Probability and Statistics with Examples using R

Siva Athreya, Deepayan Sarkar, and Steve Tanner

November 19, 2024

Version: – November 19, 2024

CONTENTS

Preface v
1 Basic Concepts 1
1.1 Definitions and Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Basic Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Equally Likely Outcomes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3 Conditional Probability and Bayes’ Theorem . . . . . . . . . . . . . . . . . 16
1.3.1 Bayes’ Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.4 Independence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.5 Using R for Computation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2 Sampling and Repeated Trials 37
2.1 Bernoulli Trials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.1.1 Using R to Compute Probabilities . . . . . . . . . . . . . . . . . . . 43
2.2 Poisson Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.3 Sampling With and Without Replacement . . . . . . . . . . . . . . . . . . . 54
2.3.1 The Hypergeometric Distribution . . . . . . . . . . . . . . . . . . . . 55
2.3.2 Hypergeometric Distribution as a Series of Dependent Trials . . . . 56
2.3.3 Binomial Approximation to the Hypergeometric Distribution . . . . 58
3 Discrete Random Variables 63
3.1 Random Variables as Functions . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.1.1 Common Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.2 Independent and Dependent Variables . . . . . . . . . . . . . . . . . . . . . 70
3.2.1 Independent Variables . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.2.2 Conditional, Joint, and Marginal Distributions . . . . . . . . . . . . 72
3.2.3 Memoryless Property of the Geometric Random Variable . . . . . . 76
3.2.4 Multinomial Distributions . . . . . . . . . . . . . . . . . . . . . . . . 77
3.3 Functions of Random Variables . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.3.1 Distribution of f (X ) and f (X1 , X2 , . . . , Xn ) . . . . . . . . . . . . . . 82
3.3.2 Functions and Independence . . . . . . . . . . . . . . . . . . . . . . 87
4 Summarizing Discrete Random Variables 93
4.1 Expected Value . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.1.1 Properties of the Expected Value . . . . . . . . . . . . . . . . . . . . 96

Version: – November 19, 2024

ii CONTENTS

4.1.2 Expected Value of a Product . . . . . . . . . . . . . . . . . . . . . . 99

4.1.3 Expected Values of Common Distributions . . . . . . . . . . . . . . 100
4.1.4 Expected Value of f (X1 , X2 , . . . , Xn ) . . . . . . . . . . . . . . . . . . 105
4.2 Variance and Standard Deviation . . . . . . . . . . . . . . . . . . . . . . . . 110
4.2.1 Properties of Variance and Standard Deviation . . . . . . . . . . . . 112
4.2.2 Variances of Common Distributions . . . . . . . . . . . . . . . . . . 114
4.2.3 Standardized Variables . . . . . . . . . . . . . . . . . . . . . . . . . . 118
4.3 Standard Units . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
4.3.1 Markov and Chebychev Inequalities . . . . . . . . . . . . . . . . . . 123
4.4 Conditional Expectation and Conditional Variance . . . . . . . . . . . . . . 127
4.5 Covariance and Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
4.5.1 Covariance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
4.5.2 Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
4.6 Exchangeable Random Variables . . . . . . . . . . . . . . . . . . . . . . . . 141
5 Continuous Probabilities and Random Variables 147
5.1 Uncountable Sample Spaces and Densities . . . . . . . . . . . . . . . . . . . 147
5.1.1 Probability Densities on R . . . . . . . . . . . . . . . . . . . . . . . 149
5.2 Continuous Random Variables . . . . . . . . . . . . . . . . . . . . . . . . . . 155
5.2.1 Common Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . 157
5.2.2 A Word About Individual Outcomes . . . . . . . . . . . . . . . . . . 166
5.3 Transformations of Continuous Random Variables . . . . . . . . . . . . . . 172
5.4 Multiple Continuous Random Variables . . . . . . . . . . . . . . . . . . . . 180
5.4.1 Marginal Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . 184
5.4.2 Independence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
5.4.3 Conditional Density . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
5.5 Functions of Independent Random Variables . . . . . . . . . . . . . . . . . . 195
5.5.1 Distributions of Sums of Independent Random variables . . . . . . . 196
5.5.2 Distributions of Quotients of Independent Random Variables . . . . 202
6 Summarising Continuous Random Variables 211
6.1 Expectation and Variance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
6.2 Covariance, Correlation, Conditional Expectation and Conditional Variance 221
6.3 Moment Generating Functions . . . . . . . . . . . . . . . . . . . . . . . . . 230
6.4 Bivariate Normals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
7 Sampling and Descriptive Statistics 247
7.1 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
7.1.1 Sample Mean and Sample Variance . . . . . . . . . . . . . . . . . . . 249
7.1.2 Sample proportion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251

Version: – November 19, 2024

 CONTENTS iii

7.2 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255

7.3 Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
7.3.1 Empirical Distribution Plot for Discrete Distributions . . . . . . . . 262
7.3.2 Histograms for Continuous Distributions . . . . . . . . . . . . . . . . 263
7.3.3 Hanging Rootograms for Comparing with Theoretical Distributions 265
7.3.4 Q-Q Plots for Continuous Distributions . . . . . . . . . . . . . . . . 267
8 Sampling Distributions and Limit Theorems 271
8.1 Multi-dimensional Continuous Random Variables . . . . . . . . . . . . . . . 271
8.1.1 Order Statistics and their Distributions . . . . . . . . . . . . . . . . 274
8.1.2 χ2 , F and t . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278
8.1.3 Distribution of Sampling Statistics from a Normal Population . . . . 281
8.2 Weak Law of Large Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . 286
8.3 Convergence in Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
8.4 Central Limit Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
8.4.1 Normal Approximation . . . . . . . . . . . . . . . . . . . . . . . . . 306
8.4.2 Continuity Correction . . . . . . . . . . . . . . . . . . . . . . . . . . 307
8.5 Delta Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
8.5.1 Variance Stabilizing Transformation . . . . . . . . . . . . . . . . . . 313
8.6 Limiting Distribution of Sample Median . . . . . . . . . . . . . . . . . . . . 316
9 Estimation 321
9.1 Method of Moments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
9.2 Maximum Likelihood . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
9.3 Confidence Intervals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 332
9.3.1 Pivotal Quantity approach . . . . . . . . . . . . . . . . . . . . . . . 334
9.3.2 Empirical Coverage Probability of Confidence Intervals . . . . . . . 337
9.3.3 Approximate Confidence Intervals using CLT . . . . . . . . . . . . . 346
9.3.4 Confidence Intervals for the Population Median . . . . . . . . . . . . 349
10 Hypothesis Testing 355
10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355
10.2 The Goodness of Fit Problem in the Multinomial Model . . . . . . . . . . . 358
10.3 Independence of Two Categorical Attributes . . . . . . . . . . . . . . . . . . 359
10.3.1 A Multinomial Model for Two-way Tables . . . . . . . . . . . . . . . 360
10.3.2 Independence in the Multinomial Model . . . . . . . . . . . . . . . . 362
10.4 Testing in the Parametric Setup : The Intuitive Approach . . . . . . . . . . 362
10.4.1 Finding a Test Statistic . . . . . . . . . . . . . . . . . . . . . . . . . 363
10.4.2 The Normal Distribution: Test for Sample Mean when σ is Known 363
10.4.3 The Normal Distribution: Test for Sample Mean when σ is Unknown 366

Version: – November 19, 2024

iv CONTENTS

10.4.4 An Alternative Test Based on the Median . . . . . . . . . . . . . . . 369

10.5 The General Approach: Likelihood Ratio Test . . . . . . . . . . . . . . . . . 371
10.5.1 The Likelihood Ratio Statistic . . . . . . . . . . . . . . . . . . . . . 371
10.5.2 Type I and Type II Error . . . . . . . . . . . . . . . . . . . . . . . . 372
10.5.3 The p-value for the Likelihood Ratio Test . . . . . . . . . . . . . . . 373
10.6 Specific Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
10.6.1 Binomial Test for Proportion . . . . . . . . . . . . . . . . . . . . . . 375
10.6.2 Normal Test for Mean When Variance is Known . . . . . . . . . . . 377
10.6.3 One-sided Test for Normal Mean when Variance is Known . . . . . . 379
10.6.4 Normal Test for Mean When Variance is Unknown . . . . . . . . . . 380
10.6.5 The Two-sample Test for Equality of Population Means . . . . . . . 384
10.6.6 Equality of Population Means with Different Variances . . . . . . . . 388
10.7 Testing for Goodness of Fit . . . . . . . . . . . . . . . . . . . . . . . . . . . 389
10.7.1 Asymptotic Distribution of the Likelihood Ratio Test Statistic . . . 391
10.7.2 The Standard χ2 Test for Goodness of Fit . . . . . . . . . . . . . . . 392
10.8 Testing for Independence of Categorical Attributes . . . . . . . . . . . . . . 394
10.8.1 The Multinomial Model . . . . . . . . . . . . . . . . . . . . . . . . . 394
10.8.2 Binomial Model with Fixed Row Margins . . . . . . . . . . . . . . . 395
10.8.3 Fisher’s Exact Test of Independence . . . . . . . . . . . . . . . . . . 396
a Some mathematical details 399
a.1 Transformation of Continuous random Variables- Jacobian . . . . . . . . . . 399
a.1.1 Multiple Continuous Random Variables and the Jacobian . . . . . . 401
a.2 Strong Law of Large Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . 405
b Tables 411

Version: – November 19, 2024

P R E FA C E

We believe that many foundational ideas of Probability and Statistics are best understood
when their natural connection is emphasised. We feel that the interested student should
learn the mathematical rigour of Probability, the motivating examples and techniques from
Statistics, and an instructive technology to perform computations relating to both in an
inclusive manner. These formed our main motivations for writing this book. We have
chosen to use the R software environment to demonstrate an available computational tool.
The book is intended to be an undergraduate text for a course on Probability Theory.
We had in mind courses such as the one year (two semester) Probability course at many
universities in India such as the Indian Statistical Institute or Chennai Mathematical
Institue, or a one semester (or two quarter) Probability course as is commonly offered
as an upper division, post-calculus elective at many North American universities. The
Statistics material and the package R are introduced so as to emphasise motivations and
applications of the probabilistic material. We assume that our readers are well-versed in
calculus, have a basic understanding of the theory of sets and functions, combinatorics,
and proof techniques, and have at least a passing awareness of the distinction between
countable and uncountable infinities. We do not assume any particular experience of Linear
Algebra or Real Analysis.
In Chapter 1 of this book we begin with an introduction to Outcomes, Sample Space,
Events and the axiomatic definition of Probability. Then we discuss the concepts of
conditional probability, independence and Bayes’ Theorem. We conclude this chapter with
a basic introduction to R. R is a Free Open Source software environment that runs on
all major software platforms, and instructions to download and install it are available at
[Link]
We begin Chapter 2 by applying the notion of independence to repeated trials (Bernoulli
Trials) and discuss the Binomial and Geometric distributions. We introduce the Poisson
distribution as a limiting approximation of the Binomial. We conclude this section with a
discussion on Sampling with and without replacement. The Hypergeometric Distribution
is thus introduced here and we prove its approximation to the Binomial. Throughout
this chapter and later in the book we provide the R code for calculating the probabilities
associated with common distributions.
In Chapters 3 and 4 we introduce discrete random variables (functions on a sample space
whose range is countable) and related concepts. In Chapter 3, we define the probability

Version: – November 19, 2024

vi CONTENTS

mass function, distribution function, and independence for random variables. We introduce
the Multinomial distribution and show the memoryless property of the Geometric random
variable. The chapter concludes by providing a method to compute the distributions of
functions of one and several random variables, defining the concept of joint distribution
along the way. In Chapter 4, we define Expectation, Variance, Covariance, Conditional
Expectation and Conditional Variance for discrete random variables. Results involving
these quantities for standard distributions are presented (and proved) as well. We also state
and prove the Markov and Chebyshev inequalities along with the notion of standardising
random variables to mean zero and variance one.
Working with uncountable spaces and understanding the probability density function of
an absolutely continuous random variable are challenging without assuming a background
in Real Analysis but we make a modest attempt towards this in Chapter 5. We begin with
a description of uncountable sample spaces. After having described events in a temporary
manner in Chapter 1 we provide a precise definition here but comfort the reader that we
shall avoid the most general events and at most consider countable union/intersection
of intervals. This allows us to be fairly rigorous with random variables having piecewise
continuous probability density functions using results from basic calculus. After this we
imitate the program conducted in Chapter 3. Standard distributions such as Uniform,
Exponential, and Normal are discussed. While computing densities of sums and ratios of
independent random variables we introduce the Gamma distribution and use it to derive
the Beta distribution as an example of ratio of dependent Gamma random variables.
In Chapter 6 we define Variance, Covariance, Conditional Expectation and Conditional
Variance for continuous random variables and summarise their properties. Moment gener-
ating functions for random variables are defined. At this point, to respect the minimal
background assumption on our reader we state a few important results without proof such
as the fact that the moment generating functions characterise distribution of a random
variable. The chapter ends with a section on Bivariate Normal random variables. Here we
have done all computations in this section without using Linear Algebra but the notational
efficiency of using Linear Algebra is explained via exercises for the interested reader.
With the foundational ideas of Probability laid out we proceed in Chapter 7 with
Sampling and Descriptive Statistics. The empirical distribution, the sample mean, variance
and proportion are defined along with their properties. Simulation is used to develop
intuition regarding sampling variability, and plots such as Histograms, Hanging Rootograms,
and Q-Q Plots are introduced and illustrated using R.
Limit Theorems for Sampling Distributions discussed in Chapter 7 are the objective
of Chapter 8. We begin with a brief description of multivariate joint densities and Order
statistics. The t-distribution and χ2 (chi-square) distributions are introduced in this chapter.

Version: – November 19, 2024

 CONTENTS vii

The sample mean and variance from a normal population are discussed in relation to t and
χ2 . We prove the Weak Law of Large numbers and the Central Limit Theorem for random
variables possessing a moment generating function. We do state a more general version of
the Central Limit Theorem and also the Strong Law of Large numbers, providing a proof
of the latter in the Appendix. Along with R code we discuss the continuity correction
and applications of the Central Limit Theorem via examples. We then discuss the delta
method and its application to variance stabilising transformations. The chapter concludes
with a derivation of the Central Limit Theorem for the median.
We end the book with two chapters focused solely on results and techniques from
statistics. In Chapter 9 we discuss Estimation and Confidence Intervals. We briefly
describe Method of Moments Estimators and Maximum Likelihood Estimators. We then
introduce Confidence Intervals using the Pivotal Quantity approach. For cases when
natural pivotal quantities do not exist, we illustrate the use of the Central Limit Theorem
to obtain approximate confidence intervals. Finally we derive confidence intervals based on
the sample median and compare its performance with intervals based on the sample mean
via simulations.
In Chapter 10 we explore a non-traditional approach to Hypothesis Testing based on
p-values rather than pre-determined significance levels. We first formulate the multinomial
goodness of fit and independence problems in the familiar parametric set up. After that
we describe an intuitive approach to derive suitable test statistics for various hypothesis
testing examples. We then proceed to outline a likelihood ratio based approach to derive
test statistics systematically. We conclude this chapter with a discussion of the goodness
of fit and independence tests.
R code for most of the computations done are given in the book itself, and the reader
should be able to reproduce and extend them easily. Code for figures are not given in the
book, but are available at a website accompanying the book.
The Appendix includes some relevant mathematical details not covered in the main
matter of the book. The topics included are the Jacobian method for computing distribution
of transformations of random variables and the Strong Law of Large Numbers.

Siva Athreya, Deepayan Sarkar, and Steve Tanner

November 19, 2024

Version: – November 19, 2024

BASIC CONCEPTS
1
1.1 definitions and properties

Most of the problems in probability and statistics involve determining how likely it is that
certain things will occur. Before we can talk about what is likely or unlikely, we need to
know what is possible. In other words, we need some framework in which to discuss what
sorts of things have the potential to occur. To that end, we begin by introducing the basic
concepts of “sample space”, “experiment”, “outcome”, and “event”. We also define what
we mean by a “probability” and provide some examples to demonstrate the consequences
of the definition.

1.1.1 Definitions

Definition 1.1.1. (Sample Space) A sample space S is a set. The elements of the
set S will be called “outcomes” and should be viewed as a listing of all possibilities
that might occur. We will call the process of actually selecting one of these outcomes
an “experiment”.

For its familiarity and simplicity we will frequently use the example of rolling a die. In
that case our sample space would be S = {1, 2, 3, 4, 5, 6}, a complete listing of all of the
outcomes on the die. Performing an experiment in this case would mean rolling the die and
recording the number that it shows. However, sample space outcomes need not be numeric.
If we are flipping a coin (another simple and familiar example) experiments would result in
one of two outcomes and the appropriate sample space would be S = {Heads, T ails}.
For a more interesting example, if we are discussing which country will win the next
World Cup, outcomes might include Brazil, Spain, Canada, and Thailand. Here the set
S might be all the world’s countries. An experiment in this case requires waiting for the
next World Cup and identifying the country which wins the championship game. Though
we have not yet explained how probability relates to a sample space, soccer fans amongst
our readers may regard this example as a foreshadowing that not all outcomes of a sample
space will necessarily have the same associated probabilities.

Version: – November 19, 2024

2 basic concepts

Definition 1.1.2. (Temporary Definition of Event) Given a sample space S,

an “event” is any subset E ⊂ S.

This definition will allow us to talk about how likely it is that a range of possible outcomes
might occur. Continuing our examples above we might want to talk about the probability
that a die rolls a number larger than two. This would involve the event {3, 4, 5, 6} as a
subset of {1, 2, 3, 4, 5, 6}. In the soccer example we might ask whether the World Cup will
be won by a South American country. This subset of our list of all the world’s nations
would contain Brazil as an element, but not Spain.
It is worth noting that the definition of “event” includes both S, the sample space itself,
and ∅, the empty set, as legitimate examples. As we introduce more complicated examples
we will see that it is not always necessary (or even possible) to regard every single subset
of a sample space as a legitimate event, but since the reasons for that may be distracting
at this point we will use the above as a temporary definition of “event” and refine the
definition when it becomes necessary.
To each event, we want to assign a chance (or “probability”) which will be a number
between 0 and 1. So if the probability of an event E is 0.72, we interpret that as saying,
“When an experiment is performed, it has a 72% chance of resulting in an outcome contained
in the event E”. Probabilities will satisfy two axioms stated and explained below. This
formal definition is due to Andrey Kolmogorov (1903-1987).

Definition 1.1.3. (Probability Space Axioms) Let S be a sample space and let
F be the collection of all events.
A “probability” is a function P : F → [0, 1] such that

(1) P (S ) = 1; and

(2) If E1 , E2 , ... are a countable collection of disjoint events

(that is, Ei ∩ Ej = ∅ if i ̸= j), then

∞ ∞
P (Ej ). (1.1.1)
[ X
P( Ej ) =
j =1 j =1

The first axiom is relatively straight forward. It simply reiterates that S did, indeed,
include all possibilities, and therefore there is a 100% chance that an experiment will result
in some outcome included in S. The second axiom is not as complicated as it looks. It
simply says that probabilities add when combining a countable number of disjoint events.
It is implicit that the series on right hand side of the equation (1.1.1) converges. Further

Version: – November 19, 2024

 1.1 definitions and properties 3

(1.1.1) also holds when combining finite number of disjoint events (see Theorem 1.1.4
below).
Returning to our World Cup example, suppose A is a list of all North American
countries and E is a list of all European countries. If it happens that P (A) = 0.05 and
P (E ) = 0.57 then P (A ∪ E ) = 0.62. In other words, if there is a 5% chance the next
World Cup will be won by a North American nation and a 57% chance that it will be
won by a European nation, then there is a 62% chance that it will be won by a nation
from either Europe or North America. The disjointness of these events is obvious as (if we
discount island territories) there is no country that is in both North America and Europe.
The requirement of axiom two that the collection of events be countable is important.
We shall see shortly that, as a consequence of axiom two, disjoint additivity also applies
to any finite collection of events. It does not apply to uncountably infinite collections
of events, though that fact will not be relevant until later in the text when we discuss
continuous probability spaces.

1.1.2 Basic Properties

There are some immediate consequences of these probability axioms which we will state
and prove before returning to some simple examples.

Theorem 1.1.4. Let P be a probability on a sample space S. Then,

(1) P (∅) = 0;

(2) If E1 , E2 , ...En are a finite collection of disjoint events, then

 
n n
P (Ej );
[ X
P Ej  =
j =1 j =1

(3) If E and F are events with E ⊂ F , then P (E ) ≤ P (F );

(4) If E and F are events with E ⊂ F , then P (F \ E ) = P (F ) − P (E );

(5) Let E c be the complement of event E. Then P (E c ) = 1 − P (E ); and

(6) If E and F are events then P (E ∪ F ) = P (E ) + P (F ) − P (E ∩ F ).

Proof. (1) - The empty set is disjoint from itself, so ∅,!∅, . . . is a countable disjoint
∞ ∞
collection of events. From the second axiom, P P (Ej ). When this is
S P
Ej =
j =1 j =1

Version: – November 19, 2024

4 basic concepts

∞
applied to the collection of empty sets we have P (∅) = P (∅). If P (∅) had any
P
j =1
non-zero value, the right hand side of this equation would be a divergent series while the
left hand side would be a number. Therefore, P (∅) = 0.

Proof of (2) - To use axiom two we need to make this a countable collection of events.
We may do so while preserving disjointness by including copies of the empty set. Define
Ej = ∅ for j > n. Then E!1 , E2 , . . . , En , ∅, ∅, . . . is a countable collection of disjoint
∞ ∞
sets and therefore P P (Ej ). However, the empty sets add nothing to the
S P
Ej =
j =1 j =1
∞ n
union and so Ej . Likewise since we have shown P (∅) = 0 these sets also add
S S
Ej =
j =1 j =1
∞ n
nothing to the sum, so P (Ej ).
P P
P (Ej ) =
j =1 j =1

Combining these gives the result:

   
n ∞ ∞ n
P (Ej ).
[ [ X X
P Ej  = P  Ej  = P (Ej ) =
j =1 j =1 j =1 j =1

Proof of (3) - If E ⊂ F , then E and F \ E are disjoint events with a union equal to F .
Using (2) above gives P (F ) = P (E ∪ (F \ E )) = P (E ) + P (F \ E ).
Since probabilities are assumed to be positive, it follows that P (F ) ≥ P (E ).

Proof of (4) - As with the proof of (3) above, E and F \ E are disjoint events with
E ∪ (F \ E ) = F . Therefore P (F ) = P (E ) + P (F \ E ) from which we get the result.

Proof of (5) - This is simple a special case of (4) where F = S.

Proof of (6) - We may disassemble E ∪ F disjointly as E ∪ F = E ∪ (F \ E ). Then from

(2) we have P (E ∪ F ) = P (E ) + P (F \ E ).
Next, since F \ E ⊂ F and since F \ (F \ E ) = E ∩ F we can use (4) to write P (E ∪ F ) =
P (E ) + P (F ) − P (E ∩ F ). ■

Example 1.1.5. A coin flip can come up either “heads” or “tails”, so S = {heads, tails}.
A coin is considered “fair” if each of these outcomes is equally likely. Which axioms or
properties above can be used to reach the (obvious) conclusion that both outcomes have a
50% chance of occurring?
Each outcome can also be regarded as an event. So E = {heads} and F = {tails}
are two disjoint events. If the coin is fair, each of these events is equally likely, so
P (E ) = P (F ) = p for some value of p. However, using the second axiom, 1 = P (S ) =
P (E ∪ F ) = P (E ) + P (F ) = 2p. Therefore, p = 0.5, or in other words each of the two
possibilities has a 50% chance of occurring on any flip of a fair coin. ■

Version: – November 19, 2024

 1.1 definitions and properties 5

In the examples above we have explicitly described the sample space S, but in many cases
this is neither necessary nor desirable. We may still use the probability space axioms and
their consequences when we know the probabilities of certain events even if the sample
space is not explicitly described.

Example 1.1.6. A certain sea-side town has a small fishing industry. The quantity of fish
caught by the town in a given year is variable, but we know there is a 35% chance that the
town’s fleet will catch over 400 tons of fish, but only a 10% chance that they will catch
over 500 tons of fish. How likely is it they will catch between 400 and 500 tons of fish?
The answer to this may be obvious without resorting to sets, but we use it as a first
example to illustrate the proper use of events. Note, though, that we will not explicitly
describe the sample space S.
There are two relevant events described in the problem above. We have F representing
“the town’s fleet will catch over 400 tons of fish” and E representing “the town’s fleet will
catch over 500 tons of fish”. We are given that P (E ) = 0.1 while P (F ) = 0.35.
Of course E ⊂ F since if over 500 tons of fish are caught, the actual tonnage will be
over 400 as well. The event that the town’s fleet will catch between 400 and 500 tons of
fish is F \ E since E hasn’t occurred, but F has. So using property (4) from above we
have P (F \ E ) = P (F ) − P (E ) = 0.35 − 0.1 = 0.25. In other words there is a 25% chance
that between 400 and 500 tons of fish will be caught. ■

Example 1.1.7. Suppose we know there is a 60% chance that it will rain tomorrow and a
70% chance the high temperature will be above 30◦ C. Suppose we also know that there is
a 40% chance that the high temperature will be above 30◦ C and it will rain. How likely is
it tomorrow will be a dry day that does not go above 30◦ C?
The answer to this question may not be so obvious, but our first step is still to view
the pieces of information in terms of events and probabilities. We have one event E which
represents “It will rain tomorrow” and another F which represents “The high will be
above 30◦ C tomorrow”. Our given probabilities tell us P (E ) = 0.6, P (F ) = 0.7, and
P (E ∩ F ) = 0.4. We are trying to determine P (E c ∩ F c ). We can do so using properties
(5) and (6) proven above, together with the set-algebraic fact that E c ∩ F c = (E ∪ F )c .
From (5) we know P (E ∪ F ) = P (E ) + P (F ) − P (E ∩ F ) = 0.7 + 0.6 − 0.4 = 0.9.
(This is the probability that it either will rain or be above 30 0 C).
Then from (6) and the set-algebraic fact, P (E c ∩ F c ) = P ((E ∪ F )c ) = 1 − P (E ∪ F ) =
1 − 0.9 = 0.1. So there is a 10% chance tomorrow will be a dry day that does not reach
30◦ C. ■

Version: – November 19, 2024

6 basic concepts

Temperature
Rain above 30℃

0.6 0.4 0.7

Figure 1.1: A Venn diagram that describes the probabilities from Example 1.1.7.

exercises

Ex. 1.1.1. Consider the sample space Ω = {a, b, c, d, e}. Given that {a, b, e}, and {b, c}
are both events, what other subsets of Ω must be events due to the requirement that the
collection of events is closed under taking unions, intersections, and compliments?
Ex. 1.1.2. There are two positions - Cashier and Waiter - open at the local restaurant.
There are two male applicants (David and Rajesh) two female applicants (Veronica and
Megha). The Cashier position is chosen by selecting one of the four applicants at random.
The Waiter position is then chosen by selecting at random one of the three remaining
applicants.

(a) List the elements of the sample space S.

(b) List the elements of the event A that the position of Cashier is filled by a female
applicant.

(c) List the elements of the event B that exactly one of the two positions is filled by a
female applicant.

(d) List the elements of the event C that neither position was filled by a female applicant.

(e) Sketch a Venn diagram to show the relationship among the events A, B, C and S.

Ex. 1.1.3. A jar contains a large collection of red, green, and white marbles. Marbles are
drawn from the jar one at a time. The color of the marble is recorded and it is put back in

Version: – November 19, 2024

 1.1 definitions and properties 7

the jar before the next draw. Let Rn denote the event that the n-th draw is a red marble
and let Gn denote the event that the n-th draw is a green marble. For example, R1 ∩ G2
would denote the event that the first marble was red and the second was green. In terms of
these events (and appropriate set-theoretic symbols – union, intersection, and complement)
find expressions for the events in parts (a), (b), and (c) below.

(a) The first marble drawn is white. (We might call this W1 , but show that it can be
written in terms of the Rn and Gn sets described above).

(b) The first marble drawn is green and the second marble drawn is not white.

(c) The first and second draws are different colors.

(d) Let E = R1 ∪ G2 and let F = R1c ∩ R2 . Are E and F disjoint? Why or why not?

Ex. 1.1.4. Suppose there are only thirteen teams with a non-zero chance of winning the
next World Cup. Suppose those teams are Spain (with a 14% chance), the Netherlands
(with a 11% chance), Germany (with a 11% chance), Italy (with a 10% chance), Brazil
(with a 10% chance), England (with a 9% chance), Argentina (with a 9% chance), Russia
(with a 7% chance), France (with an 6% chance), Turkey (with a 4% chance), Paraguay
(with a 4% chance), Croatia (with a 4% chance) and Portugal (with a 1% chance).

(a) What is the probability that the next World Cup will be won by a South American
country?

(b) What is the probability that the next World Cup will be won by a country that is
not from South America? (Think of two ways to do this problem - one directly and
one using part (5) of Theorem 1.1.4. Which do you prefer and why?)

Ex. 1.1.5. If A and B are disjoint events and P (A) = 0.3 and P (B ) = 0.6, find P (A ∪ B ),
P (Ac ) and P (Ac ∩ B ).
Ex. 1.1.6. Suppose E and F are events in a sample space S. Suppose that P (E ) = 0.7
and P (F ) = 0.5.

(a) What is the largest possible value of P (E ∩ F )? Explain.

(b) What is the smallest possible value of P (E ∩ F )? Explain.

Ex. 1.1.7. A biologist is modeling the size of a frog population in a series of ponds. She is
concerned with both the number of egg masses laid by the frogs during breeding season
and the annual precipitation into the ponds. She knows that in a given year there is an
86% chance that there will be over 150 egg masses deposited by the frogs (event E) and
that there is a 64% chance that the annual precipitation will be over 17 inches (event F ).

Version: – November 19, 2024

8 basic concepts

(a) In terms of E and F , what is the event “there will be over 150 egg masses and an
annual precipitation of over 17 inches”?

(b) In terms of E and F , what is the event “there will be 150 or fewer egg masses and
the annual precipitation will be over 17 inches”?

(c) Suppose the probability of the event from (a) is 59%. What is the probability of the
event from (b)?

Ex. 1.1.8. In part (6) of Theorem 1.1.4 we showed that

P (E ∪ F ) = P (E ) + P (F ) − P (E ∩ F ).

Versions of this rule for three or more sets are explored below.

(a) Prove that P (A ∪ B ∪ C ) is equal to

P (A) + P (B ) + P (C ) − P (A ∩ B ) − P (A ∩ C ) − P (B ∩ C ) + P (A ∩ B ∩ C )

for any events A, B, and C.

(b) Use part (a) to answer the following question. Suppose that in a certain United
States city 49.3% of the population is male, 11.6% of the population is sixty-five
years of age or older, and 13.8% of the population is Hispanic. Further, suppose 5.1%
is both male and at least sixty-five, 1.8% is both male and Hispanic, and 5.9% is
Hispanic and at least sixty-five. Finally, suppose that 0.7% of the population consists
of Hispanic men that are at least sixty-five years old. What percentage of people in
this city consists of non-Hispanic women younger than sixty-five years old?

(c) Find a four-set version of the equation. That is, write P (A ∪ B ∪ C ∪ D ) in terms of
probabilities of intersections of A, B, C, and D.

(d) Find an n-set version of the equation.

Ex. 1.1.9. A and B are two events. P(A)=0.4, P(B)=0.3, P(A∪B)=0.6. Find the following
probabilities:

(a) P (A ∩ B );

(b) P(Only A happens); and

(c) P(Exactly one of A or B happens).

Version: – November 19, 2024

 1.1 definitions and properties 9

Ex. 1.1.10. In the next subsection we begin to look at probability spaces where each of the
outcomes are equally likely. This problem will help develop some early intuition for such
problems.

(a) Suppose we roll a die and so S = {1, 2, 3, 4, 5, 6}. Each outcome separately
{1}, {2}, {3}, {4}, {5}, {6} is an event. Suppose each of these events is equally likely.
What must the probability of each event be? What axioms or properties are you
using to come to your conclusion?

(b) With the same assumptions as in part (a), how would you determine the probability
of an event like E = {1, 3, 4, 6}? What axioms or properties are you using to come
to your conclusion?

(c) If S = {1, 2, 3, ..., n} and each single-outcome event is equally likely, what would be
the probability of each of these events?

(d) Suppose E ⊂ S is an event in the sample space from part (c). Explain how you could
determine P (E ).

Ex. 1.1.11. Suppose A and B are subsets of a sample space Ω.

(a) Show that (A − B ) ∪ B = A when B ⊂ A.

(b) Show by example that the equality doesn’t always hold if B is not a subset of A.

Ex. 1.1.12. Let A and B be events.

(a) Suppose P (A) = P (B ) = 0. Prove that P (A ∪ B ) = 0.

(b) Suppose P (A) = P (B ) = 1. Prove that P (A ∩ B ) = 1.

Ex. 1.1.13. Let An be a sequence of events.

(a) Suppose An ⊆ An+1 for all n ≥ 1. Show that

∞
!
= lim P (An )
[
P An
n→∞
n=1

(b) Suppose An ⊇ An+1 for all n ≥ 1. Show that

∞
!
= lim P (An )
\
P An
n→∞
n=1

Version: – November 19, 2024

10 basic concepts

1.2 equally likely outcomes

When a sample space S consists of only a countable collection of outcomes, describing the
probability of each individual outcome is sufficient to describe the probability of all events.
This is because if A ⊂ S we may simply compute

P ({ω}).
[ X
P (A) = P ( {ω}) =
ω∈A ω∈A

This assignment of probabilities to each outcome is called a “distribution” since it describes

how probability is distributed amongst the possibilities. Perhaps the simplest example
arises when there are a finite collection of equally likely outcomes. Think of examples such
as flipping a fair coin (“heads” and “tails” are equally likely to occur), rolling a fair die (1, 2,
3, 4, 5, and 6 are equally likely), or drawing a set of numbers for a lottery (many possibilities,
but in a typical lottery, each outcome is as likely as any other). Such distributions are
common enough that it is useful to have shorthand notations for them. In the case of a
sample space S = {ω1 , . . . , ωn } where each outcome is equally likely, the probability is
referred to as a “uniform distribution” and is denoted by Uniform({ω1 , . . . , ωn }). In such
situations, computing probabilities simply reduces to computing the number of outcomes
in a given event and consequently becomes a combinatorial problem.

Theorem 1.2.1. Uniform({ω1 , . . . , ωn }) : Let S = {ω1 , ω2 , . . . , ωn } be a non-

|E|
empty, finite set. If E ⊂ S is any subset of S, let P (E ) = |S| (where |E| represents
the number of elements of E). Then P defines a probability on S and P assigns
equal probability to each individual outcome in S.

Proof. As E ⊂ S, we know that |E| ≤ |S| and so 0 ≤ P (E ) ≤ 1. So, we must prove that
P satisfies the two probability axioms.
|S|
The first axiom is satisfied because P (S ) = = 1. To verify the second axiom,
|S|
suppose E1 , E2 , ... is a countable collection of disjoint events. As S is finite, only finitely
many of these Ej can be non-empty, so we may list the non-empty events as E1 , E2 , . . . , En .
For j > n we know Ej = ∅ and so P (Ej ) = 0 by definition. As the events are disjoint,
to find the number of elements in their union we simply add the elements of each event
separately. That is, |E1 ∪ E2 ∪ · · · ∪ En | = |E1 | + |E2 | + · · · + |En |, and so

∞ n
|Ej |
P
∪ Ej
 
∞ n n ∞
j =1 j =1 |Ej |
P (Ej ).
[ X X X
P Ej  = = = = P ( Ej ) =
j =1
|S| |S| j =1
|S| j =1 j =1

Version: – November 19, 2024

 1.2 equally likely outcomes 11

Finally, let ω ∈ S be any single outcome and let E = {ω}. Then P (E ) = |S| ,
1
so every
outcome in S is equally likely. ■

Example 1.2.2. A deck of twenty cards labeled 1, 2, 3, . . . , 20 is shuffled and a card selected
at random. What is the probability that the number on the card is a multiple of six?
The description of the scenario suggests that each of the twenty cards is as likely to
be chosen as any other. In this case S = {1, 2, 3, ..., 20} while E = {6, 12, 18}. Therefore,
|E|
P (E ) = |S| = 3
20 = 0.15. There is a 15% chance that the card will be a multiple of six. ■
Example 1.2.3. Two dice are rolled. How likely is it that their sum will equal eight?
Since we are looking at a sum of dice, it might be tempting to regard the sample space
as S = {2, 3, 4, ..., 11, 12}, the collection of possible sums. While this is a possible approach
(and one we will return to later), it is not the case that all of these outcomes are equally
likely. Instead we can view an experiment as tossing a first die and a second die and
recording the pair of numbers that occur on each of the dice. Each of these pairs is as
likely as any other to occur. So
 


 (1, 1), (1, 2), (1, 3), (1, 4), (1, 5), (1, 6) 


(2, 1), (2, 2), (2, 3), (2, 4), (2, 5), (2, 6)

 


 


 

(3, 1), (3, 2), (3, 3), (3, 4), (3, 5), (3, 6)

 

S=


 (4, 1), (4, 2), (4, 3), (4, 4), (4, 5), (4, 6) 


 
(5, 1), (5, 2), (5, 3), (5, 4), (5, 5), (5, 6)

 


 

 
(6, 1), (6, 2), (6, 3), (6, 4), (6, 5), (6, 6)

 


and |S| = 6 × 6 = 36. The event that the sum of the dice is an eight is

E = {(2, 6), (3, 5), (4, 4), (5, 3), (6, 2)} .

|E|
Therefore, P (E ) = |S| = 36 .
5
■
Example 1.2.4. A seven letter code is selected at random with every code as likely to
be selected as any other code (so AQRVTAS and CRXAOLZ would be two possibilities).
How likely is it that such a code has at least one letter used more than once? (This would
happen with the first code above with a repeated A - but not with the second).
As with the examples above, the solution amounts to counting numbers of outcomes.
However, unlike the examples above the numbers involved here are quite large and we will
need to use some combinatorics to find the solution. The sample space S consists of all
seven-letter codes from AAAAAAA to ZZZZZZZ. Each of the seven spots in the code
could be any of twenty-six letters, so |S| = 267 = 8, 031, 810, 176. If E is the event for
which there is at least one letter used more than once, it is easier to count E c , the event

Version: – November 19, 2024

12 basic concepts

where no letter is repeated. Since in this case each new letter rules out a possibility for the
next letter there are 26 × 25 × 24 × 23 × 22 × 21 × 20 = 3, 315, 312, 000 such possibilities.
This lets us compute P (E c ) = 3,315,312,000
8,031,810,176 from which we find P (E ) = 1 − P (E c ) =
4,716,498,176
8,031,810,176 ≈ 0.587. That is, there is about a 58.7% chance that such a code will have a
repeated letter. ■
Example 1.2.5. A group of twelve people includes Grant and Dilip. A group of three
people is to be randomly selected from the twelve. How likely is it that this three-person
group will include Grant, but not Dilip?
Here, S is the collection of all three-person groups, each of which is as likely to be
selected as any other. The number of ways of selecting a three-person group from a pool of
twelve is |S| = (12
3 ) = 220. The event E consists of those three-person groups that include
Grant, but not Dilip. Such groups must include two people other than ! Grant and there
10
are ten people remaining from which to select the two, so |E| = = 45. Therefore,
2
45
P (E ) = 220 9
= 44 . ■

exercises

Ex. 1.2.1. A day is selected at random from a given week with each day as likely to be
selected as any other.

(a) What is the sample space S? What is the size of S?

(b) Let E be the event that the selected day is a Saturday or a Sunday. What is the
probability of E.

Ex. 1.2.2. A box contains 500 envelopes, of which 50 contain Rs 100 in cash, 100 contain
Rs 50 in cash and 350 contain Rs 10. An envelope can be purchased at Rs 25 from the
owner, who will pick an envelope at random and give it to you. Write down the sample
space for the net money gained by you. If each envelope is as likely to be selected as any
other envelope, what is the probability that the first envelope purchased contains less than
Rs 100?
Ex. 1.2.3. Three dice are tossed.

(a) Describe (in words) the sample space S and give an example of an object in S.

(b) What is the size of S?

(c) Let E be the event that the first two dice both come up “1”. What is the size of E?
What is the probability of E?

Version: – November 19, 2024

 1.2 equally likely outcomes 13

(d) Let G be the event that the three dice show three different numbers. What is the
size of G? What is the probability of G?

(e) Let F be the event that the third die is larger than the sum of the first two. What
is the size of F ? What is the probability of F ?

Ex. 1.2.4. Suppose that each of three women at a party throws her hat into the center of
the room. The hats are first mixed up and then each one randomly selects a hat. Describe
the probability space for the possible selection of hats. If all of these selections are equally
likely, what is the probability that none of the three women selects her own hat?

Ex. 1.2.5. A group of ten people includes Sona and Adam. A group of five people is to be
randomly selected from the ten. How likely is it that this group of five people will include
neither Sona nor Adam?

Ex. 1.2.6. There are eight students with two females and six males. They are split into
two groups A and B, of four each.

(a) In how many different ways can this be done?

(b) What is the probability that two females end up in group A?

(c) What is the probability that there is one female in each group?

Ex. 1.2.7. Sheela has lost her key to her room. The security officer gives her 50 keys and
tells her that one of them will open her room. She decides to try each key successively
and notes down the number of the attempt at which the room opens. Describe the sample
space for this experiment. Do you think it is realistic that each of these outcomes is equally
likely? Why or why not?

Ex. 1.2.8. Suppose that n balls, of which k are red, are arranged at random in a line.
What is the probability that all the red balls are next to each other?

Ex. 1.2.9. Consider a deck of 50 cards. Each card has one of 5 colors (black, blue, green,
red, and yellow), and is printed with a number (1,2,3,4,5,6,7,8,9, or 10) so that each of the
50 color/number combinations is represented exactly once. A hand is produced by dealing
out five different cards from the deck. The order in which the cards were dealt does not
matter.

(a) How many different hands are there?

(b) How many hands consist of cards of identical color? What is the probability of being
dealt such a hand?

Version: – November 19, 2024

14 basic concepts

(c) What is the probability of being dealt a hand that contains exactly three cards with
one number, and two cards with a different number?

(d) What is the probability of being dealt a hand that contains two cards with one
number, two cards with a different number, and one card of a third number?

Ex. 1.2.10. Suppose you are in charge of quality control for a light bulb manufacturing
company. Suppose that in the process of producing 100 light bulbs, either all 100 bulbs
will work properly, or through some manufacturing error twenty of the 100 will not work.
Suppose your quality control procedure is to randomly select ten bulbs from a 100 bulb
batch and test them to see if they work properly. How likely is this procedure to detect if
a batch has bad bulbs in it?
Ex. 1.2.11. A fair die is rolled five times. What is the probability of getting at least two
5’s and at least two 6’s among the five rolls.
Ex. 1.2.12. (The “Birthday Problem”) For a group of N people, if their birthdays were
listed one-by-one, there are 365N different ways that such a list might read (if we ignore
February 29 as a possibility). Suppose each of those possible lists is as likely as any other.

(a) For a group of two people, let E be the event that they have the same birthday.
What is the size of E? What is the probability of E?

(b) For a group of three people, let F be the event that at least two of the three have
the same birthday. What is the size of F ? What is the probability of F ? (Hint: It
is easier to find the size of F c than it is to find the size of F ).

(c) For a group of four people, how likely is it that at least two of the four have the same
birthday?

(d) How large a group of people would you need to have before it becomes more likely
than not that at least two of them share a birthday?

Ex. 1.2.13. A coin is tossed 100 times.

(a) How likely is it that the 100 tosses will produce exactly fifty heads and fifty tails?

(b) How likely is it that the number of heads will be between 50 and 55 (inclusive)?

Ex. 1.2.14. Suppose I have a coin that I claim is “fair” (equally likely to come up heads or
tails) and that my friend claims is weighted towards heads. Suppose I flip the coin twenty
times and find that it comes up heads on sixteen of those twenty flips. While this seems to
favor my friend’s hypothesis, it is still possible that I am correct about the coin and that

Version: – November 19, 2024

 1.2 equally likely outcomes 15

1.0

0.8
Probability of common birthday

0.6

0.4

0.2

0.0

0 10 20 30 40 50

Number of people in group

Figure 1.2: The birthday problem discussed in Exercise 1.2.12

just by chance the coin happened to come up heads more often than tails on this series of
flips. Let S be the sample space of all possible sequences of flips. The size of S is then 220 ,
and if I am correct about the coin being “fair”, each of these outcomes are equally likely.

(a) Let E be the event that exactly sixteen of the flips come up heads. What is the size
of E? What is the probability of E?

(b) Let F be the event that at least sixteen of the flips come up heads. What is the size
of F ? What is the probability of F ?

Note that the probability of F is the chance of getting a result as extreme as the one I
observed if I happen to be correct about the coin being fair. The larger P (F ) is, the more
reasonable seems my assumption about the coin being fair. The smaller P (F ) is, the more
that assumption looks doubtful. This is the basic idea behind the statistical concept of
“hypothesis testing” which we will revisit in Chapter 9.

Ex. 1.2.15. Suppose that r indistinguishable balls are placed in n distinguishable boxes so
that each distinguishable arrangement is equally likely. Find the probability that no box
will be empty.

Version: – November 19, 2024

16 basic concepts

Ex. 1.2.16. Suppose that 10 potato sticks are broken into two - one long and one short
piece. The 20 pieces are now arranged into 10 random pairs chosen uniformly.

(a) Find the probability that each of pairs consists of two pieces that were originally
part of the same potato stick.

(b) Find the probability that each pair consists of a long piece and a short piece.

Ex. 1.2.17. Let S be a non-empty, countable (finite or infinite) set such that for each
ω ∈ S, 0 ≤ pω ≤ 1. Let F be the collection of all events. Suppose P : F → [0, 1] is given by

pω ,
X
P (E ) =
ω∈E

for any event E.

(a) Show that P satisfies Axiom 2 in Definition 1.1.3.

(b) Further, conclude that if P (S ) = 1 then P defines a probability on S.

1.3 conditional probability and bayes’ theorem

In the previous section we introduced an axiomatic definition of “probability” and discussed

the concept of an “event”. Now we look at ways in which the knowledge that one event has
occurred may be used as information to inform and alter the probability of another event.
Example 1.3.1. Consider the experiment of tossing a fair coin three times with sample
space S = {hhh, hht, hth, htt, thh, tht, tth, ttt}. Let A be the event that there are two or
more heads. As all outcomes are equally likely,

|A| |{hhh, hht, hth, thh}| 1

P (A) = = = .
|S| 8 2

Let B be the event that there is a head in the first toss. As above,

|B| |{hhh, hht, hth, htt}| 1

P (B ) = = = .
|S| 8 2

Now suppose we are asked to find the probability of at least two or more heads among
the three tosses, but we are also given the additional information that the first toss was a
head. In other words, we are asked to find the probability of A, given the information that
event B has definitely occurred. As the additional information guarantees that B is now a
list of all possible outcomes, it makes intuitive sense to view the event B as a new sample

Version: – November 19, 2024

 1.3 conditional probability and bayes’ theorem 17

space and then identify the subset A ∩ B = {hhh, hht, hth} of B consisting of outcomes
for which there are at least two heads. We conclude that the probability of at least two or
more heads in three tosses given that the first toss was a head is

|A ∩ B| 3
= . ■
|B| 4

This is a legitimate way to view the problem and it leads to the correct solution. However,
this method has one very serious drawback—it requires us to change both our sample
space and our probability function in order to carry out the computation. It would be
preferable to have a method that allows us to work within the original framework of the
sample space S and to talk about the “conditional probability” of A given that the result
of the experiment will be an outcome in B. This is denoted as P (A | B ) and is read as
“the (conditional) probability of A given B.”
Suppose S is a finite set of equally likely outcomes from a given experiment. Then for
any two non-empty events A and B, the conditional probability of A given B is given by

|A∩B|
|A ∩ B| |S| P (A ∩ B )
= |B|
= .
|B| P (B )
|S|

This leads us to a formal definition of conditional probability for general sample spaces.

Definition 1.3.2. (Conditional Probability) Let S be a sample space with

probability P . Let A and B be two events with P (B ) > 0. Then the conditional
probability of A given B written as P (A | B ) and is defined by

P (A ∩ B )
P (A | B ) = .
P (B )

This definition makes it possible to compute a conditional probability in terms of the

original (unconditioned) probability function.
Example 1.3.3. A pair of dice are thrown. If it is known that one die shows a 4, what is
the probability the other die shows a 6?
Let B be the event that one of the dice shows a 4. So,
n o
B = (4, 1), (4, 2), (4, 3), (4, 4), (4, 5), (4, 6), (1, 4), (2, 4), (3, 4), (5, 4), (6, 4) .

Let A be the event that one of the dice is a 6. So,

n o
A = (6, 1), (6, 2), (6, 3), (6, 4), (6, 5), (6, 6), (1, 6), (2, 6), (3, 6), (4, 6), (5, 6) .

Version: – November 19, 2024

18 basic concepts

then n
A ∩ B = (4, 6), (6, 4)}.

Hence

P (one die shows 6 | one die shows 4)

P (A ∩ B )
= P (A | B ) =
P (B )
P (one die shows 6 and the other shows 4)
=
P (one die shows 4)
2/36 2
= = .
11/36 11
■

In many applications, the conditional probabilities are implicitly defined within the context
of the problem. In such cases, it is useful to have a method for computing non-conditional
probabilities from the given conditional ones. Two such methods are given by the next
results and the subsequent examples.

Example 1.3.4. An economic model predicts that if interest rates rise, then there is a 60%
chance that unemployment will increase, but that if interest rates do not rise, then there is
only a 30% chance that unemployment will increase. If the economist believes there is a
40% chance that interest rates will rise, what should she calculate is the probability that
unemployment will increase?

Let B be the event that interest rates rise and A be the event that unemployment
increases. We know the values

P (B ) = 0.4, P (B c ) = 0.6, P (A | B ) = 0.6, and P (A | B c ) = 0.3.

Using the axioms of probability and definition of conditional probability we have

P (A) = P ((A ∩ B ) ∪ (A ∩ B c ))
= P ((A ∩ B )) + P (A ∩ B c ))
= P (A | B )P (B ) + P (A | B c )P (B c )
= 0.6 × 0.4 + 0.3 × 0.6 = 0.42.

So there is a 42% chance that unemployment will increase. ■

Version: – November 19, 2024

 1.3 conditional probability and bayes’ theorem 19

Theorem 1.3.5. Let A be an event and let {Bi : 1 ≤ i ≤ n} be a disjoint collection

n
of events for which P (Bi ) > 0 for all i and such that A ⊂
S
Bi . Suppose P (Bi )
i=1
and P (A | Bi ) are known. Then P (A) may be computed as
n
P ( A | Bi ) P ( Bi ) .
X
P (A) =
i=1

Proof. The events (A ∩ Bi ) and (A ∩ Bj ) are disjoint if i ̸= j and

n n
!
[ [
( A ∩ Bi ) = A ∩ Bi = A.
i=1 i=1

So,
n
!
[
P (A) = P ( A ∩ Bi )
i=1
n
X
= P ( A ∩ Bi )
i=1
n
P ( A | Bi ) P ( Bi ) .
X
=
i=1
■

A nearly identical proof holds when there are only countably many Bi (see Exercise 1.3.11).

Example 1.3.6. Suppose we have coloured balls distributed in three boxes in quantities as
given by the table below:

Box 1 Box 2 Box 3

Red 4 3 3
Green 3 3 4
Blue 5 2 3

A box is selected at random. From that box a ball is selected at random. How likely is it
that a red ball is drawn?
Let B1 , B2 , and B3 be the events that Box 1, 2, or 3 is selected, respectively. Note
that these events are disjoint and cover all possibilities in the sample space. Let R be the
event that the selected ball is red. Then by Theorem 1.3.5,

P ( R ) = P ( R | B1 ) P ( B1 ) + P ( R | B2 ) P ( B2 ) + P ( R | B3 ) P ( B3 )
4 1 3 1 3 1 121
= · + · + · = . ■
12 3 8 3 10 3 360

Version: – November 19, 2024

20 basic concepts

Example 1.3.7. (Polya’s Urn Scheme) Suppose there is an urn that contains r red balls
and b black balls. A ball is drawn at random and its colour noted. It is replaced with c > 0
balls of the same colour. The procedure is then repeated. For j = 1, 2, . . . , let Rj and Bj
be the events that the j-th ball drawn is red and black respectively. Clearly P (R1 ) = r
b+r
and P (B1 ) = b+r .
b
When the first ball is replaced, c new balls will be added to the urn, so
that when the second ball is drawn there will be r + b + c balls available. From this it can
easily be checked that P (R2 | R1 ) = r +c
b+r +c and P (R2 | B1 ) = b+r +c .
r
Noting that R1 and
B1 are disjoint and together represent the entire sample space, P (R2 ) can be computed as

P (R2 ) = P (R1 )P (R2 | R1 ) + P (B1 )P (R2 | B1 )

r r+c b r
= · + ·
b+r b+r+c b+r b+r+c
r (r + b + c) r
= = P ( R1 ) .
(r + b + c)(b + r ) b + r

One can show that P (Rj ) = r

b+r for all j ≥ 1. ■

The urn schemes were originally developed by George Polya (1887–1985). Various modifi-
cations to Polya’s urn scheme are discussed in the exercises.
Above we have described how conditioning on an event B may be viewed as modifying
the original probability based on the additional information provided by knowing that
B has occurred. Frequently in applications, we gain information more than once in the
process of an experiment. The following theorem shows how to deal with such a situation.

Theorem 1.3.8. For an integer n ≥ 2, let A1 , A2 , . . . , An be a collection of events

n−1
T
for which Aj has positive probability. Then,
j =1

   
n n j−1
Ak  .
\ Y \
P Aj  = P (A1 ) · P Aj
j =1 j =2 k =1

The proof of this theorem is left as Exercise 1.3.14, but we will provide a framework in
which to make sense of the equality. Usually the events A1 , . . . , An are viewed as a sequence
in time for which we know the probability of a given event provided that all of the others
before it have already occurred. Then we can calculate P (A1 ∩ A2 ∩ · · · ∩ An ) by taking the
product of the values P (A1 ), P (A2 | A1 ), P (A3 | A1 ∩ A2 ), . . . , P (An | A1 ∩ · · · ∩ An−1 ).

Example 1.3.9. A probability class has fifteen students - four seniors, eight juniors, and
three sophomores. Three different students are selected at random to present homework

Version: – November 19, 2024

 1.3 conditional probability and bayes’ theorem 21

problems. What is the probability the selection will be a junior, a sophomore, and a junior
again, in that order?
Let A1 be the event that the first selection is a junior. Let A2 be the event that the
second selection is a sophomore, and let A3 be the event that the third selection is a junior.
The problem asks for P (A1 ∩ A2 ∩ A3 ) which we can calculate using Theorem 1.3.8.

P ( A1 ∩ A2 ∩ A3 ) = P ( A1 ) P ( A2 | A1 ) P ( A3 | A1 ∩ A2 )
8 3 7 4
= · · = . ■
15 14 13 65

1.3.1 Bayes’ Theorem

It is often the case that we know the conditional probability of A given B, but want to
know the conditional probability of B given A instead. It is possible to calculate one
quantity from the other using a formula known as Bayes’ theorem. We introduce this with
a motivating example.

Example 1.3.10. We return to Example 1.3.6. In that example we had three boxes
containing balls given by the table below.

Box 1 Box 2 Box 3

Red 4 3 3
Green 3 3 4
Blue 5 2 3

A box is selected at random. From the box a ball is selected at random. When we looked
at conditional probabilities we saw how to determine the probability of an event such as
{the ball drawn is red}. Now suppose we know the ball is red and want to determine
the probability of the event {the ball was drawn from box 3}. That is, if R is the event
that a red ball is chosen and if B1 , B2 , and B3 are the events that boxes 1, 2, and 3 are
selected, we want to determine the conditional probability P (B3 | R). The difficulty is
that while the conditional probabilities P (R | B1 ), P (R | B2 ), and P (R | B3 ) are easy to
determine, calculating the conditional probability with the order of the events reversed is
not immediately obvious.
Using the definition of conditional probability we have that
  P (B3 ∩ R)
P B3 | R = ,
P (R )

Version: – November 19, 2024

22 basic concepts

so we can rewrite P (B3 ∩ R) = P (R | B3 )P (B3 ) = 3

= 101
10 . On the other hand, we
× 1
3
can decompose the event R over which box was chosen. This is exactly what we did to
360 . Hence,
solve Example 1.3.6 where we found that P (R) = 121

P ( B3 ∩ R ) 1/10 36
P ( B3 | R ) = = = ≈ 0.298.
P (R ) 121/360 121

So if we know that a red ball was drawn, there is slightly less than a 30% chance that it
came from Box 3. ■

In the above example the description of the experiment allowed us to determine P (B1 ),
P (B2 ), P (B3 ), P (R | B1 ), P (R | B2 ), and P (R | B3 ). We were then able to use the
definition of conditional probability to find P (B3 | R). Such a computation can be done in
general.

Theorem 1.3.11. (Bayes’ Theorem) Suppose A is an event, {Bi : 1 ≤ i ≤ n}

are a collection of disjoint events whose union contains all of A. Further assume
that P (A) > 0 and P (Bi ) > 0 for all 1 ≤ i ≤ n. Then for any 1 ≤ i ≤ n,

P (A | Bi )P (Bi )
P ( Bi | A ) = P
n . (1.3.1)
P ( A | Bj ) P ( Bj )
j =1

Proof. The left hand side of (1.3.1) can be written as

P ( Bi ∩ A ) P ( A | Bi ) P ( Bi )
P ( Bi | A ) = = !
P (A) n
A ∩ Bj
S
P
j =1
P (A | Bi )P (Bi ) P ( A | Bi ) P ( Bi )
= n = P
n .
P (A ∩ Bj ) P (A | Bj )P (Bj )
P
j =1 j =1 ■

Equation (1.3.1) is sometimes referred to as “Bayes’ formula” or “Bayes’ rule” as well. This
result is originally due to Thomas Bayes (1701–1761).

Example 1.3.12. Shyam is randomly selected from the citizens of Hyderabad by the Health
authorities. A laboratory test on his blood sample tells Shyam that he has tested positive
for Swine Flu. It is found that 95% of people with Swine Flu test positive but 2% of people
without the disease will also test positive. Suppose that 1% of the population has the
disease. What is the probability that Shyam indeed has the Swine Flu ?

Version: – November 19, 2024

 1.3 conditional probability and bayes’ theorem 23

Consider the events A = { Shyam has Swine Flu } and B = { Shyam tested postive
for Swine Flu }. We are given:

P (B | A) = 0.95, P (B | Ac ) = 0.02, and P (A) = 0.01.

Using Bayes’ Theorem, we have

P (B | A)P (A)
P (A | B ) =
P (B | A)P (A) + P (B | Ac )P (Ac )
(0.95)(0.01)
= = 0.324
(0.95)(0.01) + (0.02)(0.99)

Despite testing positive, there is only a 32.4 percent chance that Shyam has the disease. ■

exercises

Ex. 1.3.1. There are two dice, one red and one blue, sitting on a table. The red die is a
standard die with six sides while the blue die is tetrahedral with four sides, so the outcomes
1, 2, 3, and 4 are all equally likely. A fair coin is flipped. If that coin comes up heads, the
red die will be rolled, but if the coin comes up tails the blue die will be rolled.

(a) Find the probability that the rolled die will show a 1.

(b) Find the probability that the rolled die will show a 6.

Ex. 1.3.2. A pair of dice are thrown. It is given that the outcome on one die is a 3. what
is the probability that the sum of the outcomes on both dice is greater than 7?
Ex. 1.3.3. Box A contains four white balls and three black balls and Box B contains three
white balls and five black balls.

(a) Suppose a box is selected at random and then a ball is chosen from the box. If the
ball drawn is black then what is the probability that it was from Box A?

(b) Suppose instead that one ball is drawn at random from Box A and placed (unseen)
in Box B. What is the probability that a ball now drawn from Box B is black?

Ex. 1.3.4. Tomorrow the weather will either be sunny, cloudy, or rainy. There is a 60%
chance tomorrow will be cloudy, a 30% chance tomorrow will be sunny, and a 10% chance
that tomorrow will be rainy. If it rains, I will not go on a walk. But if it is cloudy, there
is a 90% chance I will take a walk and if it’s sunny there is a 70% chance I will take a
walk. If I take a walk on a cloudy day, there is an 80% chance I will walk further than

Version: – November 19, 2024

24 basic concepts

five kilometers, but if I walk on a sunny day, there’s only a 50% chance I will walk further
than five kilometers. Using the percentages as given probabilities, answer the following
questions:

(a) How likely is it that tomorrow will be cloudy and I will walk over five kilometers?

(b) How likely is it I will take a walk over five kilometers tomorrow?

Ex. 1.3.5. A box contains B black balls and W balls, where W ≥ 3, B ≥ 3. A sample of
three balls is drawn at random with each drawn ball being discarded (not put back into
the box) after it is drawn. For j = 1, 2, 3 let Aj denote the event that the ball drawn on
the j th draw is white. Find P (A1 ), P (A2 ) and P (A3 ).
Ex. 1.3.6. There are two sets of cards, one red and one blue. The red set has four cards -
one that reads 1, two that read 2, and one that reads 3. An experiment involves flipping a
fair coin. If the coin comes up heads a card will be randomly selected from the red set
(and its number recorded) while if the coin comes up tails a card will be randomly selected
from the blue set (and its number recorded). You can construct the blue set of cards in
any way you see fit using any number of cards reading 1, 2, or 3. Explain how to build the
blue set of cards to make each of the experimental outcomes 1, 2, 3 equally likely.
Ex. 1.3.7. There are three tables, each with two drawers. Table 1 has a red ball in each
drawer. Table 2 has a blue ball in each drawer. Table 3 has a red ball in one drawer and a
blue ball in the other. A table is chosen at random, then a drawer is chosen at random
from that table. Find the conditional probability that Table 1 is chosen, given that a red
ball is drawn.
Ex. 1.3.8. In the G.R.E advanced mathematics exam, each multiple choice question has 4
choices for an answer. A prospective graduate student taking the test knows the correct
answer with probability 34 . If the student does not know the answer, she guesses randomly.
Given that a question was answered correctly, find the conditional probability that the
student knew the answer.
Ex. 1.3.9. You first roll a fair die, then toss as many fair coins as the number that showed
on the die. Given that 5 heads are obtained, what is the probability that the die showed 5?
Ex. 1.3.10. Manish is a student in a probability class. He gets a note saying, “I’ve organized
a probability study group tonight at 7pm in the coffee shop. Come if you want.” The note
is signed “Hannah”. However, Manish has class with two different Hannahs and he isn’t
sure which one sent the note. He figures that there is a 75% chance that Hannah A. would
have organized such a study group, but only a 25% chance that Hannah B. would have
done so. However, he also figures that if Hannah A. had organized the group, there is an

Version: – November 19, 2024

 1.3 conditional probability and bayes’ theorem 25

80% chance that she would have planned to meet on campus and only a 20% chance that
she would have planned to meet in the coffee shop. While if Hannah B. had organized the
group there is a 10% chance she would have planned for it on campus and a 90% chance
she would have chosen the coffee shop. Given all this information, determine whether it is
more likely that Manish should think the note came from Hannah A. or from Hannah B.
Ex. 1.3.11. State and prove a version of

(a) Theorem 1.3.5 when {Bi } is a countably infinite collection of disjoint events.

(b) Theorem 1.3.11 when {Bi } is a countably infinite collection of disjoint events.

Ex. 1.3.12. A bag contains 100 coins. Sixty of the coins are fair. The rest are biased to
land heads with probability p (where 0 ≤ p ≤ 1). A coin is drawn at random from the bag
and tossed.

(a) Given that the outcome was a head what is the conditional probability that it is a
biased coin?

(b) Evaluate your answer to (a) when p = 0. Can you explain why this answer should
be intuitively obvious?

(c) Evaluate your answer to (a) when p = 12 . Can you explain why this answer should
be fairly intuitive as well?

(d) View your answer to part (a) as a function f (p). Show that f (p) is an increasing
function when 0 ≤ p ≤ 1. Give an interpretation of this fact in the context of the
problem.

Ex. 1.3.13. An urn contains b black balls and r red balls. A ball is drawn at random. The
ball is replaced into the urn along with c balls of its colour and d balls of the opposite
colour. Then another random ball is drawn and the procedure is repeated.

(a) What is the probability that the second ball drawn is a red ball?

(b) Assume c = d. What is the probability that the second ball drawn is a black ball?

(c) Still assuming c = d, what is the probability that the nth ball drawn is a black ball?

(d) Assume c > 0 and d = 0, what is the probability that the nth ball drawn is a black
ball?

(6) Can you comment on the answers to (b) and/or (c) if the assumption that c = d was
removed?

Version: – November 19, 2024

26 basic concepts

Ex. 1.3.14. Use the following steps to prove Theorem 1.3.8.

(a) Prove Theorem 1.3.8 for the n = 2 case. (Hint: The proof should follow immediately
from the definition of conditional probability).

(b) Prove Theorem 1.3.8 for the n = 3 case. (Hint: Rewrite the conditional probabilities
in terms of ordinary probabilities).

(c) Prove Theorem 1.3.8 generally. (Hint: One method is to use induction, and parts (a)
and (b) have already provided a starting point).

1.4 independence

In the previous section we have seen instances where the probability of an event may
change given the occurrence of a related event. However it is instructive and useful to
study the case of two events where the occurrence of one has no effect on the probability
of the other. Such events are said to be “independent”.

Example 1.4.1. Suppose we toss a coin three times. Then the sample space

S = {hhh, hht, hth, htt, thh, tht, tth, ttt} .

Define A = {hhh, hht, hth, htt} = {the first toss is a head}, and similarly define B =
{hhh, hht, thh, tht} = {the second toss is a head}. Note that P (A) = 1
2 = P (B ), while

P (A ∩ B ) |A ∩ B| 2 1
P (A | B ) = = = =
P (A) |B| 4 2

and
P (A ∩ B c ) |A ∩ B c | 2 1
P (A | B c ) = = = = .
c
P (B ) c
|B | 4 2
We have shown that P (A) = P (A | B ) = P (A | B c ). Therefore we conclude that the
occurrence (or non-occurrence) of B has no effect on the probability of A. ■
This is the sort of condition we would want in a definition of independence. However, since
defining P (A | B ) requires that P (B ) > 0, our formal definition of “independence” will
appear slightly different.

Definition 1.4.2. (Independence) Two events A and B are independent if

P (A ∩ B ) = P (A)P (B ).

Version: – November 19, 2024

 1.4 independence 27

Example 1.4.3. Suppose we roll a die twice and denote as an ordered pair the result of
the rolls. Suppose

E = { a six appears on the first roll } = {(6, 1), (6, 2), (6, 3), (6, 4), (6, 5), (6, 6)}

and

F = { a six appears on the second roll } = {(1, 6), (2, 6), (3, 6), (4, 6), (5, 6), (6, 6)} .

As E ∩ F = {(6, 6)}, it is easy to see that

1 6 1 6 1
P (E ∩ F ) = , P (E ) = = , P (F ) = = .
36 36 6 36 6

So E, F are independent as P (E ∩ F ) = P (E )P (F ). ■

Using the definition of conditional probability it is not hard to show (see Exercise 1.4.9)
that if A and B are independent, and if 0 < P (B ) < 1 then

P (A | B ) = P (A) = P (A | B c ). (1.4.1)

If P (A) > 0 then the equations of (1.4.1) also hold with the roles of A and B reversed.
Thus, independence implies four conditional probability equalities.
If we want to extend our definition of independence to three events A1 , A2 , and A3 , we
would certainly want

P (A1 ∩ A2 ∩ A3 ) = P (A1 )P (A2 )P (A3 ) (1.4.2)

to hold. We would also want any pair of the three events to be independent of each other.
It is tempting to hope that pairwise independence is enough to imply (1.4.2). However,
this is not true, as shown by the next example.

Example 1.4.4. Suppose we toss a fair coin two times. Consider the three events A1 =
{hh, tt}, A2 = {hh, ht}, and A3 = {hh, th}. Then it is easy to calculate that

1
P ( A1 ) = P ( A2 ) = P ( A3 ) = ,
2
1
P (A1 ∩ A2 ) = P (A1 ∩ A3 ) = P (A2 ∩ A3 ) = , and
4
1
P ( A1 ∩ A2 ∩ A3 ) = .
4

So even though A1 , A2 and A3 are pairwise independent, they do not satisfy (1.4.2). ■

Version: – November 19, 2024

28 basic concepts

It may also be tempting to hope that (1.4.2) is enough to imply pairwise independence,
but that is not true either (see Exercise 1.4.6). The root of the problem is that, unlike the
two event case, (1.4.2) does not imply that equality holds if any of the Ai are replaced by
their complements. One solution is to insist that the multiplicative equality hold for any
intersection of the events or their complements, which gives us the following definition.

Definition 1.4.5. (Mutual Independence) A finite collection of events

A1 , A2 , . . . , An is mutually independent if

P (E1 ∩ E2 ∩ · · · ∩ En ) = P (E1 )P (E2 ) . . . P (En ). (1.4.3)

whenever Ej is either Aj or Acj .

An arbitrary collection of events At where t ∈ I for some index set I is mutually
independent if every finite subcollection is mutually independent.

Thus, mutual independence of n events is defined in terms of 2n equations. It is a fact

(see Exercise 1.4.10) that if a collection of events is mutually independent, then so is any
subcollection.

exercises

Ex. 1.4.1. In the first semifinal of an international volleyball tournament Brazil has a 60%
chance to beat Pakistan. In the other semifinal Poland has a 70% chance to beat Mexico.
If the results of the two matches are independent, what is the probability that Pakistan
will meet Poland in the tournament final?

Ex. 1.4.2. A manufacturer produces nuts and markets them as having 50mm radius. The
machines that produce the nuts are not perfect. From repeated testing, it was established
that 15% of the nuts have radius below 49mm and 12% have radius above 51mm. If two
nuts are randomly (and independently) selected, find the probabilities of the following
events:

(a) The radii of both the nuts are between 49mm and 51mm;

(b) The radius of at least one nut exceeds 51mm.

Ex. 1.4.3. Four tennis players (Avinash, Ben, Carlos, and David) play a single-elimination
tournament with Avinash playing David and Ben playing Carlos in the first round and the
winner of each of those contests playing each other in the tournament final. Below is the

Version: – November 19, 2024

 1.4 independence 29

chart giving the percentage chance that one player will beat the other if they play. For
instance, Avinash has a 30% chance of beating Ben if they happen to play.

Avinash Ben Carlos David

Avinash - 30% 55% 40%
Ben - - 80% 45%
Carlos - - - 15%
David - - - -

Suppose the outcomes of the games are independent. For each of the four players,
determine the probability that player wins the tournament. Verify that the calculated
probabilities sum to 1.

Ex. 1.4.4. Let A and B be events with P (A) = 0.8 and P (B ) = 0.7.

(a) What is the largest possible value of P (A ∩ B )?

(b) What is the smallest possible value of P (A ∩ B )?

(c) What is the value of P (A ∩ B ) if A and B are independent?

Ex. 1.4.5. Suppose we toss two fair dice. Let E1 denote the event that the sum of the dice
is six. E2 denote the event that sum of the dice equals seven. Let F denote the event that
the first die equals four. Is E1 independent of F ? Is E2 independent of F ?

Ex. 1.4.6. Suppose a bowl has twenty-seven balls. One ball is black, two are white, and
eight each are green, red, and blue. A single ball is drawn from the bowl and its color is
recorded. Define

A = {the ball is either black or green}

B = {the ball is either black or red}
C = {the ball is either black or blue}

(a) Calculate P (A ∩ B ∩ C ).

(b) Calculate P (A)P (B )P (C ).

(c) Are A, B, and C mutually independent? Why or why not?

Version: – November 19, 2024

30 basic concepts

Ex. 1.4.7. There are 150 students in the Probability 101 class. Of them, ninety are female,
sixty use a pencil (instead of a pen), and thirty are wearing eye glasses. A student is chosen
at random from the class. Define the following events:

A1 = {the student is a female}

A2 = {the student uses a pencil}
A3 = {the student is wearing eye glasses}

(a) Show that it is impossible for these events to be mutually independent.

(b) Give an example to show that it may be possible for these events to be pairwise
independent.

Ex. 1.4.8. When can an event be independent of itself? Do parts (a) and (b) below to
answer this question.

(a) Prove that if an event A is independent of itself then either P (A) = 0 or P (A) = 1.

(b) Prove that if A is an event such that either P (A) = 0 or P (A) = 1 then A is
independent of itself.

Ex. 1.4.9. This exercise explores the relationship between independence and conditional
probability.

(a) Suppose A and B are independent events with 0 < P (B ) < 1. Prove that P (A |
B ) = P (A) and that P (A | B c ) = P (A).

(b) Suppose that A and B are independent events. Prove that A and B c are also
independent.

(c) Suppose that A and B are events with P (B ) > 0. Prove that if P (A | B ) = P (A),
then A and B are independent.

(d) Suppose that A and B are events with 0 < P (B ) < 1. Prove that if P (A | B ) = P (A),
then P (A | B c ) = P (A) as well.

Ex. 1.4.10. In this section we mentioned the following theorem: “If E1 , E2 , . . . , En is a

collection of mutually independent events, then any subcollection of these events is mutually
independent”. Follow the steps below to prove the theorem.

(a) Suppose A, B, and C are mutually independent. In particular, this means that

P (A ∩ B ∩ C ) = P (A) · P (B ) · P (C ), and
P (A ∩ B ∩ C c ) = P (A) · P (B ) · P (C c ).

Version: – November 19, 2024

 1.5 using r for computation 31

Use these two facts to conclude that A and B are pairwise independent.

(b) Suppose E1 , E2 , . . . , En is a collection of mutually independent events. Prove that

E1 , E2 , . . . , En−1 is also mutually independent.

(c) Use (b) and induction to prove the full theorem.

1.5 using r for computation

As we have already seen, and will see throughout this book, the general approach to solve
problems in probability and statistics is to put them in an abstract mathematical framework.
Many of these problems eventually simplify to computing some specific numbers. Usually
these computations are simple and can be done using a calculator. For some computations
however, a more powerful tool is needed. In this book, we will use a software called R to
illustrate such computations. R is freely available open source software that runs on a
variety of computer platforms, including Windows, macOS, and GNU/Linux.
R is many different things to different people, but for our purposes, it is best to think
of it as a very powerful calculator. Once you install and start R,1 you will be presented
with a prompt that looks like the “greater than” sign (>). You can type expressions that
you want to evaluate here and press the Enter key to obtain the answer. For example,

9 / 44

[1] 0.2045455

0.6 * 0.4 + 0.3 * 0.6

[1] 0.42

log(0.6 * 0.4 + 0.3 * 0.6)

[1] -0.8675006

It may seem odd to see a [1] at the beginning of each answer, but that is there for a
good reason. R is designed for statistical computations, which often require working with
a collection of numbers, which following standard mathematical terminology are referred
to as vectors. For example, we may want to do some computations on a vector consisting
1
Visit [Link] to download R and learn more about it.

Version: – November 19, 2024

32 basic concepts

of the first 5 positive integers. Specifically, suppose we want to compute the squares of
these integers, and then sum them up. Using R, we can do

c(1, 2, 3, 4, 5)ˆ2

[1] 1 4 9 16 25

sum(c(1, 2, 3, 4, 5)ˆ2)

[1] 55

Here the construct c(...) is used to create a vector containing the first five integers. Of
course, doing this manually is difficult for larger vectors, so another useful construct is m:n
which creates a vector containing all integers from m to n. Just as we do in mathematics, it
is also convenient to use symbols (called “variables”) to store intermediate values in long
computations. For example, to do the same operations as above for the first 40 integers,
we can do

x <- 1:40
x

[1] 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
[23] 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40

xˆ2

[1] 1 4 9 16 25 36 49 64 81 100 121 144 169

[14] 196 225 256 289 324 361 400 441 484 529 576 625 676
[27] 729 784 841 900 961 1024 1089 1156 1225 1296 1369 1444 1521
[40] 1600

sum(xˆ2)

[1] 22140

We can now guess the meaning of the number in square brackets at the beginning of
each line in the output: when R prints a vector that spans multiple lines, it prefixes each

Version: – November 19, 2024

 1.5 using r for computation 33

line by the index of the first element printed in that line. The prefix appears for scalars
too because R treats scalars as vectors of length one.
In the example above, we see two kinds of operations. The expression xˆ2 is interpreted
as an element-wise squaring operation, which means that the result will have the same
length as the input. On the other hand, the expression sum(x) takes the elements of a
vector x and computes their sum. The first kind of operation is called a vectorized operation,
and most mathematical operations in R are of this kind.
To see how this can be useful, let us use R to compute factorials and binomial coefficients,
which will turn up frequently in this book. Recall that the binomial coefficient
!
n n!
=
k k!(n − k )!

represents the number of ways of choosing k items out of n, where for any positive integer
m, m! is the product of the first m positive integers. Just as sum(x) computes the sum
! of
10
the elements of x, prod(x) computes their product. So, we can compute 10! and as
4

prod(1:10)

[1] 3628800

prod(1:10) / (prod(1:4) * prod(1:6))

[1] 210

Unfortunately, factorials can quickly become quite big, and may be beyond R’s ability to
compute
! precisely even for moderately large numbers. For example, trying to compute
200
, we get
4

prod(1:200)

[1] Inf

prod(1:200) / (prod(1:4) * prod(1:196))

[1] NaN

Version: – November 19, 2024

34 basic concepts

The first computation yields Inf because at some point in the computation of the prod-
uct, the result becomes larger than the largest number R can store (this is often called
“overflowing”). The second computation essentially reduces to computing Inf/Inf, and the
resulting NaN indicates that the answer is ambiguous. The trivial mathematical fact that
m
log m! = log i
X

i=1

comes to our aid here because it lets us do our computations on much smaller numbers.
Using this, we can compute

logb <- sum(log(1:200)) - sum(log(1:4)) - sum(log(1:196))

logb

[1] 17.98504

exp(logb)

[1] 64684950

R actually has the ability to compute binomial coefficients built into it.

choose(200, 4)

[1] 64684950

These named operations, such as sum(), prod() log(), exp(), and choose(), are known
as functions in R. They are analogous to mathematical functions in the sense that they
map some inputs to an output. Vectorized functions map vectors to vectors, whereas
summary functions like sum() and prod() map vectors to scalars. It is common practice
in R to make functions vectorized whenever possible. For example, the choose() function
is also vectorized:

choose(10, 0:10)

[1] 1 10 45 120 210 252 210 120 45 10 1

choose(10:20, 4)

Version: – November 19, 2024

 1.5 using r for computation 35

[1] 210 330 495 715 1001 1365 1820 2380 3060 3876 4845

choose(2:15, 0:13)

[1] 1 3 6 10 15 21 28 36 45 55 66 78 91 105

A detailed exposition of R is beyond the scope of this book. In this book, we will only use
relatively basic R functions, which we will introduce as and when needed. There are many
excellent introductions available for the interested reader. In particular, R is very useful
for producing statistical plots, and most figures in this book are produced using R. We do
not describe how to create these figures in the book itself, but R code to reproduce them
is available on the website.

exercises

Ex. 1.5.1. In R suppose we type in the following

x <- c(-15, -11, -4, 0, 7, 9, 16, 23)

Find out the output of the built-in functions given below:

sum(x) length(x) mean(x) var(x) sd(x) max(x) min(x) median(x)

Ex. 1.5.2. Obtain a six-sided die, and throw it ten times, keeping a record of the face
that comes up each time. Store these values in a vector variable x. Find the output of the
built-in functions given in the previous exercise when applied to this vector.
Ex. 1.5.3. Use R to verify the calculations done in Example 1.2.4.
Ex. 1.5.4. We return to the Birthday Problem given in Exercise 1.2.12. Using R, calculate
the Probability that at least two from a group of N people share the same birthday, for
N = 10, 12, 17, 26, 34, 40, 41, 45, 75, 105.

Version: – November 19, 2024

36 basic concepts

Version: – November 19, 2024

S A M P L I N G A N D R E P E AT E D T R I A L S
2
Consider an experiment and an event A within the sample space. We say the experiment is
a success if an outcome from A occurs and failure otherwise. Let us consider the following
examples:

Experiment Sample Space Event Description Event A P(A)

Toss a fair coin {H, T } Head appears {H} 1

2

Roll a die {1, 2, 3, 4, 5, 6} Six appears {6} 1

6

Roll a die {1, 2, 3, 4, 5, 6} A multiple of 3 appears {3, 6} 1

3

In typical applications we would repeat an experiment several times independently and

would be interested in the total number of successes achieved, a process that may be viewed
as sampling from a large population. For instance, a manager in a factory making nuts and
bolts, may devise an experiment to choose uniformly from a collection of manufactured
bolts and call the experiment a success if the bolt is not defective. Then she would want
to repeat such a selection every time and quantify the number of successes.

2.1 bernoulli trials

We will now proceed to construct a mathematical framework for independent trials of an

experiment where each trial is either a success or a failure. Let p be the probability of
success at each trial. The sequence so obtained is called a sequence of Bernoulli trials with
parameter p. The trials are named after James Bernoulli (1654-1705).
We will occasionally want to consider a single Bernoulli trial, so we will use the notation
Bernoulli(p) to indicate such a distribution. Since we are only interested in the result of
the trial, we may view this as a probability on the sample space S = {success, f ailure}

37

Version: – November 19, 2024

38 sampling and repeated trials

where P ({success}) = p, but more often we will be interested in multiple independent

trials. We discuss this in the next example.
Example 2.1.1. Suppose we roll a die twice and ask how likely it is that we observe exactly
one 6 between the two rolls. In the previous chapter (See Example 1.4.3) we would have
viewed the sample space S as thirty-six equally likely outcomes, each of which was an
ordered pair of results of the rolls. But since we are only concerned with whether the die
roll is a 6 (success) or not a 6 (failure) we could also view it as two Bernoulli( 16 ) trials.
Using notation from Example 1.4.3, note that P (success on the first roll) = P (E ) = 1
6
and P (success on the second roll) = P (F ) = 16 . So

P ({success, success})
= P (E ∩ F )
(using independence)
= P (E )P (F )
= P (success on the first roll) · P (success on the second roll)
1 1 1
= · = .
6 6 36

We could alternately view S as having only four elements - (success,success), (success,failure),

(failure,success), and (failure,failure). The four outcomes are not equally likely, but the
fact that the trials are independent allows us to easily compute the probability of each.
Through similar computations,

P ({(success, f ailure)}) = 5/36,

P ({(f ailure, success)}) = 5/36,
and P ({(f ailure, f ailure)}) = 25/36.

To complete the problem, the event of rolling exactly one 6 among the two dice requires
exactly one success and exactly one failure. From the list above, this can happen in either
of two orders, so the probability of observing exactly one 6 is 5
36 + 5
36 = 36 .
10
■
For any two real numbers a, b and any integer n ≥ 1, it is well known that
n
!
n k n−k
. (2.1.1)
X
(a + b)n = a b
k =0
k

This is the binomial expansion due to Blaise Pascal(1623-1662). It turns out when a and
b are positive numbers with a + b = 1, the terms in the right hand side above have a
probabilistic interpretation. We illustrate it in the example below.

Version: – November 19, 2024

 2.1 bernoulli trials 39

50

40
Number of successes (cumulative)

30

20

10

0 10 20 30 40 50 0.00 0.04 0.08 0.12

Trial

Figure 2.1: The Binomial distribution as number of successes in fifty Bernoulli ( 13 ) trials. The
paths on the left count the cumulative successes in the fifty trials. The graph on
the right show the actual probability given by the Binomial(50, 31 ) distribution.

Version: – November 19, 2024

40 sampling and repeated trials

Example 2.1.2. After performing n independent Bernoulli(p) trials we are typically

interested in the following questions:

(a) What is the probability of observing exactly k successes?

(b) What is the most likely number of successes?

(c) How many attempts must be made before the first success is observed?

(d) On average how many successes will there be?

Ans (a) - Binomial(n,p): If n = 1, then the answer is clear, namely P ({one success}) = p
and P ({zero successes}) = 1 − p. For, n > 1 let ω = (ω1 , ω2 , . . . , ωn ) be an n-tuple of
outcomes. So we may view the sample space S as the set of all ω where each ωi is allowed
to be either “success” or “failure”. Let Ai represent either the event {the ith trial is a
success} or {the ith trial is a failure}. Then by independence
n
P ( Ai ) . (2.1.2)
Y
P ( A1 ∩ A2 ∩ . . . ∩ An ) =
i=1

Let Bk denote the event that there are k successes among the n trials. Then

P ({ω}).
X
P ( Bk ) =
ω∈Bk

But if ω ∈ Bk , then in notation (2.1.2), exactly k of the Ai represent success trials and
the other n − k represent the failure trials. The order in which the successes and failures
appear does not matter since the probabilities are being multiplied together. So for every
ω ∈ Bk ,
P ({ω}) = pk (1 − p)n−k .

Consequently, we have
P (Bk ) = |Bk |pk (1 − p)n−k .

But Bk is the event of all outcomes for which there are k successes and the number of ways
in which k successes can occur in n trials is known to be (nk). Therefore, for 0 ≤ k ≤ n,
!
n k
P ( Bk ) = p (1 − p)n−k . (2.1.3)
k

Note that if we are only interested in questions involving the number of successes, we could
ignore the set S described above and simply use {0, 1, 2, . . . , n} as our sample space with
P ({k}) = (nk)pk (1 − p)n−k . We call this a Binomial distribution with parameters n and p

Version: – November 19, 2024

 2.1 bernoulli trials 41

(or a Binomial(n, p) for short). It is also worth noting that the binomial expansion (2.1.1)
shows
n
!
n k
p (1 − p)n−k = (p + (1 − p))n = 1,
X

k =0
k

which simply provides additional confirmation that we have accounted for all possible
outcomes in our list of Bernoulli trials. See Figure 2.1 for a simulated example of fifty
replications of Bernoulli( 13 ) trials.

Ans (b) - Mode of a Binomial: The problem is trivial if p = 0 or p = 1, so assume

0 < p < 1. Using the same notation for Bk as in part (a), pick a particular number of
successes k for which 0 ≤ k < n. We want to determine the value of k that makes P (Bk )
as large as possible; such a value is called the “mode” . To find this value, it is instructive
to compare the probability of (k + 1) successes to the probability of k successes –

n k +1 (1 − p)n−(k +1)
P ( Bk + 1 ) (k + 1)p
=
P (Bk ) n k
( k )p (1 − p)n−k
n! k!(n − k )! pk+1 (1 − p)n−(k+1)
= · ·
(k + 1)!(n − (k + 1))! n! pk (1 − p)n−k
p n−k
= · .
1−p k+1

If this ratio were to equal 1 we could conclude that {(k + 1) successes} was exactly
as likely as {k successes}. Similarly if the ratio were bigger than 1 we would know that
{(k + 1) successes} was the more likely of the two and if the ratio were less than 1 we
P ( Bk + 1 )
would see that {k successes} was the more likely case. Setting P ( Bk )
≥ 1 and solving for
k yields the following sequence of equivalent inequalities:

P (Bk+1 )
≥ 1
P ( Bk )
p n−k
· ≥ 1
1−p k+1
pn − pk ≥ k + 1 − pk − p
k ≤ p(n + 1) − 1.

In other words if k starts at 0 and begins to increase, the probability of achieving exactly
k successes will increase while k < p(n + 1) − 1 and then will decrease once k > p(n + 1) − 1.
As a consequence the most likely number of successes is the integer value of k for which

Version: – November 19, 2024

42 sampling and repeated trials

k − 1 ≤ p(n + 1) − 1 < k. This gives the critical value of k = ⌊p(n + 1)⌋, the greatest
integer less than or equal to p(n + 1).
An unusual special case occurs if p(n + 1) is already an integer. Then the sequence of
inequalities above is equality throughout, so if we let k = ⌊p(n + 1)⌋ = p(n + 1) we find a
ratio P (Bk )/P (Bk−1 ) exactly equal to 1. In this case {k − 1 successes} and {k successes}
share the distinction of being equally likely.
Ans (c) - Geometric(p): It is possible we could see the first success as early as the first
trial and, in fact, the probability of this occurring is just p, the probability that the first
trial is a success. The probability of the first success coming on the k th trial requires that
the first k − 1 trials be failures and the k th trial be a success. Let Ai be the event {the ith
trial is a success} and let Ck be the event {the first success occurs on the k th trial}. So,

P (Ck ) = P (Ac1 ∩ Ac2 ∩ . . . ∩ Ack−1 ∩ Ak ).

As usual P (Ai ) = p and P (Aci ) = 1 − p, so by independence

P (Ck ) = P (Ac1 )P (Ac2 ) . . . P (Ack−1 )P (Ak ) = (1 − p)k−1 p

for k > 0. If we view these as probabilities of the outcomes of a sample space {1, 2, 3, . . . },
we call this a geometric distribution with parameter p (or a Geometric(p) for short).
Ans (d) - Average: This is a natural question to ask but it requires a precise definition of
what we mean by “average” in the context of probability. We shall do this in Chapter 4
and return to answer (d) at that point in time.
■

Bernoulli trials may also be used to determine probabilities associated with who will
win a contest that requires a certain number of individual victories. Below is an example
applied to a “best two out of three” situation.

