rS'_[_e_tX_]_________ ~
(b - a)2 fLr = 0 for r odd

12 (b - a)t
(b-aY
fLr = 2r(r + 1) for r even
r! U
fLr = 0, r odd; fLr = (r/2)! 2r/2' r even; exp[fLt + I (J2 t 2]
Kr= 0, r > 2
1 r(r+ 1)
n'-
,r- Ar " for t<"
"2 A-t
, r(r + j)
fLl = "lr(r)
ab B(r + a, b)
not usefu1
(a + b + 1)(a + b)2 fLr = B(a, b)
Does not exist Do not exist Characteristic function

is e'lZt-Pltl
exp[2fL + 2a2 ] not usefu1

-exp[2fL +a 2 ]
fLr = 0 for r odd; e lZt

fLr = r! f3r for r even 1 - (f3t)2
( continued)
542 PARAMETRIC FAMIUES OF DISTRIBUTIONS APPPENDIX B
Table 2 CONTINUOUS DISTRIBUTIONS (continued)
Name of
parametric
family of Cumulative distribution function F(') Parameter Mean
distributions or probability density functionf(·) space p. = rS'[X]
a>O
Weibul1 f(x) = abx"-lexp[-axbJI(o,oo)(x) a- 1IbI'(1 + b- 1 )
b>O
e -(x- Cl)/II]-1 -oo<ex<OO

Logistic F(x) = [1 {3> 0
ex
8x& Xo> 0 8xo

Pareto f(x) = x'+1 1(Jl:o'oo)(x) 8>0 8-1
for 8> 1
Gumbel or F(x) = exp (_e-(JI:-IX)/II) -oo<ex<OO ex + {3y,

extreme value {3>0 y ~ .577216
I'[(k + 1)/2] 1 1 p.=O

t distribution f(x) = k>O
I'(k/2) V k1r (1 + x 2/k)(k +1)12 for k> 1
,
f(x) = I'[(m + n)/2] (m)m/2 n

F distribution m, n = 1. 2, ... n-2
I'(m/2)I'(n/2) n
x(m-Z)/2 for n > 2
x [1 + (m/n)x](m+n)/Z l(o.oo)(x)
~~
7~~
Chi-square f(x)= 1 (~kIZx"IZ-le-(l/2)XI (x) k 1,2, ... k
,
distribution I'(kj2) (0,00)
, 1lI
"
Moments p.: = 8[X']

Variance or p., = 8[(X - p.)"] Moment generating
a2 = t8'[(X - p.)Z] and/or cumuJants K, function &[e tx ]
a- Z1b [r(1 + 2b- 1)

p.: = a-r1br( 1 +~) t8'[Xt] = a-rfbr( 1 + i)
- rZ(1 + b- 1)]
fl Z1J'2
e llt 1J'fll csc(1J'flt)
3
8xS 8x'O does not

p.:=O for 8> r exist
(8 - 1)2(8 - 2) -r
for 8> 2
1J'Z{32 Kr = ( - fly~(r-l)(l)
for r 2, ellfrO - fl/)
6 where ~(.) is digamma function for t < 1/{3
p'r = 0 for k> r and r odd

k
kr/ZB«r + 1)/2, (k - r)/2)
i
does not (
k-2 p.r= exist
B(!, k/1J
for k > 2
for k> rand r even
2n Z(m+n 2) , =(!:)"
p., m
r(m/2+ r)r(n/2
r(m/2)r(n/2)
r) does not
exist
m(n - 2)2(n - 4)
n
for n > 4 for r <2
2Jr(k/2 + j)
f _
(~r/2
2k p.) - r(k/2) 1 21
for t< 1/2
APPENDIX C
REFERENCES AND RELATED READING
MATHEMATICS BOOKS
1. PR01TER and MORREY: "Calculus with Analytic Geometry: A Second Course,"
Addison-Wesley Publishing Company, Inc., Reading, Mass., 1971.
2. THOMAS: "Calculus and Analytic Geometry," alternate ed., Addison-Wesley Pub-
lishing Company, Inc., Reading, Mass., 1972.
3. WIDDER: "Advanced Calculus," 2d ed., Prentice-Hall, Inc., Englewood Cliffs,
N.J., 1961.
4. WYLIE: "Advanced Engineering Mathematics," 3d ed., McGraw-Hill Book
Company, New York, 1966.
PROBABILITY BOOKS
5. "Basic Probability Theory," John Wiley & Sons, Inc., New York, 1970.
ASH:
6. DRAKE: "Fundamentals of Applied Probability Theory," McGraw-Hill Book
Company, New York, 1967.
7. DWASS: "Probability: Theory and Applications," W. A. Benjamin, Inc., New York,
1970.
PROBABILITY AND STATISTICS BOOKS 545
8. FELLER: "An Introduction to Probability Theory and Its Applications," Vol. 1,

3d ed., John Wiley & Sons, Inc., New York, 1968.
9. FELLER; "An Introduction to Probability Theory and Its Applications," Vol. 2,
John Wiley & Sons, Inc., New York, 1966.
10. PARZEN: "Modern Probability Theory and Its Applications," John Wiley & Sons,
Inc., New York, 1960.
PROBABILITY AND STATISTICS BOOKS
Advanced (more advanced than MOB)
11. CRAMER: "Mathematical Methods of Statistics," Princeton University Press,

Princeton, N.J., 1946.
12. FERGUSON: "Mathematical Statistics, A Decision Theoretic Approach," Academic
Press, Inc., New York, 1967.
13. KENDALL and STUART: "The Advanced Theory of Statistics, Vol. 1, Distribution
Theory," 2d ed., Hafner Publishing Company, Inc., New York, 1963.
14. KENDALL and sTUART: "The Advanced Theory of Statistics, Vol. 2, Inference and
Relationship," Hafner Publishing Company, Inc., New York, 1961.
15. KENDALL and STUART: "The Advanced Theory of Statistics, Vol. 3, Design and
Analysis, and Time Series," Hafner Publishing Company, Inc., New York, 1966.
16. LEHMANN: "Testing Statistical Hypotheses," John Wiley & Sons, Inc., New York,
1959.
17. RAO:" Linear Statistical Inference and Its Applications," John Wiley & Sons, Inc.,
New York, 1965.
18. WILKS: "Mathematical Statistics," John Wiley & Sons, Inc., New York, 1962.
19. ZACKS:" The Theory of Statistical Inference," John Wiley & Sons, Inc., New York,
1971.
Intermediate (about the same level as MOB)
20. BRUNK: "An Introduction to Mathematical Statistics," 2d ed., Blaisdell Publishing

Company, a division of Ginn and Company, Waltham, Mass., 1965.
21. DWASS: "Probability and Statistics," W. A. Benjamin, Inc., New York, 1970.
22. HOGG and CRAIG: " Introduction to Mathematical Statistics," 3d ed., The Macmillan
Co. of Canada, Limited, Toronto, 1970.
23. LINDGREN: "Statistical Theory," 2d ed., The Macmillan Company, New York,
1968.
24. TUCKER: "An Introduction to Probability and Mathematical Statistics," Academic
Press, Inc., New York, 1962.
546 REFERENCES AND RELATED READING APPENDIX C
Elementary (less advanced than MOB, but calculus prerequisite)

25. "Mathematical Statistics," 2d ed., Prentice-Hall, Inc., Englewood Cliffs,
FREUND:
N.J., 1971.
26. GUlTMAN, WILKS, and HUNTER: " Introductory Engineering Statistics," 2d ed., John
Wiley & Sons, Inc., New York, 1971.
27. HOEL:" Introduction to Mathematical Statistics," 3d ed., John Wiley & Sons, Inc.,
New York, 1971.
28. HOEL, PORT, and STONE: "Introduction to Statistical Theory," Houghton Mifflin
Company, Boston, 1971.
SPECIAL BOOKS
29. DANIEL and WOOD: "Fitting Equations to Data; Computer Analysis of Multi-
factor Data for Scientists and Engineers," Interscience Publishers, a division of
John Wiley & Sons, Inc., New York, 1971.
30. DAVID: "Order Statistics," John Wiley & Sons, Inc., New York, 1970.
31. DRAPER and SMITH: "Applied Regression Analysis," John Wiley & Sons, Inc.,
New York, 1966.
32. GIBBONS: "Nonparametric Statistical Inference," McGraw..Hill Book Company,
New York, 1971.
33. GRAYBILL: "An Introduction to Linear Statistical Models," Vol. 1, McGraw-Hill
Book Company, New York, 1961.
34. JOHNSON and KOTZ:" Discrete Distributions," Houghton Mifflin Company, Boston,
1969.
35. JOHNSON and KOTZ: "Continuous Univariate Distributions-I," Houghton Mifflin
36. JOHNSON and KOTZ: "Continuous Univariate Distributions-2," Houghton Mifflin
37. KEMPTHORNE and FOLKS: U Probability, Statistics, and Data Analysis," The Iowa
State University Press, Ames, 1971.
38 . .MORRISON: "Multivariate Statistical Methods," McGraw-Hill Book Company,
New York, 1967.
39. RAJ: "Sampling Theory," McGraw-Hill Book Company, New York, 1968.
PAPERS
40. JOINER and ROSENBLATT: "Some Properties of the Range in Samples from Tukey's
Symmetric Lambda Distributions," Journal 0/ the American Statistical Associa·
lion, Vol. 66 (1971), 394-399.
BOOKS OF TABLES 547
41. PITMAN: "The Estimation of the Location and Scale Parameters of a Continuous
Population of Any Given Form," Biometrika, Vol. 30 (1939), pp. 391--421.
42. WOLFOWITZ:" The Minimum Distance Method," Annalso/Mathematical Statistics,
Vol. 28(1) (1957), pp. 75-88.
43. ZEHNA:" Invariance of Maximum Likelihood Estimation ,"Annals 0/ Mathematical
Statistics, Vol. 37 (1966), p. 744.
BOOKS OF. TABLES