Example 2.1.3. Jed and Sania play a tennis match. The match is won by the first player to
win two sets. Sania is a bit better than Jed and she will win any given set with probability
3. How likely is it that Sania will win the match? (Assume the results of each set are
2

independent).
This can almost be viewed as three Bernoulli( 23 ) trials where we view a success as a set
won by Sania. One problem with that perspective is that an outcome such as (win,win,loss)
never occurs since two wins put an end to the match and the third set will never be played.
Nevertheless, the same tools used to solve the earlier problem can be used for this one as
well. Sania wins the match if she wins the first two sets (which happens with probability

Version: – November 19, 2024

 2.1 bernoulli trials 43

9 ). She also wins the match with either a (win,loss,win) or a (loss,win,win) sequence of
4

sets, each of which has probability 4

27 of occurring.
So the total probability of Sania winning the series is 4
+ 27
9
4 4
+ 27 27 .
= 20
Alternatively, it is possible to view this somewhat artificially as a genuine sequence
of three Bernoulli( 23 ) trials where we pretend the players will play a third set even if the
match is over by then. In effect the (win, win) scenario above is replaced by two different
outcomes - (win, win, win) and (win, win, loss). Sania wins the match if she either wins
all three sets (which has probability 278
) or if she wins exactly two of the three (which has
probability 3 · 27 ).
4

This perspective still leads us to the correct answer as 27 8

+ 3 · 27
4
27 .
= 20 ■

2.1.1 Using R to Compute Probabilities

R can be used to compute probabilities of both the Binomial and Geometric distribution
quite easily. We can compute them directly from the respective formulas. For example,
with n = 10 and p = 0.25, all Binomial probabilities are given by

k <- 0:5
choose(5, k) * 0.25ˆk * 0.75ˆ(5-k)

[1] 0.2373046875 0.3955078125 0.2636718750 0.0878906250 0.0146484375

[6] 0.0009765625

Similarly, the Geometric probabilities with p = 0.25 for k = 0, 1, 2, . . . , 10 are given by

k <- 0:10
0.25 * 0.75ˆk

[1] 0.25000000 0.18750000 0.14062500 0.10546875 0.07910156 0.05932617

[7] 0.04449463 0.03337097 0.02502823 0.01877117 0.01407838

Actually, as both Binomial and Geometric are standard distributions, R has built-in
functions to compute these probabilities. These can be used as follows.

dbinom(0:5, size = 5, prob = 0.25)

[1] 0.2373046875 0.3955078125 0.2636718750 0.0878906250 0.0146484375

[6] 0.0009765625

Version: – November 19, 2024

44 sampling and repeated trials

dgeom(0:10, prob = 0.25)

[1] 0.25000000 0.18750000 0.14062500 0.10546875 0.07910156 0.05932617

[7] 0.04449463 0.03337097 0.02502823 0.01877117 0.01407838

exercises

Ex. 2.1.1. Three dice are rolled. How likely is it that exactly one of the dice shows a 6?
Ex. 2.1.2. A fair die is rolled repeatedly.

(a) What is the probability that the first 6 appears on the fifth roll?

(b) What is the probability that no 6’s appear in the first four rolls?

(c) What is the probability that the second 6 appears on the fifth roll?

Ex. 2.1.3. Suppose that airplane engines operate independently in flight and fail with
probability p (0 ≤ p ≤ 1). A plane makes a safe flight if at least half of its engines
are running. Kingfisher Air lines has a four–engine plane and Paramount Airlines has
a two–engine plane for a flight from Bangalore to Delhi. Which airline has the higher
probability for a successful flight?
Ex. 2.1.4. Two intramural volleyball teams have eight players each. There is a 10% chance
that any given player will not show up to a game, independently of any another. The game
can be played if each team has at least six members show up. How likely is it the game
can be played?
Ex. 2.1.5. Mark is a 70% free throw shooter. Assume each attempted free throw is
independent of every other attempt. If he attempts ten free throws, answer the following
questions.

(a) How likely is it that Mark will make exactly seven of ten attempted free throws?

(b) What is the most likely number of free throws Mark will make?

(c) How do your answers to (a) and (b) change if Mark only attempts 9 free throws
instead of 10?

Ex. 2.1.6. Continuing the previous exercise, Kalyani isn’t as good a free throw shooter as
Mark, but she can still make a shot 40% of the time. Mark and Kalyani play a game where
the first one to sink a free throw is the winner. Since Kalyani isn’t as skilled a player, she
goes first to make it more fair.

Version: – November 19, 2024

 2.1 bernoulli trials 45

(a) How likely is it that Kalyani will win the game on her first shot?

(b) How likely is it that Mark will win this game on his first shot? (Remember, for Mark
even to get a chance to shoot, Kalyani must miss her first shot).

(c) How likely is it that Kalyani will win the game on her second shot?

(d) How likely is it that Kalyani will win the game?

Ex. 2.1.7. Recall from the text above that the R code

dbinom(0:5, size = 5, prob = 0.25)

[1] 0.2373046875 0.3955078125 0.2636718750 0.0878906250 0.0146484375

[6] 0.0009765625

produces a vector of six outputs corresponding to the probabilities that a Binomial(5, 0.25)
distribution takes on the six values 0-5. Specifically, the output indicates that the probability
of the value 0 is approximately 0.2373046875, the probability of the value 1 is approximately
0.3955078125 and so on. In Example 2.1.2 we derived a formula for the most likely outcome
of such a distribution. In the case of a Binomial(5, 0.25) that formula gives the result
⌊(5 + 1)(0.25)⌋ = 1. We could have verified this via the R output above as well, since the
second number on the list is the largest of the probabilities.

(a) Find the most likely outcome of a Binomial(7, 0.34) distribution using the formula
from example 2.1.2.

(b) Type an appropirate command into R to produce a vector of values corresponding to

the probabilities that a Binomial(7, 0.34) distribution takes on the possible values in
its range. Use this list to verify your answer to part (a).

(c) Find the most likely outcome of a Binomial(8, 0.34) distribution using the formula
from Example 2.1.2.

(d) Type an appropirate command into R to produce a vector of values corresponding to

the probabilities that a Binomial(8, 0.34) distribution takes on the possible values in
its range. Use this list to verify your answer to part (c).

Ex. 2.1.8. It is estimated that 0.8% of a large shipment of eggs to a certain supermarket are
cracked. The eggs are packaged in cartons, each with a dozen eggs, with the cracked eggs
being randomly distributed. A restaurant owner buys 10 cartons from the supermarket.
Call a carton “defective” if it contains at least one cracked egg.

Version: – November 19, 2024

46 sampling and repeated trials

(a) If she notes the number of defective cartons, what are the possible outcomes for this
experiment?

(b) If she notes the total number of cracked eggs, what are the possible outcomes for
this experiment?

(c) How likely is it that she will find exactly one cracked egg among all of her cartons?

(d) How likely is it that she will find exactly one defective carton?

(e) Explain why your answer to (d) is close to, but slightly larger than, than your answer
to (c).

(f) What is the most likely number of cracked eggs she will find among her cartons?

(g) What is the most likely number of defective cartons she will find?

(h) How do you reconcile your answers to parts (g) and (h)?

Ex. 2.1.9. Steve and Siva enter a bar with $30 each. A round of drinks cost $10. For each
round, they roll a die. If the roll is even, Steve pays for the round and if the roll is odd,
Siva pays for it. This continues until one of them runs out of money.

(a) What is the Probability that Siva runs out of money?

(b) What is the Probability that Siva runs out of money if Steve has cheated by bringing
a die that comes up even only 40% of the time?

Ex. 2.1.10. Let 0 < p < 1. Show that the mode of a Geometric(p) distribution is 1.
Ex. 2.1.11. Scott is playing a game where he rolls a standard die until it shows a 6. The
number of rolls needed therefore has a Geometric( 16 ) distribution. Use the appropriate R
commands to do the following:

(a) Produce a vector of values for j = 1, . . . , 6 corresponding to the probabilities that it

will take Scott j rolls before he observes a 6.

(b) Scott figures that since each roll has a 1

6 probability of producing a 6, he’s bound
to get that result at some point after six rolls. Use the results from part (a) to
determine the probability that Scott’s expectations are met and a 6 will show up in
one his first six rolls.

Ex. 2.1.12. Suppose a fair coin is tossed n times. Compute the following:

(a) P ({4 heads occur }|{3 or 4 heads occur});

Version: – November 19, 2024

 2.1 bernoulli trials 47

(b) P ({k − 1 heads occur}|{k − 1 or k heads occur}); and

(c) P ({k heads occur}|{k − 1 or k heads occur}).

Ex. 2.1.13. At a basketball tournament, each round is on a “best of seven games” basis.
That is, Team I and Team 2 play until one of the teams has won four games. Suppose
each game is won by Team I with probability p, independently of all previous games. Are
the events A = {Team I wins the round} and B = {the round lasts exactly four games}
independent?
Ex. 2.1.14. Two coins are sitting on a table. One is fair and the other is weighted so that
it always comes up heads.

(a) If one coin is selected at random (each equally likely) and flipped, what is the
probability the result is heads?

(b) One coin is selected at random (each equally likely) and flipped five times. Each
flip shows heads. Given this information about the coin flip results, what is the
conditional probability that the selected coin was the fair one?

Ex. 2.1.15. For 0 < p < 1 we defined the geometric distribution as a probability on the set
{1, 2, 3, . . . } for which P ({k}) = p(1 − p)k−1 . Show that these outcomes account for all
P∞
possibilities by demonstrating that k =1 P ({k}) = 1.
Ex. 2.1.16. The geometric distribution described the waiting time to observe a single
success. A “Negative Binomial” distribution with parameters n and p (NegBinomial(n, p))
is defined the number of Bernoulli(p) trials needed before observing n successes. The
following problem builds toward calculating some associated probabilities.

(a) If a fair die is rolled repeatedly and a number is recorded equal to the number of
rolls until the second 6 is observed, what is the sample space of possible outcomes
for this experiment?

(b) For k in the sample space you identified in part (a), what is P ({k})?

(c) If a fair die is rolled repeatedly and a number is recorded equal to the number of
rolls until the nth 6 is observed, what is the sample space of possible outcomes for
this experiment?

(d) For k in the sample space you identified in part (c), what is P ({k})?

(e) If a sequence of Bernoulli(p) trials (with 0 < p < 1) is performed and a number is
recorded equal to the number of trials until the nth success is observed, what is the
sample space of possible outcomes for this experiment?

Version: – November 19, 2024

48 sampling and repeated trials

(f) For k in the sample space you identified in part (e), what is P ({k})?

(g) Show that you have accounted for all possibilities in part (f) by showing

P ({k}) = 1.
X

k∈S

2.2 poisson approximation

Calculating Binomial probabilities can be challenging when n is large. Let us consider the
following example:

Example 2.2.1. A small college has 1460 students. Assume that birthrates are constant
throughout the year and that each year has 365 days. What is the probability that five or
more students were born on Independence day?

The probability that any given student was born on Independence day is 365 .
1
So the
exact probability is
4
1460  1 k  364 1460−k
!
1− .
X

k =0
k 365 365

Repeatedly dealing with large powers of fractions or large combinatorial computations is

not so easy, so it would be convenient to find a faster way to estimate such a probability. ■

The example above can be thought of as a series of Bernoulli trials where a success
means finding a student whose birthday is Independence day. In this case p is small ( 365
1
)
and n is large (1460). To approximate we will consider a limiting procedure where p → 0
and n → ∞, but with limits carried out in such a way that np is held constant. The
computation below is called a Poisson approximation.

Theorem 2.2.2. Let λ > 0, k ≥ 1, n ≥ λ and p = nλ . Defining Ak as

Ak = {k successes in n Bernoulli(p) Trials},

it then follows that

e−λ λk
lim P (Ak ) = . (2.2.1)
n→∞ k!

Version: – November 19, 2024

 2.2 poisson approximation 49

Proof -
!  
k n−k
n λ λ
P (Ak ) = 1−
k n n
n ( n − 1 ) . . . ( n − k + 1 ) λk λ n−k
 
= 1 −
k! nk n
λ n(n − 1) . . . (n − k + 1)
k λ n λ −k
   
= 1 − 1 −
k! nk n n
1 k−1 λ −k
k n 
λ λ
 
= 1(1 − ) . . . (1 − ) 1− 1−
k! n n n n
λk k−1
Y r λ n λ −k
   
= 1− 1− 1− .
k! r =1 n n n

Standard limit results imply that

r
lim (1 − )=1 for all r ≥ 1;
n→∞ n
−k
λ

lim 1−=1 for all λ ≥ 0, k ≥ 1; and
n→∞ n
λ n
 
lim 1 − = e−λ for all λ ≥ 0.
n→∞ n

As P (Ak ) is a finite product of such expressions, the result is now immediate using the
properties of limits. ■
Returning to Example 2.2.1 and using the above approximation, we would take λ =
pn = 1460
365 = 4. So if E is the event {five or more Independence day birthdays},

4
1460  1 k  364 1460−k
!
P (E ) = 1 −
X

k =0
k 365 365
42 43 44
" #
−4 −4
≈ 1− e + 4e + e−4 + e−4 + e−4 .
2 6 24

Calculation demonstrates this is a good approximation. To seven digits of accuracy, the

correct value is 0.37116294 while the Poisson approximation gives an answer of 0.37116306.
These can be obtained using R as follows:

1 - sum(dbinom(0:4, size = 1460, prob = 1/365))

Version: – November 19, 2024

50 sampling and repeated trials

[1] 0.3711629

lambda <- 1460 / 365

1 - sum(exp(-lambda) * lambdaˆ(0:4) / factorial(0:4))

[1] 0.3711631

It also turns out that the right hand side of (2.2.1) defines a probability on the sample space
of non-negative integers. The distribution is named after Siméon Poisson (1781–1840).
Poisson (λ): Let λ ≥ 0 and S = {0, 1, 2, 3, . . .} with probability P given by

e−λ λk
P ({k}) =
k!

for k ∈ S. This distribution is called Poisson with parameter λ (or Poisson(λ) for short).
As with Binomial and Geometric, R has a built-in function to evaluate Poisson proba-
bilities as well. An alternative to the calculation above is the following.

1 - sum(dpois(0:4, lambda = 1460 / 365))

[1] 0.3711631

It is important to note that for this approximation to work well, p must be small and n
must be large. For example, we may modify our question as follows:

Example 2.2.3. A class has 48 students. Assume that birthrates are constant throughout
the year and that each year has 365 days. What is the probability that five or more
students were born in September? ■
The correct answer to this question is

1 - sum(dbinom(0:4, size = 48, prob = 1/12))

[1] 0.3710398

However, the Poisson approximation remains unchanged at 0.3711631, because np =

48/12 = 1460/365 = 4, and only matches the correct answer up to 3 digits rather than
6. Figure 2.2 shows a point-by-point approximation of both Binomial distributions by
Poisson.
At this point we have defined many named distributions. Frequently a problem will
require the use of more than one of these as evidenced in the next example.

Version: – November 19, 2024

 2.2 poisson approximation 51

Example 2.2.4. A computer transmits three digital messages of 12 million bits of infor-
mation each. Each bit has a probability of one one-billionth that it will be incorrectly
received, independent of all other bits. What is the probability that at least two of the of
the three messages will be received error free?
Since n = 12, 000, 000 is large and since p = 1
1,000,000,000 is small it is appropriate to
use a Poisson approximation where λ = np = 0.012. A message is error free if there isn’t
a single misread bit, so the probability that a given message will be received without an
error is e−0.012 .
Now we can think of each message being like a Bernoulli trial with probability e−0.012 ,
so the number of messages correctly received is then like a Binomial (3, e−0.012 ). Therefore
the probability of receiving at least two error-free messages is

3 3
! !
(e−0.012 )3 (1 − e−0.012 )0 + (e−0.012 )2 (1 − e−0.012 )1 ≈ 0.9996.
3 2

There is about a 99.96% chance that at least two of the messages will be correctly
received. ■

exercises

Ex. 2.2.1. Do the problems below to familiarize yourself with the “sum” command in R.

(a) If a fair coin is tossed 100 times, what is the probability exactly 55 of the tosses show
heads?

(b) Example 2.2.3 showed how to use R to add the probabilities of a range of outcomes
for common distributions. Use this code as a guide to calculate the probability at
least 55 tosses show heads.

Ex. 2.2.2. Consider an experiment described by a Poisson( 12 ) distribution and answer the
following questions.

(a) What is the probability the experiment will produce a result of 0?

(b) What is the probability the experiment will produce a result larger than 1?

Ex. 2.2.3. Suppose we perform 500 independent trials with probability of success being
0.02.

(a) Use R to compute the probability that there are six or fewer successes. Obtain a
decimal approximation accurate to five decimal places.

Version: – November 19, 2024

52 sampling and repeated trials

0.20

0.15
Probability

0.10

0.05

0.00

0 5 10 15 20

0.20

0.15
Probability

0.10

0.05

0.00

0 5 10 15 20

Figure 2.2: The Poisson approximation to the Binomial distribution. In both plots above,
the points indicate Binomial probabilities for k = 0, 1, 2, . . . , 20; the top plot for
Binomial(1460, 365 1
), and the bottom for Binomial(48, 12 1
). The lengths of the
vertical lines, “hanging” from the points, represent the corresponding probabilities
for Poisson(4). For a good approximation, the bottom of the hanging lines should
end up at the x-axis. As we can see, this happens in the top plot but not for
the bottom plot, indicating that Poisson(4) is a good approximation for the first
Binomial distribution, but not as good for the second.

Version: – November 19, 2024

 2.2 poisson approximation 53

(b) Use the Poisson approximation to estimate the probability that there are six or fewer
successes and compare it to your answer to (a).

Now suppose we perform 5000 independent trials with probability of success being
0.002.

(c) Use R to compute the probability that there are six or fewer successes. Obtain a
decimal approximation accurate to five decimal places.

(d) Use the Poisson approximation to estimate the probability that there are six or fewer
successes and compare it to your answer to (c).

(e) Which approximation (b) or (d) is more accurate? Why?

Ex. 2.2.4. For a certain daily lottery, the probability is 1

10000 that you will win. Suppose
you play this lottery every day for three years. Use the Poisson approximation to estimate
the chance that you will win more than once.
Ex. 2.2.5. A book has 200 pages. The number of mistakes on each page has a Poisson(1)
distribution, and is independent of the number of mistakes on all other pages.

(a) What is the chance that there are at least 2 mistakes on the first page?

(b) What is the chance that at least eight of the first ten pages are free of mistakes?

Ex. 2.2.6. Let λ > 0. For the problems below, assume the probability space is a Poisson(λ)
distribution.
P ({k +1})
(a) Let k be a non-negative integer. Calculate the ratio P ({k})
.

(b) Use (a) to calculate the mode of a Poisson(λ).

Ex. 2.2.7. A number is to be produced as follows. A fair coin is tossed. If the coin comes
up heads the number will be the outcome of an experiment corresponding to a Poisson(1)
distribution. If the coin comes up tails the number will be the outcome of an experiment
corresponding to a Poisson(2) distribution. Given that the number produced was a 2,
determine the conditional probability that the coin came up heads.
Ex. 2.2.8. Suppose that the number of earthquakes that occur in a year in California has
a Poisson distribution with parameter λ. Suppose that the probability that any given
earthquake has magnitude at least 6 on the Richter scale is p.

(a) Given that there are exactly n earthquakes in a year, find an expression (in terms of
n and p) for the conditional probability that exactly one of them is magnitude at
least 6.

Version: – November 19, 2024

54 sampling and repeated trials

(b) Find an expression (in terms of λ and p) for the probability that there will be exactly
one earthquake of magnitude at least 6 in a year.

(c) Find an expression (in terms of n, λ, and p) for the probability that there will be
exactly k earthquakes of magnitude at least 6 in a year.

Ex. 2.2.9. We defined a Poisson distribution as a probability on S = {0, 1, 2, . . . } for which

e−λ λk
P ({k}) = ,
k!

for k ≥ 1. Prove that this completely accounts for all possibilities by proving that
∞ −λ k
e λ
= 1.
X

k =0
k!

(Hint: Consider the power series expansion of the exponential function).

Ex. 2.2.10. Consider n vertices labeled {1, 2, . . . , n}. Corresponding to each distinct pair
{i, j} we perform an independent Bernoulli (p) experiment and insert an edge between i
and j with probability p. The graph constructed this way is denoted as G(n, p).

(a) Let 1 ≤ i ≤ n. We say j is a neighbour of i if there is an edge between i and j. For

some 1 ≤ k ≤ n determine the probability that i has k neighbours ?

(b) Let λ > 0 and n large enough so that 0 < p = λ

n < 1 and let Ak = { vertex 1 has k neighbours}
what is the
lim P (Ak )?
n→∞

2.3 sampling with and without replacement

Imagine a small town with 5000 residents, exactly 1000 of whom are under the age of
eighteen. Suppose we randomly select four of these residents and ask how many of the
four are under the age of eighteen. There is some ambiguity in how to interpret this idea
of selecting four residents. One possibility is “sampling with replacement” where each
selection could be any of the 5000 residents and the selections are all genuinely independent.
With this interpretation, the sample is simply a series of four independent Bernoulli( 15 )
trials, in which case the answer may be found using techniques from the previous sections.
Note, however, that the assumption of independence allows for the possibility that the
same individual will be chosen two or more times in separate trials. This is a situation that
might seem peculiar when we think about choosing four people from a population of 5000,

Version: – November 19, 2024

 2.3 sampling with and without replacement 55

since we may not have four different individuals at the end of the process. To eliminate
this possibility consider “sampling without replacement” where it is assumed that if an
individual is chosen for inclusion in the sample, that person is no longer available to be
picked in a later selection. Equivalently we can consider all possible groups of four which
might be selected and view each grouping as equally likely. This change means the problem
can no longer be solved by viewing the situation as a series of independent Bernoulli trials.
Nevertheless, other tools that have been previously developed will serve to answer this new
problem.

Example 2.3.1. For the town described above, what is the probability that, of four
residents randomly selected (without replacement), exactly two of them will be under the
age of eighteen?
Since we are selecting four residents from the town of 5000, there are (5000
4 ) ways this
may be done. If each of these is equally likely, the desired probability may be calculated
by determining how many of these selections result in exactly two people under the age of
eighteen. This requires selecting two of the 1000 who are in that younger age group and
also selecting two of the 4000 who are older. So there are (1000
2 )( 2 ) ways to make such
4000

choices and therefore the probability of selecting exactly two residents under age eighteen
is (1000
2 )( 2 ) / ( 4 ).
4000 5000

It is instructive to compare this to the solution if it is assumed the selection is done with
replacement. In that case, the answer is the simply the probability that a Binomial(4, 15 )
produces a result of two. From the previous sections, the answer is (42)( 15 )2 ( 45 )2 .
To compare these answers we give decimal approximations of both. To six digits of
accuracy

1000 4000
! !

2 2 4 1 2 4 2
!   
≈ 0.153592 and = 0.1536,
5000 2 5 5
!

so while the two answers are not equal, they are very close. This is a reflection of an
important fact in statistical analysis—when samples are small relative to the size of the
populations they came from, the two methods of sampling give very similar results. ■

2.3.1 The Hypergeometric Distribution

Analyzing such problems more generally, consider a population of N people. Suppose

r of these N share a common characteristic and the remaining N − r do not have this

Version: – November 19, 2024

56 sampling and repeated trials

characteristic. We take a sample of size m (without replacement) from the population

and count the number of people among the sample that have the specified characteristic.
This experiment is described by probabilities known as a hypergeometric distribution.
Notice that the largest possible result is min{m, r} since the number cannot be larger than
the size of the sample nor can it be larger than the number of people in the population
with the characteristic. On the other extreme, it may be that the sample is so large it is
guaranteed to select some people with the characteristic simply because the number of
people without has been exhausted. More precisely, for every selection over N − r in the
sample we are guaranteed to select at least one person who has the characteristic. So the
minimum possible result is the larger of 0 or (m − (N − r )).
HyperGeo(N , r, m): Let r and m be non-negative integers and let N be an integer
with N > max{r, m}. Let S be the set of integers ranging from max{0, m − (N − r )} to
min{m, r} inclusive with probability P given by
! !
r N −r
k m−k
P ({k}) = !
N
m

for k ∈ S. Such a distribution is called hypergeometric with parameters N , r, and m (or

HyperGeo(N , r, m)).
Of course, R can be used to compute hypergeometric probabilities as well. Example
2.3.1 can be phrased in terms of a HyperGeo(5000, 1000, 4) distribution, with P ({k}) being
the desired answer. This probability can be computed as:

dhyper(2, 4000, 1000, 4)

[1] 0.1535923

Note however, that instead of N , the parameter used by R is N − r.

2.3.2 Hypergeometric Distribution as a Series of Dependent Trials

It is also possible (and useful) to view sampling without replacement as a series of dependent
Bernoulli trials for which each trial reduces the possible outcomes of subsequent trials. In
this case each trial is described in terms of conditional probabilities based on the results of
the preceding observations. We illustrate this by revisiting the previous example.

Version: – November 19, 2024

 2.3 sampling with and without replacement 57

Example 2.3.1 Continued: We first solved this problem by considering every group of four
as equally likely to be selected. Now consider the sampling procedure as a series of four
separate Bernoulli trials where a success corresponds to the selection of a person under
eighteen and a failure as the selection of someone older. We still want to determine the
probability that a sample of size four will produce exactly two successes. One complication
with this perspective is that the successes and failures could come in many different orders,
so first consider the event where the series of selections follow the pattern “success-success-
failure-failure”. More precisely, for j = 1, 2, 3, 4 let

Aj = {The j th selection is a person younger than eighteen}.

Clearly P (A1 ) = 5000 .

1000
Given that the first selection is someone under eighteen, there are
now only 4999 people remaining to choose among, and only 999 of them are under eighteen.
Therefore P (A2 |A1 ) = 4999 .
999
Continuing with that same reasoning,

4000
P (Ac3 |A1 ∩ A2 ) =
4998

and
3999
P (Ac4 |A1 ∩ A2 ∩ Ac3 ) = .
4997
From those values, Theorem 1.3.8 may be used to calculate

P (success − success − f ailure − f ailure) = P (A1 ∩ A2 ∩ Ac3 ∩ Ac4 )

1000 999 4000 3999
= · · · .
5000 4999 4998 4997

Next we must account for the fact that this figure only considers the case where the two
younger people were chosen as the first two selections. There are (42) different orderings
that result in two younger and two older people, and it happens that each of these has the
same probability calculated above. For example,

P (f ailure − success − success − f ailure) = P (Ac1 ∩ A2 ∩ A3 ∩ Ac4 )

4000 1000 999 3999
= · · · .
5000 4999 4998 4997

The individual fractions are different, but their product is the same. This will always
happen for different orderings of a specific number of successes since the denominators
(5000 through 4997) reflect the steady reduction of one available choice with each additional
selection. Similarly the numerators (1000 and 999 together with 4000 and 3999) reflect the
number of people available from each of the two different categories and their reduction

Version: – November 19, 2024

58 sampling and repeated trials

as previous choices eliminate possible candidates. Therefore the total probability is the
product of the number of orderings and the probability of each ordering.

4 4000 1000 999 3999

!
P (two under eighteen) = · · · · .
2 5000 4999 4998 4997

We leave it to the reader to verify that this is equal to (1000

2 )( 2 ) / ( 4 ), the answer we
4000 5000

found when we originally solved the problem via a different method.

The following theorem generalizes this previous example.

Theorem 2.3.2. Let S be a sample space with a hypergeometric distribution with

parameters N , r, and m. Then P ({k}) equals

r r−1 r − (k − 1) N −r N −r−1 N − r − (m − 1 − k )
!
m
 
... ...
k N N −1 N − (k − 1) N −kN −k−1 N − (m − 1)

for any k ∈ S.

Proof. Following the previous example as a model, this can be proven by viewing the
hypergeometric distribution as a series of dependent trials. The first k fractions are the
probabilities the first k trials each result in successes conditioned on the successes of
the preceding trials. The remaining m − k fractions are the conditional probabilities the
remaining trials result in failures. The leading factor of (m
k ) accounts for the number of
different patterns of k successes and m − k failures, each of which is equally likely. It is
also possible to prove the equality directly using combinatorial identities and we leave this
as Exercise 2.3.4. ■

2.3.3 Binomial Approximation to the Hypergeometric Distribution

We saw with Example 2.3.1 that sampling with and without replacement may give very
similar results. The following theorem makes a precise statement to this effect.

Version: – November 19, 2024

 2.3 sampling with and without replacement 59

Theorem 2.3.3. Let N , m, and r be positive integers for which m < r < N and
let k be a positive integer between 0 and m. Define

r r−k r−k
p= , p1 = , and p2 = .
N N −k N −m

Letting H denote the probability that a hypergeometric distribution with parameters

N , r, and m takes on the value k, the following inequalities give bounds on this
probability: ! !
m k m k
p (1 − p2 )m−k < H ≤ p (1 − p1 )m−k .
k 1 k

Proof- The inequalities may be verified by comparing p, p1 , and p2 to the fractions from
Theorem 2.3.2. Specifically note that the k fractions

r r−1 r − (k − 1)
, ,...,
N N −1 N − (k − 1)

are all less than or equal to p. Likewise the m − k fractions

N −r N −r−1 N − r − (m − 1 − k )
, ,...,
N −k N −k−1 N − (m − 1)

are all less than or equal to N −r

N −k which itself equals 1 − p1 . Combining these facts proves
the right hand inequality. The left hand inequality may be similarly shown by noting that
the fractions
r r−1 r − (k − 1)
, ,...,
N N −1 N − (k − 1)
are all greater than p1 while the fractions

N −r N −r−1 N − r − (m − 1 − k )
, ,...,
N −k N −k−1 N − (m − 1)

N −r−(m−k )
all exceed N −m which equals 1 − p2 . ■

When m is small relative to r and N , both fractions p1 and p2 are approximately

equal to p. So this theorem justifies the earlier statement that sampling with and without
replacement yield similar results when samples are small relative to the populations from
which they were derived.

Version: – November 19, 2024

60 sampling and repeated trials

exercises

Ex. 2.3.1. Suppose there are thirty balls in an urn, ten of which are black and the
remaining twenty of which are red. Suppose three balls are selected from the urn (without
replacement).

(a) What is the probability that the sequence of draws is red-red-black?

(b) What is the probability that the three draws result in exactly two red balls?

Ex. 2.3.2. This exercise explores how to use R to investigate the Binomial approximation
to the Hypergeometric distribution.

(a) A jar contains forty marbles – thirty white and ten black. Ten marbles are drawn
at random from the jar. Use R to calculate the probability that exactly five of the
marbles drawn are black. Do two separate computations, one under the assumption
that the draws are with replacement and the other under the assumption that the
draws are without replacement.

(b) Repeat part (a) except now assume the jar contains 400 marbles – 300 wihite and
100 black.

(c) Repeat part (a) excpet now assume the jar contains 4000 marbles – 3000 white and
1000 black.

(d) Explain what you are observing with your results of parts (a), (b), and (c).

Ex. 2.3.3. Consider a room of one hundred people – forty men and sixty women.

(a) If ten people are selected from the room, find the probability that exactly six are
women. Calculate this probability with and without replacement and compare the
decimal approximations of your two results.

(b) If ten people are selected from the room, find the probability that exactly seven are
women. Calculate this probability with and without replacement and compare the
decimal approximations of your two results.

(c) If 100 people are selected from the room, find the probability that exactly sixty are
women. Calculate this probability with and without replacement and compare the
two answers.

(d) If 100 people are selected from the room, find the probability that exactly sixty-one
are women. Calculate this probability with and without replacement and compare
the two answers.

Version: – November 19, 2024

 2.3 sampling with and without replacement 61

Ex. 2.3.4. Use the steps below to prove Theorem 2.3.2

r!(N −k )!
(a) Prove that N !(r−k )!
equals

r r−1 r − (k − 1)
· ... .
N N −1 N − (k − 1)

(N −r )!(N −m)!
(b) Prove that (N −k )!((N −r−(m−k ))! equals

N −r N −r−1 N − r − (m − 1 − k )
· ... .
N −k N −k−1 N − (m − 1)

(c) Use (a) and (b) to prove Theorem 2.3.2.

Ex. 2.3.5. A box contains W white balls and B black balls. A sample of n balls is drawn
at random for some n ≤ min(W , B ). For j = 1, 2, · · · , n, let Aj denote the event that the
ball drawn on the j th draw is white. Let Bk denote the event that the sample of n balls
contains exactly k white balls.

(a) Find P (Aj |Bk ) if the sample is drawn with replacement.

(b) Find P (Aj |Bk ) if the sample is drawn without replacement.

Ex. 2.3.6. For the problems below, assume a HyperGeo(N , r, m) distribution.

P ({k +1})
(a) Calculate the ratio P ({k})
.
(Assume that max{0, m − (N − r )} ≤ k ≤ min{r, m} to avoid zero in the denomina-
tor).

(b) Use (a) to calculate the mode of a HyperGeo(N , r, m).

Ex. 2.3.7. Biologists use a technique called “capture-recapture” to estimate the size of the
population of a species that cannot be directly counted. The following exercise illustrates
the role a hypergeometric distribution plays in such an estimate.
Suppose there is a species of unknown population size N . Suppose fifty members of
the species are selected and given an identifying mark. Sometime later a sample of size
twenty is taken from the population and it is found that four of the twenty were previously
marked. The basic idea behind mark-recapture is that since the sample showed 4
= 20%
20
marked members, that should also be a good estimate for the fraction of marked members
of the species as a whole. However, for the whole species that fraction is 50
N which provides
a population estimate of N ≈ 250.

Version: – November 19, 2024

62 sampling and repeated trials

Looking more deeply at the problem, if the second sample is assumed to be done at
random without replacement and with each member of the population equally likely to be
selected, the resulting number of marked members should follow a HyperGeo(N , 50, 20)
distribution.
Under these assumptions use the formula for the mode calculated in the previous
exercise to determine which values of N would cause a result of four marked members to
be the most likely of the possible outcomes.
Ex. 2.3.8. The geometric distribution was first developed to determine the number
of independent Bernoulli trials needed to observe the first success. When viewing the
hypergeometric distribution as a series of dependent trials, the same question may be
asked. Suppose we have a population of N people for which r have a certain characteristic
and the remaining N − r do not have that characteristic. Suppose an experiment consists
of sampling (without replacement) repeatedly and recording the number of the sample
that first corresponds to selecting someone with the specified characteristic. Answer the
questions below.

(a) What is S, the list of possible outcomes of this experiment?

(b) For each k ∈ S, what is P ({k})?

r−(k−1)
(c) Define p = r
N and p1 = N −(k−1)
. Using the result from (b) prove the following
bounds on the probability distribution:

p(1 − p1 )k−1 ≤ P ({k}) ≤ p1 (1 − p)k−1

(As a consequence, when k is much smaller than r and N , the values of p1 and p are
approximately equal and the probabilities from (b) are closely approximated by a geometric
distribution).

Version: – November 19, 2024

D I S C R E T E R A N D O M VA R I A B L E S
3
In the previous chapter many different distributions were developed out of Bernoulli trials.
In that chapter we proceeded by creating new sample spaces for each new distribution, but
when faced with many questions related to the same basic framework, it is usually clearer to
maintain a single sample space and to define functions on that space whose outputs relate
them to questions under consideration. Such functions are known as “random variables”
and they will be the focus of this chapter.

3.1 random variables as functions

Example 3.1.1. Suppose a coin is flipped three times. Consider the probabilities associated
with the following two questions:

(a) How many coins will come up heads?

(b) Which will be the first flip (if any) that shows heads?

At this point the answers to these questions should be easy to determine, but the purpose
of this example is to emphasize how functions could be used to answer both questions
within the context of a single sample space. Let S be a listing of all eight possible orderings
of heads and tails on the three flips, so that S = {hhh, hht, hth, htt, thh, tht, tth, ttt}. Now
define two functions on S. Let X be the function that describes the total number of heads
among the three flips and let Y be the function that describes the first flip that produces
heads. Then X and Y are given by the table

ω X (ω ) Y (ω )
hhh 3 1
hht 2 1
hth 2 1
htt 1 1
thh 2 2
tht 1 2
tth 1 3
ttt 0 none

where Y (ttt) is defined as “none” as there is no first time the coin produces heads.

63

Version: – November 19, 2024

64 discrete random variables

Suppose we want to know the probability that exactly two of the three coins will be
heads. The relevant event is E = {hht, hth, thh}, but in the pre-image notation of function
theory this set may also be described as X −1 ({2}), the elements of S for which X produces
an output of 2. This allows us to describe the probability of the event as:

3
P (two heads) = P (X −1 ({2})) = P ({hht, hth, thh}) = .
8

Rather than use the standard pre-image notation, it is more common in probability to
write (X = 2) for the set X −1 ({2}) as this emphasizes that we are considering outcomes
for which the function X equals 2.
Similarly, if we wanted to know the probability that the first result of heads showed
up on the third flip, that is a question that involves the function Y . Using the notation
(Y = 3) in place of Y −1 ({3}) the probability may be calculated as

1
P (first heads on flip three) = P (Y = 3) = P ({tth}) = .
8

As above we can compute the

1 3 1
P (X = 0) = , P (X = 1) = , and P (X = 3) =
8 8 8

and
1 1 1
P (Y = 1) = , P (Y = 2) = , and P (Y = none) = ,
2 4 8
thus giving a complete description of how X and Y distribute the probabilities onto their
range. For both cases only a single sample space was needed. Two different questions were
approached by defining two different functions on that sample space. ■

The following theorem explains how the mechanism of the previous example may be more
generally applied.

Theorem 3.1.2. Let S be a sample space with probability P and let X : S → T be

a function. Then X generates a probability Q on T given by

Q(B ) = P (X −1 (B )).

The probability Q is called the “distribution of X” as it describes how X distributes the

probability from S onto T . The proof relies on two set-theoretic facts that we will take as
∞ ∞
given. The first is that X −1 ( X −1 (Bi ) and the second is the fact that if Bi
S S
Bi ) =
i=1 i=1
and Bj are disjoint, then so are X −1 (Bi ) and X −1 (Bj ).

Version: – November 19, 2024

 3.1 random variables as functions 65

Proof. Let B ⊂ T . As P is known to be a probability, 0 ≤ P (X −1 (B )) ≤ 1, and so Q

maps subsets of T into [0, 1]. As X is a function into T , we know X −1 (T ) = S. Therefore
Q(T ) = P (X −1 (T )) = P (S ) = 1 and Q satisfies the first probability axiom.
To show that Q satisfies the second axiom, suppose B1 , B2 , . . . are a countable collection
of disjoint subsets of T .
∞ ∞
Bi ) = P (X −1 (
[ [
Q( Bi ))
i=1 i=1
∞
X −1 (Bi ))
[
= P(
i=1
∞
P (X −1 (Bi ))
X
=
i=1
∞
Q ( Bi ) .
X
=
i=1
■

As in the previous example, it is typical to write (X ∈ B ) in place of the notation

X −1 (B ) to emphasize the fact that we are computing the probability that the function
X takes a value in the set B. In practice, the new probability Q would rarely be used
explicity, but would be calculated in terms of the original probability P via the relationship
described in the theorem.

Example 3.1.3. A board game has a wheel that is to be spun periodically. The wheel can
stop in one of ten equally likely spots. Four of these spots are red, three are blue, two are
green, and one is black. Let X denote the color of the spot. Determine the distribution of
X.
The function X is defined on a sample space S that consists of the ten spots the
wheel could stop, and it takes values on the set of colors T = {red, blue, green, black}. Its
distribution is a probability Q on the set of colors which can be determined by calculating
the probability of each color.
For instance Q({red}) = P (X = red) = P (X −1 ({red})) = 4
10 as four of the ten spots
on the wheel are red and all spots are equally likely. Similarly,

3
Q({blue}) = P (X = blue) =
10
2
Q({green}) = P (X = green) =
10
1
Q({black}) = P (X = black ) =
10

completing the description of the distribution. ■

Version: – November 19, 2024

66 discrete random variables

Example 3.1.4. For a certain lottery, a three-digit number is randomly selected (from 000
to 999). If a ticket matches the number exactly, it is worth $200. If the ticket matches
exactly two of the three digits, it is worth $20. Otherwise it is worth nothing. Let X be
the value of the ticket. Find the distribution of X.
The function X is defined on S = {000, 001, . . . , 998, 999} - the set of all one thousand
possible three digit numbers. The function takes values on the set {0, 20, 200}, so the
distribution Q is a probability on T = {0, 20, 200}.
First, Q({200}) = P (X = 200) = 1
1000 as only one of the one thousand three digit
numbers is going to be an exact match.
Next, Q({20}) = P (X = 20), so it must be determined how many of the one thousand
possibilities will have exactly two matches. There are (32) = 3 different ways to choose the
two digits that will match. Those digits are determined at that point and the remaining
digit must be one of the nine digits that do not match the third spot, so there are 3 · 9 = 27
three digit numbers that match exactly two digits. So Q({20}) = P (X = 20) = 1000 .
27

Finally, as Q is a probability, Q({0}) = 1 − Q({20}) − Q({200}) = 1000 .

972
■
It is frequently the case that we are interested in functions that have real-valued outputs
and we reserve the term “random variable” for such a situation.

Definition 3.1.5. A “discrete random variable” is a function X : S → T where

S is a sample space equipped with a probability P , and T is a countable (or finite)
subset of the real numbers.
From Theorem 3.1.2, P generates a probability on T . As it is a discrete space, the
distribution may be determined by knowing the likelihood of each possible value of X.
Because of this we define a function fX : T → [0, 1] given by

fX (t) = P (X = t)

referred to as a “probability mass function”. Then for any event A ⊂ T the quantity
P (X ∈ A) may be computed via

P (X = t).
X X
P (X ∈ A) = fX (t) =
t∈A t∈A

The function from Example 3.1.4 is a discrete random variable because it takes on one of
the real values 0, 20, or 200. We calculated its probability mass function when describing
its distribution and it is given by

972 27 1
fX (0) = , fX (20) = , fX (200) = .
1000 1000 1000

Version: – November 19, 2024

 3.1 random variables as functions 67

The function from Example 3.1.3 is not a discrete random variable by the above definition
because its range is a collection of colors, not real numbers.

3.1.1 Common Distributions

When studying random variables it is often more important to know how they distribute
probability onto their range than how they actually act as functions on their domains. As
such it is useful to have a notation that recognizes the fact that two functions may be very
different in terms of where they map domain elements, but nevertheless have the same
range and produce the same distribution on this range.

Definition 3.1.6. Let X : S → T and Y : S → T be discrete random variables. We

say X and Y have equal distriubtion provided P (X = t) = P (Y = t) for all t ∈ T .

There are many distributions which appear frequently enough they deserve their own
special names for easy identification. We shall use the symbol ∼ to mean “is distributed
as” or “is equal in distribution to”. For example, in the definition below X ∼ Bernoulli(p)
should be read as “X has a Bernoulli(p) distribution”. This says nothing explicit about
how X behaves as a function on its domain, but completely describes how X distributes
probability onto its range.

The following are common discrete distributions which we have seen arise previously in
the text.

Definition 3.1.7. X ∼ Uniform({1, 2, . . . , n}): Let n ≥ 1 be an integer. If X is

a random variable such that P (X = k ) = 1
n for all 1 ≤ k ≤ n then we say that X is
a uniform random variable on the set {1, 2, . . . , n}.

Definition 3.1.8. X ∼ Bernoulli(p): Let 0 ≤ p ≤ 1. When X is a random

variable such that P (X = 1) = p and P (X = 0) = 1 − p we say that X is a
Bernoulli random variable with parameter p. This takes the concept of a “Bernoulli
trial” which we have previously discussed and puts it in the context of a random
variable where 1 corresponds to success and 0 corresponds to failure.

Version: – November 19, 2024

68 discrete random variables

Definition 3.1.9. X ∼ Binomial(n, p): Let 0 ≤ p ≤ 1 and let n ≥ 1 be an integer.

If X is a random variable taking values in {0, 1, . . . , n} having a probability mass
function !
n k
P (X = k ) = p (1 − p)n−k
k
for all 0 ≤ k ≤ n, then X is a Binomial random variable with parameters n and p.
We have seen that such a quantity describes the number of successes in n Bernoulli
trials.

Definition 3.1.10. X ∼ Geometric(p): Let 0 < p < 1. If X is a random variable

with values in {1, 2, 3, . . . } and a probability mass function

P (X = k ) = p · (1 − p)k−1

for all k ≥ 1, then X is a geometric random variable with parameter p. Such

a random variable arises when determining how many Bernoulli trials must be
attempted before seeing the first success.

Definition 3.1.11. X ∼ Negative Binomial(r, p): Let 0 < p < 1. If X is a

random variable with values in {r, r + 1, r + 2, . . . } and a probability mass function

k−1 r
!
P (X = k ) = p · (1 − p)k−r
r−1

for all k ≥ r, then X is a Negative Binomial random variable with parameters (r, p).
Such a random variable arises when determining how many Bernoulli trials must be
attempted before seeing r successes.

Definition 3.1.12. X ∼ Poisson(λ): Let λ > 0. When X is a random variable

with values in {0, 1, 2, . . . } such that its probability mass function is

e−λ λk
P (X = k ) =
k!

for all k ≥ 0, then X is called a Poisson random variable with parameter λ. We first
used these distributions as approximations to a Binomial (n, p) when n was large
and p was small.

Version: – November 19, 2024

 3.1 random variables as functions 69

Definition 3.1.13. X ∼ HyperGeo(N , r, m): Let N , r, and m be positive integers

for which r < N and m < N . Let X be a random variable taking values in the
integers between min{m, r} and max{0, m − (N − r )} inclusive with probability mass
function
N −r
(kr )(m−k )
P (X = k ) =
(N
m)

The random variable X is called hypergeoemtric with parameters N , r, and m. Such

quantities occur when sampling without replacement.

exercises

Ex. 3.1.1. Consider the experiment of flipping a coin four times and recording the sequence
of heads and tails. Let S be the sample space of all sixteen possible orderings of the results.
Let X be the function on S describing the number of tails among the flips. Let Y be the
function on S describing the first flip (if any) to come up tails.

(a) Create a table as in Example 3.1.1 describing functions X and Y .

(b) Use the table to calculate P (X = 2).

(c) Use the table to calculate P (Y = 3).

Ex. 3.1.2. A pair of fair dice are thrown. Let X represent the larger of the two values on
the dice and let Y represent the smaller of the two values.

(a) Describe S, the domain of functions X and Y . How many elements are in S?

(b) What are the ranges of X and Y . Do X and Y have the same range? Why or why
not?

(c) Describe the distribution of X and describe the distribution of Y by finding the
probability mass function of each. Is it true that X and Y have the same distribution
?

Ex. 3.1.3. A pair of fair dice are thrown. Let X represent the number of the first die and
let Y represent the number of the second die.

(a) Describe S, the domain of functions X and Y . How many elements are in S?

(b) Describe T , the range of functions X and Y . How many elements are in T ?

Version: – November 19, 2024

70 discrete random variables

(c) Describe the distribution of X and describe the distribution of Y by finding the
probability mass function of each. Is it true that X and Y have the same distribution
?

(d) Are X and Y the same function? Why or why not?

Ex. 3.1.4. Use the ∼ notation to classify the distributions of the random variables described
by the scenarios below. For instance, if a scenario said, “let X be the number of heads
in three flips of a coin” the approrpriate answer would be X ∼ Binomial(3, 12 ) as that
describes the number of successes in three Bernoulli trials.

(a) Let X be the number of 5’s seen in four die rolls. What is the distribution of X?

(b) Each ticket in a certain lottery has a 20% chance to be a prize-winning ticket. Let Y
be the number of tickets that need to be purchased before seeing the first prize-winner.
What is the distribution of Y ?

(c) A class of ten students is comprised of seven women and three men. Four students
are randomly selected from the class. Let Z denote the number of men among the
four randomly selected students. What is the distribution of Z?

Ex. 3.1.5. Suppose X and Y are random variables.

(a) Explain why X + Y is a random variable.

(b) Theorem 3.1.2 does not require that X be real-valued. Why do you suppose that our
definition of “random variable” insisted that such functions should be real-valued?

Ex. 3.1.6. Let X : S → T be a discrete random variable. Suppose {Bi }i≥1 are sequence of
∞ ∞
events in T then show that X −1 ( X −1 (Bi ) and that if Bi and Bj are disjoint,
S S
Bi ) =
i=1 i=1
then so are X −1 (Bi ) and X −1 (Bj ).

3.2 independent and dependent variables

Most interesting problems require the consideration of several different random variables
and an analysis of the relationships among them. We have already discussed what it means
for a collection of events to be independent, and it is useful to extend this notion to random
variables as well. As with events, we will first describe the notion of pairwise independence
of two objects, before defining mutual independence of an arbitrary collection of objects.

Version: – November 19, 2024

 3.2 independent and dependent variables 71

3.2.1 Independent Variables

Definition 3.2.1. (Independence of a Pair of Random Variables) Two random

variables X and Y are independent if (X ∈ A) and (Y ∈ B ) are independent for
every event A in the range of X and every event B in the range of Y .

As events become more complicated and involve multiple random variables, a notational
shorthand will become useful. It is common in probability to write (X ∈ A, Y ∈ B ) for
the event (X ∈ A) ∩ (Y ∈ B ) and we will begin using this convention at this point.
Further, even though the definition of X : S → T and Y : S → U being independent
random variables requires that (X ∈ A) and (Y ∈ B ) be independent for all events A ⊂ T
and B ⊂ U , for discrete random variables it is enough to verify the events (X = t) and
(Y = u) are independent events for all t ∈ T and u ∈ U to conclude they are independent
(See Exercise 3.2.12).

Example 3.2.2. When we originally considered the example of rolling a pair of dice, we
viewed the results as thirty-six equally likely outcomes. However, it is also possible to view
the result of each die as a random variable in its own right, and then consider the possible
results of the pair of random variables. Let X, Y ∼ Uniform({1, 2, 3, 4, 5, 6}) and suppose
X and Y are independent. If x, y ∈ {1, 2, 3, 4, 5, 6} what is P (X = x, Y = y )?
By indpendence P (X = x, Y = y ) = P (X = x)P (Y = y ) = 1
6
1
·
= 36 . Therefore,
1
6
the result is identical to the original perspective – each of the thirty-six outcomes of the
pair of dice is equally likely. ■

Definition 3.2.3. (Mutual Independence of Random Variables) A finite

collection of random variables X1 , X2 , . . . , Xn is mutually independent if the sets
(Xj ∈ Aj ) are mutually independent for all events Aj in the ranges of the corre-
sponding Xj .
An arbitrary collection of random variables Xt where t ∈ I for some index set I is
mutually independent if every finite sub-collection is mutually independent.

For many problems it is useful to think about repeating a single experiment many times
with the results of each repetition being independent from every other. Though the results
are assumed to be independent, the experiment itself remains the same, so the random
variables produced all have the same distribution. The resulting sequence of random
variables X1 , X2 , X3 , . . . is referred to as “i.i.d.” (standing for “independent and identically
distributed”). When considering such sequences we will sometimes write X1 , X2 , X3 , . . .

Version: – November 19, 2024

72 discrete random variables

are i.i.d. with distribution X, where X is a random variable that shares their common
distribution.

Example 3.2.4. Let X1 , X2 , . . . , Xn be i.i.d. with a Geometric(p) distribution. What is

the probabilty that all of these random variables are larger than some positive integer j?
As a preliminary calculation, if X ∼ Geometric(p) and if j ≥ 1 is an integer we may
determine P (X > j ).
∞ ∞
p(1 − p)i−1
X X
P (X > j ) = P (X = i) =
i=j +1 i=j +1
p · (1 − p)j
= = (1 − p)j .
1 − (1 − p)

But each of X1 , X2 , . . . , Xn have this distribution, so using the computatation above,

together with independence, we have

P (X1 > j, X2 > j, . . . , Xn > j ) = P (X1 > j )P (X2 > j ) . . . P (Xn > j )
= (1 − p)j · (1 − p)j · · · · · (1 − p)j
= (1 − p)nj . ■

3.2.2 Conditional, Joint, and Marginal Distributions

Consider a problem involving two random variables. Let X be the number of centimeters
of rainfall in a certain forest in a given year, and let Y be the number of square meters
of the forest burned by fires that same year. It seems these variables should be related;
knowing one should affect the probabilities associated with the values of the other. Such
random variables are not independent of each other and we now introduce several ways to
compute probabilities under such circumstances. An important concept toward this end is
the notion of a “conditional distribution” which reflects the fact that the occurrence of an
event may affect the likely values of a random variable.

Definition 3.2.5. Let X be a random variable on a sample space S and let A ⊂ S

be an event such that P (A) > 0. Then the probability Q described by

Q(B ) = P (X ∈ B | A) (3.2.1)

is called the “conditional distribution” of X given the event A.

Version: – November 19, 2024

 3.2 independent and dependent variables 73

As with any discrete random variable, the distribution is completely determined by the
probabilities associated with each possible value the random variable may assume. This
means the conditional distribution may be considered known provided the values of
P (X = a|A) are known for every a ∈ Range(X ). Though this definition allows for A to be
any sort of event, in this section we will mainly consider examples where A describes the
outcome of some random variable. So a notation like P (X|Y = b) will be the conditional
distribution of the random variable X given that the random variable Y is known to have
the value b.
In many cases random variables are dependent in such a way that the distribution of
one variable is known in terms of the values taken on by another.
Example 3.2.6. Let X ∼ Uniform({1, 2}) and let Y be the number of heads in X tosses
of a fair coin. Clearly X and Y should not be independent. In particular, a result of Y = 0
could occur regardless of the value of X, but a result of Y = 2 guarantees that X = 2 as
two heads could not be observed with just one flip on the coin. Any information regarding
X or Y may influence the distribution of the other, but the description of the variables
makes it clearest how Y depends on X. If X = 1 then Y is the number of heads in one
flip of a fair coin. Letting A be the event (X = 1) and using the terminology of (3.2.1)
from Definition 3.2.5, we can say the conditional distribution of Y given that X = 1 is a
Bernoulli( 12 ). We will use the notation

1
(Y | X = 1) ∼ Bernoulli( )
2

to indicate this fact. In other words, this notation means the same thing as the pair of
equations

1
P (Y = 0 | X = 1) =
2
1
P (Y = 1 | X = 1) =
2

If X = 2 then Y is the number of heads in two flips of a fair coin and therefore (Y | X = 2) ∼
Binomial(2, 21 ) which means the following three equations hold:

1
P (Y = 0 | X = 2) =
4
1
P (Y = 1 | X = 2) =
2
1
P (Y = 2 | X = 2) = ■
4
The conditional probabilities of the previous example were easily determined in part
because the description of Y was already given in terms of X, but frequently random

Version: – November 19, 2024

74 discrete random variables

variables may be dependent in some way that is not so explicitly described. A more
general method of expressing the dependence of two (or more) variables is to present the
probabilities associated with all combinations of possible values for every variable. This is
known as their joint distribution.

Definition 3.2.7. If X and Y are discrete random variables, the “joint distribution”
of X and Y is the probability Q on pairs of values in the ranges of X and Y defined
by
Q((a, b)) = P (X = a, Y = b).

The definition may be expanded to a finite collection of discrete random variables

X1 , X2 , . . . , Xn for which the joint distribution of all n variables is the probability
defined by

Q((a1 , a2 , . . . , an )) = P (X1 = a1 , X2 = a2 , . . . , Xn = an ).

In the above definition as discussed before for any event D,

Q((a1 , a2 , . . . , an )).
X
Q(D ) =
(a1 ,a2 ,...,an )∈D

For a pair of random variables with few possible outcomes, it is common to describe the
joint distribution using a chart for which the columns correspond to possible X values, the
rows to possible Y values, and for which the entries of the chart are probabilities.

Example 3.2.8. Let X and Y be the dependent variables described in Example 3.2.6. The
X variable will be either 1 or 2. The Y variable could be as low as 0 (if no heads are flipped)
or as high as 2 (if two coins are flipped and both show heads). As Range(X ) = {1, 2}
and as Range(Y ) = {0, 1, 2}, the pair (X, Y ) could potentially be any of the six possible
pairings (though, in fact, one of the pairings has probability zero). To find the joint
distribution of X and Y we must calculate the probabilities of each possibility. In this case
the values may be obtained using the definition of conditional probability. For instance,

1 1 1
P (X = 1, Y = 0) = P (Y = 0|X = 1) · P (X = 1) = · =
2 2 4

and
1
P (X = 1, Y = 2) = P (Y = 2|X = 1) · P (X = 1) = 0 · = 0.
2
The entire joint distribution P (X = a, Y = b) is described by the following chart.

Version: – November 19, 2024

 3.2 independent and dependent variables 75

X=1 X=2
Y =0 1/4 1/8
Y =1 1/4 1/4
Y =2 0 1/8 ■

Knowing the joint distribution of random variables gives a complete picture of the proba-
bilities associated with those variables. From that information it is possible to compute
all conditional probabilities of one variable from another. For instance, in the example
analyzed above, the variable Y was originally described in terms of how it depended on X.
However, this also means that X should be dependent on Y . The joint distribution may
be used to determine how.

Example 3.2.9. Let X and Y be the variables of Example 3.2.6. How may the conditional
distributions of X given values of Y be determined?
There will be three different conditional distributions depending on whether Y = 0,
Y = 1, or Y = 2. Below we will solve the Y = 0 case. The other two cases will be left as
exercises. The conditional distribution of X given that Y = 0 is determined by the values
of P (X = 1|Y = 0) and P (X = 2|Y = 0) both of which may be computed using Bayes’
rule.

P (Y = 0|X = 1) · P (X = 1)
P (X = 1|Y = 0) =
P (Y = 0)
P (Y = 0|X = 1) · P (X = 1)
=
P (Y = 0|X = 1) · P (X = 1) + P (Y = 0|X = 2) · P (X = 2)
(1/2)(1/2) 2
= =
(1/2)(1/2) + (1/4)(1/2) 3

As the only values for X are 1 and 2 it must be that P (X = 2|Y = 0) = 13 . ■

Just because X and Y are dependent on each other doesn’t mean they need to be
thought of as a pair. It still makes sense to talk about the distribution of X as a random
variable in its own right while ignoring its dependence on the variable Y . When there are
two or more variables under discussion, the distribution of X alone is sometimes called the
“marginal” distribution of X because it can be computed using the margins of the chart
describing the joint distribution of X and Y .

Example 3.2.10. Continue with X and Y as described in Example 3.2.6. Below is the
chart describing the joint distribution of X and Y that was created in Example 3.2.8, but
with the addition of one column on the right and one row at the bottom. The entries in
the extra column are the sums of the values in the corresponding row; likewise the entries
in the extra row are the sums of the values in the corresponding column.

Version: – November 19, 2024

76 discrete random variables

X=1 X=2 Sum

Y =0 1/4 1/8 3/8
Y =1 1/4 1/4 4/8
Y =2 0 1/8 1/8
Sum 1/2 1/2

The values in the right hand margin (column) exactly describe the distribution of Y . For
instance the event (Y = 0) can be partitioned into two disjoint events (X = 1, Y =
0) ∪ (X = 2, Y = 0) each of which is already described in the joint distribution chart.
Adding them together gives the result that P (Y = 0) = 8.
3
In a similar fashion, the
bottom margin (row) describes the distribution of X. This extended chart also makes it
numerically clearer why these two random variables are dependent. For instance,

1 3
P (X = 1, Y = 0) = while P (X = 1) · P (Y = 0) =
4 16

As these quantities are unequal, the random variables cannot be independent. ■

In general, knowing the marginal distributions of X and Y is not sufficient information

to reconstruct their joint distribution. This is because the marginal distributions do not
provide any information about how the random variables relate to each other. However, if
X and Y happen to be independent, then their joint distribution may easily be computed
from the marginals as

P (X = x, Y = y ) = P (X = x)P (Y = x)

3.2.3 Memoryless Property of the Geometric Random Variable

It is also possible to calculate conditional probabilities of a random variable based on

subsets of its own values. A particularly important example of this is the “memoryless
property” of geometric random variables.

Example 3.2.11. Suppose we toss a fair coin until the first head appears. Let X be the
number of tosses performed. We have seen in Example 2.1.2 that X ∼ Geometric( 12 ). Note
that if m is a positive integer,
∞ ∞
X X 1 1
P (X > m) = P (X = k ) = = m
k =m+1 k =m+1
2 k 2

Version: – November 19, 2024

 3.2 independent and dependent variables 77

Now let n be a positive integer and suppose we take the event (X > n) as given. In other
words, we assume we know that none of the first n flips resulted in heads. What is the
conditional distribution of X given this new information? A routine calculation shows

P (X > n + m) 1
2m+n 1
P (X > n + m | X > n) = = =
P (X > n) 1
2n
2m

As a consequence,
P (X > n + m | X > n) = P (X > m). (3.2.2)

Given that a result of heads has not occurred by the n-th flip, the probability that such a
result will require at least m more flips is identical to the (non-conditional) probability
the result would have required more than m flips from the start. In other words, if we
know that the first n flips have not yet produced a head, the number of additional flips
required to observe the first head still is a Geometric( 21 ) random variable. This is called
the “memoryless property” of the geometric distribution as it can be interpreted to mean
that when waiting times are geometrically distributed, no matter how long we wait for an
event to occur, the future waiting time always looks the same given that the event has not
occurred yet. The result remains true of geometric variables of any parameter p, a fact
which we leave as an exercise. ■

3.2.4 Multinomial Distributions

Consider a situation similar to that of Bernoulli trials, but instead of results of each attempt
limited to success or failure, suppose there are many different possible results for each
trial. As with the Bernoulli trial cases we assume that the trials are mutually independent,
but identically distributed. In the next example we will show how to calculate the joint
distribution for the random variables representing the number of times each outcome
occurs.

Example 3.2.12. Suppose we perform n i.i.d. trials each of which has k different possible
outcomes. For j = 1, 2, . . . , k, let pj represent the probability any given trial results in
the j-th outcome and let Xj represent the number of the n trials that result in the j-th
outcome. The joint distribution of all of the random variables X1 , X2 , . . . , Xk is called a
“multinomial distribution”.
Let B (x1 , x2 , . . . , xk ) = {X1 = x1 , X2 = x2 , . . . , Xk = xk }. Then,

P (B (x1 , x2 , . . . , xk )) =
X
P ({ω})
ω∈B (x1 ,x2 ,...,xk )

Version: – November 19, 2024

78 discrete random variables

Each ω ∈ B (x1 , x2 , . . . , xk ) is an element in the sample space corresponding to the j-th

outcome occuring exactly xj times. As the trials are independent, and as an outcome j
occurs in xj trials, each of which had probability pj , this means

k
Y xj
P ({ω}) = pj
j =1

Consequently, each outcome in B (x1 , x2 , . . . , xk ) has the same probability. So to determine

the likelihood of the event, we need only determine |B (x1 , x2 , . . . , xk )|, the number of
outcomes the event contains. The calculation of this quantity is a combinatorial problem;
it is the number of ways of allocating n balls in k boxes, such that xj of them fall into box
j. We leave it as an exercise to prove that

n!
|B (x1 , x2 , . . . , xk )| =
x1 ! x2 ! . . . xk !

With that computation complete, the joint distribution of X1 , X2 , . . . Xk is given by


k x
n!
pj j if xj ∈ {0, 1, . . . , n}

 Q

 x1 ! x 2 ! ... x k !
j =1



k


and
 P
P (X1 = x1 , . . . , Xk = xk ) = xj = n

 j =1





0 otherwise

■



exercises

Ex. 3.2.1. An urn has four balls labeled 1, 2, 3, and 4. A first ball is drawn and its number
is denoted by X. A second ball is then drawn from the three remaining balls in the urn
and its number is denoted by Y .

(a) Calculate P (X = 1).

(b) Calculate P (Y = 2 | X = 1).

(c) Calculate P (Y = 2).

(d) Calculate P (X = 1, Y = 2).

(e) Are X and Y independent? Why or why not?

Ex. 3.2.2. Two dice are rolled. Let X denote the sum of the dice and let Y denote the
value of the first die.

Version: – November 19, 2024

 3.2 independent and dependent variables 79

(a) Calculate P (X = 7) and P (Y = 4).

(b) Calculate P (X = 7, Y = 4).

(c) Calculate P (X = 5) and P (Y = 4).

(d) Calculate P (X = 5, Y = 4).

(e) Are X and Y independent? Why or why not?

Ex. 3.2.3. Let X and Y be the variables described in Example 3.2.6.

(a) Determine the conditional distribution of X given that Y = 1.

(b) Determine the conditional distribution of X given that Y = 2.

Ex. 3.2.4. Let X and Y be random variables with joint distribution given by the chart
below.
X=0 X=1 X=2
Y =0 1/12 0 3/12
Y =1 2/12 1/12 0
Y =2 3/12 1/12 1/12

(a) Compute the marginal distributions of X and Y .

(b) Compute the conditional distribution of X given that Y = 2.

(c) Compute the conditional distribution of Y given that X = 2.

(d) Carry out a computation to show that X and Y are not independent.

Ex. 3.2.5. Let X be a random variable with range {0, 1} and distribution

1 2
P (X = 0) = and P (X = 1) =
3 3

and let Y be a random variable with range {0, 1, 2} and distribution

1 1 3
P (Y = 0) = , P (Y = 1) = , and P (Y = 2) =
5 5 5

Suppose that X and Y are independent. Create a chart describing the joint distribution of
X and Y .
Ex. 3.2.6. Consider six independent trials each of which are equally likely to produce
a result of 1, 2, or 3. Let Xj denote the number of trials that result in j. Calculate
P (X1 = 1, X2 = 2, X3 = 3).

Version: – November 19, 2024

80 discrete random variables

Ex. 3.2.7. Prove the combinatorial fact from Example 3.2.12 in the following way. Let
An (x1 , x2 , . . . , xk ) denote the number of ways of putting n balls into k boxes in such a way
that exactly xj balls wind up in box j for j = 1, 2, . . . , k.

(a) Prove that An (x1 , x2 , . . . , xk ) = (xn1 )An−x1 (x2 , x3 , . . . , xk ).

(b) Use part (a) and induction to prove that An (x1 , x2 , . . . , xk ) = x1 ! x2 ! ... xk ! .
n!

Ex. 3.2.8. Let X be the result of a fair die roll and let Y be the number of heads in X
coin flips.

(a) Both X and (Y |X = n) can be written in terms of common distributions using the
∼ notation. What is the distribution of X? What is the distribution of (Y |X = n)
for n = 1, . . . 6?

(b) Determine the joint distribution for X and Y .

(c) Calculate the marginal distribution of Y .

(d) Compute the conditional distribution of X given that Y = 6.

(e) Compute the conditional distribution of X given that Y = 0.

(f) Perform a computation to prove that X and Y are not independent.

Ex. 3.2.9. Suppose the number of earthquakes that occur in a year, anywhere in the
world, is a Poisson random variable with mean λ. Suppose the probability that any given
earthquake has magnitude at least 5 on the Richter scale is p independent of all other quakes.
Let N ∼ Poisson(λ) be the number of earthquakes in a year and let M be the number of
earthquakes in a year with magnitude at least 5, so that (M |N = n) ∼ Binomial(n, p).

(a) Calculate the joint distribution of M and N .

(b) Show that the marginal distribution of M is determined by

∞
1 −λ λn−m
(1 − p)n−m
X
P (M = m) = e (λp)m
m! n=m ( n − m ) !

for m > 0.

(c) Perform a change of variables (where k = n − m) in the infinite series from part (b)
to prove
∞
1 −λ X (λ(1 − p))k
P (M = m) = e (λp)m
m! k =0
k!

Version: – November 19, 2024

 3.3 functions of random variables 81

∞ k
(d) Use part (c) together with the infinite series equality ex = k! to conclude that
x
P
k =0
M ∼ Poisson(λp).

Ex. 3.2.10. Let X be a discrete random variable which has N = {1, 2, 3, . . . } as its
range. Suppose that for all positive integers m and n, X has the memoryless property –
P (X > n + m | X > n) = P (X > m). Prove that X must be a geometric random variable.
[Hint: Define p = P (X = 1) and use the memoryless property to calculate P (X = n)
inductively].
Ex. 3.2.11. A discrete random variable X is called “constant” if there is a single value c
for which P (X = c) = 1.

(a) Prove that if X is a constant discrete random variable then X is independent of

itself.

(b) Prove that if X is a discrete random variable which is independent of itself, then X
must be constant. [Hint: It may help to look at Exercise 1.4.8].

Ex. 3.2.12. Let X : S → T and Y : S → U be discrete random variables. Show that if

P (X = t, Y = u) = P (X = t)P (Y = u)

for all t ∈ T and u ∈ U then X and Y are independent random variables.

3.3 functions of random variables

There are many circumstances where we want to consider functions applied to random
variables as inputs of functions. For a simple geometric example, suppose a rectangle
is selected in such a way that its width X and its length Y are both random variables
with known joint distribution. The area of the rectangle is A = XY , and as X and Y
are random, A should be random as well. How may the distribution of A be calculated
from the joint distribution of X and Y ? In general, if a new random variable Z depends
on random variables X1 , X2 , . . . , Xn which have a given joint distribution, how may the
distribution of Z be calculated from what is already known? In this section we discuss the
answers to such questions and also address related issues surrounding independence.
If X : S → T is a random variable and if f : T → R is a function, then the quantity
f (X ) makes sense as a composition of functions f ◦ X : S → R. In fact, as f (X ) is defined
on the sample space S, this new composition is itself a random variable.
The same reasoning holds for functions of more than one variable. If X1 , X2 , . . . , Xn
are random variables then f (X1 , X2 , . . . , Xn ) is a random variable provided f is defined for

Version: – November 19, 2024

82 discrete random variables

the values the Xj variables produce. Below we illustrate how to calculate the distribution
of f (X1 , X2 , . . . , Xn ) in terms of the joint distribution of the Xj input variables. We
demonstrate the method with several examples followed by a general theorem.

3.3.1 Distribution of f (X ) and f (X1 , X2 , . . . , Xn )

The distribution of f (X ) involves the probability of events such as (f (X ) = a) for values

of a that the function may produce. The key to calculating this probability is that these
events may be rewritten in terms of the input values of X instead of the output values of
f (X ).

Example 3.3.1. Let X ∼ Uniform({−2, −1, 0, 1, 2}) and let f (x) = x2 . Determine the
range and distribution of f (X ).
As f (X ) = X 2 , the values that f (X ) produces are the squares of the values that X
produces. Squaring the values in {−2, −1, 0, 1, 2} shows the range of f (X ) is {0, 1, 4}.
The probabilities that f (X ) takes on each of these three values determine the distribution
of f (X ) and these probabilities can be easily calculated from the known probabilities
associated with X.

1
P (f (X ) = 0) = P (X = 0) =
5
1 1 2
P (f (X ) = 1) = P ((X = 1) ∪ (X = −1)) = + =
5 5 5
1 1 2
P (f (X ) = 4) = P ((X = 2) ∪ (X = −2)) = + =
5 5 5 ■

A complication with this method is that there may be many different inputs that produce
the same output. Sometimes a problem requires careful consideration of all ways that a
given output may be produced. For instance,

Example 3.3.2. What is the probability the sum of three dice will equal six? Let X, Y ,
and Z be the results of the first, second, and third die respectively. These are i.i.d. random
variables each distributed as Uniform({1, 2, 3, 4, 5, 6}). A sum of six can be arrived at in
three distinct ways:

Case I: through three rolls of 2;

Case II: through one roll of 3, one roll of 2, and one roll of 1; or
Case III: through one roll of 4 and two rolls of 1

Version: – November 19, 2024

 3.3 functions of random variables 83

The first of these is the simplest to deal with as independence gives

1 1 1 1
P (X = 2, Y = 2, Z = 2) = P (X = 2) · P (Y = 2) · P (Z = 2) = · · =
6 6 6 216

The other cases involve a similar computation, but are complicated by the consideration of
which number shows up on which die. For instance, both events (X = 1, Y = 2, Z = 3)
and (X = 3, Y = 2, Z = 1) are included as part of Case II as are four other permutations
of the numbers. Likewise Case III includes three permutations, one of which is (X =
4, Y = 1, Z = 1). Putting all three cases together,

P (sum of 6) = P (Case I) + P (Case II) + P (Case III)

1 1 1 5
= +6· +3· = .
216 216 216 108

So there is slightly less than a 5% chance three rolled dice will produce a sum of six. ■

This method may also be used to show relationships among the common (named) distribu-
tions that have been previously described, as in the next two examples.

Example 3.3.3. Let X, Y ∼ Bernoulli(p) be two independent random variables. If

Z = X + Y , show that Z ∼ Binomial(2, p).

This result should not be surpirsing given how Bernoulli and Binomial distributions
arose in the first place. Each of X and Y produces a value of 0 if the corresponding
Bernoulli trial was a failure and 1 if the trial was a success. Therefore Z = X + Y equals
the total number of successes in two independent Bernoulli trials, which is exactly what
led us to the Binomial distribution in the first place. However, it is instructive to consider
how this problem relates to the current topic of discussion.

As each of X and Y is either 0 or 1 the possible values of Z are in the set {0,1,2}. A
result of Z = 0 can only occur if both X and Y are zero. So,

P (Z = 0) = P (X = 0, Y = 0)
= P (X = 0) · P (Y = 0)
= (1 − p)(1 − p)
= (1 − p)2 .

Similarly, P (Z = 2) = P (X = 1) · P (Y = 1) = p2 .

Version: – November 19, 2024

84 discrete random variables

There are two different ways that Z could equal 1, either X = 1 and Y = 0, or X = 0
and Y = 1. So,

P (Z = 1) = P ((X = 1, Y = 0) ∪ (X = 0, Y = 1))
= P (X = 1, Y = 0) + P (X = 0, Y = 1)
= p(1 − p) + (1 − p)p
= 2p(1 − p)

These values of P (Z = 0), P (Z = 1), and P (Z = 2) are exactly what define Z ∼

Binomial(2, p). ■

Two of the previous three examples involve adding random variables together. In fact,
addition is one of the most common examples of applying functions to random quantities.
In the previous situations, calculating the distribution of the sum was relatively simple
because the component variables only had finitely many outcomes. But now suppose X
and Y are random variables taking values in {0, 1, 2, . . . } and suppose Z = X + Y . How
could P (Z = n) be calculated?
As both X and Y are non-negative and as Z = X + Y , the value of Z must be at
least as large as either X or Y individually. If Z = n, then X could take on any value
j ∈ {0, 1, . . . , n}, but once that value is determed, the value of Y is compelled to be n − j to
give the appropriate sum. In other words, the event (Z = n) partitions into the following
union. n
(X = j, Y = n − j ).
[
(Z = n) =
j =0

When X and Y are independent, this means

 
n
[
P (Z = n) = P  (X = j, Y = n − j )
j =0
n
X
= P (X = j, Y = n − j )
j =0
X n
= P (X = j ) · P (Y = n − j )
j =0

Such a computation is usually referred to as a “convolution” which will be addressed more

generally later in the text. It occurs regularly when determining the distribution of sums
of independent random variables.

Version: – November 19, 2024

 3.3 functions of random variables 85

Example 3.3.4. Let X ∼ Poisson(λ1 ) and Y ∼ Poisson(λ2 ) be independent random

variables. Let Z = X + Y .

(a) Find the distribution of Z.

(b) Find the conditional distribution of X | Z.

For x, y ∈ {0, 1, 2, . . . }, we have

P (X = x, Y = y ) = P (X = x) · P (Y = y )
λx λy
= e−λ1 1 · e−λ2 2 .
x! y!

(a) As computed above, the distribution of Z is given by the convolution. For any
n = 0, 1, 2, . . . we have

P (Z = n) = P (X + Y = n)
n
X
= P (X = j ) · P (Y = n − j )
j =0
n
λj1 −λ2 λ2n−j
e−λ1
X
= ·e
j =0
j! (n − j ) !
n
λj1 λ2n−j
= e−(λ1 +λ2 )
X

j =0
j!(n − j )!
n
1 X n!
= e−(λ1 +λ2 ) λj λn−j
n! j =0 j!(n − j )! 1 2
(λ1 + λ2 )n
= e−(λ1 +λ2 )
n!

where in the last line we have used the binomial expansion (2.1.1). Hence we can conclude
that Z ∼ Poisson (λ1 + λ2 ).
The above calculation is easily extended by an induction argument to obtain the fact
that if λi > 0, Xi , 1 ≤ i ≤ k are independent Poisson(λi ) distributed random variables
(respectively). Then Z = has Poisson ( distribution. Thus if we have k
Pk Pk
i=1 Xi i = 1 λi )
independent Poisson (λ) random variables then has Poisson(kλ) distribution.
Pk
i=1 Xi

(b) We readily observe that X and Z are dependent. We shall now try to understand
the conditional distribution of (X|Z = n). As the ranges of X and Y do not have any
negative numbers, given that Z = X + Y = n, X can only take values in {0, 1, 2, 3, . . . , n}.

Version: – November 19, 2024

86 discrete random variables

For k ∈ {0, 1, 2, 3, . . . , n} we have,

P (X = k, X + Y = n) P (X = k, Y = n − k )
P (X = k | Z = n) = =
P (X + Y = n) P (X + Y = n)
P (X = k )P (Y = n − k )
=
P (X + Y = n)
λk1 λn−k
e−λ1 k! · e−λ2 (n−k
2
)! n! λk1 λn−k
2
= )n
=
e−(λ1 +λ2 ) (λ1 +n!λ2 k!(n − k )! (λ1 + λ2 )n
! k  n−k
n λ1 λ2
= .
k λ1 + λ2 λ1 + λ2

Hence (X | Z = n) ∼ Binomial(n, λ1λ+1λ2 ). ■

The point of the examples above is that a probability associated with a functional value
f (X1 , X2 , . . . , Xn ) may be calculated directly from the probabilities associated with the
input variables X1 , X2 , . . . , Xn . The following theorem explains how this may be accom-
plished generally for any number of variables.

Theorem 3.3.5. Let X1 , X2 , . . . , Xn be random variables defined on a single

sample space S. Let f be a function of n variables for which f (X1 , X2 , . . . , Xn ) is
defined in the range of the Xj variables. Let B be a subset of the range of f . Then,

P (f (X1 , X2 , . . . , Xn ) ∈ B ) = P ((X1 , X2 , . . . , Xn ) ∈ f −1 (B )).

Proof. Note that the events (f (X1 , X2 , . . . , Xn ) ∈ B ) and ((X1 , X2 , . . . , Xn ) ∈ f −1 (B ))

are both subsets of S, as outcomes s ∈ S determine the values of the Xj variables which in
turn determine the output of f . The theorem follows immediately from the set theoretic
fact that

f (X1 (s), X2 (s), . . . , Xn (s)) ∈ B ⇐⇒ (X1 (s), X2 (s), . . . Xn (s)) ∈ f −1 (B )

This is because the expression f (X1 (s), X2 (s), . . . , Xn (s)) ∈ B is what defines s to be an
outcome in the event (f (X1 , X2 , . . . , Xn ) ∈ B ). Likewise, the expression

(X1 (s), X2 (s), . . . Xn (s)) ∈ f −1 (B )

defines s to be in the event ((X1 , X2 , . . . Xn ) ∈ f −1 (B )). As these events are equal, they
have the same probability. ■

Version: – November 19, 2024

 3.3 functions of random variables 87

3.3.2 Functions and Independence

If X and Y are independent random variables, does that guarantee that functions f (X )
and g (Y ) of these random variables are also indpendent? If we take the intuitive view of
independence as saying “knowing information about X does not affect the probabilities
associated with Y ” then it seems the answer should be “yes”. After all, X determines
the value of f (X ) and Y determines the value of g (Y ). So information about f (X )
should translate to information about X and infromation about g (Y ) should translate to
information about Y . Therefore if information about f (X ) affected probabilities associated
with g (Y ), then it seems there should be information about X that would affect the
probability assoicated with Y . We generalize this argument and make it more rigorous in
the following result.

Theorem 3.3.6. Fix n ≥ 1. For each j ∈ {1, 2, . . . , n} let i ∈ {1, 2, . . . , mj } for

some positive integer mj . Suppose Xi,j is an array of mutually independent discrete
random variables and we define

Yj = fj (X1,j , X2,j , . . . Xmj ,j ),

where fj : Rmj → R are continuous functions. Then the resulting variables

Y1 , Y2 , . . . , Yn are mutually independent.

Informally this theorem says that random quantities produced from independent inputs
will, themselves, be independent.

Proof. Let B1 , B2 , . . . , Bn be sets in the ranges of Y1 , Y2 , . . . , Yn respectively. Using inde-

pendence and some set-theoretic identities, we have

P (Y1 ∈ B1 , . . . , Yn ∈ Bn )
= P (f1 (X1,1 , . . . , Xm1 ,1 ) ∈ B1 , . . . , fn (X1,n , . . . , Xmn ,n ) ∈ Bn )
= P ((X1,1 , . . . , Xm1 ,1 ) ∈ f1−1 (B1 ), . . . , (X1,n , . . . , Xmn ,n ) ∈ fn−1 (Bn ))
n
P ((Xi,1 , . . . , Xmi ,i ) ∈ fi−1 (Bi ))
Y
=
i=1
n
P (fi (Xi,1 , . . . , Xmi ,i ) ∈ Bi )
Y
=
i=1
= P (Y1 ∈ B1 ) · · · P (Yn ∈ Bn )

It follows that Y1 , Y2 , . . . , Yn are mutually independent. ■

Version: – November 19, 2024

88 discrete random variables

exercises

Ex. 3.3.1. Let X ∼ Uniform({1, 2, 3}) and Y ∼ Uniform({1, 2, 3}) be independent and let
Z = X +Y.

(a) Determine the range of Z.

(b) Determine the distriubtion of Z.

(c) Is Z uniformly distributed over its range?

Ex. 3.3.2. Consider the experiment of rolling three dice and calculating the sum of the
rolls. Answer the following questions.

(a) What is the range of possible results of this experiment?

(b) Calculate the probability the sum equals three.

(c) Calculate the probability the sum equals four.

(d) Calculate the probability the sum equals five.

(e) Calculate the probability the sum equals ten.

Ex. 3.3.3. Let X ∼ Bernoulli(p) and Y ∼ Bernoulli(q ) be independent.

(a) Prove that XY is a Bernoulli random variable. What is its parameter?

(b) Prove that (1 − X ) is a Bernoulli random variable. What is its parameter?

(c) Prove that X + Y − XY is a Bernoulli random variable. What is its parameter?

Ex. 3.3.4. Let X ∼ Binomial(n, p) and Y ∼ Binomial(m, p). Assume X and Y are
independent and let Z = X + Y . Prove that Z ∼ Binomial(m + n, p).
Ex. 3.3.5. Let X ∼ Negative Binomial(r, p) and Y ∼ Negative Binomial(s, p). Assume X
and Y are independent and let Z = X + Y . Prove that Z ∼ Negative Binomial(r + s, p).
Ex. 3.3.6. Consider one flip of a single fair coin. Let X denote the number of heads on the
flip and let Y denote the number of tails on the flip.

(a) Show that X, Y ∼ Bernoulli( 12 ).

(b) Let Z = X + Y and explain why P (Z = 1) = 1.

(c) As (b) clearly says that Z cannot be a Binomial (2, 12 ), explain why this result does
not conflict with the conclusion of Example 3.3.3.

Version: – November 19, 2024

 3.3 functions of random variables 89

Ex. 3.3.7. Let X ∼ Geometric(p) and Y ∼ Geometric(p) be independent. Let Z = X + Y .

(a) Determine the range of Z.

(b) Use a convolution to prove that P (Z = n) = (n − 1)p2 (1 − p)n−2 .

(c) Recall from the discussion of Geometric distributions that (X = 1) is the most likely
result for X and (Y = 1) is the most likely result for Y . This does not imply that
(Z = 2) is the most likely outcome for Z. Determine the values of p for which
P (Z = 3) is larger than P (Z = 2).

Ex. 3.3.8. Let X1 , X2 , X3 , X4 be an i.i.d. sequence of Bernouli(p) random variables. Let

Y = X1 + X2 + X3 + X4 . Prove that P (Y = 2) = 6p2 (1 − p)2 .
Ex. 3.3.9. Let X1 , X2 , . . . , Xn be an i.i.d. sequence of Bernoulli(p) random variables. Let
Y = X1 + X2 + · · · + Xn . Prove that Y ∼ Binomial(n, p).
Ex. 3.3.10. Let X1 , X2 , . . . , Xr be an i.i.d. sequence of Geometric (p) random variables.
Let Y = X1 + X2 + · · · + Xr . Prove that Y ∼ Negative Binomial(r, p).
Ex. 3.3.11. Let X1 , X2 , X3 , X4 be an i.i.d. sequence of Bernoulli(p) random variables. Let
Y = X1 + X2 and let Z = X3 + X4 . Note that Example 3.3.3 guarantees that Y , Z ∼
Binomial(2, p).

(a) Create a chart describing the joint distribution of Y and Z.

(b) Use the chart from (a) to explain why Y and Z are independent.

(c) Explain how you could use Theorem 3.3.6 to reach the conclusion that Y and Z are
independent without calculating their joint distribution.

Ex. 3.3.12. Let X1 , X2 , X3 be an i.i.d. sequence of Bernoulli(p) random variables. Let

Y = X1 + X2 and let Z = X2 + X3 . Note that Example 3.3.3 guarantees that Y , Z ∼
Binomial(2, p).

(a) Create a chart describing the joint distribution of Y and Z.

(b) Use the chart from (a) to explain why Y and Z are not independent.

(c) Explain why the conclusion from (b) is not inconsistant with Theorem 3.3.6.

Ex. 3.3.13. Let X1 , X2 , . . . , Xn be an i.i.d. sequence of discrete random variables and let
Z be the maximum of these n variables. Let r be a real number and let R = P (X1 ≤ r ).
Prove that P (Z ≤ r ) = Rn .

Version: – November 19, 2024

90 discrete random variables

Ex. 3.3.14. Let X1 , X2 , . . . , Xn be an i.i.d. sequence of discrete random variables and let
Z be the minimum of these n variables. Let r be a real number and let R = P (X1 ≤ r ).
Prove that P (Z ≤ r ) = 1 − (1 − R)n .
Ex. 3.3.15. Let X ∼ Geometric(p) and let Y ∼ Geometric(q ) be independent random
variables. Let Z be the smaller of X and Y . It is a fact that Z is also geometrically
distributed. This problem asks you to prove this fact using two different methods.
METHOD I:

(a) Explain why the event (Z = n) can be written as the disjoint union

(Z = n) = (X = n, Y = n) ∪ (X = n, Y > n) ∪ (X > n, Y = n)

(b) Recall from the proof of the memoryless property of geometric random variables that
2m . Use this fact and part (a) to prove that
1
P (X > m) =

P (Z = n) = [(1 − p)(1 − q )]n−1 (pq + p(1 − q ) + (1 − p)q )

(c) Use (b) to conclude that Z ∼ Geometric(r ) for some quantity r and calculate the
value of r in terms of the p and q.

METHOD II: Recall that geometric random variables first arose from noting the time it
takes for a sequence of Bernoulli trials to first produce a success. With that in mind, let
A1 , A2 , . . . be Bernoulli(p) random variables and let B1 , B2 , . . . be Bernoulli(q ) random
variables. Further assume the Aj and Bk variables collectively are mutually independent.
The variable X may be viewed as the number of the first Aj that produces a result of 1
and the variable Y may be viewed similarly for the Bk sequence.

(a) Let Cj be a random variable that is 1 if either Aj = 1 or Bj = 1 (or both), and is

equal to 0 otherwise. Prove that Cj ∼ Bernoulli(r ) for some quantity r and calculate
the value of r in terms of p and q.

(b) Explain why the sequence C1 , C2 , . . . are mutually independent random variables.

(c) Let Z be the random variable that equals the number of the first Cj that results in a
1 and explain why Z is the smaller of X and Y .

(d) Use (c) to conclude that Z ∼ Geometric(r ) for the value of r calculated in part (a).

Ex. 3.3.16. Each day during the hatching season along the Odisha and Northern Tamil
Nadu coast line a Poisson (λ) number of turtle eggs hatch giving birth to young turtles. As
these turtles swim into the sea the probability that they will survive each day is p. Assume

Version: – November 19, 2024

 3.3 functions of random variables 91

that number of hatchings on each day and the life of the turtles born are all independent.
Let X1 = 0 and for i ≥ 2, Xi be the total number of turtles alive at sea on the ith morning
of the hatching season before the hatchings on the i-th day. Find the distribution of Xn .

Version: – November 19, 2024

92 discrete random variables

Version: – November 19, 2024

S U M M A R I Z I N G D I S C R E T E R A N D O M VA R I A B L E S
4
When we first looked at Bernoulli trials in Example 2.1.2 we asked the question “On
average how many successes will there be after n trials?” In order to answer this question,
a specific definition of “average” must be developed.
To begin, consider how to extend the basic notion of the average of a list of numbers to
the situation of equally likely outcomes. For instance, if we want to know what the average
roll of a die will be, it makes sense to declare it to be 3.5, the average value of 1, 2, 3, 4, 5,
and 6. A motivation for a more general definition of average comes from a rewriting of this
calculation.

1+2+3+4+5+6 1 1 1 1 1 1
= 1( ) + 2( ) + 3( ) + 4( ) + 5( ) + 6( ).
6 6 6 6 6 6 6
From the perspective of the right hand side of the equation, the results of all outcomes
are added together after being weighted, each according to its probability. In the case of a
die, all six outcomes have probability 16 .

4.1 expected value

Definition 4.1.1. Let X : S → T be a discrete random variable (so T is countable).

Then the expected value (or average) of X is written as E [X ] and is given by
X
E [X ] = t · P (X = t)
t∈T

provided that the sum converges absolutely. In this case we say that X has “finite
expectation”. If the sum diverges to ±∞ we say the random variable has infinite
expectation. If the sum diverges, but not to infinity, we say the expected value is
undefined.

Example 4.1.2. In the previous chapter, Example 3.1.4 described a lottery for which a
ticket could be worth nothing, or it could be worth either $20 or $200. What is the average
value of such a ticket?

93

Version: – November 19, 2024

94 summarizing discrete random variables

We calculated the distribution of ticket values as P (X = 200) = 1000 ,

1
P (X = 20) =
1000 , and P (X = 0) = 1000 . Applying the definition of expected value results in
27 972

1 27 972
E [X ] = 200( ) + 20( ) + 0( ) = 0.74,
1000 1000 1000

so a ticket has an expected value of 56 cents. ■

It is possible to think of a constant as a random variable. If c ∈ R then we could define
a random variable X with a distribution such that P (X = c) = 1. It is a slight abuse of
notation, but in this case we will simply write c for both the real number as well as the
constant random variable. Such random variables have the obvious expected value.

Theorem 4.1.3. Let c be a real number. Then E [c] = c.

Proof - By definition E [c] is a sum over all possible values of c, but in this case that is just
a single value, so E [c] = c · P (c = c) = c · 1 = c. ■
When the range of X is finite, E [X ] always exists since it is a finite sum. When the
range of X is infinite there is a possibility that the infinite series will not be absolutely
convergent and therefore that E [X ] will be infinite or undefined. In fact, when proving
theorems about how expected values behave, most of the complications arise from the fact
that one must know that an infinite sum converges absolutely in order to rearrange terms
within that sum with equality. The next examples explore ways in which expected values
may misbehave.

Example 4.1.4. Suppose X is a random variable taking values in the range T =

{2, 4, 8, 16, . . . } such that P (X = 2n ) = 1
2n for all integers n ≥ 1.
This is the distribution of a random variable since
∞ ∞
1
P (X = 2n ) = = 1.
X X

n=1 n=1
2n

But note that ∞ ∞ ∞

1
2n · P (X = 2n ) = 2n 1
X X X
=
n=1 n=1
2n n=1

which diverges to infinity, so this random variable has an infinite expected value. ■
Example 4.1.5. Suppose X is a random variable taking values in the range T =
{−2, 4, −8, 16, . . . } such that P (X = (−2)n ) = 1
2n for all integers n ≥ 1.

∞ ∞ ∞
1
(−2)n · P (X = 2n ) = (−1)n .
X X X
(−2)n =
n=1 n=1
2n n=1

Version: – November 19, 2024

 4.1 expected value 95

This infinite sum diverges (not to ±∞), so the expected value of this random variable is
undefined. ■

The examples above were specifically constructed to produce series which clearly
diverged, but in general it can be complicated to check whether an infinite sum is absolutely
convergent or not. The next technical lemma provides a condition that is often simpler
to check. The convenience of this lemma is that, since |X| is always positive, the terms
of the series for E [|X|] may be freely rearranged without changing the value of (or the
convergence of) the sum.

Lemma 4.1.6. E [X ] is a real number if and only if E [|X|] < ∞.

Proof - Let T be the range of X. So U = {|t| : t ∈ T } is the range of |X|. By definition

u · P (|X| = u), while

X
E [|X|] =
u∈U

t · P (X = t).
X
E [X ] =
t∈T

To more easilly relate these two sums, define T̂ = {t : |t| ∈ U }. Since every u ∈ U came
from some t ∈ T the new set T̂ contains every element of T . For every t ∈ T̂ for which
/ T , the element is outside of the range of X and so P (X = t) = 0 for such elements.
t∈
Because of this E [X ] may be written as
X
E [X ] = t · P (X = t)
t∈T̂

since any additional terms in the series are zero.

Note that for each u ∈ U , the event (|X| = u) is equal to (X = u) ∪ (X = −u) where
each of u and −u is an element of T̂ . Therefore,

u · P (U = u) = u · (P (X = u) + P (X = −u))
= u · P (X = u) + u · P (X = −u)
= |u| · P (X = u) + | − u| · P (X = −u)

Version: – November 19, 2024

96 summarizing discrete random variables

(When u = 0 the quantities P (|X| = 0) and P (X = 0) + P (X = −0) are typically not

equal, but the equation is still true since both sides of the equation are zero). Summing
over all u ∈ U then yields
X X
u · P (|X| = u) = |u| · P (X = u) + | − u| · P (X = −u)
u∈U u∈U
X
= |t| · P (X = t)
t∈T̂
X
= |t · P (X = t)|.
t∈T

Therefore the series describing E [X ] is absolutely convergent exactly when E [|X|] < ∞. ■

4.1.1 Properties of the Expected Value

We will eventually wish to calculate the expected values of functions of multiple random
variables. Of particular interest to statistics is an understanding of expected values of sums
and averages of i.i.d. sequences. That understanding will be made easier by first learning
something about how expected values behave for simple combinations of variables.

Theorem 4.1.7. Suppose that X and Y are discrete random variables, both with
finite expected value and both defined on the same sample space S. If a and b are
real numbers then

(1) E [aX ] = aE [X ];

(2) E [X + Y ] = E [X ] + E [Y ]; and

(3) E [aX + bY ] = aE [X ] + bE [Y ].

(4) If X ≥ 0 then E [X ] ≥ 0.

Proof of (1) - If a = 0 then both sides of the equation are zero, so assume a ̸= 0. We know
that X is a function from S to some range U . So aX is also a random variable and its
range is T = {au : u ∈ U }.

Version: – November 19, 2024

 4.1 expected value 97

By definition E [aX ] = t · P (aX = t), but because of how T is defined, adding values
P
t∈T
indexed by t ∈ T is equivalent to adding values indexed by u ∈ U where t = au. In other
words
X
E [aX ] = t · P (aX = t)
t∈T
X
= au · P (aX = au)
u∈U
X
= a· u · P (X = u)
u∈U
= aE [X ].

Proof of (2) - We are assuming that X and Y have the same domain, but they typically
have different ranges. Suppose X : S → U and Y : S → V . Then the random variable
X + Y is also defined on S and takes values in T = {u + v : u ∈ U , v ∈ V }. Therefore,
adding values indexed by t ∈ T is equivalent to adding values indexed by u and v as they
range over U and V respectively. So,
X
E [X + Y ] = t · P (X + Y = t)
t∈T
X
= (u + v ) · P (X = u, Y = v )
u∈U ,v∈V
X X
= (u + v ) · P (X = u, Y = v )
u∈U v∈V
X X X X
= u · P (X = u, Y = v ) + v · P (X = u, Y = v )
u∈U v∈V u∈U v∈V
X X X X
= u · P (X = u, Y = v ) + v · P (X = u, Y = v )
u∈U v∈V v∈V u∈U

where the rearrangement of summation is legitimate since the series converges absolutely.
Notice that as u ranges over all of U the sets (X = u, Y = v ) partition the set (Y = v )
into disjoint pieces based on the value of X. Likewise the event (X = u) is partitioned by
(X = u, Y = v ) as v ranges over all values of v ∈ V . Therefore, as a disjoint union,

and (X = u, Y = v ),
[ [
(Y = v ) = (X = u, Y = v ) (X = u) =
u∈U v∈V

and so

P (X = u, Y = v ) and P (X = u) = P (X = u, Y = v ).
X X
P (Y = v ) =
u∈U v∈V

Version: – November 19, 2024

98 summarizing discrete random variables

From there the proof may be completed, since

X X X X
E [X + Y ] = u P (X = u, Y = v ) + v P (X = u, Y = v )
u∈U v∈V v∈V u∈U
X X
= u · P (X = u) + v · P (Y = v )
u∈U v∈V
= E [X ] + E [Y ].

Proof of (3) - This is an easy consequence of (1) and (2). From (2) the expected value
E [aX + bY ] may be rewritten as E [aX ] + E [bY ]. From there, applying (1) shows this is
also equal to aE [X ] + bE [Y ]. (Using induction this theorem may be extended to any finite
line ar combination of random variables, a fact which we leave as an exercise below).
Proof of (4) - We know that X is a function from S to T where t ∈ T implies that t ≥ 0.
As,

t · P (X = t),
X
E [X ] =
t∈T

it follows by definition of series (in the case T is countable) that E [X ] ≥ 0. ■

Example 4.1.8. What is the average value of the sum of a pair of dice?
To answer this question by appealing to the definition of expected value would require
summing over the eleven possible outcomes {2, 3, . . . , 12} and computing the probabilities
of each of those outcomes. Theorem 4.1.7 makes things much simpler. We began this
section by noting that a single die roll has an expected value of 3.5. The sum of two dice
is X + Y where each of X and Y represents the outcome of a single die. So the average
value of the sum of a pair of dice is E [X + Y ] = E [X ] + E [Y ] = 3.5 + 3.5 = 7. ■
Example 4.1.9. Consider a game in which a player might either gain or lose money based
on the result. A game is considered “fair” if it is described by a random variable with an
expected value of zero. Such a game is fair in the sense that, on average, the player will
have no net change in money after playing.
Suppose a particular game is played with one player (the roller) throwing a die. If the
die comes up an even number, the roller wins that dollar amount from his opponent. If
the die is odd, the roller wins nothing. Obviously the game as stated is not “fair” since
the roller cannot lose money and may win something. How much should the roller pay his
opponent to play this game in order to make it a fair game?
Let X be the amount of money the rolling player gains by the result on the die. The
set of possible outcomes is T = {0, 2, 4, 6} and it should be routine at this point to verify
that E [X ] = 2. Let c be the amount of money the roller should pay to play in order to

Version: – November 19, 2024

 4.1 expected value 99

make the game fair. Since X is the amount of money gained by the roll, the net change of
money for the roller is X − c after accounting for how much was paid to play. A fair game
requires
0 = E [X − c] = E [X ] − E [c] = 2 − c.

So the roller should pay his opponent $2 to make the game fair. ■

4.1.2 Expected Value of a Product

Theorem 4.1.7 showed that E [X + Y ] = E [X ] + E [Y ]. It is natural to ask whether a

similar rule exists for the product of variables. While it is not generally the case that the
expected value of a product is the product of the expected values, if X and Y happen to
be independent, the result is true.

Theorem 4.1.10. Suppose that X and Y are discrete random variables, both with
finite expected value and both defined on the same sample space S. If X and Y are
independent, then E [XY ] = E [X ]E [Y ].

Proof - Suppose X : S → U and Y : S → V . Then the random variable XY takes values

in T = {uv : u ∈ U , v ∈ V }. So,
X
E [XY ] = t · P (XY = t)
t∈T
X X
= (uv ) · P (X = u, Y = v )
u∈U v∈V
X X
= (uv ) · P (X = u)P (Y = v )
u∈U v∈V
X X
= u · P (X = u) v · P (Y = v )
u∈U v∈V
! !
X X
= u · P (X = u) v · P (Y = v )
u∈U v∈V
= E [X ]E [Y ].

Before showing an example of how this theorem might be used, we provide a demon-
stration that the result will not typically hold without the assumption of independence.

Example 4.1.11. Let X ∼ Uniform({1, 2, 3}) and let Y = 4 − X. It is easy to verify Y ∼

Uniform({1, 2, 3}) as well, but X and Y are certainly dependent. A routine computation
shows E [X ] = E [Y ] = 2, and so E [X ]E [Y ] = 4.

Version: – November 19, 2024

100 summarizing discrete random variables

However, the random variable XY can only take on two possible values. It may equal
3 (if either X = 1 and Y = 3 or vica versa) or it may equal 4 (if X = Y = 2). So,
P (XY = 3) = 2
3 and P (XY = 4) = 31 . Therefore,

2 1 10
E [XY ] = 3( ) + 4( ) = ̸= 4.
3 3 3

The conclusion of Theorem 4.1.10 fails since X and Y are dependent. ■

Example 4.1.12. Suppose an insurance company assumes that, for a given month, both
the number of customer claims X and the average cost per claim Y are independent
random variables. Suppose further the company is able to estimate that E [X ] = 100 and
E [Y ] = $1, 250. How should the company estimate the total cost of all claims that month?
The total cost should be the number of claims times the average cost per claim, or XY .
Using Theorem 4.1.10 the expected value of XY is simply the product of the separate
expected values.

E [XY ] = E [X ]E [Y ] = 100 · $1, 250 = $125, 000.

Notice, though, that the assumption of independence played a critical role in this computa-
tion. Such an assumption might not be valid for many practical problems. Consider, for
example, if a weather event such as a tornado tends to cause both a larger-than-average
number of claims and also a larger-than-average value per claim. This could cause the
variables X and Y to be dependent and, in such a case, estimating the total cost would
not be as simple as taking the product of the separate expected values. ■

4.1.3 Expected Values of Common Distributions

A quick glance at the definition of expected value shows that it only depends on the
distribution of the random variable. Therefore one can compute the expected values for
the various common distributions we defined in the previous chapter.

Example 4.1.13. (Expected Value of a Bernoulli(p))

Let X ∼ Bernoulli(p). So P (X = 0) = 1 − p and P (X = 1) = p.
Therefore E [X ] = 0(1 − p) + 1(p) = p. ■

Example 4.1.14. (Expected Value of a Binomial(n,p))

We will show two ways to calculate this expected value – the first is more computationally
complicated, but follows from the definition of the Binomial distribution directly; the

Version: – November 19, 2024

 4.1 expected value 101

second is simpler, but requires using the relationship between the Binomial and Bernoulli
random variables. In algebraic terms, if Y ∼ Binomial(n, p) then
n
X
E [Y ] = k · P (Y = k )
k =0
n
!
n k
p (1 − p)n−k
X
= k·
k =1
k
n
n!
pk (1 − p)n−k
X
= k·
k =1
k!(n − k )!
n
(n − 1) !
pk−1 (1 − p)(n−1)−(k−1)
X
= np ·
k =1
( k − 1 ) ! (( n − 1 ) − ( k − 1 )) !
n
n − 1 k−1
!
p (1 − p)(n−1)−(k−1)
X
= np ·
k =1
k−1
n−1
n−1 k
!
p (1 − p)(n−1)−k
X
= np ·
k =0
k

where the last equality is a shift of variables. But now, by the binomial theorem, the sum
n−1
(n−1
k )p (1 − p)
k (n−1)−k is equal to 1 and therefore E [Y ] = np.
P
k =0

Alternatively, recall that the Binomial distribution first came about as the total number
of successes in n independent Bernoulli trials. Therefore a Binomial(n, p) distribution results
from adding together n independent Bernoulli(p) random variables. Let X1 , X2 , . . . , Xn
be i.i.d. Bernoulli(p) and let Y = X1 + X2 + · · · + Xn . Then Y ∼ Binomial(n, p) and

E [Y ] = E [X1 + X2 + · · · + Xn ]
= E [X1 ] + E [X2 ] + · · · + E [Xn ]
= p + p + · · · + p = np.

This also provides the answer to part (d) of Example 2.1.2. The expected number of
successes in a series of n independent Bernoulli(p) trials is np. ■

In the next example we will calculate the expected value of a geometric random variable.
The computation illustrates a common technique from calculus for simplifying power series
by differentiating the sum term-by-term in order to rewrite a complicated series in a simpler
way.

Example 4.1.15. (Expected Value of a Geometric(p))

Version: – November 19, 2024

102 summarizing discrete random variables

If X ∼ Geometric(p) and 0 < p < 1, then

∞
k · p(1 − p)k−1
X
E [X ] =
k =1

To evaluate the sum of the series we will need to work the partial sums of the same. For
any n ≥ 1, let
n n
kp(1 − p)k−1 = k (1 − (1 − p))(1 − p)k−1
X X
Tn =
k =1 k =1
n n
k (1 − p)k−1 − k (1 − p)k
X X
=
k =1 k =1
n
1 − (1 − p)n
(1 − p)k−1 − n(1 − p)n = − n(1 − p)n .
X
=
k =1
p

Using standard results from analysis we know that for 0 < p < 1,

lim (1 − p)n = 0 and lim n(1 − p)n = 0.

n→∞ n→∞

Therefore Tn → 1
p as n → ∞. Hence

1
E [X ] = .
p

For instance, suppose we wanted to know on average how many rolls of a die it would
take before we observed a 5. Each roll is a Bernoulli trial with a probability 1
6 of success.
The time it takes to observe the first success is distributed as a Geometric( 16 ) and so
has expected value 1
1/6 = 6. On average it should take six rolls before observing this
outcome. ■

Example 4.1.16. (Expected Value of a Poisson(λ))

We can make a reasonable guess at the expected value of a Poisson(λ) random variable
by recalling that such a distribution was created to approximate a Binomial when n was
large and p was small. The parameter λ = np remained fixed as we took a limit. Since we
showed above that a Binomial (n, p) has an expected value of np, it seems plausible that
a P oisson(λ) should have an expected value of λ. This is indeed true and it is possible
to prove the fact by using the idea that the Poisson random variable is the limit of a
sequence of Binomial random variables. However, this proof requires an understanding of
how limits and expected values interact, a concept that has not yet been introduced in the
text. Instead we leave a proof based on a direct algebraic computation as Exercise 4.1.12.

Version: – November 19, 2024

 4.1 expected value 103

Taking the result as a given, we will illustrate how this expected value might be used
for an applied problem. Suppose an insurance company wants to model catastrophic floods
using a Poisson(λ) random variable. Since floods are rare in any given year, and since
the company is considering what might occur over a long span of years, this may be a
reasonable assumption.

As its name implies, a “50-year flood” is a flood so substantial that it should occur, on
average, only once every fifty years. However, this is just an average; it may be possible to
have two “50-year floods” in consecutive years, though such an event would be quite rare.
Suppose the insurance company wants to know how likely it is that there will be two or
more “50-year floods” in the next decade, how should this be calculated?

There is an average of one such flood every fifty years, so by proportional reasoning, in
the next ten years there should be an average of 0.2 floods. In other words, the number of
floods in the next ten years should a random variable X ∼ P oisson(0.2) and we wish to
calculate P (X ≥ 2).

P (X ≥ 2) = 1 − P (X = 0) − P (X = 1)
= 1 − e−0.2 − e−0.2 (0.2)
≈ 0.0002.

So assuming the Poisson random variable is an accurate model, there is only about a 0.02%
chance that two or more such disastrous floods would occur in the next decade. ■

For a Hypergeometric random variable, we will demonstrate another proof technique

common to probability. An expected value may involve a complicated (or infinite) sum
which must be computed. However, this sum includes within it the probabilities of each
outcome of the random variable, and those probabilities must therefore add to 1. It is
sometimes possible to simplify the sum describing the expected value using the fact that a
related sum is already known.

Example 4.1.17. (Expected Value of a HyperGeo(N , r, m)) Let m and r be positive

integers an d let N be an integer for which N > max{m, r}. Let X be a random variable

Version: – November 19, 2024

104 summarizing discrete random variables

with X ∼ HyperGeo(N , r, m). To calculate the expected value of X, we begin with two
facts. The first is an identity involving combinations. If n ≥ k > 0 then
!
n n!
=
k k!(n − k )!
n (n − 1) !
=
k (k − 1)!((n − 1) − (k − 1))!
n n−1
!
= .
k k−1

The second comes from the consideration of the probabilities associated with a HyperGeo(N −
1, r − 1, m − 1) distribution. Specifically, as k ranges over all possible values of such a
distribution, we have
r−1 (N −1)−(r−1)
X ( k )( (m−1)−k )
−1
=1
k (N
m−1)

since this is the sum over all outcomes of the random variable.

To calculate E [X ], let j range over the possible values of X. Recall that the minimum
value of j is max{0, m − (N − r )} and the maximum value of j is min{r, m}. Now let
k = j − 1. This means that the maximum value for k is min{r − 1, m − 1}. If the
minimum value for j was m − (N − r ) then the minimum value for k is m − (N − r ) − 1 =
((m − 1) − ((N − 1) − (r − 1))). If the minimum value for j was 0 then the minimum
value for k is −1.

The key to the computation is to note that as j ranges over all of the values of X, the
values of k cover all possible values of a HyperGeo(N − 1, m − 1, r − 1) distribution. In
fact, the only possible value k may assume that is not in the range of such a distribution is
if k = −1 as a minimum value. Now,

−r
(rj )(N
m−j )
,
X
E [X ] = j·
j (N
m)

and if j = 0 is in the range of X, then that term of the sum is zero and it may be
deleted without affecting the value. That is equivalent to deleting the k = −1 term, so

Version: – November 19, 2024

 4.1 expected value 105

the remaining values of k exactly describe the range of a HyperGeo(N − 1, m − 1, r − 1)

distriubtion. From there, the expected value may be calculated as

−r
X (rj )(N
m−j )
E [X ] = j·
j (N
m)
r r−1 (N −1)−(r−1)
X j (j−1)( (m−1)−(j−1) )
= j· N N −1
j m (m−1)
r−1 (N −1)−(r−1)
rm X (j−1)( (m−1)−(j−1) )
= ( )· −1
N j (N
m−1)
r−1 (N −1)−(r−1)
rm X ( k )( (m−1)−k )
= ( )· −1
N k (Nm−1)
rm rm
= ( ) · (1) = .
N N

This nearly completes the goal of calculating the expected values of hypergeometric
distributions. The only remaining issues are the cases when m = 0 and r = 0. Since the
hypergoemetric distribution was only defined when m and r were non-negative integers,
and since the proof above requires the consideration of such a distribution for the values
m − 1 and r − 1, the remaining cases must be handled separately. However, they are fairly
easy and yield the same result, a fact we leave it to the reader to verify. ■

4.1.4 Expected Value of f (X1 , X2 , . . . , Xn )

As we have seen previously, if X is a random variable and if f is a function defined on the

possible outputs of X, then f (X ) is a random variable in its own right. The expected value
of this new random variable may be computed in the usual way from the distribution of
f (X ), but it is an extremely useful fact that it may also be computed from the distribution
of X itself. The next example and theorems illustrate this fact.

Example 4.1.18. Returning to a setting first seen in Example 3.3.1 we will let X ∼
Uniform({−2, −1, 0, 1, 2}), and let f (x) = x2 . How may E [f (X )] be calculated?
We will demonstrate this in two ways – first by appealing directly to the definition, and
then using the distribution of X instead of the distribution of f (X ). To use the definition
of expected value, recall that f (X ) = X 2 takes values in {0, 1, 4} with the following
probabilities: P (f (X ) = 0) = 1
5 while P (f (X ) = 1) = P (f (X ) = 4) = 25 . Therefore,

1 2 2
E [f (X )] = 0( ) + 1( ) + 4( ) = 2.
5 5 5

Version: – November 19, 2024

106 summarizing discrete random variables

However, the values of f (X ) are completely determined from the values of X. For
instance, the event (f (X ) = 4) had a probability of 2
5 because it was the disjoint union of
two other events (X = 2) ∪ (X = −2), each of which had probability 15 . So the term 4( 25 )
in the computation above could equally well have been thought of in two pieces

4 · P (f (X ) = 4) = 4 · P ((X = 2) ∪ (X = −2))
= 4 · (P (X = 2) + P (X = −2))
= 4 · P (X = 2) + 4 · P (X = −2)
= 22 · P (X = 2) + (−2)2 · P (X = −2),

where the final expression emphasizes that the outcome of 4 resulted either from 22 or
(−2)2 depending on the value of X. Following a similar plan for the other values of f (X )
allows E [f (X )] to be calcualted directly from the probabilities of X as

E [f (X )] = (−2)2 · P (X = −2) + (−1)2 · P (X = −1) + 02 · P (X = 0)

+12 · P (X = 1) + 22 · P (X = 2)
1 1 1 1 1
= 4( ) + 1( ) + 0( ) + 1( ) + 4( )
5 5 5 5 5
= 2,

which gives the same result as the previous computation. ■

The technique of the example above works for any functions as demonstrated by the
next two theorems. We first state and prove a version for functions of a single random
variable and then deal with the multivariate case.

Theorem 4.1.19. Let X : S → T be a discrete random variable and define a

function f : T → U . Then the expected value of f (X ) may be computed as

f (t) · P (X = t).
X
E [f (X )] =
t∈T

Proof - By definition E [f (X )] = u · P (f (X ) = u). However, as in the previous

P
u∈U
example, the event (f (X ) = u) may be partitioned according to the input values of X
which cause f (X ) to equal u. Recall that f −1 (u) describes the set of values in T which,
when input into the function f , produce the value u. That is, f −1 (u) = {t ∈ T : f (t) = u}.
Therefore,
(X = t), and so
[
(f (X ) = u) =
t∈f −1 (u)

Version: – November 19, 2024

 4.1 expected value 107

P (X = t).
X
P (f (X ) = u) =
t∈f −1 (u)

Putting this together with the definition of E [f (X )] shows

X
E [f (X )] = u · P (f (X ) = u)
u∈U
X X
= u· P (X = t)
u∈U t∈f −1 (u)
X X
= u · P (X = t)
u∈U t∈f −1 (u)
X X
= f (t) · P (X = t)
u∈U t∈f −1 (u)

f (t) · P (X = t),
X
=
t∈T

where the final step is simply the fact that T = f −1 (U ) and so summing over the values of
t ∈ T is equivalent to grouping them together in the sets f −1 (u) and summing over all
values in U that may be achieved by f (X ). ■

Theorem 4.1.20. Let X1 , X2 , . . . Xn be random variables defined on a common

sample space S. The Xj variables may have different ranges, so let Xj : S → Tj .
Let f be a function defined for all possible outputs of the Xj variables. Then

f (t1 , . . . , tn ) · P (X1 = t1 , . . . , Xn = tn ).
X
E [f (X )] =
t1 ∈T1 ,... tn ∈Tn

The proof is nearly the same as for the one-variable case. The only diference is that f −1 (u)
is now a set of vectors of values (t1 , . . . , tn ), so that the event (f (X ) = u) decomposes into
events of the form (X1 = t1 , . . . , Xn = tn ). However, this change does not interfere with
the logic of the proof. We leave the details to the reader.

exercises

Ex. 4.1.1. Let X, Y be discrete random variables. If X ≤ Y , then show that E [X ] ≤ E [Y ].

Ex. 4.1.2. A lottery is held every day, and on any given day there is a 30% chance that
someone will win, with each day independent of every other. Let X denote the random
variable describing the number of times in the next five days that the lottery will be won.

Version: – November 19, 2024

108 summarizing discrete random variables

(a) What type of random variable (with what parameter) is X?

(b) On average (expected value), how many times in the next five days will the lottery
be won?

(c) When the lottery occurs for each of the next five days, what is the most likely number
(mode) of days there will be a winner?

(d) How likely is it the lottery will be won in either one or two of the next five days?

Ex. 4.1.3. A game show contestant is asked a series of questions. She has a probability of
0.88 of knowing the answer to any given question, independently of every other. Let Y
denote the random variable describing the number of questions asked until the contestant
does not know the correct answer.

(a) What type of random variable (with what parameter) is Y ?

(b) On average (expected value), how many questions will be asked until the first question
for which the contestant does not know the answer?

(c) What is the most likely number of questions (mode) that will be asked until the
contestant does not know a correct answer?

(d) If the contestant is able to answer twelve questions in a row, she will win the grand
prize. How likely is it that she will know the answers to all twelve questions?

Ex. 4.1.4. Sonia sends out invitations to eleven of her friends to join her on a hike she’s
planning. She knows that each of her friends has a 59% chance of deciding to join her
independently of each other. Let Z denote the number of friends who join her on the hike.

(a) What type of random variable (with what parameter) is Z?

(b) What is the average (expected value) number of her friends that will join her on the
hike?

(c) What is the most likely number (mode) of her friends that will join her on the hike?

(d) How do your answers to (b) and (c) change if each friend has only a 41% chance of
joining her?

Ex. 4.1.5. A player rolls three dice and earns $1 for each die that shows a 6. How much
should the player pay to make this a fair game?

Version: – November 19, 2024

 4.1 expected value 109

Ex. 4.1.6. (“The [Link] Paradox”) Suppose a game is played whereby a player
begins flipping a fair coin and continues flipping it until it comes up heads. At that time
the player wins a 2n dollars where n is the total number of times he flipped the coin. Show
that there is no amount of money the player could pay to make this a fair game. (Hint:
See Example 4.1.4).
Ex. 4.1.7. Two different investment strategies have the following probabilities of return on
$10,000.
Strategy A has a 20% chance of returning $14,000, a 35% chance of returning $12,000,
a 20% chance of returning $10,000, a 15% chance of returning $8,000, and a 10% chance of
returning only $6,000.
Strategy B has a 25% chance of returning $12,000, a 35% chance of returning $11,000,
a 25% chance of returning $10,000, and a 15% chance of returning $9,000.

(a) Which strategy has the larger expected value of return?

(b) Which strategy is more likely to produce a positive return on investment?

(c) Is one strategy clearly preferable to the other? Explain your reasoning.

Ex. 4.1.8. Calculate the expected value of a Uniform({1, 2, . . . , n}) random variable by
following the steps below.
n
n2 + n
(a) Prove the numerical fact that 2 . (Hint: There are many methods to do
P
j=
j =1
this. One uses induction).

(b) Use (a) to show that if X ∼ Uniform({1, 2, . . . , n}), then E [X ] = 2 .

n+1

Ex. 4.1.9. Use induction to extend the result of Theorem 4.1.7 by proving the following:
If X1 , X2 , . . . , Xn are random variables with finite expectation all defined on the same
sample space S and if a1 , a2 , . . . an are real numbers, then

E [a1 X1 + a2 X2 + · · · + an Xn ] = a1 E [X1 ] + a2 E [X2 ] + · · · + an E [Xn ].

Ex. 4.1.10. Suppose X and Y are random variables for which X has finitie expected value
and Y has infinite expected value. Prove that X + Y has infinite expected value.
Ex. 4.1.11. Suppose X and Y are random variables. Suppose E [X ] = ∞ and E [Y ] = −∞.

(a) Provide an example to show that E [X + Y ] = ∞ is possible.

(b) Provide an example to show that E [X + Y ] = −∞ is possible.

Version: – November 19, 2024

110 summarizing discrete random variables

(c) Provide an example to show that E [X + Y ] may have finite expected value.

Ex. 4.1.12. Let X ∼ P oisson(λ).

(a) Write an expression for E [X ] as an infinite sum.

(b) Every non-zero term in your answer to (a) should have a λ in it. Factor this λ out
and explain why the remaining sum equals 1. (Hint: One way to do this is through
the use of infinite series. Another way is to use the idea from Example 4.1.17).

Ex. 4.1.13. A daily lottery is an event that many people play, but for which the likelihood
of any given person winning is very small, making a Poisson approximation appropriate.
Suppose a daily lottery has, on average, two winners every five weeks. Estimate the
probability that next week there will be more than one winner.

4.2 variance and standard deviation

As a single number, the average of a random variable may or may not be a good approxi-
mation of the values that variable is likely to produce. For example, let X be defined such
that P (X = 10) = 1, let Y be defined so that P (Y = 9) = P (Y = 11) = 12 , and let Z be
defined such that P (Z = 0) = P (Z = 20) = 21 . It is easy to check that all three of these
random variables have an expected value of 10. However the number 10 exactly describes
X, is always off from Y by an absolute value of 1 and is always off from Z by an absolute
value of 10.
It is useful to be able to quantify how far away a random variable typically is from
its average. Put another way, if we think of the expected value as somehow measuring
the “center” of the random variable, we would like to find a way to measure the size of the
“spread” of the variable about its center. Quantities useful for this are the variance and
standard deviation.

Definition 4.2.1. Let X be a random variable with finite expected value. Then the
variance of the random variable is written as V ar [X ] and is defined as

V ar [X ] = E [(X − E [X ])2 ]

The standard deviation of X is written as SD [X ] and is defined as

q
SD [X ] = V ar [X ]

Version: – November 19, 2024

 4.2 variance and standard deviation 111

Notice that V ar [X ] is the average of the square distance of X from its expected value.
So if X has a high probability of being far away from E [X ] the variance will tend to be
large, while if X is very near E [X ] with high probability the variance will tend to be small.
In either case the variance is the expected value of a squared quantity, and as such is
always non-negative. Therefore SD [X ] is defined whenever V ar [X ] is defined.
If we were to associate units with the random variable X (say meters), then the units
of V ar [X ] would be meters2 and the units of SD [X ] would be meters. We will see that
the standard deviation is more meaningful as a measure of the “spread” of a random
variable while the variance tends to be a more useful quantity to consider when carrying
out complex computations.
Informally we will view the standard deviation as a typical distance from average. So
if X is a random variable and we calculate that E [X ] = 12 and SD [X ] = 3, we might
say, “The variable X will typically take on values that are in or near the range 9 − 15,
one standard deviation either side of the average”. A goal of this section is to make that
language more precise, but at this point it will help with intuition to understand this
informal view.
The variance and standard deviation are described in terms of the expected value.
Therefore V ar [X ] and SD [X ] can only be defined if E [X ] exists as a real number. However,
it is possible that V ar [X ] and SD [X ] could be infinite even if E [X ] is finite (see Exercises).
In practical terms, if X has a finite expected value and infinite standard deviation, it
means that the random variable has a clear average, but is so spread out that any finite
number underestimates the typical distance of the random variable from its average.
Example 4.2.2. As above, let X be a constant varaible with P (X = 10) = 1. Let Y be such
that P (Y = 9) = P (Y = 11) = 1
2 and let Z be such that P (Z = 0) = P (Z = 20) = 12 .
Since X always equals E [X ], the quantity (X − E [X ])2 is always zero and we can
conclude that V ar [X ] = 0 and SD [X ] = 0. This makes sense given the view of SD [X ] as
an estimate of how spread out the variable is. Since X is constant it is not at all spread
out and so SD [X ] = 0.
To calculate V ar [Y ] we note that (Y − E [Y ])2 is always equal to 1. Therefore V ar [Y ] =
1 and SD [Y ] = 1. Again this reaffirms the informal description of the standard deviation;
the typical distance between Y and its average is 1.
Likewise (Z − E [Z ])2 is always equal to 100. Therefore V ar [Z ] = 100 and SD [Z ] = 10.
The typical distance between Z and its average is 10. ■
Example 4.2.3. What are the variance and standard deviation of a die roll?
Before we carry out the calculation, let us use the informal idea of standard deviation
to estimate an answer and help build intuition. We know the average of a die roll is 3.5.
The closest a die could possibly be to this average is 0.5 (if it were to roll a 3 or a 4) and

Version: – November 19, 2024

112 summarizing discrete random variables

the furthest it could possibly be is 2.5 (if it were to roll a 1 or a 6). Therefore the standard
deviation, a typical distance from average, should be somewhere between 0.5 and 2.5.
To calculate the quantity exactly, let X represent the roll of a die. By definition,
V ar [X ] = E [(X − 3.5)2 ], and the values that (X − 3.5)2 may assume are determined by
the six values X may take on.

V ar [X ] = E [(X − 3.5)2 ]
1 1 1 1 1 1
= (2.5)2 + (1.5)2 + (0.5)2 + (−0.5)2 + (−1.5)2 + (−2.5)2
6 6 6 6 6 6
35
= .
12
q
So, SD [X ] = 35
12 ≈ 1.71 which is near the midpoint of the range of our estimate above. ■

4.2.1 Properties of Variance and Standard Deviation

Theorem 4.2.4. Let a ∈ R and let X be a random variable with finite variance
(and thus, with finite expected value as well). Then,

(a) V ar [aX ] = a2 · V ar [X ];

(b) SD [aX ] = |a| · SD [X ];

(c) V ar [X + a] = V ar [X ]; and

(d) SD [X + a] = SD [X ].

Proof of (a) and (b) - V ar [aX ] = E [(aX − E [aX ])2 ]. Using known properties of expected
value this may be rewritten as

V ar [aX ] = E [(aX − aE [X ])2 ]

= E [a2 (X − E [X ])2 ]
= a2 E [(X − E [X ])2 ]
= a2 V ar [X ].

That concludes the proof of (a). The result from (b) follows by taking square roots of both
sides of this equation.
Proof of (c) and (d) - (See Exercises) ■

Version: – November 19, 2024

 4.2 variance and standard deviation 113

The variance may also be computed using a different (but equivalent) formula if E [X ]
and E [X 2 ] are known.

Theorem 4.2.5. Let X be a random variable for which E [X ] and E [X 2 ] are both
finite. Then
V ar [X ] = E [X 2 ] − (E [X ])2 .

Proof -

V ar [X ] = E [(X − E [X ])2 ]
= E [X 2 − 2XE [X ] + (E [X ])2 ]
= E [X 2 ] − 2E [XE [X ]] + E [(E [X ])2 ].

But E [X ] is a constant, so

V ar [X ] = E [X 2 ] − 2E [XE [X ]] + E [(E [X ])2 ]

= E [X 2 ] − 2E [X ]E [X ] + (E [X ])2
= E [X 2 ] − (E [X ])2 .

■
In statistics we frequently want to consider the sum or average of many random variables.
As such it is useful to know how the variance of a sum relates to the variances of each
variable separately. Toward that goal we have

Theorem 4.2.6. If X and Y are independent random variables, both with finite
expectation and finite variance, then

(a) V ar [X + Y ] = V ar [X ] + V ar [Y ]; and
q
(b) SD [X + Y ] = (SD [X ])2 + (SD [Y ])2 .

Proof - Using Theorem 4.2.5,

V ar [X + Y ] = E [(X + Y )2 ] − (E [X + Y ])2
 
= E [X 2 + 2XY + Y 2 ] − (E [X ])2 + 2E [X ]E [Y ] + (E [Y ])2
= E [X 2 ] + 2E [XY ] + E [Y 2 ] − (E [X ])2 − 2E [X ]E [Y ] − (E [Y ])2 .

Version: – November 19, 2024

114 summarizing discrete random variables

But by Theorem 4.1.10, E [XY ] = E [X ]E [Y ] since X and Y are independent. So,

V ar [X + Y ] = E [X 2 ] − (E [X ])2 + E [Y 2 ] − (E [Y ])2
= V ar [X ] + V ar [Y ].

Part (b) follows immediately after rewriting the variances in terms of standard deviations
and taking square roots. As with expected values, this theorem may be generalized to a
sum of any finite number of independent random variables using induction. The proof of
that fact is left as Exercise 4.2.11. ■
Example 4.2.7. What is the standard deviation of the sum of two dice?
We previously found that if X represents one die, then V ar [X ] = 12 . If X
35
and Y are
two independent
q dice, then V ar [X + Y ] = V ar [X ] + V ar [Y ] =
35
12 + 12 = 6 .
35 35
Therefore
SD [X + Y ] = 35
6 ≈ 2.42. ■

4.2.2 Variances of Common Distributions

As with expected value, the variances of the common discrete random variables can be
calculated from their corresponding distributions.
Example 4.2.8. (Variance of a Bernoulli(p))
Let X ∼ Bernoulli(p). We have already calculated that E [X ] = p. Since X only takes
on the values 0 or 1 it is always true that X 2 = X. Therefore E [X 2 ] = E [X ] = p.
So, V ar [X ] = E [X 2 ] − (E [X ])2 = p − p2 = p(1 − p). ■
Example 4.2.9. (Variance of a Binomial(n,p))
We will calculate the variance of a Binomial random variable using the fact that it may
be viewed as the sum of n independent Bernoulli random variables. A strictly algebraic
computation is also possible (see Exercises).
Let X1 , X2 , . . . , Xn be independent Bernoulli(p) random variables. Therefore, if
Y = X1 + X2 + · · · + Xn then Y ∼ Binomial (n, p) and

V ar [Y ] = V ar [X1 + X2 + · · · + Xn ]
= V ar [X1 ] + V ar [X2 ] + · · · + V ar [Xn ]
= p(1 − p) + p(1 − p) + · · · + p(1 − p)
= np(1 − p).

For an application of this computation we return to the idea of sampling from a

population where some members of the population have a certain characteristic and others

Version: – November 19, 2024

 4.2 variance and standard deviation 115

do not. The goal is to provide an estimate of the number of people in the sample that have
the characteristic. For this example, suppose we were to randomly select 100 people from
a large city in which 20% of the population works in a service industry. How many of the
100 people from our sample should we expect to be service industry workers?
If the sampling is done without replacement (so we cannot pick the same person twice),
then strictly speaking the desired number would be described by a Hypergeometric random
variable. However, we have also seen that there is little difference between the Binomial
and Hypergeometric distributions when the size of the sample is small relative to the size
of the population. So since the sample is only 100 people from a “large city”, we will
assume this situation is modeled by a binomial random variable. Specifically, since 20% of
the population consits of service workers, we will assume X ∼ Binomial (100, 0.2).
The simplest way to answer to the question of how many service industy workers to
expect within the sample is to compute the expected value of X. In this case E [X ] =
100(0.2) = 20, so we should expect around 20 of the 100 people in the sample to be service
workers. However, this is an incomplete answer to the question since it only provides an
average value; the actual number of service workers in the sample is probably not going to
be exactly 20, it’s only likely to be around 20 on average. A more complete answer to the
question would give an estimate as to how far away from 20 the actual value is likely to
be. But this is precisely what the standard deviation describes – an estimate of the likely
difference between the actual result of the random variable and its expected value.
√
In this case V ar [X ] = 100(0.2)(0.8) = 16 and so SD [X ] = 16 = 4. This means that
the actual number of service industry workers in the sample will typically be about 4 or so
away from the expected value of 20, so a more complete answer to the question would be
“The sample is likely to have around 16 − 24 service workers in it”. That is not to say that
the actual number of service workers is guaranteed to fall in the that range, but the range
provides s a sort of likely error associated with the estimate of 20. Results in the 16 − 24
range should be considered fairly common. Results far outside that range, while possible,
should be considered fairly unusual. ■
Recall in Example 4.1.17 we calculated E [X ] using a technique in which the sum
describing E [X ] was computed based on another sum which only involved the distribution
of X directly. This second sum equalled 1 since it simply added up the probabilities that
X assumed each of its possible values. In a similar fashion, it is sometimes possible to
calculate a sum describing E [X 2 ] in terms of a sum for E [X ] which is already known. From
that point, Theorem 4.2.5 may be used to calculate the variance and standard deviation of
X. This technique will be illustrated in the next example in which we calculate the spread
associated with a geometric random variable.
Example 4.2.10. (Variance of a Geometric(p))

Version: – November 19, 2024

116 summarizing discrete random variables

Let 0 < p < 1. X ∼ Geometric(p) for which we know E [X ] = p1 . Then,

∞
k 2 p(1 − p)k−1
X
E [X 2 ] =
k =1

To evaluate the sum of the series we will need to work the partial sums of the same. For
any n ≥ 1, let
n n
k 2 p(1 − p)k−1 = k 2 (1 − (1 − p))(1 − p)k−1
X X
Sn =
k =1 k =1
n n
k 2 (1 − p)k−1 − k 2 (1 − p)k
X X
=
k =1 k =1
n
= 1+ (2k − 1)(1 − p)k−1 − n2 (1 − p)n
X

k =2
n n
= 1− (1 − p)k−1 + 2 k (1 − p)k−1 − n2 (1 − p)n
X X

k =2 k =2
n n
= 2− (1 − p)k−1 + 2(−1 + k (1 − p)k−1 ) − n2 (1 − p)n
X X

k =1 k =1
n
1 − (1 − p)n 2
kp(1 − p)k−1 − n2 (1 − p)n
X
= − +
p p k =1

Using standard results from analysis and result from Example 4.1.15 we know that for
0 < p < 1,
n
1
lim kp(1 − p)k−1 = , lim (1 − p)n = 0, and lim n2 (1 − p)n = 0.
X
n→∞
k =1
p n→∞ n→∞

Therefore Sn → − p1 + 2
p2
as n → ∞. Hence

1 2
E [X 2 ] = − + 2 .
p p

Using Theorem 4.2.5 the variance may then be calculated as

V ar [X ] = E [X 2 ] − (E [X ])2
2 1 1
= 2 − − ( )2
p p p
1 1
= 2−
p p

Version: – November 19, 2024

 4.2 variance and standard deviation 117

A similar technique may be used for calculating the variance of a Poisson random
variable, a fact which is left as an exercise. We finish this subsection with a computation
of the variance of a hypergeometric distribution using an idea similar to how we calculated
its expected value in Example 4.1.17.

Example 4.2.11. Let m and r be positive integers and let N be an integer with N >
max{m, r} and let X ∼ HyperGeo(N , r, m). To calculate E [X 2 ], as j ranges over the
values of X,

−r
2
X
2
(rj )(N
m−j )
E [X ] = j ·
j (N
m)
r r−1 (N −1)−(r−1)
X
2 j (j−1)( (m−1)−(j−1) )
= j · N N −1
j m (m−1)
r−1 (N −1)−(r−1)
rm X (j−1)( (m−1)−(j−1) )
= ( ) j· −1
N j (N
m−1)
r−1 (N −1)−(r−1)
rm X ( k )( (m−1)−k )
= ( ) · (k + 1) −1
N k (N
m−1)

where k ranges over the values of Y ∼ HyperGeo(N − 1, r − 1, m − 1). Therefore,

rm
E [X 2 ] = ( )E [Y + 1]
N
rm
= ( )(E [Y ] + 1)
N
rm (r − 1)(m − 1)
= ( )( + 1).
N (N − 1)

Now the variance may be easily computed as

V ar [X ] = E [X 2 ] − (E [X ])2
rm (r − 1)(m − 1) rm 2
= ( )( + 1) − ( )
N (N − 1) N
N 2 rm − N rm2 − N r2 m + r2 m2
= .
N 2 (N − 1)

As with the computation of expected value, the cases of m = 0 and r = 0 must be handled
separately, but yield the same result. ■

Version: – November 19, 2024

118 summarizing discrete random variables

4.2.3 Standardized Variables

Many random variables may be rescaled into a standard format by shifting them so that
they have an average of zero and then rescaling them so that they have a variance (and
standard deviation) of one. We introduce this idea now, though its chief importance will
not be realized until later.

Definition 4.2.12. A standardized random variable X is one for which

E [X ] = 0 and V ar [X ] = 1.

Theorem 4.2.13. Let X be a discrete random variable with finite expected value
X−E [X ]
and finite, non-zero variance. Then Z = SD [X ]
is a standardized random variable.

Proof - The expected value value of Z is

X − E [X ]
E [Z ] = E [ ]
SD [X ]
E [X − E [X ]]
=
SD [X ]
E [X ] − E [X ]
= =0
SD [X ]

while the variance of Z is

X − E [X ]
V ar [Z ] = V ar [ ]
SD [X ]
V ar [X − E [X ]]
=
(SD [X ])2
V ar [X ]
= = 1.
V ar [X ]

For easy reference we finish off this section by providing a chart of values associated
with common discrete distributions.

Version: – November 19, 2024

 4.2 variance and standard deviation 119

Distribution Expected Value Variance

Bernoulli(p) p p(1 − p)
Binomial(n, p) np np(1 − p)
Geometric(p) 1
p
1−p
p2
N 2 rm−N rm2 −N r2 m+r2 m2
HyperGeo(N , r, m) rm
N N 2 (N −1)
Poisson(λ) λ λ
n2 −1
Uniform({1, 2, . . . , n}) n+1
2 12

exercises

Ex. 4.2.1. A random variable X has a probability mass function given by

P (X = 0) = 0.2, P (X = 1) = 0.5, P (X = 2) = 0.2, and P (X = 3) = 0.1.

Calculate the expected value and standard deviation of this random variable. What is the
probability this random variable will produce a result more than one standard deviation
from its expected value?

Ex. 4.2.2. Answer the following questions about flips of a fair coin.

(a) Calculate the standard deviation of the number of heads that show up in 100 flips of
a fair coin.

(b) Show that if the number of coins is quadrupled (to 400) the standard deviation only
doubles.

Ex. 4.2.3. Suppose we begin rolling a die, and let X be the number of rolls needed before
we see the first 3.

(a) Show that E [X ] = 6.

(b) Calculate SD [X ].

(c) Viewing SD [X ] as a typical distance of X from its expected value, would it seem
unusual to roll the die more than nine times before seeing a 3?

(d) Calculate the actual probability P (X > 9).

(e) Calculate the probability X produces a result within one standard deviation of its
expected value.

Version: – November 19, 2024

120 summarizing discrete random variables

Ex. 4.2.4. A key issue in statistical sampling is the determination of how much a sample
is likely to differ from the population it came from. This exercise explores some of these
ideas.

(a) Suppose a large city is exactly 50% women and 50% men and suppose we randomly
select 60 people from this city as part of a sample. Let X be the number of women in
the sample. What are the expected value and standard deviation of X? Given these
values, would it seem unusual if fewer than 45% of the individuals in the sample were
women?

(b) Repeat part (a), but now assume that the sample consists of 600 people.

Ex. 4.2.5. Calculate the variance and standard deviation of the value of the lottery ticket
from Example 3.1.4.
Ex. 4.2.6. Prove parts (c) and (d) of Theorem 4.2.4.
Ex. 4.2.7. Let X ∼ Binomial (n, p). Show that for 0 < p < 1, this random variable has
the largest standard deviation when p = 12 .
Ex. 4.2.8. Follow the steps below to calculate the variance of a random variable with a
Uniform({1, 2, . . . , n}) distribution.
n
n(n+1)(2n+1)
(a) Prove that k2 = . (Induction is one way to do this).
P
6
k =1

(b) Let X ∼ Uniform({1, 2, . . . , n}). Use (a) to calculate E [X 2 ].

(c) Use (b) and the fact that E [X ] = n+1

2 to calculate V ar [X ].

Ex. 4.2.9. This exercise provides an example of a random variable with finite expected
2n
value, but infinite variance. Let X be a random variable for which P (X = n(n+1)
) = 1
2n
for all integers n ≥ 1.
∞
2n
(a) Prove that X is a well-defined variable by showing = 1.
P
P (X = n(n+1)
)
n=1

(b) Prove that E [X ] = 1.

(c) Prove that V ar [X ] is infinite.

Ex. 4.2.10. Recall that the hypergeometric distribution was first developed to answer
questions about sampling without replacement. With that in mind, answer the following
questions using the chart of expected values and variances.

(a) Use the formula in the chart to calculate the variance of a hypergeometric distribution
if m = 0. Explain this result in the context of what it means in terms of sampling.

Version: – November 19, 2024

 4.2 variance and standard deviation 121

(b) Use the formula in the chart to calculate the variance of a hypergeometric distribution
if r = 0. Explain this result in the context of what it means in terms of sampling.

(c) Though we only defined a hypergeometric distrbiution if N > max{r, m}, the
definition could be extended to N = max{r, m}. Use the chart to calculate the
variance of a hypergeometric distribution if N = m. Explain this result in the context
of what it means in terms of sampling without replacement.

Ex. 4.2.11. Prove the following facts about independent random variables.

(a) Use Theorem 4.2.6 and induction to prove that if X1 , X2 , . . . , Xn are independent,
then
V ar [X1 + · · · + Xn ] = V ar [X1 ] + · · · + V ar [Xn ].

(b) Suppose X1 , X2 , . . . , Xn are i.i.d. Prove that

√
E [X1 + · · · + Xn ] = n · E [X1 ] and SD [X1 + · · · + Xn ] = n · SD [X1 ].

(c) Suppose X1 , X2 , . . . , Xn are mutually independent standardized random variables

(not necessarilly identically distributed). Let

X1 + X2 + · · · + Xn
Y = √ .
n

Prove that Y is a standardized random variable.

Ex. 4.2.12. Let X be a discrete random variable which takes on only non-negative values.
Show that if E [X ] = 0 then P (X = 0) = 1.
Ex. 4.2.13. Suppose X is a discrete random variable with finite variance (and thus finite
expected value as well) and suppose there are two different numbers a, b ∈ R for which
P (X = a) and P (X = b) are both positive. Prove that V ar [X ] > 0.
Ex. 4.2.14. Let X be a discrete random variable with finite variance (and thus finite
expected value as well).

(a) Prove that E [X 2 ] ≥ (E [X ])2 .

(b) Suppose there are two different numbers a, b ∈ R for which P (X = a) and P (X = b)
are both positive. Prove that E [X 2 ] > (E [X ])2 .

Ex. 4.2.15. Let X ∼ Binomial(n, p) for n > 1 and 0 < p < 1. Using the steps below,
provide an algebraic proof of the fact that V ar [X ] = np(1 − p) without appealing to the
fact that such a variable is the sum of Bernoulli trials.

Version: – November 19, 2024

122 summarizing discrete random variables

(a) Begin by writing an expression for E [X 2 ] in summation form.

n−1
(b) Use (a) to show that E [X 2 ] = np · (k + 1)(n−1
k )p (1 − p)
k (n−1)−k .
P
k =0

(c) Use (b) to explain why E [X 2 ] = np · E [Y + 1] where Y ∼ Binomial(n − 1, p).

(d) Use (c) together with Theorem 4.2.5 to prove that V ar [X ] = np(1 − p).

4.3 standard units

When there is no confusion about what random variable is being discussed, it is usual to
use the Greek letter µ in place of E [X ] and σ in place of SD [X ]. When more than one
variable is involved the same letters can be used with subscripts (µX and σX ) to indicate
which variable is being described.
In statistics one frequently measures results in terms of “standard units” – the number
of standard deviations a result is from its expected value. For instance if µ = 12 and σ = 5,
then a result of X = 20 would be 1.6 standard units because 20 = µ + 1.6σ. That is, 20 is
1.6 standard deviations above expected value. Similarly a result of X = 10 would be −0.4
standard units because 10 = µ − 0.4σ.
Since the standard deviation measures a typical distance from average, results that
are within one standard deviation from average (between −1 and +1 standard units) will
tend to be fairly common, while results that are more than two standard deviations from
average (less than −2 or greater than +2 in standard units) will usually be relatively rare.
The likelihoods of some such events will be calculated in the next two examples. Notice
that the event (|X − µ| ≤ kσ ) describes those outcomes of X that are within k standard
deviations from average.

Example 4.3.1. Let Y represent the sum of two dice. How likely is it that Y will be
within one standard deviation of its average? How likely is it that Y will be more than
two standard deviations from its average? q
We can use our previous calculations that µ = 7 and σ = 35
6 ≈ 2.42. The achievable
values that are within one standard deviation of average are 5, 6, 7, 8, and 9. So the
probability that the sum of two dice will be within one standard deviation of average is

P (|Y − µ| ≤ σ ) = P (Y ∈ {5, 6, 7, 8, 9})

4 5 6 5 4
= + + + +
36 36 36 36 36
2
= .
3

Version: – November 19, 2024

 4.3 standard units 123

There is about a 66.7% chance that a pair of dice will fall within one standard deviation of
their expected value.
q
Two standard deviations is 2 35
6 ≈ 4.83. Only the results 2 and 12 further than this
distance from the expected value, so the probability that X will be more than two standard
deviations from average is

P (|Y − µ| > 2σ ) = P (Y ∈ {2, 12})

2
= ≈ 0.056.
36

There is only about a 5.6% chance that a pair of dice will be more than two standard
deviations from expected value. ■

Example 4.3.2. If X ∼ U nif orm{(1, 2, . . . , 100)}, what is the probability that X will be
within one standard deviation of expected value? What is the probability it will be more
than two standard deviations from expected value?

q Again, based on earlier calculations we know that µ = 2 = 50.5 and that σ =

101

9999
12 ≈ 28.9. Of the possible values that X can achieve, only the numbers 22, 23, . . . , 79
fall within one standard deviation of average. So the desired probability is

P (|X − µ| ≤ σ ) = P (X ∈ {22, 23, . . . , 79})

58
= .
100

There is a 58% chance that this random variable will be within one standard deviation of
expected value.
q
Similarly we can calculate that two standard deviations is 2 9999
12 ≈ 57.7. Since
µ = 50.5 and since the minimal and maximal values of X are 1 and 100 respectively, results
that are more than two or more standard deviations from average cannot happen at all for
this random variable. In other words P (|X − µ| > 2σ ) = 0. ■

4.3.1 Markov and Chebychev Inequalities

The examples of the previous section show that the exact probabilities a random variable
will fall within a certain number of standard deviations of its expected value depend on the
distribution of the random variable. However, there are some general results that apply to
all random variables. To prove these results we will need to investigate some inequalities.

Version: – November 19, 2024

124 summarizing discrete random variables

Theorem 4.3.3. (Markov’s Inequality) Let X be a discrete random variable

which takes on only non-negative values and suppose that X has a finite expected
value. Then for any c > 0,
µ
P (X ≥ c) ≤ .
c

Proof - Let T be the range of X, so T is a countable subset of the positive real numbers.
By dividing T into those numbers smaller than c and those numbers that are at least as
large as c we have
X
µ = t · P (X = t)
t∈T

t · P (X = t).
X X
= t · P (X = t) +
t∈T ,t<c t∈T ,t≥c

The first sum must be non-negative, since we assumed that T consisted of only non-negative
numbers, so we only make the quantity smaller by deleting it. Likewise, for each term in
the second sum, t ≥ c so we only make the quantity smaller by replacing t by c. This gives
us
X X
µ = t · P (X = t) + t · P (X = t)
t∈T ,t<c t∈T ,t≥c
X
≥ c · P (X = t)
t∈T ,t≥c

P (X = t).
X
= c·
t∈T ,t≥c

The events (X = t) indexed over all values t ∈ T for which t ≥ c are a countable collection
of disjoint sets whose union is (X ≥ c). So,
X
µ ≥ c· P (X = t)
t∈T ,t≥c
= cP (X ≥ c).

Dividing by c gives the desired result.

Markov’s theorem can be useful in its own right for producing an upper bound on the
liklihood of certain events, but for now we will use it simply as a lemma to prove our next
result.

Version: – November 19, 2024

 4.3 standard units 125

Theorem 4.3.4. (Chebychev’s Inequality) Let X be a discrete random variable

with finite, non-zero variance. Then for any k > 0,

1
P (|X − µ| ≥ kσ ) ≤ .
k2

Proof - The event (|X − µ| ≥ kσ ) is the same as the event ((X − µ)2 ≥ k 2 σ 2 ). The
random variable (X − µ)2 is certainly non-negative and its expected value is the variance
of X which we have assumed to be finite. Therefore we may apply Markov’s inequality to
(X − µ)2 to get

P (|X − µ| ≥ kσ ) = P ((X − µ)2 ≥ k 2 σ 2 )

E [(X − µ)2 ]
≤
k2 σ2
V ar [X ]
=
k2 σ2
σ2
=
k2 σ2
1
= .
k2

Though the theorem is true for all k > 0, it doesn’t give any useful information unless
k > 1.
Example 4.3.5. Let X be a discrete random variable. Find an upper bound on the
likelihood that X will be more than two standard deviations from its expected value.
For the question to make sense we need to assume that X has finite variance to begin
with. In which case we may apply Chebychev’s inequality with k = 2 to find that

1
P (|X − µ| > 2σ ) ≤ P (|X − µ| ≥ 2σ ) ≤ .
4

There is at most a 25% chance that a random variable will be more than two standard
deviations from its expected value. ■

exercises

Ex. 4.3.1. Let X ∼ Binomial (10, 21 ).

(a) Calculate µ and σ.

(b) Calculate P (|X − µ| ≤ σ ), the probability that X will be within one standard
deviation of average. Approximate your answer to the nearest tenth of a percent.

Version: – November 19, 2024

126 summarizing discrete random variables

(c) Calculate P (|X − µ| > 2σ ), the probability that X will be more than two standard
deviations from average. Approximate your answer to the nearest tenth of a percent.

Ex. 4.3.2. Let X ∼ Geometric( 14 ).

(a) Calculate µ and σ.

(b) Calculate P (|X − µ| ≤ σ ), the probability that X will be within one standard
deviation of average. Approximate your answer to the nearest tenth of a percent.

(c) Calculate P (|X − µ| > 2σ ), the probability that X will be more than two standard
deviations from average. Approximate your answer to the nearest tenth of a percent.

Ex. 4.3.3. Let X ∼ P oisson(3).

(a) Calculate µ and σ.

(b) Calculate P (|X − µ| ≤ σ ), the probability that X will be within one standard
deviation of average. Approximate your answer to the nearest tenth of a percent.

(c) Calculate P (|X − µ| > 2σ ), the probability that X will be more than two standard
deviations from average. Approximate your answer to the nearest tenth of a percent.

Ex. 4.3.4. Let X ∼ Binomial (n, 12 ). Determine the smallest value of n for which P (|X −
µ| > 4σ ) > 0. That is, what is the smallest n for which there is a positive probability that
X will be more than four standard deviations from average.
Ex. 4.3.5. For k ≥ 1 there are distributions for which Chebychev’s inequality is an equality.

(a) Let X be a random variable with probability mass function P (X = 1) = P (X =

−1) = 12 . Prove that Chebychev’s inequality is an equality for this random variable
when k = 1.

(b) Let X be a random variable with probability mass function P (X = 1) = P (X =

−1) = p and P (X = 0) = 1 − 2p. For any given value of k > 1, show that it is
possible to select a value of p for which Chebychev’s inequality is an equality when
applied to this random variable.

Ex. 4.3.6. Let X be a discrete random variable with finite expected value µ and finite
variance σ 2 .

(a) Explain why P (|X − µ| > σ ) = P ((X − µ)2 > σ 2 ).

(b) Let T be the range of the random variable (X − µ)2 .

Explain why P ((X − µ)2 = t) = 1.
P
t∈T

Version: – November 19, 2024

 4.4 conditional expectation and conditional variance 127

(c) Explain why V ar [X ] = t · P ((X − µ)2 = t).

P
t∈T

(d) Prove that if P (|X − µ| > σ ) = 1, then

σ 2 · P ((X − µ)2 = t). (Hint: Use (a) to explain why replacing t by σ 2

P
V ar [X ] >
t∈T
in the sum from (c) will only make the quantity smaller).

(e) Use parts (b) and (d) to derive a contradiction. Note that this proves that the
assumption that was made in part (d), namely that P (|X − µ| > σ ) = 1, cannot be
true for any discrete random variable where µ and σ are finite quantities. In other
words, no random variable can produce only values that are more than one standard
deviation from average.

Ex. 4.3.7. Let X be a discrete random variable with finite expected value and finite
variance.

(a) Prove P (|X − µ| ≥ σ ) = 1 ⇐⇒ P (|X − µ| = σ ) = 1. (A random variable that

assumes only values one or more standard deviations from average must only produce
values that are exactly one standard deviation from average).

(b) Prove that if P (|X − µ| > σ ) > 0 then P (|X − µ| < σ ) > 0. (If a random variable is
able to produce values more one standard deviation from average, it must also be
able to produce values that are less than one standard deviation from average).

4.4 conditional expectation and conditional variance

In previous chapters we saw that information that a particular event had occurred could
substantially change the probability associated with another event. That realization led us
to the notion of conditional probability. It is also reasonable to ask how such information
might affect the expected value or variance of a random variable.

Version: – November 19, 2024

128 summarizing discrete random variables

Definition 4.4.1. Let X : S → T be a discrete random variable and let A ⊂ S

be an event for which P (A) > 0. The “conditional expected value” is defined from
conditional probabilities in the same way the (ordinary) expected value is defined
from (ordinary) probabilities. Likewise the “conditional variance” is described in
terms of the conditional expected value in the same way the (ordinary) variance is
described in terms of the (ordinary) expected value. Specificially, the “conditional
expected value” of X given A is

t · P (X = t|A),
X
E [X|A] =
t∈T

and the “conditional variance” of X given A is

V ar [X|A] = E [(X − E [X|A])2 |A].

Example 4.4.2. A die is rolled. What are the expected value and variance of the result
given that the roll was even?
Let X be the die roll. Then X ∼ Uniform({1, 2, 3, 4, 5, 6}), but conditioned on the
event A that the roll was even, this changes so that

P (X = 1|A) = P (X = 3|A) = P (X = 5|A) = 0 while

1
P (X = 2|A) = P (X = 4|A) = P (X = 6|A) = .
3
Therefore,
1 1 1
E [X|A] = 2( ) + 4( ) + 6( ) = 4.
3 3 3
Note that the (unconditioned) expected value of a die roll is E [X ] = 3.5, so the knowledge
of event A slightly increases the expected value of the die roll.
The conditional variance is

1 1 1 8
V ar [X|A] = (2 − 4)2 ( ) + (4 − 4)2 ( ) + (6 − 4)2 ( ) = .
3 3 3 3

This result is slightly less than 12 ,

35
the (unconditional) variance of a die roll. This means
that knowledge of event A sligthly decreased the typical spread of the die roll results. ■

In many cases the event A on which an expected value is conditioned will be described in
terms of another random variable. For instance E [X|Y = y ] is the conditional expectation
of X given that variable Y has taken on the value y.

Version: – November 19, 2024

 4.4 conditional expectation and conditional variance 129

Example 4.4.3. Cards are drawn from an ordinary deck of 52, one at a time, randomly
and with replacement. Let X and Y denote the number of draws until the first ace and
first king are drawn, respectively. We are interested in say, E [X|Y = 3]. When Y = 3 an
ace was seen of draw 3, but not on draws 1 or 2. Hence

 4
 48
 if n = 1 or 2
P (king on draw n|Y = 3) = 0 if n = 3

 4 if n > 3

52

so that   n−1
 44 4
if n = 1 or 2
 48 48


P (X = n|Y = 5) = 0 if n = 3
 44 2 48 n−4 4
    
if n > 3


48 52 52

For example, when n > 3, in order to have X = n a non-king must have been seen on
draws 1 and 2 (each with probability 48 ),
44
a non-king must have resulted on draw 3 (which
is automatic, since an ace was drawn), a non-king must have been seen on each of draws 4
through n − 1 (each with probability 52 ),
48
and finally a king was produced on draw n (with
probability 52 ).
4
Hence,

2  44 n−1 4 ∞  44 2  48 n−4 4
E [X|Y = 3] =
X X
n + n
n=1
48 48 n=4
48 52 52
2  44 n−1 4 ∞  44 2  48 m 4
(m + 4) .
X X
= n +
n=1
48 48 m=0
48 52 52

But
∞ ∞ 
d m+1 
(m + 4)r m = 3rm +
X X
r
m=0 m=0
dr
3 d  r 
= +
1 − r dr 1 − r
3 1
= + ,
1 − r (1 − r )2

Version: – November 19, 2024

130 summarizing discrete random variables

so

4  44  4   44 2  4  3 1 
E [X|Y = 3] = +2 + +
48 48 48 48 52 1 − (48/52) (1 − (48/52))2
4  44  4   44 2  4  3 × 52 522 
= +2 + + 2
48 48 48 48 52 4 4
1  11  1   11 2 52  11 2
= +2 +3 +
12 12 12 12 4 12
985
= ≈ 13.68.
72

Given that the first ace appeared on draw 3, it takes an average of between 13 and 14
draws until the first king appears. Compare this to the unconditional E [X ]. Since X ∼
Geometric( 52
4
) we know E [X ] = = 13. In other words, on average it takes 13 draws
52
4
to observe the first king. But given that the first ace appeared on draw three, we should
expect to need about 0.68 draws more (on average) to see the first king. ■
Recall how Theorem 1.3.2 described a way in which a non-conditional probability could
be calculated in terms of conditional probabilities. There is an analogous theorem for
expected value.

Theorem 4.4.4. Let X : S → T be a discrete random variable and let {Bi : i ≥ 1} be

∞
a disjoint collection of events for which P (Bi ) > 0 for all i and such that
S
Bi = S.
i=1
Suppose P (Bi ) and E [X|Bi ] are known. Then E [X ] may be computed as
∞
E [X|Bi ]P (Bi ).
X
E [X ] =
i=1

Proof - Using Theorem 1.3.2 and the definition of conditional expectation,

∞
X ∞ X
X
E [X|Bi ]P (Bi ) = t · P (X = t|Bi )P (Bi )
i=1 i=1 t∈T
XX ∞
= t · P (X = t|Bi )P (Bi )
t∈T i=1

t · P (X = t) = E [X ].
X
=
t∈T

■
Example 4.4.5. A venture capitalist estimates that regardless of whether the economy
strengthens, weakens, or remains the same in the next fiscal quarter, a particular investment
could either gain or lose money. However, he figures that if the economy strengthens,

Version: – November 19, 2024

 4.4 conditional expectation and conditional variance 131

the investment should, on average, earn 3 million dollars. If the economy remains the
same, he figures the expected gain on the investment will be 1 million dollars, while if the
economy weakens, the investment will, on average, lose 1 million dollars. He also trusts
economic forcasts which predict a 50% chance of a weaker economy, a 40% chance of a
stagnant economy, and a 10% chance of a stronger economy. What should he calculate is
the expected return on the investment?
Let X be the return on investment and let A, B, and C represent the events that the
economy will be stronger, the same, and weaker in the next quarter, respectively. Then
the estimates on return give the following information in millions:

E [X|A] = 3; E [X|B ] = 1; and E [X|C ] = −1.

Therefore,

E [X ] = E [X|A]P (A) + E [X|B ]P (B ) + E [X|C ]P (C )

= 3(0.1) + 1(0.4) + (−1)(0.5) = 0.2

The expected return on investment is $200,000. ■

When the conditioning event is described in terms of outcomes of a random variable,
Theorem 4.4.4 can be written in another useful way.

Theorem 4.4.6. Let X and Y be two discrete random variables on a sample space
S with Y : S → T . Let g : T → R be defined as g (y ) = E [X|Y = y ]. Then

E [g (Y )] = E [X ].

It is common to use E [X|Y ] to denote g (Y ) after which the theorem may be expressed
as E [E [X|Y ]] = E [X ]. This can be slightly confusing notation, but one must keep
in mind that the exterior expected value in the expression E [E [X|Y ]] refers to the
averge of E [X|Y ] viewed as a function of Y .

Proof - As y ranges over T , the events (Y = y ) are disjoint and cover all of S. Therefore,
by Theorem 4.4.4,
X
E [g (Y )] = g (y )P (Y = y )
y∈T
X
= E [X|Y = y ]P (Y = y )
y∈T
= E [X ].

Version: – November 19, 2024

132 summarizing discrete random variables

Example 4.4.7. Let Y ∼ Uniform({1, 2, . . . , n}) and let X be the number of heads on Y
flips of a coin. What is the expected value of X?

Without Theorem 4.4.6 this problem would require computing many complicated
probabilities. However, it is made much simpler by noting that the distribution of X is
given conditionally by (X|Y = j ) ∼ Binomial(j, 12 ). Therefore we know E [X|Y = j ] = 2j .
Using the notation above, this may be written as E [X|Y ] = Y
2 after which

Y 1n+1 n+1
E [X ] = E [E [X|Y ]] = E [ ]= = .
2 2 2 4

Though it requires a somewhat more complicated formula, the variance of a random

variable can be computed from conditional information.

Theorem 4.4.8. Let X : S → T be a discrete random variable and let {Bi : i ≥ 1} be

∞
a disjoint collection of events for which P (Bi ) > 0 for all i and such that
S
Bi = S.
i=1
Suppose E [X|Bi ] and V ar [X|Bi ] are known. Then V ar [X ] may be computed as
∞
X 
V ar [X ] = (V ar [X|Bi ] + (E [X|Bi ])2 )P (Bi ) − (E [X ])2 .
i=1

Proof- First note that V ar [X|Bi ] = E [X 2 |Bi ] − (E [X|Bi ])2 , and so

V ar [X|Bi ] + (E [X|Bi ])2 = E [X 2 |Bi ].

Therefore,
∞ ∞
E [X 2 |Bi ]P (Bi ),
X X
(V ar [X|Bi ] + (E [X|Bi ])2 )P (Bi ) =
i=1 i=1

but the right hand side of this equation is E [X 2 ] from Theorem 4.4.4. The fact that
V ar [X ] = E [X 2 ] − (E [X ])2 completes the proof of the theorem. ■

As with expected value, this formula may be rewritten in a different form if the
conditioning events describe the outcomes of a random variable.

Version: – November 19, 2024

 4.4 conditional expectation and conditional variance 133

Theorem 4.4.9. Let X and Y : S → T be two discrete random variables on a sample

space S. As in Theorem 4.4.6 let g (y ) = E [X|Y = y ]. Let h(y ) = V ar [X|Y = y ].
Denoting g (Y ) by E [X|Y ] and denoting h(Y ) by V ar [X|Y ], then

V ar [X ] = E [V ar [X|Y ]] + V ar [E [X|Y ]].

Proof - First consdier the following three facts:

(1) V ar [X|Y = t]P (Y = t) = E [V ar [X|Y ]];

P
t∈T

(2) (E [X|Y = t])2 P (Y = t) = E [(E [X|Y ])2 ]; and

P
t∈T

(3) V ar [E [X|Y ]] = E [(E [X|Y ])2 ] − (E [E [X|Y ]])2 = E [(E [X|Y ])2 ] − (E [X ])2 .

Then from Theorem 4.4.8,

X
V ar [ X ]= (V ar [X|Y = t] + (E [X|Y = t])2 )P (Y = t) − (E [X ])2
t∈T
X X
= V ar [X|Y = t]P (Y = t) + (E [X|Y = t])2 P (Y = t) − (E [X ])2
t∈T t∈T
= E [V ar [X|Y ]] + E [(E [X|Y ]) ] − (E [X ])2
2

= E [V ar [X|Y ]] + V ar [E [X|Y ]].

■
Example 4.4.10. The number of eggs N found in nests of a certain species of turtles has
a Poisson distribution with mean λ. Each egg has probability p of being viable and this
event is independent from egg to egg. Find the mean and variance of the number of viable
eggs per nest.
Let N be the total number of eggs in a nest and X the number of viable ones. Then if
N = n, X has a Binomial distribution with number of trials n and probability p of success
for each trial. Thus, if N = n, X has mean np and variance np(1 − p). That is,

E [X|N = n] = np; V ar [X|N = n] = np(1 − p)

or
E [X|N ] = pN ; V ar [X|N ] = p(1 − p)N .

Hence
E [X ] = E [E [X|N ]] = E [pN ] = pE [N ] = pλ

Version: – November 19, 2024

134 summarizing discrete random variables

and

V ar [X ] = E [V ar [X|N ]] + V ar [E [X|N ]]
= E [p(1 − p)N ] + V ar [pN ] = p(1 − p)E [N ] + p2 V ar [N ].

Since N is Poisson we know that E [N ] = V ar [N ] = λ, so that

E [X ] = pλ and V ar [X ] = p(1 − p)λ + p2 λ = pλ.

exercises

Ex. 4.4.1. Let X ∼ Geometric(p) and let A be event (X ≤ 3). Calculate E [X|A] and
V ar [X|A].

Ex. 4.4.2. Calculate the variance of the quantity X from Example 4.4.7.

Ex. 4.4.3. Return to Example 4.4.5. Suppose that, in addition to the estimates on average
return, the investor had estimates on the standard deviations. If the economy strengthens
or weakens, the estimated standard deviation is 3 million dollars, but if the economy stays
the same, the estimated standard deviation is 2 million dollars. So, in millions of dollars,

SD [X|A] = 3; SD [X|B ] = 2; and SD [X|C ] = 3.

Use this information, together with the conditional expectations from Example 4.4.5 to
calculate V ar [X ].

Ex. 4.4.4. A standard light bulb has an average lifetime of four years with a standard
deviation of one year. A Super D-Lux lightbulb has an average lifetime of eight years
with a standard devaition of three years. A box contains many bulbs – 90% of which are
standard bulbs and 10% of which are Super D-Lux bulbs. A bulb is selected at random
from the box. What are the average and standard deviation of the lifetime of the selected
bulb?

Ex. 4.4.5. Let X and Y be described by the joint distribution

X = −1 X=0 X=1
Y = −1 1/15 2/15 2/15
Y =0 2/15 1/15 2/15
Y =1 2/15 2/15 1/15

Version: – November 19, 2024

 4.5 covariance and correlation 135

and answer the following questions.

(a) Calculate E [X|Y = −1].

(b) Calculate V ar [X|Y = −1].

(c) Describe the distribution of E [X|Y ].

(d) Describe the distribution of V ar [X|Y ].

Ex. 4.4.6. Let X and Y be discrete random variables. Let x be in the range of X and let
y be in the range of Y .
(a) Suppose X and Y are independent. Show that E [X|Y = y ] = E [X ] (and so
E [X|Y ] = E [X ]).

(b) Show that E [X|X = x] = x (and so E [X|X ] = X). (From results in this section we
know E [X|Y ] is always a random variable with expected value equal to E [X ]. The
results above in some sense show two extremes. When X and Y are independent,
E [X|Y ] is a constant random variable E [X ]. When X and Y are equal, E [X|X ] is
just X itself).
Ex. 4.4.7. Let X ∼ Uniform {1, 2, . . . , n} be independent of Y ∼ Uniform {1, 2, . . . , n}.
Let Z = max(X, Y ) and W = min(X, Y ).
(a) Find the joint distribution of (Z, W ).

(b) Fine E [Z | W ].

4.5 covariance and correlation

When faced with two different random variables, we are frequently interested in how the two
different quantities relate to each other. Often the purpose of this is to predict something
about one variable knowing information about the other. For instance, if rainfall amounts
in July affect the quantity of corn harvested in August, then a farmer, or anyone else keenly
interested in the supply and demand of the agriculture industry, would like to be able to
use the July information to help make predictions about August costs.

4.5.1 Covariance

Just as we developed the concepts of expected value and standard deviation to summarize
a single random variable, we would like to develop a number that describes something
about how two different random variables X and Y relate to each other.

Version: – November 19, 2024

136 summarizing discrete random variables

Definition 4.5.1. (Covariance of X and Y ) Let X and Y be two discrete random

variables on a sample space S. Then the “covariance of X and Y ” is defined as

Cov [X, Y ] = E [(X − E [X ])(Y − E [Y ])]. (4.5.1)

Since it is defined in terms of an expected value, there is the possibility that the
covariance may be infinite or not defined at all because the sum describing the
expectation is divergent.

Notice that if X is larger than its average at the same time that Y is larger than
its average (or if X is smaller than its average at the same time Y is smaller than its
average) then (X − E [X ])(Y − E [Y ]) will contribute a positive result to the expected
value describing the covariance. Conversely, if X is smaller than E [X ] while Y is larger
than E [Y ] or vica versa, a negative result will be contributed toward the covariance. This
means that when two variables tend to be both above average or both below average
simultaneously, the covariance will typically be positive (and the variables are said to be
positively correlated ), but when one variable tends to be above average when the other
is below average, the covariance will typically be negative (and the variables are said to
be negatively correlated ). When Cov [X, Y ] = 0 the variables X and Y are said to be
“uncorrelated”.
For example, suppose X and Y are the height and weight, respectively, of an individual
randomly selected from a large population. We might expect that Cov [X, Y ] > 0 since
people who are taller than average also tend to be heavier than average and people who are
shorter than average tend to be lighter. Conversely suppose X and Y represent elevation
and air density at a randomly selected point on Earth. We might expect Cov [X, Y ] < 0
since locations at a higher elevation tend to have thinner air.

Example 4.5.2. Consider a pair of random variables X and Y with joint distribution

X = −1 X=0 X=1
Y = −1 1/15 2/15 2/15
Y =0 2/15 1/15 2/15
Y =1 2/15 2/15 1/15

By a routine calculation of the marginal distributions it can be shown that X, Y ∼

Uniform({−1, 0, 1}) and therefore that E [X ] = E [Y ] = 0. However, it is clear from the
joint distribution that when X = −1, then Y is more likely to be above average than below,

Version: – November 19, 2024

 4.5 covariance and correlation 137

while when X = 1, then Y is more likely to be below average than above. This suggests
the two random variables should have a negative correlation. In fact, we can calculate

4 9 2 2
E [XY ] = (−1)( ) + 0( ) + 1( ) = − ,
15 15 15 15

and therefore Cov [X, Y ] = E [XY ] − E [X ]E [Y ] = − 15

2
. ■

As its name suggests, the covariance is closely related to the variance.

Theorem 4.5.3. Let X be a discrete random variable. Then

Cov [X, X ] = V ar [X ].

Proof - Cov [X, X ] = E [(X − E [X ])(X − E [X ])) = E [(X − E [X ])2 ) = V ar [X ]. ■

With Theorem 4.2.5 it was shown that V ar [X ] = E [X 2 ] − (E [X ])2 , which provided

an alternate formula for the variance. There is an analogous alternate formula for the
covariance.

Theorem 4.5.4. Let X and Y be discrete random variables with finite mean for
which E [XY ] is also finite. Then

Cov [X, Y ] = E [XY ] − E [X ]E [Y ].

Proof - Using the linearity properties of expected value,

Cov [X, Y ] = E [(X − E [X ])(Y − E [Y ])]

= E [XY − XE [Y ] − E [X ]Y + E [X ]E [Y ]]
= E [XY ] − E [XE [Y ]] − E [E [X ]Y ] + E [E [X ]E [Y ]]
= E [XY ] − E [Y ]E [X ] − E [X ]E [Y ] + E [X ]E [Y ]
= E [XY ] − E [X ]E [Y ].

As with the expected value, the covariance is a linear quantity. It is also related to the
concept of independence.

Version: – November 19, 2024

138 summarizing discrete random variables

Theorem 4.5.5. Let X, Y , and Z be discrete random variables, and let a, b ∈ R.

Then,

(a) Cov [X, Y ] = Cov [Y , X ];

(b) Cov [X, aY + bZ ] = a · Cov [X, Y ] + b · Cov [X, Z ];

(c) Cov [aX + bY , Z ] = a · Cov [X, Z ] + b · Cov [Y , Z ]; and

(d) If X and Y are independent with a finite covariance, then Cov [X, Y ] = 0.

Proof of (1) - This follows immediately from the definition.

Cov [X, Y ] = E [(X − E [X ])(Y − E [Y ])]

= E [(Y − E [Y ])(X − E [X ])] = Cov [Y , X ].

Therefore, reversing the roles of X and Y does not change the correlation.

Proof of (2) - This follows from linearity properties of expected value. Using Theorem
4.5.4

Cov [X, aY + bZ ] = E [X (aY + bZ )] − E [X ]E [aY + bZ ]

= a · E [XY ] + b · E [XZ ] − a · E [X ]E [Y ] − b · E [X ]E [Z ]
= a · (E [XY ] − E [X ]E [Y ]) + b · (E [XZ ] − E [X ]E [Z ])
= a · Cov [X, Y ] + b · Cov [X, Z ]

Proof of (3) - This proof is essentially the same as that of (2) and is left as an exercise.

Poof of (4) - We have previously seen that if X and Y are independent, then E [XY ] =
E [X ]E [Y ]. Using Theorem 4.5.4 it follows that

Cov [X, Y ] = E [XY ] − E [X ]E [Y ] = 0.

Though independence of X and Y guarantees that they are uncorrelated, the converse
is not true. It is possible that Cov [X, Y ] = 0 and yet that X and Y are dependent, as the
next example shows.

Example 4.5.6. Let X, Y be two discrete random variables taking values {−1, 1}. Suppose
their joint distribution P (X = x, Y = y ) is given by the table

Version: – November 19, 2024

 4.5 covariance and correlation 139

x=-1 x=1

y=-1 0.3 0.2

y=1 0.3 0.2

By summing the columns and rows respectively,

P (X = 1) = 0.4 and P (X = −1) = 0.6, while

P (Y = 1) = 0.5 and P (Y = −1) = 0.5.

Moreover,

E [XY ] = (1)(−1)P (X = 1, Y = −1) + (−1)(1)P (X = −1, Y = 1)

+(1)(1)P (X = 1, Y = 1) + (−1)(−1)P (X = −1, Y = −1)
= −0.3 − 0.2 + 0.2 + 0.3 = 0,
E [X ] = (1)0.4 + (−1)0.6 = −0.2,
E [Y ] = (1)0.5 + (−1)0.5 = 0,

implying that Cov[X, Y ] = 0. As

P (X = 1, Y = 1) = 0.2 ̸= 0.1 = P (X = 1)P (Y = 1),

they are not independent random variables. ■

4.5.2 Correlation

The possible size of Cov [X, Y ] has upper and lower bounds based on the standard deviations
of the two variables.

Theorem 4.5.7. Let X and Y be two discrete random variables both with finite
variance. Then
−σX σY ≤ Cov [X, Y ] ≤ σX σY ,
Cov [X,Y ]
and therefore −1 ≤ σX σY ≤ 1.

Version: – November 19, 2024

140 summarizing discrete random variables

Proof - Standardize both variables and consider the expected value of their sum squared.
Since this is the expected value of a non-negative quantity,

X − µX Y − µY 2
0 ≤ E [( + ) ]
σX σY
( X − µX ) 2 (X − µX )(Y − µY ) (Y − µY )2
= E[ 2 + 2 + ]
σX σX σY σY2
E [(X − µX )2 ] 2E [(X − µX )(Y − µY )] E [(Y − µY )2 ]
= 2 + +
σX σX σY σY2
Cov [X, Y ]
= 1+2 + 1.
σX σY

Sovling the inequality for the covariance yields

Cov [X, Y ] ≥ −σX σY .

A similar computation (see Exercises) for the expected value of the squared difference of
the standardized variables shows

Cov [X, Y ] ≤ σX σY .

Putting both inequalities together proves the theorem. ■

Cov [X,Y ]
Definition 4.5.8. The quantity σX σY from Theorem 4.5.7 is known as
the“correlation” of X and Y and is often denoted as ρ[X, Y ]. Thinking in terms
of dimensional analysis, both the numerator and denominator include the units of
X and the units of Y . The correlation, therefore, has no units associated with it.
It is thus a dimensionless rescaling of the covariance and is frequently used as an
absolute measure of trends between the two variables.

exercises

Ex. 4.5.1. Consider the experiment of flipping two coins. Let X be the number of heads
among the coins and let Y be the number of tails among the coins.

(a) Should you expect X and Y to be posivitely correlated, negatively correlated, or

uncorrelated? Why?

(b) Calculate Cov [X, Y ] to confirm your answer to (a).

Version: – November 19, 2024

 4.6 exchangeable random variables 141

Ex. 4.5.2. Let X ∼ Uniform({0, 1, 2}) and let Y be the number of heads in X flips of a
coin.

(a) Should you expect X and Y to be positively correlated, negatively correlated, or

uncorrelated? Why?

(b) Calculate Cov [X, Y ] to confirm your answer to (a).

Ex. 4.5.3. Prove part (3) of Theorem 4.5.5.

Ex. 4.5.4. Prove the missing inequality from the proof of Theorem 4.5.7. Specifically, use
the inequality
X − µX Y − µY 2
0 ≤ E [( − ) ]
σX σY
to prove that Cov [X, Y ] ≤ σX σY .
Ex. 4.5.5. Prove that the inequality of Theorem 4.5.7 is an equality if and only if there are
a, b ∈ R with a ̸= 0 for which P (Y = aX + b) = 1. (Put another way, the correlation of
X and Y is ±1 exactly when Y can be expressed as a non-trivial linear function of X).
Ex. 4.5.6. In previous sections it was shown that if X and Y are independent, then
V ar [X + Y ] = V ar [X ] + V ar [Y ]. If X and Y are dependent, the result is typically not
true, but the covariance provides a way relate the variances of X and Y to the variance of
their sum.

(a) Show that for any discrete random variables X and Y ,

V ar [X + Y ] = V ar [X ] + V ar [Y ] + 2Cov [X, Y ].

(b) Use (a) to conclude that when X and Y are positively correlated, then V ar [X + Y ] >
V ar [X ] + V ar [Y ], while when X and Y are negatively correlated, V ar [X + Y ] <
V ar [X ] + V ar [Y ].

(c) Suppose Xi 1 ≤ i ≤ n are discrete random variables with finite variance and
covariances. Use induction and (a) to conclude that
n n
V ar [Xi ] + 2 Cov [Xi , Xj ].
X X X
V ar [ Xi ] =
i=1 i=1 1≤i<j≤n

4.6 exchangeable random variables

We conclude this section with a discussion on exchangeable random variables. In brief

we say that a collection of random variables is exchangeable if the joint probability mass

Version: – November 19, 2024

142 summarizing discrete random variables

function of (X1 , X2 , . . . , Xn ) is a symmetric function. In other words, the distribution of

(X1 , X2 , . . . , Xn ) is independent of the order in which the Xi′ s appear. In particular any
collection of mutually independent random variables is exchangeable.

Definition 4.6.1. Let n ≥ 2 and σ : {1, 2, . . . , n} → {1, 2, . . . , n} be a bijection. We

say that a subset T of Rn is symmetric if

(x1 , x2 , . . . , xn ) ∈ T ⇐⇒ (xσ (1) , xσ (2) , . . . , xσ (n) ) ∈ T

for all (x1 , x2 , . . . , xn ) ∈ Rn . For any symmetric set T , a function f : T → R is

symmetric if
f (x1 , x2 , . . . , xn ) = f (xσ (1) , xσ (2) , . . . , xσ (n) )

for all (x1 , x2 , . . . , xn ) ∈ Rn .

A bijection σ : {1, 2, . . . , n} → {1, 2, . . . , n} is often referred to as a permutation of

{1, 2, . . . , n}. When n = 2 the function f would be symmetric if f (x, y ) = f (y, x) for all
x, y ∈ R.

Definition 4.6.2. Let n ≥ 1 and X1 , X2 , . . . , Xn be discrete random variables. We

say that X1 , X2 , . . . , Xn is a collection of exchangeable random variables if the joint
probability mass function given by

f (x1 , x2 , . . . , xn ) = P (X1 = x1 , . . . Xn = xn )

is a symmetric function.

In particular, X1 , X2 , . . . , Xn are exchangeable then for any one of the possible n!

permutations, σ, of {1, 2, . . . , n}, X1 , X2 , . . . , Xn and Xσ (1) , Xσ (2) , . . . , Xσ (n) have the
same distribution.

Example 4.6.3. Suppose we have an urn of m distinct objects labelled {1, 2, . . . , m}.
Objects are drawn at random from the urn without replacements till the urn is empty.
Let Xi be the label of the i-th object that is drawn. Then X1 , X2 , . . . , Xm is a particular
ordering of the objects in the urn. Since each ordering is equally likely and there are m!
possible orderings we have that the joint probability mass function

1
f (x1 , x2 , . . . , xm ) = P (X1 = x1 , X2 = x2 , . . . , Xm = xm ) = ,
m!

Version: – November 19, 2024

 4.6 exchangeable random variables 143

whenever xi ∈ {1, 2, . . . m} with xi ̸= xj . As the function is a constant function on the sym-

metric set {1, 2, . . . , m}, it is clearly symmetric. So the random variables X1 , X2 , . . . , Xm
are exchangeable. ■

Theorem 4.6.4. Let X1 , X2 , . . . , Xn be a collection of exchangeable random variables

on a sample space S. Then for any i, j ∈ {1, 2, . . . , n}, Xi and Xj have the same
marginal distribution.

Proof - The random variables (X1 , X2 , . . . , Xn ) are exchangeable. Then we have for
any permutation σ and xi ∈ Range(Xi )

P (X1 = x1 , X2 = x2 , . . . , Xn = xn ) = P (Xσ (1) = x1 , Xσ (2) = x2 , . . . , Xσ (n) = xn ).

As this is true for all permutations σ all the random variables must have same range.
Otherwise if any two of them differ the we could get a contradiction by choosing an
appropriate permutation.

Let T denote the common range. Let i ∈ {2, . . . , n}, a, b ∈ T . Let

A = {xj ∈ T : 1 ≤ j ̸= 1, i ≤ n}

By using the exchangeable property with the permutation σ that is given by σ (i) =
1, σ (1) = i and σ (j ) = j for all j ̸= 1, i. We have that for any x2 , . . . , xi−1 , xi+1 , . . . , xn ∈ A

P (X1 = a, X2 = x2 , . . . , Xi−1 = xi−1 , Xi = b, Xi+1 = xi+1 , . . . Xn = xn )

= P (X1 = b, X2 = x2 , . . . , Xi−1 = xi−1 , Xi = a, Xi+1 = xi+1 , . . . Xn = xn ).