44. "Handbook of Tables for Probability and Statistics," Chemical Rubber Company,
Cleveland, 1966.
45. MOLINA: "Poisson's Exponential Binomial Limit," D. Van Nostrand Company,
Inc., Princeton, N.J., 1942.
46. OWEN: "Handbook of Statistical Tables," Addison-Wesley Publishing Company,
Inc., Reading, Mass., 1962.
APPENDIX D
TABLES
1 DESCRIPTION OF TABLES
Table 1 Ordinates of the Normal Density Function

This table gives values of
for values of x between 0 and 4 at intervals of .01. For negative values of x one uses
the fact that r$( - x) = r$(x).
Table 2 Cumulative Normal Distribution

This table gives values of
for values of x beteenn 0 and 3.5 at intervals of .01. For negative values of x, one uses
1 DESCRIPTION OF TABLES 549
the relation tP( - x) = 1 - tP(x). Values of x corresponding to a few special values of tP

are given separately beneath the main table.
Table 3 Cumulative Chi-Square Distribution

This table gives values of u corresponding to a few selected values of F(u), where
" x<n-2){2e-X/2 dx
F(u)
f
o 2n/2r(n/2)
for n, the number of degrees of freedom, equal to 1, 2, ... , 30. For larger values of n,
a normal approximation is quite accurate. The quantity V2u V2n 1 is nearly
normal1y distributed with mean 0 and unit variance. Thus U/t, the exth quantile point
of the distribution, may be computed by
where Zit is the exth quantile point of the cumulative normal distribution. As an ilJustra·
tion, we may compute the .95 value of U for n = 30 degrees of freedom:
U.95 ,(1.645 + 1/59)2

43.S,
which is in error by Jess than 1 percent.
Table 4 Cumulative F Distribution

This table gives values of F corresponding to five values of
for selected values of m and n; m is the number of degrees of freedom in the numerator
of F, and n is the number of degrees of freedom in the denominator of F. The table
also provides values corresponding to G = .10, .OS, .025, .01, and .005 because F 1 - 1t
for m and n degrees of freedom is the reciprocal of Fa for nand m degrees of freedom.
Thus for G = .OS with three and six degrees of freedom, one finds
1 1
F.os(3, 6) = F.9s(6, 3) = 8.94 = .112
One should interpolate on the reciprocals of m and n as in Table S for good accuracy.
550 TABLES APPENDIX D
Table S Cumulative Students t Distribution

This table gives values of t corresponding to a few selected values of
with n = 1, 2, ... , 30, 40, 60, 120, 00. Since the density is symmetrical in t, it follows
that F( - t) = 1 - F(t). One should not interpolate linearly between degrees of
freedom but On the reciprocal of the degrees of freedom, if good accuracy in the last
digit is desired. As an illustration, we shall compute the .975th quantile point for
40 degrees of freedom. The values for 30 and 60 are 2.042 and 2.000. Using the
reciprocals of n, the interpolated value is
1 1
ao--Co
2.042 - 1 1 (2.042 - 2.000) = 2.021,
30-eo
which is the correct value. Interpolating linearly, one would have obtained 2.028.
DESCRIPTION OF TABLES 551
Table 1 ORDINATES OF THE NORMAL DENSITY FUNCTION

1
.(x) = -= e- x 1./2
V21T
x .00 .01 .02 .03 .04 .05 .06 .07 .08 .09
.0 .3989 .3989 .3989 .3988 .3986 .3984 .3982 .3980 .3977 .3973
.1 .3970 .3965 .3961 .3956 .3951 .3945 .3939 .3932 .3925 .3918
.2 .3910 .3902 .3894 .3885 .3876 .3867 .3857 .3847 .3836 .3825
.3 .3814 ' .3802 .3790 .3778 .3765 .3752 .3739 .3725 .3712 .3697
.4 .3683 .3668 .3653 .3637 .3621 .3605 .3589 .3572 .3555 .3538
.5 .3521 .3503 .3485 .3467 .3448 .3429 .3410 .3391 .3372 .3352
.6 .3332 .3312 .3292 .3271 .3251 .3230 .3209 .3187 .3166 .3144
.7 .3123 .3101 .3079 .3056 .3034 .3011 .2989 .2966 .2943 .2920
.8 .2897 .2874 .2850 .2827 .2803 .2780 .2756 .2732 .2709 .2685
.9 .2661 .2637 ,2613 .2589 .2565 .2541 .2516 .2492 .2468 .2444
1.0 .2420 .2396 .2371 .2347 .2323 .2299 .2275 .2251 .2227 .2203
1.1 .2179 .2155 .2131 .2107 .2083 .2059 .2036 .2012 .1989 .1965
1.2 .1942 .1919 .1895 .1872 .1849 .1826 .1804 .1781 .1758 .1736
1.3 .1714 .1691 .1669 .1647 .1626 .1604 .1582 .1561 .1539 .15'18
1.4 .1497 .1476 .1456 .1435 .1415 .1394 .1374 .1354 .1334 .1315
1.5 .1295 .1276 .1257 .1238 .1219 .1200 .1182 .1163 .1145 .1127
1.6 .1109 .1092 .1074 .1057 .1040 .1023 .1006 .0989 .0973 .0957
1.7 .0940 .0925 .0909 .0893 .0878 .0863 .0848 .0833 .0818 .0804
1.8 .0790 .0775 .0761 .0748 .0734 .0721 .0707 .0694 .0681 .0669
1.9 .0656 .0644 .0632 .0620 .0608 .0596 .0584 .0573 .0562 .0551
2.0 .0540 .0529 .0519 .0508 .0498 .0488 .0478 .0468 .0459 .0449
2.1 .0440 .0431 .0422 .0413 .0404 .0396 .0387 .0379 .0371 .0363
2.2 .0355 .0347 .0339 .0332 .0325 .0317 .0310 .0303 .0297 .0290
2.3 .0283 •0277 .0270 .0264 .0258 . .0252 .0246 .0241 .0235 .0229
2.4 .0224 .0219 .0213 .0208 .0203 .0198 .0194 .0189 .0184 .0180
2.5 .0175 .0171 .0167 .0163 .0158 .0154 .0151 .0147 .0143 .0139
2.6 .0136 .0132 .0129 .0126 .0122 .0119 .0116 .0113 .0110 .0107
2.7 .0104 .0101 .0099 .0096 .0093 .0091 .0088 .0086 .0084 .0081
2.8 .0079 .0077 .0075 .0073 .0071 .0069 .0067 .0065 .0063 .0061
2.9 .0060 .0058 .0056 .0055 .0053 .0051 .0050 .0048 .0047 .0046
3.0 .0044 .0043 .0042 .0040 .0039 .0038 .0037 .0036 .0035 .0034
3.1 .0033 .0032 .0031 .0030 .0029 .0028 .0027 .0026 .0025 .0025
3.2 .0024 .0023 .0022 .0022 .0021 .0020 .0020 .0019 .0018 .0018
3.3 .0017 .0017 .0016 .0016 .0015 .0015 .0014 .0014 .0013 .0013
3.4 .0012 .0012 .0012 .0011 .0011 .0010 .0010 .0010 .0009 .0009
3.5 .0009 .0008 .0008 .0008 .0008 .0007 .0007 .0007 .0007 .0006
3.6 .0006 .0006 .0006 .0005 .0005 ,0005 .0005 .0005 .0005 ,0004
3.7 .0004 .0004 .0004 .0004 .0004 ,0004 .0003 .0003 .0003 .0003
3.8 .0003 .0003 .0003 .0003 .0003 .0002 .0002 .0002 .0002 .0002
3.9 .0002 .0002 .0002 .0002 .0002 .0002 .0002 .0002 .0001 .0001
552 TABLES APPENDIX D
Table 2 CUMULATIVE NORMAL DISTRIBUTION
x .00 .01 .02 .03 .04 .05 .06 .07 .08 .09
.0 .5000 .5040 .5080 .5120 .5160 .5199 .5239 .5279 .5319 .5359
.1 .5398 .5438 .5478 .5517 .5557 .5596 .5636 .5675 .5714 .5753
.2 .5793 .5832 .5871 .5910 .5948 .5987 .6026 .6064 .6103 .6141
.3 .6179 .6217 .6255 .6293 .6331 .6368 .6406 .6443 .6480 .6517
.4 .6554 .6591 .6628 .6664 .6700 .6736 .6772 .6808 .6844 .6879
.5 .6915 .6950 .6985 .7019 .7054 .7088 .7123 .7157 .7190 .7224
.6 .7257 .7291 .7324 .7357 .7389 .7422 .7454 .7486 .7517 .7549
.7 .7580 .7611 .7642 .7673 .7704 .7734 .7764 .7794 .7823 .7852
.8 .7881 .7910 .7939 .7967 .7995 .8023 .8051 .8078 .8106 .8133
.9 .8159 .8186 .8212 .8238 .8264 .8289 .8315 .8340 .8365 .8389
1.0 .8413 .8438 .8461 .8485 .8508 .8531 .8554 .8577 .8599 .8621
1.1 .8643 .8665 .8686 .8708 .8729 .8749 .8770 .8790 .8810 .8830
1.2 .8849 .8869 .8888 .8907 .8925 .8944 .8962 .8980 .82~7 . .9015
1.3 .9032 .9049 .9066 .9082 .9099 .9115 . 9131 .9147 .9162 .9177
1.4 .9192 .9207 .9222 .9236 .9251 .9265 .9279 .9292 .9306 .9319
1.5 .9332 .9345 .9357 .9370 .9382 .9394 .9406 .9418 .9429 .9441
1.6 .9452 .9463 .9474 .9484 .9495 .9505 .9515 .9525 .9535 .9545
1.7 .9554 .9564 .9573 .9582 .9591 .9599 .9608 .9616 «:9625 .9633
1.8 .9641 .9649 .9656 .9664 .9671 .9678 .9686 .9693 .9699 .9706
1.9 .9713 .9719 .9726 .9732 .9738 .9744 .9750 .9756 .9761 .9767
2.0 .9772 .9778 .9783 .9788 .9793 .9798 .9803 .9808 .9812 .9817
2.1 .9821 .9826 .9830 .9834 .9838 .9842 .9846 .9850 .9854 .9857
2.2 .9861 .9864 .9868 .9871 .9875 .9878 .9881 .9884 .9887 .9890
2.3 .9893 .9896 .9898 .9901 .9904 .9906 .9909 .9911 .9913 .9916
2.4 .9918 ~9920 .9922 .9925 .9927 .9929 .9931 .9932 .9934 .9936
2.5 .9938 .9940 .9941 .9943 .9945 .9946 .9948 .9949 .9951 .9952
2.6 .9953 .9955 .9956 .9957 .9959 .9960 .9961 .9962 .9963 .9964
2.7 .9965 .9966 .9967 .9968 .9969 .9970 .9971 .9972 .9973 .9974
2.8 .9974 .9975 .9976 .9977 .9977 .9978 .9979 .9979 .9980 .9981
2.9 .9981 .9982 .9982 .9983 .9984 .9984 .9985 .9985 .9986 .9986
3.0 .9987 .9987 .9987 .9988 .9988 .9989 .9989 .9989 .9990 .9990
3.1 .9990 .9991 .9991 .9991 .9992 .9992 .9992 .9992 .9993 .9993
3.2 .9993 .9993 .9994 .9994 .9994 .9994 .9994 .9995 .9995 .9995
3.3 .9995 .9995 .9995 .9996 .9996 .9996 .9996 .9996 .9996 .9997
3.4 .9997 .9997 .9997 .9997 .9997 .9997 .9997 .9997 .9997 .9998
x 1.282 1.645 1.960 2.326 2.576 3.090 3.291 3.891 4.417
(x) .90 .95 .975 .99 .995 .999 .9995 .99995 .999995
2[1 - (x)) .20 .10 .05 .02 .01 .002 .001 .0001 .00001
Table 3 CUMULATIVE CHI-SQUARE DISTRIBUTION·
F(u) = f: x(n~ 2.)~2~- X/~ dx
~ .005 .010 .025 .050 .100 .250 .500 .750 .900 .950 .975 .990 .995
1 .04 393 .0 3 157 .0 3 982 .0 2 393 .0158 .102 .455 1.32 2.71 5.02
2
3
.0100
.0717
.0201
.115
.0506
.216
.103
.352
.211
.584
.575
1.21
1.39
2.37
2.77
4.11
4.61
6.25
1*
.99
7.81
7.38
9.35
6.63
9.21
11.3
7.88
10.6
12.8
4 .207 .297 .484 .711 1.06 1.92 3.36 5.39 7.78 9.49 11.1 13.3 14.9
5 .412 .554 .831 1.15 1.61 2.67 4.35 6.63 9.24 11.1 12.8 15.1 16.7
6 .676 .872 1.24 1.64 2.20 3.45 5.35 7.84 10.6 12.6 14.4 16.8 18.5
7 .989 1.24 1.69 2.17 2.83 4.25 6.35 9.04 12.0 14.1 16.0 18.5 20.3
8 1.34 1.65 2.18 2.73 3.49 5.07 7.34 10.2 13.4 15.5 17.5 20.1 22.0
9 1.73 2.09 2.10, 3.33 4.17 5.90 8.34 11.4 14.7 16.9 19.0 21.7 23.6
10 2.16 2.56 3.25 . 3.94 4.87 6.74 9.34 12.5 16.0 18.3 20.5 23.2 25.2
11 2.60 3.05 3.82 4.57 5.58 7.58 10.3 13.7 17.3 19.7 21.9 24.7 26.8
12 3.07 3.57 4.40 5.23 6.30 8.44 11.3 14.8 18.5 21.0 23.3 26.2 28.3
13 3.57 4.11 5.01 5.89 7.04 9.30 12.3 16.0 19.8 22.4 24.7 27.7 29.8
14 4.07 4.66 5.63 6.57 7.79 10.2 13.3 17.1 21.1 23.7 26.1 29.1 31.3
15 4.60 5.23 6.26 7.26 8.55 11.0 14.3 18.2 22.3 25.0 27.5 30.6 32.8
16 5.14 5.81 6.91 7.96 9.31 11.9 15.3 19.4 23.5 26.3 28.8 32.0 34.3
17 5.70 6.41 7.56 8.67 10.1 12.8 16.3 20.5 24.8 27.6 30.2 33.4 35.7
18 6.26 7.01 8.23 9.39 10.9 13.7 17.3 21.6 26.0 28.9 31.5 34.8 37.2
19 6.84 7.63 8.91 10.1 11.7 14.6 18.3 22.7 27.2 30.1 32.9 36.2 38.6
20 7.43 8.26 9.59 10.9 12.4 15.5 19.3 23.8 28.4 31.4 34.2 37.6 40.0
21 8.03 8.90 10.3 11.6 13.2 16.3 20.3 24.9 29.6 32.7 35.5 38.9 41.4
22 8.64 9.54 11.0 12.3 14.0 17.2 21.3 26.0 30.8 33.9 36.8 40.3 42.8
23 9.26 10.2 11.7 13.1 14.8 18.1 22.3 27.1 32.0 35.2 38.1 41.6 44.2
24 9.89 10.9 12.4 13.8 15.7 19.0 23.3 28.2 33.2 36.4 39.4 43.0 45.6
25 10.5 11.5 13.1 14.6 16.5 19.9 24.3 29.3 34.4 37.7 40.6 44.3 46.9
26 11.2 12.2 13.8 15.4 17.3 20.8 25.3 30.4 35.6 38.9 41.9 45.6 48.3
27 11.8 12.9 14.6 16.2 18.1 21.7 26.3 31.5 36.7 40.1 43.2 47.0 49.6
28 12.5 13.6 15.3 16.9 18.9 22.7 27.3 32.6 37.9 41.3 44.5 48.3 51.0
29 13.1 14.3 16.0 17.7 19.8 23.6 28.3 33.7 39.1 42.6 45.7 49.6 52.3
30 13.8 15.0 16.8 18.5 20.6 24.5 29.3 34.8 40.3 43.8 47.0 50.9 53.7 v.
- ~- --
v.
!...oJ
• This table is abridged from "Tables of percentage points of the incomplete beta function and of the chi-square distribution," Biometrika,
Vol. 32 (1941). It is here published with the kind permission of its author, Catherine M. Thompson, and the editor of Biometrika.
Table 4 CUMULATIVE F DISTRIBUTION· (m degrees of freedom in numerator; n in denominator)
F r(m + n)m"'12 rf I2 x <m-2)/2(n + mx)-(m+n)/2dx