Version: – November 19, 2024

144 summarizing discrete random variables

Therefore,
[
P (X1 = a) = P ( X1 = a, Xi = b)
b∈T
X
= P (X1 = a, Xi = b)
b∈T

X1 = a, X2 = x2 , . . . Xi−1 = xi−1 , Xi = b, Xi+1 = xi+1 , . . . Xn = xn )

X [
= P(
b∈T xj ∈A

P (X1 = a, X2 = x2 , . . . , Xi = b, . . . Xn = xn )
X X
=
b∈T xj ∈A

P (X1 = b, X2 = x2 , . . . , Xi = a, . . . Xn = xn )
X X
=
b∈T xj ∈A

X1 = b, X2 = x2 , . . . Xi−1 = xi−1 , Xi = a, Xi+1 = xi+1 , . . . Xn = xn )

X [
= P(
b∈T xj ∈A
X
= P (X1 = b, Xi = a)
b∈T
[
= P( X1 = b, Xi = a)
b∈T
= P (Xi = a)

So the distribution of Xi is the same as the distribution of X1 and hence all of them have
the same distribution. ■

Example 4.6.5. (Sampling without Replacement) An urn contains b black balls and
r red balls. A ball is drawn at random and its colour noted. This procedure is repeated
n times. Assume that n ≤ b + r. Let max 0, n − r ≤ k ≤ min(n, b). In this example we
examine the random variables Xi given by

1 if i-th ball drawn is black

(
Xi =
0 otherwise

We have already seen that (See Theorem 2.3.2 and Example 2.3.1)
!Q
k−1 Qm−k−1
n i=0 (b − i ) i=0 (r − i)
P (k black balls are drawn in n draws) = Qm−1 .
k i=0 (r + b − i )

Using the same proof we see that the joint probability mass function of (X1 , X2 , . . . , Xn )
is given by

Q ni=1 xi −1 Q ni=1 xi −k−1

P P
( b − i ) (r − i)
f (x1 , x2 , . . . , xn ) = P (X1 = x1 , X2 = x2 . . . Xn = xn ) = i=0 QPn i=0
,
x −1
i=1 i
i=0 (r + b − i)

Version: – November 19, 2024

 4.6 exchangeable random variables 145

where xi ∈ {0, 1}. It is clear from the right hand side of the above that the function f
depends only on the i = 1 xi . Hence any permutation of the xi ’s will not change the value
Pn

of f . So f is a symmetric function and the random variables are exchangeable. Therefore,

by Theorem 4.6.4 we know that for any 1 ≤ i ≤ n,

b
P (Xi = 1) = P (X1 = 1) = .
b+r

So we can conclude that they are all identically distributed as Bernoulli ( b+b r ) and the
probability of choosing a black ball in the i-th draw is b
b+r (See Exercise 4.6.4 for a similar
result). Further for any i, j

Cov [Xi , Xj ] = E [Xi Xj ] − E [Xi ]E [Xj ]

2
b

= E [X1 X2 ] −
b+r
b(b − 1)
2
b

= −
(b + r )(b + r − 1) b+r
−br
=
(b + r )2 (b + r − 1)

Finally, we observe that Y = is a Hypergeometric (b + r, b, m). Exchangeability

Pn
i=1 Xi
thus provides another alternative way to compute the mean and variance of Y . Using the
linearity of expectation provided by Theorem 4.1.7, we have
n n
b
.
X X
E [Y ] = E [ Xi ] = E [Xi ] = n
i=1 i=1
b+r

and by Exercise 4.5.6,

n n n
V ar [Xi ] + 2 Cov [Xi , Xj ]
X X X
V ar [Y ] = V ar [ Xi ] =
i=1 i=1 1≤i<j≤n
= nV ar [X1 ] + n(n − 1)Cov [X1 , X2 ]
br −br
= n + n(n − 1)( )
(b + r )2 (b + r )2 (b + r − 1)
br b+r−n
= n .
(b + r ) b + r − 1
2

Version: – November 19, 2024

146 summarizing discrete random variables

exercises

Ex. 4.6.1. Suppose X1 , X2 , . . . , Xn are exchangeable random variables. For any 2 ≤ m < n,
show that X1 , X2 , . . . , Xm are also a collection of exchangeable random variables.
Ex. 4.6.2. Suppose X1 , X2 , . . . , Xn are exchangeable random variables. Let T denote their
common range. Suppose b : T → R. Show that b(X1 ), b(X2 ), . . . , b(Xn ) is also a collection
of exchangeable random variables.
Ex. 4.6.3. Suppose n cards are drawn from a standard pack of 52 cards without replacement
(so we will assume n ≤ 52). For 1 ≤ i ≤ n, let Xi be random variables given by

1 if i-th card drawn is black in colour

(
Xi =
0 otherwise

(a) Suppose n = 52. Using Example 4.6.3 and the Exercise 4.6.2 show that (X1 , X2 , X3 , . . . Xn )
are exchangeable.

(b) Show that (X1 , X2 , X3 , . . . , Xn ) are exchangeable for any 2 ≤ n ≤ 52. Hint: If
n < 52 extend the sample to exhause the deck of cards. Use (a) and Exercise 4.6.1

(c) Find the probability that the second and fourth card drawn have the same colour.

Ex. 4.6.4. (Polya Urn Scheme) An urn contains b black balls and r red balls. A ball is
drawn at random and its colour noted. Then it is replaced along with c ≥ 0 balls of the
same colour. This procedure is repeated n times.

(a) Let 1 ≤ k ≤ m ≤ n. Show that

!Q
k−1 Qm−k−1
m i=0 (b + ci) i=0 (r + ci)
P (k black balls are drawn in m draws) = Qm−1
k i=0 (r + b + ci)

(b) Let 1 ≤ i ≤ n and random variables Xi be given by

1 if i-th ball drawn is black

(
Xi =
0 otherwise

Show that the collection of random variables is exchangeable.

(c) Let 1 ≤ m ≤ n. Let Bm be the event that the m-th ball drawn is black. Show that

b
P (Bm ) = .
b+r

Version: – November 19, 2024

C O N T I N U O U S P R O B A B I L I T I E S A N D R A N D O M VA R I A B L E S
5
We have thus far restricted discussion to discrete spaces and discrete random variables—
those consisting of at most a countably infinite number of outcomes. This is not because it
is not possible, interesting, or useful to consider probabilities on an uncountably infinite
set such as the real line or the interval (0, 1). Instead, there are a few technicalities that
arise when discussing such probabilities that are best avoided until they are needed. That
time is now.

5.1 uncountable sample spaces and densities

Suppose we want to randomly select a number on the interval (0, 1) in some uniform way.
In the discrete setting we would have said that “uniform” meant that every outcome in
our sample space S = (0, 1) was equally likely. Suppose we took that same approach here
and declared that there was some value p for which P ({x}) = p for every x ∈ (0, 1). Then
if we let E be the event E = { n1 : n = 2, 3, 4, . . . } ⊂ S, we find that

1 1 1
 
P (E ) = P ( , , ,... )
2 3 4
1 1 1
= P( ) + P( ) + P( ) + ...
2 3 4
= p+p+p+...

If p > 0 this sum diverges to infinity, which cannot be since it describes a probability.
Therefore it must be that p = 0. If every individual outcome in S = (0, 1) is equally likely,
then each outcome must have a probability of zero. After several chapters considering only
discrete probabilities many readers may suspect that this, in and of itself, is a contradiction.
How is it possible for P (S ) = 1 when every single element of S has probability zero? Could
not one then show
[
P (S ) = P ( {s})
s∈S
X
= P ({s})
s∈S

0
X
=
s∈S
= 0

147

Version: – November 19, 2024

148 continuous probabilities and random variables

using the probability axioms? The answer to that question is “no”. The probability space
axiom that allows us to write the probability of a disjoint union as the sum of separate
probabilities only applies to countable collections of events. But the events {s} that
combine to create (0, 1) are an uncountable collection. If S is uncountable, we could still
have P (S ) = 1 even if every individual element of s ∈ S has probability zero.
However, all of that does not yet explain how to define a uniform probability on
(0, 1). Knowing that each individual outcome has probability zero does not tell us how to
calculate P ([ 14 , 34 ]), for example, since we cannot simply add up the probabilities of each
of the constituent outcomes individually. Instead we need to reinterpret what we mean
by “uniform” in this situation. It would make sense to suggest that the event [ 14 , 34 ] should
have a probability of 12 since its length is exactly half of the length of (0, 1). Indeed it is
tempting (and essentially correct) to declare that P (A) should be the length of the set
A. The complication with making such a statement is that, although length is easy to
define if A is an interval or even a countable collection of disjoint intervals, it is not even
possible to consistently define a length for every single subset of (0, 1). Because of this
unfortunate fact, we will need to reconsider which subsets of S are actually events which
will be assigned a probability.
At a minimum we will want events to include any interval. The axioms and basic
properties of probability spaces also require that for any collection of events we must be
able to consider complements and countable unions of these events. Further, the entire
sample space S should also be considered a legitimate event. Consequently we make the
following definition.

Definition 5.1.1. (σ-field) If S is a sample space, then a σ-field F is a collection

of subsets of S such that

(1) S ∈ F

(2) If A ∈ F then Ac ∈ F
∞
(3) If A1 , A2 , . . . is a countable collection of sets in F then An ∈ F
S
n=1

We shall refer to an element of the σ-field as an event.

If S happens to be the set of real numbers there is a smallest σ-field that contains all
intervals, and this collection of subsets of R is known as the Borel sets. It happens that
the concept of the “length” of a set can be consistantly described for such sets. Because of
this we will modify our definition of probability space slightly at this point.

Version: – November 19, 2024

 5.1 uncountable sample spaces and densities 149

Definition 5.1.2. (Probability Space Axioms) Let S be a sample space and let
F be a σ-field of S. A “probability” is a function P : F → [0, 1] such that

(1) P (S ) = 1;

(2) If E1 , E2 , ... are a countable collection of disjoint events in F, then

∞ ∞
P (Ej ).
[ X
P( Ej ) =
j =1 j =1

The triplet (S, F, P ) is referred to as a probability space.

Our old definition is simply a special case where the σ-field was the collection of all
subset of S, so all results we have previously seen in the discrete setting are still legitimate
in this new framework. There are many technicalities that arise due to the fact that not
every set may be viewed as an event, but these issues would be distracting from the primary
goal of this text. Thus we give the definitions above only to provide the modern definition
of probability space.

Throughout the remainder of the sections on continuous probability spaces we will

restrict our attention to the sample space being R. Whenever we state or prove anything
for an event A (a Borel set) we shall restrict ourselves to the case the event A is a finite or
countable unions of intervals. This will enable us to use standard results from calculus and
thereby avoid technicalities. A thorough study of the Borel sets and the related theory of
integration is beyond the scope of this text (the interested reader may see [AS09] in the
bibliography for additional information).

5.1.1 Probability Densities on R

The primary way we will define continuous probabilities on R is through a “density function”.
We begin by providing an example of what should be meant by a uniform distribution on
(0, 1).

Example 5.1.3. Let f : R → R be a function defined by


1 if 0 < x < 1
f (x) =
0 otherwise.

Version: – November 19, 2024

150 continuous probabilities and random variables

For an event A define P (A) = Note that for an interval A = [a, b] ⊂ (0, 1) it
R
A f (x) dx.
happens that P (A) is just the length of the interval.
Z
P (A) = f (x) dx
A
Z b
= 1 dx
a
= b−a

For disjoint unions of intervals, the lengths simply add. For instance if A = [ 51 , 25 ] ∪ [ 53 , 45 ],
then
Z
P (A) = f (x) dx
[ 51 , 25 ]∪[ 53 , 45 ]
Z 2 Z 4
5 5
= 1 dx + 1 dx
1 3
5 5
1 1 2
= + =
5 5 5

which is the sum of the lengths of the two component intervals. In particular note that
P ((0, 1)) = 1 while P ({c}) = 0 for any c since a single point has no length. Similarly, if
A = [a, b] is an interval that is disjoint from (0, 1), then
Z
P (A) = f (x) dx
A
Z b
= 0 dx
a
= 0

We will soon see that P defines a probability on R. From the computation above this
probability gives equal likelihood to all equal-width intervals within (0, 1) and assigns zero
probability to any interval outside of (0, 1). Therefore it is consistant with the properties a
uniform probability on (0, 1) should have. ■

The function f from the example above is known as a density. What properties must be
required of such a function in order for it to define a probability? The fact that probabilities
cannot be negative suggests we will need to require f (x) to be non-negative for all real
R∞
numbers x. The fact that P (S ) = 1 means that −∞ f (x) dx has to be 1. It turns out that
these two requirements are essentially all that are needed. The only other assumption we
will make is that a density funciton be piecewise continuous. Though this final requirement
is more restrictive than necessary, the assumption will help avoid technicalities and will

Version: – November 19, 2024

 5.1 uncountable sample spaces and densities 151

include all densities of interest to us in the remainder of the text. We give a precise
definition.

Definition 5.1.4. Let f : R → R is called a density function if f satisfies the

following:

(i) f (x) ≥ 0,

(ii) f is piecewise-continuous, and

R∞
(iii) −∞ f (x) dx = 1.

We proceed to state and prove a result that will help us construct probabilities on R
with the help of density functions. This will also ensure that in Example 5.1.3 we indeed
constructed a probability on R.

Theorem 5.1.5. Let f (x) be a density function. Define

Z
P (A) = f (x) dx, (5.1.1)
A

for any event A ⊂ R. Then P defines a probability on R. The function f is called

the “density function” for the probability P .

Proof - First note

Z
P (R) = f (x) dx
ZR∞
= f (x) dx = 1
−∞

by assumption, so the entire sample space has probability 1. Now let A be a Borel subset
of R. Since f (x) is non-negative,
Z
P (A) = f (x) dx ≥ 0, and
ZA Z
P (A) = f (x) dx ≤ f (x) dx = 1,
A R

Version: – November 19, 2024

152 continuous probabilities and random variables

so P (A) ∈ [0, 1]. Finally, if E1 , E2 , . . . are a countable collection of disjoint events, then
∞
[ Z
P( En ) = S∞ f (x) dx
n=1 n=1
En
∞ Z
X
= f (x) dx
n=1 En
∞
P (En ).
X
=
n=1

Therefore P satisfies the conditions of a probability space on R. ■

Example 5.1.6. Let f : R → R be defined by

3x2 if 0 < x < 1

(
f (x) =
0 otherwise

f is piecewise continuous, f (x) is non-negative for all x and

Z Z 1
1
f (x) dx = 3x2 dx = x3 = 1.
R 0 0

As it satisfies (i) − (iii) in Definition 5.1.4, f is a density function. Let P be as defined in

(5.1.1). As with the uniform example, f (x) is zero outside of (0, 1), so events lying outside
this interval will have zero probability. However, note that
2
1 2 7
Z
5
P ([ , ]) = 3x2 dx =
5 5 1
5
125

while 4
3 4 37
Z
5
P ([ , ]) = 3x2 dx = .
5 5 3
5
125
In other words, intervals of the same length do not have equal probabilities; this probability
is not uniform on (0, 1).
The probability of individual points is still zero, so P ({ 15 }) = P ({ 25 }) = 0, but in
terms of the density function, f ( 25 ) is four times as large as f ( 15 ). What does this mean in
practical terms?
Let ϵ be a small positive quantity (certianly less than 15 ). Then

1 1 2 2
P ([ − ϵ, + ϵ]) = ϵ + 2ϵ3 ≈ ϵ while
5 5 25 25
2 2 8 8
P ([ − ϵ, + ϵ]) = ϵ + 2ϵ3 ≈ ϵ.
5 5 25 25

Version: – November 19, 2024

 5.1 uncountable sample spaces and densities 153

The fact that f ( 25 ) is four times as large as f ( 15 ) essentially means that a tiny interval
around 2
5 has approximately four times the probability of a similarly sized interval around
5.
1
■

exercises

Ex. 5.1.1. Let f : R → R be defined by

2x if 0 < x < 1
(
f (x) =
0 otherwise

(a) Show that f is a probability density function.

(b) Use f to calculate P ((0, 12 )).

Ex. 5.1.2. Let f : R → R be defined by




 x if 0 < x < 1
f (x) = 2−x if 1 ≤ x < 2

0 otherwise



(a) Sketch a graph of the function f .

(b) Show that f is a probability density function.

(c) Use f to calculate : P ((0, 14 ), P (( 32 , 2)), P ((−3, −2)) and P (( 12 , 32 )).

Ex. 5.1.3. Let f : R → R be defined by




 k if 0 < x < 1
4

 2k

if 1
≤x< 3
f (x) = 4 4


 3k if 3
4 ≤x<1

0 otherwise



(a) Find k that makes f a probability density function.

(b) Sketch a graph of the function f .

(c) Use f to calculate : P ((0, 14 ), P (( 14 , 34 )), P (( 34 , 1)).

Ex. 5.1.4. Let f : R → R be defined by

k · sin(x) if 0 < x < π

(
f (x) =
0 otherwise

Version: – November 19, 2024

154 continuous probabilities and random variables

(a) Determine the value of k that makes f a probability density function.

(b) Calculate P ((0, 21 )) and P (( 12 , 1)).

(c) Which will be larger, P ((0, 14 )) or P (( 14 , 12 ))? Explain how you can answer this
question without actually calculating either probability.

(d) A game is played in the following way. A random variable X is selected with a density
described by f above. You must select a number r and you win the game if the
random variable results in an outcome in the interval (r − 0.01, r + 0.01). Explain
how you should choose r to maximize your chance of winning the game. (A formal
proof requires only basic calculus, but it should take very little computation to
determine the correct answer).

Ex. 5.1.5. Let λ > 0 and f : R → R be defined by

if 0 < x
(
λe−λx
f (x) =
0 otherwise

(a) Show that f is a probability density function.

(b) Let a > 0. Find P ((a, ∞)).

Ex. 5.1.6. Let a, b ∈ R and f : R → R be defined by

if a < x < b
(
1
f (x) = b−a
0 otherwise

(a) Show that f is a probability density function.

(b) Show that if I, J ⊂ [a, b] are two intervals that have the same length, then P (I ) =
P (J ).

Ex. 5.1.7. Let f : R → R be defined by

if 1 < x
(
1
f (x) = x2
0 otherwise

(a) Show that f is a probability density function.

(b) Let a > 1. Find P ((a, ∞)).

Version: – November 19, 2024

 5.2 continuous random variables 155

Ex. 5.1.8. Let f : R → R be defined by

if 0 < x
(
1 2 −x
f (x) = 6x e
0 otherwise

Show that f is a probability density function.

Ex. 5.1.9. For any x ∈ R, the hyperbolic secant is defined as sech x = (ex +e−x ) .
2
Let
f : R → R be defined by
1 π
f (x) = sech( x), x ∈ R
2 2
Show that f is a probability density function.
Ex. 5.1.10. Let f : R → R be defined by

1 (x−µ)2
f (x) = √ e− 2σ2 x ∈ R
σ 2π

Follow the steps below to show that the function f is a density function.
R∞ −x2 /2 dx
(a) Let I = −∞ e and then explain why
Z ∞ Z ∞
2 +y 2 ) /2
I2 = e− ( x dx dy
−∞ −∞

(Hint: Write I 2 as a product of two integrals each over a different variable and explain
why the resulting expression may be written as the double integral above).

(b) Explain why

Z 2π Z ∞
2 /2
2
I = r · e−r dr dθ
0 0

after switching from rectangular to polar coordinates. (Hint: Use the fact from
multivariate calculus that after the change of variables (dx dy ) becomes (r dr dθ )
and explain the new limits of integration based on the region being described in the
plane).

(c) Compute the integral from (b) to find the value of I.

(x−µ)2
R∞ − x−µ
(d) Use (c) to show that √1
−∞ σ 2π e
2σ 2 dx = 1. (Hint: Use a u-substitution u = σ ).

5.2 continuous random variables

Just as the move from discrete to continuous spaces required a slight change in the definition
of probability space, so it also requires a slight change in the definition of random variable.

Version: – November 19, 2024

156 continuous probabilities and random variables

In the discrete setting we frequently needed to consider the preimage X −1 (A) of a set.
Now we need to make sure that such a preimage is a legitimate event.

Definition 5.2.1. Let (S, F, P ) be a probability space and let X : S → R be a

function. Then X is a random variable provided that whenever B is an event in R
(i.e. a Borel set), X −1 (B ) is also an event in F.

Note that in the discrete setting this extra condition was met trivially as every subset of
S was an event. Therefore the discrete setting is simply a special case of this new definition.
As with the introduction of σ-fields, we include this definition for completeness. We will
only consider functions which meet this criterion. In this section we shall consider only
continuous random variables. These are defined next.

Definition 5.2.2. Let (S, F, P ) be a probability space. A random variable X :

S → R is called a continuous random variable if there exists a density function
fX : R → R such that for any event A in R,
Z
P (X ∈ A) = fX (x) dx. (5.2.1)
A

fX is called the probability density function of X.

The following lemma demonstrates an elementary property of continous random vari-

ables that distinguishes them from discrete random variables.

Lemma 5.2.3. Let X be a continuous random variable. Then for any a ∈ R,

P (X = a) = 0 (5.2.2)

Ra
Proof- Let a ∈ R, then P (X = a) = a f (x)dx = 0. ■
Random variables may also be described using a “distribution function” (also commonly
known as a “cumulative distribution function”).

Definition 5.2.4. If X is a random variable then its distribution funciton F : R →

[0, 1] is defined by
F (x) = P (X ≤ x). (5.2.3)

When it must be emphasized that a distribution function belongs to a particular

random variable X the notation FX (x) will be used to indicate the random variable.

Version: – November 19, 2024

 5.2 continuous random variables 157

Though a distribution function is defined for any real-valued random variable, there is
a special relationship between fX (x) and FX (x) when the random variable has a density.

Theorem 5.2.5. Let X be a random variable with a piecewise continuous density

function f (x). If F (x) denotes the distriubtion function of X then
Z x
F (x) = f (x) dx. (5.2.4)
−∞

Moreover, where f (x) is continuous, F (x) is differentiable and F ′ (x) = f (x).

Proof - By definition F (x) = P (X ≤ x) = P (X ∈ (−∞, x]), but this probability is

Rx
described in terms of an integral over the density of X, so F (x) = −∞ f (x) dx.

The result that F ′ (x) = f (x) then follows from the fundamental theorem of calculus
after taking derivatives of both sides of the equation (when such a derivative exists).
Note, in particular, that since densities are assumed to be piecewise continuous, their
corresponding distribution functions are piecewise differentiable. ■

This theorem will be useful for computation, but it also shows that the distribution of
a continuous random variable X is completely determined by its distribution function FX .
That is, if we know FX (x) and want to calculate P (X ∈ A) for some set A we could do so
by differentiating FX (x) to find fX (x) and then integrating this density over the set A. In
fact FX (x) always completely determines the distribution of X (regardless of whether or
not X is a continuous random variable), but a proof of that fact is beyond the scope of the
course and will not be needed for subsequent results.

5.2.1 Common Distributions

In the literature random variables whose distributions satisfy (5.2.1) are called absolutely
continuous random variables and those that satisfy (5.2.2)are referred to as continous
random variables. Since we shall only consider continuous random variables that satisfy
(5.2.1) we refer to them as continous random variables.

There are many continuous distributions that commonly arise. Some of these are
continuous analogs of discrete random variables we have already studied. We will define
these in the context of continuous random variables having the corresponding distributions.
We begin with the already discussed uniform distribution but on an arbitrary interval.

Version: – November 19, 2024

158 continuous probabilities and random variables

Definition 5.2.6. X ∼ Uniform(a, b): Let (a, b) be an open interval. If X is a

random variable with its probabilty density function given by

1

(b−a) if a < x < b
f (x) =
 0 otherwise

then X is said to be uniformly distributed on (a, b). Note that this is consistant with
the example at the beginning of the section since the density of a Uniform(0, 1) is
one on the interval (0, 1) and zero elsewhere. Further, recall that in Exercise 5.1.6
we have shown that f is indeed a probability density function.
Since X only takes values on (a, b) if x < a then P (X ≤ x) = 0 while if x > b then
P (X ≤ x) = 1. So let a ≤ x ≤ b. Then,

1
Z x Z a Z x
x−a
P (X ≤ x) = fX (y ) dy = 0 dy + dy = .
−∞ −∞ a b−a b−a

Therefore the distribution function for X is

0


 if x < a






FX (x) = x−a
b−a if a ≤ x ≤ b






1

if x > b


Exponential Random Variable

The next continuous distribution we introduce is called the exponential distribution. It

is well known from physical experiments that the radioactive isotopes decay to its stable
form. Suppose there were N (0) atoms of a certain radioctive material at time 0 then one
is interested in the fraction of radioactive material that have not decayed at time t > 0. It
is observed from experiments that if N (t) is the number of atoms of radioactive material
that has not decayed by time t then the fraction

N (t)
≈ e−λt ,
N (0)

for some λ > 0. One can introduce a probability model for the above experiment in the
following manner. Suppose X represented the time taken by a randomly chosen radioactive

Version: – November 19, 2024

 5.2 continuous random variables 159

atom to decay to its stable form. The distribution of the random variable X needs to
satisfy
P (X ≥ t) = e−λt , (5.2.5)

for t > 0. It is possible to define such a random variable.

Definition 5.2.7. X ∼ Exp(λ): Suppose λ > 0. If X is a random variable with

its probabilty density function given by

λe−λx if x > 0
f (x) =
0 otherwise

it is said to be distributed exponentially with parameter λ. Recall that in Exercise

5.1.5 we have shown that f is indeed a probability density function. Since X only
takes values on (0, ∞) if x < 0 then P (X ≤ x) = 0. So let x ≥ 0. Then,
Z x Z 0 Z x
P (X ≤ x) = fX (x) dx = 0 dx + λe−λy dy = −e−λy |x0 = 1 − e−λx .
−∞ −∞ 0

Therefore the distribution function for X is

0 if x < 0
(
FX (x) =
1 − e−λx if 0 ≤ x

Exp(1) Exp(2)

Density f(x) Distribution F(x)

1.0

1.5
0.8

0.6
1.0

0.4
0.5
0.2

0.0 0.0

0 1 2 3 4 5 0 1 2 3 4 5

Figure 5.1: The shape of typical Exponential density and cumulative distribution functions.

Version: – November 19, 2024

160 continuous probabilities and random variables

We have previously seen that geometric random variables have the memoryless property
(See (3.2.2)). It turns out that the exponential random variable also possess the memoryless
property in continuous time. Clearly if X ∼ Exp(λ) then P (X ≥ 0) = 1 and
Z ∞
P (X ≥ t) = P (X ∈ [t, ∞)) = λe−λx dx = −e−λx |∞
t =e
−λt
,
t

for t > 0. Further if s, t > 0, X > s + t imples X > s. So

P ((X > s + t) ∩ (X > s))

P (X > s + t|X > s) =
P (X > s)
P (X > s + t) e−λ(s+t)
= = = e−λt .
P (X > s) e−λs

Therefore for all s, t > 0

P (X > s + t|X > s) = P (X > t) (5.2.6)

Thinking of the variables s and t in terms of time, this says that if an exponential random
variable has not yet occurred by time s, then its distribution from that time onward
continues to be distributed like an exponential random variable with the same parameter.
Situations that involve waiting times such as the lifetime of a light bulb or the time spent
in a queue at a service counter are often modelled with the exponential distribution. It is a
fact (see Exercise 5.2.12) that if a positive continuous random variable has the memoryless
property then it necessarily is an exponential random variable.

Example 5.2.8. Let X ∼ Exp(2). Calculate the probability that X produces a value
larger than 4.

The density of X is
2e−2x if x > 0
(
f (x) =
0 otherwise

So, P (X > 4) may be calculated via an integral.

Z ∞
P (X > 4) = 2e−2x dx
4
= −e−2x |∞ −8
4 = 0 − (−e ) = e
−8
≈ 0.000335

So there is only about a 0.0335% chance of such a result. ■

Version: – November 19, 2024

 5.2 continuous random variables 161

Normal Random Variable

Of all continuous distributions, The normal distribution (also sometimes called a “Gaussian
distribution”) is the most fundamental for applictions of statistics as it frequently arises as
a limiting distribution of sampling procedures.

Definition 5.2.9. X ∼ Normal(µ, σ 2 ): Let µ ∈ R and let σ > 0. Then X is said

to be normally distributed with parameters µ and σ 2 if it has the density

1 (x−µ)2
f (x) = √ e− 2σ2 (5.2.7)
σ 2π

for all x ∈ R. We will prove that µ and σ are, respectively, the mean and standard
deviation of such a random variable (See Definiton 6.1.1, Definition 6.1.9, Example
6.1.11). Recall that in Exercise 5.1.10 we have seen that f is a probability density
function.

Normal(0, 1) Normal(1, 2)

Density f(x) Distribution F(x)

0.4 1.0

0.8
0.3

0.6
0.2
0.4

0.1
0.2

0.0 0.0

−2 0 2 4 6 −2 0 2 4 6

Figure 5.2: The shape of typical Normal density and cumulative distribution functions.

It is observed during statistical experiments that if X were to denote the number of

leaves in an apple tree or the height of adult men in a population then X would be close
to a normal random variable with approrpriate parameters µ and σ 2 . It also arises as a
limiting distribution. We shall discuss this aspect in general in Chapter 8, but here we will
briefly mention one such limit that appears as an approximation for Binomial probabilities.

Version: – November 19, 2024

162 continuous probabilities and random variables

n = 10 n = 50 n = 200
1.0
0.8
p = 0.5

0.6
0.4
Cumulative distribution function

0.2
0.0
1.0
0.8
p = 0.25

0.6 Binomial(n,p)
Normal approximation
0.4
0.2
0.0
1.0
0.8
p = 0.1

0.6
0.4
0.2
0.0

0 2 4 6 8 10 0 10 20 30 40 50 0 50 100 150 200

Figure 5.3: The Normal approximation to Binomial.

Suppose we have X1 , X2 , . . . , Xn are i.i.d Bernoulli (p) random variables. Then we

know that Sn = is a Binomial (n, p) random variable. In Theorem 2.2.2 we saw
Pn
i=1 Xi
that for λ > 0, k ≥ 1, 0 ≤ p = λ
n < 1,

e−λ λk
lim P (Sn = k ) =
n→∞ k!

Such an approximation was useful when p was decreasing to zero while n grew to infinity
with np remaining constant. The De Moivre-Laplace Central Limit Theorem allows us to
consider another form of limit where p remains fixed, but n increases.

Theorem 5.2.10. (De Moivre-Laplace Central Limit Theorem) Suppose

Sn ∼ Binomial (n, p), where 0 < p < 1. Then for any a < b

1
Z b
Sn − np x2
lim P (a < q ≤ b) = √ e− 2 dx (5.2.8)
n→∞
np(1 − p) 2π a

We shall omit the proof of the above Theorem for now. We prove it in a more general
setting in Chapter 8. For the students well versed with Real Analysis the proof is sketched
in Exercise 5.2.16. We refer the reader to [Ram97] for a detailed discussion of the Theorem
5.2.10.

Version: – November 19, 2024

 5.2 continuous random variables 163

Calculating Normal Probabilities and Necessity of Normal Tables

In a standard introduction to integral calculus one learns many different techniques for
calculating integrals. But there are some functions whose indefinite integral has no closed-
form solution in terms of simple functions. The density of a normal random variable is one
such function. Because of this if X ∼ Normal(0, 1) the probability

1
Z x
2
P (X ≤ x) = √ e−x /2 dx
−∞ 2π

cannot be expressed exactly in terms of standard functions. Many scientific calculators

will have a feature that allows this expression to be evaluated. For example, in R, the
command pnorm(x) evaluates the integral above. Another common solution in statistical
texts is to provide a table of values.

0.00 0.02 0.04 0.06 0.08 0.10 0.12 0.14 0.16 0.18
0.0 0.500 0.508 0.516 0.524 0.532 0.540 0.548 0.556 0.564 0.571
0.2 0.579 0.587 0.595 0.603 0.610 0.618 0.626 0.633 0.641 0.648
0.4 0.655 0.663 0.670 0.677 0.684 0.691 0.698 0.705 0.712 0.719
0.6 0.726 0.732 0.739 0.745 0.752 0.758 0.764 0.770 0.776 0.782
0.8 0.788 0.794 0.800 0.805 0.811 0.816 0.821 0.826 0.831 0.836
1.0 0.841 0.846 0.851 0.855 0.860 0.864 0.869 0.873 0.877 0.881
1.2 0.885 0.889 0.893 0.896 0.900 0.903 0.907 0.910 0.913 0.916
1.4 0.919 0.922 0.925 0.928 0.931 0.933 0.936 0.938 0.941 0.943
1.6 0.945 0.947 0.949 0.952 0.954 0.955 0.957 0.959 0.961 0.962
1.8 0.964 0.966 0.967 0.969 0.970 0.971 0.973 0.974 0.975 0.976
2.0 0.977 0.978 0.979 0.980 0.981 0.982 0.983 0.984 0.985 0.985

Table 5.1: Table of Normal(0, 1) probabilities. For X ∼ Normal(0, 1), the table gives values of
P (X ≤ z ) for various values of z between 0 and 2.18 upto three digits. The value
of z for each entry is obtained by adding the corresponding row and column labels.

Table 5.1 gives values only for positive values of z because for negative z, P (X ≤ z ) can
be easily computed using the symmetry of the Normal(0, 1) distribution as (see Figure 5.4)

1 1
Z z Z ∞
2 2
P (X ≤ z ) = √ e−x /2 dx = √ e−x /2 dx = 1 − P (X ≤ −z ). (5.2.9)
−∞ 2π −z 2π

A more complete version of this table is given in the Appendix. A similar computation can
be made for other normally distributed random variables by normalizing them. Suppose

Version: – November 19, 2024

164 continuous probabilities and random variables

0.4

0.3

0.2

0.1

0.0

−3 −2 −z −1 0 1 z 2 3

Figure 5.4: Computation of Normal(0, 1) probabilities as area under the normal density curve.
For Normal(0, 1) and in fact for any symmetric distribution in general, it is enough
to know the distribution function for positive values (see Exercise 5.2.8).

Y ∼ Normal (µ, σ 2 ) and we were interested in finding the distribution function of Y .

Observe that

1
Z y
2 2
P (Y ≤ y ) = √ e−(z−µ) /2σ dz.
−∞ σ 2π
z−µ
Now perform a change of variables u = σ so that du = 1
σ dz. This integral then becomes

y−µ
1 y−µ
Z
σ 2
P (Y ≤ y ) = √ e−u /2 du = P (X ≤ ), (5.2.10)
−∞ 2π σ

where X ∼ Normal (0, 1). Now we may use Table 5.1 to compute the distribution function
of Y . We conclude this section with two examples.

Example 5.2.11. If X ∼ Normal(0, 1), how likely is it that X will be within one standard
deviation of its expected value?
In this case the expected value of the random variable is zero and the standard deviation
is one. Therefore the answer is given by

1
Z 1
2
P (−1 ≤ X ≤ 1) = √ e−x /2 dx
−1 2π
1 1
Z 1 Z −1
2 2
= √ e−x /2 dx − √ e−x /2 dx
−∞ 2π −∞ 2π
= P (X ≤ 1) − P (X ≤ −1)

Version: – November 19, 2024

 5.2 continuous random variables 165

R tells us that

pnorm(1)

[1] 0.8413447

pnorm(-1)

[1] 0.1586553

pnorm(1) - pnorm(-1)

[1] 0.6826895

Alternatively, using Table 5.1, we see that P (X ≤ 1) = 0.841 (upto three decimal places),
and by symmetry P (X ≤ −1) = P (X ≥ 1) = 1 − P (X ≤ 1) = 1 − 0.841 = 0.159.
Therefore, P (−1 ≤ X ≤ 1) ≈ 0.841 − 0.159 = 0.682. In other words, there is roughly a
68% chance that a standardized normal random variable will produce a value within one
standard deviation of expected value. ■

Example 5.2.12. A machine fills bags with cashews. The intended weight of cashews in
the bag is 200 grams. Assume the machine has a tolerance such that the actual weight
of the cashews is a normally distributed random variables with an expected value of 200
grams and a standard deviation of 4 grams. How likely is it that a bag filled by this
machine will have fewer than 195 grams of cashews in it?
We know Y ∼ Normal(200, 42 ) and we want the probability P (Y < 195). By above
computation, (5.2.10),

195 − 200 5
P (Y < 195) = P (X < ) = P (X < − )
4 4

where X ∼ Normal(0, 1). If we were to use Table 5.1, we would first obtain

5 5
P (X < − ) = 1 − P (X < ) = 1 − P (X < 1.25) = 1 − 0.896 = 0.104
4 4

Using the R command pnorm(-5/4), we obtain the value 0.1056498. That is, there is
slightly more than a 10% chance of a bag this light being produced by the machine. ■

Version: – November 19, 2024

166 continuous probabilities and random variables

5.2.2 A Word About Individual Outcomes

We began this section by noting that continuous random variables must necessarily give
probability zero to any single outcome. It is an awkward consequence of this that two
different densities may give rise to exactly the same probabilities. For instance, the functions

1 if 0 < x < 1
(
f (x) =
0 otherwise

and
1 if 0 ≤ x ≤ 1
(
g (x) =
0 otherwise
are different because they assign different values to the points x = 0 and x = 1. However,
these individual points cannot affect the computation of probabilities so both f (x) and
g (x) give rise to the same probability distribution. The same thing would occur even if
f (x) and g (x) differed in a countably infinite number of points, since these will still have
probability zero when taken collectively.
Because of this we will describe f (x) and g (x) as the same density (and sometimes
even write f (x) = g (x)) when the two densities produce the same probabilities. We do this
even when f and g may technically be different functions. Though it is a more restirctive
assumption than is necessary, we have required densities to be piecewise continuous. As a
consquence of the explanation above, altering the values of the function at the endpoints
of intervals of continuity will not change the resulting probabilities and will result in the
same density.

exercises

Ex. 5.2.1. Suppose X was continuous random variable with distribution function F .
Express the following probabilities in terms of F :

(a) P (a < X ≤ b), where −∞ < a < b < ∞

(b) P (a < X < ∞) where a ∈ R.

(c) P (| X − a |≥ b) where a, b ∈ R and b > 0.

Ex. 5.2.2. Let R > 0 and X ∼ Uniform [0, R]. Let Y = min(X, 10
R
). Find the distribution
function of Y .

Version: – November 19, 2024

 5.2 continuous random variables 167

Ex. 5.2.3. Let X be a random variable with distribution function given by




 0 if x < 0


if 0 < x < 1

x



 4
F (x) =

x
2 8+ 1
if 1
4 ≤x< 3
4

2x − 1 if 3
≤x<1




 4

1

if x ≥ 1

(a) Sketch a graph of the function F .

(b) Use F to calculate : P ([0, 14 )), P ([ 81 , 32 ]), P (( 34 , 87 ]).

(c) Find the probabilty density function of X.

Ex. 5.2.4. Let X be a continuous random variable with distribution function F : R → [0, 1].
Then G : R → [0, 1] given by
G(x) = 1 − F (x)

is called the reliability function of X or the right tail distribution function of X. Suppose
T ∼ Exponential(λ) for some λ > 0, then find the reliability function of T .

Ex. 5.2.5. Let X be a random variable whose probability density function f : R → [0, 1] is
given by 
kxk−1 e−xk if x > 0
f (x) =
0 otherwise

(a) Find the distribution function of X for k = 2.

(b) Find the distribution function of X for general k.

The distribution of X is called the Weibull distribution. Figure 5.5 plots the Weibull
distribution for selected values of k.

Ex. 5.2.6. Let X be a random variable whose probability density function f : R → [0, 1] is
given by  √
 2 2 R2 − x2
πR
if − R < x < R
f (x) =
0 otherwise

Find the distribution function of X. The distribution of X is called the semicircular

distribution (see Figure 5.6).

Version: – November 19, 2024

168 continuous probabilities and random variables

Weibull(1) Weibull(2) Weibull(3) Weibull(4)

Density f(x) Distribution F(x)

1.5 1.0

0.8
1.0
0.6

0.4
0.5
0.2

0.0 0.0

0 1 2 3 0 1 2 3

Figure 5.5: The shape of typical Weibull density and cumulative distribution functions.

Semicircular(1) Semicircular(2)

Density f(x) Distribution F(x)

1.0
0.6
0.8

0.4 0.6

0.4
0.2
0.2

0.0 0.0

−1 0 1 −1 0 1

Figure 5.6: The shape of the semicircular density and cumulative distribution functions.

Version: – November 19, 2024

 5.2 continuous random variables 169

0.8

0.6

0.4

0.2

0.0

−1.0 −0.5 0.0 0.5 1.0

Figure 5.7: Computation of probabilities as area under the density curve. For symmetric
distributions, it is enough to know the (cumulative) distribution function for
positive values.

Ex. 5.2.7. Let X be a random variable whose distribution function F : R → [0, 1] is given
by 



 0 if x ≤ 0

√
F (x) = π2 arcsin( x) if 0 < x < 1


1 if x ≥ 1



Find the probability density function of X. The distribution of X is called the standard
arcsine law.

Ex. 5.2.8. Let X be a continuous random variable with probability density function f
and distribution function F . Suppose f is a symmetric function, i.e. f (x) = f (−x) for all
x ∈ R. Then show that

(a) P (X ≤ 0) = P (X ≥ 0) = 12 ,

(b) for x ≥ 0, F (x) = 1

2 + P (0 ≤ X ≤ x),

(c) for x ≤ 0, F (x) = P (X ≥ −x) = 1

2 + P (0 ≤ X ≤ −x).

We have observed this fact for the normal distribution earlier (see Figure 5.7).

Ex. 5.2.9. Let X ∼ Exp(λ). The “90th percentile” is a value a such that X is larger than
a 90% of the time. Find the 90th percentile of X by determining the value of a for which
P (X < a) = 0.9.

Version: – November 19, 2024

170 continuous probabilities and random variables

Ex. 5.2.10. Let X be a continuous random variable such that its distribution function F is
strictly increasing on the set {x ∈ R : 0 < F (x) < 1}. The “median” of X is the value of x
for which P (X > x) = P (X < x) = 12 .

(a) If X ∼ Uniform(a, b) calculate the median of X.

(b) If Y ∼ Exp(λ) calcluate the median of Y .

(c) Let Z ∼ Normal(µ, σ 2 ). Show that the median of Z is µ.

Ex. 5.2.11. Let X ∼ Normal(µ, σ 2 ). Show that P (|X − µ| < kσ ) does not depend on the
values of µ or σ. (Hint: Use a change of variables for the appropriate integral).
Ex. 5.2.12. Above we saw that exponential random variables satisfied the memoryless
property, (5.2.6). It can be shown that any positive, continuous random variable with
the memoryless property must be exponential. Follow the steps below to prove a slightly
weakened version of this result. For all parts, suppose X is a positive, continuous random
variable with the memoryless property for which the distribution function FX (t) has a
continuous derivative for t > 0. Suppose further that limt→0+ F ′ (t) exists and call this
quantity α. Let G(t) = 1 − FX (t) = P (X > t) and do the following.

(a) Use the memoryless property to show that G(s + t) = G(s) · G(t) for all postiive s
and t.

(b) Use part (a) to conclude that G′ (t) = −αG(t). (Hint: Take a derivative with respect
to s and then take an appropriate limit).

(c) It is a fact (which you may take as granted) that the differential equation from (b)
has solutions of the form G(t) = Ce−αt . Use the fact that X is positive to explain
why it must be that C = 1.

(d) Use part (c) to calculate FX (t) and then differentiate to find fX (t).

(e) Conclude that X must be exponentially distributed and determine the associated
parameter in terms of α.

Ex. 5.2.13. Let X be a random variable with density f (x) = 2x for 0 < x < 1 (and
f (x) = 0 otherwise). Calculate the distribution function of X.
Ex. 5.2.14. Let X ∼ Uniform({1, 2, 3, 4, 5, 6}). Despite the fact this is a discrete random
variable without a density, the distribution function FX (x) is still defined. Find a piecewise
defined expression for FX (x) (see Figure 5.8 for a plot).
Ex. 5.2.15. Suppose F : R → [0, 1] is given by (5.2.3). Then show that

Version: – November 19, 2024

 5.2 continuous random variables 171
Cumulative distribution function

1.0

0.8

0.6

0.4

0.2

0.0

1 2 3 4 5 6

Figure 5.8: The cumulative distribution function for Exercise 5.2.14.

1. F is a monotonically increasing function.

2. limx→∞ F (x) = 1.

3. limx→−∞ F (x) = 0.

4. if, in addition, F is given by (5.2.4) then F is continuous.

Ex. 5.2.16. We use the notation as in Theorem 5.2.10.

(a) Let
 q q 
An = k : 0 ≤ k ≤ n, np + a np(1 − p) ≤ k ≤ np + a np(1 − p) .

Show that
Sn − np
P ( Sn = k ) .
X
P (a ≤ q ≤ b) =
np(1 − p) k∈An

(b) Let
k − np
ξk,n = q .
np(1 − p)

Using the definition of the Riemann integral show that

ξ2
k,n Z b − x2
e− 2 e 2
lim
X
= √
2π
q
n→∞
k∈An 2πnp(1 − p) a

Version: – November 19, 2024

172 continuous probabilities and random variables

(c) Using Stirling’s approximation show that

(nk)pk (1 − p)n−k
lim sup ξ2
=1
n→∞ k∈A q k,n
2πnp(1 − p)e −
n
2

(d) Prove Theorem 5.2.10 by observing

Sn − np
P (a ≤ q ≤ b) =
np(1 − p)
 
ξ2 ξ2
− k,n − k,n  (n)pk (1 − p)n−k
e 2 e 2

1
X X  k
+ −

2
q q  
ξ
k∈An 2πnp(1 − p) k∈An 2πnp(1 − p) √ 1

− k,n2 e

2πnp(1−p)

5.3 transformations of continuous random variables

In Section 3.3 we have discussed functions of discrete random variables and how to find
their distributions. Suppose g : R → R and Y = g (X ), to find the distribution of Y
we converted events associated with Y with events of X by inverting the function g. In
the setting of continuous random variables distribution functions are used for calculating
probabilities associated with functions of a known random variable. We next present a
simple example for which g (x) = x2 followed by a result that covers situations when g (x)
is any linear function.
Example 5.3.1. Let X ∼ Uniform(0, 1) and let Y = X 2 . What is the density for Y ?
Since X takes values on (0, 1) and since Y = X 2 , it will also be the case that Y
takes values on (0, 1). However, though X is uniform on the interval, there should be no
expectation that Y will also be uniform. In fact, since squaring a positive number less
than one results in a smaller number than the original, it should seem intuitive that results
of Y will be more likely to be near to zero than they are to be near to one.
It is not easy to see how to calculate the density of Y directly from the density of X.
However, it is a much easier task to compute the distribution of Y from the distribution of
X. Therefore we will use the following plan in the calculation below – integrate fX (x) to
find FX (x); use FX (x) to determine FY (y ); then differentiate FY (y ) to calculate fY (y ).
For the first step, note

 0 if 0 < x
Z x 

FX (x) = fX (x) dx = x if 0 ≤ x ≤ 1
−∞ 
 1 if x > 1


Version: – November 19, 2024

 5.3 transformations of continuous random variables 173

Next, since Y takes values in (0, 1), if y ≤ 0 then FY (y ) = P (Y ≤ y ) = 0. But if y > 0

then
√ √
FY (y ) = P (Y ≤ y ) = P (X 2 ≤ y ) = P (− y ≤ X ≤ y ).
√
Since X is always positive, the event (X < − y ) has zero probability we may connect this
to the distribution of X by writing

√ √
FY ( y ) = P ( − y ≤ X ≤ y )
√ √ √
= P (X < − y ) + P (− y ≤ X ≤ y )
√ √ √
= P ((X < − y ) ∪ (− y ≤ X ≤ y ))
√ √
= P (X ≤ y ) = FX ( y ).

Therefore, 
 0 if 0 ≤ y


√
FY (y ) = y if 0 < y < 1

 1

if y ≥ 1

and finally by using the fact that F ′ (y ) = f (y ) we can determine that


 1
√
2 y if 0 < y < 1
fY (y ) =
 0 otherwise

As noted in the beginning of this example, this distribution is far from uniform and gives
much more weight to intervals close to zero than it does intervals close to one. ■

Lemma 5.3.2. Let a ̸= 0 and b ∈ R. Suppose X is a continuous random variable with

probability density function fX . Let g (x) = ax + b be any non-constant linear function
(so a ̸= 0) and let Y = g (X ) then Y is also a continuous random variable whose density
function fY is given by
1 y−b
fY (y ) = fX ( ), (5.3.1)
|a| a
for all y ∈ R.

Proof- Let y ∈ R. Assume first that a > 0. Then

y−b
y−b
Z
a
P (Y ≤ y ) = P (aX + b ≤ y ) = P (X ≤ )= fX (z )dz
a −∞

By a simple change of variable z = u−b

a we obtain that

1
Z y
u−b
P (Y ≤ y ) = fX ( )du. (5.3.2)
−∞ a a

Version: – November 19, 2024

174 continuous probabilities and random variables

If a < 0 then
Z ∞
y−b
P (Y ≤ y ) = P (aX + b ≤ y ) = P (X ≥ )= fX (z )dz
a y−b
a

Again a simple change of variable z = a ,

u−b
with a < 0, we obtain that

1
Z y
u−b
P (Y ≤ y ) = fX ( )du. (5.3.3)
−∞ −a a

Using (5.3.2) and (5.3.3) we have that Y is a continuous random varable with density as in
(5.3.1). ■
Lemma 5.3.2 provides a method to standardize the normal random variable.

Corollary 5.3.3. (a) Let X ∼ N ormal (0, 1) and let Y = aX + b with a, b ∈ R, a ̸= 0.

Then, Y ∼ N ormal (b, a2 ).

X−µ
(b) Let X ∼ N ormal (µ, σ 2 ) and let Z = σ . Then Z ∼ N ormal (0, 1).

Proof - X has a probability density function given by (5.2.7).

(a)By Lemma 5.3.2, we have that the density of Y is given by

1 y−b 1 (z−b)2
 
fY (y ) = fX =√ e− 2a2 ,
|a| a 2π | a |

for all y ∈ R. Hence Y ∼ Normal (b, a2 ).

(b) By Lemma 5.3.2, with a = 1

σ and b = − σµ we have that the density of Z is given by

µ 1
 
z2
fZ (z ) = σfX σ (z + ) = √ e− 2 ,
σ 2π

for all z ∈ R. Hence Z ∼ Normal (0, 1). ■

Example 5.3.4. Consider the two parallel lines in R2 , given by y = 0 and y = 1. Piku is
standing at the origin in the plane. She chooses an angle θ uniformly in (0, π ) and she
draws a line segment between the lines y = 0 and y = 1 at an angle θ from the origin in R2 .
Suppose the line segment meets the line y = 1 at the point (X, 1). Find the probability
density function of X.

Version: – November 19, 2024

 5.3 transformations of continuous random variables 175

(0,1) (X, 1)

θ
(0,0)

Figure 5.9: Illustration of Example 5.3.4.

First observe that X = tan( π2 − θ ). We shall first find the distribution function of X.
Let x ∈ R. Observe that tan(x) is a strictly increasing function in the interval (− π2 , π2 )
and has an inverse denoted by arctan(x). So

π
P (X ≤ x) = P (tan( − θ ) ≤ x)
2
π
= P (( − θ ) ≤ arctan(x))
2
π
= P (θ ≥ − arctan(x))
2
π
= 1 − P (θ ≤ − arctan(x))
2

For any x ∈ R, π
2 − arctan(x) ∈ (0, π ). As θ has Uniform (0, π ) distribution, the above is

1 π
= 1 − ( − arctan(x))
π 2
1 1
= + arctan(x)
2 π

Hence the distribution function of X is differentiable and therefore the probability density
function of X is given by
1 1
fX (x) = ,
π 1 + x2
for all x ∈ R. Such a random variable is an example of a Cauchy distribution which we
define more generally next. ■

Version: – November 19, 2024

176 continuous probabilities and random variables

Cauchy(0, 1) Cauchy(1, 2)

Density f(x) Distribution F(x)

1.0
0.3
0.8

0.2 0.6

0.4
0.1

0.2

0.0

0 5 0 5

Figure 5.10: The shape of Cauchy density and cumulative distribution functions for selected
parameter values.

Definition 5.3.5. X ∼ Cauchy(θ, α2 ): Let θ ∈ R and let α > 0. Then X is said

to have a Cauchy distribution with parameters θ and α2 if it has the density

1 α
f (x) = (5.3.4)
π α2 + ( x − θ ) 2

for all x ∈ R. Here θ is referred to as the location parameter and α is referred to as

the scale parameter. The distribution function of X is given by

1 x−θ
F (x) = arctan( ) (5.3.5)
π α

Figure 5.10 gives plots of the Cauchy density and distribution functions.
Similar computations as above are useful for simulations. Most computer progam-
ming languages and spreadsheets have a “Random” function designed to approximate a
Uniform(0, 1) random variable. How could one use such a feature to simulate random
variables with other densities? We start with an example.

Example 5.3.6. If X ∼ Uniform(0, 1), our goal is to find a function g : (0, 1) → R for
which Y = g (X ) ∼ Exponential (λ). We will try to find such a g : (0, 1) → R which
is strictly increasing so that it has an inverse. This will be important when it comes to
relating the distributions of X and Y .

Version: – November 19, 2024

 5.3 transformations of continuous random variables 177

We require Y to Exponential(λ). So the distribution function of Y is

0 if y ≤ 0
(
FY (y ) =
1 − e−λy if y > 0

But
FY (y ) = P (Y ≤ y ) = P (g (X ) ≤ y ) = P (X ≤ g −1 (y ))

where the final equality comes from our decree that the function g should be strictly
increasing. Therefore,
FY (y ) = FX (g −1 (y )).

But the distribution function of a uniform random variable has previously been computed.
Hence, 


 0 if g −1 (y ) ≤ 0
FX (g −1 (y )) = g −1 (y ) if 0 < g −1 (y ) < 1

1 if g −1 (y ) ≥ 1



Thus we are forced to have

g −1 (y ) = 1 − e−λy

for y > 0. So inverting the above formula, we get g : (0, 1) → (0, ∞) is given by

1
g (x) = − log(1 − x),
λ

for x ∈ (0, 1). Hence,

1
X ∼ Uniform(0, 1) =⇒ − log(1 − X ) ∼ Exponential(λ).
λ

In conclusion one could view g as the inverse of FY , on (0, ∞). It turns out that this is
a general result. We state a special case of this in the lemma below. ■

Lemma 5.3.7. Let U ∼ Uniform (0, 1) random variable. Let X be a continuous random
variable such that its distribution function, FX , is a strictly increasing continous function.
Then

(a) Y = FX−1 (U ) has the same distribution as X.

(b) Z = FX (X ) has the same distribution as U .

Proof- We observe that as F is strictly increasing continuous distribution function

F : R → (0, 1) and Range (F ) = (0, 1).

Version: – November 19, 2024

178 continuous probabilities and random variables

(a) We shall verify that Y and X have the same distribution function. Let y ∈ R, then

FY (y ) = P (Y ≤ y ) = P (FX−1 (U ) ≤ y ) = P (U ≤ FX (y )) = FX (y )

Hence X and Y have the same distribution.

(b) We shall verify that Z and U have the same distribution function. Let z ∈ R. If
z ≤ 0 then

P (Z ≤ z ) = P (F (X ) ≤ z ) = 0

as F : R → (0, 1). If z ≥ 1 then

P (Z ≤ z ) = P (F (X ) ≤ z ) = 1

as F : R → (0, 1). If 0 < z < 1 then F −1 (z ) is well defined as Range (F ) = (0, 1) and

P (Z ≤ z ) = P (F (X ) ≤ z ) = P (X ≤ F −1 (z )) = F (F −1 (z )) = z.

Hence Z and U have the same distribution. ■

The previous lemma may be generalized even to the case when F is not strictly
increasing. It requires a concept called the generalized inverse. The interested reader will
find it discussed in Exercise 5.3.12.

exercises

√
Ex. 5.3.1. Let X ∼ Uniform(0, 1) and let Y = X. Determine the density of Y .
Ex. 5.3.2. Let X ∼ Uniform(0, 1) and let Z = X.
1
Determine the density of Z.
Ex. 5.3.3. Let X ∼ Uniform(0, 1). Let r > 0 and define Y = rX. Show that Y is uniformly
distributed on (0, r ).
Ex. 5.3.4. Let X ∼ Uniform(0, 1). Let Y = 1 − X. Show that Y ∼ Uniform(0, 1) as well.
Ex. 5.3.5. Let X ∼ Uniform(0, 1). Let a and b be real numbers with a < b and let
Y = (b − a)X + a. Show that Y ∼ Uniform(a, b).
Ex. 5.3.6. Let X ∼ Uniform(0, 1). Find a function g (x) (which is strictly increasing) such
that the random variable Y = g (X ) has density fY (y ) = 3y 2 for 0 < y < 1 (and fY (y ) = 0
otherwise).
Ex. 5.3.7. Let X ∼ N ormal (µ, σ 2 ). Let g : (−∞, ∞) → R be given by g (x) = x2 . Find
the probability density function of Y = g (X ).

Version: – November 19, 2024

 5.3 transformations of continuous random variables 179

Pareto(1) Pareto(2)

Density f(x) Distribution F(x)

1.0

1.5 0.8

0.6
1.0
0.4
0.5
0.2

0.0 0.0

1 2 3 4 1 2 3 4

Figure 5.11: The shape of the pareto density and cumulative distribution functions.

Ex. 5.3.8. Let α > 0 and X be a random variable with the p.d.f given by

 αα+1
x
1≤x<∞
f (x) =
0 otherwise

The random variable X is said to have Pareto (α) distribution (see Figure 5.11).

(a) Find the distribution of X1 = X 2

(b) Find the distribution of X2 = 1

X

(c) Find the distribution of X3 = log(X )

In the above exercises we assume that the transformation function is defined as above
when the p.d.f of X is positive and zero otherwise.
Ex. 5.3.9. Let X be a continuous random variable with probability density function
fX : R → R. Let a > 0, b ∈ R Y = a1 (X − b)2 . Show that Y is also a continuous random
variable with probability density function fY : R → R given by
√
a √ √
fY (y ) = √ [fX ( ay + b) + fX (− ay + b)]
2 y

for y > 0.
Ex. 5.3.10. Let −∞ ≤ a < b ≤ ∞ and I = (a, b) and g : I → R. Let X be a continuous
random variable whose density fX is zero on the complement of I. Set Y = g (X ).

(a) Let g be a differentiable strictly increasing function.

(i) Show that inverse of g exists and g −1 is strictly increasing on g (I ).

Version: – November 19, 2024

180 continuous probabilities and random variables

(ii) For any y ∈ R, show that P (Y ≤ y ) = P (X ≤ g −1 (y ))

(iii) Show that Y has a density fY (·) given by

d −1
fY (y ) = fX (g −1 (y )) g (y ).
dy

(b) Let g be a differentiable strictly decreasing function.

(i) Show that inverse of g exists and g −1 is strictly decreasing on g (I ).

(ii) For any y ∈ R, show that P (Y ≤ y ) = 1 − P (X ≤ g −1 (y ))
(iii) Show that Y has a density fY (·) given by

d
 
−1
fY (y ) = fX (g (y )) − g −1 (y ) .
dy

Ex. 5.3.11. Let X be a random variable having an exponential density. Let g : [0, ∞) → R
1
be given by g (x) = x β , for some β ̸= 0. Find the probability density function of Y = g (X ).
Ex. 5.3.12. Let U ∼ Uniform (0, 1). Let X be a continuous random variable with a
distribution function F . Extend F : R → R to F : R ∪ {−∞} ∪ {∞} → R by setting
F (∞) = 1 and F (−∞) = 0. Define the generalised inverse of F , G : [0, 1] → R ∪ {−∞} ∪
{∞} by
G(y ) = inf{x ∈ R : F (x) ≥ y}.

Show that

(a) Show that for all y ∈ [0, 1], F (G(y )) = y.

(b) Show that for all x ∈ R and y ∈ [0, 1]

F (x) ≥ y ⇐⇒ x ≥ G(y ).

(c) Y = G(U ) has the same distribution as X.

(d) Z = F (X ) has the same distribution as U .

5.4 multiple continuous random variables

When analyzing multiple random variables at once, one may consider a “joint density”
analogous to the joint distribution of the discrete variable case. In this section we will
restrict considerations to only two random variables, but we shall see in Chapter 8 that
the definitions and results all generalize to any finite collection of variables.

Version: – November 19, 2024

 5.4 multiple continuous random variables 181

Theorem 5.4.1. Let f : R2 → R be a non-negative function, piecewise-continuous

in each variable for which
Z ∞ Z ∞
f (x, y ) dx dy = 1.
−∞ −∞

For a Borel set A ⊂ R2 define

Z
P (A) = f (x, y ) dx dy.
A

Then P is a probability on R2 and f is called the density for P .

Proof- The proof of the theorem is essentially the same as in the one-variable version
of Theorem 5.1.5. We will not reproduce it here. As in the discrete case we will typically
associate such densities with random variables. ■

Definition 5.4.2. A pair of random variables (X, Y ) is said to have a joint density
f (x, y ) if for every Borel set A ⊂ R2
Z
P ((X, Y ) ∈ A) = f (x, y ) dx dy.
A

As in the one-variable case we describe this in terms of “Borel sets” to be precise, but
in practice we will only consider sets A which are simple regions in the plane. In fact
regions such as (−∞, a] × (−∞, b], for all real numbers a, b are enough to characterise the
joint distribution. As in the one variable case we can define a “joint distribution function”
of (X, Y ) as
Z a Z b
F(X,Y ) (a, b) = P ((X ≤ a) ∩ (Y ≤ b)) = f (z, w )dwdz (5.4.1)
−∞ −∞

for all a, b ∈ R. We will usually denote the joint distribution function by F omiting
the subscripts unless it is particularly needed. One can state and prove a similar type
of result as Theorem 5.2.5 for F (a, b) when (X, Y ) have a joint density. In particular,
we can conclude that since the joint densities are assumed to be piecewise continuous,
the corresponding distribution functions are piecewise differentiable. Further, the joint
distribution of two continuous random variables (X, Y ) are completely determined by their
joint distribution function F . That is, if we know the value of F (a, b) for all a, b ∈ R, we
could use multivariable calculus to differentiate F (a, b) to find f (a, b). Then P ((X, Y ) ∈ A)

Version: – November 19, 2024

182 continuous probabilities and random variables

for any event A is obtained by integrating the joint density f over the event A. We illustrate
this with a couple of examples.

Example 5.4.3. Consider the open rectangle in R2 given by R = (0, 1) × (3, 5) and
| R |= 2 denote its area. Let (X, Y ) have a joint density f : R2 → R given by

1
2 if (x, y ) ∈ R
f (x, y ) =
0 otherwise.

The above is clearly a density function. So for any recntangle A = (a, b) × (c, d) ⊂ R,
Z dZ b
(b − a)(d − c) |A|
P ((X, Y ) ∈ A) = f (x, y )dxdy = = .
c a 2 |R|

In general one can use the following definition to define a uniform random variable on
the plane.

Definition 5.4.4. Let D ⊂ R2 be non-empty and with positive area (assume D

is a Borel set or in particular f or any simple region whose area is well defined).
Then (X, Y ) ∼Uniform (D ) if it has a joint probability density function given by
f : R2 → R given by

1

|D| if (x, y ) ∈ D
f (x, y ) =
 0 otherwise,

where | D | denotes the area of D.

When (X, Y ) ∼ Uniform (D ) then the probability that (X, Y ) lies in a region A ⊂ D
is proportional to the area of A.

Example 5.4.5. Let (X, Y ) have a joint density f : R2 → R given by

if 0 < x < 1, 0 < y < 1

(
x+y
f (x, y ) =
0 otherwise

Version: – November 19, 2024

 5.4 multiple continuous random variables 183

We note that this really does describe a density. The function f (x, y ) is non-negative and
Z ∞ Z ∞ Z 1Z 1
f (x, y ) dx dy = x + y dx dy
−∞ −∞ 0 0
1
Z 1
= ( x2 + xy ) |xx= 1
=0 dy
0 2
1
Z 1
= + y dy
0 2
1 1
= y + y 2 |yy = 1
=0 = 1.
2 2

Calculating a probability such as P ((X < 12 ) ∩ (Y < 12 )) requires integrating over the
appropriate region.

1 1
Z 1/2 Z 1/2
P ((X < ) ∩ (Y < )) = f (x, y ) dx dy
2 2 −∞ −∞
Z 1/2 Z 1/2
= x + y dx dy
0 0
1 1
Z 1/2
= + y dy
0 8 2
1
= .
8

A probability only involving one variable may still be calculated from the joint density.
For instance P (X < 12 ) does not appear to involve Y , but this simply means that Y is
unrestircted and the corresponding integral should range over all possible values of Y .
Therefore,

1
Z ∞ Z 1/2
P (X < ) = f (x, y ) dx dy
2 −∞ −∞
3
Z 1 Z 1/2
= x + y dx dy = .
0 0 8

It is just as easy to compute that P (Y < 12 ) = 38 . Note that these computations also
demonstrate that X and Y are not independent since

1 1 9 1 1
P (X < ) · P (Y < ) = ̸= P ((X < ) ∩ (Y < )).
2 2 64 2 2

Version: – November 19, 2024

184 continuous probabilities and random variables

(0,1)

(0,0) (1,0)

Figure 5.12: The subset A of the unit square that represents the region x + y < 1.

A probability such as P (X + Y < 1) can be found by integrating over a non-rectangular

region in the plane, as shown in Figure 5.12. Let A = {(x, y )|x + y < 1}. Then
Z
P (X + Y < 1) = f (x, y ) dx dy
A
Z 1 Z 1−y
= x + y dx dy
0 0
1 2
Z 1
= x + xy |01−y dy
0 2
1
Z 1
= (1 − y )2 + (1 − y )y dy
0 2
1 1 2
Z 1
= − y dy
0 2 2
1
= .
3

5.4.1 Marginal Distributions

As in the discrete case, when we begin with the joint density of many random variables,
but want to speak of the distribution of an individual variable we will frequently refer to it
as a “marginal distribution” .

Version: – November 19, 2024

 5.4 multiple continuous random variables 185

Suppose (X, Y ) are random variables and have a joint probability density function
f : R2 → R. Then we obseve that
Z x Z ∞
P (X ≤ x) = P (X ≤ x, −∞ < Y < ∞) = f (u, y )dydu.
−∞ −∞

If g : R → R is given by Z ∞
g (u) = f (u, y )dy
−∞

then Z x
P (X ≤ x) g (u)du.
−∞

Using Theorem 5.2.5, by the continuity assumptions on f , we find that the random variable
X is also a continuous random variable with probability density function of X given by
Z ∞
fX (x) = g (x) = f (x, y )dy. (5.4.2)
−∞

As it was derived from a joint probability density function, the density of X is referred to
as the marginal density of X. Similarly one can show that Y is also a continuous random
variable and its marginal density is given by
Z ∞
fY (y ) = f (x, y )dx. (5.4.3)
−∞

Example 5.4.6. (Example 5.4.3 contd.) Going back to Example 5.4.3, we can compute
the marginal density of X and Y . The marginal density of X is given by
( R5
if 0 < x < 1 1 if 0 < x < 1
(
Z ∞ 1
3 2
fX (x) = f (x, y )dy = =
−∞ 0 otherwise. 0 otherwise.

The marginal density of Y is given by

( R1
if 3 < y < 5 if 3 < y < 5
(
Z ∞ 1 1
0 2 2
fY (y ) = f (x, y )dx = =
−∞ 0 otherwise. 0 otherwise.

So we observe that X ∼ Uniform (0, 1) and Y ∼ Uniform (3, 5). ■

While it is routine to find the marginal densities from the joint density there is no
standard way to get to the joint from the marginals. Part of the reason for this difficulty
is that the marginal desnitieis offer no information about how the varaibles relate to each
other, which is critical information for determining how they behave jointly. However,
in the case that the random variables happen to be independent there is a convenient
relationship between the joint and marginal densities.

Version: – November 19, 2024

186 continuous probabilities and random variables

5.4.2 Independence

Theorem 5.4.7. Let f be the joint density of random variables X and Y and let
fX and fY be the respective marginal densities. Then

f (x, y ) = fX (x)fY (y )

if and only if X and Y are independent.

Proof - First suppose X and Y are independent and consider the quantity P ((X ≤
x) ∩ (Y ≤ y )). On one hand independnece gives

P ((X ≤ x) ∩ (Y ≤ y )) = P (X ≤ x)P (Y ≤ y ) = FX (x)FY (y ) (5.4.4)

On the other hand, integrating the joint density yields

Z x Z y
P ((X ≤ x) ∩ (Y ≤ y )) = f (x, y ) dx dy. (5.4.5)
−∞ −∞

Since equations 5.4.4 and 5.4.5 are equal we may differentiate both with respect to each of
the variables x and y and they remain equal. However, differentiating the former gives
fX (x)fY (y ) because of the relationship between the distribution and the density, while
differentiating the latter yields f (x, y ) by a two-fold application of the fundamental theorem
of calculus.
To prove the opposite direction, suppose f (x, y ) = fX (x)fY (y ). Let A and B be Borel
sets in R. Then
Z Z
P ((X ∈ A) ∩ (Y ∈ B )) = f (x, y ) dx dy
ZB ZA
= fX (x)fY (y ) dx dy
B A
Z  Z 
= fX (x) dx fY (y ) dy
A B
= P (X ∈ A)P (Y ∈ B )

Since this is true for all sets such sets A and B, the variables X and Y are independent. ■

Example 5.4.8. (Example 5.4.3 contd.) We had observed that if (X, Y ) ∼ Uniform (R)
then X ∼ Uniform (0, 1) and Y ∼ Uniform (3, 5). Note further that

f (x, y ) = fX (x)fY (y )

Version: – November 19, 2024

 5.4 multiple continuous random variables 187

for all x, y ∈ R. Consequently X, Y are independent as well. ■

It is tempting to generalise and say that (X, Y ) ∼ Uniform (D ) for a region D with
non-trivial area then X and Y would be independent. This is not the case, we illustrate in
the example below.
Example 5.4.9. Consider the open disk in R2 given by C = {(x, y ) : x2 + y 2 < 25} and
| C |= 25π denote its area. Let (X, Y ) have a joint density f : R2 → R given by

 1
|C| if (x, y ) ∈ C
f (x, y ) =
 0 otherwise.

As before for any Borel A ⊂ C,

|A|
P ((X, Y ) ∈ A) = ,
|C|

and the probability that (X, Y ) lies in A is proportional to the area of A. However the
marginal density calculation is a little different. The marginal density of X is given by
 Z √
25−x2 1
if − 5 < x < 5
Z ∞ 

√ dy
fX (x) = f (x, y )dy = − 25−x2 |C|
−∞
0 otherwise.


√
25 − x2 if − 5 < x < 5
(
2
= 25π
0 otherwise.

The distribution of X is the Semi-circular law described in Exercise 5.2.6. As the joint
density f is symmetric in x and y (i.e f (x, y ) = f (y, x)) the marginal density of Y is the
same as that of X (why ?). It is easy to see

1 4
= f (0, 0) ̸= fX (0)fY (0) =
25π 25π 2

Consequently X, Y are not independent. This fact should make intuitive sense as well, for
if X happens to take a value near 5 or −5 the range of possible values of Y is much more
restricted than if X takes a value near 0. ■
We shall see the utility of independence when computing distributions of various
functions of independent random variables (see Section 5.5). Independence of random
variables also makes it easier to compute their joint density and hence probabilites. For
instance, consider the following example.
Example 5.4.10. Suppose X ∼ Exponential(λ1 ), Y ∼ Exponential(λ2 ) are independent
random variables. Find P (X − Y < 0).

Version: – November 19, 2024

188 continuous probabilities and random variables

The joint density of (X, Y ) is given by

if x > 0 and y > 0

(
λ1 λ2 e−(λ1 x+λ2 y )
f (x, y ) = fX (x)fY (y ) =
0 otherwise

Therefore
Z ∞Z y Z ∞ Z y
− ( λ1 x + λ2 y ) −λ2 y
P (X − Y < 0) = λ1 λ2 e dxdy = λ1 λ2 e [ e−λ1 x dx]dy
0 0 0 0
1
Z ∞
= λ1 λ2 e−λ2 y [1 − e−λ1 y ]dy
0 λ1
Z ∞ 
= λ2 e−λ2 y − e−(λ1 +λ2 )y dy
 0
−1 −λ2 y ∞ 1

= λ2 (e |0 ) + (e−(λ1 +λ2 )y |∞
0 )
λ2 λ1 + λ2
1 1
 
= λ2 −
λ2 λ1 + λ2
λ1
= .
λ1 + λ2

Similarly one can also compute P (Y − X < 0) = λ1 + λ2 .

λ2
This fact is quite useful when using
exponential random variables to model waiting times, for P (X − Y < 0) = P (X < Y ), so
we have determined the probability that one waiting time will be shorter than another. ■

5.4.3 Conditional Density

In Section 3.2.2 we have seen the notion of conditional distributions for discrete random
variables and in Section 4.4 we have seen the notions of conditional expectation and
variance for discrete random variables. Suppose X measures the parts per million of a
particulate matter less than 10 microns in the air and Y is the incidence rate of asthma in
the population. It is clear that X and Y ought to be related; for the distribution of one
affects the other. Towards this, in this section we shall discuss conditional distributions for
two continuous random variables having a joint probability density function. We recall
from Definition 3.2.5 that if X is a random variable on a sample space S and A ⊂ S be an
event such that P (A) > 0, then the probability Q described by

Q(B ) = P (X ∈ B|A)

is called the conditional distribution of X given the event A.

Version: – November 19, 2024

 5.4 multiple continuous random variables 189

Suppose X and Y have a joint probability density function f . Given our discussion
for discrete random variables it is natural to characterise the conditional distribution of
X given some information on Y . In the discrete setting we typically considered an event
A = {Y = b} for some real number b in the range of Y . In the continuous setting such an
event A would have zero probability, so the usual way of conditioning on an event would
not be possible. However, there is a way to make such a conditioning meaningful and
precise provided fY (b) > 0, where fY is the marginal density of Y .

Suppose we wish to find the following :

P (X ∈ [3, 4] | Y = b).

We shall argue heuristically and arrive at an expression for the above probability. Suppose
the marginal density of X is fX (·), and that of Y is fY (·). Assume first that fY is piecewise
continuous and fY (b) > 0. Then it is a standard fact from real analysis to see that

1
P (Y ∈ [b, b + )) > 0,
n

for all n ≥ 1. One can then view the conditional probability as before, that is

1 P (X ∈ [3, 4] ∩ X ∈ [b, b + n1 ))
P (X ∈ [3, 4] | X ∈ [b, b + )) =
n P (X ∈ [b, b + n1 ))
 
R 4 R b+ n1
3 b f (u, v )du dv
= R b+ n1
fX (u)du
 b 
R4 R b+ n1
3 n b f (u, v )du dv
= R b+ n1
n b fX (u)du

From facts in real analysis (under some mild assumptions on f ) the following can be
established,
Z b+ 1
n
lim n f (u, v )du = f (b, v ),
n→∞ b
for all real numbers v and
Z b+ 1
n
lim n fX (u)du = fX (b).
n→∞ b

Version: – November 19, 2024

190 continuous probabilities and random variables

We have seen earlier (see Exercise 1.1.13 (b))

1
lim P (Y ∈ [b, b + )) = P (Y = b).
n→∞ n

Hence it would be reasonable to argue that P (X ∈ [3, 4] | Y = b) ought to be defined as

R4
f (b, v )dv
P (X ∈ [3, 4] | Y = b) = 3
.
fY (b)

With the above motivation we are now ready to define conditional densities for two random
variables.

Definition 5.4.11. Let (X, Y ) be random variables having joint density f . Let the
marginal density of Y be fY (·). Suppose b is a real number such that fY (b) > 0 and
is continuous at b then conditional density of X given Y = b is given by

f (x, b)
fX|Y =b (x) = (5.4.6)
fY (b)

for all real numbers x. Similarly, let the marginal density of X be fX (·). Suppose a
is a real number such that fX (a) > 0 and is continuous at a then conditional density
of Y given X = a is given by

f (a, y )
fY |X =a (y ) =
fX (a)

for all real numbers y.

This definition genuinely defines a probability density function, for fX|Y =b (x) ≥ 0 since
it is the ratio of a non-negative quantity and a positive quantity. Moreover,
Z ∞ Z ∞
f (x, b)
fX|Y =b (x)dx = dx
−∞ −∞ fY (b)
1 1
Z ∞
= f (x, b)dx = fY (b) = 1
fY (b) −∞ fY (b)

Note that if X and Y are independent then

f (x, b) fX (x)fY (b)

fX|Y =b (x) = = = fX (x).
fY (b) fY (b)

Version: – November 19, 2024

 5.4 multiple continuous random variables 191

One can use the conditional density to compute the conditional probabilities, namely if
(X, Y ) are random variables having joint density f and b is a real number such that its
marginal density has the property fY (b) > 0 then

f (x, b)
Z Z
P (X ∈ A | Y = b) = fX|Y =b (x)dx = dx.
A A fY (b)

We conclude this section with two examples where we compute conditional densities.
In both the examples the dependencies between the random variables imply that the
conditional distributions are different from the marginal distributions.

Example 5.4.12. Let (X, Y ) have joint probability density function f given by
√
3 − 1 (x2 −xy +y2 )
f (x, y ) = e 2 − ∞ < x, y < ∞.
4π

Let x ∈ R, then the marginal density of X at x is given by

Z ∞ √
3 − 1 (x2 −xy +y2 )
Z ∞
fX (x) = f (x, y )dy = e 2 dy
−∞ −∞ 4π

3x2
By a standard completing the square computation, 1
2 (x
2 − xy + y 2 ) = 8 + 21 (y − x2 )2 .
Therefore, √
3 − 3x2
Z ∞
1 x 2
fX (x) = e 8 e− 2 (y− 2 ) dy
4π −∞
R∞ 1 x 2
Observing that −∞
√1 e− 2 (y− 2 ) dy
2π
= 1 (why ?), we have
√
3 − 3x2 √ 3 1 − 3x2
r
fX (x) = e 8 2π = √ e 8
4π 4 2π

Hence X is a Normal random variable with mean 0 and variance 3.

4
By symmetry (or
calculating similarly as above) we can also show that Y is a Normal random variable with
mean 0 and variance 43 . Also, we can easily see that
√
3 − 3 ( x2 + y 2 ) 3 − 1 (x2 −xy +y2 )
fX (x)fY (y ) = e 8 ̸= e 2 = f (x, y )
8π 4π

Version: – November 19, 2024

192 continuous probabilities and random variables

for many x, y ∈ R. Hence X and Y are not independent. Note that fX (x) ̸= 0 for all real
numbers x and is continuous at all x ∈ R. Fix x ∈ R, the conditional density of Y given
X = x is given by
√
3 − 12 (x2 −xy +y 2 )
f (x, y ) 4π e 1 1 x 2
fY |X =x (y ) = = = √ e− 2 (y− 2 ) ∀ y ∈ R.
2π
q 2
fX (x) 3 √1 − 3x8
4 2π e

Hence though the marginal distribution of Y is Normal(0, 43 , the)the conditional distribution

of Y given X = x is Normal with mean x
2 and variance 1. Put another way, if we are given
that X = x the mean of Y changes from 0 to x and the variance reduces from 4
3 to 1.
Such a pair (X, Y ) is an example of a bivariate normal random variable and will be
discussed in detail in Section 6.4. ■

(0,4) (4,4)
X=1

Y=2

(0,0)

Figure 5.13: The region T = {(x, y ) | 0 < x < y < 4} from Example 5.4.13.

Example 5.4.13. Suppose T = {(x, y ) | 0 < x < y < 4} and let (X, Y ) ∼ Uniform (T ).
Therefore its joint density is given by (see Figure 5.13)

if (x, y ) ∈ T
(
1
f (x, y ) = 8
0 otherwise.

The marginal density of X is given by

( R4
if 0 < x < 4 if 0 < x < 4
(
Z ∞ 1 4−x
x 8 dy 8
fX (x) = f (x, y )dy = =
−∞ 0 otherwise. 0 otherwise.

Version: – November 19, 2024

 5.4 multiple continuous random variables 193

The marginal density of Y is given by

( Ry ( y
1
if 0 < y < 4 if 0 < y < 4
Z ∞
0 8 dy 8
fY (y ) = f (x, y )dx = =
−∞ 0 otherwise. 0 otherwise.

Let us fix 0 < b < 4. So fY (·) is non-zero at b and is continuous at b. The conditional
density of (X | Y = b) is given by

if 0 < x < b if 0 < x < b

( (
1/8 1
f (x, b) b/8 b
fX|Y =b (x) = = =
fY (b) 0 otherwise. 0 otherwise.

Therefore (X | Y = b) ∼ Uniform (0, b). Similarly if we fix 0 < a < 4, we observe fX (·) is
non-zero at a and is continuous at a. The conditional density of (Y | X = a) is given by
 
f (a, y )  1/8
if a < y < 4  1 if a < y < 4
fY |X =a (y ) = = (4−a)/8 = 4−a
fX (a) 0 otherwise. 0 otherwise.

Therefore (Y | X = a) ∼ Uniform (a, 4).

Clearly X and Y are continuous random variables with distributions that are not
uniform, but the conditional distributions turn out to be uniform. ■

exercises

Ex. 5.4.1. Let (X, Y ) be random variables whose probability density function is given by
f : R2 → R. Find the probability density function of X and probability density function
of Y in each of the following cases:-

(a) f (x, y ) = (x + y ) if 0 ≤ x ≤ 1, 0 ≤ y ≤ 1 and 0 otherwise

(b) f (x, y ) = 2(x + y ) if 0 ≤ x ≤ y ≤ 1 and 0 otherwise

(c) f (x, y ) = 6x2 y if 0 ≤ x ≤ 1, 0 ≤ y ≤ 1 and 0 otherwise

(d) f (x, y ) = 15x2 y if 0 ≤ x ≤ y ≤ 1 and 0 otherwise

Ex. 5.4.2. Let c > 0. Suppose that X and Y are random variables with joint probability
density
c(xy + 1) if 0 ≤ x ≤ 1 and 0 ≤ y ≤ 1
(
f (x, y ) =
0 otherwise

(a) Find c.

Version: – November 19, 2024

194 continuous probabilities and random variables

(b) Compute the marginal densities fX (·) and fY (·) and the conditional density fX|Y =b (·)

Ex. 5.4.3. Let A = {(x, y ) ∈ R2 : x > 0, y > 0, x + y < 1} and let X and Y be random
variables defined by the joint density f (x, y ) = 24xy if (x, y ) ∈ A (and f (x, y ) = 0
otherwise).

(a) Verify the claim that f (x, y ) is a density.

(b) Show that X and Y are dependent random variables.

(c) Explain why (b) doesn’t violate Theorem 5.4.7 despite the fact that 24xy is a product
of a function of x with a function of y.

Ex. 5.4.4. Consider the set D = [−1, 1] × [−1, 1]. Let

L = {(x, y ) ∈ D : x = 0 or or x = −1 or x = 1 or y = 0 or y = 1 or y = −1}

be the lines that create a tiling of D. Suppose we drop a coin of radius R at a uniformly
chosen point in D what is the probability that it will intersect the set L ?
Ex. 5.4.5. Let X and Y be two independent uniform (0, 1) random variables. Let
U = max(X, Y ) and V = min(X, Y ).

(a) Find the joint distribution of U , V .

(b) Find the conditional distribution of (V | U = 0.5)

Ex. 5.4.6. Suppose X is a random variable with density

cx2 (1 − x) for 0 ≤ x ≤ 1,
(
f (x) =
0 otherwise.

Find:

(a) the value of c.

(b) the distribution function of X.

(c) the conditional probability P (X > 0.2 | X < 0.5).

Ex. 5.4.7. Suppose g : R → R be a continuous probability density function, such that

g (x) = 0 when x ̸∈ [0, 1]. Let D ⊂ R2 be given by

D = {(x, y ) : x ∈ R and 0 ≤ y ≤ g (x)}

Let (X, Y ) be uniformly distributed on D. Find the probability density function of X.

Version: – November 19, 2024

 5.5 functions of independent random variables 195

Ex. 5.4.8. Continuous random variables X and Y have a joint density

24 , for 0 < x < 6, 0 < y < 4

(
1
f (x, y ) =
0, elsewhere.

(a) Find P (2Y > X ).

(b) Are X and Y independent?

Ex. 5.4.9. Let

if 0 ≤ x < y ≤ 1
(
η (y − x)γ
f (x, y ) =
0 otherwise
(a) For what values of γ can η be chosen so that f be a joint probability density function
of X, Y .

(b) Given a γ from part (a), what is the value of η ?

(c) Given a γ and η from parts (a) and (b), find the marginal densities of X and Y .

Ex. 5.4.10. Let D = {(x, y ) : x3 ≤ y ≤ x}. A point (X, Y ) is chosen uniformly from D.
Find the joint probability density function of X and Y .
Ex. 5.4.11. Let X and Y be two random variables with the joint p.d.f given by

 ae


−by 0≤x≤y
f (x, y ) =

 0

otherwise

Find a conditions on a and b that make this a joint probability density function.
Ex. 5.4.12. Suppandi and Meera plan to meet at Gopalan Arcade between 7pm and 8pm.
Each will arrive at a time (independent of each other) uniformly between 7pm and 8pm
and will wait for 15 minutes for the other person before leaving. Find the probability that
they will meet ?

5.5 functions of independent random variables

In Section 5.3 we have seen how to compute the distribution of Y = g (X ) from the
distribution of X for various g : R2 → R. Suppose (X, Y ) are random variables having
a joint probability density function f : R2 → R. Let h : R2 → R. A natural follow up
objective is then to determine the distribution of

Z = h(X, Y ).

Version: – November 19, 2024

196 continuous probabilities and random variables

In Section 3.3 we discussed an approach to this question when the random variables where
discrete.
One could prove a result as attained in Exercise 5.3.10 for functions of two variables
but this will require knowledge of Linear Algebra and multivariable calculus. Here we limit
our objective and shall focus on two specific functions namely the sum and the product.

5.5.1 Distributions of Sums of Independent Random variables

Let X and Y be two independent continous random variables with densities fX and fY .
In this section we shall see how to compute the distribution of Z = X + Y . We first prove
a proposition that describes the probability density function of Z.

Proposition 5.5.1. (Sum of two independent random variables) Let X and Y be two
independent random variables with marginal densities given by fX : R → R and fY : R → R.
Then Z = X + Y has a probability density function fZ : R → R given by
Z ∞
fZ (z ) = fX (x)fY (z − x)dx. (5.5.1)
−∞

Proof- Let us first find an expression for the distribution function of Z.

F (z ) = P (Z ≤ z )
= P (X + Y ≤ z )
Z Z
= fX (x)fY (y )dydx
{(x,y ):x+y≤z}
Z ∞ Z z−x
= fX (x)fY (y )dydx
−∞ −∞
Z zZ ∞
= [ fX (x)fY (u − x)dx]du.
−∞ −∞

As fX (·) and fY (·) are densities, it can be shown that the integrand is a piecewise continuous
function. Hence F is of the form (5.2.4) and Theorem 5.2.5 implies that the probability
density function of Z is given by (5.5.1). ■
The integral expression on the right hand side of (5.5.1) is referred to as the convolution
of fX and fY and is denoted by fX ⋆ fY (z ). It is a property of convolutions that fX ⋆
fY (z ) = fY ⋆ fX (z ) for all z ∈ R. Thus if we view the sum of X and Y as Z = X + Y or
Z = Y + X the distribution will be the same (See Exercise 5.5.8).

Version: – November 19, 2024

 5.5 functions of independent random variables 197

1.0

0.8

0.6

0.4

0.2

0.0

0.0 0.5 1.0 1.5 2.0

Figure 5.14: The region T = {(x, y ) | 0 < x < y < 4} from Example 5.5.2.

Example 5.5.2. (Sum of Uniforms) Let X and Y be two independent Uniform (0, 1)
random variables. Let Z = X + Y . From the above proposition that Z has a density given
by (5.5.1). Note that

1 if 0 < x < 1, 0 < z − x < 1 and 0 < z < 2
fX (x)fY (z − x) =
0 otherwise

Therefore fX (x)fY (z − x) is non-zero if and only if max{0, z − 1} < x < min{1, z}, 0 <
z < 2. So for 0 < z < 2,
Z min{1,z} Z min{1,z}
fZ (z ) = fX (x)fY (z − x)dx = 1dx = min{1, z} − max{0, z − 1}.
max{0,z−1} max{0,z−1}

Therefore,

 z
 if 0 < z ≤ 1
min{1, z} − max{0, z − 1} if 0 < z < 2
( 
fZ (z ) = = 2−z if 1 < z < 2
0 otherwise 
 0 otherwise


A graph of this density is displayed in Figure 5.14. ■

Our next example will deal with sum of two independent exponential random variables.
This will lead us to the Gamma distribution which is of significant interest in statistics.

Version: – November 19, 2024

198 continuous probabilities and random variables

Example 5.5.3. (Sum of Exponentials) Let λ > 0, X and Y be two independent Exponen-
tial (λ) random variables. Let Z = X + Y . Then we know and Z has a density given by
(5.5.1). Further,
 
λ2 e−λx e−λ(z−x) if x ≥ 0, z − x ≥ 0 λ2 e−λz if x ≥ 0, x ≤ z, z ≥ 0
fX (x)fY (z − x) = =
0 otherwise 0 otherwise

Hence fX (x)fY (z − x) is non-zero if and only if 0 ≤ x ≤ z. So

Z z Z z
fZ (z ) = fX (x)fY (z − x)dx = λ2 e−λz 1dx = λ2 ze−λz ,
0 0

for z ≥ 0 and fZ (z ) = 0 otherwise. This is known as Gamma (2, λ) distribution. ■

Before we define the Gamma distribution more generally we prove a lemma in real analysis,
the proof of which can be skipped upon first reading.

Lemma 5.5.4. For n ≥ 1, and λ > 0,

(n − 1) !
Z ∞
xn−1 e−λx = (5.5.2)
0 λn

Proof. For all n ≥ 1, λ > 0, a > 0 define u : [0, a] → R and v : [0, a] → R by

u(x) = xn−1 and v (x) = e−λx .