G(F) = fo 2 ..
G n m 1 2 3 4 5 6 7 8 9 10 12 15 20 30 60 120 00
.90 39.9 49.5 53.6 55.8 57.2 58.2 58.9 59.4 59.9 60.2 60.7 61.2 61.7 62.3 62.8 63.1 63.3
.95 161 200 216 225 230 234 237 239 241 242 244 246 248 250 252 253 254
.975 1 648 800 864 900 922 937 948 957 963 969 977 985 993 1000 1010 1010 1020
.99 4,050 5,000 5,400 5,620 5,760 5,860 5,930 5,980 6,020 6,060 6,110 6,160 6,210 6,260 6,310 6,340 6,370
.995 16,200 20,000 21,600 22,500 23,100 23,400 23,700 23,900 24,100 24,200 24,400 24,600 24,800 25,000 25,200 25,400 25,500
.90 8.53 9.00 9.16 9.24 9.29 9.33 9.35 9.37 9.38 9.39 9.41 9.42 9.44 9.46 9.47 9.48 9.49
.95 18.5 19.0 19.2 19.2 19.3 19.3 19.4 19.4 19.4 19.4 19.4 19.4 19.5 19.5 19.5 19.5 19.5
.975 2 38.5 39.0 39.2 39.2 39.3 39.3 39.4 39.4 39.4 39.4 39.4 39.4 39.4 39.5 39.5 39.5 39.5
.99 98.5 99.0 99.2 99.2 99.3 99.3 99.4 99.4 99.4 99.4 Q9.4 99.4 99.4 99.5 99.5 99.5 99.5
.995 199 199 199 199 199 199 199 199 199 199 199 199 199 199 199 199 199
.90 5.54 5.46 5.39 5.34 5.31 5.28 5.27 5.25 5.24 5.23 5.22 5.20 5.18 5.17 5.15 5.14 5.13
.95 10.1 9.55 9.28 9.12 9.01 8.94 8.89 8.85 8.81 8.79 8.74 8.70 8.66 8.62 8.57 8.55 8.53
.975 3 17.4 16.0 15.4 15.1 14.9 14.7 14.6 14.5 14.5 14.4 14.3 14.3 14.2 14.1 14.0 13.9 13.9
.99 34.1 30.8 29.5 28.7 28.2 27.9 27.7 27.5 27.3 27.2 27.1 26.9 26.7 26.5 26.3 26.2 26.1
.995 55.6 49.8 47.5 46.2 45.4 44.8 44.4 44.1 43.9 43.7 43.4 43.1 42.8 42.5 42.1 42.0 41.8
.90 4.54 4.32 4.19 4.11 4.05 4.01 3.98 3.95 3.93 3.92 3.90 3.87 3.84 3.82 3.79 3.78 3.76
.95 7.71 6.94 6.59 6.39 6.26 6.16 6.09 6.04 6.00 5.96 5.91 5.86 5.80 5.75 5.69 5.66 5.63
.975 4 12.2 10.6 9.98 9.60 9.36 9.20 9.07 8.98 8.90 8.84 8.75 8.66 8.56 8.46 8.36 8.31 8.26
.99 21.2 18.0 16.7 16.0 15.5 15.2 15.0 14.8 14.7 14.5 14.4 14.2 14.0 13.8 13.7 13.6 13.5
.995 31.3 26.3 24.3 23.2 22.5 22.0 21.6 21.4 21.1 21.0 20.7 20.4 20.2 19.9 19.6 19.5 19.3
.90 4.06 3.78 3.62 3.52 3.45 3.40 3.37 3.34 3.32 3.30 3.27 3.24 3.21 3.17 3.14 3.12 3.11
.95 6.61 5.79 5.41 5.19 5.05 4.95 4.88 4.82 4.77 4.74 4.68 4.62 4.56 4.50 4.43 4.40 4.37
.975 5 10.0 8.43 7.76 7.39 7.15 6.98 6.85 6.76 6.68 6.62 6.52 6.43 6.33 6.23 6.12 6.07 6.02
.99 16.3 13.3 12.1 11.4 11.0 10.7 10.5 10.3 10.2 10.1 9.89 9.72 9.55 9.38 9.20 9.11 9.02
.995 22.8 18.3 16.5 15.6 14.9 14.5 14.2 14.0 13.8 13.6 13.4 13.1 12.9 12.7 12.4 12.3 12.1
.90 3.78 3.46 3.29 3.18 3.11 3.05 3.01 2.98 2.96 2.94 2.90 2.87 2.84 2.80 2.76 2.74 2.72
.95 5.99 5.14 4.76 4.53 4.39 4.28 4.21 4.15 4.10 4.06 4.00 3.94 3.87 3.81 3.74 3.70 3.67
.975 6 8.81 7.26 6.60 6.23 5.99 5.82 5.70 5.60 5.52 5.46 5.37 5.27 5.17 5.07 4.96 4.90 4.85
.99 13.7 10.9 9.78 9.15 8.75 8.47 8.26 8.10 7.98 7.87 7.72 7.56 7.40 7.23 7.06 6.97 6.88
.995 18.6 14.5 12.9 12.0 11.5 11.1 10.8 10.6 10.4 10.2 10.0 9.81 9.59 9.36 9.12 9.00 8.88
.90 3.59 3.26 3.07 2.96 2.88 2.83 2.78 2.75 2.72 2.70 2.67 2.63 2.59 2.56 2.51 2.49 2.47
.95 5.59 4.74 4.35 4.12 3.97 3.87 3.79 3.73 3.68 3.64 3.57 3.51 3.44 3.38 3.30 3.27 3.23
.975 7 8.07 6.54 5.89 5.52 5.29 5.12 4.99 4.90 4.82 4.76 4.67 4.57 4.47 4.36 4.25 4.20 4.14
.99 12.2 9.55 8.45 7.85 7.46 7.19 6.99 6.84 6.72 6.62 6.47 6.31 6.16 5.99 5.82 5.74 5.65
.995 16.2 12.4 10.9 10.1 9.52 9.16 8.89 8.68 8.51 8.38 8.18 7.97 7.75 7.53 7.31 7.19 7.08
.90 3.46 3.11 2.92 2.81 2.73 2.67 2.62 2.59 2.56 2.54 2.50 2.46 2.42 2.38 2.34 2.31 2.29
.95 5.32 4.46 4.07 3.84 3.69 3.58 3.50 3.44 3.39 3.35 3.28 3.22 3.15 3.08 3.01 2.97 2.93
.975 8 7.57 6.06 5.42 5.05 4.82 4.65 4.53 4.43 4.36 4.30 4.20 4.10 4.00 3.89 3.78 3.73 3.67
.99 11.3 8.65 7.59 7.01 6.63 6.37 6.18 6.03 5.91 5.81 5.67 5.52 5.36 5.20 5.03 4.95 4.86
.995 14.7 11.0 9.60 8.81 8.30 7.95 7.69 7.50 7.34 7.21 7.01 6.81 6.61 6.40 6.18 6.06 5.95
.90 3.36 3.01 2.81 2.69 2.61 2.55 2.51 2.47 2.44 2.42 2.38 2.34 2.30 2.25 2.21 2.18 2.16
.95 5.12 4.26 3.86 3.63 3.48 3.37 3.29 3.23 3.18 3.14 3.07 3.01 2.94 2.86 2.79 2.75 2.71
.975 9 7.21 5.71 5.08 4.72 4.48 4.32 4.20 4.10 4.03 3.96 3.87 3.77 3.67 3.56 3.45 3.39 3.33
.99 10.6 8.02 6.99 6.42 6.06 5.80 5.61 5.47 5.35 5.26 5.11 4.96 4.81 4.65 4.48 4.40 4.31
.995 13.6 10.1 8.72 7.96 7.47 7.13 6.88 6.69 6.54 6.42 6.23 6.03 5.83 5.62 5.41 5.30 5.19
.90 3.29 2.92 2.73 2.61 2.52 2.46 2.41 2.38 2.35 2.32 2.28 2.24 2.20 2.15 2.11 2.08 2.06
.95 4.96 4.10 3.71 3.48 3.33 3.22 3.14 3.07 3.02 2.98 2.91 2.84 2.77 2.70 2.62 2.58 2.54
.975 10 6.94 5.46 4.83 4.47 4.24 4.07 3.95 3.85 3.78 3.72 3.62 3.52 3.42 3.31 3.20 3.14 3.08
.99 10.0. 7.56 6.55 5.99 5.64 5.39 ~.20 5.06 4.94 4.85 4.71 4.56 4.41 4.25 4.08 4.00 3.91
.995 12.8 9.43 8.08 7.34 6.87 6.54 6.30 6.12 5.97 5.85 5.66 5.47 5.27 5.07 4.86 4.75 4.64
.90 3.18 2.81 2.61 2.48 2.39 2.33 2.28 2.24 2.21 2.19 2.15 2.10 2.06 2.01 1.96 1.93 1.90
.95 4.75 3.89 3.49 3.26 3.11 3.00 2.91 2.85 2.80 2.75 2.69 2.62 2.54 2.47 2.38 2.34 2.30
.975 12 6.55 5.10 4.47 4.12 3.89 3.73 3.61 3.51 3.44 3.37 3.28 3.18 3.07 2.96 2.85 2.79 2.72
.99 9.33 6.93 5.95 5.41 5.06 4.82 4.64 4.50 4.39 4.30 4.16 4.01 3.86 3.70 3.54 3.45 3.36
.995 11.8 8.51 7.23 6.52 6.07 5.76 5.52 5.35 5.20 5.<)9 4.91 4.72 4.53 4.33 4.12 4.01 3.90
.90 3.07 2.70 2.49 2.36 2.27 2.21 2.16 2.12 2.09 2.06 2.02 1.97 1.92 1.87 1.82 1.79 1.76
.95 4.54 3.68 3.29 3.06 2.90 2.79 2.71 2.64 2.59 2.54 2.48 2.40 2.33 2.25 2.16 2.1 J. 2.07
.975 15 6.20 4.77 4.15 3.80 3.58 3.41 3.29 3.20 3.12 3.06 2.96 2.86 2.76 2.64 2.52 2.46 2.40
.99 8.68 6.36 5.42 4.89 4.56 4.32 4.14 4.00 3.89 3.80 3.67 3.52 3.37 3.21 3.05 2.96 2.87
.995 10.8 7.70 6.48 5.80 5.37 5.07 4.85 4.67 4.54 4.42 4.25 4.07 3.88 3.69 3.48 3.37 3.26
.90 2.97 2.59 2.38 2.25 2.16 2.<)9 2.04 2.00 1.96 1.94 1.89 1.84 1.79 1.74 1.68 1.64 1.61
.95 4.35 3.49 3.10 2.87 2.71 2.60 2.51 2.45 2.39 2.35 2.28 2.20 2.12 2.04 1.95 1.90 1.84
.975 20 5.87 4.46 3.86 3.51 3.29 3.13 3.0t 2.91 2.84 2.77 2.68 2.57 2.46 2.35 2.22 2.16 2.09
.99 8.10 5.85 4.94 4.43 4.10 3.87 3.70 3.56 3.46 3.37 3.23 3.09 2.94 2.78 2.61 2.52 2.42
.995 9.94 6.99 5.82 5.17 4.76 4.47 4.26 4.()9 3.96 3.85 3.68 3.50 3.32 3.12 2.92 2.81 2.69
.90 2.88 2.49 2.28 2.14 2.05 1.98 1.93 1.88 1.85 1.82 1.77 1.72 1.67 1.61 1.54 1.50 1.46
.95 4.17 3.32 2.92 2.69 2.53 2.42 2.33 2.27 2.21 2.16 2.09 2.01 1.93 1.84 1.74 1.68 1.62
.975 30 5.57 4.18 3.59 3.25 3.(\3 2.87 2.75 2.65 2.57 2.51 2.41 2.31 2.20 2.07 1.94 1.87 1.79
.99 7.56 5.39 4.51 4.02 3.70 3.47 3.30 3.17 3.07 2.98 2.84 2.70 2.55 2.39 2.21 2.11 2.01
.995 9.18 6.35 5.24 4.62 4.23 3.95 3.74 3.58 3.45 3.34 3.18 3.01 2.82 2.63 2.42 2.30 2.18
.90 2.79 2.39 2.18 2.04 1.95 1.87 1.82 1.77 1.74 1.71 1.66 1.60 1.54 1.48 . 1.40 1.35 1.29
.95 4.00 3.15 2.76 2.53 2.37 2.25 2.17 2.10 2.04 1.99 1.92 1.84 1.75 1.65 1.53 1.47 1.39
.975 60 5.29 3.93 3.34 3.01 2.79 2.63 2.51 2.41 2.33 2.27 2.17 2.06 1.94 1.82 1.67 1.58 1.48
.99 7.08 4.98 4.13 3.65 3.34 3.12 2.95 2.82 2.72 2.63 2.50 2.35 2.20 2.03 1.84 1.73 1.60
.995 8.49 5.80 4.73 4.14 3.76 3.49 3.29 3.13 3.01 2.90 2.74 2.57 2.39 2.19 1.96 1.83 1.69
.90 2.75 2.35 2.13 1.99 1.90 1.82 1.77 1.12 1.68 1.65 1.60 1.54 1.48 1.41 1.32 1.26 1.19
.95 3.92 3.07 2.68 2.45 2.29 2.18 2.09 2.02 1.96 1.91 1.83 1.75 1.66 1.55 1.43 1.35 1.25
.975 120 5.15 3.80 3.23 2.89 2.67 2.52 2.39 2.30 2.22 2.16 2.05 1.94 1.82 1.69 1.53 1.43 1.31
.99 6.85 4.79 3.95 3.48 3.17 2.96 2.79 2.66 2.56 2.47 2.34 2.19 2.03 1.86 1.66 1.53 1.38
.995 8.18 5.54 4.50 3.92 3.55 3.28 3.09 2.93 2.81 2.71 2.54 2.37 2.19 1.98 1.75 1.61 1.43
.90 2.71 2.30 2.08 1.94 1.85 1.77 1.72 1.67 1.63 1.60 1.55 1.49 1.42 1.34 1.24 1.17 1.00
.95 3.84 3.00 2.60 2.37 2.21 2.10 2.01 1.94 1.88 1.83 1.75 1.67 1.57 1.46 1.32 1.22 1.00
.975 «> 5.02 3.69 3.12 2.79 2.57 2.41 2.29 2.19 2.11 2.05 1.94 1.83 1.71 1.57 1.39 1.27 1.00
.99 6.63 4.61 3.78 3.32 3.02 2.80 2.64 2.51 2.41 2.32 2.18 2.04 1.88 1.70 1.47 1.32 1.00
.995 7.88 5.30 4.28 3.72 3.35 3.09 2.90 2.74 2.62 2.52 2.36 2.19 2.00 1.79 1.53 1.36 1.00
* This table is abridged from "Tables of percentage points of the inverted beta distribution," Biometrika, Vol. 33 (1943). It is here published
with the kind permission of its authors, Maxine Merrington and Catherine M. Thompson, and the editor of Biometrika.
556 TABLES APPENDIX 0
Table 5 CUMULATIVE STUDENT'S t DISTRIBUTION*
, r(n11)
Fr.t) =f - <:D
_ (1 + x'),O+'>l
r(n/2)V -rrn _
2 dx
n
~ 1
.75 .90 .95 .975 .99 .995 .9995
1.000 3.078 6.314 12.706 31.821 63.657 636.619