As u, v are continuous functions, clearly In,λ

a given by

Z a
a
In,λ = xn−1 e−λx .
0

is well defined finite positive number. As xα e−βx → 0 as x → ∞ for any α, β > 0 there is
a K > 0 such that
λx
0 ≤ xn−1 e−λx < e− 2 ,

for all K > 0. Therefore b > a > k we have

Z b Z b
λx λb λa
a
| In,λ b
− In,λ |= xn−1 e−λx ≤ e− 2 dx = 2(e− 2 − e− 2 ).
a a

From this it is standard to note that

Z ∞
In,λ := xn−1 e−λx = lim In,λ
a
0 a→∞

Version: – November 19, 2024

 5.5 functions of independent random variables 199

is a well defined finite positive number. Now, as u, v are differentiable we have by the
integration by parts formula
Z a Z a
′
u(x)v (x)dx = u(a)v (a) − u(0)v (0) − u′ (x)v (x)dx.
0 0

Substituting for u, v above we get

a
−λIn,λ = an−1 e−λa − (n − 1)In−1,λ
a
.

Taking limits as a → ∞ we have

λIn,λ = (n − 1)In−1,λ .

Applying the above inductively we have

n−1
(n − i) (n − 1) !
I1,λ .
Y
In,λ = I1,λ =
i=1
λ λn−1

Using the fact that I1 = 1

λ we have the result. ■

Definition 5.5.5. X ∼ Gamma(n, λ): Let λ > 0 and n ∈ N. Then X is said to

be Gamma distributed with parameters n and λ if it has the density

λn
f (x) = xn−1 e−λx , (5.5.3)
(n − 1) !

where x ≥ 0. The parameter n is referred to as the shape parameter and λ as the

rate parameter. By (5.5.2) we know that f given by (5.5.3) is a density function.

We saw in Example 5.5.3 that sum of two exponential distributions resulted in a gamma
distribution. If X ∼ Exponential (λ) then it can also be viewed as a Gamma(1, λ)
distribution. The result in Example 5.5.3 could be rephrased as follows: the sum of two
gamma random variables with shape parameter 1 and rate parameter λ is distributed as a
gamma random variable with shape parameter 2 and rate parameter λ. This holds more
generally as we show in the next example.

Version: – November 19, 2024

200 continuous probabilities and random variables

Gamma(2, 1) Gamma(3, 2) Gamma(4.5, 1.5)

Density f(x) Distribution F(x)

1.0
0.5
0.8
0.4

0.6
0.3

0.2 0.4

0.1 0.2

0.0 0.0

0 2 4 6 0 2 4 6

Figure 5.15: The Gamma density and cumulative distribution functions for various shape and
rate parameters.

Example 5.5.6. (Sum of Gammas) Let n ∈ N, m ∈ N, λ > 0, X and Y be two independent

Gamma(n, λ) and Gamma(m, λ) random variables respectively. Let Z = X + Y . Then
we know that Z has a density given by (5.5.1). Further,

λn n−1 e−λx λm (z

(n−1)! x (m−1)! − x)m−1 e−λ(z−x) if x ≥ 0, z − x ≥ 0
fX (x)fY (z − x) =
 0 otherwise

 e−λz λn+m xn−1 (z − x)m−1 if x ≥ 0, x ≤ z, z ≥ 0
(n−1)!(m−1)!
=
 0 otherwise

For z ≥ 0, we have
Z ∞ Z z
fZ (z ) = fX1 (x)fX2 (z − x)dx = fX1 (x)fX2 (z − x)dx
−∞ 0
e−λz λn+m
Z z
= xn−1 (z − x)m−1 dx
(n − 1) ! (m − 1) ! 0

We now make a change of variable x = zu so that dx = zdu to obtain

z n+m−1 e−λz λn+m

Z 1
fZ (z ) = un−1 (1 − u)m−1 du
(n − 1) ! (m − 1) ! 0

Define R 1 n−1
0 u (1 − u)m−1 du
c(n, m) = .
(n − 1) ! (m − 1) !

Version: – November 19, 2024

 5.5 functions of independent random variables 201

Thus we have the probability density of Z is given by,

if z ≥ 0
(
c(n, m) · λn+m z n+m−1 e−λz
fZ (z ) =
0 otherwise

To evaluate c(n, m) we use the following fact. From Proposition 5.5.1 fZ (·) (given by
(5.5.1)) is a Probability density function. Therefore,
Z ∞
1= fZ (z )dz
−∞
Z ∞
= c(n, m)λn+m z n+m−1 e−λz dz
0
= c(n, m)[(n + m − 1)!],

where in the last line we have used (5.5.2) with n replaced by n + m. So c(n, m) = (n+m−1)! .
1

Hence Z has Gamma (n + m, λ) distribution. From the definition of c(n, m) we also have

(n + m − 1) !
Z 1
un−1 (1 − u)m−1 du = .
0 (n − 1) ! (m − 1) !

The above calculation is easily extended by an induction argument to obtain the fact
that if λ > 0, Xi , 1 ≤ i ≤ n are independent Gamma(ni , λ) distributed random variables
n n
(respectively). Then Z = Xi has Gamma ( ni , λ) distribution.
P P
i=1 i=1
As Exponential (λ) is the same as Gamma(1, λ) random variable, the above implies
that the sum of n independent Exponential (λ) random variables is a Gamma(n, λ) random
variable. ■

It is possible to define the Gamma distribution when the shape parameter is not necessarily
an integer.

Definition 5.5.7. X ∼ Gamma(α, λ): Let λ > 0 and α > 0. Then X is said to
be Gamma distributed with shape parameter α and rate parameter λ if it has the
density
λα α−1 −λx
f (x) = x e , (5.5.4)
Γ (α )
where x ≥ 0 and for α > 0
Z ∞
Γ (α ) = xα−1 e−x dx (5.5.5)
0

One can imitate the calculation done in Example 5.5.6 as well for such a Gamma distribution.

Version: – November 19, 2024

202 continuous probabilities and random variables

The distribution function of a gamma random variable involves an indefinite form of the
integral in (5.5.5). Such integrals are known as incomplete gamma functions, and have no
closed-form solution in terms of simple functions. In R, F (x) for the gamma distribution

λα
Z x
F (x) = P (X ≤ x) = z α−1 e−λz dz , x > 0
0 Γ (α )

can be evaluated numerically with a function call of the form pgamma(x, alpha, lambda).
For example,

pgamma(1, 2, 1)

[1] 0.2642411

pgamma(3, 4.5, 1.5)

[1] 0.5627258

Similarly, the density function f (x) in (5.5.4) involves the normalising constant Γ(α) (also
known as the gamma function) which usually cannot be computed explicitly when α is not
an integer. Using R, one can evaluate f (x) numerically using the dgamma() function as

dgamma(1, 2, 1)

[1] 0.3678794

dgamma(3, 4.5, 1.5)

[1] 0.2769272

5.5.2 Distributions of Quotients of Independent Random Variables

Let X and Y be two independent continous random variables with densities fX and fY . In
this section we shall find out the probability density function of Z = Y .
X
As P (Y = 0) = 0,
Z is well defined random variable.

Version: – November 19, 2024

 5.5 functions of independent random variables 203

Proposition 5.5.8. (Quotient of two independent random variables) Let X and Y be

two independent random variables with marginal densities given by fX : R → R and
fY : R → R. Then Z = X
Y has a probability density function fZ : R → R given by
Z ∞
fZ (z ) = | y | fX (zy )fY (y )dy. (5.5.6)
−∞

Proof- Let us find an expression for the distribution function of Z.

F (z ) = P (Z ≤ z )
X
= P ( ≤ z)
Z ZY
= fX (x)fY (y )dydx
{(x,y ):y̸=0, x
y
≤z}
Z Z Z Z
= fX (x)fY (y )dydx + fX (x)fY (y )dydx
{(x,y ):y<0, x
y
≤z} {(x,y ):y>0, x
y
≤z}
Z Z Z Z
= fX (x)fY (y )dydx + fX (x)fY (y )dydx
{(x,y ):y<0,x≥yz} {(x,y ):y>0,x≤yz}
Z 0 Z ∞ Z ∞ Z yz
= fX (x)fY (y )dxdy + fX (x)fY (y )dxdy
−∞ yz 0 −∞
= I + II

Let us make a u-substituion x = yu in both I and II. For I, y < 0, so we will obtain,
Z 0 Z −∞
I = yfX (yu)fY (y )dudy
−∞ z
Z 0 Z z
= (−y )fX (yu)fY (y )dudy
−∞ −∞
Z z Z 0
= (−y )fX (yu)fY (y )dydu,
−∞ −∞

where in the last line we have changed the order of integration1 . For II, y > 0 so we will
obtain (similarly as in I),
Z ∞Z z
II = yfX (yu)fY (y )dudy
0 −∞
Z z Z ∞
= yfX (yu)fY (y )dydu,
−∞ 0

1
The change of order of integration is justifiable under certain hypothesis for the integrand. We shall assume
these are satisfied, as it is not possible to state and verify them within the scope of this book

Version: – November 19, 2024

204 continuous probabilities and random variables

Therefore

F (z ) = I + II
Z z Z 0 Z z Z ∞
= (−y )fX (yu)fY (y )dydu + yfX (yu)fY (y )dydu
−∞ −∞ −∞ 0
Z z Z ∞
= | y | fX (yu)fY (y )dydu
−∞ −∞

As fX (·) and fY (·) are densities, it can be shown that the integrand is a piecewise continuous
function. Hence the F is of the form (5.2.4) and Theorem 5.2.5 implies that the probability
density function of Z is given by (5.5.6). ■

Using the above method for finding the distribution of quotient of two random variables,
we shall present three examples that will lead us to standard continuous distributions
that are useful in applications. We begin with an example that constructs the Cauchy
distribution.

Example 5.5.9. Let X and Y be two independent Normal random variables with mean 0
and variance σ 2 ̸= 0. Let Z = Y .
X
We know that the probability density function of Z is
given by (5.5.6). Further, for any y, z ∈ R

1 1 1 1 + z2
! !
z2 y2 y2
fX (zy )fY (y ) = √ e− 2σ2 √ e− 2σ2 = exp − y 2
2πσ 2πσ 2πσ 2 2σ 2

Fix z ∈ R.

1 1 + z2
Z ∞ ! !
fZ (z ) = |y| exp − y 2
dy
−∞ 2πσ 2 2σ 2
1 1 + z2 1 + z2
"Z ! ! Z ∞ ! ! #
0
= | y | exp − y 2
dy + | y | exp − y 2
dy
2πσ 2 −∞ 2σ 2 0 2σ 2
1 1 + z2 1 + z2
"Z ! ! Z ∞ ! ! #
0
= (−y ) exp − y 2 dy + y exp − y 2 dy
2πσ 2 −∞ 2σ 2 0 2σ 2

It is easy to see that two integrals are the same (perform a substitution of u = −y in the
first integral). So the above is

1 1 + z2
Z ∞ ! !
= y exp − y 2
dy.
πσ 2 0 2σ 2

Version: – November 19, 2024

 5.5 functions of independent random variables 205

 
1+z 2 1+z 2
Now perform a substitution 2σ 2
y 2 = t, so σ2
ydy = dt.

1 σ2 ∞
Z
fZ (z ) = exp(−t)dt.
πσ 2 1 + z 2 0
1 1
= (−e−t |∞
0 ) = .
π (1 + z 2 ) π (1 + z 2 )

Therefore Z has the Cauchy distribution, which we first saw in the context of Example
5.3.4. ■
The next example considers the ratio of two gamma random variables. This motivates a
standard distribution called the F -distribution, which we will encounter in Chapter 8.

Example 5.5.10. Let m ∈ N, n ∈ N, λ > 0, X and Y be two independent Gamma

(m, λ) and Gamma (n, λ) random variables respectively. Let Z = Y .
X
We know that the
probability density function of Z is given by (5.5.6). Further,

λm m−1 e−λ(zy ) λn y n−1 e−λy

(m−1)! (zy ) (n−1)! if y ≥ 0, zy ≥ 0
fX (zy )fY (y ) =
 0 otherwise

λn + m

(n−1)!(m−1)! y
n+m−2 z m−1 e−λ(1+z )y if y ≥ 0, z ≥ 0
=
 0 otherwise

Fix z > 0,

λn+m
Z ∞
fZ (z ) = y y n+m−2 z m−1 e−λ(1+z )y dy
0 (n − 1) ! (m − 1) !
z m−1 λn+m
Z ∞
= y n+m−1 e−λ(1+z )y dy
(n − 1) ! (m − 1) ! 0

Now perform a substition (1 + z )y = t, so (1 + z )dy = dt and the above is

z m−1 λm+n
Z ∞
= tm+n−1 e−λt dt
(1 + z )m+n (m − 1 ) ! (n − 1 ) ! 0

Using (5.5.2) we have that


(m+n−1)! m−1 (1 + z )−(m+n)

(m−1)!(n−1)! z if z ≥ 0
fZ (z ) = (5.5.7)
 0 otherwise

■
Our next example is a construction of the Beta-distribution.

Version: – November 19, 2024

206 continuous probabilities and random variables

Example 5.5.11. Let m ∈ N, n ∈ N, λ > 0. Let X and Y be two independent Gamma

(m, λ) and Gamma (n, λ) random variables respectively. Let Z = X
X +Y .

Let W = X.
Y
Note that Z = 1+W .
1
In Example 5.5.10 we found the probability density
function of W . We shall use this to find the distribution funciton of Z. As P (W ≥ 0) = 1,

0 if z < 0
(
P (Z ≤ z ) =
1 if z > 1.

For 0 < z < 1,

1 1−z
P (Z ≤ z ) = P ( ≤ z ) = P (W ≥ )
1+W z
1−z
= 1 − P (W ≤ )
z

Using (5.5.7) we obtain that the above is

1−z
(m + n − 1)! m−1
Z
z
= 1− u (1 + u)−(m+n) du
0 (m − 1) ! (n − 1) !
1−z
(m + n − 1) !
Z
z
= 1− um−1 (1 + u)−(m+n) du
(m − 1) ! (n − 1) ! 0

For 0 < z < 1, by the fundamental theorem of calculus, differentiating in z

1 (m + n − 1) ! 1 − z m−1 1−z
   −(m+n)
fZ (z ) = · 1+
z 2 (m − 1) ! (n − 1) ! z z
(m + n − 1)! n−1
= z (1 − z )m−1
(m − 1) ! (n − 1) !

Z is said to have the Beta(m, n) distribution. ■

We define the distribution in general next.

Definition 5.5.12. X ∼ Beta(α, β): Let α > 0 and β > 0. Then X is said to be
Beta distributed with parameters α and β if it has the density

 Γ(α+β ) xα−1 (1 − x)β−1


Γ (α) Γ (β )
0<x<1
f (x) = (5.5.8)
0 otherwise.

Version: – November 19, 2024

 5.5 functions of independent random variables 207

Beta(0.5, 0.5) Beta(1, 1) Beta(2, 2) Beta(3, 6)

Density f(x) Distribution F(x)

2.5 1.0

2.0 0.8

1.5 0.6

1.0 0.4

0.5 0.2

0.0 0.0

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Figure 5.16: The Beta density and cumulative distribution functions for selected shape param-
eters.

The distribution function of a beta random variable is given by an indefinite integral

which in general has no closed-form solution in terms of simple functions. In R, F (x) for
the beta distribution

Γ(α + β ) α−1
Z x
F (x) = P (X ≤ x) = u (1 − u)β−1 du , 0 < x < 1
0 Γ (α ) Γ (β )

can be evaluated numerically with a function call of the form pbeta(x, alpha, beta).
For example,

pbeta(0.5, 0.5, 0.5)

[1] 0.5

pbeta(0.5, 3, 6)

[1] 0.8554688

pbeta(0.2, 6, 1)

[1] 6.4e-05

Version: – November 19, 2024

208 continuous probabilities and random variables

pbeta(0.2, 1, 6)

[1] 0.737856

In the special case where either α or β equals 1, the distribution function of X can
be computed explicitly. Another special case is the standard arcsine law we previously
encountered in Exercise 5.2.7 in terms of its explicit distribution function; it is easy to
see that this is the same as the Beta( 21 , 12 ) distribution. The semicircular distribution
encountered in Exercise 5.2.6 is also related, in the sense that it can be viewed as a location
and scale transformed beta random variable.

exercises

Ex. 5.5.1. Suppose that X and Y are random variables with joint probability density

 4 (xy + 1) if 0 ≤ x ≤ 1 and 0 ≤ y ≤ 1
5
f (x, y ) =
 0 otherwise

(a) Compute the marginal densities of X and Y ?

(b) Compute the conditional density (X|Y = y ) (for appropriate y).

(c) Are X and Y independent?

Ex. 5.5.2. Let X and Y be two random variables with the joint p.d.f given by

2 −λy
 λ e

 0≤x≤y
f (x, y ) =

 0

otherwise

(a) Find the marginal distribution of X and Y .

(b) Find the conditional distribution of (Y | X = x) for some x > 0

Ex. 5.5.3. Let a, b > 0. Let X ∼ Gamma (a, b) and Y ∼ Exponential (X ).

(a) Find the joint density of X and Y .

(b) Find the marginal density of Y .

(c) Find the conditional density of (X | Y = y ).

Version: – November 19, 2024

 5.5 functions of independent random variables 209

Ex. 5.5.4. Let X1 , X2 , X3 be independent and identically distributed Uniform (0, 1) random
variables. Let A = X1 X3 and B = X22 . Find the P (A < B ).
Ex. 5.5.5. Let X and Y be two independent exponential random variables each with mean
1.
1
(a) Find the density of U1 = X 2 .

(b) Find the density of U2 = X + Y + 1.

(c) Find P (max{X, Y } > 1).

Ex. 5.5.6. Suppose X is a uniform random variable in the interval (0, 1) and Y is an
independent exponential(2) random variable. Find the distribution of Z = X + Y .
Ex. 5.5.7. Let α > 0, β > 0, λ > 0, X and Y be two independent Gamma(α, λ) and
Gamma(β, λ) random variables respectively. Then Z = X + Y is distributed as a Gamma
(α + β, λ).
Ex. 5.5.8. Let X and Y be two independent random variables with probability density
function fX (·) and fY (·). Show that X + Y and Y + X have the same distribution by
showing that the integral expression defining fX ⋆ fY (·) is equal to the integral expression
defining fY ⋆ fX (·)).
Ex. 5.5.9. Let α > 0 and Γ(α) as in (5.5.5).

(a) Using the same technique as in Lemma 5.5.4, show that 0 < Γ(α) < ∞.
R ∞ −0.5 −x √
(b) Show that Γ( 12 ) = 0 x e dx = π.

Ex. 5.5.10. Let α > 0, δ > 0, λ > 0. Let X and Y be two independent Gamma (α, λ) and
Gamma (δ, λ) random variables respectively.

(a) Let W = X.
Y
Find the probability density function of W .

(b) Let Z = X
X +Y . Find the probability density function of Z.

(c) Are X and Z independent ?

Hint: Compute the joint density and see if it is a product of the marginals.

Ex. 5.5.11. Suppose X, Y are independent random variables each normally distributed
with mean 0 and variance 1.
√
(a) Find the probability density function of R = X2 + Y 2

(b) Find the probability density function of Z = X

Y

Version: – November 19, 2024

210 continuous probabilities and random variables

 
(c) Find the probability density function of θ = arctan X
Y

(d) Are R, θ independent random variables ?

Hint: Compute the joint density using the change of variable indicated in Exercise
5.1.10. Decide if it is a product of the marginals

Version: – November 19, 2024

S U M M A R I S I N G C O N T I N U O U S R A N D O M VA R I A B L E S
6
In this chapter we shall revisit concepts that have been discussed for discrete random
variables and see their analogues in the continuous setting. We then introduce generating
functions and conclude this chapter with a discussion on bivariate normal random variables.

6.1 expectation and variance

The notion of expected value carries over from discrete to continuous random variables,
but instead of being described in terms of sums, it is defined in terms of integrals.

Definition 6.1.1. Let X be a continuous random variable with piecewise continuous

density f (x). Then the expected value of X is given by

Z∞
E [X ] = xf (x) dx.
−∞

provided that the integral converges absolutely.a In this case we say that X has
“finite expectation”. If the integral diverges to ±∞ we say the random variable has
infinite expectation. If the integral diverges, but not to ±∞ we say the expected value
is undefined.
ZN
a
That is, lim |x| f (x) dx < ∞.
M →−∞
N →∞
M

The next three examples illustrate the three posibilities: the first is an example where
expectation exists as a real number; the next is an example of an infinite expected value;
and the final example shows that the expected value may not be defined at all.

Example 6.1.2. Let X ∼ Uniform(a, b). Then the expected value of X is given by

1 1
Z ∞ Z b
b+a
E [X ] = x · f (x) dx = x· dx = (b2 − a2 ) = .
−∞ a b−a 2(b − a) 2

This result is intuitive since it says that the average value of a Uniform(a, b) random
variable is the midpoint of its interval. ■

211

Version: – November 19, 2024

212 summarising continuous random variables

Example 6.1.3. Let 0 < α < 1 and X ∼ Pareto(α) which is defined to have the probability
density function 
α
 xα + 1

 1≤x<∞
f (x) =

 0

otherwise
Z ∞ Z M
α α
E [X ] = x· dx = α lim x−α dx = (−1 + lim M −α+1 ) = ∞
1 xα + 1 M →∞ 1 −α + 1 M →∞

as 0 < α < 1.
Thus this Pareto random variable has an infinite expected value. ■

Example 6.1.4. Let X ∼ Cauchy(0, 1). Then the probability density function of X is
given by
1 1
f (x) = for all x ∈ R.
π 1 + x2
Now,
1
Z ∞
E [X ] = x· dx
−∞ π ( 1 + x2 )
RN
Now by Exercise 6.1.10, we know that as M → −∞, N → ∞ the x
M 1+x2 dx does not
converge or diverge to ±∞. So E [X ] is not defined for this Cauchy random variable. ■

Expected values of functions of continuous random variables may be computed using

their respective probability density function by the following theorem.

Theorem 6.1.5. Let X be continuous random variables with probability density

function fX : R → R.

(a) Let g : R → R be piecewise continuous and Z = g (X ) Then the expected value

of Z given by
Z ∞
E [g (X )] = g (x)fX (x) dx
−∞

(b) Let Y be a continuous random variable such that (X, Y ) have a joint probability
density function f : R2 → R. Suppose h : R2 → R be piecewise continuous.
Then, Z ∞ Z ∞
E [h(X, Y )] = h(x, y )f (x, y ) dx dy.
−∞ −∞

Proof- The proof is beyond the scope of this book. For (a) when g is as in Exercise
5.3.10 then one can provide the proof using only the tools of basic calculus (we will leave
this case as an exercise to the reader) ■

Version: – November 19, 2024

 6.1 expectation and variance 213

We illustrate the use of the above theorem with a couple of examples.

Example 6.1.6. A piece of equipment breaks down after a functional lifetime that is a
random variable T ∼ Exp( 51 ). An insurance policy purchased on the equipment pays a
dollar amount equal to 1000 − 200t if the equipment breaks down at a time 0 ≤ t ≤ 5
and pays nothing if the equipment breaks down after time t = 5. What is the expected
payment of the insurance policy?

For t ≥ 0 the policy pays g (t) = max{1000 − 200t, 0} so,

1 (1/5)t
Z ∞
E [g (T )] = e max{1000 − 200t, 0} dt
0 5
1 (1/5)t
Z 5
= e (1000 − 200t) dt
0 5
= 1000e−1 ≈ $367.88

Example 6.1.7. Let X, Y ∼ Uniform(0, 1). What is the expected value of the larger of
the two variables?

We offer two methods of solving this problem. The first is to define Z = max{X, Y } and
then determine the density of Z. To do so, we first find its distribution. FZ (z ) = P (Z ≤ z ),
but max{X, Y } is less than or equal to z exactly when both X and Y are less than or
equal to z. So for 0 ≤ z ≤ 1,

FZ (z ) = P ((X ≤ z ) ∩ (Y ≤ z ))
= P (X ≤ z ) · P (Y ≤ z )
= z2

Therefore fZ (z ) = FZ′ (z ) = 2z after which the expected value can be obtained through
integration
2 3 1 2
Z 1
E [Z ] = z · 2z dz = z |0 = .
0 3 3

An alternative method is to use Theorem 6.1.5 (b) to calculate the expectation directly
without finding a new density. Since X and Y are independent, their joint distribution is
the product of their marginal distributions. That is,

1 if 0 ≤ x ≤ 1 and 0 ≤ y ≤ 1
(
f (x, y ) = fX (x)fY (y ) =
0 otherwise

Version: – November 19, 2024

214 summarising continuous random variables

Therefore,
Z ∞ Z ∞
E [max{X, Y }] = max{x, y} · f (x, y ) dx dy
−∞ −∞
Z 1Z 1
= max{x, y} · 1 dx dy
0 0

The value of max{x, y} is x if 0 < y ≤ x < 1 and it is y if 0 < x ≤ y < 1. So,

Z 1Z y Z 1Z 1
E [max{X, Y }] = y dx dy + x dx dy
0 0 0 y
1 2 x=1
Z 1 Z 1
= xy |xx= y
dy + x |x=y dy
0 2
=0
0
1 1 2
Z 1 Z 1
= y 2 dy + − y dy
0 0 2 2
1 1 2
= + = .
3 3 3

Results from calculus may be used to show that the linearity properties from Theorem
4.1.7 such as apply to continuous random variables as well as to discrete ones. We restate
it here for completeness.

Theorem 6.1.8. Suppose that X and Y are continuous random variables with
piecewise continuous joint density function function f : R2 → R. Assume that both
have finite expected value. If a and b are real numbers then

(a) E [aX ] = aE [X ];

(b) E [aX + b] = aE [X ] + b

(c) E [X + Y ] = E [X ] + E [Y ]; and

(d) E [aX + bY ] = aE [X ] + bE [Y ].

(e) If X ≥ 0 then E [X ] ≥ 0.

Proof- See Exercise 6.1.11. ■

We will use these now-familiar properties in the continuous setting. As in the discrete
setting we can define the variance and standard deviation of a continuous random variable.

Version: – November 19, 2024

 6.1 expectation and variance 215

Definition 6.1.9. Let X be a random variable with probability density function

f : R → R. Suppose X has finite expectation. Then

(a) the variance of the random variable is written as V ar [X ] and is defined as

Z ∞
V ar [X ] = E [(X − E [X ])2 ] = (x − E [X ])2 fX (x)dx,
−∞

(b) the standard deviation of X is written as SD [X ] and is defined as

q
SD [X ] = V ar [X ]

Since the above terms are expected values, there is the possibility that they may be
infinite because the integral describing the expectation diverges to infinity. As the
integrand is strictly positive, it isn’t possible for the integral to diverge unless it
diverges to infinity.

The properties of variance and standard deviation of continuous random variables

match those of their discrete counterparts. A list of these properties follows below.

Theorem 6.1.10. Let a ∈ R and let X be a continuous random variable with finite
variance (and thus, with finite expected value as well). Then,

(a) V ar [X ] = E [X 2 ] − (E [X ])2 .

(b) V ar [aX ] = a2 · V ar [X ];

(c) SD [aX ] = |a| · SD [X ];

(d) V ar [X + a] = V ar [X ]; and

(e) SD [X + a] = SD [X ].

If Y is another independent continuous random variable with finite variance (and

thus, with finite expected value as well) then

(f) E [XY ] = E [X ]E [Y ];

(g) V ar [X + Y ] = V ar [X ] + V ar [Y ]; and
q
(h) SD [X + Y ] = (SD [X ])2 + (SD [Y ])2 .

Version: – November 19, 2024

216 summarising continuous random variables

Proof- The proof is essentially an imitation of the proofs presented in Theorem 4.1.10,
Theorem 4.2.5, Theorem 4.2.4, and Theorem 4.2.6. One needs to use the respective
densities, integrals in lieu of sums, and use Theorem 6.1.11 and Theorem 6.1.5 when needed.
We will leave this as an exercise to the reader. ■

Example 6.1.11. Let X ∼ Normal (0, 1). In this example we shall show that E [X ] = 0
and V ar [X ] = 1. Before that we collect some facts about the probability density function
of X, given by (5.2.7). Using (5.2.9) with z = 0, we can conclude that

1 1
Z ∞
x2
√ e− 2 dx = (6.1.1)
0 2π 2

Observe that there exists c1 > 0 such that

x2
max{| x |, x2 }e− 2 ≤ c1 e−c1 |x|

for all x ∈ R. Hence

1
Z ∞ ∞
x2
Z
| x | √ e− 2 dx ≤ c1 e−c1 |x| < ∞
−∞ 2π −∞
2 1
Z ∞ 2
Z ∞
− x2
x √ e dx ≤ c1 e−c1 |x| < ∞ (6.1.2)
−∞ 2π −∞

Using the above we see that

1
Z ∞
x2
E [X ] = x √ e− 2 dx < ∞
−∞ 2π

So we can split integral expression in definition of E [X ] as

1 1
Z 0 Z ∞
x2 x2
E [X ] = x √ e− 2 dx + x √ e− 2 dx.
−∞ 2π 0 2π

Further the change of variable y = −x will imply that

1 1
Z 0 Z ∞
x2 y2
x √ e− 2 dx = − y √ e− 2 dy.
−∞ 2π 0 2π

So E [X ] = 0. Again by (6.1.2),

1 1
Z ∞ Z ∞
x2 x2
V ar [X ] = (x − E [X ]) √ e− 2 dx =
2
x2 √ e− 2 dx < ∞
−∞ 2π −∞ 2π

Version: – November 19, 2024

 6.1 expectation and variance 217

To evaluate the integral we make a change of variable to obtain

1 1 1 1
Z ∞ Z 0 Z ∞ Z ∞
x2 x2 x2 x2
x2 √ e− 2 = x2 √ e− 2 dx + x2 √ e− 2 dx = 2 x2 √ e− 2 dx.
−∞ 2π −∞ 2π 0 2π 0 2π

x2
Then we use integration by parts like Lemma 5.5.4. Set u(x) = x and v (x) = e− 2 , which
2
− x2
imply u′ (x) = 1 and v ′ (x) = −xe . Therefore for a > 0,
Z a Z a Z a
x2
x2 e − 2 dx = u(x)(−v ′ (x))dx = u(x)(−v (x)) |a0 − u′ (x)(−v (x))dx
0 0 0
2
Z a
x2
2 − a2
= a e + e− 2
0

a2
Using the fact that lima→∞ a2 e− 2 = 0 and (6.1.1) we have

1 ∞ 1 a 1
 Z a 
x2 x2 a2 x2
Z Z
V ar [X ] = 2 √ x2 e− 2 dx = √ lim x2 e− 2 dx = √ lim a2 e− 2 + e− 2 dx
2π 0 π a→∞ 0 π a→∞ 0
1 1
Z ∞
√
 
x2
= √ 0+ e− 2 dx = √ [0 + π ] = 1
π 0 π

Y −µ
Suppose Y ∼ Normal (µ, σ 2 ) then we know by Corollary 5.3.3 that W = σ ∼ Normal
(0, 1). By Example 6.1.11, E [W ] = 0 and V ar [W ] = 1. Also Y = σW + µ, so by
Theorem 6.1.8(b) E [Y ] = σE [W ] + µ = µ and by Theorem 6.1.10 (d) and (b) V ar [Y ] =
σ 2 V ar [W ] = σ 2 . ■

Example 6.1.12. Let X ∼ Uniform(a, b). To calculate the variance of X first note that
Theorem 6.1.5(a) gives

1 1 b2 + ab + a2
Z ∞ Z b
E [X ] = 2 2
x · f (x) dx = x2 · dx = (b3 − a3 ) = .
−∞ a b−a 3(b − a) 3

Now, since E [X ] = b+a

2 (see Example 6.1.2), the variance may be found as

b2 + ab + a2 b+a 2 (b − a)2
V ar [X ] = E [X 2 ] − (E [X ])2 = −( ) = .
3 2 12

Taking square roots, we obtain SD [X ] = b−a

√
12
. So the standard deviation of a continuous,
uniform random variable is √1
12
times of the length of its interval. ■

The Markov and Chebychev inequalities also apply to continuous random variables. As
with discrete variables, these help to estimate the probabilities that a random variable will
fall within a certain number of standard deviations from its expected value.

Version: – November 19, 2024

218 summarising continuous random variables

Theorem 6.1.13. Let X be a continuous random variable with probability density

function f and finite non-zero variance.

(a) (Markov’s Inequality) Suppose X is supported on non-negative values, i.e.

f (x) = 0 for all x < 0. Then for any c > 0,

µ
P (X ≥ c) ≤ .
c

(b) (Chebychev’s Inequality) For any k > 0,

1
P (|X − µ| ≥ kσ ) ≤ .
k2

Proof - (a) By definition of µ and assumptions on f , we have

Z ∞ Z ∞
µ= xf (x)dx = xf (x)dx.
−∞ 0

Using an elementary fact from integrals we know that

Z ∞ Z c Z ∞
xf (x)dx = xf (x)dx + xf (x)dx
0 0 c

We note that the first integral is non-negative so we have

Z ∞
µ ≥ xf (x)dx.
c

As f (·) ≥ 0, we have xf (x) ≥ cf (x) whenever x > c. So again using facts about integrals
Z ∞ Z ∞
µ ≥ cf (x)dx = c f (x)dx = cP (X > c).
c c

The last equality follows from definition. Hence we have the result.
(b) The event (|X − µ| ≥ kσ ) is the same as the event ((X − µ)2 ≥ k 2 σ 2 ). The random
variable (X − µ)2 is certainly non-negative, is continuous by Exercise 5.3.9, and its expected
value is the variance of X which we have assumed to be finite. Therefore we may apply
Markov’s inequality to (X − µ)2 to get

E [(X − µ)2 ] V ar [X ] σ2 1
P (|X − µ| ≥ kσ ) = P ((X − µ)2 ≥ k 2 σ 2 ) ≤ 2 2
= 2 2
= 2 2
= 2.
k σ k σ k σ k

Though the theorem is true for all k > 0, it doesn’t give any useful information unless
k > 1.

Version: – November 19, 2024

 6.1 expectation and variance 219

exercises

Ex. 6.1.1. Suppose X has probability density function given by


1− | x | −1 ≤ x ≤ 1
fX (x) =
0 otherwise

(a) Compute the distribution function of X.

(b) Compute E [X ] and V ar [X ].

Ex. 6.1.2. Suppose X has probability density function given by


 cos(x) − π2 ≤ x ≤ π
2 2
fX (x) =
0 otherwise

(a) Compute the distribution function of X.

(b) Compute E [X ] and V ar [X ].

Ex. 6.1.3. Find E [X ] and V ar [X ] in the following situations:

(a) X ∼ Normal(µ, σ 2 ), with µ ∈ R and σ > 0.

(b) X has probability density function given by





 x 0≤x≤1

fX (x) =

2−x 1≤x≤2

0 otherwise



Ex. 6.1.4. Let 1 < α and X ∼ Pareto(α). Calculate E [X ] to show that it is finite.
Ex. 6.1.5. Let X be a random variable with density f (x) = 2x for 0 < x < 1 (and f (x) = 0
otherwise).

(a) Calculate E [X ]. You should get a result larger than 12 . Explain why this should be
expected even without computations.

(b) Calculate SD [X ].

Version: – November 19, 2024

220 summarising continuous random variables

Ex. 6.1.6. Let X ∼ Uniform(a, b) and let k > 0. Let µ and σ be the expected value and
standard deviation calculated in Example 6.1.12.

(a) Calculate P (|X − µ| ≤ kσ ). Your final answer should depend on k, but not on the
values of a or b.

(b) What is the value of k such that results of more than k standard deviations from
expected value are unachievable for X?

Ex. 6.1.7. Let X ∼ Exponential(λ).

(a) Prove that E [X ] = 1

λ and SD [X ] = λ1 .

(b) Let µ and σ denote the mean and standard deviation of X respectively. Use your
computations from (a) to calculate P (|X − µ| ≤ kσ ). Your final answer should
depend on k, but not on the value of λ.

(c) Is there a value of k such that results of more than k standard deviations from
expected value are unachievable for X?

Ex. 6.1.8. Let X ∼ Gamma(n, λ) with n ∈ N and λ > 0. Using Example 5.5.3, Exercise
6.1.7(a) and Theorem 6.1.8(c) calculate E [X ]. Using Theorem 6.1.10 calculate V ar [X ].
Ex. 6.1.9. Let X ∼ Uniform(0, 10) and let g (x) = max{x, 4}. Calculate E [g (X )].
RN
Ex. 6.1.10. Show that as M → −∞, N → ∞ x
M 1+x2 dx does not have a limit.
Ex. 6.1.11. Using the hints provided below prove the respective parts of Theorem 6.1.8.

(a) For a = 0 the result is clear. Let a ̸= 0 and fX : R → R be the probability density
function of X. Use Lemma 5.3.2 to find the probability density function of aX.
Compute the expectation of aX to obtain the result. Alternatively use Theorem
6.1.5(a).

(b) Use Theorem 6.1.5(b).

(c) Use the joint density of (X, Y ) to write E [X + Y ]. Then use (5.4.2) an (5.4.3) to
prove the result.

(d) Use the same technique as in (b).

(e) If X ≥ 0 then its marginal density fX : R → R is positive only when the x ≥ 0. The
result immediately follows from definition of expectation.

Ex. 6.1.12. Prove Theorem 6.1.10.

Version: – November 19, 2024

6.2 covariance, correlation, conditional expectation and conditional variance 221

6.2 covariance, correlation, conditional expectation and condi-

tional variance

Covariance of continuous random variables (X, Y ) is used to describe how the two random
variables relate to each other. The properties proved about covariances for discrete random
variables in Section 4.5 apply to continuous random variables as well via essentially the
same arguments. We define covariance and state the properties next.

Definition 6.2.1. Let X and Y be random variables with joint probability density
function f : R2 → R. Suppose X and Y have finite expectation. Then the covariance
of X and Y is defined as
Z ∞ Z ∞
Cov [X, Y ] = E [(X − E [X ])(Y − E [Y ])] = (x − E [X ])(y − E [Y ])f (x, y )dxdy,
−∞ −∞
(6.2.1)

Since it is defined in terms of an expected value, there is the possibility that the covariance
may be infinite or not defined at all. We now state the properties of Covariance.

Theorem 6.2.2. Let X, Y be continuous random variables such that they have
joint probability density function. Assume that 0 ̸= σx2 = Var(X ) < ∞, 0 ̸= σy2 =
Var(Y ) < ∞. Then

(a) Cov [X, Y ] = E [XY ] − E [X ]E [Y ].

(b) Cov [X, Y ] = Cov [Y , X ];

(c) Cov [X, X ] = V ar [X ].

(d) −σX σY ≤ Cov [X, Y ] ≤ σX σY

(e) If X and Y are independent then Cov [X, Y ] = 0.

Let a, b be real numbers. Suppose Z is another continuous random variable, and

σz = Var(Z ) < ∞. Further (X, Z ), (Y , Z ), (X, aY + bZ ), and (aX + bY , Z ) all
have (their respective) joint probability functions. Then

(f) Cov [X, aY + bZ ] = a · Cov [X, Y ] + b · Cov [X, Z ];

(g) Cov [aX + bY , Z ] = a · Cov [X, Z ] + b · Cov [Y , Z ];

Proof. See Exercise 6.2.13. ■

Version: – November 19, 2024

222 summarising continuous random variables

Definition 6.2.3. Let (X, Y ) be continuous random variables both with finite
Cov [X,Y ]
variance and covariance. From Theorem 6.2.2(d) the quantity ρ[X, Y ] = σX σY is
in the interval [−1, 1]. It is known as the “correlation” of X and Y . As discussed
earlier, both the numerator and denominator include the units of X and the units of Y .
The correlation, therefore, has no units associated with it. It is thus a dimensionless
rescaling of the covariance and is frequently used as an absolute measure of trends
between the two continuous random variables as well.

Example 6.2.4. Let X ∼ Uniform (0, 1) and be independent of Y ∼ Uniform (0, 1). Let
U = min(X, Y ) and V = max(X, Y ). We wish to find ρ[U , V ]. First, 0 < u < 1

P (U ≤ u) = 1 − P (U > u) = 1 − P (X > u, Y > u) = 1 − P (X > u)P (Y > u) = 1 − (1 − u)2 ,

as X, Y are independent uniform random variables. Second, for 0 < v < 1,

P (V ≤ v ) = P (X ≤ v, Y ≤ v ) = P (X ≤ v )P (Y ≤ v ) = v 2 ,

as X, Y are independent uniform random variables. Therefore the distribution function of

U and V are given by
 
0 if u < 0 0 if v < 0

 


 

FU (u) =

1 − (1 − u)2 if 0 < u < 1 and FV (v ) =  v2 if 0 < v < 1
 
1 if u ≥ 1. 1 if v ≥ 1.

 


As FU , FV are piecewise differentiable, the probability density function of U and V are

obtained by differentiating FU and FV respectively.
 
2 ( 1 − u ) if 0 < u < 1 2v if 0 < v < 1
fU (u) = and fV (v ) =
0 otherwise 0 otherwise.

Thirdly, 0 < u < v < 1

P (U ≤ u, V ≤ v ) = P (V ≤ v ) − P (U > u, V ≤ v )
= v 2 − P (u < X ≤ v, u < Y ≤ v )
= v 2 − P (u < X ≤ v )P (u < Y ≤ v )
= v 2 − (v − u)2 ,

Version: – November 19, 2024

6.2 covariance, correlation, conditional expectation and conditional variance 223

where we have used the formula for distribution function of V and the fact that X, Y are
independent uniform random variables. It is easily seen that P (U ≤ u, V ≤ v ) = 0 for all
other possibilities of (u, v ). As the joint distribution function is piecewise differentiable in
each variable, the joint probability density function of U and V , f : R2 → R, exists and is
obtained by differentiating it partially in u and v.

2 if 0 < u < v < 1
f (u, v ) =
0 otherwise

Now,

u3 1 1
Z 1
E [U ] = u2(1 − u)du = u2 − 2 | =
0 3 0 3
v3 1 2
Z 1
E [V ] = v2vdv = 2 | =
0 3 0 3
u3 u4 1 1
Z 1
E [U 2 ] = u2 2(1 − u)du = 2 − 2 | =
0 3 4 0 6
v4 1
Z 1
E [V 2 ] = v 2 2vdv = 2 |10 =
0 4 2
v4 1 1
" #
u2 1 v2
Z 1 Z v  Z 1 Z 1
E [U V ] = uv2du dv = 2v |0 dv = 2v dv = | =
0 0 0 2 0 2 4 0 4

Therefore

2 1 5
V ar [U ] = E [U 2 ] − (E [U ])2 = − =
3 9 9
2 2 1 4 1
V ar [V ] = E [V ] − (E [V ]) = − =
2 9 18
1 12 5
Cov [U , V ] = E [U V ] − E [U ]E [V ] = − =
4 33 36
Cov [U , V ] 5
1
ρ[U , V ] = q = q 36 = √
2 2
q q
5 1
V ar [V ] V ar [U ] 9 18

As seen in Theorem 6.2.2 (e), independence of X and Y guarantees that they are uncorre-
lated (i.e ρ[X, Y ] = 0). The converse is not true (See Example 4.5.6 for discrete case). It
is possible that Cov [X, Y ] = 0 and yet that X and Y are dependent, as the next example
shows.

Version: – November 19, 2024

224 summarising continuous random variables

Example 6.2.5. Let X ∼ Uniform (−1, 1). Let Y = X 2 . Note from Example 6.1.2 and
Example 6.1.12 we have E [X ] = 0, E [Y ] = E [X 2 ] = 13 . Further using the probability
density function of X,

1 x4 1
Z 1
E [XY ] = E [X ] =3
x3 = | = 0.
−1 2 8 −1

So ρ[X, Y ] = 0. Clearly X and Y are not independent. We verify this precisely as well.
Consider the

1 1 1 1 1 1 1
P (X ≤ − , Y ≤ ) = P (X ≤ − , X 2 ≤ ) = P (− ≤ X ≤ − ) = ,
4 4 4 4 2 4 8

as X ∼ Uniform (−1, 1). Whereas,

1 1 1 1 1 1 1 31 3
P (X ≤ − )P (Y ≤ ) = P (X ≤ − )P (X 2 ≤ ) = P (X ≤ − )P (− ≤ X ≤ ) = = .
4 4 4 4 4 2 2 82 16

Clearly
1 1 1 1
P (X ≤ − , Y ≤ ) ̸= P (X ≤ − )P (Y ≤ )
4 4 4 4
implying they are not independent. ■

We are now ready to define conditional expectation and variance.

Definition 6.2.6. Let (X, Y ) be continuous random variables with a piecewise

continuous joint probability density function f . Let fX be the marginal density of X.
Assume x is a real number for which fx (x) ̸= 0. The conditional expectation of Y
given X = x is defined by
Z ∞ Z ∞
f (x, y )
E [Y | X = x] = yfY |X =x (y )dy = y dy
−∞ −∞ fX (x)

whenever it exists. The conditional variance of Y given X = x is defined by

V ar [Y |X = x] = E [(Y − E [Y |X = x])2 |X = x]
Z ∞  Z ∞ 2
f (x, y ) f (x, y )
= y− y dy dy.
−∞ −∞ fX (x) fX (x)

The results proved in Theorem 4.4.4, Theorem 4.4.6, Theorem 4.4.8, and Theorem 4.4.9
are all applicable when X and Y are continuous random variables having joint probability
density function f . The proofs of these results in the continuous setting follow very similarly
(though using facts about integrals from analysis).

Version: – November 19, 2024

6.2 covariance, correlation, conditional expectation and conditional variance 225

Theorem 6.2.7. Let (X, Y ) be continuous random variables with joint probability
density function f : R → R. Assume that h, g : R → R be defined as
 
E [X|Y = y ] if fY (y ) > 0 V ar [X|Y = y ] if fY (y ) > 0
g (y ) = and h(y ) =
0 otherwise 0 otherwise

are well-defined piecewise continuous functions. Let k : R → R be a piecewise

continous function. Then
Z ∞
E [k (X ) | Y = y ] = k (x)fX|Y =y (x)dx, (6.2.2)
−∞

E [g (Y )] = E [X ], (6.2.3)

and
V ar [X ] = E [h(Y )] + V ar [g (Y )]. (6.2.4)

Proof- The proof of (6.2.2) is beyond the scope of this book. We shall omit it. To prove
(6.2.3) we use the definition of g and Theorem 6.1.8 (a) to write
Z ∞ Z ∞ Z ∞ 
E [g (Y )] = g (y )fY (y )dy = xfX|Y =y (x)dx fY (y )dy
−∞ −∞ −∞

Using the definition of conditional density and rearranging the order of integration we
obtain that the above is
Z ∞ Z ∞ Z ∞ Z ∞ Z ∞
f (x, y )
 
= x dx fY (y )dy = x f (x, y )dy dx = xfX (x)dx = E [X ].
−∞ −∞ fY (y ) −∞ −∞ −∞

So we are done. To prove (6.2.4), using Exercise 6.2.8

h(y ) = E [X 2 | Y = y ] − (E [X | Y = y ])2 = E [X 2 | Y = y ] − (g (y ))2

From the above we have,

E [h(Y )] = E [X 2 ] + E [g (Y )2 ]
V ar [g (Y )] = E [g (Y )2 ] − (E [g (Y )])2 = E [g (Y )2 ] − (E [X ])2

Therefore summing the two equations we have (6.2.4). ■

As before it is common to use E [X|Y ] to denote g (Y ) after which the result may be
expressed as E [E [X|Y ]] = E [X ]. This can be slightly confusing notation, but one must

Version: – November 19, 2024

226 summarising continuous random variables

keep in mind that the exterior expected value in the expression E [E [X|Y ]] refers to the
averge of E [X|Y ] viewed as a function of Y .

Similarly one denotes h(Y ) by V ar [X|Y ]. Then we can rewrite (6.2.4) as

V ar [X ] = E [V ar [X|Y ]] + V ar [E [X|Y ]].

Example 6.2.8. Let X ∼ Uniform (0, 1) and be independent of Y ∼ Uniform (0, 1). Let
U = min(X, Y ) and V = max(X, Y ). In Example 6.2.4 we found ρ[U , V ]. During that
computation we showed that the marginal densities of U and V were given by
 
2 ( 1 − u ) if 0 < u < 1 2v if 0 < v < 1
fU (u) = and fV (v ) =
0 otherwise 0 otherwise.

and the joint density of (U , V ) was given by


2 if 0 < u < v < 1
f (u, v ) =
0 otherwise

Let 0 < u < 1. The conditional density of V | U = u, is given by

f (u, v )
fV |U =u (v ) = , for v ∈ R.
fU (u)

So, 
 1
1−u if u < v < 1
fV |U =u (v ) =
0 otherwise

Therefore (V | U = u) ∼ Uniform (u, 1). So the conditional expectation is given by

1 − u2 1+u
Z 1
v
E [V | U = u] = dv = = .
u 1−u 2(1 − u) 2

The conditional variance is given by

V ar [V | U = u] = E [V 2 | U = u] − (E [V | U = u])2
v2 1+u 2
Z 1  
= dv −
u 1−u 2
1−u 3 (1 + u)2 (1 − u)2
= dv − = .
3(1 − u) 4 12

Version: – November 19, 2024

6.2 covariance, correlation, conditional expectation and conditional variance 227

We could have also concluded these from properties of Uniform distribution computed in
Example 6.1.2 and Example 6.1.12. We will use this approach in the next example. ■
Example 6.2.9. Let (X, Y ) have joint probability density function f given by
√
3 − 1 (x2 −xy +y2 )
f (x, y ) = e 2 − ∞ < x, y < ∞.
4π

These random variables were considered in Example 5.4.12. We showed there that X is
a Normal random variable with mean 0 and variance 4
3 and Y is also a Normal random
variable with mean 0 and variance 3.
4
We observed that they are not independent as well
and the conditional distribution of Y given X = x was Normal with mean x
2 and variance
1. Either by direct computation or by definition we observe that

x
E [Y | X = x] = V ar [Y | X = x] = 1.
2

We could compute the V ar [Y ] using (6.2.4), i.e

V ar [Y ] = V ar [E [Y | X ]] + E [V ar [Y | X = x]]
X
= V ar [ ] + E [1]
2
1 14 4
= V ar [X ] + 1 = +1 = .
4 43 3

exercises

Ex. 6.2.1. Let (X, Y ) be uniformly distributed on the triangle 0 < x < y < 1.

(a) Compute E [X|Y = 16 ].

(b) Compute E [(X − Y )2 ].

Ex. 6.2.2. X is a random variable with mean 3 and variance 2. Y is a random variable
with mean −1 and variance 6. The covariance of X and Y is −2. Let U = X + Y and
V = X − Y . Find the correlation coefficient of U and V .
Ex. 6.2.3. Suppose X and Y are both uniformly distributed on [0, 1]. Suppose Cov [X, Y ] =
24 . Compute the variance of X + Y .
−1

Ex. 6.2.4. A dice game between two people is played by a pair of dice being thrown. One
of the dice is green and the other is white. If the green die is larger than the white die,
player number one earns a number of points equal to the value on the green die. If the

Version: – November 19, 2024

228 summarising continuous random variables

green die is less than or equal to the white die, then player number two earns a number
of points equal to the value of the green die. Let X be the random variable representing
the number of points earned by player one after one throw. Let Y be the random variable
representing the number of points earned by player two after one throw.

(a) Compute the expected value of X and of Y .

(b) Without explicitly computing it, would you expect Cov [X, Y ] to be positive or
negative? Explain.

(c) Calculate Cov [X, Y ] to confirm your intuition.

Ex. 6.2.5. Suppose X has variance σX

2 , Y has variance σ 2 , and the pair (X, Y ) has
Y
correlation coefficient ρ[X, Y ].

(a) In terms of σX , σY , and ρ[X, Y ], find Cov [X, Y ] and Cov [X + Y , X − Y ].

(b) What must be true of σX

2 and σ 2 if X + Y and X − Y are uncorrelated?
Y

Ex. 6.2.6. Let (X, Y ) have the joint probability density function f : R2 → R given by

3(x + y ) if x > 0, y > 0, and x + y < 1

(
fX,Y (x, y ) =
0 otherwise

(a) Find E [X|Y = 12 ] and V ar [X|Y = 12 ]

(b) Are X and Y independent ?

Ex. 6.2.7. Suppose Y is uniformly distributed on (0, 1), and suppose for 0 < y < 1 the
conditional density of X | Y = y is given by

 2x2
y
if 0 < x < y
fX|Y =y (x) =
0 otherwise.

(a) Show that, as a function of x, fX|Y =y (x) is a density.

(b) Compute the joint p.d.f. of (X, Y ) and the marginal density of X.

(c) Compute the expected value and variance of X given that Y = y, with 0 < y < 1.

Ex. 6.2.8. Let (X, Y ) have joint probability density function f : R2 → R. Show that
V ar [X | Y = y ] = E [X 2 | Y = y ] − (E [X | Y = y ])2 .

Version: – November 19, 2024

6.2 covariance, correlation, conditional expectation and conditional variance 229

Ex. 6.2.9. For random variables (X, Y ) as in Exercise 5.4.1, find

(a) E [X ] and E [Y ]

(b) V ar [X ] and V ar [Y ]

(c) Cov [X, Y ] and ρ[X, Y ]

Ex. 6.2.10. From Example 5.4.12, consider(X, Y ) have joint probability density function f
given by √
3 − 1 (x2 −xy +y2 )
f (x, y ) = e 2 − ∞ < x, y < ∞.
4π
Find

(a) E [X ] and E [Y ]

(b) V ar [X ] and V ar [Y ]

(c) Cov [X, Y ] and ρ[X, Y ]

Ex. 6.2.11. From Example 5.4.13, suppose T = {(x, y ) | 0 < x < y < 4} and let (X, Y ) ∼
Uniform (T ). Find

(a) E [X ] and E [Y ]

(b) V ar [X ] and V ar [Y ]

(c) Cov [X, Y ] and ρ[X, Y ]

Ex. 6.2.12. From Example 5.4.9, consider the open disk in R2 given by C = {(x, y ) :
x2 + y 2 < 25} and | C |= 25π denote its area. Let (X, Y ) have a joint density f : R2 → R
given by 
 1
|C| if (x, y ) ∈ C
f (x, y ) =
 0 otherwise.

Find

(a) E [X ] and E [Y ]

(b) V ar [X ] and V ar [Y ]

(c) Cov [X, Y ] and ρ[X, Y ]

Ex. 6.2.13. Using the hints provided below prove the respective parts of Theorem 6.2.2

(a) Use the linearity properties of the expected value from Theorem 6.1.8.

Version: – November 19, 2024

230 summarising continuous random variables

(b) Use definition of covariance.

(c) Use the definitions of variance and covariance.

(d) Imitate the proof of Theorem 4.5.7.

(e) Use part (a) of this problem and part (f) of Theorem ??.

(f) Use the linearity properties of the expected value from Theorem 6.1.8.

(g) Use the linearity properties of the expected value from Theorem 6.1.8.

Ex. 6.2.14. Let X, Y be continuous random variable with piecewise continuous densities
f (x) and g (y ) and well-defined expected values. Suppose X ≤ Y then show that E [X ] ≤
E [Y ].

Ex. 6.2.15. Let T be the triangle bounded by the lines y = 0, y = 1 − x, and y = 1 + x.

Suppose a random vector (X, Y ) has a joint p.d.f.

3y if
f(X,Y ) (x, y ) =
0 otherwise.

Compute E [Y |X = 12 ].

Ex. 6.2.16. Let (X, Y ) be random variables with joint probability density function
f : R2 → R. Assume that both random variables have finite variances and that their
covariance is also finite.

(a) Show that V ar [X + Y ] = V ar [X ] + V ar [Y ] + 2Cov [X, Y ].

(b) Show that when X and Y are positively correlated (i.e. ρ[X, Y ] > 0) then V ar [X +
Y ] > V ar [X ] + V ar [Y ], while when X and Y are negatively correlated (i.e. ρ[X, Y ] <
0), then V ar [X + Y ] < V ar [X ] + V ar [Y ].

6.3 moment generating functions

We have already seen for the distribution of a discrete random variable or a continuous
random variable is determined by its distribution function. In this section we shall discuss
the concept of moment generating functions. Under suitable assumptions, these functions
will determine the distribution of random variables. They are also serve as tools in
computations and come in handy for convergence concepts that we will discuss.

Version: – November 19, 2024

 6.3 moment generating functions 231

The moment generating function generates or determine the moments which in turn,
under suitable hypothesis determine the distribution of the corresponding random variable.
We begin with a definition of a moment.

Definition 6.3.1. Suppose X is a random variable. For a positive integer k, the

quantity
mk = E [X k ]

is known as the “k-th moment of X”. As before the existence of a given moment is
determined by whether the above expectation exists or not.

We have previously seen many computations of the first moment E [X ] and also seen
that the second moment E [X 2 ] is related to the variance of the random variable. The
next theorem states that if a moment exists then it guarantees the existence of all lesser
moments.

Theorem 6.3.2. Let X be a random variable and let k be a positive integer. If

E [X k ] < ∞ then E [X j ] < ∞ for all positive integers j < k.

Proof - Suppose X is a continuous random variable. Suppose E [X k ] exists and is finite,

so that E [|X k |] < ∞. Divide R in two pieces by letting R1 = {x ∈ T : |x| < 1} and letting
R2 = {x ∈ T : |x| ≥ 1}. If j < k then |x|j ≤ |x|k for x ∈ R2 so,
Z Z Z
j j j
E [|X |] = |x| fX (x) dx = |x| fX (x) dx + |x|j fX (x) dx
R R1 R2
Z Z
≤ 1 · fX (x) dx + |x|k fX (x) dx
R1 R2
Z Z
≤ fX (x) dx + |x|k fX (x) dx
R1 R2

= 1 + E [|X k |] < ∞

Therefore E [X j ] exists and is finite. See Exericse 6.3.7 when X is a discrete random
variable. ■
When a random variable has finite moments for all positive integers, then these moments
provide a great deal of information about the random variable itself. In fact, in some
cases, these moments serve to completely describe the distribution of the random variable.
One way to simultaneously describe all moments of such a variable in terms of a single
expression is through the use of a “moment generating function”.

Version: – November 19, 2024

232 summarising continuous random variables

Definition 6.3.3. Suppose X is a random variable and D = {t ∈ R : E [etX ] exists}.

The function M : D → R given by

M (t) = E [etX ],

is called the moment generating function for X.

The notation MX (t) will also be used when clarification is needed as to which variable
a particular moment generating function belongs. Note that M (0) = 1 will always be true,
but for other values of t, there is no guarantee that the function is even defined as the
expected value might be infinite. However, when M (t) has derivatives defined at zero,
these values incorporate information about the moments of X. For a discrete random
variable X : S → T with T = {xi : i ∈ N}, then for t ∈ D (as in Definition 6.3.3)

etxi P (X = xi ).
X
MX ( t ) =
i≥1

For a continuous random variable X with probability density function fX : R → R then

for t ∈ D (as in Definition 6.3.3)
Z
MX ( t ) = etx fX (x)dx.
R

We compute moment generating function for a Poisson (λ) and a Gamma (n, λ), with
n ∈ N, λ > 0.

Example 6.3.4. Suppose X ∼ Poisson (λ) then for all t ∈ R,

∞ ∞ ∞ k
λk e−λ (et λ) t t
= e−λ = e−λ ee λ = e−λ(1+e ) .
X X X
MX ( t ) = etk P (X = k ) = etk
k =0 k =0
k! k =0
k!

So the moment generating function of X exists for all t ∈ R. Suppose Y ∼ Gamma (n, λ)
then t < λ,

λn Γ ( n )
n
λn n−1 −λy λn λ
Z Z 
MY ( t ) = ety y e dy = y n−1 e−(λ−t)y dy = = ,
R Γ (n) Γ (n) R Γ (n) (λ − t)n λ−t

where we have used (5.5.3). The moment generating function of Y will not be finite if
t ≥ λ. ■

We summarily compile some facts about moment generating functions. The proof of
some of the results are beyond the scope of this text.

Version: – November 19, 2024

 6.3 moment generating functions 233

Theorem 6.3.5. Suppose for a random variable X, there exists δ > 0 such that
MX (t) exists (−δ, δ ).

(a) The k-th moment of X exists and is given by

(k )
E [ X k ] = MX ( 0 ) ,

(k )
where MX denotes the k-th derivative of MX .

(b) For 0 ̸= a ∈ R such that at, t ∈ (−δ, δ ) we have

MaX (t) = MX (at).

(c) Suppose Y is another independent random variable such that MY (t) exists for
t ∈ (−δ, δ ). Then
MX + Y ( t ) = MX ( t ) MY ( t ) .

for t ∈ (−δ, δ ).

Proof - (a) A precise proof is beyond the scope of this book. We provide a sketch. Express
etX as a power series in t.

t2 X 2 tn X n
etX = 1 + tX + +···+ +...
2 n!

The expected value of the left hand side is the moment generating function for X while
linearity may be used on the right hand side. So the power series of M (t) is given by

t2 tn
M (t) = 1 + t · E [X ] + · E [X 2 ] + · · · + · E [X n ] + . . .
2 n!

Taking k derivatives of both sides of the equation (which is valid in the interval of
convergence) yields

t2
M (k ) ( t ) = E [ X k ] + t · E [ X k +1 ] + · E [ X k +2 ] + . . .
2

Finally, when evaluating both sides at t = 0 all but one term on the right hand side
vanishes and the equation becomes simply M (k) (0) = E [X k ].

(b) MaX (t) = E [e(aX )t ] = E [eX (at) ] = MX (at).

Version: – November 19, 2024

234 summarising continuous random variables

(c) Using Theorem 4.1.10 or Theorem 6.1.10 (f) we have

MX +Y (t) = E [et(X +Y ) ] = E [etX etY ] = E [etX ]E [etY ] = MX (t)MY (t).

■
Theorem 6.3.5 applies equally well for both discrete and continuous variables. A discrete
example is presented next.

Example 6.3.6. Let X ∼ Geometric(p). We shall find MX (t) and use this function to
calculate the expected value and variance X. For any t ∈ R,
∞ ∞ ∞
(et )n · p(1 − p)n−1 = pet · (et · (1 − p))n−1
X X X
MX (t) = E [etX ] = etn P (X = n) =
n=1 n=1 n=1
pet
=
1 − et (1 − p)

Having completed that computation, the expected value and variance can be computed
simply by calculating derivatives.

′ pet
MX (t) =
[ 1 − ( 1 − p ) et ] 2
p
and so E [X ] = MX
′ (0) =
p2
= p1 . Similarly,

′′ pet + p(1 − p)e2t

MX (t) =
[ 1 − ( 1 − p ) et ] 3

2p−p2
and so E [X 2 ] = MX
′′ (0) = = p22 − p1 . Therefore, V ar [X ] = E [X 2 ] − (E [X ])2 = 1−p
p3 p2
.
Both the expected value and variance are in agreement with the previous computations for
the goemetric random variable.
2 2
Let Y ∼ Normal(µ, σ 2 ). The density of Y is fY (y ) = √1 e−(y−µ) /2σ .
σ 2π
For any t ∈ R,

1 1
Z ∞ ∞ Z
2 2 2 2 2 2
tY
MY ( t ) = E [ e ] = e · √ e−(y−µ) /2σ dy =
ty
√ e−(y −(2µy +2σ ty )+µ )/2σ dy
−∞ σ 2π −∞ σ 2π
1
Z ∞
2 2 2 2 2
= eµt+(1/2)σ t √ e−(y−(µ+σ t)) /2σ dy
−∞ σ 2π
2 t2
= eµt+(1/2)σ (6.3.1)

where the integral in the final step is equal to one since it integrates the density of a
Normal(µ + σ 2 t, σ 2 ) random variable. One can easily verify that the MY′ (0) = µ and
MY′′ (0) = µ2 + σ 2 . ■

Version: – November 19, 2024

 6.3 moment generating functions 235

As with the expected value and variance, moment generating functions behave well
when applied to linear combinations of independent variables (courtesy Theorem 6.3.5 (b)
and (c)).

Example 6.3.7. Supppose we wish to find the moment generating function of X ∼

Binomial(n, p). We have seen that such a random variable may arise as the sum of
indpendent Bernoulli variables. That is, X = Y1 + · · · + Yn where Yj ∼ Bernoulli(p). But
it is routine to compute

MYj (t) = E [etYj ] = et·1 P (Yj = 1) + et·0 P (Yj = 0) = pet + (1 − p).

Therefore by linearity (inductively applying Theorem 6.3.5 (c)),

MX (t) = MY1 +···+Yn (t) = MY1 (t) · . . . · MYn (t) = (pet + (1 − p))n .

Moment generating functions are an extraordinarily useful tool in analyzing the distri-
butions of random variables. Two particularly useful tools involve the uniqueness and limit
properties of such generating functions. Unfortunately these theorems require analysis
beyond the scope of this text to prove. We will state the uniqueness fact (unproven)
below and the limit property in Chapter 8. First we generalize the definition of moment
generating functions to pairs of random variables.

Definition 6.3.8. Suppose X and Y are random variables. Then the function

M (s, t) = E [esX +tY ]

is called the (joint) moment generating function for X and Y . The notation
MX,Y (s, t) will be used when confusion may arise as to which random variables are
being represented.

Moment generating functions completely describe the distributions of random variables.

We state the result precisely.

Version: – November 19, 2024

236 summarising continuous random variables

Theorem 6.3.9. (M.G.F. Uniqueness Theorem)

(a) (One variable) Suppose X and Y are random variables and MX (t) = MY (t)
in some open interval containing the origin. Then X and Y are equal in
distribution.

(b) (Two variable) Suppose (X, W ) and (Y , Z ) are pairs of random variables
and suppose MX,W (s, t) = MY ,Z (s, t) in some rectangle containing the origin.
Then (X, W ) and (Y , Z ) have the same joint distribution.

An immediate application of the theorem is an alternate proof of Corollary 5.3.3 based

on moment generating functions.
X−µ
Example 6.3.10. Let X ∼ Normal(µ, σ 2 ) and let Y = σ . Show that Y ∼ Normal(0, 1).
We know X is normal, (6.3.1) shows that the moment generating function of X is
2 t2
MX (t) = eµt+(1/2)σ , for all t ∈ R. So consider the moment generating function of Y .
For all t ∈ R

t
MY (t) = E [etY ] = E [et(X−µ)/σ ] = E [etX/σ e−tµ/σ ] = e−tµ/σ · MX ( )
σ
2 (t/σ )2 t2
= e−tµ/σ · eµ(t/σ )+(1/2)σ =e2.

But this expression is the moment generating function of a Normal(0, 1) random variable.
So by the uniqueness of moment generating functions, Theorem 6.3.9 (a), the distribution
of Y is Normal(0, 1). ■

Just as the joint density of a pair of random variables factors as a product of marginal
densities exactly when the variables are independent (Theorem 5.4.7), a similar result holds
for moment generating functions.

Theorem 6.3.11. Suppose (X, Y ) are a pair of continuous random variables with
moment generating function M (s, t). Then X and Y are indpendent if and only if

M (s, t) = MX (s) · MY (t).

Proof - One direction of the proof follows from basic facts about independence. If X
and Y are independent, then by Exercise 6.3.4 , we have

M (s, t) = E [esX +tY ] = E [esX etY ] = E [esX ]E [etY ] = MX (s) · MY (t).

Version: – November 19, 2024

 6.3 moment generating functions 237

To prove the opposite direction, we shall use Theorem 6.3.9(b). Let X̂ and Ŷ be independent,
but have the same distributions as X and Y respectively. Since MX,Y (s, t) = MX (s)MY (t)
we have the following series of equalities:

MX,Y (s, t) = MX (s)MY (t) = MX̂ (s)MŶ (t) = MX̂,Ŷ (s, t).

By Theorem 6.3.9(b), this means that (X, Y ) and (X̂, Ŷ ) have the same distribution. This
would imply that

P (X ∈ A, Y ∈ B ) = P (X̂ ∈ A, Ŷ ∈ B ) = P (X̂ ∈ A)P (Ŷ ∈ B ) = P (X ∈ A)P (Y ∈ B ),

for any events A and B. Hence X and Y are independent. ■

Notice that the method employed in Example 6.3.10 did not require considering integrals
directly. Since the manipulation of integrals can be complicated (particularly when dealing
with multiple integrals), the moment generating function method will often be simpler as
the next example illustrates.

Example 6.3.12. Let a, b be two real numbers. Let X ∼ Normal(µ1 , σ12 ) and Y ∼
Normal(µ2 , σ22 ) be independent. Observe that

MaX +bY (t) = MX,Y (at, bt)

Using Theorem 6.3.11, we have that the above is

2 σ 2 t2 2 σ 2 t2 2 σ 2 +b2 σ 2 )t2
MX (at)MY (bt) = eaµ1 t+(1/2)a 1 ebµ2 t+(1/2)b 2 = e(aµ1 +bµ2 )t+(1/2)(a 1 2

which is the moment generating function of a Normal random variable with mean aµ1 + bµ2
and variance a2 σ12 + b2 σ22 ). So aX + bY ∼ Normal(aµ1 + bµ2 , a2 σ12 + b2 σ22 ). ■

We conclude this section with a result on finite linear combinations of independent

normal random variables.

Theorem 6.3.13. Let X1 , X2 , . . . , Xn be independent, normally distributed ran-

dom variables with mean µi and variance σi2 respectively for i = 1, 2, . . . n. Let
a1 , a2 , . . . , an be real-valued numbers, not all of which are zero. Then then the linear
combination Y = a1 X1 + a2 X2 + · · · + an Xn is also normally distributed with mean
Pn Pn 2 2
i=1 ai µi and variance i=1 ai σi .

Proof- This follows from the preceeding example by induction and is left as an exercise.
■

Version: – November 19, 2024

238 summarising continuous random variables

exercises

Ex. 6.3.1. Let X ∼ Normal(0, 1). Use the moment generating function of X to calcluate
E [X 4 ].

Ex. 6.3.2. Let Y ∼ Exponential(λ).

(a) Calculate the moment generating function MY (t).

(b) Use (a) to calculate E [Y 3 ] and E [Y 4 ], the third and fourth moments of an exponential
distriubtion.

Ex. 6.3.3. Let X1 , X2 , . . . , Xn be i.i.d. random variables.

(a) Let Y = X1 + · · · + Xn . Prove that MY (t) = [MX1 (t)]n .

(b) Let Z = (X1 + · · · + Xn )/n. Prove that MZ (t) = [MX1 ( nt )]n .

Ex. 6.3.4. Let X and Y be two independent discrete random variables. Let h : R → R
and g : R → R. Show that

E [h(X )g (Y )] = E [h(X )]E [g (Y )].

Show that the above holds if X and Y are independent continous random variables.

Ex. 6.3.5. Suppose X is a discrete random variable and D = {t ∈ R : E [tX ] exists}. The
function ψ : D → R given by
ψ (t) = E [tX ],

is called the probability generating function for X. Calculate the probability generating
function of X when X is

(a) X ∼ Bernoulli(p), with 0 < p < 1.

(b) X ∼ Binomial(n, p), with 0 < p < 1, n ≥ 1.

(c) X ∼ Geometric(p), with 0 < p < 1.

(d) X ∼ Poisson (λ), with 0 < λ.

Ex. 6.3.6. Let X, Y : S → T be dicrete random variables with the number of elements in
T is finite. Prove part (a) of Theorem 6.3.9 in this case.

Ex. 6.3.7. Prove Theorem 6.3.2 when X is a discrete random variable.

Version: – November 19, 2024

 6.4 bivariate normals 239

6.4 bivariate normals

In Example 6.3.12, we saw that if X and Y are independent, normally distributed random
variables, any linear combination aX + bY is also normally distributed. In such a case
the joint density of (X, Y ) is determined easily (courtesy Theorem 5.4.7). We would like
to understand random variables that are not independent but have normally distributed
marginals. Motivated by the observations in Example 6.3.12 we provide the following
definition.

Definition 6.4.1. A pair of random variables (X, Y ) is called “bivariate normal”

if aX + bY is a normally distributed random variable for all real numbers a and b.

We need to be somewhat cautious in the above definition. Since the variables are
dependent it may turn out that aX + bY = 0 or some constant. (E.g: Y = −X,or
Y = −X + 2 with a = 1, b = 1 ). We shall follow the convention that a constant c random
variable in such cases is a normal random variable with mean c and variance 0.

If (X, Y ) are bivariate normal then as X = X + 0Y and Y = 0X + Y both X and Y

individually are normal random variables. The converse if not true (See Exercise 6.4.3).
However the joint distribution of bivariate normal random variables are determined by
their means, variances and covariances. This fact is proved next.

Theorem 6.4.2. Suppose (X, Y ) and (Z, W ) are two bivariate normal random
variables. If

E [ X ] = E [ Z ] = µ1 , E [ Y ] = E [ W ] = µ2
V ar [X ] = V ar [Z ] = σ12 , V ar [Y ] = V ar [W ] = σ22
and
Cov [X, Y ] = Cov [Z, W ] = σ12 (6.4.1)

then (X, Y ) and (Z, W ) have the same joint distribution.

Version: – November 19, 2024

240 summarising continuous random variables

Proof- As (X, Y ) and (Z, W ) are bivariate normal random variables, given real numbers
s, t sX + tY and sZ + tW are normal random variables. Using (6.4.1) and the properties
of mean and covariance (see Theorem 6.2.2) we have

E [sX + tY ] = sE [X ] + tE [Y ] = sµ1 + tµ2 ,

E [sZ + tW ] = sE [Z ] + tE [W ] = sµ1 + tµ2 ,
V ar [sX + tY ] = s2 V ar [X ] + t2 V ar [Y ] + 2stCov [X, Y ]
= s2 σ12 + t2 σ22 + 2stσ12 ,
and
V ar [sZ + tW ] = s2 V ar [Z ] + t2 V ar [W ] + 2stCov [Z, W ]
= s2 σ12 + t2 σ22 + 2stσ12 .

From the above, sX + tY and sZ + tW have the same mean and variance. So they have
the same distribution (as normal random variables are determined by their mean and
variances). By Theorem 6.3.9 (a) they have the same moment generating function. So, the
(joint) moment generating function of (X, Y ) at (s,t) is

MX,Y (s, t) = E [esX +tY ] = MsX +tY (1) = MsZ +tW (1) = E [esZ +tW ] = MZ,W (s, t)

Therefore (Z, W ) has the same joint m.g.f. as (X,Y) and Theorem 6.3.9 (b) implies that
they have the same joint distribution. ■
Though, in general, two variables which are uncorrelated may not be independent, it
is a remarkable fact that the two concepts are equivalent for bivariate normal random
variables.

Theorem 6.4.3. Let (X, Y ) be a bivariate normal random variable. Then

Cov [X, Y ] = 0 if and only if X and Y are independent.

Proof - That independence implies a zero covariance is true for any pair of random variables
(use Theorem 6.1.10 (e)), so we need to only consider the reverse implication.
Suppose Cov [X, Y ] = 0. Let µX and σX
2 denote the expected value and variance of X

and µY and σY2 the corresponding values for Y . Let s and t be real numbers. Then, by
the bivariate normality of (X, Y ), we know sX + tY is normally distributed. Moreover by
properties of expected value and variance we have

E [sX + tY ] = sE [X ] + tE [Y ] = sµX + tµY

Version: – November 19, 2024

 6.4 bivariate normals 241

−0.3 0 0.7

g(y1, y2) g(y1, y2) g(y1, y2)

3 3 3
2 2 2
1 1 1
3 3 3
y20−1 1
2 y20−1 1
2 y20−1 1
2
0 0 0
−2 −1 y1 −2 −1 y1 −2 −1 y1
−3−3 −2 −3−3 −2 −3−3 −2

−0.3 0 0.7

2
0.05 0.05
0.10
1 0.10 0.10 0.15
0.15 0.15 0.20
y2

−1
0.05
−2

−2 −1 0 1 2 −2 −1 0 1 2 −2 −1 0 1 2
y1

Figure 6.1: The density function of Bivariate Normal distributions. The set of panels on top
show a three-dimensional view of the density function for various values of the
correlation ρ. The bottom set of panels show contour plots, where each ellipse
corresponds to the (y1 , y2 ) pairs corresponding to a constant value of g (y1 , y2 ).

and
V ar [sX + tY ] = s2 V ar [X ] + 2stCov [X, Y ] + t2 V ar [Y ] = s2 σX
2
+ t2 σY2 .

That is, sX + tY ∼ Normal(sµX + tµY , s2 σX

2 + t2 σ 2 ). So for all s, t ∈ R
Y

2 σ 2 +t2 σ 2
MX,Y (s, t) = E [esX +tY ] = MsX +tY (1) = e(sµX +tµY )+(1/2)(s X Y )

2 σ2 2 σ2
= esµx +(1/2)s X · etµY +(1/2)t Y

= MX ( s ) · M Y ( t ) .

Hence by Theorem 6.3.11 X and Y are independent. ■

We conclude this section by finding the joint density of a Bivariate normal random
variable. See Figure 6.1 for a graphical display of this density.

Version: – November 19, 2024

242 summarising continuous random variables

Theorem 6.4.4. Let (Y1 , Y2 ) be a bivariate Normal random variable, with µ1 =

E [Y1 ], µ2 = E [Y2 ], 0 ̸= σ12 = V ar [Y1 ], 0 ̸= σ22 = V ar [Y2 ], and σ12 = Cov [Y1 , Y2 ].
Assume that the correlation coefficient |ρ[Y1 , Y2 ]| ̸= 1. Then the joint probability
density function of (Y1 , Y2 ), g : R2 → [0, ∞) is given by
  
y1 −µ1 2 y2 −µ2 2
    
y1 −µ1 y2 −µ2
exp − 2(1−ρ
1
2) σ1 + σ2 − 2ρ σ1 σ2
g (y1 , y2 ) = (6.4.2)
2πσ1 σ2 1 − ρ2
p

Proof- Let a, b be two real numbers. We will show that

Z a Z b
P (Y1 ≤ a, Y2 ≤ b) = g (y1 , y2 )dy2 dy1 . (6.4.3)
−∞ −∞

From the discussion that follows (5.4.1), we can then conclude that the joint density of
(Y1 , Y2 ) is indeed given by g. To show (6.4.3) we find an alternate description of (Y1 , Y2 )
which is the same in distribution. Let Z1 , Z2 be two independent standard normal random
variables. Define

U = σ 1 Z 1 + µ1 (6.4.4)
q
V = σ2 (ρZ1 + 1 − ρ2 Z2 ) + µ2

Let α, β ∈ R. Then
q
αU + βV = (ασ1 + βσ2 ρ)Z1 + (βσ2 1 − ρ2 )Z2 + α1 µ1 + βµ2 .

As Z1 and Z2 are independent standard normal random variables by Theorem 6.3.13,

(ασ1 + βσ2 ρ)Z1 + (βσ2 1 − ρ2 )Z2 ∼ Normal (0, (ασ1 + βσ2 ρ)2 + (βσ2 1 − ρ2 )2 ). Fur-
p p

ther using Corollary 5.3.3 (a) we have that αU + βV ∼ Normal (α1 µ1 + βµ2 , (ασ1 +
βσ2 ρ)2 + (βσ2 1 − ρ2 )2 ). As α, β were arbitrary real numbers by Definition 6.4.1, (U , V )
p

is a bivariate normal random variable.

Using Theorem 6.1.8 and Theorem 6.1.10 (d) that,

µ1 = E [U ], µ2 = E [V ], Var[U ] = σ12 .

Version: – November 19, 2024

 6.4 bivariate normals 243

Also in addition, using Exercise 6.2.16 and Theorem 6.2.2 (f), we have
q
V ar [V ] = σ22 ρ2 V ar [Z1 ] + σ22 (1 − ρ2 )V ar [Z2 ] + 2(σ2 (ρ + 1 − ρ2 )Cov [Z1 , Z2 ]
= σ22 ρ2 + σ22 (1 − ρ2 ) + 0 = σ22
and
q
Cov [U , V ] = Cov [σ1 Z1 + µ1 , σ2 (ρZ1 + 1 − ρ2 Z2 )]
q
= σ1 σ2 ρCov [Z1 , Z1 ] + σ1 σ2 1 − ρ2 Cov [Z1 , Z2 ]
= σ1 σ2 ρ + 0 = σ12 .

As bivariate normal random variables are by their means and covariances (by Theorem
6.4.2), (Y1 , Y2 ) and (U , V ) have the same joint distribution. By the above, we have

P (Y1 ≤ a, Y2 ≤ b) = P (U ≤ a, V ≤ b). (6.4.5)

By elementary algebra we can also infer from (6.4.4)

U − µ1 V − µ2 ρZ1
Z1 = , Z2 = p −p .
σ1 σ2 1 − ρ2 1 − ρ2

So
( )
a − µ1 b − µ2 ρZ1
{U ≤ a, V ≤ b} = Z1 ≤ , Z2 ≤ p −p
σ1 σ2 1 − ρ 2 1 − ρ2

So, using this fact in (6.4.5) we get

!
a − µ1 b − µ2 ρZ1
P (Y1 ≤ a, Y2 ≤ b) = P Z1 ≤ , Z2 ≤ p −
σ2 1 − ρ2 1 − ρ2
p
σ1
a−µ1 b−µ2 ρz z 2 +z22
√ −√ 1 2 exp(− 1 2 )
Z Z
σ1 1−ρ2
= σ2 1−ρ
dz2 dz1 (6.4.6)
−∞ −∞ 2π

First performing a u-substitution in the inner integral for each fixed z1 ,

y2 − µ2 ρz1
z2 = −p
σ2 1 − ρ 1 − ρ2
p
2

Version: – November 19, 2024

244 summarising continuous random variables

yields that the inner integral in (6.4.6) for each z1 ∈ R

 2
y2 −µ2 ρz
z12 + √ −√ 1 2
b−µ2 ρz z12 +z22 σ2 1−ρ2
√ −√ 1 2 exp(− exp(− 1−ρ
Z b
) )
Z
1−ρ2 2
σ2 1−ρ
dz2 = p 2 dy2
−∞ 2π −∞ 2πσ2 1 − ρ2
Z b exp(− y −µ
1
2(1−ρ2 )
[(1 − ρ )z1 + ( 2σ2 2 − ρz1 ) ])
2 2 2
= dy2
2πσ2 1 − ρ2
p
−∞
Z b exp(− 1 [z 2 + ( y2 −µ2 )2 − 2ρ( y2 −µ2 )z ])
2(1−ρ2 ) 1 σ σ2 1
= p2 dy2 .
−∞ 2πσ2 1 − ρ 2

Substituting the above into (6.4.6), we have

y2 −µ2 2
Z a−µ1
σ1
Z b exp(− 2(1−ρ
1 2
2 ) [ z1 + ( σ ) − 2ρ( y2σ−µ 2
)z1 ])
P (Y1 ≤ a, Y2 ≤ b) = 2 2
dy2 .dz1
2πσ2 1 − ρ2
p
−∞ −∞
(6.4.7)

Performing a u-subsitution
y1 − µ1
z1 =
σ1
on the outer integral above we obtain

P (Y1 ≤ a, Y2 ≤ b)
 2  
y1 −µ1
Z a Z b exp(− 2(1−ρ
1
2) [ σ1 + ( y2σ−µ
2
) − 2ρ( y2σ−µ
2 2
2
2
) y1 −µ1
σ1 ])
= dy2 dy1
2πσ1 σ2 1 − ρ2
p
−∞ −∞

Thus we have established (6.4.3). ■

exercises

Ex. 6.4.1. Let X1 , X2 be two independent Normal random variables with mean 0 and variance 1.
Show that (X1 , X2 ) is a bivariate normal random variable.

Ex. 6.4.2. Let (X1 , X2 ) be a bivariate normal random variable. Assume that the correlation
coefficient |ρ[X1 , X2 ]| ̸= 1. Show that X1 and X2 are Normal random variables by calculating their
marginal densities.

Ex. 6.4.3. Let X1 , X2 be two independent normal random variables with mean 0 and variance
1. Let (Y1 , Y2 ) be a bivariate normal random variable with zero means, variances equal to 1 and
correlation ρ = ρ[Y1 , Y2 ], with ρ2 ̸= 1. Let f be the joint probability density function of (X1 , X2 )

Version: – November 19, 2024

 6.4 bivariate normals 245

and g be the joint probability density function of (Y1 , Y2 ). For 0 < α < 1, let (Z1 , Z2 ) be a bivariate
random variable with joint density given by

h(z1 , z2 ) = αg (z1 , z2 ) + (1 − α)f (z1 , z2 ),

for any real numbers z1 , z2 .

(a) Write down the exact expressions for f and g.

(b) Verify that h is indeed a probability density function.

(c) Show that Z1 and Z2 are Normal random variables by calculating their marginal densities.

(d) Show that (Z1 , Z2 ) is not a bivariate normal random variable.

Ex. 6.4.4. Suppose X1 , X2 , . . . , Xn are independent and normaly distributed. Let Y = c1 X1 + · · · +

cn Xn and let Z = d1 X1 + · · · + dn Xn be linear combinations of these variables (for real numbers
cj and dj ). Then (Y , Z ) is bivariate normal.
Ex. 6.4.5. Prove Theorem 6.3.13. Specifically, suppose for i = 1, 2, . . . , n that Xi ∼ Normal(µi , σi2 )
with X1 , X2 , . . . , Xn independent. Let a1 , a2 , . . . , an be real numbers, not all zero, and let Y =
a1 X1 + a2 X2 + · · · + an Xn . Prove that Y is normally distributed and find its mean and variance
in terms of the a’s, µ’s, and σ’s.
Ex. 6.4.6. Let (X1 , X2 ) be a bivariate Normal random variable. Define

Cov [X1 , X1 ] Cov [X1 , X2 ]

 

Σ=
 

Cov [X1 , X2 ] Cov [X2 , X2 ]
" #
µ1
and µ1 = E [X1 ], µ2 = E [X2 ], µ2×1 = .
µ2

Σ is referred to as the covariance matrix of (X1 , X2 ) and µ is the mean matrix of (X1 , X2 ).

(a) Compute det(Σ).

(b) Show that the joint density of (X1 , X2 ) can be rewritten in matrix notation as
" #!
1 1h i x1 − µ 1
g ( x1 , x2 ) = exp − x1 − µ 1 x2 − µ 2 Σ−1
2π det(Σ) 2
p
x2 − µ 2

(c) " # " #

a11 a12 η1
A2×2 = , η2×1 =
a21 a22 η2
such that aij are real numbers. Suppose we define
" #
a11 X1 + a12 X2 + η1
Y = AX = .
a21 X1 + a22 X2 + η2

Version: – November 19, 2024

246 summarising continuous random variables

Then (Y1 , Y2 ) is also a bivariate Normal random variable, with covariance matrix AΣAT and
mean matrix Aµ + η.
Hint: Compute means, variances and covariances of Y1 , Y2 and use Theorem 6.4.2

Version: – November 19, 2024

 7
S A M P L I N G A N D D E S C R I P T I V E S TAT I S T I C S

The distinction between Probability and Statistics is somewhat blurred, but largely has to do with
the perspective of what is known versus what is to be determined. One may think of Probability as
the study of models for (random) experiments when the model is fully known. When the model is
not fully known and one tries to infer about the unknown aspects of the model based on observed
outcomes of the experiment, this is where Statistics enters the picture. In this chapter we will be
interested in problems where we assume we know the outputs of random variables, and wish to use
that information to say what we can about their (unknown) distributions.

Suppose, for instance, we sample from a large population and record a numerical fact associated
with each selection. This may be recording the heights of people, recording the arsenic content
of water samples, recording the diameters of randomly selected trees, or anything else that may
be thought of as repeated, random measurements. Sampling an individual from a population in
this case may be viewed as a random experiment. If the sampling were done at random with
replacement with each selection independent of any other, we could view the resulting numerical
measurements as i.i.d. random variables X1 , X2 , . . . , Xn . A more common situation is sampling
without replacement, but we have previously seen (see Section 2.3) that when the sample size is
small relative to the size of the population, the two sampling methods are not dramatically different.
In this case we have the results of n samples from a distribution, but we do not actually know the
distribution itself. How might we use the samples to attempt to predict or “infer” such things as
expected value and variance?

7.1 descriptive statistics

A natural quantity we can create from the observed data, regardless of the underlying distribution
that generated it, is a discrete distribution that puts equal probability on each observed point. This
distribution is known as the empirical distribution. Inferences based on the empirical distribution
are traditionally referred to as “descriptive statistics”. In later chapters, we will see that making
additional assumptions lets us make “better” inferences, provided the additional assumptions are
valid.

We will assume that the random variables X1 , X2 , . . . , Xn are i.i.d. from some common
distribution, usually unknown. Some values of Xi can of course be repeated, so the empirical
distribution (and the empirical cummulative distribution function) is formally defined as follows.

247

Version: – November 19, 2024

248 sampling and descriptive statistics

Definition 7.1.1. Let X1 , X2 , . . . , Xn be random variables be i.i.d with distribution X.

The “empirical distribution” based on these is the discrete distribution with probability mass
function given by
1
fn (t) = |{i : Xi = t}| ,
n
for t ∈ R. Further, for x ∈ R

|{i : Xi ≤ x}|
Fn (x) = ,
n
is known as the “empirical cumulative distribution function” or ECDF of X1 , X2 , . . . , Xn .

Given a realisation of X1 , X2 , . . . , Xn ECDF are easy to compute and provide information about
the underlying distribution. One can also show that as n → ∞ the ECDF will converge to the
underlying distribution function.

Example 7.1.2. Suppose we surveyed 10 random people and asked them how many litres of water
they consume in a day. Suppose the data collected was the following:

3 4 2 5 2 4 4 6 3 4

We can compute the empirical probability mass function and the empirical cummulative distribution
function. That is,

 



2
10 if t = 2, 

 0 if x < 2,
2
if t = 3, 2
if 2 ≤ x < 3,

 

10 10

 

 
4
if t = 4, 4
if 3 ≤ x < 4,

 

10  10

 
f10 (t) = 1
10 if t = 5, and F10 (x) = 8
10 if 4 ≤ x < 5,
1
if t=6 9
if 5 ≤ x < 6,

 

 

 10 
 10
and and

 


 

 
 0

otherwise  1

if 6 ≤ x

R has an built function called ecdf which will compute the empirical cumulative distribution
function given the data with options for plotting as indicated below.

x = c(3, 4, 2 , 5 , 2 , 4 , 4 , 6 , 3 , 4 )
F= ecdf(x)
plot(F)

Note that, the empirical distribution is a random object, as it is defined in terms of random
variables. However, for any fixed realisation of these random variables X1 , X2 , . . . , Xn , the corre-
sponding empirical distribution is a fixed probability distribution, so we can now study it using

Version: – November 19, 2024

 7.1 descriptive statistics 249

the tools of probability. Doing so does not make any additional assumptions about the underlying
distribution.