2 .816 1.886 2.920 4.303 6.965 9.925 31.598
3 .765 1.638 2.353 3.182 4.541 5.841 12.941
4 .741 1.533 2.132 1.776 3.747 4.604 8.610
5 .727 1.476 2.015 2.571 3.365 4.032 6.859
6 .718 1.440 1.943 2.447 3.143 3.707 5.959

7 .711 1.415 1.895 2.365 2.998 3.499 5.405
8 .706 1.397 1.860 2.306 2.896 3.355 5.041
9 .703 1.383 1.833 2.262 2.821 3.250 4.781
10 .700 1.372 1.812 2.228 2.764 3.169 4.587
11 .697 1.363 1.796 2.201 2.718 3.106 4.437

12 .695 1.356 1.782 2.179 2.681 3.055 4.318
13 .694 1.350 1.771 2.160 2.650 3.012 4.221
14 .692 1.345 1.761 2.145 2.624 2.977 4.140
15 .691 1.341 1.753 2.131 2.602 2.947 4.073
16 .690 1.337 1.746 2.120 2.583 2.921 4.015

17 .689 1.333 1.740 2.110 2.567 2.898 3.965
18 .688 1.330 1.734 2.101 2.552 .2878 3.922
19 .688 1.328 1.729 2.093 2.539 2.861 3.883
20 .687 1.325 1.725 2.086 2.528 2.845 3.850
21 .686 1.323 1.721 2.080 2.518 2.831 3.819

22 .686 1.321 1.717 2.074 2.508 2.819 3.792
23 .685 1.319 1.714 2.069 2.500 2.807 3.767
24 .685 1.318 1.711 2.064 2.492 2.797 3.745
25 .684 1.316 1.708 2.060 2.485 2.787 3.725
26 .684 1.315 1.706 2.056 2.479 2.779 3.707

27 .684 1.314 1.703 2.052 2.473 2.771 3.690
28 .683 1.313 1.701 2.048 2.467 2.763 3.674
29 .683 1.311 1.699 2.045 2.462 2.756 3.659
30 .683 1.310 1.697 2.042 2.457 2.750 3.646
40 .681 1.303 1.684 2.021 2.423 2.704 3.551

60 .679 1.296 1.671 2:000 2.390 2.660 3.460
120 .677 1.289 1.658 1.980 2.358 2.617 3.373
00 .674 1.282 1.645 1.960 2.326 2.576 3.291
This table is abridged from the" Statistical Tables n of R. A. Fisher and Frank Yates published
III
by Oliver &, Boyd, Ltd., Edinburgh and London, 1938. It is here pubHshed with the kind permission
of the authors and their publishers.
INDEX
Absolutely continuous, 60, 61, 63,64 Complete families of densities, 321, 324, 354
Admissible estimator, 299 Complete statistic, 324
Algebra of sets, 18, 22 Completeness, 321, 354
Analysis of variance, 431 Composite hypothesis, 418
A posteriori probability, 5, 9 (See also Hypotheses)
A priori probability, 2-4, 9 Concentration, 289
Arithmetic series, 528 Conditional distributions, 129, 148
Asymptotic distribution, 196,256-258,261,359, bivariate normal, 161
440,444 continuous, 146, 141
Average sample size in sequential tests, 410--472 discrete, 143-145
Conditional expectation, 151
Conditional mean, 158
BAN estimators, 294, 296, 349, 446 Conditional probability, 32
Bayes estimation, 339, 344 Conditional variance, 159
Bayes' formula, 36 Confidence bands for c,d.f., 511
Bayes risk, 344 Confidence coefficient. 315,311,461
Bayes test, 411 Confidence intervals, 313, 315, 311, 461
Bayesian interval estimates, 396,391 c,d,f.,511
BernouUi distribution, 81, 538 difference in means, 386
Bernoulli trial, 88 general method for, 389
repeated independent, 89,101,131 large sample, 393
Best linear unbiased estimators, 499 mean of normal population, 315,381,384
Beta distribution, 115, 540 median, 512
of second kind, 215 method of finding tests, 425, 461
Beta function, 534, 535 on~sided, 318
Bias, 293 p of binomial population, 393, 395
Binomial coefficient, 529 pivotal method of obtaining, 319, 381
Binomial distribution, 81-89, 119, 120,538 regression coefficients, 491-494
confidence limits for p, 393, 395 uniformly most accurate, 464
normal approximation, 120 variance of normal population, 382, 384
Poisson approximation, 119 Confidence limits [see Confidence interval(s)]
Binomial theorem, 530 Confidencere~on,311
Birthday problem, 45 for mean and variance of normal population,
Bivariate normal distribution, 162-168 384
conditional distribution, 161 Confidence sets, 461
marginal distribution, 161 uniformly most accurate, 464
moment generating function, 164 Consistency of an estimator, 291, 294, 295. 359
moments, 165 Contagious distribution, 102, 122, 123
Boole's inequality, 25 Contingency tables, 452-461
interaction, 454
tests for independence, 452
Cauchy distribution, 111, 201, 238, 540 Continuous distributions, 60, 62
Cauchy· Schwarz inequalities, 162 (See also Distributions)
Censored, 104 Continuous random variable, 60
Central-limit theorem, Ill, 120, 195, 233, 234, Convex function, 12
258 Convolution, 186
Centroid, 65 Correlation, 155, 161
Chebyshev inequality, 11 sample, 526
multivariate, 172 Spearman's rank, 525, 526
Chi-square distribution, 241,542,549 Correlation coefficient, 155, 156
table of, 553 Covariance, 155, 156
Chi-square tests, 440, 442-461 of two Bnear combinations of random
contingency tables, 452-461 variables, 119
goodness-of-fit, 442, 441 Covariance matrix, 352, 489
Combinations and permutations, 528 Cramer-Rao inequality, 316
Combinatorial symbol, 528 Cramer-Rao lower bound, 316, 320
Complement of set, 10 Critical function, 404
SS8 INDEX
Critical region, 403 Distributions:

size of, 407 degenerate, 258
Cumulant generating function, 80 discrete, 58, 60
Cumulants, 80 discrete uniform, 86, 538
Cumulative distribution function(s), 54, 56, 63, double exponential, 117, 540
130,132,144 exponential, 111, 121,237,262,540
bivariate, 132 F,246,247,542,549,554
decomposition of, 63 gamma, 111, 123, 540
empirical,264, 506 geometric, 99, 538
joint, 130 Gumbel, 118,542
sample, 264,287,506,511 hypergeometric, 91, 538
unidimensional, 54, 56 joint, 130, 133, 138
lambda, 128
Laplace, 117
Decision theory, 297,343,415 limiting, 196,258,261,444,446,507-509
Deductive inference, 220 linear function of normal variates, 194
Degenerate distribution, 258 logarithmic, 105
Degrees of freedom, 242, 246 logistic, 118,540
De Moivre-Laplace limit theorem, 120 lognormal, 117, 540
De Morgan's law, 13,14 marginal, 132, 135, 141
Density functions, 62 MaxwelJ, 127
discrete, 57, 135 multinomial, 137
joint discrete, 133 covariance for, 196
joint probability, 138 multivariate, 129-174
probability, 60, 62 negative binomial, 99, 538
(See also Distributions) negative hypergeometric, 213
Difference between means: normal (see Normal distribution)
confidence intervals, 386 order statistics, 251, 254
tests, 432 Pareto, 118, 542
Differentiation, 532 Pearsonian system, 118-119
Discrete distributions, 57, 133, 144, 145 Poisson, 93,104,119-121,123,236,538
(See also Distributions) prior, 340, 417
Discrete random variables, 57, 63, 133 r distribution, 127
Disjoint, 14 Rayleigh, 127
Distance function, 287 rectangular, 105,238,540
Distribution-free tests, 505,509,514,518-524 sample, 224
Distribution functions, 54,56,63, 130, 132 Student's t, 249, 250,542,556
of difference, 185 symmetric, 170
of maximum, 182 table of, 538-543
of minimum, 182 truncated, 122
of order statistics, 251 Tukey's symmetrical lambda, 128
of product, 187 uniform, 105,238,540
of quotient, 187 continuous, 105, 540
of sums, 185 discrete, 86, 538
Distributions: variance ratio, 246,437,438
asymptotic, 196,256-264,359,440,444 Weib"lll, 117,542
BernouJJi, 87, 236,538 (See alsQ Sampling distributions)
beta, 115, 540
beta-binomial, 104
binomial (see Binomial distribution) Efficiency, 291
bivariate normal, 162-168 EJIipsoid of concentration, 353
Cauchy, 117,207,238,540 Empty set, 10
chi-square, 241,542,549 Equivalent sets, 10
conditional (see Conditional distributions) Error:
contagious, 102,122,123 mean-squared,291
continuous, 60, 62 size of, 405
cumulative [see Cumulative distribution Type I, 405
function (s) ] Type II, 405
INDEX 559
Estimation: Function:
interval, 372 decision, 291
point, 211 definition of, 19
Estimator ( s) ,212, 213 density (see Distributions)
admissible, 299 distance, 281
BAN, 294, 296 distribution (see Distributions)
Bayes method, 286, 339, 344 domain of, 19,53
best linear unbiased, 499 gamma, 534
better, 299 generating, 84
closeness, 288 image of, 19
concentrated,289 indicator, 20
consistent, 294, 295, 359 likelihood,218
e1Iipsoid of concentration, 353 loss, 291
least squares, 286,498 squared-error, 291
location invariant, 334 moment generating (see Moment generating
maximum likelihood (see Maximum likeH- function)
hood estimators) power, 406
mean-squared error, 291 preimage, 19
method of moments, 214 probability, 21,22,26
minimax, 299,350 regression, 158, 168
minimum chi-square, 286, 281 risk, 291, 298
minimum distance, 286, 281 set, 20, 21
Pitman, 290, 334, 336 size-of-set, 21
scale invariant, 336
unbiased, 293,315
uniformly minimum-variance unbiased, 315 Game of craps, 48
Wilks' generalized variance, 353 Gamma distribution, 111, 112, 123, 540
(See also Large samples)
Gamma function, 534
Event, 14, 15, 18,53 Gauss-Markov theorem, 500
elementary, 15 Gaussian distribution (see Normal distribution)
Event space, 15, 18,23 Generalized likelihood ratio (see Likelihood
Excess, coefficient of, 16 ratio)
Expectation, 64, 69,153,160,116 Generalized variance, 352, 353
Expected values, 64, 69, 10, 129,153,160 Generating functions (see specific generating
conditional, 151 functions)
of functions of random variables, 116 Geometric distribution, 99, 538
properties of, 10 Geometric series, 528
Exponential c1ass, 312, 313,320,326,355,422 G1ivenko-Cantelli theorem, 501
Exponential distribution, 111,121,231,262,540 Goodness-of-fit test:
Exponential family, 312, 313,320, 326, 355, 422 chi-si\uare, 442, 441
Extension theorem, 22 Kolmogorov-Smirnov, 508, 509
Extreme-value statistic, 118,258 Gumbel distribution, 118, 542
asymptotic distribution of, 261
F distribution, 246, 542, 549 Homogeneity of populations, test of, 505

table of, 554 two exponentials, 416
Factorial moment generating function, 19 two multinomials, 450
Factorial moments, 11, 19 two normals, 432, 435
Factorial notations, 528 two Poissons, 451
Factorial symbol, 528 two trinomials, 419
Factorization criterion, 301 Homogeneity of variances, test of, 438,439
Finite population, sampling from, 261 Hypergeometric distributions, 91, 538
Frequency function, 58 Hypotheses, statistical: alternative, 405
Function, 19 composite, 402
beta, 535 null, 405
convex, 12 simple. 402, 409
counterdomain of, 19, 53 (See also Tests of hypotheses)
560 INDEX
Ideal power function, 406 Likelihood ratio:

Incomplete beta function, 115,513, 516 generalized,419
Incomplete gamma function, 114 large-sample distribution for, 440
Independence, 32, 150,160 monotone, 423, 424
in contingency tables, 452 simple, 410, 423
of events, 40, 41, 46 tests, 409, 419, 440
in probability sense, 32,143, 150, 160,161 Limiting distribution, 196,258,261,444,446,
of random variables, 150 507-509
of sample mean and variance, 243, 245 Linear function of normal variates, distribution
stochastic, 150 of, 194
Index set, 13 Linear models, 482, 485
Inference, 220, 271 confidence intervals, 491
deductive, 220 point estimation, 487, 498
inductive, 220,221 tests of hypotheses, 494
Information, 300, 301 Linear regression, 482
Interquartile range, 75 Location invariance, 332~336
Intersection, 10, 13 Location parameter, 333
Interval estimation, 372 Logistic, 118,540
Bayesian, 396 Lognormal distribution, 117, 540
large sample, 393 Lossfunction,297,343,414,415
(See also Confidence intervals)
Invariance:
location, 331~336
of maximum likelihood estimators, 284, 285, Marginal distributions:
442 for bivariate normal di~tribution, 167
scale, 331, 336-338 continuous, 141
Inverse binomial sampling, 103 discrete, 132, 135
Mass, 58
Mass point, 58
Jacobian, 205 Maximum of random variables, 182
Jensen inequality, 72 Maximum likelihood, principle of, 276
Joint distributions, 129ff, Maximum-likelihood estimators, 279
Joint moments, 159 invariance property, 284,285
large-sample distribution of, 358, 359
of parameters: of normal distribution, 281
Kolmogorov-Smirnov goodness-of-fit test, of uniform distribution, 282
508-510 properties of, 284, 358
Kurtosis, 76 Mean:
definition of, 64
distribution of, 236-238
Lagrange multipliers, 501 sample, 228, 230
Large sample, 294, 358 variance of, 231
confidence limits, 393 Mean absolute deviation, 297
distribution of estimators, 359 Mean-squared error, 291, 297
distribution of generalized likelihood ratio, Mean-squared-error consistency, 294, 295
440 Median, 73,255
distribution of mean, 233 sample, 255
Law of large numbers, 231, 232,258 tests of, 521
Least squares, 498 Mendelian inheritance,445
Lehmann-Scheffe theorem, 326, 356 Method of moments, 274
Leibniz' rule, 532 Midrange, 255
I'Hospital's rule, 532 Minimal sufficient statistic, 311 ,.
Liars problem, 47 Minimax estimator, 299,350
Likelihood function, 278 Minimax test, 416
induced, 285 Minimum of random variables, 182
Likelihood ratio: Minimum chi-square estimation, 286, 287
exact "distribution of, 480 Minimum distance estimation, 286, 287
INDEX 561
Minimum-variance unbiased estimator, Normal distribution, 107-111, 120,239,540,548

uniformly (UMVUE), 315 bivariate, 162, 525
Mixture, 122, 123 conditional, 167
Mode, 74 independence of sample mean and variance,
Model: 243,245
functional, 483 marginal, 167
Jinear, 483,485 moment generating function for, 164
Moment generating function, 72, 78, 80, 159- multivariate, 162
161,164 regression functions for, 168
factorial, 79 role of, 239
of random variables, 159, 160 sample mean, 240
table of, 538-543 sample variance, 241
Moment problem, 81 table of, 552
Moments, 64, 72, 73 truncated, 124
central, 73 Normal equations, 487
cumulant, 80 Null hypothesis (see Hypotheses, statistical)
estimators of, 227
factorial, 77
joint, 159 Order statistics, 251
population, 227 asymptotic distribution of, 256-264
problem of, 81 distribution of functions of, 254
raw, 72, 73
sample, 227
Monotone likelihood ratio, 423, 424 Parameter, 85
Multinomial distribution, 137,253,443 parameter space, 273, 351
covariance of, ] 96 Parametric family, 85
tests on, 448 Pareto distribution, 118, 542
Multinomial theorem, 530, 531 Partition, 300
Multiplication rule, 37 Pearson's chi-square tests, 444, 459,461
Multivariate distributions, 129ft Pitman-closer, 290
marginal and conditional distributions for, Pitman estimator for location, 334
132,135,141,143-147 Pitman estimator for scale, 337
moment generating function for, 160 Pivotal quantity, 379
moments of~ 159 Pivotal quantity method, 379, 387
Mutua]]yexclusive, 3, 14,41 Poisson distribution, 93, 104, 119-121, 123,236,
538
compound, 123
Populations, 222, 224
Negative binomial distribution, 99, 438 sampled, 223
Negative hypergeometric distribution, 213 target, 222
Neyman~Pearson lemma, 411 Posterior Bayes estimator, 341
Nonparametric methods, 504 Posterior distribution, 340
confidence intervals, 512 Posterior risk, 346
equality of distributions, 518-524 Power function of test, 406, 411
interval estimates, 512 ideal, 406
Kolmogorov-Smirnov statistic, 508-510 Prior distribution, 340,417
median, 512 ProbabiJity,2
median test, 521 a posteriori, 5, 9
point estimation, 512 a priori, 2-4, 9
quantiles, 512 axioms of. 8, 22
rank correlation, 525 classical, 2, 3, 5
rank-sum test, 522 conditional, 32, 42
run'test, 519 properties of, 34
sign test: one-sample. 514 definition, 3, 21, 22
two-sample. 519 equa]]y likely, 3, 5, 25
tests (see Tests of hypotheses) frequency, 2, 5, 6
tolerance limits, 515 function, 21, 22
56.2 INDEX
Probability: Regression function (see Regression curve)