It is important to realize that the empirical distribution is itself a random quantity, as each
sample realisation will produce a different discrete distribution. We intuitively expect it to carry
information about the underlying distribution, especially as the sample size n grows. For example,
the expectation computed from the empirical distribution should be closely related to the true
underlying expectation, probabilities of events computed from the empirical distribution should be
related to the true probabilities of those events, and so on. In the remainder of this chapter, we
will make this intuition more precise and describe some tools to investigate the properties of the
empirical distribution.

7.1.1 Sample Mean and Sample Variance

Given a sample of observations X1 , X2 , . . . , Xn from a distribution X, we define the sample mean

to be the familiar definition of average. We shall present the precise definition first and then a
result that describes how well does the sample mean work as an estimate of the true mean of the
distribution X.

Definition 7.1.3. Let X1 , X2 , . . . , Xn be i.i.d. random variables with distribution X. The

“sample mean” of these is
X1 + X2 + · · · + Xn
X= .
n

It is easy to see that X is the expected value of a random variable whose distribution is the empirical
distribution based on X1 , X2 , . . . , Xn (see Exercise 7.1.5). Suppose the Xj random variables have a
finite expected value µ. The sample mean X is not the same as this expected value. In particular
µ is a fixed constant while X is a random variable. From the statistical perspective, µ is usually
assumed to be an unknown quantity while X is something that may be computed from the results
of the sample X1 , X2 , . . . , Xn . The next theorem is a first step in answering how well does X work
as an estimate of µ.

Theorem 7.1.4. Let X1 , X2 , . . . , Xn be an i.i.d. sample of random variables whose distri-

bution has finite expected value µ and finite variance σ 2 . Let X represent the sample mean.
Then
σ
E [X ] = µ and SD [X ] = √ .
n

Version: – November 19, 2024

250 sampling and descriptive statistics

Proof. We can write

 
  X1 + X2 + · · · + Xn
E X = E
n
E [X1 ] + E [X2 ] + · · · + E [Xn ]
=
n
nµ
= =µ
n
To calculate the standard deviation, we consider the variance and use Theorem 4.2.6 and Exercise
6.1.12 to obtain
 
  X1 + X2 + · · · + Xn
V ar X = V ar
n
V ar [X1 ] + V ar [X2 ] + · · · + V ar [Xn ]
=
n2
nσ 2 σ 2
= =
n2 n

Taking square roots then shows SD X = √σn .

 
■

The fact that E [X ] = µ implies that, on average, the quantity X is accurately describing the
unknown mean µ. In the language of statistics X is said to be an “unbiased estimator” of the
quantity µ. Note also that SD [X ] → 0 as n → ∞ meaning that the larger the sample size, the
more accurately X reflects its average of µ. In other words, if there is an unknown distribution
from which it is possible to sample, averaging a large sample should produce a value close to the
expected value of the distribution. In technical terms, this is considered as a notion of consistency
and we say that the sample mean is a “consistent estimator” of the population mean µ.
Given a sample of observations from a given distribution one may try to estimate the variance
of the distribution via the sample variance which we define below.

Definition 7.1.5. Let X1 , X2 , . . . , Xn be i.i.d. random variables. The “sample variance”

of these is
(X1 − X )2 + (X2 − X )2 + · · · + (Xn − X )2
S2 = .
n−1

Note that this definition is not universal; it is common to define sample variance with n (instead
of n − 1) in the denominator, in which case the definition matches the variance of the empirical
distribution of X1 , X2 , . . . , Xn (Exercise 7.1.5). The definition given here produces a quantity that
is unbiased for the underlying population variance, a fact that follows from the next theorem.

Theorem 7.1.6. Let X1 , X2 , . . . , Xn be an i.i.d. sample of random vairables whose distri-

bution has finite expected value µ and finite variance σ 2 . Then S 2 is an unbiased estimator
of σ 2 , i.e.,
E [S 2 ] = σ 2 .

Version: – November 19, 2024

 7.1 descriptive statistics 251

Proof. First note that

2 σ2
E [X ] = V ar [X ] + (E [X ])2 = + µ2
n
whereas
E [Xj2 ] = V ar [Xj ] + E [Xj ]2 = σ 2 + µ2 .

Now consider the quantity (n − 1)S 2 .

E [(n − 1)S 2 ] = E [(X1 − X )2 + (X2 − X )2 + · · · + (Xn − X )2 ]

= E [X12 + X22 + · · · + Xn2 ] − 2E [(X1 + X2 + · · · + Xn )X ]
2 2 2
+E [X + X + · · · + X ]

But X1 + X2 + · · · + Xn = nX, so

2 2
E [(n − 1)S 2 ] = E [X12 + X22 + · · · + Xn2 ] − 2nE [X ] + nE [X ]
2
= E [X12 + X22 + · · · + Xn2 ] − nE [X ]
σ2
= n ( σ 2 + µ2 ) − n ( + µ2 ) = ( n − 1 ) σ 2
n

Dividing by n − 1 gives the desired result, E [S 2 ] = σ 2 . ■

A more important property (than unbiasedness) is that S 2 and its variant with n in the denominator
are both “consistent” for σ 2 , just as X was for µ, in the sense that V ar [S 2 ] → 0 as n → ∞ under
some mild conditions (See Exercise 7.1.7). One may also try to estimate σ from S but due to
vagaries of averaging (in turn expectation) one will typically loose the unbiasedness property (See
Exercise 7.1.8).
Expectation and variance are commonly used summaries of a random variable, but they do not
characterize its distribution completely. In the next subsection we shall see how to use the idea of
sample proportion to understand the underlying distribution better.

7.1.2 Sample proportion

In general, the distribution of a random variable X is fully known if we can compute P (X ∈ A) for
any event A. On the other hand if the distribution is not known and we have an event A of interest
then we can use the empirical distribution to estimate the probability P (X ∈ A).
Given a sample of i.i.d. observations X1 , X2 , . . . , Xn from a common distribution defined by a
random variable X, let Y be the random variable that has the same distribution as the empirical
distribution based on sample. More precisely,

|{i : Xi = t|}
Range(Y ) = {X1 , X2 , . . . , Xn } and P (Y = t) = , for t ∈ Range(Y ).
n

Version: – November 19, 2024

252 sampling and descriptive statistics

Let A be the event of interest, then

|{i : Xi ∈ A}|
P (Y ∈ A) = .
n

In other words, P (Y ∈ A) is simply the proportion of sample observations for which the event A
happened. Not surprisingly, P (Y ∈ A) is a good estimator of P (X ∈ A).

Theorem 7.1.7. Let X1 , X2 , . . . , Xn be an i.i.d. sample of random variables with the

same distribution as a random variable X. Suppose that we are interested in the value
p = P (X ∈ A) for an event A. Let

|{i : Xi ∈ A}|
p̂n = .
n

Then, E [p̂n ] = P (X ∈ A) and V ar (p̂n ) → 0 as n → ∞.

Proof. For 1 ≤ i ≤ n, let 

1 if Xi ∈ A
Zi =
0 otherwise.

It is easy to see that

n
X
|{i : Xi ∈ A}| = Zi
i=1

and Zi ’s are independent because Xi ’s are independent (See Theorem 3.3.6 and Exercise 7.1.2) and
identically distributed with
P (Zi = 1) = P (Xi ∈ A) = p.
Pn
Thus, i=1 Zi has the Binomial distribution with parameters n and p, with expectation np and
variance np(1 − p). It is immediate that
Pn n
i = 1 Zi 1 X
E [p̂n ] = E [ ]= E[ Zi ] = p
n n
i=1
and
Pn n
i = 1 Zi 1 X
V ar [p̂n ] = V ar [ ]= 2
V ar [ Zi ] = p(1 − p)/n. (7.1.1)
n n
i=1

The result follows. ■

This result is a special case of the more general “law of large numbers” we will encounter in Section
8.2. It is important because it gives formal credence to our intuition that the probability of an
event measures the limiting relative frequency of that event over repeated trials of an experiment.

Example 7.1.8. Suppose that U and V are independent Uniform(0, 1), and we interpret (U , V )as
coordinates of a point in R2 . Let A be the event that the point (U , V ) is inside the unit circle

Version: – November 19, 2024

 7.1 descriptive statistics 253

and we wish to estimate the probability of p = P ((U , V ) ∈ A) = P (U 2 + V 2 < 1). We can

find the answer by a direct computation (See Exercise 7.1.6) that p = π4 . We can estimate this
probability using the sampling proportion. We simulate the experiment a large number of times,
and computing the proportion of cases where the simulated Z is less than 1. That is generate
samples {(Ui , Vi ) : 1 ≤ i ≤ n}

replicate(10, {
u <- runif(10000)
v <- runif(10000)
z <- sqrt(uˆ2 + vˆ2)
sum(z < 1) / 10000
})

[1] 0.7820 0.7791 0.7882 0.7834 0.7888 0.7802 0.7861 0.7816 0.7813
[10] 0.7872

We can see that our estimates are quite good with n = 10000 that p̂n is very close to p. A little
thought tells us that the true probability P (Z < 1) = π/4 ≈ 0.7854. The simulation experiment we
have performed above is in fact one way of estimating π, although it is not a particularly efficient
one. We illustrate this below by repeating the experiment with 1000000 trials and multiplying the
observed sample proportion by 4. Note that in this experiment z < 1 ⇐⇒ z 2 < 1, so calculating
the square root is unnecessary.

u <- runif(1000000)
v <- runif(1000000)
zsq <- uˆ2 + vˆ2
4 * mean(zsq < 1)

[1] 3.13966

As the variance of the sample proportion p̂ is given by n1 p(1 − p) (see (7.1.1)), increasing the
number of replications by a factor of 100 (from 104 to 106 ) leads to an improvement in the accuracy
√
(in terms of standard deviation) of the estimate of π by a factor of 100 = 10. ■
Example 7.1.9. Suppose we are given A, B, C are independent Poisson random variables with
parameters α, β and γ respectively. What is the probablity that the equation Ax2 − Bx + C = 0
has a real solution? To answer this question one would have to calculate the probability that
B 2 − 4AC > 0. That would imply evaluating
∞ X
∞ X
∞ 
(α )a (β )b (γ )c
X 
1(b − 4ac ≥ 0) exp(−α − β − γ )
2
,
a! b! c!
a=0 b=0 c=0

which would require some combinatorial effort. However we can use the strong law of large numbers
and try to estimate the number via simulations.

Version: – November 19, 2024

254 sampling and descriptive statistics

B <- rpois(10000,6) B <- rpois(10000,5)

A <- rpois(10000,3) A <- rpois(10000,5)
C <- rpois(10000,3) C <- rpois(10000,5)
D <- Bˆ2-4*A*C D <- Bˆ2-4*A*C
mean(D >=0) mean(D >=0)

[1] 0.5906 [1] 0.1397

hist(D) hist(D)

Histogram of D Histogram of D

3000
4000

2500
3000

2000
Frequency

Frequency

1500
2000

1000
1000

500
0

−200 −100 0 100 200 300 −600 −400 −200 0 200

D D

exercises

Ex. 7.1.1. Consider the following data

13 40 23 15 21 4 44 16 32 14

(a) Compute the probability mass function of the empirical distribution from the data and also
the corresponding ECDF.

(b) Using ecdf and plot command in R do part (a).

Ex. 7.1.2. Verify that the proofs of Theorem 3.3.5 and Theorem 3.3.6 hold for continuous random
variables.

Version: – November 19, 2024

 7.2 simulation 255

Ex. 7.1.3. Let X and Y be two continuous random variables having the same distribution. Let
f : R → R be a piecewise continuous function. Then show that f (X ) and f (Y ) have the same
distribution.
Ex. 7.1.4. Let X and Y be two discrete variables having the same distribution. Let f : R → R be
a piecewise continuous function. Then show that f (X ) and f (Y ) have the same distribution.
Ex. 7.1.5. Let P be the empirical distribution defined by sample observations X1 , X2 , . . . , Xn . In
other words, P is the discrete distribution with probability mass function given in Definition 7.1.1.
Let Y be a random variable with distribution P .
(a) Show that E [Y ] = X.
(b) Show that V ar [Y ] = n S .
n−1 2

Ex. 7.1.6. Suppose that U and V are independent Uniform(0, 1), and we interpret (U , V ) as
coordinates of a point in R2 .
√
1. Let Z = U 2 + V 2 . Find p := P (Z < 1).
2. Can you modify the above R-code given in Example 7.1.8 to provide an estimate for π ?
3. Can you modify the above R-code to observe that the variance of the estimator of p goes to
0?
Ex. 7.1.7. Let X1 , X2 , . . . , Xn be i.i.d. random variables with finite expectation µ, finite variance σ 2 ,
and finite γ = E [X1 − µ]4 . Compute V ar (S 2 ) in terms of µ, σ 2 , and γ and show that V ar (S 2 ) → 0
as n → ∞.
Ex. 7.1.8. Let X1 , X2 , . . . , Xn be i.i.d. random variables with finite expectation µ and finite
√
variance σ 2 . let S = S 2 , the non-negative root of the sample variance. The quantity S is called
the “sample standard deviation”. Although E [S 2 ] = σ 2 , it is not true that E [S ] = σ. In other
words, S is not an unbiased estimator for σ. Follow the steps below to see why.
(a) Let Z be a random variable with finite mean and finite variance. Prove that E [Z 2 ] ≥ E [Z ]2
and give an example to show that equality may not hold. (Hint: Consider how these quantities
relate to the variance of Z).
(b) Use (a) to explain why E [S ] ≤ σ and give an example to show that equality may not hold.

7.2 simulation

The preceding discussion gives several mathematical statements about random samples, but it is
difficult to develop any intuition about what these statements mean unless we look at actual data.
Data is of course abundant in our world; however, the problem with real data is that we do not
usually know for certain the random variable that generated it. To hone our intuition, it is therefore
useful to be able to generate random samples from a distribution we specify. The process of doing
so using a computer program is known as “simulation”.
Simulation is not an easy task, because computers are by nature not random. Simulation is
in fact not a random process at all; it is a completely deterministic process that tries to mimic

Version: – November 19, 2024

256 sampling and descriptive statistics

randomness. We will not go into how simulation is done, but simply use R to obtain simulated
random samples.
R supports simulation from many distributions, including all the ones we have encountered.
The general pattern of usage is that each distribution has a corresponding function that is called
with the sample size an argument, and further arguments specifying parameters. The function
returns the simulated observations as a vector. For example, 30 Binomial(100, 0.75) samples can be
generated by

rbinom(30, size = 100, prob = 0.75)

[1] 63 72 77 75 82 73 69 78 68 76 87 67 75 73 68 64 71 74 65 79 72 79
[23] 76 72 70 74 72 69 74 72

We usually want to do more than just print simulated data, so we typically store the result in
a variable and make further calculations with it; for example, compute the sample mean, or the
sample proportion of cases where a particular event happens.

x <- rbinom(30, size = 100, prob = 0.75)

mean(x)

[1] 75.66667

sum(x >= 75) / length(x)

[1] 0.5666667

R has a useful function called replicate that allows us to repeat such an experiment several
times.

replicate(15, {
x <- rbinom(30, size = 100, prob = 0.75)
mean(x)
})

[1] 77.06667 76.60000 74.90000 73.66667 74.53333 76.40000 74.00000

[8] 76.00000 74.33333 75.03333 74.76667 74.36667 75.16667 75.46667
[15] 75.26667

Version: – November 19, 2024

 7.2 simulation 257

replicate(15, {
x <- rbinom(30, size = 100, prob = 0.75)
sum(x >= 75) / length(x)
})

[1] 0.5666667 0.5333333 0.6000000 0.5333333 0.4333333 0.6000000

[7] 0.5666667 0.5000000 0.7000000 0.5333333 0.6000000 0.4333333
[13] 0.7000000 0.6000000 0.3666667

This gives us an idea of the variability of the sample mean and sample proportion computed from a
sample of size 30. We know of course that the sample mean has expectation 100 × 0.75 = 75, and
we can use R to compute the expected value of the proportion as follows.

1 - pbinom(74, size = 100, prob = 0.75)

[1] 0.5534708

So the correponding estimates are close to the expected values, but with some variability. We
expect the variability to go down if the sample size increases, say, from 30 to 3000.

replicate(15, {
x <- rbinom(3000, size = 100, prob = 0.75)
mean(x)
})

[1] 75.01667 75.00100 74.90767 75.08067 75.01767 75.02867 74.97900

[8] 74.98433 74.94667 75.08833 74.93900 74.94833 74.91500 74.96400
[15] 74.92700

replicate(15, {
x <- rbinom(3000, size = 100, prob = 0.75)
sum(x >= 75) / length(x)
})

[1] 0.5516667 0.5403333 0.5583333 0.5520000 0.5520000 0.5423333

[7] 0.5590000 0.5493333 0.5403333 0.5556667 0.5543333 0.5563333
[13] 0.5563333 0.5496667 0.5446667

Indeed we see that the estimates are much closer to their expected values now.
We can of course repeat this process for other events of interest, and indeed for many other
distributions. We will see in the next section how we can simulate observations following the normal
distribution using the funtion rnorm, and the exponential distribution using the funtion rexp. It is

Version: – November 19, 2024

258 sampling and descriptive statistics

also interesting to think about how one can simulate observations from a given distribution when a
function to do so is not already available.
Recall from Lemma 5.3.7, that suppose U ∼ Uniform (0, 1) random variable and X is a
continuous random variable such that its distribution function, FX , is a strictly increasing continous
−1
function then Y = FX (U ) has the same distribution as X. This approach can be used to be
simulate distributions (both discrete and continuous) from Uniform samples. The following examples
explore this approach. We begin with an example on how to simulate Poisson samples from Uniform.

Example 7.2.1. When trying to formulate a method to simulate random variables from a new
distribution, it is customary to assume that we already have a method to generate random variables
from Uniform(0, 1). Let us see this can be used to generate random observations from a Poisson(λ)
distribution using its probability mass function.
Let X denote an observation from the Poisson(λ) distribution, and U ∼ Uniform(0, 1). Denote
pi = P (X = i). An algorithm to generate a random variable with the same distribution as X is
suggested by the following observation.

p0 = P (U ≤ p0 ),
P (U ≤ p0 + p1 ) = p0 + p1 ⇒ p1 = P (p0 < U < p0 + p1 ),
P (U ≤ p0 + p1 + p2 ) = p0 + p1 + p2 ⇒ p2 = P (p0 + p1 < U < p0 + p1 + p2 ),

k−1 k
and so on. Thus, if we set Y to be 0 if U ≤ p0 , and k if U satisfies pi , then Y has
P P
pi < U <
i=0 i=0
the same distribution as X. To use this idea to generate 50 observations from Poisson(5), we can
k
X
use the following code in R, noting that pi = P (X ≤ k ).
i=0

replicate(50,
{
U <- runif(1)
Y <- 0
while (U > ppois(Y, lambda = 5)) Y <- Y + 1
Y
})

[1] 5 3 3 5 5 3 4 4 4 6 6 8 10 2 10 9 4 5 5 9 2 9
[23] 2 6 4 5 7 3 4 6 8 3 4 2 4 6 4 2 3 0 4 5 5 5
[45] 4 9 7 6 3 4

Of course, there is nothing in this procedure that is specific to the Poisson distribution. By replacing
the call to ppois() suitably, the same process can be used to simulate random observations from
any discrete distribution supported on the non-negative integers. ■

Version: – November 19, 2024

 7.2 simulation 259

The process described in the previous example cannot be used for continuous random variables.
In such cases, Lemma 5.3.7 often proves useful. We illustrate how to generate samples from
Exponential distribution using Uniform.
Example 7.2.2. Consider the case where we want X to have the Exp(1) distribution. Then,
FX (x) = 1 − e−x for x > 0. Solving for FX (x) = u, we have

1 − e−x = u
⇒ e−x = 1−u
⇒x = − log(1 − u),

−1
that is, FX (u) = −log (1 − u). Thus, we can simulate 50 observations from the Exp(1) distribution
using the following R code.

-log(1 - runif(50))

[1] 0.233065598 0.280462639 0.154108577 1.372431939 1.821648686

[6] 0.138321757 0.436222604 0.542218451 1.380248674 0.766713622
[11] 2.144642933 0.301378922 1.269994489 0.333248095 1.149509378
[16] 0.117884391 2.562483778 0.153560104 1.522516377 0.596372217
[21] 1.185683422 0.601846501 0.937989125 1.488177881 0.414706054
[26] 0.199426507 0.820649856 1.291911171 0.678224144 0.161773024
[31] 0.356192943 0.491640264 1.168934205 0.695930930 0.340648718
[36] 0.232449735 0.698323026 0.061983953 1.152038946 0.608277765
[41] 0.252472685 0.662322314 0.152055387 1.440552940 1.382534336
[46] 0.004607301 0.554657905 0.989217878 1.185984115 0.254887938

This takes advantage of the ability of runif() to generate multiple values at once, and the fact
−1
that the expression for FX (u) can be easily vectorized. We can multiply the resulting observations
by 1/λ to simulate observations from the Exp(λ) distribution. ■
The approach illustrated in the last two examples has a disadvantage when the distribution
function F or its generalised inverse cannot be computed explicitly. This is the case when one
wishes to simulate samples from standard Normal distribution. We will discuss a few instances in
the Exercises next.

exercises

Ex. 7.2.1. Let U1 and U2 be i.i.d Normal random variables.

(a) Let Θ = 2πU1 . Find the distribution of Θ
(b) R = −2 log(U2 ). Find the distribution of R.
p

(c) Let X1 = R cos(Θ) and X2 = R sin(Θ). Show that X1 , X2 are i.i.d standard Normal random
variables.

Version: – November 19, 2024

260 sampling and descriptive statistics

(d) Write a R code to simulate 100 samples from standard Normal distribution from Uniform(0, 1).

Ex. 7.2.2. Let Z1 , Z2 be i.i.d. standard normal random variables. µ1 , µ2 ∈ R, σ1 , σ2 > 0 and
−1 < ρ < 1. Suppose X1 = σ1 Z1 + µ1 and X2 = σ2 (ρZ1 + 1 − ρ2 Z2 ) + µ2 .
p

(a) Find the joint distribution X1 , X2 Hint: Proof of Theorem 6.4.4

(b) Use the Exercise 7.2.1 (d) and write an R code to simulate 100 samples from a bivariate
Normal disstribution where the correlation ρ = 12 and marginals are standard Normal random
variables.

Ex. 7.2.3. Use the approach in Example 7.2.1 write an R code to simulate from Uniform(0,1), 100
samples of

(a) Binomial (10, 13 )

(b) Geometric( 14 )

Ex. 7.2.4. Let X1 , X2 , . . . , Xn be an i.i.d. sample from the Poisson(λ) distribution, and suppose
we are interested in estimating λ.

(a) Show that both the sample mean and the sample variance of X1 , X2 , . . . , Xn are unbiased
estimators of λ.

(b) Which of these estimators is better? To answer this question, simulate random observations
from the Poisson(λ) distribution for various values of λ using the R function rpois. Explore
the behaviour of the two estimates by varying λ as well as the sample size.

Ex. 7.2.5. Exercise 2.3.7 described the technique called “capture-recapture” which biologists use
to estimate the size of the population of a species when it cannot be directly counted. Suppose
the unknown population size is N , and fifty members of the species are selected and given an
identifying mark. Sometime later a sample of size twenty is taken from the population, and it is
found to contain X of the twenty previously marked. Equating the proportion of marked members
in the second sample and the population, we have 20X
N , giving an estimate of N̂ = X .
= 50 1000

Recall that X has a hypergeometric distribution that involves N as a parameter. It is not easy
to compute E [N̂ ] and V ar [N̂ ]. However, Hypergeometric random variables can be simulated in
R using the function rhyper. For each N = 50, 100, 200, 300, 400, and 500, use this function to
simulate 1000 values of N̂ and use them to estimate E [N̂ ] and V ar [N̂ ]. Plot these estimates as a
function of N .
Ex. 7.2.6. Suppose p is the unknown probability of an event A, and we estimate p by the sample
proportion p̂ based on an i.i.d. sample of size n.

(a) Write V ar [p̂] and SD [p̂] as functions of n and p.

(b) Using the relations derived above, determine the sample size n, as a function of p, that is
required to acheive SD (p̂) = 0.01. How does this required value of n vary with p?

(c) Design and implement the following simulation study to verify this behaviour. For p = 0.01,
0.1, 0.25, 0.5, 0.75, 0.9, and 0.99,

Version: – November 19, 2024

 7.3 plots 261

(i) Simulate 1000 values of p̂ with n = 500.

(ii) Simulate 1000 values of p̂ with n chosen according to the formula derived above.

In each case, you can think of the 1000 values as i.i.d. samples from the distribution of p̂,
and use the sample standard deviation as an estimate of SD [p̂]. Plot the estimated values of
SD (p̂) against p for both choices of n. Your plot should look similar to Figure 7.1.

0.0 0.2 0.4 0.6 0.8 1.0

Fixed sample size Variable sample size

0.015
[Link]

0.010

0.005

0.0 0.2 0.4 0.6 0.8 1.0

Figure 7.1: Estimated standard deviation in estimating a probability using sample proportion
as a function of the probability being estimated. See exercise 7.2.6.

7.3 plots

As we will see in later chapters, making more assumptions about the underlying distribution of X
allows us to give concrete answers to many important questions. This is indeed a standard and
effective approach to doing statistics, but in following that approach there is a danger of forgetting
that assumptions have been made, which we should guard against by doing our best to convince
ourselves beforehand that the assumptions we are making are reasonable.
Doing this is more of an art than a science, and usually takes the form of staring at plots
obtained from the sample observations, with the hope of answering the question: “does this plot
look like what I would have expected it to look like had my assumptions been valid?” Remember
that the sample X1 , X2 , . . . , Xn is a random sample, so any plot derived from it is also a “random
plot”. Unlike simple quantities such as sample mean and sample variance, it is not clear what to
“expect” such plots to look like, and the only way to really hone our instincts to spot anomalies is
through experience. In this section, we introduce some commonly used plots and use simulated
data to give examples of how such plots might look like when the usual assumptions we make are
valid or invalid.

Version: – November 19, 2024

262 sampling and descriptive statistics

0.06
Proportion

0.04

0.02

0.00

20 30 40 50

Value

Figure 7.2: Empirical frequency distribution of 10000 random samples from the Poisson(30)
distribution.

7.3.1 Empirical Distribution Plot for Discrete Distributions

The typical assumption made about a random sample is that the underlying random variable
belongs to a family of distributions rather than a very specific one. For example, we may assume
that the random variable has a Poisson(λ) distribution for some λ > 0, without placing any further
restriction on λ, or a Binomial(n, p) distribution for some 0 < p < 1. Such families are known as
parametric families.
When the data X1 , X2 , . . . , Xn are from a discrete distribution, the simplest representation of
the data is its empirical distribution, which is essentially a table of the frequencies of each value
that appeared. For example, if we simulate 1000 samples from a Poisson distribution with mean 3,
its frequency table may look like

x <- rpois(1000, lambda = 3)

table(x)

x
0 1 2 3 4 5 6 7 8
49 168 238 215 155 99 48 18 10

[Link](table(x))

Version: – November 19, 2024

 7.3 plots 263

x
0 1 2 3 4 5 6 7 8
0.049 0.168 0.238 0.215 0.155 0.099 0.048 0.018 0.010

The simplest graphical representation of such a table is through a plot similar to Figure 7.2, which
represents a larger Poisson sample with mean 30, resulting in many more distinct values. Although
in theory all non-negative integers have positive probability of occurring, the probabilites are too
small to be relevant beyond a certain range. This plot does not have a standard name, although
it may be considered a variant of the Cleveland Dot Plot. We will refer to it as the Empirical
Distribution Plot from now on.
We can make similar plots for samples from Binomial or any other distribution. Unfortunately,
looking at this plot does not necessarily tell us whether the underlying distribution is Poisson, in
part because the shape of the Poisson distribution varies with the λ parameter. A little later, We
will discuss a modification of the empirical distribution plot, known as a rootogram, that helps
make this kind of comparison a little easier.

7.3.2 Histograms for Continuous Distributions

In the case of continuous distributions, we similarly want to make assumptions about a random
sample being from a parametric family of distributions. For example, we may assume that the
random variable has a Normal(µ, σ 2 ) distribution without placing any further restriction on the
parameters µ or σ 2 (except of course that σ 2 > 0), or that it has an Exponential(λ) distribution
with any value of the parameter λ > 0. Such families, as noted earlier, are known as parametric
families. For both these examples, the shape of the distribution does not depend on the parameters,
and this makes various diagnostic plots more useful.
The empirical distribution plot above is not useful for data from a continuous distribution,
because by the very nature of continuous distributions, all the data points will be distinct with
probability 1, and the value of the empirical distribution function will be exactly 1/n at these
points.
The plot that is most commonly used instead to study distributions is the histogram. It
is similar to the empirical distribution plot, except that it does not retain all the information
contained in the empirical distribution. Instead, it divides the range of the data into arbitrary bins
and counts the frequencies of data points falling into each bin, effectively discretizing the data.
More precisely, the histogram estimates the probability density function of the underlying random
variable by estimating the density in each bin as a quantity such that the probability of each bin
is proportional to the number of observations in that bin. By choosing the bins judiciously, for
example by having more of them as sample size increases, the histogram strikes a balance that
ensures that the histogram “converges” to the true underlying density as n → ∞.
Figure 7.3 gives examples of histograms where data are simulated from the normal and
exponential distributions for varying sample sizes. Five replications are shown for each sample size.
We can see that for large sample sizes, the shape of the histograms are recognizably similar to the
shapes of the corresponding theoretical distributions seen in Figure 5.1 and Figure 5.2 in Chapter 5.

Version: – November 19, 2024

264 sampling and descriptive statistics

−2 −1 0 1 2 −2 −1 0 1 2

20 50 100 500 1000

Density

−2 −1 0 1 2 −2 −1 0 1 2 −2 −1 0 1 2

0 1 2 3 4 0 1 2 3 4

20 50 100 500 1000

Density

0 1 2 3 4 0 1 2 3 4 0 1 2 3 4

Figure 7.3: Histograms of random samples from the Normal(0, 1) (top) and Exponential(1)
(bottom) distributions. Columns represent increasing sample sizes, and rows are
independent repetitions of the experiment.

Version: – November 19, 2024

 7.3 plots 265

Moreover, the shape is consistent over the five replications. This is not true, however, for small
sample sizes. Remember that the histograms are based on the observed data, and are therefore
random objects themselves. As we saw with numerical properties like the mean, estimates have
higher variability when the sample size is small, and get less variable as sample size increases. The
same holds for graphical estimates, although making this statement precise is more difficult.
There are several ways to create histograms in R, which we will not go into here, but one
approach is explored in the exercises.

7.3.3 Hanging Rootograms for Comparing with Theoretical Distributions

Graphical displays of data are almost always used for some kind of comparison. Sometimes these
are implicit comparisons, asking, say, “how many peaks does a density have”, or “is it symmetric?”
More often, they are used to compare samples from two subpopulations, say, the distribution of
height in males and females. Sometimes, as discussed above, they are used to compare an observed
sample to a hypothesized distribution.
In the case of the empirical distribution plot, a simple modification is to add the probability
mass function of the theoretical distribution. This, although a reasonable modification, is not
optimal. Research into human perception of graphical displays indicates that the human eye is more
adept at detecting departures from straight lines than from curves. Taking this insight into account,
John Tukey suggested “hanging” the vertical lines in an empirical distribution plot (which are after
all nothing but sample proportions) from their expected values under the hypothesized distribution.
He further suggested a transformation of what is plotted: instead of the sample proportions and
the correponding expected probabilities, he suggested plotting their square roots, thus leading to
the name hanging rootogram for the resulting plot. The reason for making this transformation is as
follows. Recall that for a proportion p̂ obtained from a sample of size n,

p(1 − p) p
V ar [p̂] = ≈
n n
provided p is close to 0. In Chapter 8, we will encounter the Central Limit Theorem and the Delta
Method, which can be used to show (see Example 8.5.4) that as the sample size n grows large,
√ √ √
V ar [ p̂] ≈ c/n for a constant c. This means that unlike p̂ − p, the variance of p̂ − p will be
approximately independent of p.
Figure 7.4 gives examples of hanging rootograms. These examples have been created using the
rootogram() function in the latticeExtra package. The following R code is an example of its
use.

library(package = "latticeExtra")
xbin30 <- rbinom(10000, 100, 0.3)
rootogram(˜ xbin30, dfun = function(x) dpois(x, lambda = 30), grid = TRUE)

This requires the latticeExtra package to be already installed on your system, which it most
likely will not be. To install it, type

Version: – November 19, 2024

266 sampling and descriptive statistics

[Link]("latticeExtra")

and follow instructions.

Samples from Poisson(30)

0.25

0.20
P(X = x)

0.15

0.10

0.05

0.00

20 30 40 50

Value

Samples from Binomial(100, 0.3)

0.25

0.20
P(X = x)

0.15

0.10

0.05

0.00

20 30 40 50

Value

Figure 7.4: Hanging rootogram of 10000 random samples compared with the Poisson(30)
distribution. In the top plot, the samples are also from Poisson(30), whereas in
the bottom plot the samples are from the Binomial(100, 0.3) distribution, which
has the same mean but different variance. Note the similarities with Figure 2.2

Version: – November 19, 2024

 7.3 plots 267

7.3.4 Q-Q Plots for Continuous Distributions

1.0

0.8
Empirical CDF

0.6

0.4

0.2

0.0

0 1 2

Value
Sorted data values

0.0 0.2 0.4 0.6 0.8 1.0

Quantiles of U(0,1)

Figure 7.5: Conventional ECDF plot (top) and its “inverted” version (bottom), with x- and
y-axes switched, and points instead of lines.

Just as histograms were binned versions of the empirical distribution plot, we can plot binned
versions of hanging rootograms for data from a continuous distribution as well. It is more common

Version: – November 19, 2024

268 sampling and descriptive statistics

however, to look at quantile-quantile plots (QQ plots), which do not bin the data, but instead plot
what is essentially a transformation of the empirical CDF.
Recall from Definition 7.1.1 that the ECDF of observations X1 , X2 , . . . , Xn is given by

#{Xi ≤ t}
F̂n (t) = P (Y ≤ t) =
n
The top plot in Figure 7.5 is a conventional ECDF plot of 200 observations simulated from a
Normal(1, 0.52 ) distribution. The bottom plot has the sorted data values on the y-axis and 200
equally spaced numbers from 0 to 1. A little thought tells us that this plot is essentially the same
as the ECDF plot, with the x- and y-axes switched, and using points instead of lines. Naturally, we
expect that for reasonably large sample sizes, the ECDF plot obtained from a random sample will
be close to the true cumulative distribution function of the underlying distribution. If we know the
shape of the distribution we expect the data to be from, we can compare it with the shape seen in
the plot.
Although this is a fine idea in principle, it is difficult in practice to detect small differences
between the observed shape and the theorized or expected shape. Here, we are helped again by the
insight that the human eye finds it easier to detect deviations from a straight line than from curves.
By keeping the sorted data values unchanged, but transforming the equally spaced probability
values to the corresponding quantile values of the theorized distribution, we obtain a plot that we
expect to be linear. We define them formally below.

Definition 7.3.1. Let F be a distribution function. For 0 < p < 1, the p-th quantile of F
is defined as
qp = inf{x ∈ R : F (x) ≥ p}.

Note that qp = F −1 (p) when F −1 exists and F (qp −) ≤ p ≤ F (qp ). q 1 is referred to as the
2
median of F . For a sample X1 , X2 , . . . , Xn from distribution F, the sample p-th quantile is
defined as the p-th quantile of the empirical distribution function Fn .

Quantiles may be thought of informally as follows, generalizing the definition of median given in
Exercise 5.2.10: For a given CDF F , the quantile corresponding to a probability value p ∈ [0, 1] is a
value x such that F (x) = p. Such an x may not exist for all p and F , or it may not be unique, and
a formal definition of quantiles needs to be modified to take this into [Link], for most
standard continuous distributions used in Q-Q plots, one may work with this informal notion. Such
a plot with Normal (0, 1) quantiles is shown in Figure 7.6 for simulated normal and exponential
random samples. More examples are explored in the exercises.

exercises

Ex. 7.3.1. The R functions histogram() and qqmath() in the lattice package can be used to
generate histograms and Q-Q plots respectively (although there are other alternatives as well).

Version: – November 19, 2024

 7.3 plots 269

−2 0 2 −2 0 2

20 50 100 500 1000

8
6
Exponential

4
2
Sorted data values

0
−2
−4
8
6
4
Normal

2
0
−2
−4
−2 0 2 −2 0 2 −2 0 2

Quantiles of N(0,1)

Figure 7.6: Normal Q-Q plots of data generated from Normal and Exponential distributions,
with varying sample size. The Q-Q plots are more or less linear for Normal data,
but exhibit curvature indicative of a relatively heavy right tail for exponential data.
Not surprisingly, the difference becomes easier to see as the sample size increases.

Version: – November 19, 2024

270 sampling and descriptive statistics

This exercise guides you through the process of simulating data from a sampling distribution and
creating corresponding histograms and Normal Q-Q plots.

(a) Suppose Z1 , Z2 , . . . , Zn are independent Normal(0, 1). Then the distribution of the mean of
Z1 , Z2 , . . . , Zn is Normal(0, 1/n). To verify this, simulate such means for n = 50 using the
following R code.

[Link] <- replicate(1000, mean(rnorm(50)))

(b) Create a histogram of the simulated values using

library(package = "lattice") # needed only once, to load the package

histogram(˜ [Link], nint = 15) # nint (optional) gives the number of bins

(c) Create a Normal Q-Q plot of the same values using

qqmath(˜ [Link], grid = TRUE)

(d) Study the behaviour of these plots over multiple repetitions, as well as by varying n and the
number of replications.

Ex. 7.3.2. If Z1 , Z2 , . . . , Zn are independent Normal(0, 1), what can you say about the distribution
of the median of Z1 , Z2 , . . . , Zn ? Use the median() function, using it to replace the call to mean()
in the previous exercise, to simulate observations from this distribution. Use histograms and Normal
Q-Q plots to study this distribution and compare it to the distribution of the mean. In particular, is
the distribution of the median also Normal? Does it have lower or higher variance than the mean?
Ex. 7.3.3. Repeat the previous exercise, replacing the median by the minimum and maximum of n
obsrvations Z1 , Z2 , . . . , Zn that are independent Normal(0, 1). What are the distingushing features
of these histograms and Normal Q-Q plots?

Version: – November 19, 2024

SAMPLING DISTRIBUTIONS AND LIMIT THEOREMS
8
For n ≥ 1, let X1 , X2 , . . . , Xn be an i.i.d. random sample from a population. Recall the sample
mean
n
1X
X= Xi
n
i=1

and sample variance

n
1 X
S2 = (Xi − X )2 .
n−1
i=1

We have seen the use of these sample statistics in the previous chapter. In this chapter, we will
discuss the distributional properties and limiting behaviour of such statistics. In Chapters 9 and
10, we will discuss how these properties can be effectively used to estimate parameters related to
the underlying population and verify specific hypotheses about them. The corresponding fields of
study are called Estimation and Hypothesis Testing.

We will spend most of our time in finding the distribution of the sample mean and the
sample variance given the distribution of X1 . One immediately observes that these are somewhat
complicated functions of independent random variables. However in Section 3.3 and Section 5.5 we
have seen examples of functions for which we were able to explicitly compute the distribution. To
understand sampling statistics we must also understand the notion of joint distribution of more
than two continuous random variables (See Section 3.3 for discrete random variables).

8.1 multi-dimensional continuous random variables

In Chapter 3, while discussing discrete random variables, we had considered a finite collection
of random variables (X1 , X2 , . . . , Xn ). In Definition 3.2.7, we had described how to define their
joint distribution and we used this to understand the multinomial distribution in Example 3.2.12.
There are many instances in the continuous setting as well where it is relevant to study the joint
distribution of a finite collection of random variables. Suppose X is a point chosen randomly inside
the unit sphere in three dimensions. Then X has three coordinates, say X = (X1 , X2 , X3 ), where
each
q Xi is a random variable in (0, 1). These coordinates are dependent because we know that
X12 + X22 + X32 ≤ 1. To reason about the properties of X, it is useful and necessary to understand
the “joint distribution” of (X1 , X2 , X3 ). Similarly, to understand the distribution of the sample
mean and the sample variance, which are functions of X1 , X2 , . . . , Xn , we first need to understand
the joint distribution of (X1 , X2 , . . . , Xn ). We begin by defining the joint distribution function.

271

Version: – November 19, 2024

272 sampling distributions and limit theorems

Definition 8.1.1. For n ≥ 1, let X1 , X2 , . . . , Xn be random variables defined on the same

probability space. The joint distribution function F : Rn → [0, 1] of X1 , X2 , . . . , Xn is given
by
F (x1 , x2 , . . . , xn ) = P (X1 ≤ x1 , X2 ≤ x2 , . . . , Xn ≤ xn ), (8.1.1)

for x1 , x2 , . . . , xn ∈ R.

As in one-variable and two-variable situations, the joint distribution function determines the entire
joint distribution of (X1 , X2 , . . . , Xn ) for discrete random variables. More precisely, if all the
random variables were discrete with Xi : S → Ti , where Ti are countable subsets of R for 1 ≤ i ≤ n,
the from the joint distribution function one can determine

P (X1 = t1 , X2 = t2 , . . . , Xn = tn )

for all ti ∈ Ti , 1 ≤ i ≤ n (See Exercise 8.1.1). The joint distribution function determines the joint
distribution in the continuous setting as well, but we need to introduce some notation before we
can state this result rigorously.
For n ≥ 1, let f : Rn → R be a non-negative function that is piecewise-continuous in each
variable, and for which Z
f (x1 , x2 , . . . , xn ) dx1 dx2 . . . dxn = 1.
Rn

If for every Borel set A ⊂ Rn we have

Z
P (A) = f (x1 , x2 , . . . , xn ) dx1 dx2 . . . dxn ,
A

then one can show as in Theorem 5.1.5 that P is a probability on Rn . In this case, f is called the
density function for P .
Density functions arise naturally from certain types of random variables. A collection of random
variables (X1 , X2 , . . . , Xn ) is said to have a joint density f : Rn → R if for every event A ⊂ Rn ,
Z
P ((X1 , X2 , . . . , Xn ) ∈ A) = f (x1 , x2 , . . . , xn ) dx1 dx2 . . . dxn .
A

In this setting, the joint distribution of (X1 , X2 , . . . , Xn ) is determined by their joint density f .
Using multivariable calculus we can can state and prove a result similar to Theorem 5.2.5 for
random variables (X1 , X2 , . . . , Xn ) that have a joint density. In particular, we can conclude that
as the joint densities are assumed to be piecewise continuous in each variable, the corresponding
distribution functions are piecewise differentiable in each variable. Further, the joint distribution
of the continuous random variables (X1 , X2 , . . . , Xn ) are completely determined by their joint
distribution function F . That is, if we know F (x1 , x2 , . . . , xn ) for all x1 , x2 , . . . , xn ∈ R we could
use multivariable calculus to differentiate F to find f . Integrating this joint density over the event
A, we can then calculate P ((X1 , X2 , . . . , Xn ) ∈ A).

Version: – November 19, 2024

 8.1 multi-dimensional continuous random variables 273

As in the n = 2 case, one can recover the marginal density of each Xi for i between 1 and n by
integrating over the other indices. So, the marginal density of Xi at a is given by
Z
f Xi ( a ) = f (x1 , . . . , xi−1 , a, xi+1 , . . . , xn ) dx1 . . . dxi−1 dxi+1 . . . dxn .
Rn−1

Further, for n ≥ 3, we can deduce the joint density for any sub-collection m ≤ n random variables
by integrating over the other variables. For instance, if we were interested in the joint density of
(X1 , X3 , X7 ), we would obtain
Z
fX1 ,X3 ,X7 (a1 , a3 , a7 ) = f (a1 , x2 , a3 , x4 , x5 , x6 , a7 , x8 . . . , xn ) dx2 dx4 dx5 dx6 . . . dxn .
Rn−3

Suppose X1 , X2 , . . . , Xn are random variables defined on a single sample space S with joint density
f : Rn → R. Let g : Rn → R be a function of n variables for which g (X1 , X2 , . . . , Xn ) is defined
on the range of the Xj variables. Let B be an event in the range of g. Then, following the proof of
Theorem 3.3.5, we can show that

P (g (X1 , X2 , . . . , Xn ) ∈ B ) = P (X1 , X2 , . . . , Xn ) ∈ g −1 (B ) .

Although the above provides an abstract method of finding the distribution of the random variable
Y = g (X1 , X2 , . . . , Xn ), it can be difficult to use for explicit calculations. For n = 1 we discussed
this question in detail in Section 5.3, and for n = 2 we explored how to find the distributions of
sums and ratios of independent random variables (see Section 5.5). This method could be extended
by induction on n in a few cases, but in general this is not possible. In Appendix B, Section A.1.1,
we discuss a more general Jacobian-based method of finding the joint density of functions of random
variable.
The notion of independence, introduced in the discrete setting, also extends to multi-dimensional
continuous random variables. As discussed in Definition 3.2.3, a finite collection of continuous
random variables X1 , X2 , . . . , Xn is mutually independent if the sets (Xj ∈ Aj ) are mutually
independent for all events Aj in the ranges of the corresponding Xj . As proved for the n = 2
case in Theorem 5.4.7, we can similarly deduce that if (X1 , X2 , . . . , Xn ) are mutually independent
continuous random variables with marginal densities fXi then their joint density is given by

n
Y
f (x1 , x2 , . . . , xn ) = fXi (xi ), (8.1.2)
i=1

for xi ∈ R and 1 ≤ i ≤ n. Further, for any finite sub-collection (Xi1 , Xi2 , . . . , Xim ) of the above
independent random variables, the joint density is given by
m
Y
f (a1 , a2 , . . . , am ) = fXi (aj ). (8.1.3)
j
j =1

We conclude this section with a result that we will repeatedly use.

Version: – November 19, 2024

274 sampling distributions and limit theorems

Theorem 8.1.2. Fix n ≥ 1. For each j ∈ {1, 2, . . . , n}, let i ∈ {1, 2, . . . , mj } for some
positive integer mj . Suppose Xi,j is an array of mutually independent continuous random
variables. Define Yj = gj (X1,j , X2,j , . . . Xmj ,j ), where gj : Rmj → R are continuous
functions. Then the resulting variables Y1 , Y2 , . . . , Yn are mutually independent.

Proof. Follows by the same proof presented in Theorem 3.3.6. ■

8.1.1 Order Statistics and their Distributions

For n ≥ 1, let X1 , X2 , . . . , Xn be an i.i.d. random sample from a population with common distribu-
tion function F . Arrange them in increasing order of magnitude, with the ordered observations
denoted by
X(1) ≤ X(2) ≤ · · · ≤ X(n) .

These ordered values are called the order statistics of the sample X1 , X2 , . . . , Xn . For, 1 ≤ r ≤ n,
X(r ) is called the r-th order statistic. The median of X1 , X2 , . . . , Xn is defined as X( n+1 ) when n
2
is odd and X( n ) when n is even.
2
One can compute F(r ) , the distribution function of X(r ) , for 1 ≤ r ≤ n in terms of n and F .
We have,
 
  n
F(1) ( x ) = P ( X(1) ≤ x ) = 1 − P X(1) > x = 1 − P ∩ (Xi > x)
i=1
n
Y n
Y
= 1− P (Xi > x) = 1 − (1 − P (Xi ≤ x))
i=1 i=1
= 1 − (1 − F (x))n ,
  n
n Y
F(n) ( x ) = P ( X(n) ≤ x ) = P ∩ (Xi ≤ x) = P (Xi ≤ x) = (F (x))n ,
i=1
i=1

and for 1 < r < n,

F(r ) ( x ) = P (X(r ) ≤ x) = P (at least r elements from the sample are ≤ x)

n
X
= P (exactly j elements from the sample are ≤ x)
j =r
n  
X n
= P (chosen j elements from the sample are ≤ x) ×
j
j =r
P ((n − j ) elements not chosen from the sample are > x)
n  
X n
= F (x)j (1 − F (x))n−j
j
j =r

Version: – November 19, 2024

 8.1 multi-dimensional continuous random variables 275

If the distribution function F had a probability density function f then each X(r ) has a probability
density function f(r ) . This can be obtained by differentiating F(r ) and is given by the following
expression.



n(1 − F (x))n−1 f (x) r=1






f(r ) (x) = nf (x)(F (x))n−1 r=n (8.1.4)







n!
f (x)(F (x))r−1 (1 − F (x))n−r 1 < r < n


(r−1)!(n−r )!

Example 8.1.3. Let n ≥ 1 and let X1 , X2 , . . . , Xn be a i.i.d. random sample from a population
whose common distribution F is an Exponential (λ) random variable. Then we know that

0 x<0
F (x) =
1 − e−λx x ≥ 0.

Therefore using (8.1.4) and substituting for F as above we have that the densities of the order
statistics are given by

r=1


 n(e−λx ))n−1 λe−λx





f(r ) ( x ) = nλe−λx (1 − e−λx )n−1 r=n





 λe−λx n!
(1 − e−λx )r−1 (e−λx )n−r 1 < r < n,

(r−1)!(n−r )!

for x > 0. Simplifying the algebra we obtain,

r=1


 nλe−nλx





f(r ) ( x ) = nλe−λx (1 − e−λx )n−1 r=n





λn!
(1 − e−λx )r−1 (e−λx )n−r +1 1 < r < n,


(r−1)!(n−r )!

for x > 0. We note from the above that X(1) , i.e minimum of exponentials, is Exponential (nλ)
random variable. However the other order statistics are not exponentially distributed. ■

In many applications, one is interested in the range of values a random variable X assumes.
A method to understand this to sample X1 , X2 , . . . , Xn i.i.d. X and examine R = X(n) − X(1) .
Suppose X has a probability density function f : R → R and distribution function F : R → [0, 1].
As before we can can calculate the joint density of X(1) , X(n) by first computing the joint distribution

Version: – November 19, 2024

276 sampling distributions and limit theorems

function. This is done by using the i.i.d. nature of the sample and the definition of the order
statistics.

P (X(1) ≤ x, X(n) ≤ y ) = P ( X(n) ≤ y ) − P ( x < X(1) , X(n) ≤ y )

   
n n
= P ∩ {Xi ≤ y} − P ∩ {x < Xi ≤ y}
i=1 i=1
= [P (X ≤ y )] − [P (x < X ≤ y )]n
n

[F (x)]n − [F (y ) − F (x)]n x < y
=
0 otherwise.

From the above, differentiating partially in x and y we see that the joint density of (X(1) , X(n) ) is
given by 
n(f (x) − f (y ))[F (y ) − F (x)]n−1 x < y
f ( x, y ) = (8.1.5)
X(1) ,X(n) 0 otherwise.

To calculate the distribution of R, we compute its distribution function. For r ≤ 0, P (R ≤ r ) = 0

and for r > 0, using the above joint density of (X(1) , X(n) ) we have

P (R ≤ r ) = P ( X(n) ≤ X(1) + r )
Z∞ Zr
 

= X(1) ,X(n) (x, z + x)dz dx

 f 
−∞ 0
Zr Z∞
 

=  fX
(1) ,X(n)
(x, z + x) dx dz,
0 −∞

where we have done a change of variable y = z + x in the second last line and a change in the order
of integration in the last line. Differentiating the above we conclude that R has a joint density
given by  ∞
(x, r + x) dx if r > 0
R
f


−∞ X(1) ,X(n)


fR (r ) = (8.1.6)



0

otherwise.

Example 8.1.4. Let X1 , X2 , . . . , Xn be i.i.d. Uniform(0, 1). The probability density function and
distribution function of a Uniform(0, 1) random variable are given by

0 if x ≤ 0
 
1 if x ∈ (0, 1)

f (x) = and F (x) = x if 0 < x < 1
0 otherwise. 
1 if x > 1.



Version: – November 19, 2024

 8.1 multi-dimensional continuous random variables 277

Let fX be the probability density function of X(r ) for 1 ≤ r ≤ n. Then, using (8.1.4), we have
(r )


n(1 − x)n−1 if x ∈ (0, 1)
fX ( x ) =
(1) 0 otherwise,

nxn−1 if x ∈ (0, 1)
fX (x) =
(n) 0 otherwise, and

 n!
(r−1)!(n−r )!
xr−1 (1 − x)n−r if x ∈ (0, 1)
for 1 < r < n, fX ( x ) =
(r ) 0 otherwise.

Using (8.1.5), the joint density of (X(1) , X(n) ) is given by


n(n − 1)(y − x)n−1 if 0 ≤ x ≤ y ≤ 1
fX
(1) ,X(n)
(x, y ) =
0 otherwise.

Using (8.1.6), the probability density function of the range R = X(n) − X(1) is given by
 1−r
n(n − 1)(x + r − x)n−1 dx if 0 < r < 1
 R

fR (r ) = 0
0

otherwise,

n(n − 1)rn−1 (1 − r ) if 0 < r < 1
=
0 otherwise.

It is easy to see by comparing density functions that X(r ) ∼ Beta(r, n − r + 1) for 1 ≤ r ≤ n, and
the range R ∼ Beta(n, 2). ■

In general, we may also be interested in the joint distribution of the order statistics. Suppose we
have an i.i.d. sample X1 , X2 , . . . , Xn having distribution X. If X has a probability density function
f : R → R then one can show that the order statistic (X(1) , X(2) , . . . , X(n) ) has a joint density
h : Rn → R given by

n!f (u )f (u ) . . . f (u )
1 2 n u1 < u2 < . . . < un ,
h(u1 , u2 , . . . , un ) =
0 otherwise.

The above fact should be intuitively clear: Any ordering u1 < u2 < . . . < un has “probability”
f (u1 )f (u2 ) . . . f (un ). Each Xi can assume any of the uk ’s. The total number of possible orderings
is n!. A formal proof involves using the Jacobian method and will be discussed in Appendix B.

Version: – November 19, 2024

278 sampling distributions and limit theorems

8.1.2 χ2 , F and t

χ2 (pronounced Chi-Square), F and t distributions arise naturally when considering functions of

i.i.d. normal random variables (X1 , X2 , . . . , Xn ) for n ≥ 1. They are useful in estimation and
hypothesis testing, which we will study in subsequent chapters. We discuss these distributions via
three examples.

The χ2 , F and t distributions arise as functions of Normal random variables. As we will

see, they are essentially special cases of distributions we have already encountered, but they are
studied separately because they come up naturally when considering the distribution of sample
variances obtained from collections of i.i.d. normal random variables. In this section, we discuss
these distributions via three examples, before discussing their connection to the sample variance in
Section 8.1.3.

Example 8.1.5. For n ≥ 1, let (X1 , X2 , . . . , Xn ) be a collection of independent Normal random

variables with mean 0 and variance 1. Then their joint density is given by

P n
n − 12 x2i
Y 1
f ( x1 , x2 , . . . , xn ) = fXi (xi ) = √ e i = 1 ,
i=1
( 2π )n

n
for xi ∈ R and 1 ≤ i ≤ n. We are interested in the distribution of Z = Xi2 .
P
i=1

We shall find this distribution in two steps. Clearly, the range of X12 is non-negative. The
distribution function for X12 at z ≥ 0 is given by

F1 ( z ) = P (X12 ≤ z )
√
= P (X1 ≤ z )
√
Zz
1 x2
= √ e− 2 dx
2π
0
Zz
1 u 1
= √ e− 2 u− 2 du
2 2π
0

Comparing it with the Gamma (α, λ) random variable defined in Definition 5.5.5 and using Exercise
5.5.9, we see that X12 is distributed as a Gamma ( 12 , 12 ) random variable. From the calculation done
n
in Example 5.5.6 for n = 2, it follows by using induction that Z = Xi2 has the Gamma n2 , 12
P

i=1
distribution. This distribution is referred to as χ2 with n degrees of freedom. We define it precisely
next. ■

Version: – November 19, 2024

 8.1 multi-dimensional continuous random variables 279

Definition 8.1.6. (χ2n (i.e. chi-square with n degrees of freedom)) A random variable X
whose distribution is Gamma n2 , 12 is said to have the chi-square distribution with n degrees

of freedom, denoted X ∼ χ2n . The density of X is given by

 n
2− 2 n x

(n
x 2 −1 e− 2 when n is even
−n 2 −1) !

2

2 n x

f (x) = 2 −1 e− 2 =
n x
Γ( 2 ) n
2n− 2 −1 ( n−1

 )! n −1 − x
√2 x2 e 2 when n is odd


(n−1)! π

for x > 0.

We show in Section 8.1.3 that the sample variance obtained from a Normal sample follows a
(scaled) χ2 random variable. The F distribution arises as the ratio of the sample variances of two
independent Normal samples, or in other words, as the ratio of two independent (scaled) χ2 random
variables, as we see in the next example.

Example 8.1.7. (F distribution) Let X1 , X2 , . . . , Xn1 be an i.i.d. random sample from the
Normal 0, σ12 population, and Y1 , Y2 , . . . , Yn2 be an independent i.i.d. random sample from


n1  2
Xi
a Normal 0, σ22 population. It follows from Example 8.1.5 that U = has the χ2n1

P
σ1
i=1
n 2  2
Yi
distribution, and V = has the χ2n2 distribution. Further, by Theorem 8.1.2 U and V are
P
σ2
i=1
n2
independent because the Xi and Yj random variables are independent. Let Z = U V
n1 / n2 = U
V · n1 .
It follows from Example 5.5.10 that the density of W = VU for w > 0 is given by
n1
w 2 −1 Γ( n1 +
2
n2
)
fW ( w ) = n1 n2
(1 + w )
n1 + n2
2 Γ( 2 )Γ( 2 )

Therefore, for z > 0,

n1
z
  Zn2
n1
FZ ( z ) = P ( Z ≤ z ) = P W ≤ z = fW (w )dw
n2
−∞

Therefore the density of Z, for z > 0 is given by

 n1 n1
2 −1 Γ( n1 + n2
)

n1 n1 n1 2 z
f (z ) = fW ( z ) = 2
.
n2 n2 n2 (1 + n1
n1 +n2
Γ( 2 )Γ( n22 )
n1
n2 z )
2

Z is said to have the F distribution with degrees freedom parameters n1 and n2 , denoted Z ∼
Fn1 ,n2 . ■

Remark 8.1.8. In the previous example, the F distribution essentially arises as the ratio U /V
where U , V are independent χ2 random variables. As the χ2 distribution is a special case of the
Gamma distribution, it follows by Example 5.5.11 that U +U
V is distributed as a Beta random

Version: – November 19, 2024

280 sampling distributions and limit theorems

variable. Further, U +
U = 1 + U , so the two distributions are simple transformations of each other.
V V

In that sense, the F distribution is not a new distribution either, and it is studied separately mainly
for its natural definition as the ratio of sample variances.
The distribution of the ratio of sample mean and sample variance plays an important role in
estimation and hypothesis testing. This forms the motivation for the next example where the t
distribution arises naturally.
Example 8.1.9. (t distribution) Let X1 be a Normal (0, 1) random variable, and let X2 be an
independent χ2n random variable. We wish to find the density of Z, where

X1
Z= √ .
X2 /n

X2
Observe that U = Z 2 is given by X2 /n
1
. Now, X12 has χ21 distribution (see Example 8.1.5), so
applying Example 8.1.7 with n1 = 1 and n2 = n, we find that U has F1,n distribution. The density
of U is given by
 1 1
1 2 u 2 −1 Γ ( n+ 1
2 )
fU (u) =
(1 + n1 u) 2 Γ( 2 )Γ( 2 )
n+1 1 n
n
1
Γ ( n+1 ) u− 2
= √ 2 n .
nπΓ( 2 ) (1 + u ) n+2 1
n
√
As X1 is a symmetric random variable and X2 /n is positive valued, we conclude that Z is a
symmetric random variable (Exercise 8.1.11). So, for u > 0,

P (U ≤ u) = P (Z 2 ≤ u)
√ √
= P (− u ≤ Z ≤ u)
√ √
= P (Z ≤ u) − P (Z ≤ − u)
√ √
= P (Z ≤ u) − P (Z ≥ u)
√
= 2P (Z ≤ u) − 1

Therefore, if fZ (·) is the density of Z then

1 √
fU (u) = √ (fZ ( u)).
u

Hence for any z ∈ R the density of Z is given by

fZ ( z ) = |z| fU (z 2 )
1
−
Γ ( n+1 ) z2 2
= |z| √ 2 n
nπΓ( 2 ) 1 + u
n+2 1
n
− n+2 1
Γ( 2 )
n+1 2

z
= √ 1+ .
nπΓ( n2 ) n

Version: – November 19, 2024

 8.1 multi-dimensional continuous random variables 281

Z is said to have the t distribution with n degrees of freedom, denoted Z ∼ tn . ■

8.1.3 Distribution of Sampling Statistics from a Normal Population

For n ≥ 2, let X1 , X2 , . . . , Xn be an i.i.d. random sample from an arbitrary population having

mean µ and variance σ 2 . Consider the sample mean
n
1X
X= Xi
n
i=1

and the sample variance

n
1 X
S2 = (Xi − X )2 .
n−1
i=1

We have already seen in Theorem 7.1.4 that E [X ] = µ and in Theorem 7.1.6 that E [S 2 ] = σ 2 . It
is unreasonable to expect that we would be able to precisely describe the distribution of X or S 2
unless the distribution of the population is known. It turns out that even in that case, it is not easy
to derive these distributions in general. However, when the population is Normal, we can obtain
the joint distribution of X and S 2 completely. The main result of this section is the following.

Theorem 8.1.10. For n ≥ 2, let X1 , X2 , . . . , Xn be an i.i.d. random sample with distribution

X ∼ Normal(µ, σ 2 ). Let X and S 2 be defined as above. Then,
σ2
(a) X is a Normal random variable with mean µ and variance n .
2
(b) (n − 1) Sσ2 has the χ2n−1 distribution.

(c) X and S 2 are independent.

Proof. (a) follows from Theorem 6.3.13. There are several proofs for (b) and (c), with the most
common ones requiring some knowledge of Linear Algebra (e.g., see [Rao73]). Here we will follow
Kruskal’s proof as illustrated in [Stig84]. The proof is by the method of induction on the sample
size n. To implement the inductive step, we shall replace X and S 2 with X n and Sn2 for the rest of
the proof. This notation also emphasizes that the distributions of X n and Sn2 depend on n, and that
as functions defined on the underlying sample space, they are in fact different random variables.
Step 1: (Proof for n = 2) Here

X1 + X2 2 X1 + X2 2 (X1 − X2 )2
   
X1 + X2
X2 = and S22 = X1 − + X2 − = . (8.1.7)
2 2 2 2

As X1 and X2 are independent Normal random variables with mean µ and variance σ 2 , by Theorem
(X −X )
6.3.13, 1 √ 2 is a Normal random variable with mean 0 and variance 1. Using Example 8.1.5,
σ 2
S2
we know that σ22 has χ21 distribution and this proves (b).
From (8.1.7), X 2 is a function of X1 + X2 and S22 is a function of X1 − X2 . Theorem 8.1.2
will imply that X 2 and S22 are independent if we show X1 + X2 and X1 − X2 are independent.

Version: – November 19, 2024

282 sampling distributions and limit theorems

Let α, β ∈ R. Then using Theorem 6.3.13 again we have that α(X1 + X2 ) + β (X1 − X2 ) =
(α + β )X1 + (α − β )X2 is normally distributed. As this is true for any α, β ∈ R, (X1 + X2 , X1 − X2 )
has a bivariate Normal distribution by Definition 6.4.1. Using Theorem 6.2.2 (f) and (g), along
with the fact that X1 and X2 are independent Normal random variables with mean µ and variance
σ 2 , we have

Cov [X1 + X2 , X1 − X2 ] = V ar [X1 ] + Cov [X2 , X1 ] − Cov [X1 , X2 ] − V ar [X2 ] = 0.

Theorem 6.4.3 then implies that X1 + X2 and X1 − X2 are independent.

Step 2: (inductive hypothesis) Let us inductively assume that (a),(b), and (c) are true when n = k
for some k ∈ N.

Step 3: (Proof for n = k + 1) We shall rewrite X k+1 and Sk2+1 using some elementary algebra.

k +1
1 X 1 1
 
k
X k − X k +1 = X k − Xi = 1 − Xk − Xk+1 = (X k − Xk+1 ). (8.1.8)
k+1 k+1 k+1 k+1
i=1

Adding and subtracting X k inside the summand of Sk2+1 , we have

k +1 k +1
1X 1X
Sk2+1 = (Xi − X k+1 )2 = (Xi − X k + X k − X k+1 )2
k k
i=1 i=1
k +1
1 X
= (Xi − X k )2 + 2(Xi − X k )(X k − X k+1 ) + (X k − X k+1 )2
k
i=1
k−1 2 1 1
(Xk+1 − X k )2 + 2(Xk+1 − X k )(X k − X k+1 ) + (k + 1)(X k − X k+1 )2


= Sk +
k k k
k−1 2 1 1 − X k ) (Xk+1 − X k )2
 
(X
= Sk + (Xk+1 − X k )2 − 2(Xk+1 − X k ) k+1 +
k k k k+1 k+1
k−1 2 1
= Sk + (Xk+1 − X k )2 ,
k k+1

where we have used (8.1.8) in the second last inequality. Dividing thoughout by σ 2 and multiplying
by k we have
k 2 k−1 2 k
Sk+1 = Sk + 2 (Xk+1 − X k )2 . (8.1.9)
σ 2 σ 2 σ (k + 1)
Part (a) follows again from Theorem 6.3.13. To prove (b), it is enough to show that
s !
k (k − 1) 2
(Xk+1 − X k ) ∼ Normal (0, 1) and is independent of Sk .
(k + 1)σ 2 σ2

This is so because k
σ 2 (k +1)
(Xk+1 − X k )2 then has the χ21 distribution by Example 8.1.5, and
(k−1) 2 (k−1)
is independent of σ2
Sk by Theorem 8.1.2; by the induction hypothesis σ2 Sk2 has the χ2k−1

Version: – November 19, 2024

 8.1 multi-dimensional continuous random variables 283

distribution, so using (8.1.9) along with Example 5.5.6 will imply that σk2 Sk2+1 has the χ2k distribution.
It is a routine calculation using Theorem 6.3.13 to verify the above distribution by noting that
s s
k
! r ! !
k (k + 1)σ 2 X 1 k
(Xk+1 − X k ) = Xk+1 − Xi .
(k + 1)σ 2 k k (k + 1)σ 2
i=1

By the induction hypothesis, X k and k−1 S 2 are independent. As X1 , . . . , Xk , Xk+1 are mutually
σ2 k
independent, Theorem 8.1.2 implies that Xk+1 is independent of X k and k−1 S 2 . Therefore,
σ2 k

k−1 2
Xk, S , Xk+1 are mutually independent random variables. (8.1.10)
σ2 k

Consequently, another application of Theorem 8.1.2 will then imply that k

σ 2 (k +1)
(Xk+1 − X k )2 and
(k−1) 2
σ2
Sk are independent random variables.
To prove (c), it is enough to show that X k+1 and Xk+1 − X k are independent. The reason is
the following:

(i) Theorem 8.1.2 then implies X k+1 is independent of k

σ 2 (k +1)
(Xk+1 − X k )2 ;

(ii) X k+1 is a function of Xk+1 and X k . So (8.1.10) and Theorem 8.1.2 will then imply X k+1 is
(k−1) (k−1)
independent of σ2 Sk2 and also σ2 (kk+1) (Xk+1 − X k )2 is independent of σ2 Sk2 ;

(k−1) 2
(iii) Using (i) and (ii) we can conclude that X k+1 , σ2
Sk , and k
σ 2 (k +1)
(Xk+1 − X k )2 are
mutually independent; and
(k−1) 2
(iv) finally Sk2+1 is a function σ2
Sk , and k
σ 2 (k +1)
(Xk+1 − X k )2 by (8.1.9). Then (iii) and
Theorem 8.1.2 will imply that Sk+1
2 and X k+1 are independent.

Let α, β ∈ R. We have

k    
X α β α
α(X k+1 ) + β (Xk+1 − X k ) = − Xi + − β Xk+1 .
k+1 k k+1
i=1

Theorem 6.3.13 will imply that α(X k+1 ) + β (Xk+1 − X k ) is is normally distributed random variable
for any α, β ∈ R. So by Definition 6.4.1 (X k+1 , Xk+1 − X k ) is a bivariate normal random variable.
Further, from Theorem 6.2.2 (f) and (g), we have

kX k + Xk+1
Cov [X k+1 , Xk+1 − X k ] = Cov [ , Xk+1 − X k ]
k+1
1 k
= V ar [Xk+1 ] − Cov [X k , Xk+1 ] − V ar [X k ]
k+1 k+1
1 k σ2
= σ2 + 0 + − = 0,
k+1 k+1 k

where we have used (8.1.10) in the last line. From Theorem 6.4.3 we conclude that X k+1 , Xk+1 − X k
are independent. ■

Version: – November 19, 2024

284 sampling distributions and limit theorems

The following important Corollary connects the sampling distributions of X and S 2 to the t distri-
bution, and will be important in the context of confidence intervals, which we discuss in Chapter 9.

Corollary 8.1.11. For n ≥ 2, let X1 , X2 , . . . , Xn be an i.i.d. random sample with distribution

X ∼ Normal µ, σ 2 . Let X and S 2 be as above. Then

√
n(X − µ)
S
has the tn−1 distribution.

Proof. From Theorem 8.1.10 it is clear that

X −µ (n − 1) 2
√ ∼ Normal (0, 1) and S ∼ χ2n−1 .
σ/ n σ2

Noting that
√ X−µ
√
n(X − µ) σ/ n
=q ,
S 1 (n−1)S 2
n−1 σ2

the result follows by Example 8.1.9. ■

exercises

Ex. 8.1.1. Let n ≥ 1. F be the joint distribution function of real valued discrete random variables
X1 , X2 , . . . , Xn as in (8.1.1).

(a) Suppose n = 2. Show that for (s, t) ∈ R2 ,

P (X1 = s, X2 = t) = lim F (u, v ) − F (s, t)

u↓s,v↓t

(b) Reformulate and prove, part (a) for general n ≥ 1.

Ex. 8.1.2. Verify that each of f : R3 → R are density functions on R3 .


 2 (x + x + x ) if 0 < x < 1, i = 1, 2, 3.
1 2 3 i
(a) f (x1 , x2 , x3 ) = 3
0 otherwise

 1 (x2 + x2 + x2 ) if 0 < x < 2, i = 1, 2, 3.
i
(b) f (x1 , x2 , x3 ) = 8 1 2 3
0 otherwise

 2
81 x1 x2 x3 if 0 < xi < 3, i = 1, 2, 3.
(c) f (x1 , x2 , x3 ) =
0 otherwise

 3 (x x + x x + x x )
1 2 1 3 2 3 if 0 < xi < 1, i = 1, 2, 3.
(d) f (x1 , x2 , x3 ) = 4
0 otherwise

Version: – November 19, 2024

 8.1 multi-dimensional continuous random variables 285

Ex. 8.1.3. Suppose (X1 , X2 , X3 ) have a joint density f : R3 → R given by


 4 (x3 + x3 + x3 )
3 1 2 3 if 0 < xi < 1, i = 1, 2, 3.
f ( x1 , x2 , x3 ) =
0 otherwise

(a) Find P (X1 < 12 , X3 > 21 ).

(b) Find the joint density of (X1 , X2 ),(X1 , X3 ), (X2 , X3 ).

(c) Find the marginal densities of X1 , X2 , and X3 .

Ex. 8.1.4. Let D be a set in R3 with a well defined volume. (X1 , X2 , X3 ) are said be uniform on a
set D if they have a joint density given by

 1
Volume(D )
if x ∈ D
f ( x1 , x2 , x3 ) =
0 otherwise.

Suppose D is a cube of dimension R.

(a) Find the joint density (X1 , X2 , X3 ) which is uniform on D.

(b) Find the marginal density of X1 , X2 , X3 .

(c) Find the joint density of (X1 , X2 ),(X1 , X3 ),(X3 , X2 ).

Ex. 8.1.5. Let X1 , X2 , . . . , Xn be i.i.d. random variables having a common distribution function
F : R → [0, 1] and probability density function f : R → R. Let X(1) < X(2) < . . . < X(n) be
the corresponding order statistics. Show that for 1 ≤ i < j ≤ n, (X(i) , X(j ) ) has a joint density
function given by

n!
fX (x, y ) = f (x)f (y )[F (x)]i−1 [F (y ) − F (x)]j−1−i [1 − F (y )]n−j ,
(i) ,X(j ) (i − 1) ! (j − 1 − i) ! (n − j ) !

for −∞ < x < y < ∞.

Ex. 8.1.6. Let X1 , X2 , . . . , Xn be i.i.d. random variables having a common distribution X ∼
X
Uniform(0, 1). Let X(1) < X(2) < . . . < X(n) be the corresponding order statistics. Show that X (1)
(n)
and X(n) are independent random variables.
Ex. 8.1.7. Let {Ui : i ≥ 1} be a sequence of i.i.d. Uniform(0, 1) random variables, and let
N ∼ Poisson(λ). Find the distribution of V = min{U1 , U2 , . . . , UN +1 }.
Ex. 8.1.8. Let −∞ < a < b < ∞. Let X1 , X2 , . . . , Xn i.i.d. X ∼ Uniform(a, b). Find the
X +X
probability density function of M = (1) 2 (n) .
Ex. 8.1.9. Let X1 , X2 be two independent standard Normal random variables. Find the distribution
of Z = X(21) .
Ex. 8.1.10. Let X1 , X2 , . . . , Xn be i.i.d. Uniform(0, 1) random variables.

(a) Find the conditional distribution of X(n) | X(1) = x for some 0 < x < 1.

Version: – November 19, 2024

286 sampling distributions and limit theorems

(b) Find E [X(n) | X(1) = x] and V ar [X(n) | X(1) = x].

Ex. 8.1.11. Suppose X is a symmetric continuous random variable. Let Y be a continuous random
Y is symmetric.
variable such that P (Y > 0) = 1. Show that X

Ex. 8.1.12. Verify (8.1.4).

Ex. 8.1.13. Suppose X1 , X2 , . . . are i.i.d. Cauchy(0, 1) random variables.

(a) Fix z ∈ R. Find a, b, c, d such that

1 1 ax + b cx + d
= + ,
1 + x 1 + (z − x)
2 2 1+x 2 1 + (z − x)2

for all x ∈ R.

(b) Show that X1 + X2 ∼ Cauchy (0, 2).

(c) Use induction to show that X1 + X2 + . . . + Xn ∼ Cauchy (0, n).

(d) Use Lemma 5.3.2 to show that X n ∼ Cauchy (0, 1).

Ex. 8.1.14. Suppose U , V are independent random variables with χ2m and χ2n respectively. Then
show that that Z = U +
U
V is distributed as Beta( 2 , 2 )
m n

8.2 weak law of large numbers

For n ≥ 1, let X1 , X2 , . . . , Xn be an i.i.d. random sample from a population whose distribution is

given by a random variable X which has mean µ. In Chapter 7 we considered the sample mean
n
1X
X= Xi
n
i=1

and showed in Theorem 7.1.4 that E [X ] = µ. We also discussed that X could be considered as an
estimate for µ. The following result makes this precise, and is referred to as the Weak law of large
numbers. To emphasise the dependence of the sample mean and its behaviour on n, we will denote
X by X n .

Theorem 8.2.1. (Weak Law of Large Numbers) Let X1 , X2 , . . . be a sequence of i.i.d.

random variables. Assume that X1 has finite mean µ and finite variance σ 2 . Then for any
ϵ>0
lim P (| X n − µ |> ϵ) = 0, (8.2.1)
n→∞

Proof. Let ϵ > 0 be given. We note that

n n
1X X1 nµ
E [X n ] = E [ Xi ] = E [Xi ] = = µ.
n n n
i=1 i=1

Version: – November 19, 2024

 8.2 weak law of large numbers 287

Using Theorem 4.2.4, Theorem 4.2.6 and Exercise 6.2.16 we have

n
1X
Var[X n ] = Var[ Xi ]
n
i=1
n
1 X
= Var[ Xi ]
n2
i=1
n
1 X
= Var[Xi ]
n2
i=1
σ2
=
n

So we have shown that the random variable X n has finite expectation and variance. By Chebychev’s
inequality (apply Theorem 6.1.13 (a) with k = σϵ ), we have

σ2
P (|X n − µ| > ϵ) ≤ .
nϵ2
2
Therefore as 0 ≤ P (|X n − µ| > ϵ) for all n ≥ 1 and nϵ
σ
2 → 0 as n → ∞, by standard results in real

analysis we conclude that

lim P (| X n − µ |> ϵ) = 0. ■
n→∞

Remark 8.2.2. The convergence

n of X noto µ actually happens with probability one. That is, if we
consider the event A = lim X n = µ , then P (A) = 1. This result is referred to as the Strong
n→∞
Law of large numbers. We state and prove it in Appendix C (see Theorem A.2.1).
We are often interested in using asymptotic results such as this as approximations when n is a
large, but finite number. To develop a sense about the reliability of such approximations, we devote
the remainder of this section to simulations of such behavior.
The Weak Law of large numbers is most commonly used to estimate probabilities of events.
However, before exploring applications, we first look at some simpler examples where we estimate
the mean of a distribution.
h  i
Example 8.2.3. Suppose X ∼ Uniform(0, 1). What is E log 1−X X
? It is easy to argue, using
symmetry, that the answer should be 0 (see Exercise 8.2.1). To verify this using simulation, we can
simply generate a large number of Uniform(0, 1) random variables, transform them, and compute
their sample mean. If the Weak Law gives a good approximation for finite samples, this should be
“close” to the true expectation.

u <- runif(10000)
mean(log(u / (1-u)))

[1] 0.006766486

Of course, however good an approximation, this estimate is still random, so we should replicate it
several times to get an idea of its general behaviour.

Version: – November 19, 2024

288 sampling distributions and limit theorems

replicate(10, {
x <- runif(10000)
mean(log(x / (1 - x)))
})

[1] 0.006112412 -0.009243912 0.019540336 0.022273225 0.003628815

[6] 0.016525060 -0.017777608 0.011400835 -0.002747576 -0.020943686

These ten replications suggest that the approximation is usually correct only up to the first decimal
place, even though the value of n = 10000 might normally be considered large. Not surprisingly,
the approximation gets worse for n = 100.

replicate(10, {
x <- runif(100)
mean(log(x / (1 - x)))
})

[1] -0.021428316 0.003939597 0.171115165 0.122717112 0.235109230

[6] -0.007103688 0.082358211 0.097755404 -0.185062653 -0.108332961

To get a sense of how the approximation improves with n, it is common to plot the cumulative or
partial means as a function of n. For example, Figure 8.1 is created using

N <- 10000
i <- seq(1, N) # to be used as denominator
x <- runif(N)
m <- cumsum(log(x / (1 - x))) / i

xyplot(m ˜ i, xlab = "Index", ylab = "Partial Mean", type = "l")

This plot suggests that the estimate gets close to zero for fairly small n, and after that improvement
is not substantial. This plot, however, only tells us about the behaviour of one particular sequence
of random variables, whose partial means are guaranteed to converge to the true mean as n → 0
according to the Strong Law, which we have stated but not proved. The Weak Law, on the other
hand, states a result about the distribution of the sample mean. To assess whether it holds, we look
at independent replications of the experiment and plot the resulting paths taken by the partial
means together. We omit the code used to do this, but show the result of one such simulation
experiment in Figure 8.2. One can observe that there is a reduction in the variance of the partial
means as n increases, which was the essential requirement in the proof of the Weak Law. ■

Example 8.2.4. We modify the previous example as follows. Suppose U , V ∼ Uniform(0, 1) are
independent, and X = max(U , V ). What is E (log 1−X
X
)?

Version: – November 19, 2024

 8.2 weak law of large numbers 289

0.8
Partial Mean

0.6
0.4
0.2
0.0
−0.2

0 2000 4000 6000 8000 10000

Index

Figure 8.1: Cumulative or partial means computed from 10000 random samples from the
population log 1−X
X
, where X follows Uniform(0, 1).

0.4
Partial Mean

0.2

0.0

−0.2

−0.4

0 2000 4000 6000 8000 10000

Index

Figure 8.2: Results of the same experiment that is shown in Figure 8.1, replicated 50 times. For
each replication, cumulative or partial means computed from 10000 random samples
are shown. The underlying population is log 1−X X
, where X follows Uniform(0, 1).

Version: – November 19, 2024

290 sampling distributions and limit theorems

The answer is not as obvious in this case. An approximate answer is easy to obtain by invoking
the Weak Law of large numbers.

replicate(10, {
u <- runif(10000)
v <- runif(10000)
x <- pmax(u, v)
mean(log(x / (1 - x)))
})

[1] 0.9981781 0.9868826 1.0355666 1.0128563 1.0030963 1.0011916

[7] 0.9939234 0.9900825 0.9905087 0.9863563

These results suggest that the expectation is 1, a fact that can be verified by explicit computation
(See Exercise 8.2.1). ■

Example 8.2.5. Suppose that U and V are independent Uniform(0, 1), and interpret them as
coordinates of a point in R2 . Suppose we want to calculate the expected norm of (U , V ). In other
√
words, if Z = U 2 + V 2 , we want to calculate E [Z ].

As before, we can estimate the expectation by simulating the experiment a large number of
times.

replicate(10, {
u <- runif(10000)
v <- runif(10000)
z <- sqrt(uˆ2 + vˆ2)
mean(z)
})

[1] 0.7642686 0.7674234 0.7660785 0.7681380 0.7608846 0.7651592

[7] 0.7621663 0.7715568 0.7613604 0.7678146

See Exercise 8.2.2 for explicit computation of E [Z ]. ■

Theorem 8.2.1 states that for any ϵ > 0, the probability P (|X n − µ| > ϵ) goes to zero as n → ∞.
This mode of convergence of the sample mean X n to the true mean µ is called “convergence in
probability” . We define it precisely below.

Version: – November 19, 2024

 8.2 weak law of large numbers 291

Definition 8.2.6. A sequence X1 , X2 , . . . is said to converge in probability to a random

variable X if for any ϵ > 0
lim P (| Xn − X |> ϵ) = 0, (8.2.2)
n→∞

The notation
p
Xn −→ X

is typically used to convey that the sequence X1 , X2 , . . . converges in probability to X.

Note that in the above definition the limit is allowed to be a non-trivial random variable X, although
in most examples we will consider, X will be a constant.

Example 8.2.7. Let X1 , X2 , . . . be i.i.d. random variables from the Uniform(0, 1) distribution.
We already know by the law of large numbers that X converges to E (X1 ) = 12 in probability. Often
we are interested in other functionals (i.e. f (X1 , X2 , . . . , Xn ) for some suitable f and n ≥ 1) of
the sample and their convergence properties. As an example, consider the n-th order statistic
X(n) = max{X1 , X2 , . . . , Xn }. Intuitively, as n increases, it is more and more likely that X(n) will
get closer to its maximum possible value 1. To see this formally, first note that for ϵ > 1,
     
P X(n) − 1 ≥ ϵ = P X(n) ≤ 1 − ϵ + P X(n) ≥ 1 + ϵ = 0.

For any 0 < ϵ < 1,

     
P X(n) − 1 ≥ ϵ = P X(n) ≤ 1 − ϵ + P X(n) ≥ 1 + ϵ
 
= P X(n) ≤ 1 − ϵ + 0
 
n
= P ∩ {Xi ≤ 1 − ϵ}
i=1

n
= 1−ϵ .
p
As lim (1 − ϵ)n = 0 for 0 < ϵ < 1, it follows from definition 8.2.6 that X(n) −→ 1 as n → ∞. ■
n→∞

An important application of the Weak Law of large numbers follows by noting that the sample
proportion discussed in Section 7.1.2 is the sample mean of Bernoulli random variables.

Example 8.2.8. Suppose we are interested in an event A and want to estimate p = P (X ∈ A).
We consider a sample X1 , X2 , . . . , Xn which is i.i.d. X. We define a sequence of random variables
{Yn }n≥1 by 
1 if X ∈ A
n
Yn =
0 if Xn ̸∈ A.

Version: – November 19, 2024

292 sampling distributions and limit theorems

Clearly Yn are independent (as the Xn are), and further they are identically distributed with
P (Yn = 1) = P (Xn ∈ A) = p. In particular, {Yn } is an i.i.d. Bernoulli(p) sequence of random
variables. We readily observe (as done in Chapter 7) that
n
1X 1
Yn = Yi = #{Xi ∈ A} = p̂.
n n
i=1

Hence the Weak Law of large numbers (applied to the sequence Yn ) implies that the sample
proportion will converge to the true proportion p in probability. This provides legitimacy, as
discussed earlier, to the relationship between probability and relative frequency. ■

exercises

Ex. 8.2.1. Find E (log 1−X

X
) when

1. X ∼ Uniform(0, 1).

2. X = max(U , V ) when U , V ∼ Uniform(0, 1) and are independent.

Ex. 8.2.2. Let U , V ∼ Uniform(0,1) and be independent.

1. Find the distribution of U 2

2. Find the distribution of U 2 + V 2

3. Find the distribution of norm of Z = (U , V ) and E [Z ]

Ex. 8.2.3. Let (U , V ) ∼ Uniform(D) where D = {(x, y ) : x2 + y 2 = 1}. Find the distribution of
norm of Z = (U , V ) and E [Z ].

Ex. 8.2.4. Let X, X1 , X2 , . . . be i.i.d. random variables that are uniformly distributed over
the interval (0, 1). Consider the first order statistic X(1) = min{X1 , · · · , Xn }. Show that X(1)
converges to 0 in probability.

Ex. 8.2.5. Let X1 , X2 , . . . , Xn , . . . be i.i.d. random variables with finite mean and variance. Define
n
2 X
Yn = iXi .
n(n + 1)
i=1

p
Show that Yn −→ E (X1 ) as n → ∞.
n
Ex. 8.2.6. Let {Xi : i ≥ 1} be a sequence of i.i.d. Normal (0, 1) random variables. Let Sn = Xi .
P
1
Design a suitable R-code as in Example 7.1.9 that will provide an estimate of the probability that
S1 , . . . , S100 all have the same sign.
p
Ex. 8.2.7. Suppose Xn and X are random variables such that Xn −→ X as n → ∞. Suppose
p
h : R → R is a continuous function. Then show that h(Xn ) −→ h(X ) as n → ∞.

Version: – November 19, 2024

 8.3 convergence in distribution 293

8.3 convergence in distribution

When discussing a collection of random variables it makes sense to think of them as a sequence of
objects, and as with any sequence in calculus we may ask whether the sequence converges in any
way. We have already seen “convergence in probability” in the previous section. Here we will be
interested in what is known as “convergence in distribution”. This type of convergence plays an
important role in understanding the limiting distribution of the sample mean, as we will see later,
particularly in the Central Limit Theorem, Theorem 8.4.1.

Definition 8.3.1. A sequence X1 , X2 , . . . is said to converge in distribution to a random

variable X if FXn (x) converges to FX (x) at every point x for which FX is continuous. The
following notation
d
Xn −→ X

is typically used to convey that the sequence X1 , X2 , . . . converges in distribution to X.

Example 8.3.2. Let Xn ∼ Uniform(0, n1 ) so that the distribution function is


 0
 if 0 ≤ x
FXn (x) = nx if 0 < x < 1
n
1 if x ≥ n1



and it is then easy to see that FXn (x) converges to

(
0 if 0 ≤ x
F (x) =
1 if x > 0

If X is the constant random variable for which P (X = 0) = 1, then X has distribution function
(
0 if 0 < x
FX (x) =
1 if x ≥ 0

It is not true that FX (x) = F (x), but the two are equal at points where they are continuous.
Therefore the sequence X1 , X2 , . . . converges in distribution to the constant random variable 0. ■

Note that this form of convergence does not generally guarantee that probabilities associated with
X can be derived as limits of probabilities associated with Xn . For instance, in the example above
P (Xn = 0) = 0 for all n while P (X = 0) = 1. However, with a few additional assumptions a
stronger claim may be made.

Theorem 8.3.3. Let fX1 , fX2 , . . . be the respective densities of continuous random variables
X1 , X2 , . . . . Suppose they converge in distribution to a continuous random variable X with
density fX . Then for every interval A we have P (Xn ∈ A) → P (X ∈ A).

Version: – November 19, 2024

294 sampling distributions and limit theorems

Proof. As X is a continuous random variable FX (x) is the integral of a density, and thus a
continuous function. Therefore convergence in distribution guarantees that FXn (x) converges to
FX (x) everywhere. Let A = (a, b) (and note that whether or not endpoints are included does not
matter as all random variables are taken to be continuous). Then

Zb
P (Xn ∈ A) = fXn (x) dx
a
= FXn (b) − FXn (a)
→ FX ( b ) − FX ( a )
Zb
= fX (x) dx = P (X ∈ A).
a ■

When a sequence Y1 , Y2 , . . . of random variables converges in probability to a constant c, one often

then tries to understand how the distribution of suitably scaled versions of the fluctuations Yn − c
behave in the limit. In many cases, we are able to identify the correct scaling at which the scaled
fluctuations converge in distribution to a non-constant random variable. The most well known
example of this is the Central Limit Theorem, to be studied in the next section, which states
√
that the fluctuations of the sample mean of n i.i.d. random variables scaled by n converges in
distribution to standard Normal, under a finite second moment hypothesis. We shall now discuss
another fundamental example.
Example 8.3.4. Let X1 , X2 , . . . be i.i.d. Uniform(0, 1) random variables. Consider Mn =
min(X1 , X2 , . . . , Xn ), the minimum value among the first n observations. Normally, we would
denote Mn simply by X(1) , but here we use a different notation to emphasize that the minimum
can change with n.
p
We saw earlier in Exercise 8.2.4 that that Mn −→ 0 and in Example 8.2.3. that convergence
in probability is at a certain rate. To understand the fluctuations around the limit we shall try
to identify the correct scaling. To see this, first note that E (Mn ) = 1/(n + 1), so (n + 1)Mn has
expected value 1 for all n. Thus we could use a scaling factor of “n” or “n + 1”. So, for x ∈ R,
 x
P (nMn > x) = P Mn >
n
 x x
= P X1 > , . . . , Xn >
n  n
  x n
= P X1 > by independence of X1 , . . . , Xn
n
 x n

= 1− → e−x as n → ∞.
n

In other words, if Z is exponentially distributed with mean 1, then we have shown that P (nMn ≤
p d
x) → P (Z ≤ x) for all x. So we have Mn −→ 0 and n(Mn − 0) −→ Z. ■
Establishing convergence in distribution using the definition, as done in the previous example,
is not always possible. There are three key results that we will use in the book. These provide
sufficient conditions that are intuitive and often easier to check. The first result deals with the

Version: – November 19, 2024

 8.3 convergence in distribution 295

case of convergence in distribution for continuous random variables, and states that pointwise
convergence of densities implies convergence in distribution.