independence, 32,40-42 Relative frequency, 3, 6, 8
integral transformation, 202 Reparameterize, 441, 442
laws of, 23-25,34-37 Risk:
mass function, 58 Bayes, 344
models, 8 function,297,298,415
properties of, 23-25,34-37 posterior, 346
space, 25, 53 Runs, 519
subjective, 9
total, 35
Probability density function, 57, 60,62,141 Sample, 222
Probability function, 58 cumulative distribution function, 264
equal1y likely, 26 distribution of, 224
Probability generating function, 64 mean, 227,228, 230
Probability integral transform, 202 median, 255
Problem of moments, 81 midrange, 255
Product of two random vaTiables, 180 moments, 226, 227
Product notation, 527 quantiles, 251
Propagation of errors, 181 random, 223
range, 255
variance, 229, 245
Quantile, 73 Sample c.d,f., 264
point and interval estimates, 512 interferences on, 506
tests of hypotheses, 514 Sample mean, 219, 227,228, 230, 240
Quotient of two random variables, 180 variance of, 231
Sample moments, 219,227
Sample point, 9
Random number, 107,202 Sample space, 9, 14
of random variables, 197 finite, 15, 31, 25ff.
Random sampling, 223 Sample variance, 229, 245
Random variables, 53 Sampled populations, 223
continuous, 60, 138 Sampling, 27,219
definition of, 53 with replacement, 27
discrete, 57, 133 without replacement, 27
joint, 133, 138 Sampling distributions, 219, 224
maximum of, 182 for difference of two means, 386
minimum of, 182 for mean: from binomial population, 236
mixed,62 from Cauchy population, 238
product, 180 from exponential, 237
quotient, 180 of large samples, 232-235
sums, 178 from normal, 241
(See also Distributions) Poisson population, 236
Range, 255 from uniform population, 238
of function, 19 for order statistics, 251ff.
interquartile, 75 for ratio of sample variances, 247
of sample, 255 for regression coefficients, 490
Rank-sum test, 522
Rao-B1ackwell theorem, 321,354
Rectangular distribution, 105,238, 540 Scale invariance, 336-338
References, 544 Scale parameter, 336
Regression, linear, 482 Semiinvariants, 80
Regression coefficient: Sequential probability ratio test, 464, 466-467
confidence interval, 491 approximate, 468
estimators, 491 expected sample size of, 468, 470
tests, 494 Sequential 'tests, 464
Regression curve, 158 for binomial, 481
for normal, 168 fundamental identity for, 470
INDEX 563
Sequential tests: t distribution (see Student's t distribution)

for mean of normal population, 471 Tables of distributions, 538-543
sample size in, 468 chi-square, 553
Set, 9 F, 554, 555
complements, 10 normal, 551, 552
confidence, 461 Student's t, 556
difference, 10 Target populations, 222
disjoint, 14 Taylor series, 533
empty, 10 Tests of hypotheses, 271,401-403
equivalence of, 10 Bayes, 417, 418
function, 21 chi-square, 440
index, 13 composite, 418
intersection, 10, 13 and confidence intervals, 461
laws of, 11 critical function of, 404
mutually exclusive, 14 critical region of, 403
null,1O distribution-free (see Nonparametric
theory,9ff. methods)
union, 10, 13 equality-of-means, 432,435
Sigma algebra, 22 equality of two distributions, 518
Significance level, 407 equality of two multinomials, 448
Significance testing, 407, 409 goodness-of-fit, 442, 447
Simple hypothesis (see Hypotheses, statistical) homogeneity (see Homogeneity of
Singular continuous c. d. f., 63 populations)
Size of critical region, 407 homogeneity of variances, 438, 439
Size of set, 29 independence in contingency tables, 452
Size of test, 407 large-sample, 440
Skewness, 75, 76 likelihood-ratio: generalized, 419
Space, 9 simple, 410,419
even~ 1, 14, 15,18,23,53 mean of normal population, 428-431
parameter, 273 median, 521
probability, 25 minimax, 416
sample, 1, 14,31 most powerful, 410, 411
Spearman's rank correlation, 525, 526 nonrandomized, 403
Standard deviation, 68 null hypothesis, 405
Statistic, 226 power of, 406
Statistical hypotheses (see Hypotheses, randomized, 403, 404
statistical) rank-sum, 522
Statistical inference, 271 ratio of variances, 438
Statistical tests (see Tests of hypotheses) relation to confidence intervals, 461
Stieltjes integral, 69 run, 519
Stirling's formula, 530 sequential (see Sequential tests)
Stochastic independence, 150 of significance, 407
Student's t distribution, 249, 250,542,550 simple, 409
table of, 556 size of, 407
Subset, 10 sufficient statistics, 408
Sufficient statistics, 299, 301, 306, 321,391 unbiased, 425
complete, 321, 324, 326 uniformly most powerful, 421
factorization criterion, 307 on variances, 431, 432, 438
jointly, 306, 307 Tests of significance, 407
minimal, 311, 312, 326 Ticktacktoe problem, 62
tests of hypotheses, 408 Tolerance limits, 505, 515, 516
Summation notation, 527 Total probabilities, theorem of, 35, 148, 149
Sums of random variables: Transformations, 175, 198,202
covariance of, 179 c.d.f. technique, 181
distribution, 192 m.g.f. technique, 189
variance of, 178 probability integral, 202, 203
Symmetrically distributed, 170 Treatment effect, 437, 519
564 INDEX
Truncated distribution, 103, 104, 122, 123 Variance, 67

normal,124 analysis of, 437
Poisson, 104 conditional, 159
Type I and II errors, 405 definition of, 67
size of, 405 distribution of sample, 241
estimator of, 229
of linear combination of random variables,
178,179
UMVUE (uniformly minimum-variance lower bound for, 315
unbiased estimator), 315 sample, 229, 245
Unbiased estimator(s), 293,315,352 of sample mean, 231
best linear, 499 of sum of random variables, 178
joint, 352 tests of, 431, 438
uniformly minimum variance, 315 Variance-covariance matrix, 352, 489
Unbiased test, 425 Variance ratio, 246, 437
Uncorrelated random variables, 161, 173, 174 Vector parameters, 351
Uniform distribution. 105,238,540 Venn diagrams. 11-13
Uniformly most accurate, 464
Uniformly most powerful test, 421
composite, 421 Waiting time, 101, 103
simple, 411 Wald'sequation,470
unbiased, 425 Weibull distribution, 117, 542
Union of sets. 10, 13 Wilks' generalized variance, 353

Math Stats Text

Uploaded by

Copyright:

Available Formats

Math Stats Text

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Math Stats Text

Uploaded by

Copyright:

Available Formats

To HARRIET A.M.M.

Library of Congress Cataloging in Publication Data

This book was set in Times Roman.

Preface to the Third Edition

Excerpts from the First and Second Edition Prefaces xv

1 Introduction and Summary 1

II Random Variables, Distribution Functions, and

ill Special Parametric Families of Univariate Distributions 85

(3.2 ) Normal Distribution 107

IV Joint and Conditional Distributions, Stochastic

V Distributions of Functions of Random Variables 175

4 Moment-generating-function Technique 189

VI Sampling and Sampling Distributions 219

1 Introduction and Summary 219

VII Parametric Point Estimation 271

1 Introduction and Summary 271

vm Parametric Interval Estimation 372

1 Introduction and Summary 372

3 Sampling from the Normal Distribution 381

IX Tests of Hypotheses 401

1 Introduction and Summary 401

7.2 Definition of Sequential Probability Ratio Test 466

X Linear Models 482

1 Introduction and Summary 482

XI N onparametric Methods 504

1 Introduction and Summary 504

Appendix A. Mathematical Addendum 527

Appendix B. Tabular Summary of Parametric Families

Appendix C. References and Related Reading 544

Appendix D. Tables 548

1 INTRODUCTION AND SUMMARY

defined on finite sample spaces. The related concepts of independence of u

2.2 Classical or A Priori Probability

Definition 1 Classical probability If a random experiment can resul t

be in games of chance. Thus we shall have to alter or extend our definition to

2.3 A Posteriori or Frequency Probability

Table 1 RESULTS OF TOSSING A COIN 100 TIMES

Total 100 1.00 1.00

Suppose, as described above, that the coin is unbalanced so that we are

Table 2 RESULTS OF TOSSING A DIE 300 TIMES

3.1 Probability Models

The two general types of probability (a priori and a posteriori) defined

3.2 An Aside-Set Theory

EXAMPLE 1 Q = R 2 , where R2 is the collection of points co in the plane and

EXAMPLE 2 n = {all United States citizens}. IIII

Definition 2 Subset If every element of a set A is also an element of a

Q but not in A. 1//1

Definition 8 Set difference Let A and B be any two subsets of n. The

Ai = {(x, y): 0 < x < 1; 0 < Y < t},

EXAMPLE 4 Let n, A 1, A 2 , and A3 be as indicated in the diagrams in Fig. I

The set operations of complement, union, and intersection have been

Theorem 1 Commutative laws A u B = B u A and A (\ B = B (\ A.

Theorem 3 Distributive laws A (\ (B u C) = (A (\ B) u (A (\ C), and

Theorem 4 (Ac)C = (1) = A; in words, the complement of A comple-

n is the entire square

Theorem 7 (A u B) = A n 13, and (A n B) = A u 13. These are known

Theorem 8 A - B = AB. Jill

We defined union and intersection of two sets; these definitions extend

Definition 9 Union and intersection of sets Let A be an index set and

UAA, = ljJ and