Theorem 8.3.5. (Scheffé’s Lemma) Let fX1 , fX2 , . . . be the respective densities of continu-
ous random variables X1 , X2 , . . . , and let fX be the density of a continuous random variable
d
X. Suppose fXn (x) → fX (x) as n → ∞ for all x ∈ R. Then, Xn −→ X as n → ∞.

This is a deceptively simple result. After all, one could argue that if fXn (·) converges to fX (·)
pointwise as n → ∞, then so should

Za Za
fXn (u)du → fX (u)du
−∞ −∞

as n → ∞ for any a ∈ R. However, such interchanging of limits and integrals is not always
valid. The result that permits it in this particular situation, known as the “dominated convergence
theorem”, is beyond the scope of this book.
d
Example 8.3.6. Suppose Xn ∼ Normal n1 , 1 . Then it is intuitively clear that Xn −→ Z, where

Z ∼ Normal (0, 1). This follows from an elementary application of Scheffé’s Theorem, as

1 1 1 2 1 1 2
fXn (x) = √ e− 2 (x− n ) → √ e− 2 x = fZ (x) for all x ∈ R.
2π 2π

A direct proof is also simple in this case. As Yn = Xn − n1 has the Normal (0, 1) distribution, we
have
FXn (x) = P (Xn ≤ x) = P (Yn ≤ x − 1/n) = FZ (x − 1/n) → FZ (x)

as n → ∞ for all x ∈ R because FZ is continuous everywhere. ■

Example 8.3.7. Let Xn ∼ tn distribution. Then

− 12 (n+1)
Γ( n + 12 ) t2

fXn (t) = √ 2 1+
nπ Γ( n2 ) n

for t ∈ R. It is straightforward to verify that (see Exercise 8.3.1)

1 t2
fXn (t) → √ exp(− ) (8.3.1)
2π 2

d
as n → ∞. Consequently by Scheffé’s Theorem Xn −→ Z as n → ∞ where Z ∼ Normal(0,1). ■

The second result, which works for both discrete and continuous random variables, formalizes the
intuition that if all moments of Xn exist and they converge to respective moments of X, then Xn
should converge in distribution to X. Unfortunately, a proof of this result is also beyond the scope
of this book.

Version: – November 19, 2024

296 sampling distributions and limit theorems

t1 t3 t5 t10 t50

0.4

0.3
Density

0.2

0.1

0.0

−3 −2 −1 0 1 2 3

Figure 8.3: Density of the tn distribution converging to that of standard Normal as n → ∞,

illustrated using the parameter values n = 1, 3, 5, 10 and 50. The thick grey line,
which represents the standard Normal density, is almost indistingushabe from the
t50 density.

Theorem 8.3.8. (M.G.F. Convergence Theorem) Let X1 , X2 , . . . be a sequence of random

variables whose moment generating functions Mn (t) exist in an interval containing zero. If
Mn (t) → M (t) on that interval, where M (t) is the moment generating function of a random
variable X, then Xn converges to X in distribution.

To illustrate the use of this result, consider an alternative proof of the limiting relationship between
Binomial and Poisson random variables (See Theorem 2.2.2).

Example 8.3.9. Let X ∼ Poisson(λ) and let Xn ∼ Binomial(n, nλ ). Then Xn converges in

distribution to X.
The moment generating function of a Binomial variable was already computed in Example
6.3.7. Therefore,
  n
λ t λ
M Xn ( t ) = e + 1−
n n
n
λ(e − 1)
t

= 1+
n

Using Exercise 8.4.4, we see that

t −1)
M Xn ( t ) → e λ ( e .

Version: – November 19, 2024

 8.3 convergence in distribution 297

On the other hand, the moment generating function of X is

MX ( t ) = E [etX ]
∞
X
= etj P (X = j )
j =0
∞
X λj e−λ
= etj
j!
j =0
∞ t
t X (λet )j e−λe
= eλe · e−λ ·
j!
j =0
t −1)
= eλ ( e

where the series equals 1 since it is simply the sum of the probabilities of a Poisson(λet ) random
variable.
Since MXn (t) → MX (t), by the M.G.F. convergence theorem (Theorem 8.3.8), Xn converges in
distribution to X. That is, Binomial(n, p) random variables converge in distribution to a Poisson(λ)
distribution when p = nλ and n → ∞. ■
The last result cannot be used to establish convergence in distribution directly. However, if
d
we already know that a sequence Xn −→ X, then this result can often be used to establish the
convergence in distribution of small “perturbations” of Xn , as long as the perturbations converge
in probability.

Lemma 8.3.10. (Slutsky’s Theorem) Let {Xn , Yn : n ∈ N} and X be random variables on a

d p
probability space (Ω, B, P ). Let Xn −→ X, and Yn −→ c for some c ∈ R. Then
d
(a) Xn + Yn −→ X + c
d
(b) Xn Yn −→ cX
d
(c) Xn
Yn −→ X
c if c ̸= 0

Proof. We prove only (a); (b) and (c) can be proved similarly. Let ϵ > 0 be given. Write
Fn = FXn +Yn . Choose t such that t, t − c + ϵ, t − c − ϵ are all continuity points of FX . This is
possible as there can be at most countably many points of discontinuity of FX . Now,

Fn (t) ≤ P (Xn + Yn ≤ t, |Yn − c| < ϵ) + P (|Yn − c| ≥ ϵ)

≤ P (Xn ≤ t − c + ϵ) + P (|Yn − c| ≥ ϵ)

and

Fn ( t ) ≥ P (Xn < t − c − ϵ) − P (|Yn − c| ≥ ϵ)

The result follows because we have shown that

FX (t − c − ϵ) ≤ lim inf Fn (t) ≤ lim sup Fn (t) ≤ FX (t − c + ϵ). ■

n→∞ n→∞

Version: – November 19, 2024

298 sampling distributions and limit theorems

Our primary application of Slutsky’s Theorem will come in Section 8.5. However, to illustrate its
usefulness, we will show that the result in Example 8.3.6 follows immediately using it below.
Example 8.3.11. Recall the tn distribution from Example 8.1.9. The convergence of the tn
distribution to the Normal (0, 1) distribution, proved in Example 8.3.7, would follow by Lemma 8.3.10
√ p
(c) if we could show that the sequence Yn /n −→ 1, where Yn is the χ2n random variable in the
denominator in the definition of the tn distribution. This is shown in two steps. Either directly
applying Chebychev’s inequality (Theorem 6.1.13) on Ynn , or by an application of the Weak Law of
Large Numbers (Theorem 8.2.1) we can show that

Yn p
−→ 1 as n → ∞. (8.3.2)
n
n
Indeed, as Yn = Xi with Xi i.i.d. χ21 random variables, and E [X12 ] = 1 < ∞, it is immediate by
P
i=1
√ p
Theorem 8.2.1 that (8.3.2) holds. It then follows from Exercise 8.2.7 that Yn /n −→ 1. ■

exercises

Ex. 8.3.1. Show (8.3.1).

d
Ex. 8.3.2. Let c ∈ R and X1 , X2 , . . . be a sequence of random variables. Show that if Xn −→ c
p
then Xn −→ c.
Ex. 8.3.3. Let Y1 , Y2 , . . . be a sequence of χ2n random variables.
d
(a) Show that Yn
n −→ 1 as n → ∞.
d
(b) Using Exercise 8.3.2 and Lemma 8.3.10 conclude that tn −→ Z as n → ∞ with Z being
standard normal random variable.
Ex. 8.3.4. Consider a sequence Xn , n ≥ 1 of random variables such that Xn ∼ Normal n1 , 1 + n1 .


d
Show that Xn −→ Z as n → ∞ where Z ∼ Normal(0,1).
Ex. 8.3.5. Suppose a sequence Xn , n ≥ 1 of random variables converges to a random variable X in
probability. Show that Xn converges in distribution to X. That is, show that

FXn (x) → FX (x) as n → ∞

for all continuity points of FX : R → [0, 1] with FXn , FX being the distribution functions of Xn
and X respectively.
Ex. 8.3.6. Let X1 , X2 , . . . be i.i.d. Uniform(0, 1) random variables. Generalize the definition
of Mn = min(X1 , X2 , . . . , Xn ) in Example 8.3.4 as follows: For fixed k ≥ 1, define Mn,k =
X(k +1) − X(k ) .

(a) Show that nMn,k = n(X(k+1) − X(k) ) also converges to Exponential(1) random variable in
distribution
(b) Show that for any fixed k ≥ 1, nX(k) converges to the Gamma(k, 1) distribution.

Version: – November 19, 2024

 8.4 central limit theorem 299

8.4 central limit theorem

For n ≥ 1, let X1 , X2 , . . . , Xn be an i.i.d. random sample from a distribution X which has mean µ
and variance σ 2 , but is otherwise unknown. Consider the sample mean
n
1X
X= Xi .
n
i=1

As observed in Theorem 7.1.4, E [X ] = µ and SD[X ] = √σ .

n
As discussed before, we might view
this information as indicating that with high probability, X is typically close to µ up to an error of
√σ . As n → ∞, √σ → 0 and this indicates that X approaches µ. We have already verified that X
n n
converges in probability to µ courtesy of the Weak Law of large numbers, and noted that in fact it
converges with probability 1 by the Strong Law of large numbers.

To get a better understanding of the limiting distribution of X, we standardize it to have mean

0 and variance 1, and consider

(X − µ) √ (X − µ)
Yn = √ = n .
σ/ n σ

Without further information about X, the common distribution of X1 , X2 , . . . , Xn , it is not possible

to find the exact probabilities of events connected with Yn . However, it turns out that one can often
find good approximate values, because for a large class of possible distributions X, the distribution
of Yn is close to that of the standard Normal random variable, particularly for large n. This
remarkable fact is referred to as the Central Limit Theorem and we prove it next.

As done earlier, we shall denote X by X n in the statement and proof of the Theorem below to
emphasise its dependence on n.

Theorem 8.4.1. (Central Limit Theorem) Let X1 , X2 , . . . be i.i.d. random variables

with finite mean µ, finite variance σ 2 , and possessing common moment generating function
MX (). Then
√ (X n − µ) d
n −→ Z, (8.4.1)
σ
where Z ∼ Normal (0, 1).

√ (X n −µ)
Proof. Let Yn = n σ . We will verify that

t2
lim MYn (t) = e 2 .
n→∞

Version: – November 19, 2024

300 sampling distributions and limit theorems

Now, using the definition of the moment generating function and some elementary algebra we have

√ (X n − µ)
  
MY n ( t ) = E [exp(tYn ))] = E exp t n
σ
n n
" !!# " !#
t√ 1 X X t
= E exp n Xi − µ = E exp √ (Xi − µ)
σ n σ n
i=1 i=1
" n  #
Y t
= E exp √ (Xi − µ) . (8.4.2)
σ n
i=1

As X1 , X2 , . . . , Xn are independent, we can conclude using Theorem 8.1.2 that

     
t t t
exp √ (X1 − µ) , exp √ (X2 − µ) , . . . , exp √ (Xn − µ)
σ n σ n σ n

are also independent. From Exercise 7.1.3 and 7.1.4, they also have the same distribution. So from
the calculation in (8.4.2) and using Exercise 6.3.4 inductively we have

n n
"  #   
Y t Y t
MYn (t) = E exp √ (Xi − µ) = E exp √ (Xi − µ)
σ n σ n
i=1 i=1
(Using Theorem 6.3.9(a))
   n
t (X − µ)
= E exp √ . (8.4.3)
n σ
  n
where X is the common distribution of X1 , X2 , . . . . In other words, MYn (t) = MU √tn , where
U = X−µ σ . As E [U ] = 0, E [U ] = 1 we have that MU (0) = 0 and MU (0) = 1. From Exercise
2 ′ ′′

8.4.5, we have that for t ∈ R

t2
MU ( t ) = 1 + + g (t), (8.4.4)
2

g (s)
where g satisfies lim = 0. Thus, we have
s→0 s2
n  n  n
t2 1 t2
    
t t t
MYn (t)) = MU √ = 1+ +g √ = 1+ + ng √ .
n 2n n n 2 n
 
t2 t2
Using the fact that for any fixed t, 2 + ng √t
n
→ 2 as n → ∞ and Exercise 8.4.4 it follows that,

t2
lim MYn (t) = e 2 .
n→∞

t2
Theorem 8.3.8 then implies the result as the limit e 2 is the moment generating function of the
standard Normal distribution. ■

Version: – November 19, 2024

 8.4 central limit theorem 301

Remark 8.4.2. The existence of moment generating function is not essential for the Central Limit
Theorem, and (8.4.1) holds as long as X1 , X2 , . . . are i.i.d. random variables with finite mean µ
and finite variance σ 2 . However, the proof of this more general statement is more complicated.

Remark 8.4.3. An equivalent formulation of the Central Limit Theorem is often useful. By
n
n −nµ
definition of X n and elementary algebra we see that Yn = S√ , where Xi . It follows
P
nσ
Sn =
i=1
that
Sn − nµ d
√ −→ Normal (0, 1) . (8.4.5)
nσ

Remark 8.4.4. The Central Limit Theorem is a remarkable result. But it perhaps bears emphasis
that the remarkable part of the result is not the specific statistic, the sample mean, but rather the
Normality of the limiting distribution, which arises in many other situations as well. Although
most such results are beyond the scope of this book, we show later in this chapter that the sample
median, when suitably standardized, also converges to the standard Normal distribution under
fairly general conditions.

The Central Limit Theorem for the sample median can be viewed as a refinement of the Weak Law
of large numbers. The weak law tells us that the sample mean X converges to the expectation µ as
n → ∞. However, for any finite n, X is still a non-constant random variable, whose distribution
we may be interested in. This distribution can be quite complicated in general. The Central Limit
Theorem is remarkable because it says that regardless of the underlying distribution, probabilities
concerning the sample mean X can be well approximated by standard Normal probabilities for
large n.

Before looking at uses of such approximations, let us consider the factors that might affect the
quality of the approximation. The Central Limit Theorem does not say anything about how well
the approximation will be for any given n, but we can guess that it will be better for larger n, and
also depend on the distribution giving rise to the data.

Example 8.4.5. As we have seen earlier, an important application of the Weak Law is to estimate
probabilities of events by sample proportion. Here the underlying distribution is Bernoulli(p),
with the probability p estimated by the sample proportion p̂n = Sn /n. Suppose X1 , . . . , Xn are
independent Bernoulli(p) random variables. Then Sn ∼ Binomial(n, p) and

p̂ − p Sn − np d
p n := p −→ Z,
p(1 − p) np(1 − p)

where Z is standard Normal. Let us see how the quality of this approximation changes with the
choice of n and p.

Instead of simulating Xi -s individually, we can simulate Sn directly using the rbinom() function.
For a specific choice of p and n, we could simulate standardized Sn values as follows.

Version: – November 19, 2024

302 sampling distributions and limit theorems

p <- 0.5
n <- 25
s <- rbinom(1000, size = n, prob = 0.5)
z <- (s - n * p) / sqrt(n * p * (1-p))
mean(z)

[1] 0.0052

sd(z)

[1] 0.9830459

The mean and standard deviation of the sample proportion, computed over these 1000 replication,
matches what we expect. To see how similar their overall distribution is to the standard Normal
distribution, the top panel in Figure 8.4 shows empirical frequency distribution plots obtained from
1000 replications for p = 0.5 and n = 10, 25, 50, and 100. Similar plots for p = 0.25 and p = 0.05
shown in the middle and bottom panels. The Normal approximation obtained using the Central
Limit Theorem are added for comparison. As the sample spaces differ substantially depending on n,
the quantity plotted on the y-axis is not the relative frequency but rather a scaled version, similar
to the scaling done in histograms, that makes the scaled quantities comparable with each other and
the Normal density. From these plots, we can conclude that the distribution of Binomial proportion
is well approximated by Normal when p is close to 12 , although for smaller sample sizes the number
of ties can become an issue as well. Values of p away from 12 can generate skewed (asymmetric)
distributions for which the Normal is not a good approximation. A general convention often used is
to consider the approximation valid if both np and n(1 − p) are at least 5.
As we saw in Chapter 7, Q-Q plots are often more useful for assessing departure from Normality.
Figure 8.5 shows Normal Q-Q plots that are analogous to the empirical frequency distribution plots
in Figure 8.4. Each plot represents 1000 replications, for p = 0.5, 0.25, 0.05 and n = 10, 25, 50, 100.
These largely confirm what we already saw from the empirical frequency distribution plots, and
suggest, in particular, that the Normal approximation may be unreliable when p is close to 0 or 1. ■

Example 8.4.6. The Central Limit Theorem applies not just to sample proportions but to general
discrete and continuous distributions if they have finite expectation and variance. For continuous
distributions, ties happen with probability 0, so empirical frequency distribution plots are not useful.
We can use histograms as an alternative, but Q-Q plots are more useful when the primary goal is
to compare with a Normal distribution.
In Figure 8.6, we show Q-Q plots similar to those in Figure 8.5, but instead of sample proportion,
we consider means of random samples from three continuous distributions, namely Uniform(0, 1),
Exp(1), and Cauchy, with sample sizes n = 5, 20, 50, 100. These plots suggest that even with a
shape very different from Normal, the distribution of the Uniform sample mean is well approximated
by a Normal distribution even for small n. For the heavily asymmetric Exponential distribution,

Version: – November 19, 2024

 8.4 central limit theorem 303

−2 0 2 4 −2 0 2 4

10 25 50 100
Scaled Frequency

0.4
0.3
0.2
0.1
0.0

−2 0 2 4 −2 0 2 4

Sample Proportion (standardized)

−2 0 2 4 −2 0 2 4

10 25 50 100
Scaled Frequency

0.4
0.3
0.2
0.1
0.0

−2 0 2 4 −2 0 2 4

Sample Proportion (standardized)

−2 0 2 4 −2 0 2 4

10 25 50 100
Scaled Frequency

0.4
0.3
0.2
0.1
0.0

−2 0 2 4 −2 0 2 4

Sample Proportion (standardized)

Figure 8.4: Empirical frequency distribution plots of standardized sample proportions when
true probability is (top) p = 0.5, (middle) p = 0.25, and (bottom) p = 0.05. In
each case, the standard Normal density has been added for comparison. To account
for the different sample spaces, the frequencies plotted on the y-axis have been
scaled to make them comparable with each other and the Normal density.

Version: – November 19, 2024

304 sampling distributions and limit theorems

−2 0 2 −2 0 2

10 25 50 100
Sample Proportion
(standardized)

−2

−2 0 2 −2 0 2

Quantiles of Standard Normal

−2 0 2 −2 0 2

10 25 50 100
Sample Proportion

4
(standardized)

−2

−2 0 2 −2 0 2

Quantiles of Standard Normal

−2 0 2 −2 0 2

10 25 50 100
Sample Proportion
(standardized)

−2

−2 0 2 −2 0 2

Quantiles of Standard Normal

Figure 8.5: Q-Q plot of standardized sample proportions when true probability is (top) p = 0.5,
(middle) p = 0.25, and (bottom) p = 0.05.

Version: – November 19, 2024

 8.4 central limit theorem 305

−2 0 2 −2 0 2

5 20 50 100
(standardized)
Sample Mean

−2

−2 0 2 −2 0 2

Quantiles of Standard Normal

−2 0 2 −2 0 2

5 20 50 100
(standardized)
Sample Mean

2
1
0
−1
−2

−2 0 2 −2 0 2

Quantiles of Standard Normal

−2 0 2 −2 0 2

5 20 50 100
(standardized)
Sample Mean

50

−50

−2 0 2 −2 0 2

Quantiles of Standard Normal

Figure 8.6: Q-Q plot of standardized sample mean of random sample from (top) Uniform(0, 1),
(middle) Exponential(1), and (bottom) Cauchy.

Version: – November 19, 2024

306 sampling distributions and limit theorems

this convergence requires a larger sample size. For the Cauchy distribution, which does not have
finite mean, the Central Limit Theorem does not hold at all. ■

Before moving on, let us summarize the main conclusions from the last two examples. Although the
sample mean converges to the population mean, the convergence is not necessarily immediate. Thus,
although we can expect that for large n the sample proportion or sample mean will be “close” to
the population proportion or mean, we cannot expect it to be exactly the same. The Central Limit
Theorem assures us that under fairly mild assumptions, the difference will behave like a Normal
random variable. As we see in the next chapter, this knowledge allows us to make useful statements
about the population proportion or mean, when it is unknown, based on what we observe.

8.4.1 Normal Approximation

A typical application of the Central Limit Theorem is to find approximate value of the probability of
events related to Sn or X. For instance, suppose we were interested in calculating for any a, b ∈ R,
P (a < Sn ≤ b) for large n. We would proceed in the following way. We know from (8.4.5) that
 
Sn − nµ
P √ ≤x → P (Z ≤ x) (8.4.6)
nσ

as n → ∞ for all x ∈ R.
 
a − nµ Sn − nµ b − nµ
P (a < Sn ≤ b) = P √ < √ ≤ √
nσ nσ nσ
Sn − nµ b − nµ Sn − nµ a − nµ
= P( √ ≤ √ ) −P( √ ≤ √ )
nσ nσ nσ nσ
from (8.4.6) for large enough n
b − nµ a − nµ
≈ P (Z ≤ √ ) − P (Z ≤ √ )
nσ nσ
a − nµ b − nµ
= P( √ <Z≤ √ ),
nσ nσ

where in the second last line we have used the notation ≈ to indicate that the right hand side is an
approximation. Therefore we would conclude that for large n,
 
a − nµ b − nµ
P (a < Sn ≤ b) ≈ P √ <Z≤ √ . (8.4.7)
nσ nσ

We would then use the R function pnorm() or Normal Tables (See Table B.1) to compute the right
hand side. A similar computation would also yield
√ √ 
n(a − µ) n(b − µ)
. (8.4.8)


P a<X≤b ≈P <Z≤
σ σ

Example 8.4.7. Let Y be a random variable distributed as Gamma(100, 4). Suppose we were
interested in finding P (20 < Y ≤ 30). Suppose X1 , X2 , . . . , X100 are independent Exponential (4)

Version: – November 19, 2024

 8.4 central limit theorem 307

100
random variables then Y and S100 = Xi have the same distribution. Therefore, applying the
P
i=1
Central Limit Theorem with µ = E [X1 ] = 14 , σ = SD[X1 ] = 41 , we have

P (20 < Y ≤ 30) = P (20 < S100 ≤ 30)

20 − 100(0.25) 30 − 100(0.25)
 
≈ P √ <Z≤ √ by (8.4.7)
100(0.25) 100(0.25)
−5 5
= P( <Z≤ )
2.5 2.5
= P (−2 < Z ≤ 2)
= P (Z ≤ 2) − P (Z ≤ −2)
using symmetry of Normal distribution
= P (Z ≤ 2) − (1 − P (Z ≤ 2))
= 2P (Z ≤ 2) − 1

Looking up Table B.1, we see that this value comes out to be approximately 2 × 0.9772 − 1 = 0.9544.
A more precise answer is given by R as

2 * pnorm(2) - 1

[1] 0.9544997

Using R, we can also compare this with the exact probability that we are approximating.

pgamma(30, 100, 4) - pgamma(20, 100, 4)

[1] 0.9550279

In this example, the approximation is correct to three decimal places. ■

8.4.2 Continuity Correction

n
Suppose X1 , X2 , X3 , . . . are all integer valued random variables. Then Sn = Xi is also an integer
P
i=1
valued random variable. Now, for any integer k, P (Sn ≤ k ) = P (Sn ≤ k + h) for all 0 < h < 1.
However it is easy to see that two distinct values of h will lead to two different answers if we use
the Normal approximation provided by the Central Limit Theorem. It is customary to use h = 12
when computing such probabilities using the Normal approximation, as

P ( Sn ≤ a ) = P (Sn ≤ a + 0.5)
a + 0.5 − nµ
 
≈ P Z≤ √ (8.4.9)
nσ

whenever a is a possible value of Sn . This convention is referred to as the “continuity correction”.

Version: – November 19, 2024

308 sampling distributions and limit theorems

Example 8.4.8. Two types of coin are produced at a factory: a fair coin and a biased one that
comes up heads 55% of the time. Priya is the quality control scientist at the factory. She wants to
design an experiment that will test whether a coin is fair or biased. In order to ascertain which
type of coin she has, she prescribes the following experiment as a test: Toss the given coin 1000
times, if the coin comes up heads 525 or more times conclude that it is a biased coin. Otherwise
conclude that it is fair. Factory manager Ayesha is interested in the following question: What is
the probability that Priya’s test shall reach a false conclusion for a fair coin ?
Let S1000 be the number of heads in 1000 tosses of a coin. As discussed in earlier chapters, we
1000
know that S1000 = Xi where each Xi are i.i.d. Bernoulli random variables with parameter p. If
P
i=1
the coin is fair, then p = 0.5 and E [X1 ] = 0.5, V ar [X1 ] = 0.25, and therefore E [S1000 ] = 500 and
√
SD [S1000 ] = 250 = 15.8114. We want to approximate

P (S1000 ≥ 525) = 1 − P (S1000 ≤ 524) = 1 − P (S1000 ≤ 524.5)

Without the continuity correction, we would approximate this probability as

24
 
1−P Z ≤ = 1 − P (Z ≤ 1.52)
15.8114

which can be computed using Table B.1 as 1 − 0.9357 = 0.0643, or using R as

1 - pnorm(24 / sqrt(250))

[1] 0.06452065

With the continuity correction, the approximation would instead use z = 24.5/15.8114 = 1.55 ,
giving 1 − 0.9394 = 0.0606 using Table B.1 or

1 - pnorm(24.5 / sqrt(250))

[1] 0.06062886

in R. We can also compute the exact probability that we are trying to approximate, namely
P (S1000 ≥ 525), in R as

1 - pbinom(524, 1000, 0.5)

[1] 0.06060713

As we can see, the continuity correction gives us a slightly better approximation. These calculations
tell us that the probability of Priya’s test reaching a false conclusion if the coin is fair is approximately
0.061. We shall examine the topic of Hypothesis testing, which is what Priya was trying to do, in
more detail in Chapter 10. ■

Version: – November 19, 2024

 8.4 central limit theorem 309

Example 8.4.9. We return to the Birthday problem. Suppose a small town has 1460 students.
What is the probability that five or more students were born on independence day ? Assume that
birthrates are constant throughout the year and that each year has 365 days.

The probability that any given student was born on independence day is 365 .
1
So the exact
probability that five or more students were born on independence day is

4 
1460 1 k 364 1460−k
X    
1− .
k 365 365
k =0

In Example 2.2.1 we have used the Poisson approximation with λ = 4 to estimate the above as

4 
1460 1
k 364
1460−k
X 
1−
k 365 365
k =0
42 1 1
 
≈ 1 − e−4 + 4e−4 + e−4 + 43 e−4 + 44 e−4
2 6 24
= 0.3711631

We can do another approximation using Central Limit Theorem, which is typically called the
Normal approximation. For 1 ≤ i ≤ 1460, define

1 if i-th person’s birthday is on independence day
Xi =
0 otherwise

Given the assumptions above on birthrates we know Xi are i.i.d. random variables distributed as
1460
Bernoulli( 365
1
). Note that S1460 = Xi is the number of people born on independence day

P
i=1
and we are interested in calculating
P (S1460 ≥ 5).

Observe that E (X1 ) = 365 , Var(X1 )

1
= 365 (1 − 365 )
1 1
= 364
3652
. By the Central Limit Theorem, we
know that

P (S1460 ≥ 5) = 1 − P (S1460 ≤ 4) = 1 − P (S1460 ≤ 4.5)

 
4.5 − (1460)( 365
1
)
≈ 1 − P Z ≤ q 
(1460)( 365
364
2)

0.5
= 1 − P (Z ≤ )
1.9973
= 0.401.

Recall from the calculations done in Example 2.2.1 that the exact answer for this problem is
0.3711629. So in this example, the Poisson approximation seems to work better then the Normal
approximation. This is due to the fact that more asymmetry in the underlying Bernoulli distribution

Version: – November 19, 2024

310 sampling distributions and limit theorems

worsens the normal approximation, just as it improves the Poisson approximation as we saw in
Figure 2.2. ■

exercises

Ex. 8.4.1. Suppose Sn is binomially distributed with parameters n = 200 and p = 0.3 Use the
Central Limit Theorem to find an approximation for P (99 ≤ Sn ≤ 101).
Ex. 8.4.2. Toss a fair coin 400 times. Use the Central Limit Theorem to

(a) find an approximation for the probability of at most 190 heads.

(b) find an approximation for the probability of at least 70 heads.

(c) find an approximation for the probability of at least 120 heads.

(d) find an approximation for the probability that the number of heads is between 140 and least
160.

Ex. 8.4.3. Suppose that the weight of open packets of daal in a home is uniformly distributed from
200 to 600 gms. In random survey of 64 homes, find the (approximate) probability that the total
weight of open boxes is less than 25 kgs.
Ex. 8.4.4. Let {an }n≥1 be a sequence of real numbers such that an → a as n → ∞. Show that
 an n
lim 1+ = ea .
n→∞ n

Ex. 8.4.5. Suppose U is a random variable (discrete or continuous) and MU (t) = E (etU ) exists for
all t. Show that
t2
MU (t) = 1 + tMU′ (0) + MU′′ (0) + g (t),
2
g (t)
where lim 2 = 0.
t→0 t
Ex. 8.4.6. Let X1 , X2 , . . . be a sequence of i.i.d. random variables with X1 ∼ Exp(1). Find
√ n √ !
n n X n n
lim P − √ ≤ [1 − exp(−Xi )] ≤ + √ .
n→∞ 2 2 3 2 2 3
i=1

n
nk −n
Ex. 8.4.7. Let an = , n ≥ 1. Using the Central Limit Theorem, evaluate lim an .
P
k! e n→∞
k =0
Ex. 8.4.8. How many times should you toss a coin:

(a) to be at least 90% sure that your estimate of the P(head) is within 0.1 of its true value ?

(b) to be at least 90% sure that your estimate of the P(head) is within 0.01 of its true value ?

Ex. 8.4.9. To forecast the outcome of the election in which two parties are contesting, an internet
poll via Facebook is conducted. How many people should be surveyed to be at least 95% sure that
the estimated proportion is within 0.05 of the true value?

Version: – November 19, 2024

 8.5 delta method 311

Ex. 8.4.10. A medical study is conducted to estimate the proportion of people suffering from April
allergies in Bangalore. How many people should be surveyed to be at least 99% sure that the
estimate is within 0.02 of the true value?

8.5 delta method

In many situations one is interested in knowing whether convergence properties are preserved
d
under transformations. Given random variables X1 , X2 , . . . and Z such that Xn −→ Z, and a
function g : R → R, we may be interested in knowing the limiting distribution of g (Xn ). In earlier
chapters, we have learnt techniques to calculate the distribution of g (X ) from the distribution of
X (see Section 3.3, Section 5.3), which may be helpful in studying this problem. In this section, we
discuss the Delta method, which answers this question in a specific situation, where g is a smooth
transformation that can be effectively approximated by a linear function in the region of interest.
Slutsky’s theorem (Lemma 8.3.10) is an important tool in proving the following result.

Theorem 8.5.1. Let µ ∈ R and σ ̸= 0. Let g : R → R be differentiable at µ, with g ′ (µ) ̸= 0

and g ′ (·) continuous in a neighbourhood of µ. Suppose Z ∼ Normal (0, 1) and X1 , X2 , . . . is
a sequence of random variables such that

√ (Xn − µ) d
n −→ Z as n → ∞.
σ
Then
√ (g (Xn ) − g (µ)) d
n −→ Z as n → ∞.
σg ′ (µ)

Proof. By the fundamental theorem of integral calculus, we have

Zx Z1
′
g (x) − g (µ) = g (t) dt = (x − µ) g ′ (µ + s(x − µ)) ds.
µ 0

For n ≥ 1, using the above, we have

Z1
√ (g (Xn ) − g (µ)) √ (Xn − µ) 1
n ′
= n · ′ g ′ (µ + s(Xn − µ)) ds. (8.5.1)
σg (µ) σ g (µ)
0

√ (X −µ) d
By our hypothesis on Xn , we know that n nσ −→ Z as n → ∞. Slutsky’s theorem (Lemma
8.3.10) will imply the result if we can show that

Z1
1 p
g ′ (µ + s(Xn − µ)) ds −→ 1 as n → ∞. (8.5.2)
g ′ (µ)
0

Version: – November 19, 2024

312 sampling distributions and limit theorems

Let ϵ > 0 be given. Then

Z1
 
1
P g ′ (µ + s(Xn − µ)) ds − 1 > ϵ
g ′ (µ)
0
 1 
Z
= P  g ′ (µ + s(Xn − µ)) ds − g ′ (µ) > g ′ (µ) ϵ
0
!
′ ′ ′
≤ P sup g (µ + s(Xn − µ)) − g (µ) > g (µ) ϵ (8.5.3)
s∈[0,1]

As g ′ is continuous in a neighbourhood of µ there is a δ > 0 such that for |x − µ| < δ we have

|g ′ (x) − g ′ (µ)| < |g ′ (µ)|ϵ. (8.5.4)

Using (8.5.4) in (8.5.3) we have for all n ≥ 1,

Z1
 
1
P ′ g ′ (µ + s(Xn − µ)) ds − 1 > ϵ ≤ P (|Xn − µ| > δ ) . (8.5.5)
g (µ)
0

Let M > 0 be such that

P (|Z| > M ) < ϵ. (8.5.6)

Let N ≥ 1 be such that for all n ≥ N we have

√  √ 
δ n n(Xn − µ)
<M and P >M < P (|Z| > M ) + ϵ. (8.5.7)
σ σ

Using (8.5.6) and (8.5.7) we have for all n ≥ N

√ 
√ (Xn − µ)

δ n
P (|Xn − µ| > δ ) = P n >
σ |σ|
 √ 
n(Xn − µ)
≤ P >M
σ
< P (|Z| > M ) + ϵ
< 2ϵ. (8.5.8)

Therefore substituting (8.5.8) in (8.5.5) we have for all n ≥ N ,

Z1
 
1
P ′ g ′ (µ + s(Xn − µ)) ds − 1 > ϵ < 2ϵ.
g (µ)
0

Thus we have proved (8.5.2) ■

Version: – November 19, 2024

 8.5 delta method 313

Remark 8.5.2. The particular transformation g affects the rate of convergence to normality of the
sequence g (Xn ). Indeed Theorem 8.5.1 shows that
√ d √ d
if n(Xn − µ) −→ Normal 0, σ 2 , then n(g (Xn ) − g (µ)) −→ Normal 0, σ 2 (g ′ (µ))2

The value of g ′ (µ) determines how large or small the variance of the limiting normal distribution is.

√ d
Example 8.5.3. Suppose n(Xn − µ) −→ Normal 0, σ 2 as n → ∞ for some µ ̸= 0 and σ > 0.

(a) If g : R → R is given by g (x) = x2 then by Theorem 8.5.1 we have that

√ d
n(Xn2 − µ2 ) −→ Normal 0, 4σ 2 µ2 as n → ∞.

(b) If g : R → R is given by g (x) = x1 for x ̸= 0 and g (0) = [Link] by Theorem 8.5.1 we have
that
√ 1 1 σ2
   
d
n − −→ Normal 0, 4 as n → ∞.
Xn µ µ

Note that the transformed random variables g (Xn ) need not have finite expectation for any n ≥ 1,
e.g., take Xn ∼ Normal µ, nσ with g (Xn ) = X1n as in (b).


■

8.5.1 Variance Stabilizing Transformation

A natural application of the Weak Law of Large Numbers and the Central Limit Theorem is to
estimate the unknown mean µ of a population by the sample mean X n , provided we have a random
sample from the population. Here the assumption is that E [Xi ] = µ and V ar [Xi ] = σ 2 for each
1 ≤ i ≤ n, so by the Central Limit Theorem,
√ d
n(X n − µ) −→ Normal 0, σ 2 .

In many applications, σ 2 ≡ σ 2 (µ) is a function of µ, but calculations become more convenient if

the variance of the limiting normal distribution does not depend on the parameter of interest µ.
This can often be achieved by carefully choosing a transformation g (X n ) and applying the Delta
method. More precisely, suppose we find a g : R → R such that g ′ (µ)σ (µ) = c for some c ∈ R then
Theorem 8.5.1 will imply that
√ d
n(g (X n ) − g (µ)) −→ Normal 0, c2 .

Such a transformation is called a variance stabilizing transformation. In the following example we

present three applications.

Example 8.5.4. Suppose X1 , X2 , X3 , . . . are i.i.d. X with E [X ] = p and Var[X ] = σ 2 . Let p̂ = X n .

Version: – November 19, 2024

314 sampling distributions and limit theorems

(a) Suppose X ∼ Bernoulli(p). Then σ 2 = p(1 − p), and we need a transformation such that
√
g ′ (p) p(1 − p) = c for some c ∈ R. Indeed if g (x) = arcsin( x) then
p

√ 1
 
d
n(g (X n ) − g (p)) −→ Normal 0, .
4

√
(b) Suppose X ∼ Poisson(p). Then σ 2 = p, and we need a transformation such that g ′ (p) p = c
√
for some c ∈ R. Indeed if g (x) = x then

√ 1
 
d
n(g (X n ) − g (p)) −→ Normal 0, .
4

(c) Suppose X ∼ Normal 0, σ 2 and we are interested in estimating σ 2 . We will then use


n
1 P
n Xi2 as the estimate for σ 2 and the Central Limit Theorem implies that
i=1

n
!
√ 1X 2 d
Xi − σ 2 −→ Normal 0, 2σ 4 .


n
n
i=1
√
Thus we need a transformation such that g ′ (σ 2 ) 2σ = c for some c ∈ R. Indeed if
g (x) = log x then
n
! !
√ 1X 2 d
n g Xi − g (σ ) −→ N(0, 2).
2
n
i=1

exercises

Ex. 8.5.1. Prove Lemma 8.3.10 (a) and (b).

Ex. 8.5.2. Supppose {Xn }n≥1 , X are a sequence of random variables and {an }n≥1 , a are a sequence
d d
of real numbers such that Xn −→ X and an → a then show that an Xn −→ aX. Hint: Use Lemma
8.3.10.

Ex. 8.5.3. Supppose {Xn }n≥1 , X and {Yn }n≥1 , Y are a sequence of random variables such that
d d d
Xn −→ X and Yn −→ Y . Show that λXn + (1 − λ)Yn −→ λX + (1 − λ)Y . Hint: Use Lemma
8.3.10.
d d
Ex. 8.5.4. Suppose Xn −→ X. Show that Xn2 −→ X 2 .

Ex. 8.5.5. Let α, µ > 0. Let {Xi }i≥1 be i.i.d. random variables following Pareto (α, µ) distribution.
That is, the probability density function of Xi , for any i ≥ 1, is given by

αµα
(
xα + 1
x ≥ µ,
f(α,µ) (x) =
0 otherwise.

Version: – November 19, 2024

 8.5 delta method 315

n  
Xi
(a) Let Y n = 1
log . Show that
P
n µ
i=1

√ 1
 
d
−→ Normal 0, α2 .


n −α
Yn

n
(b) Let Z n = 1
log(Xi ) and Mn = max{X1 , X2 , . . . , Xn }.
P
n
i=1

√ p
(i) Show that n(log(Mn ) − log(µ)) −→ 0.
(ii) Using (a) and Lemma 8.3.10 show that

√ 1 1
   
d
n Z n − log(Mn ) − −→ Normal 0, 2 .
log(µ) α

(iii) Show that

√ 1
 
d
−→ Normal 0, α2 .


n −α
Z n − log(Mn )

Ex. 8.5.6. For α, µ > 0, let {Xi }i≥1 be i.i.d. random variables with the probability density function
of Xi , for any i ≥ 1, given by
(
µe−µ(x−α) x ≥ α,
f(α,µ) (x) =
0 otherwise.

n
Let X n = 1
Xi and Mn = max{X1 , X2 , . . . , Xn }.
P
n
i=1

(a) Show that E [X1 ] = α + 1

µ and Var[X1 ] = 1
µ2
.
√ p
(b) Show that n(Mn − α) −→ 0.
√ d
 
(c) Show that n(X n − Mn − µ1 ) −→ Normal 0, µ12 .
√ d
(d) Show that 1
−→ Normal 0, µ2 .


n( )
X n −Mn −µ

n
Ex. 8.5.7. Let Xi , i ≥ 1 be i.i.d. Bernoulli(p) random variables. Let X n = 1
Xk . Show that
P
n
k =1

√
   
Xn p d p
n − −→ Normal 0, .
1 − Xn 1 − p (1 − p)3

p
The statistic Xn
is typically used to estimate the odds ratio 1−p .
1−X n
n
Ex. 8.5.8. Let Xi , i ≥ 1 be i.i.d. Bernoulli(p) random variables. Let X n = 1
Xk . Show that
P
n
k =1

√ p(1 − p)
 
d
n X n (1 − X n ) − p(1 − p) −→ Normal 0, ,


(1 − 2p)2

Version: – November 19, 2024

316 sampling distributions and limit theorems

for p ̸= 12 . The statistic X n (1 − X n ) is typically used to estimate the variance p(1 − p).
n
Ex. 8.5.9. Let Xi , i ≥ 1 be i.i.d. Exp(λ) random variables. Let X n = 1
Xk . Show that
P
n
k =1

√ 1 1 1
   
d
n − −→ Normal 0, 2 .
Xn λ λ

Ex. 8.5.10. Consider the same set up as in Example 8.5.3 with µ = 0. Then show that
√ p
nXn2 −→ 0
√ 2 d
as n → ∞, and that the correct scaling is n as opposed to n, that is, n X
σ2
n
−→ χ21 as n → ∞.

8.6 limiting distribution of sample median

The sample median is a natural alternative to the sample mean as a measure of centrality. For
continuous symmetric distributions, the median is the same as the mean when the mean exists. The
median of a distribution always exists, even if the mean does not. It is invariant under monotone
transformations, making it a more appealing measure of centrality for skewed distributions. The
asymptotic distribution of the sample median is therefore of natural interest.
The Central Limit Theorem establishes the Normal distribution as the limiting distribution
of the standardized sample mean for all distributions that have finite second moment. This is a
universality result which says that all sums and averages from a random sample are asymptotically
Normal. However, from a sample one can derive many other summary statistics, each with different
sampling distributions, and it is natural to ask about their asymptotic behaviour. In this section,
we show that the limiting distribution of the sample median is also Normal. We do this in two
stages. First we prove it for Uniform(0, 1) random variables, and then use the Delta method to
prove it for more general distributions.

Lemma 8.6.1. Suppose that U1 , U2 , . . . are i.i.d. Uniform(0, 1) random variables, and let U
en be
the sample median obtained from U1 , . . . , Un . Then,

en − 1
√
 
d
2 n U −→ Z, (8.6.1)
2

where Z has a standard Normal distribution.

Proof. To begin with and to keep the definition of the median unambiguous, we consider odd
samples sizes such that n = 2k − 1 for some positive integer k and we shall let k → ∞. In this case
the median Uen = U .
(k )
As seen in Example 8.1.4, U(k) has the Beta(k, k ) distribution with density

 k 2k uk−1 (1 − u)k−1
  
0 < u < 1,

fk ( u ) = 2 k
0

otherwise.

Version: – November 19, 2024

 8.6 limiting distribution of sample median 317

−3 −2 −1 0 1 2 3

5 11 19
0.4

0.3
Density

0.2

0.1

0.0

−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3

Figure 8.7: Density of standardized sample median for a Uniform(0, 1) population, for sample
sizes 5, 11, and 19. The grey curve represents the standard Normal density. As the
sample size increases, the density function of the standardized median converges
to the standard Normal density.

It is easy to verify that E [U(k) ] = 1

2 and Var[U(k) ] = 1
4(2k +1)
. It follows that the density of the
standardized median
√ 1
 
Zk = 2 2k + 1 U(k) −
2
is given, after simplification, by
 k−1
2k! z2 √ √

k
√ 1− if − 2k + 1 < z < 2k + 1,


gk (z ) = 4k 2k + 1 k!2 2k + 1 (8.6.2)
0 otherwise.



Figure 8.7, which plots gk (z ) for k = 3, 6, 10 which correspond to n = 5, 11, 19 shows that the
Normal approximation is good even for small values of n.
By Stirling’s approximation for factorials we have that for all k ≥ 1

√ √ 1
 
k k 2πk exp (−k ) < k! < k k 2πk exp −k + . (8.6.3)
12k

Therefore, using the bounds provided in (8.6.3) in (8.6.2) we have

√ k−1 √ k−1
1 z2 1 z2
 
1
− 6k k 1 k
e √ 1− ≤ gk (z ) ≤ e 6k q √ 1− . (8.6.4)
2π 2k 1 2π 2k + 1
q
k+ 1 + k+ 1
2 2

Using the facts that

√
(a) lim p k
=1
k→∞ k + 21

Version: – November 19, 2024

318 sampling distributions and limit theorems

(b) lim exp( 6k

1
) = lim exp(− 6k
1
)=1
k→∞ k→∞
h ik−1 1 2
z2
(c) for all z ∈ R, lim 1 − 2k +1 = e− 2 z
k→∞

in (8.6.4) we obtain for all z ∈ R

1 1 2
gk (z ) → √ e− 2 z as k → ∞. (8.6.5)
2π

Theorem 8.3.5 then implies that

d
Zk −→ Normal (0, 1) as k → ∞. (8.6.6)
√ e √
n
As n(U n − 2 ) = √n+2 Z n+1 , an application of Lemma 8.3.10 (see Exercise 8.5.2) yields the result
1
2
when n is odd.
For n = 2k even, the median may be defined as any value between U(k) and U(k+1) , i.e., any
U +U
convex combination of U(k) and U(k+1) . For example, we could take (k) 2 (k+1) . Imitating the
above argument in n odd case one can prove the result in this case as well (see Exercise 8.6.1). ■

Having established the result for the special case of Uniform(0, 1), we next wish to prove that a
similar result holds for a much wider class of distributions. For this, we will use the Delta method,
generalizing the result to all continuous distributions which have strictly positive density at the
population median.

Theorem 8.6.2. Let X1 , X2 , . . . be i.i.d. random variables with probability density function
en be the sample median obtained from X1 , X2 , . . . , Xn . Assume that f (µ) > 0,
f . Let X
where µ denotes the median of the random variable X. Then,
√  
d
2 nf (µ) Xen − µ −→ Z, (8.6.7)

where Z has a standard Normal distribution.

Proof. Define Ui = F (Xi ) for i ≥ 1. By Lemma 5.3.7, Exercise 5.3.12 and Theorem 8.1.2, U1 , U2 , . . .
are i.i.d. Uniform(0, 1). By Lemma 8.6.1

en − 1 −→
√
 
d
2 n U Z, (8.6.8)
2

where Z is standard Normal and Uen is the median of U1 , . . . , Un . Let F be the distribution function
of X. Now define G : [0, 1] → R ∪ {−∞} ∪ {∞} as

G(u) = inf{x ∈ R : F (x) ≥ u}.

Recall from Exercise 5.3.12 that G is the generalised inverse of F . First note, since Xi are sampled
from f , Xi = G(Ui ) for i ≥ 1 with probability 1. By definition G is increasing, and so X
en = G(Uen ).

Version: – November 19, 2024

 8.6 limiting distribution of sample median 319

Further, since f (µ) > 0 then F is strictly monotone and F ′ (exists, is continuous) is strictly positive
in a neighbourhood of µ. This will imply G( 12 ) = µ, G is differentiable at µ with G′ ( 12 ) = f (1µ) > 0
and G′ (·) is continuous in a neighbourhood of µ.
As G satisfies the hypothesis of Theorem 8.5.1, using (8.6.8) we have

en ) − G( 1 ) d
√ G(U 2
n 1 ′ 1 −→ Z,
2 G ( 2 )

with Z being standard Normal. The result follows. ■

We conclude this section with a couple of examples.

Example 8.6.3. Suppose that X1 , X2 , . . . are i.i.d. Normal(µ, σ 2 ) random variables, and let X
en
be the sample median obtained from X1 , . . . , Xn . Then, it follows from Theorem 8.6.2 that

2 (X
r
√ en − µ) d
n −→ Z, (8.6.9)
π σ

where Z has a standard Normal distribution.  

Note that although the sample mean X n ∼ Normal µ, √σn , the distribution of the sample
median Xen , which can be calculated from (8.1.3), is not Normal. Asymptotically however, the
standardized sample median also has a limiting Normal distribution. One can compare their
asymptotic efficiency in terms of limiting variances. As noted, we have

2
r
√ (X
en − µ) d √ (X n − µ) d
n −→ Z and n −→ Z (8.6.10)
σ π σ

The asymptotic efficiency of X en over X n is defined as the inverse of the ratio of the limiting
variances, i.e., π2 ≈ 0.64. This number can be interpreted in the following manner: If one uses Xen
to estimate µ, then one could instead have used use X m with sample size m ≈ 0.64n to get an
estimator with the same variance. ■
The previous example suggests that one should use the sample mean rather than the sample
median when the population is Normal. In practice, however, the underlying distribution can be
rarely known with certainty, and we will see in Chapter 9 that under even fairly mild departures
from Normality, the sample median may become much more useful than the sample mean in the
sense of asymptotic efficiency as defined above. An extreme case is the following example, where
the sample mean does not have finite variance, and hence the asymptotic efficiency of the sample
mean over the sample median is zero.
Example 8.6.4. Suppose that X1 , X2 , . . . are i.i.d. Cauchy(θ, α2 ) random variables. It is easy to
see (by symmetry) that θ is the median. As the Cauchy distribution has no finite moments, one
way to estimate θ is using the sample median. We can apply Theorem 8.6.2 to get

√ 1
2 n en − θ ) → Z
(X as n → ∞,
πα
where Z is standard Normal. ■

Version: – November 19, 2024

320 sampling distributions and limit theorems

exercises

Ex. 8.6.1. Suppose that U1 , U2 , . . . are i.i.d. Uniform(0, 1) random variables, and let U
en be the
sample median obtained from U1 , . . . , Un , with n = 2k for some k ≥ 1.

(a) Using Example 8.1.4 find the distribution of U(k) and U(k+1) and Compute E [U(k) ], Var[U(k) ],
E [U(k+1) ] and Var[U(k+1) ].

(b) As in the proof of Lemma 8.6.1 find ak and bk such that both Zk = ak (U(k) − 12 ), Z
ek =
bk (U(k+1) − 21 ) converge in distribution to Normal (0, 1) as k → ∞.
√ d
(c) Using Lemma 8.3.10 For 0 < λ < 1 show that n(λUk + (1 − λ)Uk+1 − 12 ) −→ Normal (0, 1)
as k → ∞.

Ex. 8.6.2. Suppose that X1 , X2 , . . . , Xn are i.i.d. Exp(λ) random variables. Find the distribution
of sample median X en and also identify the standardization required to obtain a Normal distribution
as the limiting distribution of the standardized sample median.
Ex. 8.6.3. Under the conditions of Theorem 8.6.2, show that the sample median converges to the
population median in probability. Thus, the Weak Law of large numbers holds for the sample
median.

Version: – November 19, 2024

E S T I M AT I O N
9
In Chapter 7 we discussed how an i.i.d. sample X1 , X2 , . . . , Xn from an unknown distribution may
be used to estimate aspects of that distribution. In Chapter 8 we saw how some sample statistics
behave asymptotically. In this chapter we look at some specific examples where various parameters
of the distribution such as the mean µ and the standard deviation σ are unknown, and the sample
statistics that are used to estimate these parameters.
For instance, suppose there is a coin which we assume has a probability p of showing heads
each time it is flipped. To gather information about p the coin is flipped 100 times. The results of
these flips are viewed as i.i.d. random variables X1 , X2 , . . . , X100 with a Bernoulli(p) distribution.
100
Suppose Xn = 67, meaning that 67 of the 100 flips showed heads. How might we use this
P
n=1
information to infer something about the value of p?
The first two topics we will consider are the “method of moments” and the “method of maximum
likelihood”. Both of these are direct forms of estimation in the sense that they produce a single-value
estimate for p. A benefit of such methods is that they produce a unique prediction, but a downside
is that the prediction they make is most likely not exactly correct. These methods amount to a
statement like “Since 67 of the 100 flips came up heads, we predict that the coin should come
up heads 67% of the time in the long run”. In some sense the 67% prediction may be the most
reasonable one given what was observed in the 100 flips, but it should also be recognised that 0.67
is unlikely to be the true value of p.
Another potential approach is that of the “confidence interval”. In using this method we
recognise that a specific estimate is unreasonable, and instead produce a range of values which is
expected to contain the true value of the parameter. This approach could yield a statement such
as this: “With 90% confidence, the actual probability that the coin will show heads is between
0.59 and 0.75”. Of course, the true p is not random and will either lie between 0.59 and 0.75 or it
will not; there is nothing probabilistic about that event. The 90% confidence here refers to our
professed confidence in the procedure, in the sense that we believe that the procedure produces a
“correct” interval with probability 0.9.
Yet another approach is based on the idea of a “hypothesis test.” In this case we make a
conjecture about the value of the parameter and carry out a computation to test the credibility of
the conjecture. There is an obvious link between hypothesis tests and confidence intervals: we can
define a confidence interval to consist of values of the parameter that are credible according to a
test, and vice versa. We will discuss the hypothesis testing approach in the next chapter.
For the purposes of this chapter, we will assume that the sample X1 , X2 , . . . , Xn are i.i.d. copies
of a random variable X with a probability mass function or probability density function f (x). For
brevity, we shall often refer to the distribution X, by which we will mean the distribution of the
random variable X. We shall further assume that f (x) depends on one or more unknown parameters
p1 , p2 , . . . , pd and emphasise this using the notation f (x | p1 , p2 , . . . , pd ). We may abbreviate this

321

Version: – November 19, 2024

322 estimation

as f (x | p), where p = (p1 , p2 , . . . , pd ) represents the vector of all the parameters. We will assume
that the set P of all possible values p can take is known, where P ⊂ Rd for some d ≥ 1. The set P
may be all of Rd or some proper subset depending on the nature of the parameters.
We now fix some notations and terminology for estimators.

Definition 9.0.1. Let X1 , X2 , . . . , Xn be an i.i.d. sample from a population with distribution

f (x | p). Suppose we are interested in estimating θ (p) for some θ : P → R. Then
g (X1 , X2 , . . . , Xn ) for any g : Rn → R can be considered as a “point estimator” of θ (p),
and its value from a particular realisation is called an “estimate”.

In practice the function g is chosen keeping in mind the parameter θ (p) of interest. We have seen
the following in Chapter 7.

Example 9.0.2. Let µ = E [X ]. Let g : Rn → R be given by

n
1X
g (x) = xi .
n
i=1

Then g (X1 , X2 , . . . , Xn ) is the (now familiar) sample mean and it is an estimator for µ. Further,
E [g (X1 , X2 , . . . , Xn )] = µ regardless of the true value of µ. We called such an estimator an
unbiased estimator. Finally we also know by the strong law of large numbers, Theorem 8.2.1, that
p
g (X1 , X2 , . . . , Xn ) −→ µ as n → ∞. ■

Recall from Chapter 6 that E [X ] is the first moment of X. As noted in Chapter 7, we can thus
view the sample mean, which is the first moment of the empirical distribution based on a sample,
as estimating the first moment of the underlying distribution. A generalization of this method is
known as the method of moments.

9.1 method of moments

For n ≥ 1, let X1 , X2 , . . . , Xn be a sample from a population with distribution X. Assume that

X has either probability mass function or probability density function f (x | p) depending on
parameter(s) p = (p1 , p2 , . . . , pd ), for some d ≥ 1. Let for k ≥ 1, mk : Rn → R be given by

n
1X k
mk (x) = xi .
n
i=1

Notice that mk (X1 , X2 , . . . , Xn ) is the k-th moment of the empirical distribution based on the
sample X1 , X2 , . . . , Xn , which we will refer to simply as the k-th sample moment.
Let µk = E [X k ], the k-th moment of the distribution X. As the distribution of X depends on
(p1 , p2 , . . . , pd ) one can view µk ≡ µk (p1 , p2 , . . . , pd ) as a function of p. The method of moments

Version: – November 19, 2024

 9.1 method of moments 323

estimator for (p1 , p2 , . . . , pd ) is obtained by equating the first d sample moments to the corresponding
moments of the distribution. Specifically, it requires solving the d equations in d unknowns given by

µk (p1 , p2 , . . . , pd ) = mk (X1 , X2 , . . . , Xn ) , k = 1, 2, . . . , d.

for p1 , p2 , . . . , pd . There is no guarantee in general that these equations have a unique solution or
that it can be computed, but in practice it is often possible to do so. The solution will be denoted
by p̂1 , p̂2 , . . . , p̂d which will be writen in terms of the realised values for mk , k = 1, 2, . . . , d. We will
now explore this method for two examples.

Example 9.1.1. Suppose X1 , X2 , . . . , X10 is an i.i.d. sample with distribution Binomial(N , p)

where neither N nor p is known. Suppose the empirical realisation of these variables is 8, 7, 6, 11,
8, 5, 3, 7, 6, 9. One can check that the average of these values is m1 = 7 while the average of their
squares is m2 = 53.4. Since X ∼ Binomial (N , p) the probability mass function is given by
 
N k
f (k | N , p) = p (1 − p)N −k , k = 0, 1, . . . , N .
k

We have previously shown that

E [X ] = N p and E [X 2 ] = Var[X ] + E [X ]2 = N p(1 − p) + N 2 p2 .

Thus, the method of moments estimator for (N , p) is obtained by solving

7 = m1 = N̂ p̂ and 53.4 = m2 = N̂ p̂(1 − p̂) + N̂ 2 p̂2 .

Using elementary algebra we see that

m21
N̂ = ≈ 19
m1 − (m2 − m21 )
m1 − (m2 − m21 )
p̂ = ≈ 0.371.
m1

Thus, according to the method of moments, the distribution from which the sample came from is
estimated to be the Binomial(19, 0.371) distribution. In practice, we usually wish to restrict the
estimates of the parameters based on the context of the problem. Since the N value is surely some
integer, the estimate of N̂ was rounded to the nearest meaningful value in this case. ■

Example 9.1.2. Suppose our distribution of interest X has a Normal (µ, σ 2 ) distribution. Therefore
our probability density function is given by

1 (x−µ) 2
−
f (x | µ, σ 2 ) = √ e 2σ2 , x ∈ R.
2πσ

Let X1 , X2 , . . . , Xn be an i.i.d. sample from the distribution X. We have shown that

E [X ] = µ and E [X 2 ] = Var[X ] + E [X ]2 = µ2 + σ 2 .

Version: – November 19, 2024

324 estimation

The method of moments estimator for µ, σ is found by solving

m1 = µ and m2 = µ2 + σ 2 .

from which

µ̂ = m1 = X and
n
!
1X 2 2 n−1 2
σ̂ 2
= m2 − m21 = Xi −X = S .
n n
i=1

Here X and S 2 are, respectively, the sample mean and sample variance defined in Chapter 7. ■
The method of moment estimators may not always be very reliable, in the sense that it might
give implausible estimates. For instance, in Example 9.1.1 above, you can check that the estimate
for p would be negative if the sample mean X happened to be smaller than n−1 n S . Such defects
2

can be somewhat rectified using moment matching and other techniques (see [CasBer90]).

exercises

Ex. 9.1.1. Suppose X1 , . . . , X5 is an i.i.d. sample with Uniform(a, b) distribution for some unknown
a and b. Suppose the empirical realisation of these variables is 3.5, 2.1, 5.7, 4.8, 3.9. Use the method
of moments to estimate a and b.
Ex. 9.1.2. Suppose X1 , X2 , . . . , Xn is an i.i.d. sample with Uniform(a, b) distribution for some
unknown a and b. Let m1 and m2 be the empirical realisation of the first and second moments of
the X1 , X2 , . . . , Xn data. Find an expression for the estimates of a and b given by the method of
moments in terms of the quantities m1 and m2 .
Ex. 9.1.3. Suppose X1 , X2 , . . . , Xn is an i.i.d. sample with Uniform(a, b) distribution for some
unknown a and b. Prove that the method of moments produces estimates â and b̂ such that â = b̂
if and only if every data point in the empirical realisation has exactly the same value.
Ex. 9.1.4. Suppose X1 , . . . , X4 is an i.i.d. sample with Binomial(N , p) distribution for some
unknown N and p. Suppose the empirical realisation of these variables is 1, 2, 5, 12. Show that the
method of moments for estimating N and p gives negative (and therefore meaningless) results.
Ex. 9.1.5. Suppose X1 , X2 , . . . , Xn is an i.i.d. sample with Binomial(N , p) distribution for some
unknown N and p. Prove that the method of moments will produce a negative estimate for p if an
only if it also produces a negative estimate for N .
Ex. 9.1.6. Suppose X1 , . . . , X6 is an i.i.d. sample with Gamma(α, λ) distribution for some unknown
α and λ. Suppose the empirical realisation of these variables is 5.3, 2.4, 2.8, 7.6, 6.9, 4.2. Use the
method of moments to estimate α and λ.
Ex. 9.1.7. Suppose X1 , X2 , . . . , Xn is an i.i.d. sample with Gamma(α, λ) distribution for some
unknown α and λ. Let m1 and m2 be the empirical realisations of the first and second moments of
X1 , X2 , . . . , Xn .

Version: – November 19, 2024

 9.2 maximum likelihood 325

(a) Find an expression for the estimates of α and λ given by the method of moments in terms of
the quantities m1 and m2 .

(b) Show that the estimates of α and λ from part (a) can never be negative.

Ex. 9.1.8. The following code simulates 100 samples from a population with distribution Binomial
(20, 0.4) and computes the method of moments estimate for the sample size parameter n and the
success probability p (n = 20 and p = 0.4 in the simulation).

n <- 100
N <- 20
p <- 0.4
x <- rbinom(n, N, p)
m1 <- mean(x)
m2 <- mean(xˆ2)
Nhat <- m1ˆ2 / (m1 -(m2-m1ˆ2))
phat <- (m1 - (m2-m1ˆ2)) / m1

(a) Run the code in R and compute phat - p and Nhat - N.

(b) Change the code suitably to simulate 1000 samples from Binomial (20, 0.4) and see if the
answer to (a) changes.

(c) Change the code suitably to simulate samples from Binomial (10, 0.1) and Binomial (10, 0.9),
and repeat (a) and (b).

Ex. 9.1.9. Using the method from Exercise 9.1.2, and by suitably modifying the R-code in Exercise
9.1.8, write R-code that computes the method of moments estimate for a and b in Uniform(a, b)
when a = 3 and b = 5 by generating 100 samples from Uniform(3, 5)

Ex. 9.1.10. Using the method from Example 9.1.2, and by suitably modifying the R-code in Exercise
9.1.8, write R code that computes the method of moments estimate for µ and σ 2 in Normal(µ, σ 2 )
when µ = 4 and σ 2 = 10 by generating 100 samples from Normal(4, 10)

Ex. 9.1.11. Using the method from Exercise 9.1.6, and by suitably modifying the R code in Exercise
9.1.8, write R code that computes the method of moments estimate for a and b in Gamma (a, b)
when a = 10 and b = 0.5 by generating 100 samples from Gamma (10, 0.5)

9.2 maximum likelihood

For n ≥ 1, let X1 , X2 , . . . , Xn be an i.i.d. sample from the distribution X. Assume that X has
either probability mass function or probability density function denoted by f (x | p) depending on
parameter(s) p = (p1 , p2 , . . . , pd ) ∈ P ⊂ Rd .

Version: – November 19, 2024

326 estimation

Definition 9.2.1. The “likelihood function” for the sample X1 , X2 , . . . , Xn is the function
L : P × Rn → R given by
n
Y
L(p; X1 , X2 , . . . , Xn ) = f (Xi | p).
i=1

For a given sample X1 , X2 , . . . , Xn , suppose p̂ ≡ p̂(X1 , X2 , . . . , Xn ) is a point at which

L(p; X1 , X2 , . . . , Xn ) attains its maximum as a function of p. Then p̂ is called a “maximum
likelihood estimator” of p (abbreviated as MLE of p) given the sample X1 , X2 , . . . , Xn .

One observes readily that the likelihood function is the joint density or joint mass function of
(X1 , X2 , . . . , Xn ) when p is fixed. Assuming that the MLE p̂ as defined above is unique, it can be
thought of as the most “likely” value of the parameter p for the given realisation of X1 , X2 , . . . , Xn ,
as for any other parameter value pe, the corresponding joint density or joint mass function has a
lower value at (X1 , X2 , . . . , Xn ). If p̂ is not unique, the same is true for any pe which is not an MLE.

Example 9.2.2. Let p ∈ R and X1 , X2 , . . . , Xn be an i.i.d. sample from a population distributed

as Normal with mean p and variance 1. Then the likelihood function is given by
n
P
n − 12 (Xi −p)2
Y 1 − (Xi −p)2 1
L(p; X1 , X2 , . . . , Xn ) = √ e 2 = √ n e i=1 .
i=1
2π ( 2π )

To find the MLE, treating the given the realisation X1 , X2 , . . . , Xn as fixed, one needs to maximise
L as a function of p. Noting that maximising L is equivalent to maximising loge L (as logarithm is
an increasing function), the problem is then to find the minimum of g : R → R given by
n
X
g (p) = (Xi − p)2 .
i=1

n
Method 1: Since g (p) = (Xi − X )2 + (X − p)2 (see Exercise 9.2.2) and the first term does not
P
i=1
depend on p, the minimum of g will occur at p̂ = X.
Method 2: An alternative approach is to use differential calculus. As g is a quadratic function of p,
it is differentiable at all p, and
n
X
g ′ (p) = −2 (Xi − p) and g ′′ (p) = 2n.
i=1

n
As g ′′ (·) > 0, the minimum will occur when g ′ (p) = 0. This occurs when p is equal to 1
Xi . So
P
n
i=1
the MLE of p is given by p̂ = X. ■

Version: – November 19, 2024

 9.2 maximum likelihood 327

Example 9.2.3. Let p ∈ (0, 1) and X1 , X2 , . . . , Xn be an i.i.d. sample from a population distributed
as Bernoulli(p). The probability mass function f can be written as

 p if x = 1 (
px (1 − p)1−x if x ∈ {0, 1}

f (x | p) = 1−p if x = 0 =
0 otherwise.
0 otherwise.



Then the likelihood function is given by

 n
  n

P P
n
Y Xi n− Xi
L(p ; X1 , X2 , . . . , Xn ) = pXi (1 − p)1−Xi = p i=1 (1 − p) i=1 .
i=1

To find the MLE, treating the given the realisation X1 , X2 , . . . , Xn as fixed, one needs to maximise
L as a function of p. We can use calculus to do this, but differentiating L is cumbersome, so as
before we look at loge L, which is called the log likelihood function.

ℓ(p ; X1 , X2 , . . . , Xn ) = loge L(p ; X1 , X2 , . . . , Xn )

n n
   P
 log p
Xi ) + n loge (1 − p) if 0 <
P
e 1−p ( Xi < n


i=1 i=1


n


n loge (1 − p) if Xi = 0
P
=
 i=1
n


 n loge (p) if
 P

 Xi = n
i=1

n
As Xi is fixed for the purpose of this maximisation problem, we can approach the problem
P
i=1
separately for the three cases above.

n
In the first case with 0 < Xi < n, we can re write
P
i=1

n
X n
X
ℓ(p ; X1 , X2 , . . . , Xn ) = loge (p)( Xi ) + (n − Xi ) loge (1 − p).
i=1 i=1

It is easy to see that

n n
1 X 1 X
ℓ′ (p ; X1 , X2 , . . . , Xn ) = ( Xi ) − (n − Xi )
p 1−p
i=1 i=1

and
n n
1 X 1 X
ℓ′′ (p ; X1 , X2 , . . . , Xn ) = − ( Xi ) + ( n − Xi ).
p2 (1 − p)2
i=1 i=1
n
As 0 < Xi < n, ℓ′′ (p ; X1 , X2 , . . . , Xn ) < 0 for all p. So the global maximum will occur at the
P
i=1
n
point where ℓ′ (p ; X1 , X2 , . . . , Xn ) = 0. This happens when p = 1
Xi .
P
n
i=1

Version: – November 19, 2024

328 estimation

n
In the second case with Xi = 0, ℓ is a decreasing function of p and the maximum occurs at
P
i=1
n
p = 0 which can be trivially re-written as p = 1
Xi in this case.
P
n
i=1
n
In the third case with Xi = n, ℓ is an increasing function of p and maximum occurs at p = 1
P
i=1
n
which can be trivially re-written as p = 1
Xi in this case.
P
n
i=1
n
Combining the three cases, we can conclude that the MLE of p, p̂ = 1 P
n Xi = X. ■
i=1

At times we may wish to maximize the likelihood as a function of a parameter that takes values in
a discrete set. Consider a collection of empirical measurements of waiting times. Suppose we know
that each waiting time is the sum of some fixed number of i.i.d. Exp(λ) distributions, but we are
not certain how many such distributions are in each sum. We might let m represent that unknown
number, and attempt to find the m which maximizes the likelihood. As we have previously seen,
such sums have a Gamma (m, λ) distribution. In the example below we will assume λ is known.

Example 9.2.4. Let λ > 0 and let m be an unknown positive integer. Let X1 , X2 , . . . , Xn be an
i.i.d. sample from a population distributed as Gamma(m, λ). Then the likelihood function is given
by
n
Y λm
L(m) = L(m ; X1 , . . . , Xn ) = X m−1 e−λXi .
(m − 1) ! i
i=1

Now consider the ratio

n
L(m + 1) λn Y
= n Xi .
L(m) m
i=1

This ratio is a decreasing function of m, so L(m) is maximimized at the smallest value of m for
which this ratio is less than 1. Therefore
1/nthe maximum likelihood estimate for m is the smallest
1/n
n
 n
integer which is larger than λ . The quantity is known as the “geometric
Q Q
Xi Xi
i=1 i=1
mean” of the X1 , X2 , . . . , Xn values. ■

As a final example, let us revisit Example 9.1.1, where we considered a Binomial distribution with
both parameters unknown.

Example 9.2.5. Suppose X1 , X2 , . . . , Xn is an i.i.d. sample with distribution Binomial(N , p) where

neither N nor p is known. The likelihood function is given by
n  
Y N
L(N , p ; X1 , X2 , . . . , Xn ) = pXi (1 − p)N −Xi .
Xi
i=1

To obtain MLEs for N and p, we need to maximise this expression as a function of N and p, for a
fixed set of empirical observations X1 , X2 , . . . , Xn .
This is unfortunately not easy to do explicitly. We can simplify the problem by observing that
n
we have already calculated an estimate of p if N is known. In that case, Xi is the sum of N n
P
i=1

Version: – November 19, 2024

 9.2 maximum likelihood 329

independent Bernoulli(p) random variables, for which the MLE of p is (see Example 9.2.3 and
Exercise 9.2.9)
Pn
Xi
i=1
p̂ = .
Nn
By plugging in this estimator in the expression for L(N , p ; X1 , X2 , . . . , Xn ), we obtain the so called
“profiled likelihood function”

e (N ) ≡ L(N , p̂ ; X1 , X2 , . . . , Xn ),
L

which can now be viewed as a function of N only. Such profiled likelihood functions, where some
parameters in the likelihood function are replaced by estimators that depend on the remaining
parameters, are useful because they reduce the number of parameters over which the maximization
problem needs to be solved. It is easy to see that maximizing the profiled likelihood is equivalent
to maximizing the original likelihood function.

Unfortunately, further theoretical analysis of this function is difficult. Numerically, however, this
problem is not difficult to solve. Consider again the empirical realisations given in Example 9.1.1,
with n = 10 and observations 8, 7, 6, 11, 8, 5, 3, 7, 6, 9. Clearly, N must be at least 11, the largest
of the observations. Let us use R to compute the logarithm of the profiled likelihood for values of
N from 11 to 50.

x <- c(8, 7, 6, 11, 8, 5, 3, 7, 6, 9)

n <- length(x)
N <- 11:50
phat <- sum(x) / (N * n)
logL <- rep(0, length(N))
for (i in seq(1, length(N)))
logL[i] <- sum(dbinom(x, size = N[i],
prob = phat[i], log = TRUE))
d <- [Link](N = N,
phat = phat,
logL = logL)

We can now plot these log-likelihood values, as we do in Figure 9.1, or look at them directly.

head(d, 15)

Version: – November 19, 2024

330 estimation

−21.6

−21.8

−22.0
logL

−22.2

−22.4

−22.6

−22.8

10 20 30 40 50

Figure 9.1: Profiled log-likelihood for the Binomial(N , p) distribution as a function of N .

N phat logL
11 0.6363636 -22.77295
12 0.5833333 -22.14081
13 0.5384615 -21.86794
14 0.5000000 -21.73123
15 0.4666667 -21.65932
16 0.4375000 -21.62190
17 0.4117647 -21.60412
18 0.3888889 -21.59799
19 0.3684211 -21.59894
20 0.3500000 -21.60426
21 0.3333333 -21.61225
22 0.3181818 -21.62184
23 0.3043478 -21.63233
24 0.2916667 -21.64328
25 0.2800000 -21.65438

By inspecting the first few rows of the table, we see that the likelihood is maximized at N̂ = 18
and p̂ = 0.389. These estimates are not very different from the ones we obtained using the method
of moments. ■
Similar numerical methods are required to compute the maximum likelihood estimate in many
other examples (See Exercise 9.2.10).
Remark 9.2.6. In general, one may not be able to compute the sampling distribution (i.e. probability
distribution of the random-sample-based statistic) of the maximum likelihood estimate. However
there is a well-understood theory of limiting distributions of maximum likelihood estimators, whether

Version: – November 19, 2024

 9.2 maximum likelihood 331

they are available in closed form solutions or not. They are “better” than other estimators in terms
of variance, and follow a normal distribution, asymptotically. A detailed discussion and proof of
these results are beyond the scope of this book.
Sampling distributions of these estimates do play an important role in obtaining confidence
interval (see Section 9.3) and test of hypotheses (see Chapter 10). For this one must understand
the limiting behaviour of the sampling distributions.

exercises

Ex. 9.2.1. In the examples above we have used the fact that an exponential with negative exponent
may be maximized by minimizing the exponent. Prove that this is generally true. Suppose f (x) is
a function which achieves a minimum when x = a. Let g (x) = e−f (x) . Prove that g (x) achieves a
maximum when x = a.
Ex. 9.2.2. Show that for any real numbers p, x1 , x2 , . . . , xn ,
n
X n
X
( xi − p ) 2 = ( xi − x ) 2 + ( x − p ) 2 .
i=1 i=1

Ex. 9.2.3. Let λ > 0 and X1 , X2 , . . . , Xn be an i.i.d. sample from a population with Exponential(λ)
distribution.

(a) Find the likelihood function L(λ ; X1 , X2 , . . . , Xn ).

(b) Prove that the maximum likelihood estimate for λ is 1/X.

Ex. 9.2.4. Let X1 , X2 , . . . , Xn be an i.i.d. sample from a population with Poisson(λ) distribution,
where λ is known to be strictly positive.

(a) Find the likelihood function L(λ ; X1 , X2 , . . . , Xn ).

(b) Prove that if at least one of the Xj values in non-zero, then the maximum likelihood estimate
for λ is X.

(c) Prove that if all of the Xj values are zero, then L(λ ; X1 , X2 , . . . , Xn ) has no maximum value
for λ > 0.

Ex. 9.2.5. Let 0 < p < 1 and X1 , X2 , . . . , Xn be an i.i.d. sample from a population with
Geometric(p) distribution.

(a) Find the likelihood function L(p ; X1 , X2 , . . . , Xn ).

(b) Let ℓ(p ; X1 , X2 , . . . , Xn ) = loge (L(p ; X1 , X2 , . . . , Xn )). Find the value of p for which
ℓ(p ; X1 , X2 , . . . , Xn ) is maximized.

(c) Prove that the maximum likelihood estimate for p is 1/X.

Ex. 9.2.6. Let σ > 0 and X1 , X2 , . . . , Xn be an i.i.d. sample from a population with Normal(0, σ 2 )
distribution.

Version: – November 19, 2024

332 estimation

(a) Find the likelihood function L(σ ; X1 , X2 , . . . , Xn ).

n
(b) Prove that the maximum likelihood estimate for σ 2 is 1
Xi2 .
P
n
i=1

(c) Prove that this maximum likelihood estimate is also an unbiased estimator for σ 2 in this
case.
Ex. 9.2.7. Let µ ∈ R, σ > 0 and X1 , X2 , . . . , Xn be an i.i.d. sample from a population with
Normal(µ, σ 2 ) distribution.
(a) Find the likelihood function L(µ, σ ; X1 , X2 , . . . , Xn ).
(b) Find the maximum likelihood estimators of µ and σ 2 .
Ex. 9.2.8. Let a < b and X1 , X2 , . . . , Xn be an i.i.d. sample from a population with Uniform(a, b)
distribution. Prove that the maximum likelihood estimates for a and b are min{X1 , X2 , . . . , Xn }
and max{X1 , X2 , . . . , Xn } respectively.
Ex. 9.2.9. Let m be a known integer and p ∈ (0, 1) be an unknown parameter. Let X1 , X2 , . . . , Xn
be an i.i.d. sample from a population with Binomial(m, p) distribution.
(a) Find the likelihood function L(p, X1 , X2 , . . . , Xn ).
(b) Let ℓ(p ; X1 , X2 , . . . , Xn ) = loge (L(p ; X1 , X2 , . . . , Xn )). Find the value of p that maximizes
ℓ(p ; X1 , X2 , . . . , Xn ).
(c) Prove that the maximum likelihood estimate for p is X/m.
Ex. 9.2.10. Let θ > 0 be unknown and X1 , X2 , . . . , Xn be an i.i.d. sample from a population with
Cauchy(θ, 1) distribution.

(a) Find the likelihood function L(θ, X1 , X2 , . . . , Xn ).

(b) Let ℓ(θ ; X1 , X2 , . . . , Xn ) = loge (L(θ ; X1 , X2 , . . . , Xn )). Decide that one cannot explicitly
compute the critical points on ℓ(·).
(c) Modify the R code provided in Example 9.2.5 and try to find the maximum likelihood
estimate for θ.

Ex. 9.2.11. Suppose we have a sample of size n from Multinomial distribution with parameters
k, (p1 , p2 , . . . , pk ). Let Xj , 1 ≤ j ≤ k represent the number of samples that correspond to j-th
X
outcome. Show that for each 1 ≤ j ≤ k the MLE for pj , p̂j = nj .

9.3 confidence intervals

In the previous sections, we have considered data X1 , X2 , . . . , Xn whose distributions are governed
by parameters and described two general methods (namely the methods of moments and of maximum
likelihood) to estimate the parameters of the model from this data. In this section, we will try to
understand how we can quantify the accuracy of estimates.
We will start with the simple model considered in Example 9.2.2 where data are distributed as
Normal with unknown mean but known variance.

Version: – November 19, 2024

 9.3 confidence intervals 333

Example 9.3.1. Let p ∈ R and X1 , X2 , . . . , Xn be an i.i.d. sample from a population distributed

as Normal with mean p and variance 1. Both the method of moments and the method of maximum
likelihood give p̂ = X as the estimator of p.
We know from Chapter 8 that p̂ has a normal distribution with mean p and variance 1/n.
This tells us that p̂ is more “likely” to be close to the true mean p for larger sample size n. It
is conventional to present this information in a slightly different way, in the form of a confidence
interval. The idea of a confidence interval is that instead of a point estimate for a (scalar) parameter
p, we report an interval, depending on the observed data, that is “likely” to contain the unknown p.
√
As p̂ ∼ Normal(p, 1/n), we have n(p̂ − p) ∼ Normal(0, 1). So we can write, for instance,
√
P (−3 ≤ n(p̂ − p) ≤ 3) = Φ(3) − Φ(−3) = 0.9973, (9.3.1)

where the probability value on the right hand side can be computed using R.

pnorm(3) - pnorm(-3)

[1] 0.9973002

By manipulating the inequalities, we can write the equation above in its more standard form,
namely,
3 3
 
P p̂ − √ ≤ p ≤ p̂ + √ = 0.9973.
n n
or
3 3
  
P p ∈ p̂ − √ , p̂ + √ = 0.9973.
n n
It is common practice to express this interval more concisely as

3
p̂ ± √ .
n

When viewed as a random interval, the probability statement above implies that the interval
contains p with probability 0.9973. This property is usually conveyed by the statement that the
interval is a 99.73% “confidence interval.” The factor of 3 used above, which led to the confidence
level of 99.73%, is arbitrary. It is more common to specify a desired confidence level and then
calculate the factor accordingly. For example, to get a confidence level of 95%, we want a factor z
such that
P (Z > z ) = P (Z < −z ) = 0.05/2 = 0.025,

where Z ∼ Normal(0, 1). Such a z is easily calculated as

-qnorm(0.025)

[1] 1.959964

Similarly, for a confidence level of 80%, the factor can be obtained as

Version: – November 19, 2024

334 estimation

-qnorm(0.10)

[1] 1.281552

If we consider a specific empirical realization of X1 , X2 , . . . , Xn , this process leads to a specific

interval. For example, consider the following n = 15 data points: 11.22, 9.56, 10.06, 10.21, 10.95,
10.03, 10.75, 10.71, 11.42, 9.61, 10.91, 8.14, 8.95, 10.57, 11.1. Here, p̂ = 10.279, so the 99.73%
confidence interval is given by

3
10.279 ± √ = [9.5047, 11.054]
15

Of course, this specific interval may or may not contain the true mean. Our “confidence” is in
the procedure which was used to produce the interval, in the sense that it will yield an interval
containing the true mean p with probability 99.73%, as long as the data are from the postulated
Normal model. ■

We could simulate data using R to verify that these confidence intervals do in fact contain the true
parameter as often as expected. It would be also interesting to evaluate the statistical properties of
the intervals when the data are from a different distribution. We do this in Section 9.3.2

9.3.1 Pivotal Quantity approach

The key observation that allows us to write the probability statement (9.3.1) in Example 9.3.1
√
is that n(p̂ − p) ∼ Normal(0, 1). In other words, we have found a function of the data and the
parameter of interest, namely,
√
T (X1 , X2 , . . . , Xn , p) = n(p̂ − p),

so that regardless of the value of p, T (X1 , X2 , . . . , Xn , p) has a completely known distribution. Such
functions are sometimes called pivotal quantities. The derived confidence interval is completely
specified, in principle, once we choose an interval in the support of the known distribution that
has the desired probability. For instance, in the previous example, to get a confidence interval
with coverage probability β = 0.95, we require an interval [a, b] such that P (Z ∈ [a, b]) = β for a
standard Normal random variable Z. There are many such intervals, but intuitively, the interval
[−1.959964, 1.959964] ≈ [−1.96, 1.96] is a good choice because it is the shortest such interval, as
the density of Z is symmetric and decreases away from 0.
In general, we could choose any interval that has probability β. Popular alternative choices in the
standard normal case are (−∞, Φ−1 (β )] and [Φ−1 (1 − β ), ∞), which give “one-sided” confidence
intervals. Once we choose a suitable interval [a, b], the confidence “interval” for p is given by the set

{p : T (X1 , X2 , . . . , Xn , p) ∈ [a, b]} .

Version: – November 19, 2024

 9.3 confidence intervals 335

This set is random as it depends on the random sample X1 , X2 , . . . , Xn , but will have a specific
realisation for any particular empirical sample. We will try to follow this approach in a few other
situations.

Example 9.3.2. Let λ > 0 and X1 , X2 , . . . , Xn be an i.i.d. sample from an Exponential population
with rate λ. We are interested in obtaining a confidence interval for λ. Take

T (X1 , X2 , . . . , Xn , λ) = nλX.

It is easy to see that T (X1 , X2 , . . . , Xn , λ) follows a Gamma(n, 1) distribution. Thus, to obtain

a confidence interval for λ with coverage probability β, we need to find an interval [a, b] ⊂ [0, ∞)
which has probability β under the Gamma(n, 1) distribution.
Such an interval will depend on n. For example, with n = 15 and β = 0.95, one-sided intervals
can be obtained as follows.

beta <- 0.95

c(0, qgamma(beta, shape = 15))

[1] 0.00000 21.88649

c(qgamma(1 - beta, shape = 15), Inf)

[1] 9.24633 Inf

As the Gamma distribution is not symmetric, the choice of a two-sided interval is not obvious. A
simple choice is given by taking the exclusion probabilities on both tails to be equal, as follows.

alpha <- 1 - beta

c(qgamma(alpha / 2, shape = 15), qgamma(1 - alpha / 2, shape = 15))

[1] 8.395386 23.489621

This will typically not be the shortest interval. The shortest interval cannot be obtained in closed
form, but can be computed numerically by varying the left and right tail exclusion probabilites
together so that they add up to 1 − β, and choosing the one giving the shortest interval. ■

Example 9.3.3. Let θ > 0 and X1 , X2 , . . . , Xn be from the the Uniform(0, θ) distribution. We are
interested in a confidence interval for θ. Take
X(n)
T (X1 , X2 , . . . , Xn , θ ) = .
θ

From Example 8.1.4, we know that T (X1 , X2 , . . . , Xn , θ ) has a Beta(n, 1) distribution, which has
an increasing density supported on (0, 1). Thus the shortest interval of probability β will have right

Version: – November 19, 2024

336 estimation

endpoint 1. The left endpoint will depend on n. For example, with n = 15 and β = 0.95, the left
endpoint is given by

qbeta(0.95, 15, 1)

[1] 0.9965863

The corresponding confidence interval for θ is therefore given by

X(n) X(n)
   
θ : 0.9965863 ≤ ≤ 1 = θ : X(n) ≤ θ ≤ . ■
θ 0.9965863

In the next example we discuss two situations where the pivotal quantity approach will not work.

Example 9.3.4. Suppose we want to find a procedure to obtain confidence intervals for the mean
parameter when the underlying distribution is Bernoulli or Poisson.

(a) Let p > 0 and X1 , X2 , . . . , Xn be an i.i.d. sample from the Bernoulli(p) distribution. We are
interested in a confidence interval for p. Unfortunately, there is no obvious pivotal quantity
T (X1 , X2 , . . . , Xn , p) in this example, so this approach does not work.

(b) Let λ > 0 and X1 , X2 , . . . , Xn be an i.i.d. sample from the Poisson(λ) distribution. We
are interested in a confidence interval for λ. We see immediately that T (X1 , . . . , Xn , λ) =
X1 + · · · + Xn has a Poisson(nλ) distribution and consequently depends on λ. We could thus
take n = 1 without loss of generality. However, there is again no obvious pivotal quantity
T (X1 , λ) in this example, so this approach does not work.

In both these examples, it is common practice to obtain approximate confidence intervals using the
Central Limit Theorem. We discuss this approach in the next section. ■

We conclude this section with a final example, returning to the Normal distribution in Example
9.3.1, where we assume that both mean and variance are unknown.

Example 9.3.5. Let X1 , X2 , . . . , Xn be an i.i.d. sample from a population distributed as Normal

with unknown mean µ ∈ R and unknown variance σ 2 > 0. Suppose we are interested in estimating
only the mean parameter µ.
If σ 2 were known, then a natural candidate for our pivotal quantity would have been

√ (X − µ)
T (X1 , X2 , . . . , Xn , µ, σ ) = n .
σ
With a factor of 1.96 for a 95% coverage probability, the corresponding confidence interval for µ
would then be
σ
X ± 1.96 √ ,
n

Version: – November 19, 2024

 9.3 confidence intervals 337

which agrees with our intuition. Unfortunately, σ 2 is not known, so this approach is not valid.
However, as we do have a natural estimator S 2 for σ 2 (recall Theorem 7.1.6), we can try replacing
σ 2 by this estimate and arrive at

√ (X − µ)
T (X1 , X2 , . . . , Xn , µ) = n .
S

Recall from Corollary 8.1.11 that the exact distribution of T (X1 , X2 , . . . , Xn , µ) is tn−1 and thus
is a pivotal quantity (see also Exercise 9.3.1). We can now proceed as we did in Example 9.3.1.
For the standard normal, the shortest interval with probability 0.95 was the symmetric interval
[−1.96, 1.96]. For the tn−1 distribution, similar quantiles can be computed using the qt() function.
For example, with n = 15, the right endpoint is given by

qt(0.975, df = 14)

[1] 2.144787

Considering again the n = 15 data points 11.22, 9.56, 10.06, 10.21, 10.95, 10.03, 10.75, 10.71, 11.42,
9.61, 10.91, 8.14, 8.95, 10.57, 11.1, we have µ̂ = X = 10.279 and σ̂ = S = 0.9112, so the confidence
interval for µ is given by

σ̂ 0.9112
µ̂ ± 2.145 √ = 10.279 ± 2.145 √ = [9.775, 10.784].
n 15

The ability to derive these kinds of confidence intervals, where we can control the confidence level
exactly as long as the model assumptions hold, is one of the main reasons for studying the t
distribution. ■

9.3.2 Empirical Coverage Probability of Confidence Intervals

The theoretical calculations in the previous section guarantee that the confidence intervals we
derived will satisfy the target coverage, as long as the data come from the distribution assumed. It
is still a good idea to verify this using simulation. Doing so will also allow us test how the coverage
probabilities change when the data do not come from the postulated distribution.

We can use R to construct confidence intervals using the formulas derived above. For example,
we can generate a random sample from the Normal distribution and then construct 80% and 95%
confidence intervals for the mean using the following code.

Version: – November 19, 2024

338 estimation

x <- rnorm(20, mean = 10, sd = 3)

n <- length(x)
m <- mean(x)
s <- sd(x)
q80 <- qt(0.9, df = n - 1)
q95 <- qt(0.975, df = n - 1)
ci80 <- c(m + c(-1, 1) * q80 * s / sqrt(n))
ci95 <- c(m + c(-1, 1) * q95 * s / sqrt(n))
ci80

[1] 8.230541 10.388727

ci95

[1] 7.608558 11.010710

We can now repeat this process using replicate(), but to do so, it would be convenient to put
the code above into an R function. Functions in R encapsulate repetitive calculations as code that
can be run with different values of certain variables that are provided to the function as arguments.
In this case, we would like to repeat the above process many times, but with different data, and
possibly different confidence levels. So, we can create a function that takes two arguments, x and
level, and repeats the calculations above.

normalMeanCI <- function(x, level)

{
n <- length(x)
m <- mean(x)
s <- sd(x)
alpha <- 1 - level
qlevel <- qt(1 - alpha / 2, df = n - 1)
ci <- c(m + c(-1, 1) * qlevel * s / sqrt(n))
ci
}

When the function is called with these two arguments, the value of the last expression in the
function is returned as its value. So, we can now repeat our earlier calculations using this more
general function as follows.

normalMeanCI(x, level = 0.8)

[1] 8.230541 10.388727

Version: – November 19, 2024

 9.3 confidence intervals 339

normalMeanCI(x, level = 0.95)

[1] 7.608558 11.010710

We can also repeat this experiment multiple times using replicate().

t(replicate(10, normalMeanCI(rnorm(20, mean = 10, sd = 3), level = 0.8)))

[,1] [,2]
[1,] 8.843671 10.18292
[2,] 8.803891 10.46128
[3,] 9.327093 10.80719
[4,] 9.176958 10.63078
[5,] 9.625395 11.23960
[6,] 8.745298 10.96407
[7,] 10.237023 12.58567
[8,] 10.148164 11.66396
[9,] 9.636105 11.04266
[10,] 9.099662 10.42231

Of course, to estimate the coverage probability we need to repeat this process a much larger number
of times and check how many times the interval contains the true mean. We can do this for 10000
replications as follows.

## each experiment samples 20 independent Normal random variables, and

## produces 80% Confidence interval
cirep <- t(replicate(10000, # repeat experiment 1000 times.
normalMeanCI(rnorm(20, mean = 10, sd = 3), level = 0.8)))
## count proportion of confidence interval that contain the true mean, 10.
mean(cirep[,1] <= 10 & cirep[,2] >= 10)

[1] 0.795

We leave it to the reader to verify that the coverage probabilities seen in simulation match the
target level regardless of the mean, standard deviation or sample sizes.
A related question is how the length of the confidence intervals, which are random quanities,
behave. The empirical distribution of the lengths obtained in the 10000 replications above can be
summarized as follows.

summary(cirep[,2] - cirep[,1])

Version: – November 19, 2024

340 estimation

Excludes true mean Includes true mean

12

11
Confidence interval

10

50 100 150 200

Sample size

Figure 9.2: Simulated 80% confidence intervals for the mean of a Normal population with mean
10 and variance 32 , computed assuming mean and variance are both unknown.
The sample sizes used are from 10, 11, . . . , 200. The interval widths decrease on
average as sample size increases. We expect roughly 1 in 5 intervals to exclude
the true mean (shown in a different color) regardless of sample size as data are
generated from a Normal distribution.

Min. 1st Qu. Median Mean 3rd Qu. Max.

0.7437 1.5547 1.7458 1.7562 1.9491 3.0018
√
From the formula it is clear that the length is inversely proportional to n and proportional to S.
As S 2 /σ 2 follows a χ2n−1 distribution, the square of the length will have a scaled χ2n−1 distribution,
where the scaling factor depends on σ and n. Figure 9.2 plots simulated confidence intervals
obtained from random samples generated from the normal distribution, with varying sample sizes
but fixed mean and varince.
In real life, we cannot expect data to follow a normal distribution exactly. If we knew the
true distribution, we could perhaps derive an appropriate confidence interval, but this is rarely
the case. Even if we knew the true distribution, a suitable confidence interval may not be easy
to derive. Thus, an important practical question is how the normal confidence interval performs
as a confidence interval for the population mean when the data do not arise from the normal
distribution.
The answer obviously depends on the distribution from which the data do arise. Below we
consider four specific examples that represent different types of departures from normality

Example 9.3.6. Consider data X1 , X2 , . . . , Xn from the Bernoulli(p) distribution, for some unknown
p ∈ (0, 1). We saw earlier that the pivotal quantity approach did not work for this model. However,
as the parameter of interest p is the mean, we could blindly apply the normal confidence interval
formula and expect to get something reasonable.

Version: – November 19, 2024

 9.3 confidence intervals 341

Excludes true mean Includes true mean

p = 0.05

0.4

0.3

0.2

0.1

0.0

p = 0.20

0.6
Confidence interval

0.4

0.2

0.0

p = 0.50
1.0

0.8

0.6

0.4

0.2

50 100 150 200

Sample size

Figure 9.3: Simulated 80% confidence intervals for Bernoulli with probability p = 0.05, 0.2,
and 0.5, computed using the normal confidence interval formula for the mean
when mean and variance are both unknown. The sample sizes used are from
10, 11, . . . , 200. Notice that some intervals go below zero, especially for p = 0.05,
where in addition, some intervals consist of the single point {0}. This happens
when all outcomes are 0, which may happen for small n and small p.

Version: – November 19, 2024

342 estimation

We have already generated one Bernoulli sample above, in the experiment where we computed
confidence intervals from normal data 10000 times. If we denote by Xi whether the i-th interval
contained the true mean, then the vector of observed Xi values is given by

x <- cirep[,1] <= 10 & cirep[,2] >= 10

We had previously estimated P (X1 = 1) by the sample proportion

mean(x)

[1] 0.795

which is close to the nominal coverage probability 0.8. However, in view of the current discussion,
we would be more reassured if we see that the value 0.8 is included in a reasonable confidence
interval. We obtain the following 95% interval by applying the normal confidence interval formula.

normalMeanCI(x, level = 0.95)

[1] 0.7870862 0.8029138

This is of course just one example, but we can evaluate the performance of the method using the
same techniques as above, replacing the data generating process to simulate Bernoulli data instead
of normal. Figure 9.3 plots confidence intervals for data generated using Bernoulli(p) for p = 0.05,
0.2, and 0.5, with the setup otherwise similar to Figure 9.2. This plot illustrates some problems with
small p, which are also present for large p close to 1, but otherwise suggests reasonable performance.

To estimate coverage probability and average length for specific combinations of p and n, we
can use the replication approach.

cisummaryBernoulli <- function(p, n, level) {

cirep <- t(replicate(10000,
normalMeanCI(rbinom(n, size = 1, prob = p),
level = level)))
[Link](n = n, p = p,
coverage = mean(cirep[,1] <= p & cirep[,2] >= p),
avg_length = mean(cirep[,2] - cirep[,1]))
}

Version: – November 19, 2024

 9.3 confidence intervals 343

rbind(cisummaryBernoulli(p = 0.05, n = 10, level = 0.95),

cisummaryBernoulli(p = 0.05, n = 100, level = 0.95),
cisummaryBernoulli(p = 0.05, n = 1000, level = 0.95),
cisummaryBernoulli(p = 0.20, n = 10, level = 0.95),
cisummaryBernoulli(p = 0.20, n = 100, level = 0.95),
cisummaryBernoulli(p = 0.20, n = 1000, level = 0.95),
cisummaryBernoulli(p = 0.50, n = 10, level = 0.95),
cisummaryBernoulli(p = 0.50, n = 100, level = 0.95),
cisummaryBernoulli(p = 0.50, n = 1000, level = 0.95))

n p coverage avg length

10 0.05 0.4059 0.19823628
100 0.05 0.8781 0.08462128
1000 0.05 0.9463 0.02701204
10 0.20 0.8879 0.53350536
100 0.20 0.9438 0.15823556
1000 0.20 0.9466 0.04963809
10 0.50 0.9778 0.71138220
100 0.50 0.9400 0.19840636
1000 0.50 0.9471 0.06205443

It is reassuring to see that for large n the coverage probability is close to the target of 95%, with
the average interval length decreasing with n. However, for small n and small p, the observed
coverage probability is substantially smaller than the target. In the next section, we will compare
these results with other approximate confidence intervals obtained using asymptotic results. ■

Example 9.3.7. Consider data X1 , X2 , . . . , Xn from the Exponential(λ) distribution, for some
unknown λ > 0. Here we have an exact confidence interval for λ based on the pivotal quantity
nλX which follows a Gamma(n, 1) distribution. However, as λ is the mean, we can try using the
normal confidence interval as well. Below we contrast the coverage probability and average interval
length for the two methods for λ = 1.

exponentialMeanCI <- function(x, level)

{
alpha <- 1 - level
qgamma(c(alpha / 2, 1 - alpha / 2), shape = length(x)) / sum(x)
}

Version: – November 19, 2024

344 estimation

cisummaryExponential <- function(n, level) {

cirepNorm <- t(replicate(10000, normalMeanCI(rexp(n), level = level)))
cirepExp <- t(replicate(10000, exponentialMeanCI(rexp(n), level = level)))
[Link](n = n,
coverageNorm = mean(cirepNorm[,1] <= 1 & cirepNorm[,2] >= 1),
coverageExp = mean(cirepExp[,1] <= 1 & cirepExp[,2] >= 1),
lengthNorm = mean(cirepNorm[,2] - cirepNorm[,1]),
lengthExp = mean(cirepExp[,2] - cirepExp[,1]),
negative = mean(cirepNorm[,1] < 0))
}

rbind(cisummaryExponential(n = 10, level = 0.95),

cisummaryExponential(n = 100, level = 0.95),
cisummaryExponential(n = 1000, level = 0.95))

n coverageNorm coverageExp lengthNorm lengthExp negative

10 0.8998 0.9506 1.3237152 1.3655739 0.0425
100 0.9397 0.9507 0.3934190 0.3955935 0.0000
1000 0.9461 0.9505 0.1239441 0.1240712 0.0000

Here again we observe that for n = 10, the normal interval has lower than desired coverage probability,
possibly stemming from the lower interval length on average. As in the Bernoulli case with small p,
the normal interval goes below zero in a small proportion of cases. As n increases, however, the
performance of the normal interval becomes comparable with the exponential interval. ■

Example 9.3.8. Consider data X1 , X2 , . . . , Xn from the Cauchy(θ, α2 ) distribution, for some
unknown location θ ∈ R that we are interested in estimating and an unknown scale α > 0. The
Cauchy distribution does not have finite mean, so it is unclear whether it makes sense to use a
confidence interval designed for the population mean. However, we can still blindly apply it and
see what happens. In the simulation below, we take θ = 0 and α2 = 1.

cisummaryCauchy <- function(n, level) {

cirepNorm <- t(replicate(10000, normalMeanCI(rcauchy(n), level = level)))
[Link](n = n,
coverage = mean(cirepNorm[,1] <= 0 & cirepNorm[,2] >= 0),
meanLength = mean(cirepNorm[,2] - cirepNorm[,1]),
medianLength = median(cirepNorm[,2] - cirepNorm[,1]))
}
rbind(cisummaryCauchy(n = 10, level = 0.95),
cisummaryCauchy(n = 100, level = 0.95),
cisummaryCauchy(n = 1000, level = 0.95))

Version: – November 19, 2024

 9.3 confidence intervals 345

n coverage meanLength medianLength

10 0.9823 25.31168 5.319377
100 0.9789 24.08324 4.687978
1000 0.9809 25.20705 4.681542
Here we observe a phenomenon that we have not seen in the ealier examples. Although the coverage
probabilities look reasonable, the average length of the confidence intervals (as measured either by
the mean or the median) do not decrease as sample size increases, and in fact can be quite unstable.
In other words, we we have no improvement in the accuracy of estimation with increasing sample
size. The reason for this seemingly unexpected behaviour is that the sample mean is not a reliable
estimator of the Cauchy location parameter. We will see an alternative confidence interval with
better performance in the next section. ■
Example 9.3.9. Real life data do not always follow theoretical assumptions precisely, but as we
have seen in the previous examples, Normal confidence intervals often work well for estimating the
population mean regardless, especially for large sample sizes. This is essentially a consequence the
Central Limit Theorem, which we will use with more finesse in the next section.
Another kind of departure from theoretical assumptions we sometimes see in real data are
outliers or other kinds of contaminations that affect individual data points rather than the full sample.
In this example, we will consider data X1 , X2 , . . . , Xn that mostly come from the Normal(0, 1)
distribution, but sometimes, with probability 0.01 or 1%, come from the Normal(0, 100) distribution.
We can write an R function to generate such a mixture as follows.

rmixnorm <- function(n, prob = 0.01) {

rnorm(n, mean = 0, sd = ifelse(runif(n) < prob, 100, 1))
}

We can now simulate Normal confidence intervals as before, and summarize their properties for
various values of n.

cisummaryMixture <- function(n, level) {

cirepNorm <- t(replicate(10000, normalMeanCI(rmixnorm(n), level = level)))
[Link](n = n,
coverage = mean(cirepNorm[,1] <= 0 & cirepNorm[,2] >= 0),
meanLength = mean(cirepNorm[,2] - cirepNorm[,1]))
}
rbind(cisummaryMixture(n = 10, level = 0.95),
cisummaryMixture(n = 100, level = 0.95),
cisummaryMixture(n = 1000, level = 0.95))

n coverage meanLength
10 0.9535 4.817333
100 0.9779 2.771039
1000 0.9593 1.195399

Version: – November 19, 2024

346 estimation

Here again the coverage probabilities are close to the target and the confidence intervals decrease
in size as sample size increases. ■

9.3.3 Approximate Confidence Intervals using CLT

We have seen in the previous section that the Normal confidence intervals obtained from data
X1 , X2 , . . . , Xn perform quite reasonably even when the underlying distribution is not Normal, with
the exception of the Cauchy distribution. This essentially follows from the Central Limit Theorem
(Theorem 8.4.1), which can be interpreted as saying that for large n we have the approximation

√ (X − µ)
 
P n ≤ 1.96 ≈ 0.95
σ

In the simulation examples in the previous section, we have replaced σ by the sample standard
deviation S, and the normal quantile 1.96 by the corresponding tn−1 quantile. As n → ∞, the tn−1
becomes unnecessary, and it can be shown that

√ (X − µ) d
n −→ Normal(0, 1)
S
under reasonable moment conditions on the underlying distribution (see Example 8.3.7). In this
section, we will see whether applying the Central Limit Theorem to specific models can provide
improved confidence intervals.

Example 9.3.10. Consider data X1 , X2 , . . . , Xn from the Bernoulli(p) distribution, for some
unknown p ∈ (0, 1). We saw in Example 9.3.6 that the normal confidence interval performs
reasonably well for large n but not for small n, especially when p is small. Let us see if the Central
Limit Theorem for this specific model leads us to a better interval.
The Normal confidence interval we have been using so far assumes that both mean and variance
are unknown. However, the Bernoulli model has only one parameter p, on which both mean and
variance depend. Specifically, the Central Limit Theorem (See Example 8.4.5) gives the asymptotic
distribution of the sample proportion p̂ as

√ (p̂ − p) d
T (X1 , X2 , . . . , Xn , p) = np −→ Normal(0, 1).
p(1 − p)

This T (X1 , X2 , . . . , Xn , p) can be viewed as an approximately pivotal quantity, and following our
earlier approach, an approximate 95% confidence interval for p is given by
( )
√ (p̂ − p)
p : −Q ≤ np ≤Q , (9.3.2)
p(1 − p)

where Q = 1.96 is the 0.975 quantile of standard Normal. The bounds will be achieved by p that
satisfy the quadratic equation
Q2
n(p̂ − p)2 = p(1 − p).
n

Version: – November 19, 2024

 9.3 confidence intervals 347

It it easy to verify that this equation has two real solutions, given by
!
1 Cn2 + 4Cn p̂(1 − p̂)
p
Cn
+ p̂ ±
( Cn + 1 ) 2 2

where Cn = Q2 /n. These are called Wilson Confidence intervals [Wil27]. Ignoring the possibility
that these solutions may lie outside [0, 1], we can compute this confidence interval in R as follows

bernoulliQuadraticCI <- function(x, level)

{
n <- length(x)
phat <- mean(x)
Q <- qnorm(1 - (1 - level) / 2)
C <- Qˆ2 / n
ci <- ((C + 2 * phat) + c(-1, 1) *
sqrt(Cˆ2 + 4 * C * phat * (1-phat))) / (2 * (C + 1))
ci
}

To avoid the complication of solving a quadratic equation, a common alternative is to replace

the unknown p in the denominator of (9.3.2) by the sample proportion p̂. This is equivalent to
using a corollary of the Central Limit Theorem for the sample proportion, which says that (see
Exercise 9.3.5)
√ (p̂ − p) d
Te(X1 , X2 , . . . , Xn , p) = n p −→ Normal(0, 1). (9.3.3)
p̂(1 − p̂)
It is then trivial to see that the confidence interval for p is given by solving linear equations as
q
p̂ ± Q p̂(1 − p̂)/n.

This simpler interval can be computed in R as follows.

bernoulliLinearCI <- function(x, level)

{
n <- length(x)
phat <- mean(x)
Q <- qnorm(1 - (1 - level) / 2)
ci <- phat + c(-1, 1) * Q * sqrt(phat * (1 - phat) / n)
ci
}

We can now repeat our earlier simulation experiment from Example 9.3.6, this time with these
intervals instead of the normal interval.

Version: – November 19, 2024

348 estimation

cisummaryBernoulli <- function(p, n, level) {

cirepQ <- t(replicate(10000,
bernoulliQuadraticCI(rbinom(n, size = 1, prob = p),
level = level)))
cirepL <- t(replicate(10000,
bernoulliLinearCI(rbinom(n, size = 1, prob = p),
level = level)))
[Link](n = n, p = p,
coverageQ = mean(cirepQ[,1] <= p & cirepQ[,2] >= p),
coverageL = mean(cirepL[,1] <= p & cirepL[,2] >= p),
meanLengthQ = mean(cirepQ[,2] - cirepQ[,1]),
meanLengthL = mean(cirepL[,2] - cirepL[,1]))
}

rbind(cisummaryBernoulli(p = 0.05, n = 10, level = 0.95),

cisummaryBernoulli(p = 0.05, n = 100, level = 0.95),
cisummaryBernoulli(p = 0.05, n = 1000, level = 0.95),
cisummaryBernoulli(p = 0.20, n = 10, level = 0.95),
cisummaryBernoulli(p = 0.20, n = 100, level = 0.95),
cisummaryBernoulli(p = 0.20, n = 1000, level = 0.95),
cisummaryBernoulli(p = 0.50, n = 10, level = 0.95),
cisummaryBernoulli(p = 0.50, n = 100, level = 0.95),
cisummaryBernoulli(p = 0.20, n = 1000, level = 0.95))

n p coverageQ coverageL meanLengthQ meanLengthL

10 0.05 0.9088 0.3992 0.32787000 0.15990182
100 0.05 0.9686 0.8821 0.08827661 0.08290597
1000 0.05 0.9487 0.9408 0.02712679 0.02697052
10 0.20 0.9636 0.8842 0.43447816 0.43615822
100 0.20 0.9375 0.9319 0.15429259 0.15563230
1000 0.20 0.9483 0.9473 0.04948471 0.04953288
10 0.50 0.9801 0.8900 0.50689814 0.58510371
100 0.50 0.9449 0.9452 0.19141843 0.19503036
1000 0.20 0.9456 0.9449 0.04951000 0.04955135

Comparing with the results for the Normal interval, we see that the simpler linear interval has
roughly similar performance. However, the more complicated quadratic interval has substantially
better performance for small n and small p. It is important to remember that all these intervals
are based on asymptotic results, and it is difficult to analyse performance in small samples except
through simulation. Nonetheless, it is not surprising that the methods which replace the variance
term by an estimate do worse than the method that does not. ■

Version: – November 19, 2024

 9.3 confidence intervals 349

9.3.4 Confidence Intervals for the Population Median

Just as the Central Limit Theorem for the sample mean allows us to formulate approximate
confidence intervals for the population mean, we can try to use the limiting distribution of
the sample median to obtain a confidence interval for the population median. For symmetric
distributions, the population median coincides with the population mean, so the two intervals are
for the same population parameter.

Let X1 , X2 , . . . be i.i.d. random variables with probability density function f with median µ.
Theorem 8.6.2 tells us that if f (µ) > 0, then
√  
d
2 nf (µ) Xen − µ −→ Z,

where X en is the sample median obtained from X1 , X2 , . . . , Xn . We cannot use this result directly
to get a confidence interval for µ, because f (µ) is unknown. To estimate it, we shall proceed with
the following approximations. First from the fact that F −1 (F (µ)) = 1 and an application of chain
rule

f (µ) = F ′ (µ)
1
=
(F −1 )′ (F (µ))
2
≈√  
n F −1 ( 21 + √1 ) − F −1 ( 1
n 2 − √1 )
n

Now if X(r ) is the r-th order statistic derived from X1 , X2 , . . . , Xn (see Section 8.1.1) then it is
intuitively clear that

1 1 1 1
F −1 ( + √ ) ≈ X([n( 1 + √1 )]) and F −1 ( − √ ) ≈ X([n( 1 − √1 )]) .
2 n 2 n 2 n 2 n

Thus we arrive at our plausible estimate

2
fˆn (µ) = √ ,
n(X([n( 1 + √1 )]) − X([n( 1 − √1 )]) )
2 n 2 n

for f (µ). This approximation above can be made rigorous and can be used to provide confidence
intervals by using the lemma below. We refer the reader to [Ser09] for a complete treatment of
convergences of sample quantiles.

Lemma 9.3.11. Suppose X1 , X2 , . . . be an i.i.d. sequence from X. Suppose X is a continuous

random variable with probability density function f : R → R. Suppose f (µ) > 0 where µ is the
median of f . Then
√  
d
2 nfˆn (µ) X en − µ −→ Z, (9.3.4)

Version: – November 19, 2024

350 estimation

Proof. From [Ser09, Corolllary, Section 2.5.2] we have

p
fˆn (µ) −→ f (µ)

as n → ∞. An application of the Central Limit Theorem for the median (Theorem 8.6.2) along
with Slutsky’s Theorem (Lemma 8.3.10) yields the result. ■

Now using (9.3.4) we have the confidence interval

en ± √1.96
X
2 nfˆn (µ)

for confidence level 0.95, with the numerator on the right hand side adjusted for other confidence
levels. We can calculate this confidence interval using R as follows.

medianCI <- function(x, level)

{
n <- length(x)
m <- median(x)
delta <- diff(quantile(x, 0.5 + c(-1, 1) / sqrt(n),
names = FALSE))
fhat <- 2 / (delta * sqrt(n))
qlevel <- qnorm(1 - (1 - level) / 2)
ci <- c(m + c(-1, 1) * qlevel / (2 * sqrt(n) * fhat))
ci
}

We apply this in three examples below.

Example 9.3.12. Let X1 , X2 , . . . , Xn be i.i.d. Normal µ, σ 2 . In the simulation study below, we

keep the true mean and variance fixed at 0 and 1, but assume that they are both unknown. The
sample size is varied to see how well the median approximation works for small samples.

Version: – November 19, 2024

 9.3 confidence intervals 351

cisummaryMedianNormal <- function(n, level) {

cirepMean <- t(replicate(10000, normalMeanCI(rnorm(n), level = level)))
cirepMedian <- t(replicate(10000, medianCI(rnorm(n), level = level)))
[Link](n = n,
coverageMean = mean(cirepMean[,1] <= 0 & cirepMean[,2] >= 0),
coverageMedian = mean(cirepMedian[,1] <= 0 & cirepMedian[,2] >= 0),
lengthMean = mean(cirepMean[,2] - cirepMean[,1]),
lengthMedian = mean(cirepMedian[,2] - cirepMedian[,1]))
}

rbind(cisummaryMedianNormal(n = 10, level = 0.95),

cisummaryMedianNormal(n = 100, level = 0.95),
cisummaryMedianNormal(n = 1000, level = 0.95))

n coverageMean coverageMedian lengthMean lengthMedian

10 0.9501 0.9252 1.3931430 1.5160993
100 0.9467 0.9396 0.3958827 0.4913522
1000 0.9532 0.9463 0.1240568 0.1554546

The coverage of the median interval is reasonable, although a little lower than the target for smaller
n. Observe that on average the length of the median interval is larger than the mean interval, in
spite of the lower coverage probability. This is related to the asymptotic efficiency of the median
discussed in the previous chapter. ■

Example 9.3.13. Median intervals are potentially more useful when the underlying distribution is
not Normal. An extreme case illustrating this is the Cauchy distribution. Continuing Example 9.3.8,
consider data X1 , X2 , . . . , Xn from the Cauchy(θ, α2 ) distribution, where we are interested in
estimating the unknown location θ ∈ R, and the unknown scale α > 0 is a nuisance parameter. As
we saw in Example 9.3.8, the Normal mean confidence inteval completely fails in this case.
Repeating the experiment with the median interval, we get the following.

cisummaryMedianCauchy <- function(n, level) {

cirepNorm <- t(replicate(10000, medianCI(rcauchy(n), level = level)))
[Link](n = n,
coverage = mean(cirepNorm[,1] <= 0 & cirepNorm[,2] >= 0),
meanLength = mean(cirepNorm[,2] - cirepNorm[,1]),
medianLength = median(cirepNorm[,2] - cirepNorm[,1]))
}
rbind(cisummaryMedianCauchy(n = 10, level = 0.95),
cisummaryMedianCauchy(n = 100, level = 0.95),
cisummaryMedianCauchy(n = 1000, level = 0.95))

Version: – November 19, 2024

352 estimation

n coverage meanLength medianLength

10 0.9741 3.6774258 2.8444431
100 0.9504 0.6397202 0.6304624
1000 0.9457 0.1951665 0.1942360

With the median confidence interval, we get reasonable coverage with the average interval lengths
decreasing with sample size, as we would expect. ■
Example 9.3.14. Continuing Example 9.3.9, consider the following model that mimics low-
probability contamination of the data by outliers. Suppose X1 , X2 , . . . , Xn come from the Normal(0, 1)
distribution with probability 0.99, but with probability 0.01, they arise from the Normal(0, 100)
distribution.
Although we have not compared properties of the sample mean and the sample median, one
intuitively obvious property of the median is that changing the values of a few extreme data points
will not affect it, whereas it might affect the sample mean. We can thus expect the median to be
more stable under this contamination model, in the sense that it will have lower variance. Although
we will not do it here, we can formalize this property in terms of the relative asymptotic efficiency
of the sample median over the sample median. In the last chapter, we noted that the sample mean
was more efficient than the sample mean when data are obtained from the normal distribution, and
this explains the wider median confidence intervals in Example 9.3.12. However, the situation is
reversed in this model, where the median is more efficient, and will lead to narrower confidence
intervals on average. We can see this in the results of the following simulation.

cisummaryMedianMixture <- function(n, level) {

cirepMean <- t(replicate(10000, normalMeanCI(rmixnorm(n), level = level)))
cirepMedian <- t(replicate(10000, medianCI(rmixnorm(n), level = level)))
[Link](n = n,
coverageMean = mean(cirepMean[,1] <= 0 & cirepMean[,2] >= 0),
coverageMedian = mean(cirepMedian[,1] <= 0 & cirepMedian[,2] >= 0),
lengthMean = mean(cirepMean[,2] - cirepMean[,1]),
lengthMedian = mean(cirepMedian[,2] - cirepMedian[,1]))
}
rbind(cisummaryMedianMixture(n = 10, level = 0.95),
cisummaryMedianMixture(n = 100, level = 0.95),
cisummaryMedianMixture(n = 1000, level = 0.95))

n coverageMean coverageMedian lengthMean lengthMedian

10 0.9563 0.9343 4.957189 1.5930168
100 0.9763 0.9367 2.794230 0.4943884
1000 0.9590 0.9483 1.201950 0.1569595

As we can see, the coverage probabilities in both intervals are reasonable close to the target level of
95%, but the median intervals are much narrower on average. ■

Version: – November 19, 2024

 9.3 confidence intervals 353

exercises

Ex. 9.3.1. Let X1 , X2 , . . . , Xn be an i.i.d. sample from a Normal population with unknown mean
µ ∈ R and unknown variance σ 2 > 0. Show that the function

√ (X − µ)
T (X1 , X2 , . . . , Xn , µ) = n .
S

is a pivotal quantity. (Hint: define Zi = Xiσ−µ and show that T (X1 , X2 , . . . , Xn , µ) = T (Z1 , Z2 , . . . , Zn , 0).
The distribution of T (Z1 , Z2 , . . . , Zn , 0) clearly does not depend on µ or σ 2 .)

Ex. 9.3.2. X1 , X2 , . . . , Xn i.i.d. U nif orm(0, θ ), Confidence interval for θ (one-sided? shortest
length of the form [aX(n) , bX(n) ])

Ex. 9.3.3. Let X1 , X2 , . . . , Xn be an i.i.d. sample from a Normal µ, σ 2 distribution. Take

n
X (Xi − X )2
T (X1 , X2 , . . . , Xn , σ 2 ) = .
σ2
i=1

It is easy to see from Theorem 8.1.10 that T (X1 , X2 , . . . , Xn , σ 2 ) has a χ2n−1 distribution. Use this
to obtain a confidence interval for σ 2 .

Ex. 9.3.4. Let σ > 0 and X1 , X2 , . . . , Xn be an i.i.d. sample from the Normal(0, σ 2 ) distribution.

(a) Find the distribution of

n
X X2
T (X1 , X2 , . . . , Xn , σ 2 ) = i
,
σ2
i=1

(b) Using qchisq() in R, find a 95% confidence interval for σ 2 with equal tail exclusion proba-
bilites.

(c) Using qchisq() in R, find a 95% one-sided confidence interval for σ 2 .

Ex. 9.3.5. Consider an i.i.d. sample X1 , X2 , . . . from Bernoulli(p) distribution. Let

Pn
i=1 Xi
p̂ = .
n
p
(a) Show that p̂ −→ p as n → ∞.

(b) Using Slutsky’s theorem (Lemma 8.3.10) show (9.3.3).

Ex. 9.3.6. Suppose X1 , X2 , X3 , . . . are i.i.d. X with E [X ] = p and Var[X ] = σ 2 . Let p̂ = X n .

√
(a) Suppose X ∼ Bernoulli(p) then using the variance stabilising formula g (x) = arcsin( x)
construct a 95% confidence interval for p.

(b) Construct confidence intervals using the variance stablisling formula from (a), R-code below
and cisummaryBernoulli discussed in Example 9.3.10.

Version: – November 19, 2024

354 estimation

bernoulliVSTCI <- function(x, level)

{
n <- length(x)
phat <- mean(x)
Q <- qnorm(1 - (1 - level) / 2)
tci <- asin(sqrt(phat)) + c(-1, 1) * Q / (2 * sqrt(n))
ci <- sin(tci)ˆ2
ci
}

(c) Are the intervals larger in comparison to those in Example 9.3.10?

√
Ex. 9.3.7. Repeat Exercise 9.3.6 when X ∼ Poisson(10) with g (x) = x.

Version: – November 19, 2024

HYPOTHESIS TESTING
10
In Chapter 9 we looked at the problem of estimating the value of unknown parameters in a
probability model, where the probability model prescribes a distribution for an attribute or quantity
of interest in the population. The estimation is based on a sample of data from the population
that we assume have been generated from that model. In many situations however, estimating the
parameters is not the primary problem of interest. Rather, we may be interested in whether two or
more subpopulations are different, or determining whether two or more measured attributes are
independent of others. Examples of such problems abound:

• Are temperatures on average higher now than they were a hundred years ago?

• Are people with higher blood glucose levels at age 30 more likely to develop diabetes by age
60?

• Does smoking decrease life expectancy? Does eating organic food increase life expectancy?

One way in which we can formulate and answer such questions is by viewing as the “test of a
hypothesis”, which is a central problem in statistics. A hypothesis test makes a conjecture or
hypothesis about the population (e.g., an attribute has the same distribution in two subpopulations,
or two attributes are independent) and then carries out a computation to test the credibility
of the conjecture. Probability theory is important in hypothesis testing because hypotheses are
based on probability models, and the computations required to arrive at a decision are done using
probabilistic techniques.
In this chapter we will discuss simplified prototype hypothesis testing problems. These may not
be as nuanced as many real life hypothesis testing problems but will convey their essence. The
techniques described are in fact practically useful in a wide range of situations.

10.1 introduction

We begin our discussion on testing with a simple example.

Example 10.1.1. Recall the example from Chapter 9 of a coin which we assume has a probability
p of showing heads each time it is flipped. The results of flipping the coin 100 times are viewed as
i.i.d. random variables X1 , X2 , . . . , X100 , each with a Bernoulli(p) distribution. This is a special
case of the multinomial distribution, a fact that will become useful in a later example. As there are
only two possible outcomes here, there is only one parameter, the probability p of seeing heads.
100
Suppose Xn = 67 of these flips showed heads. In the previous chapter we considered what
P
n=1
inferences we can make about the value of p. However, suppose our primary interest is only in
deciding whether p = 0.5; that is, whether or not the coin is fair. Testing the conjecture (or
hypothesis) that p = 0.5 involves making a judgement about whether the observed data is consistent
with the conjecture. Given what we have learned in the previous chapter, one simple approach

355

Version: – November 19, 2024

356 hypothesis testing

could be to say “yes” if a suitable confidence interval for p includes the value 0.5. This is a perfectly
reasonable approach for this problem, but for now we will take an alternative, more direct approach
that is easier to generalize to other testing problems. In this case, for example, we may argue as
follows: “If the coin had an equal chance of showing heads or tails, then the probability of observing
67 heads or more in 100 flips is around 0.0004 (this can be verified using R). This number is small
enough to suggest that the hypothesis of heads and tails being equally likely is inaccurate”. ■

We will try to make this intuition concrete in a way that can be generalized to other situations. In
Chapter 9 we introduced the idea of maximum likelihood as a systematic approach to estimation.
In this chapter, we will continue using it to develop a unified approach to hypothesis testing. We
will see that this approach leads to useful testing procedures in many situations. However, we will
also come across situations where it does not. In practice we often resort to ad hoc procedures
instead. Such procedures, while important and useful, are beyond the scope of this book.
As before, we assume that the sample X1 , X2 , . . . , Xn are i.i.d. copies of a random variable
X with a probability mass function or probability density function f (x), where f (x) = f (x |
p1 , p2 , . . . , pd ) = f (x | p) depends on one or more unknown parameters p = (p1 , p2 , . . . , pd ) ∈ P ⊂
Rd for some d ≥ 1. A conjecture or “null hypothesis” about X restricts the possible values that
p can take, and is represented by the statement that p ∈ P0 , where P0 is a proper subset of P.
Just as we computed the maximum likelihood estimator p̂ as the value of p that maximizes the
likelihood function, we can also compute an estimator p̂0 that maximizes the likelihood function
within this smaller subset P0 . If the null hypothesis holds, we expect the likelihoods at p̂ and p̂0 to
be close to each other, whereas we expect the likelihood at p̂0 to be substantially smaller if the null
hypothesis does not hold.
Our goal is to develop this idea into an approach to hypothesis testing. In the formal testing
problem described above, this approach will lead to a “test statistic”, which is a function of the data
X1 , X2 , . . . , Xn . Naturally, the probability distribution of this statistic will depend on the unknown
parameter p ∈ P. To “test” the conjecture that p ∈ P0 for a specified P0 ⊂ P, one essentially asks
whether the observed value of the test statistic could conceivably have come from a value of p ∈ P0 ,
quantifying the degree of this possibility through a probability that is referred to as the “p-value”.
The only requirement, albeit a very important one, is that this p-value not depend on the unknown
parameter p, which is equivalent to saying that the distribution of the test statistic should be fully
known when p ∈ P0 , so that probability calculations involving the test statistic can be performed
explicitly leading to a numerical answer.
The precise notion of the p-value is important, but somewhat difficult to convey. We will get
to a formal definition only in Section 10.5.3, but loosely speaking, it can be thought of as the
probability of the test statistic being “at least as extreme” as the observed statistic if the null
hypothesis was true. This makes intuitive sense: an observed value of the test statistic that is
“likely” to occur if the null hypothesis were true supports the null hypothesis, so conversely one
that is “unlikely” contradicts it. However, only considering the probability of the observed statistic
under the null is not enough, and we need to consider the probabilities of “more extreme” outcomes
as well. To get a sense of why this is important, let us revisit Example 10.1.1 above.

Version: – November 19, 2024

 10.1 introduction 357

100
Example 10.1.1. (Continued) Suppose we use the natural test statistic S = Xi for this example.
P
i=1
Under the null hypothesis that the coin is fair (P0 = { 12 }), S ∼ Binomial(100, 12 ), so P (S = 67)
and P (S ≥ 67) are given by

dbinom(67, 100, 0.5)

[1] 0.0002324713

1 - pbinom(66, 100, 0.5)

[1] 0.0004368599

Both of these probabilities are small, and provide evidence against the null. But consider the
situation where the observed sum was 50 rather 67. Clearly, this outcome is the strongest possible
evidence in favour of the null hypothesis. Here, the corresponding probabilities P (S = 50) and
P (S ≥ 50) are

dbinom(50, 100, 0.5)

[1] 0.07958924

1 - pbinom(49, 100, 0.5)

[1] 0.5397946

Here P (S = 50) is arguably quite small, and if we make a decision based on P (S = 50) alone, we
ignore the fact that while small, it is still the highest probability for any individual outcome. The
contrast is more extreme if we consider a similar problem with 1000 instead of 100 tosses; in this
case, P (S = 500) and P (S ≥ 500) are

dbinom(500, 1000, 0.5)

[1] 0.02522502

1 - pbinom(499, 1000, 0.5)

[1] 0.5126125

Version: – November 19, 2024

358 hypothesis testing

In general, the probability of individual outcomes decreases as the size of the sample space
“increases”, and in fact for continuous distributions, the probability of any singleton outcome is 0. It
is thus more natural to define the p-value as P (S ≥ s) where s is the observed quantity. ■
We will see later that the above notion of the p-value defined in terms of the probability of the test
statistic being “at least as extreme as the observed test statistic” is generally useful regardless of
the underlying model.

10.2 the goodness of fit problem in the multinomial model

A natural generalization of the previous example is to situations where there are more than one
possible outcomes, and we want to “test”, based on observed data, whether the probabilities
associated with each outcome are as expected. A prototype of this problem is given by the following
example.
Example 10.2.1. Suppose we roll a six-sided die n times, and record the outcome of the i-th roll
as Ri . Assuming that the rolls are independent, the distribution of R1 , R2 , . . . , Rn is determined
by the vector of probabilities of each outcome, which we denote by p = (p1 , p2 , . . . , p6 ). Formally,
the parameter space for the problem is

6
( )
X
P= p = (p1 , p2 , . . . , p6 ) : 0 ≤ pi ≤ 1 for all i, pi = 1 .
i=1

We wish to test whether the die is fair, or in other words, pi = 1/6 for all i. This happens if p
belongs to the singleton subset P0 = {(1/6, 1/6, . . . , 1/6)} of P. ■
Recall the multinomial distribution discussed in Example 3.2.12. If we define the vector X =
(X1 , X2 , . . . , X6 ) as the number of times each outcome occurs, that is,
n
X
Xj = 1{Ri = j}, j = 1, 2, . . . , 6
i=1

then we can identify the distribution of X as the multinomial distribution with size n and probability
vector p. Another useful way to represent X is using indicator vectors as follows. For j = 1, 2, . . . , 6,
let ej be the 6-dimensional unit vector with 1 in the j-th position and 0 otherwise. Define the
i.i.d. random variables Yi = ej if Ri = j. Then it is easy to verify that the distribution of Yi is
Multinomial with parameter 1 and probabilities p = (p1 , p2 . . . p6 ) and X = ni=1 Yi is Multinomial
P

with parameter n and probabilities p. The distribution of individual Yi ’s is refered to as the

“categorical distribution” which generalizes the Bernoulli distribution to multiple categories.
In the multinomial framework, we have seen in Exercise 9.2.11 that the MLE of p̂k = Xk /n
with X = Ri , then
P

There is no obvious test statistic, although intutively it seems that large values of (Xi − n6 )2
P

would suggest that the equal probability hypothesis does not hold.
In its more general form, when the number of outcomes is some fixed number m (not necessarily
6) and P0 is a singleton set consisting of a fixed element of P (not necessarily with all components

Version: – November 19, 2024

 10.3 independence of two categorical attributes 359

equal), this problem is known as the “goodness of fit” problem. Despite the simplicity of the
problem, we will see that it is difficult to find a simple solution.

10.3 independence of two categorical attributes

Another simple but important problem that arises in the context of categorical data is to decide
whether two categorical attributes are independent. Once suitably framed, this problem can also
be formulated in terms of the multinomial distribution, but with a more complicated P0 . As with
the goodness of fit problem, there is no simple solution, yet the simplicity and wide applicability
of the problem makes it important to study. We use the remainder of this section to formulate
the problem precisely. We then move on to a discussion of the testing problem in general and
some specfic problems that are simpler to analyze. We will come back to the goodness of fit and
independence of categorical attributes problems towards the end of the chapter.

To motivate the problem we want to solve, consider the following example.

Example 10.3.1. Suppose a medical research team has come up with a potential vaccine for a
dangerous disease, and we have been asked to design an experiment to determine whether it is
effective. There are of course established protocols to design such studies (commonly known as
‘clinical trials’), but the following strategy is reasonable as a first attempt. Choose n individuals
from a vulnerable population, give the vaccine to n1 of them (giving the remaining n2 = n − n1
individuals a placebo as “control”). Then, wait for a reasonable period of time (which may depend
on the features of the disease) to see how many individuals are affected by the disease in each
group. The result may be summarized in a 2 × 2 table as follows, where X11 denotes the number of
vaccinated individuals who were affected, X12 denotes the number of vaccinated individuals who
were not affected and so on.

Affected Not Affected Total

Vaccine X11 X12 n1
Placebo X21 X22 n2

If the vaccine is effective, we expect a smaller proportion of the vaccinated group to be affected. If
the chance of getting affected does not depend on whether the vaccine was given, then the vaccine
is ineffective. The principle of scientific skepticism suggests that we should start by supposing
independence (i.e., the vaccine has no effect), unless convinced otherwise by evidence. ■
It is easy to see that the setup in the previous example can apply to a wide range of problems
of a similar nature. The essential aspects are that either of two “treatments” are applied to a group
of experimental units, and one of two possible “outcomes” is then recorded for each unit. In the
example above, treatments are vaccine and placebo, and outcomes are affected and not affected,
but in general they can represent any binary categorical variable. The number of participants
given treatment k who have outcome ℓ is denoted by Xkℓ for k, ℓ = 1, 2. In general, neither the
number of possible treatments nor the number of possible outcomes need to be two, and both can
be categorical attributes with an arbitrary number of categories.

Version: – November 19, 2024

360 hypothesis testing

In the example above, we have not explicitly introduced any randomness. Indeed, many aspects
of the experiment could be random, such as the way the participants are selected from the population
of interest, the number of participants, the number of participants given each treatment, and so on.
We will now consider ways in which parametric probability models can describe such experiments.

To use our usual parametric setup, we need a sample of i.i.d. observations. The natural
independent units in Example 10.3.1 are the participants in the study; in general, we may assume
that individuals or units are selected randomly from a population of interest. Each such individual
has two categorical attributes (coded by the numbers 1 and 2 for convenience), which we will
generally refer to as treatment and outcome, even if they are not actually treatments or outcomes
in the literal sense. There are four possible treatment-outcome combinations for each unit, namely
(1, 1), (1, 2), (2, 1), and (2, 2). If we identify these combinations respectively with the four matrices
" # " # " # " #
1 0 0 0 0 1 0 0
, , , and .
0 0 1 0 0 0 0 1

then it is easy to see that the summary table is in fact simply the sum of these matrices over all
units in the experiment.

10.3.1 A Multinomial Model for Two-way Tables

It is now quite simple to formulate the problem parametrically. The information available for each
individual unit (namely, the values of the “treatment” and “outcome” attributes) can be one of four
possible 2 × 2 matrices, which can thus be thought of as the sample space of outcomes in a random
experiment. If we assume that these outcomes, say Mi for the i-th individual, are distributed
independently and identically for each individual unit, then we are back in our usual setup where
we have n i.i.d. random observations (i.e., a random sample) M1 , M2 , . . . , Mn from some unknown
distribution. The unknown distribution is discrete with four possible outcomes, so the most general
parametric model is to assign four unknown probabilities to each outcome, with the only constraint
that they must add up to "1. In #other words, the" parameters # in the model
" are
# the probabilities
1 0 0 0 0 1
p11 of seeing the outcome , p21 of seeing , p12 of seeing , and p22 of seeing
0 0 1 0 0 0
" #
0 0
, with the constraints that 0 ≤ p11 , p21 , p12 , p22 ≤ 1 and p11 + p21 + p12 + p22 = 1. This
0 1
model is precisely the categorical distribution we saw in the previous section, with the number of
possible outcomes being four rather than six. As the n random outcomes are independent, the
individuals outcomes
" Mi can
# be combined and summarized by their sum, giving the 2 × 2 table
n X11 X12
. The distribution of the table X can thus be identified as a multinomial
P
X= Mi =
i=1 X21 X22

Version: – November 19, 2024

 10.3 independence of two categorical attributes 361

distribution, with size n and probability vector (p11 , p21 , p12 , p22 ). The parameter space is a subset
of R4 , specifically
 
 X 
P= p = (p11 , p21 , p12 , p22 ) : 0 ≤ pij ≤ 1, pij = 1 .
 
i,j

This multinomial model is appropriate when individuals or units are sampled from the population
independently, with the total sample size n fixed in advance. This assumption is usually reasonable
in observational studies where both attributes are intrinsic properties of the units being sampled,
for example, in a study of college students that record each individual’s gender and whether they
need corrective lenses. For controlled trials such as the one described in Example 10.3.1, it is not
immediately clear whether this model is appropriate because one of the attributes, namely the
treatment, is assigned by the experimenter and not an intrinsic property of the individual subjects.
Consider this alternative probability model for such experiments: Suppose individuals are
chosen independently at random from the population of interest. Each chosen individual is assigned
treatment 1 with probability π1 and treatment 2 with probability π2 = 1 − π1 . Let q11 be the
conditional probability of observing outcome 1 given treatment 1 and q12 = 1 − q11 . Similarly,
suppose the conditional probability of outcome 1 given treatment 2 is q21 and the conditional
probability of outcome 2 given treatment 2 is q22 . Then, the unconditional probability that
individual i gets treatment 1 and has outcome 1 can be calculated as
" #!
1 0
P Mi = = P (treatment = 1, outcome = 1)
0 0
= P (treatment = 1) P (outcome = 1 | treatment = 1)
= π1 q11

Similarly, we have
" #!
0 0
P Mi = = π2 q21 ,
1 0
" #!
0 1
P Mi = = π1 q12 , and
0 0
" #!
0 0
P Mi = = π2 q22 .
0 1

By construction, M1 , M2 , . . . , Mn are independent. Thus, the multinomial model is appropriate

as a model for this setup as well if we identify the parameter vector p = (p11 , p21 , p12 , p22 )
with (π1 q11 , π2 q21 , π1 q12 , π2 q22 ). It is easy to see that in fact the two formulations are simple
reparameterizations of each other.
Another alternative experimental setup, which may be more appropriate depending on how the
experiment was conducted, is to fix the sample sizes corresponding to treatment 1 and treatment 2
in advance. We consider this model in Section 10.8.2.

Version: – November 19, 2024

362 hypothesis testing

10.3.2 Independence in the Multinomial Model

Our current interest is in testing whether the treatment and outcome attributes are independent.
We will need to understand how this “hypothesis” translates to restrictions on P that define P0 ,
the parameter space under the null hypothesis. This is given by the following lemma.

Lemma 10.3.2. Let {pkℓ , πk and qkℓ : k, ℓ = 1, 2} be as in Section 10.3.1. Let pk◦ = pk1 + pk2
and p◦ℓ = p1ℓ + p2ℓ for k, ℓ = 1, 2. For an individual, let T denote the treatment and Y denote the
outcome. Then, the following are equivalent.

(a) T and Y are independent

(b) pkℓ = pk◦ p◦ℓ for k, ℓ = 1, 2.

(c) q11 = q21 .

Proof. The proof is left as an exercise. ■

Thus the conjecture or hypothesis of independence that we are interested in testing can be stated
as a constraint on the parametric model, specifically, that the true parameter p is in

P0 = {p = (p11 , p21 , p12 , p22 ) ∈ P : pkℓ = pk◦ p◦ℓ for k, ℓ = 1, 2}.

" #
X11 X12
It is easy to verify that the maximum likelihood estimator of p is p̂ = 1
=
nX
1
n (see
X21 X22
Exercise 9.2.11) It can be similarly verified that the likelihood as a function of p ∈ P0 is maximized
at p̂k◦ = n1 (Xk1 + Xk2 ) and p̂◦ℓ = n1 (X1ℓ + X2ℓ ), or equivalently, q̂11 = q̂21 = n1 (X11 + X21 ).
As we will see later, test statistics are usually a function of maximum likelihood estimators.
Unfortunately, unlike in Example 10.1.1, it is not obvious how we can combine these estimators into
a summary statistic that can be used for testing of independence. The difference in this case is that
even under the null hypothesis, when p ∈ P0 , the actual value of p is unknown, so the distribution
of the MLEs obtained above cannot be fully determined.
There is in fact a widely used ‘solution’ for this problem: Pearson’s χ2 test of independence. The
test statistic used in this test is not difficult to motivate intuitively. However, its exact distribution
is unknown and can be only approximately calculated. We will return to this test towards the end
of this chapter, after discussing some simpler parametric problems and introducing some general
terminology and results.

10.4 testing in the parametric setup : the intuitive approach

So far, in this chapter, we have discussed testing problems with one common theme: the primary
attribute we were interested in was a categorical outcome, and could be modeled using a categorical
distribution (or the Bernoulli distribution in the simplest case). Although we have not yet been able
to derive a satisfactory general approach to testing in this problem, the setup is useful in motivating

Version: – November 19, 2024

 10.4 testing in the parametric setup : the intuitive approach 363

the next set of examples, where instead of a categorical outcome we consider a continuous outcome,
which we model using a Normal distribution.

In this section, we consider the continuous analog of Example 10.1.1, with one set of univariate
i.i.d. observations X1 , X2 , . . . , Xn from the Normal distribution.

10.4.1 Finding a Test Statistic

As we will soon see in Definition 10.4.1, any statistic, or function of the available data, can be
a test statistic in principle. To be of practical use, the distribution of a test statistic should be
completely known under the null hypothesis. In the Bernoulli coin toss example above, P is the
interval [0, 1], P0 is the singleton set {0.5} and the test statistic is the total number of heads in say
n
n tosses is S = Xi . Although the distribution of S (Binomial) generally involves the unknown
P
i=1
p, it is completely known for p = 0.5.

Ideally one wants to find the test statistic that is also “best” in some sense, but often the
optimality of a test is difficult to establish. In Section 10.5, we will describe one principled approach
that often, but not always, leads to a test statistic with some optimality properties. Sometimes,
however, a suitable test statistic is easy to derive intuitively, as in the coin toss example above. In
this section, we consider some specific examples where such tests are available.

Intuitively, a good test statistic should satisfy two main criteria. Firstly, the sampling distribu-
tion of the test statistic should be known when p ∈ P0 . Secondly, it should have the “power” to
distinguish whether the conjecture is true or false (i.e., its distribution should vary substantially
when p ∈ P0 and when it is not). We have seen one important example, namely the Bernoulli
model, where an intuitively appealing test statistic was easy to find, but we have also seen closely
related multinomial models where such test statistics were not available. We will now investigate a
few examples involving the Normal distribution, which has two parameters, to see whether we can
come up with similarly simple and intuitive test statistics. Our goal is to get a sense of what test
procedures typically look like, before discussing a more systematic but somewhat abstract approach
in the next section. Before proceeding, let us formally define the notion of a test statistic in this
context.

Definition 10.4.1. Given observed data X1 , X2 , . . . , Xn , a test statistic is any real-valued

function T (X1 , X2 , . . . , Xn ) of X1 , X2 , . . . , Xn .

10.4.2 The Normal Distribution: Test for Sample Mean when σ is Known

Suppose X ∼ N ormal (µ, σ 2 ) where µ is an unknown mean, but σ is a known standard deviation.
Thus, here the parameter of interest p ≡ µ and parameter space P = R.

Version: – November 19, 2024

364 hypothesis testing

Two-sided test
Suppose we want to test the null hypothesis that µ = a, where a is some known value that has
special meaning in the context of the problem. Thus, here P0 = {a}.
Example 10.4.2. Let X1 , X2 , . . . , Xn be an i.i.d. sample from the distribution X ∼ Normal(µ, 1),
where µ ∈ P = R. We are interested in testing the null hypothesis that µ = 0, or equivalently, that
µ ∈ P0 = {0}. Clearly, an intuitively appealing test statistic is the observed sample mean X. The
further X is from 0, the less confident we would be that the null hypothesis is true. The question is
how far away should it be before we decide that the evidence against the null hypothesis is strong
enough.
Note that although we have assumed σ = 1 and a = 0, this does not lead to any loss in generality.
Suppose the true distribution of X was Normal(µ, σ 2 ), and we were interested in testing the null
hypothesis that µ = a. Consider the transformed observations Zi = (Xi − a)/σ. Then their common
distribution is Z ∼ Normal((µ − a)/σ, 1), and the null hypothesis µ = a ⇐⇒ (µ − a)/σ = 0. ■
n
The sample mean X can thus be viewed as the test statistic T (X1 , X2 , . . . , Xn ) = n1 Xi . In
P
i=1
general, a value of X close to a tends to support the conjecture that µ = a, and a value away from
a tends to make us suspect it. But the strength of the evidence against the conjecture depends
not only on the difference X − a, but also on σ. It is easy to see from results we have already
encountered that if the null hypothesis is true, that is, µ = a, then X ∼ Normal(a, σ 2 /n), and so

√
 
X −a
Ta (X1 , X2 , . . . , Xn ) = n ∼ Normal(0, 1).
σ

The if in the last sentence is important, and bears emphasizing. Another way to express the same
statement is to say that if Y1 , Y2 , . . . , Yn are i.i.d. observations from the Normal(a, σ 2 ) distribution,
then
√
 
Y −a
Ta (Y1 , Y2 , . . . , Yn ) = n ∼ Normal(0, 1).
σ
This distinguishes between the observed data X1 , X2 , . . . , Xn whose assumed distribution depends
on the unknown µ, and the hypothetical random variables Y1 , Y2 , . . . , Yn whose distribution is
assumed to satisfy the null hypothesis.
The known distribution of the test statistic, calculated from the distribution of the hypothetical
data for which the null hypothesis is true, is known as the “null distribution”. It then remains
to compare the observed value of Ta to this known distribution to obtain a p-value. The formal
definition of p-value will be given later, but as noted earlier, we can think of it as the probability
that under the null distribution we will see a value of Ta (Y1 , Y2 , . . . , Yn ) at least as extreme as
the value of Ta (X1 , X2 , . . . , Xn ) observed from the data at hand. Here, “more extreme” is to be
interpreted as values that are even less likely, and thus would have provided even more evidence
against the null hypothesis. In this case, large positive and large negative values of Ta are both
evidence against the null hypothesis, and the symmetry of the distribution of Ta (Y1 , Y2 , . . . , Yn )
suggests that the p-value may be computed as

2P Ta (Y1 , Y2 , . . . , Yn ) ≥ |Ta (X1 , X2 , . . . , Xn )| (10.4.1)

Version: – November 19, 2024

 10.4 testing in the parametric setup : the intuitive approach 365

given the observed value of the test statistic Ta (X1 , X2 , . . . , Xn ). We would reject the null hypothesis
if the p-value obtained in (10.4.1) is small enough (say smaller than 0.05).

One-sided test
The test above conjectures that µ = a, and considers deviations on either side to be departures
that invalidate the null hypothesis. A common variation of the above test is interested in deviations
from the null hypothesis in only one direction. Suppose, as before, that the distribution of X
is Normal(µ, σ 2 ) with known σ 2 and µ ∈ P = R. Further, X1 , X2 , . . . , Xn is an i.i.d. sample
from X, and we are interested in testing the null hypothesis that µ ≤ a, or equivalently, that
µ ∈ P0 = (−∞, a]. Intuitively, the data support the null hypothesis if X ≤ a, and oppose it more
and more strongly the larger X is than a. Here, we would expect the null hypothesis to be rejected
if X − a is large. It would be natural to compute

P (Ta (Y1 , Y2 , . . . , Yn ) ≥ Ta (X1 , X2 , . . . , Xn ) ) ,

where Y1 , Y2 , . . . , Yn ∼ i.i.d. Normal(µ, σ 2 ) and reject if this probability is uniformly small for all
µ ≤ a. We thus define the p-value to be

max P (Ta (Y1 , Y2 , . . . , Yn ) ≥ Ta (X1 , X2 , . . . , Xn ) ) : Y1 , Y2 , . . . , Yn i.i.d. Normal(µ σ 2 ), µ ≤ a .



One can in fact show that this maximum is achieved for µ = a . Let us look at a specific numerical
example to make these ideas concrete.
Example 10.4.3. Students in a probability class were surveyed to obtain the sex and heights of
one of their parents, along with those of that parent’s oldest sibling, provided they had one. When
the siblings compared were both male, the difference in heights (oldest sibling - parent) rounded to
the nearest centimeter were as follows: -5 -5 -4 -2 -2 1 2 2 3 4 4 5 7.
Assume that these values are i.i.d. observations from a Normal distribution with unknown
mean µ and known variance σ 2 = 32 . We want to test the null hypothesis that µ = 0.
√
The observed value of X is 0.769, and thus T0 (X1 , X2 , . . . , Xn ) = 13(0.769/3) = 0.925. The
probability that the random variable T0 (Y1 , Y2 , . . . , Yn ), which follows Normal(0, 1), is larger than
0.925 can be calculated in R as follows.

pnorm(0.925, mean = 0, sd = 1, [Link] = FALSE)

[1] 0.177483

This is the p-value for the one-sided null hypothesis µ ≤ 0. For the two-sided null hypothesis the
p-value would then be 2 × 0.1775 = 0.354. As both p-values are larger than 0.05 we do not reject
either of the null hypotheses. ■
We will discuss the interpretation of the p-value in more detail in Section 10.5, but the value 0.05 is
generally considered a reasonable cutoff between “weak” and “strong” evidence against the null
hypothesis.

Version: – November 19, 2024

366 hypothesis testing

10.4.3 The Normal Distribution: Test for Sample Mean when σ is Unknown

Despite all the elaborate notation, at its core the previous example is not very different from the
Binomial example we started with. In both cases, there is only one unknown parameter of interest,
and the null hypothesis fixes it to a particular known value, making the distribution of the data
essentially known if the null hypothesis holds. Obtaining a test statistic whose distribution would
be known under the null hypothesis is then almost trivial, as literally any function of the data
would satisfy this requirement.

This is rarely the case in realistic hypothesis testing scenarios. Suppose again that X ∼
Normal µ, σ 2 , but now both the mean µ and the standard deviation σ are unknown. Thus, here

parameter of interest is p ≡ (µ, σ ) and parameter space P = R × (0, ∞). Let X1 , X2 , . . . , Xn be

an i.i.d. sample from the distribution X. As before, we want to test the null hypothesis that
µ = a, where a is some known value that has special meaning in the context of the problem.
However, the null hypothesis makes no specific conjecture about the possible values of σ 2 . As
P0 = {a} × (0, ∞) is no longer a singleton set, the distribution of X is not completely specified
when the null hypothesis holds. As the null hypothesis makes a conjecture only about µ and none
about σ, the latter is often referred to as a “nuisance parameter”.

The essential approach remains the same as before. We wish to find a test statistic whose
distribution does not depend on the unknown parameters when the null hypothesis is true. The
test statistic in the earlier case was

√
 
X −a
Ta (X1 , X2 , . . . , Xn ) = n ∼ Normal (0, 1) .
σ

However, this is not a valid test statistic in this case because it cannot even be calculated, as it
involves σ, which is unknown. Intuitively, we may expect to fix this problem by replacing σ with
an estimate, say the sample standard deviation S (see Definition 7.1.5 and Exercise 7.1.8). This
leads to a proper statistic

√
 
X −a
Ta (X1 , X2 , . . . , Xn ) = n .
S

For this modified statistic Ta (X1 , X2 , . . . , Xn ) to be useful as a test statistic, its distribution under
the null hypothesis should be completely known. In general, the distribution of Ta would depend on
µ and σ 2 . If the null hypothesis µ = a is true, then µ is known, but σ 2 is still unknown. However,

Version: – November 19, 2024

 10.4 testing in the parametric setup : the intuitive approach 367

it is easy to see that the distribution of Ta (Y1 , Y2 , . . . , Yn ) does not depend on σ 2 in this case. To
see this, let Y1 , Y2 , . . . , Yn be an i.i.d. sample from Normal(a, σ 2 ) for some arbitrary σ 2 > 0. Note,
 
√ Y − a
Ta (Y1 , Y2 , . . . , Yn ) = n q 
1 Pn 2
n−1 i=1 (Yi − Y )
 
√  Y −a
σ

= n
r 

2 
1 Pn Yi −Y
n−1 i=1 σ

Define Zi = (Yi − a)/σ. So from the above we have

 
√ (Z − 0)
Ta (Y1 , Y2 , . . . , Yn ) = n q Pn
 = T0 (Z1 , Z2 , . . . , Zn ).
1 2
n−1 i=1 ( Z i − Z )

Clearly, Zi ∼ Normal(0, 1) has distribution free of σ 2 . Thus the distribution of Ta (Y1 , Y2 , . . . , Yn )

which is equal to T0 (Z1 , Z2 , . . . , Zn ) is also free of σ 2 . So we can calculate p-values involving it, at
least in principle.
Continuing Example 10.4.3, we can obtain the value of Ta (X1 , X2 , . . . , Xn ), with a = 0, using
the following R code, by first estimating X and S 2 from the data itself.

a <- 0
x <- c(-5, -5, -4, -2, -2, 1, 2, 2, 3, 4, 4, 5, 7)
n <- length(x)
mu_x <- mean(x)
sigma_x <- sqrt(sum((x - mu_x)ˆ2) / (n-1))
T_x <- sqrt(n) * ((mu_x - a) / sigma_x)
T_x

[1] 0.6964513

The value of Ta (X1 , X2 , . . . , Xn ) is thus 0.6965. To obtain the corresponding p-value for the
two-sided null hypothesis µ = 0, we must compute

2P (Ta (Y1 , Y2 , . . . , Yn ) ≥ | 0.6965 | )

when Y1 , Y2 , . . . , Yn are i.i.d. Normal a, σ 2 .

Suppose we did not know the distribution of Ta (Y1 , Y2 , . . . , Yn ), we could approximately compute
this probability by simulating values of Y1 , Y2 , . . . , Yn . The critical observation here is that it does
not matter that we do not know σ 2 , because the distribution of Ta (Y1 , Y2 , . . . , Yn ) does not depend
on σ 2 . Thus, for simulation, any value of σ 2 is as good as any other. Here we choose to simulate

Version: – November 19, 2024

368 hypothesis testing

with σ 2 = 1. Once Y1 , Y2 , . . . , Yn are available, we simply repeat the same calculations as above to
calculate Ta (Y1 , Y2 , . . . , Yn ).

n <- length(x)
T_sim <-
replicate(1000000,
{
y <- rnorm(n, mean = 0, sd = 1)
mu_y <- mean(y)
sigma_y <- sqrt(sum((y - mu_y)ˆ2) / (n-1))
T_y <- sqrt(n) * (mu_y / sigma_y)
T_y
})
uprob <- mean(T_sim >= abs(T_x))
uprob

[1] 0.249762

As before, the p-value is twice the upper tail probability, which in turn is estimated by uprob, the
proportion of cases out of a million simulations where the simulated T0 (Y1 , Y2 , . . . , Yn ) exceeds
0.6965. The estimated p-value is thus 2 × 0.2498 = 0.4995. Of course, this estimated p-value will
vary every time the simulation is run, but it should be reasonably close to the correct answer. In
fact, as this estimate is based on estimating the proportion p from a series of Bernoulli trials, we
can use what we learned in Chapter 9 to obtain a confidence interval for the true p-value. Using
the bernoulliQuadraticCI() and bernoulliLinearCI() functions defined in Section 9.3.3, we
have the following 95% confidence intervals for the p-value, which are essentially identical, which is
not surprising given the large number of replications.

2 * bernoulliQuadraticCI(T_sim >= abs(T_x), level = 0.95)

[1] 0.4978291 0.5012228

2 * bernoulliLinearCI(T_sim >= abs(T_x), level = 0.95)

[1] 0.4978272 0.5012208

In this somewhat special situation, however, this simulation approach is not necessary because the
distribution of Ta (Y1 , Y2 , . . . , Yn ) is well-studied theoretically, with its density function available
in closed form and numerical algorithms available to evaluate its cumulative distribution func-

Version: – November 19, 2024

 10.4 testing in the parametric setup : the intuitive approach 369

tion. Recall that we learned about the t distribution in Chapter 8, and note in particular that
Corollary 8.1.11 immediately implies that in this case,

√ (Y − a)
Ta (Y1 , Y2 , . . . , Yn ) = n ∼ tn−1 .
S
Tail probabilities of the t distribution can be computed in R using the pt() function, which is
analogous to the pnorm() function for the Normal distribution. Thus, the exact p-value in this case
can be computed as follows.

2 * pt(T_x, df = n - 1, [Link] = FALSE)

[1] 0.4994148

As the p-values we have computed for this example are all larger than 0.05, we would not have
rejected the null hypothesis that µ = 0, irrespective of the approach taken.

10.4.4 An Alternative Test Based on the Median

As long as we are considering tests that are intuitively appealing, the test statistics outlined above
are not the only possibilities. For example, the parameter µ in the Normal µ, σ 2 distribution is

the median as well as the mean, so it may be reasonable to define a test statistic based on the
sample median instead of the sample mean. Specifically, to test the null hypothesis that µ = a,
consider the statistic
√ median(X ) − a
 
Ta (X1 , X2 , . . . , Xn ) = n
e .
Se
where
n
1X
Se = |Xi − median(X )| .
n
i=1

Essentially, Tea (X1 , X2 , . . . , Xn ) replaces the sample mean by the sample median in the formula for
Ta (X1 , X2 , . . . , Xn ), with an analogous change in the estimate of scale. We can argue as before that
the distribution of Tea (Y1 , Y2 , . . . , Yn ), when Y1 , Y2 , . . . , Yn are independent Normal a, σ 2 , does

not depend on σ 2 , although it may still depend on n. Unfortunately, no useful representation

is available that allows us to compute its exact tail probabilities. However, we can still use the
simulation approach as before.

Continuing Example 10.4.3, we can obtain the value of Tea (X1 , X2 , . . . , Xn ), with a = 0, as
follows.

Version: – November 19, 2024

370 hypothesis testing

a <- 0
x <- c(-5, -5, -4, -2, -2, 1, 2, 2, 3, 4, 4, 5, 7)
n <- length(x)
mu_x <- median(x)
sigma_x <- sum(abs(x - mu_x)) / n
T_x <- sqrt(n) * ((mu_x - a) / sigma_x)
T_x

[1] 2.232008

The observed value of Tea (X1 , X2 , . . . , Xn ) here is 2.232. To compute the corresponding p-value, we
again simulate values of Te0 (Y1 , Y2 , . . . , Yn ) where Y1 , Y2 , . . . , Yn are i.i.d. Normal (0, 1).

T_sim_median <-
replicate(1000000,
{
y <- rnorm(n, mean = 0, sd = 1)
mu_y <- median(y)
sigma_y <- sum(abs(y - mu_y)) / n
T_y <- sqrt(n) * (mu_y / sigma_y)
T_y
})
uprob <- sum(T_sim_median >= abs(T_x)) / 1000000
uprob

[1] 0.094342

The desired approximation to the p-value in this case is again twice the upper tail probability, i.e.
2 × 0.0943 = 0.1887. As this value is larger than 0.05, we will not reject the null hypothesis that
µ = 0.

One may reasonably wonder which of the above tests is “better”. This is an important question
in general, but it is beyond the scope of this book. In this specific case, the first test can be shown
to have more “power” to identify situations where the null hypothesis does not hold: This means
that if the true value of µ differs from a, then the test based on the mean has higher probability of
rejecting the null hypothesis (by producing a p-value less than 0.05) than the test based on the
median. This holds for all values of µ ̸= a, although the actual probabilities of rejection would
certainly depend on the value of µ. However, this assurance requires Normality of the underlying
measurement. As in Chapter 9, simulation studies can usually provide helpful guidance regarding
the performance of specific tests under various scenarios.

Version: – November 19, 2024

 10.5 the general approach: likelihood ratio test 371

exercises

10.5 the general approach: likelihood ratio test

The examples in the previous section illustrate the “intuitive” approach for developing a test
statistic given a hypothesis of interest. While this approach is useful in many situations, one is often
interested in a principle that may be applied in a general setup, much as the maximum likelihood
principle served for estimation in Chapter 9. We will describe a similar principle for testing in this
section. Although we will not delve into theoretical properties of this approach, we mention two
important results about it: First, the resulting test can be easily shown to be optimal in the special
case where both P0 and P \ P0 are singleton sets (this is a fundamental result in statistics that
is known as the Neyman-Pearson Lemma (see [CasBer90]) as well as more generally for certain
families of distributions. Second, even though the distribution of the resulting test statistic under
the null hypothesis may not always be computable, it can be determined asymptotically under fairly
general conditions (see Wilks’ Theorem, Theorem 10.7.2).

10.5.1 The Likelihood Ratio Statistic

Recall from Chapter 9 that the likelihood function given the sample (X1 , X2 , . . . , Xn ) is defined as

n
Y
L(p; X1 , X2 , . . . , Xn ) = f (Xi | p).
i=1

and the maximum likelihood estimator (MLE) is given by

p̂ ≡ p̂(X1 , X2 , . . . , Xn ) = arg max L(p; X1 , . . . , Xn ),

p∈P

where we use the notation “arg max” to denote the value of the argument p for which the maximum
is obtained. We can similarly define the MLE restricted to the null hypothesis being tested as

p̂0 ≡ p̂0 (X1 , X2 , . . . , Xn ) = arg max L(p; X1 , . . . , Xn ).

p∈P0

We now define the likelihood ratio as

L(p̂0 ; X1 , . . . , Xn )
λ(X1 , . . . , Xn ) = , (10.5.1)
L(p̂; X1 , . . . , Xn )

and the likelihood ratio statistic 1 as

L(p̂; X1 , . . . , Xn )
Λ(X1 , . . . , Xn ) = −2 log λ(X1 , . . . , Xn ) = 2 log . (10.5.2)
L(p̂0 ; X1 , . . . , Xn )

1
which should perhaps be called the log-likelihood ratio statistic, and often is.

Version: – November 19, 2024

372 hypothesis testing

The intuition behind these definitions is as follows. By definition of p̂ and p̂0 , we must have
L(p̂; X1 , . . . , Xn ) ≥ L(p̂0 ; X1 , . . . , Xn ) ≥ 0, hence 0 ≤ λ(X1 , . . . , Xn ) ≤ 1 and thus Λ(X1 , . . . , Xn ) ≥
0. Equality is achieved if the unrestricted MLE p̂ ∈ P0 . If p̂ ̸∈ P0 then Λ(X1 , . . . , Xn ) > 0 gives a
measure of how far p̂ is from P0 (in terms of L). The further p̂ is away from P0 , the less likely it is
that the null hypothesis p ∈ P0 is true. The general principle then is to believe the null hypothesis
if the likelihood ratio λ is close to one, or equivalently, the likelihood ratio statistic Λ is small (close
to zero), and suspect it when Λ is large. The reason for taking log and the including the factor of 2
is that doing so makes the distribution of Λ more convenient, as will become clear in due course.

10.5.2 Type I and Type II Error

It still remains for us to determine how large Λ(X1 , . . . , Xn ) should be before we conclude that the
balance of evidence suggests that the null hypothesis is false. We next present a simple example
where the underlying distribution depends only on one parameter, but is nonetheless helpful in
understanding this question.

Example 10.5.1. Suppose X1 , X2 , . . . , Xn is an i.i.d. sample from Bernoulli(p), where p ∈

P = (0, 1). We are interested in testing the null hypothesis that p = p0 , or equivalently, that
p ∈ P0 = {p0 }. Intuitively, the further the sample proportion p̂ is from p0 , the less confident we
would be that the null hypothesis is true. It is easy to see that in this case Λ(X1 , X2 , . . . , Xn ) is
given by
T /n 1 − T /n
Λ(X1 , X2 , . . . , Xn ) = 2T log + 2(n − T ) log ,
p0 1 − p0
n
where T = Xi (see (10.6.1) below). Therefore Λ is minimised (i.e. Λ = 0) when T = np0 which
P
i=1
matches our intuition. ■

Consider a specific instance in Example 10.5.1 with n = 10 and p0 = 0.5. Then T ∼

Binomial(10, p) distribution. If T = 5, we would obviously have no reason to suspect the null
hypothesis, even though the probability of observing T = 5 when p = 0.5 is only around 0.25.
Even if T = 4 or T = 6, we would not suspect the null hypothesis, as these are still quite plausible
outcomes. However, if T = 0 or T = 10, we would possibly suspect that the null hypothesis is false,
because P (T = 0) = P (T = 10) = 0.0098 is quite low. But where exactly do we draw the line?
To reasonably answer this question, we must distinguish between two types of mistakes we risk
making. Suppose we decide that we will accept (i.e., not reject) the null hypothesis if c1 < T < c2
and reject it otherwise, where c1 and c2 need to be determined. Now imagine that the null hypothesis
(p = 0.5) is indeed true. Suppose we obtain a random sample X1 , X2 , . . . , X10 and perform the test.
Rejecting the null hypothesis would then be a mistake because we would conclude that the null
hypothesis is false even though it is actually true. By convention, such a mistake is called a “Type
I error” or “false positive”. The probability of making a Type I error in this case is

2 −1
cX    10
10 1
P ( T ≤ c1 ) + P ( T ≥ c2 ) = 1 −
k 2
k =c1 +1

Version: – November 19, 2024

 10.5 the general approach: likelihood ratio test 373

For example, if c1 = 2 and c2 = 8, the probability of Type I error is 0.109. If instead c1 = 1 and
c2 = 9, the probability of Type I error is 0.021.
On the other hand, if the null hypothesis is not true, we may still mistakenly accept the null
hypothesis. This is called a “Type II error” or a “false negative”. The probability of making a Type
II error depends on the true value of the parameter p and is given by

2 −1
cX
10 k
 
P (T ≤ c1 ) + P (T ≥ c2 ) = 1 − p (1 − p)10−k .
k
k =c1 +1

For example, if p = 0.75, then with c1 = 2 and c2 = 8, probability of Type II error is 0.474, and
with c1 = 1 and c2 = 9, probability of Type II error is 0.756. Similarly, for p = 0.9, with c1 = 2
and c2 = 8, probability of Type II error is 0.07, and with c1 = 1 and c2 = 9, probability of Type II
error is 0.264.
These calculations illustrate a general principle, that trying to decrease the probability of
Type I error (e.g., by controlling c1 and c2 in this example) generally results in an increase in the
probability of Type II error. The standard approach to resolve this trade-off is a two-step approach:
(a) Obtain a reasonable “test statistic”, such as the likelihood ratio statistic Λ(X1 , . . . , Xn ) defined
above. (b) Choose a threshold for the test statistic,2 beyond which the null hyothesis is declared
to be rejected, in a manner that ensures that the probability of Type I error does not exceed a
pre-determined limit.
An “optimal” choice of test statistic in (a) would ensure that any other test with equal or
lower probability of Type I error will always have higher probability of Type II error. Using
Λ(X1 , . . . , Xn ) as the test statistic ensures such optimality in many situations, provided that it is
possible to compute and control the probability of Type I error for the resulting test. Such a test is
known as the likelihood ratio test.
With this background, we will now proceed to outline a general test procedure, before getting
back to the specific examples cited above.

10.5.3 The p-value for the Likelihood Ratio Test

Let X be a random variable with probability mass function or probability density function f (x),
where f (x) ≡ f (x | p) depends on one or more unknown parameters p ∈ P. Let X1 , X2 , . . . , Xn be
an i.i.d. random sample with common distribution X. In our approach, the test statistic is the
likelihood ratio statistic Λ(X1 , . . . , Xn ), and we reject the null hypothesis if Λ is large. However,
rather than trying to find a rejection cutoff that controls the probability of Type I error, we will
take a more modern approach and introduce the closely related concept of p-value.
Note that Λ(X1 , . . . , Xn ) is also a random variable having a corresponding sampling distribution
of its own. When performing a test, we will work with a specific realisation of this sample. Henceforth,
we will denote this realised value of Λ(X1 , . . . , Xn ) by d, to emphasize that it is a constant for the
purposes of the test.

2
More generally, a set of possible values of the test statistic for which the null hypothesis is rejected.

Version: – November 19, 2024

374 hypothesis testing

Now, the sample X1 , X2 , . . . , Xn was generated with a particular value of p that may or may
not have belonged to P0 . The p-value is concerned with the probabilistic behaviour of the random
variable Λ(X1 , . . . , Xn ) when the underlying parameter does belong to P0 . To distinguish a sample
from such a “thought experiment” from the actual realized sample, imagine a second set of i.i.d.
observations Y1 , Y2 , . . . , Yn from the distribution of X for some p ∈ P.

Definition 10.5.2. The p-value for testing p ∈ P0 based on Λ is defined as

max Pp (Λ(Y1 , . . . , Yn ) ≥ d). (10.5.3)

p∈P0

The notation Pp emphasizes that the probability calculations are done with parameter value
p, and d = Λ(X1 , X2 , . . . , Xn ) is the likelihood ratio statistic calculated from the observed
sample.

Whether we can actually compute the p-value still remains to be seen, and will depend on the
problem. Assuming that it can be, consider the following test procedure.

Definition 10.5.3. (Level α test) Fix 0 < α < 1. Let X1 , X2 , . . . , Xn be an i.i.d. sample
from a population with distribution X. To test the null hypothesis p ∈ P0 at level α, compute
the p-value as above, and reject the null hypothesis if the p-value is less than or equal to α.
Otherwise, accept the null hypothesis.

The following result establishes that the probability of Type I error for this test procedure does not
exceed α. This property is conventionally taken to be defining characteristic of a level α test.

Theorem 10.5.4. For a level α test obtained from the likelihood ratio statistic Λ(X1 , . . . , Xn ) with
p-value computed according to (10.5.3), the probability of Type I error does not exceed α.

Proof. As we are interested in the probability of Type I error, suppose that X1 , X2 , . . . , Xn is a

random sample from X with some parameter value p0 ∈ P0 . We need to show that the probability
of rejection does not exceed α. Let Λ(X1 , . . . , Xn ) be the likelihood ratio statistic. Consider
another random sample Y1 , Y2 , . . . , Yn from X with the same parameter p0 . If the test is rejected,
the p-value must be α or less, and in particular, we must have (holding Λ(X1 , . . . , Xn ) fixed)

PY (Λ(Y1 , . . . , Yn ) ≥ Λ(X1 , . . . , Xn )) ≤ α

as this inequality must hold for all p ∈ P0 and hence in particular for p0 . Therefore, we must have

PX (null hypothesis rejected) ≤ PX (PY (Λ(Y1 , . . . , Yn ) ≥ Λ(X1 , . . . , Xn )) < α) ≤ α.

This last result follows from Lemma10.5.5 below, with Z = Λ(X1 , . . . , Xn ) and W = Λ(Y1 , . . . , Yn ).
■

Version: – November 19, 2024

 10.6 specific examples 375

Lemma 10.5.5. Let Z and W be i.i.d. random variables with distribution function F . Then

PZ (PW (W ≥ Z ) < α) ≤ α.

Proof. The proof is simple if we assume F is continuous and strictly increasing on its support.

PZ ( PW ( W ≥ Z ) < α ) = PZ ( 1 − F ( Z ) < α )
= PZ ( F ( Z ) > 1 − α )
= PZ (Z > F −1 (1 − α))
= 1 − F (F −1 (1 − α))
= 1 − (1 − α ) = α

The result also holds for more general F , but that case requires more careful manipulation, with
the first two equalities in the proof above becoming inequalities. ■

10.6 specific examples

10.6.1 Binomial Test for Proportion

Let us now return to Example 10.5.1, where X1 , X2 , . . . , Xn is an i.i.d. sample from Bernoulli(p),
where p ∈ P = (0, 1). Then T = ni=1 Xi has a Binomial(n, p) distribution. We are interested in
P

testing the null hypothesis that p = p0 .

The likelihood function for this example is

 
n T
L(p; X1 , X2 , . . . , Xn ) = p (1 − p)n−T , p ∈ P,
T

It is easy to see that p̂ = T /n and p̂0 = p0 . Thus the likelihood ratio statistic is

L(p̂; X )
Λ (X ) = 2 log
L(p̂0 ; X )
X/n 1 − X/n
= 2X log + 2(n − X ) log , (10.6.1)
p0 1 − p0

where we interpret 0 log(0) as 0. To compute the p-value for this test for a specific realization of X,
we need to first compute d = Λ(X ), and then compute

Pp ( Λ ( Y ) ≥ d ) ,
0

Version: – November 19, 2024

376 hypothesis testing

where Y has the Binomial(n, p0 ) distribution. Although we cannot express this probability in closed
form, we can explicitly write it as the following sum involving Binomial(n, p0 ) probabilities:
 
X n k
P ( Λ (Y ) ≥ d) = p (1 − p0 )n−k , (10.6.2)
k 0
0≤k≤n:Λ(k )≥d

which can be computed easily using R.

Example 10.6.1. Consider a specific instance of the above experiment where n = 100, and X = 67.
Suppose we want to test the null hypothesis p0 = 0.5. Then the observed value of d is

X/n 1 − X/n 0.67 0.33

2X log + 2(n − X ) log = 2 × 67 log + 2 × 33 log = 11.794
p0 1 − p0 0.5 0.5

The required p-value then is

P ( Λ (Y ) ≥ d)

where Y has a Binomial(100, 0.5) distribution, and can be computed as follows.

d <- 2 * 67 * log(0.67 / 0.5) + 2 * 33 * log((1 - 0.67) / 0.5)

n <- 100
p0 <- 0.5
y <- 0:n
Py <- dbinom(y, size = n, prob = p0)
Ly <- ifelse(y %in% c(0, n),
Inf,
2 * y * log((y/n) / p0) + 2 * (n-y) * log((1-y/n) / (1-p0)))
sum(Py[ Ly >= d ])

[1] 0.0006412485

In the computation above, we need to treat y = 0 and y = n specially, as direct computation of

Λ(y ) fails for operations involving log 0.
Not surprisingly, the p-value is quite small, suggesting that the observed value of X = 67 would
be very unlikely if the null hypothesis p = 0.5 were true. In particular, as the p-value is less that
0.05, the null hypothesis would be rejected at level α = 0.05.
It is instructive to investigate a little further and understand which values of k contribute to
the sum in (10.6.2). Figure 10.1 plots values of Λ(k ) for a range of k centered around 50, along
with the observed value d = Λ(67). The traditional testing approach is to fix α and obtain a
corresponding cutoff for d, usually by looking up the cutoff in a table. Suppose we fix α = 0.05. It
can be verified using R that P (Λ(Y ) ≥ Λ(60)) = 0.057 and P (Λ(Y ) ≥ Λ(61)) = 0.035. Thus, the
highest probability of Type I error not exceeding α = 0.05 that we can achieve in this example
is 0.035, if we reject the null hypothesis when the observed d ≥ Λ(61) = 2.44 (also shown in the
figure).

Version: – November 19, 2024

 10.6 specific examples 377

20

15
Λ(k)

10

Realized value of d
5

Cutoff at level 0.05

0

20 30 40 50 60 70 80

Figure 10.1: Values of Λ(k ) as a function of k for the Binomal test.

As can be seen by inspecting Figure 10.1, the null hypothesis will be rejected at level α = 0.05
either if the realized X is 61 or higher, or if X is 39 or lower. In other words, this test is a two-sided
test. ■

10.6.2 Normal Test for Mean When Variance is Known

Next we continue Example 10.4.2, where X1 , X2 , . . . , Xn is an i.i.d. sample from the Normal (µ, σ 2 )
distribution, where the variance σ 2 is known but the mean µ ∈ P = R is not. We are interested in
testing the null hypothesis µ = µ0 , or equivalently, that µ ∈ P0 = {µ0 }, for a specific value µ0 . It
is easily seen that

1 X
2 log L(µ; X1 , . . . , Xn ) = − log 2π − log σ 2 − (Xi − µ)2
σ2
i

The unrestricted MLE of µ is µ̂ = X, and the restricted MLE is µ̂0 = µ0 . It follows that

L(µ̂; X1 , . . . , Xn )
Λ(X1 , . . . , Xn ) = 2 log
L(µ̂0 ; X1 , . . . , Xn )
= 2 log L(µ̂; X1 , . . . , Xn ) − 2 log L(µ̂0 ; X1 , . . . , Xn )
1 X
 X 
2 2
= ( X i − µ 0 ) − ( X i − X )
σ2
i i
1 h 2 2
i
= nµ 0 − 2nµ 0 X + nX
σ2
 2
n 2 X − µ0
= ( X − µ0 ) = √
σ2 σ/ n

Version: – November 19, 2024

378 hypothesis testing

For a given realization of the sample X1 , X2 , . . . , Xn , we need to first compute d = Λ(X1 , X2 , . . . , Xn )

and then compute the p-value
P (Λ(Y1 , Y2 , . . . , Yn ) ≥ d)

where Y1 , Y2 , . . . , Yn have a Normal µ0 , σ 2 distribution. Now, we know that in that case Z = Yσ/−µ


√0
n
has a Normal (0, 1) distribution, and hence
 2
Y − µ0
Λ(Y1 , Y2 , . . . , Yn ) = √
σ/ n

has a χ21 distribution. The p-value can thus be easily calculated using R. The following code
computes the p-value for the data previously seen in Example 10.4.3.

x <- c(-5, -5, -4, -2, -2, 1, 2, 2, 3, 4, 4, 5, 7)

n <- length(x)
d <- (mean(x) / (3 / sqrt(n)))ˆ2
pchisq(d, df = 1, [Link] = FALSE)

[1] 0.3552259

As in the Binomial example, we can also derive a rejection cutoff for the observed d corresponding
to a fixed level α. For example, with α = 0.05, this cutoff would be the 0.95 quantile of the χ21
distribution, which is

qchisq(0.95, df = 1)

[1] 3.841459

In other words, we reject the null hypothesis that µ = µ0 if

2
√ X − µ0 √

X − µ0
d= √ ≥ 3.84, or equivalently, if n ≥ 3.84 = 1.96.
σ/ n σ

Thus our approach gives the same rule as the intuitive test derived earlier which which rejects the
√  X−µ0 
null hypothesis when the absolute value of Tµ0 (X1 , X2 , . . . , Xn ) = n σ is large. As with
the test for Binomial proportion described earlier, this is a two-sided test. Recall that we have seen
the cutoff of 1.96 before in Section 9.3 where we obtained a confidence interval for µ. This is not a
coincidence, as the 0.95 quantile of χ21 should indeed be the same as the 0.975 quantile (and the
negative of the 0.025 quantile) of Normal (0, 1).
More interestingly, it follows that given i.i.d. data X1 , X2 , . . . , Xn ∼ Normal µ, σ 2 , a particular

value of µ0 will belong to the level (1 − α) confidence interval for µ obtained in Example 9.3.1
if and only if the null hypothesis that µ = µ0 is accepted at level α. This observation applies
more generally, and every hypothesis test can be used to generate a confidence region, and vice
versa. This idea is often useful, especially in situations where a hypothesis test is available, but a
confidence region may not be easy to obtain directly.

Version: – November 19, 2024

 10.6 specific examples 379

10.6.3 One-sided Test for Normal Mean when Variance is Known

An important variation of the above test is when we are interested in testing not a single point
value of µ, but rather a range. For example, we may want to “reject” the null hypothesis only
when the true mean is larger than the conjectured value, but not when it is lower. In this case, the
null hypothsis is actually µ ≤ µ0 , or µ ∈ P0 = (−∞, µ0 ].
Although the unrestricted MLE µ̂ = X remains unchanged, the restricted MLE µ̂0 is now given
by (See Exercise 10.6.1) 
X if X ≤ µ0
µ̂0 = = min(X, µ0 )
µ0 otherwise.

The same calculations as above gives the likelihood ratio statistic as

L(µ̂; X1 , . . . , Xn )
Λ(X1 , X2 , . . . , Xn ) = 2 log
L(µ̂0 ; X1 , . . . , Xn )
1 X
 X 
2 2
= (Xi − µ̂0 ) − (Xi − X )
σ2
i i

n 0 if X ≤ µ0
2
= 2
( X − µ̂ 0 ) =
σ n(X − µ0 )2 /σ 2 otherwise.

We then need to compute the p-value

max Pµ (Λ(Y1 , Y2 , . . . , Yn ) ≥ d),

µ≤µ0

where d = Λ(X1 , X2 , . . . , Xn ) is the observed likelihood ratio statistic.

If X ≤ µ0 , d = 0 and hence the p-value is trivially 1. If X > µ0 , we must also have
d > 0, and so Λ(Y1 , . . . , Yn ) ≥ d if and only if Y > µ0 and n(Y − µ0 )2 /σ 2 ≥ d, or equivalently,
√ √
n(Y − µ0 )/σ ≥ d. Thus,
√ √
Pµ (Λ(Y1 , . . . , Yn ) ≥ d) = Pµ ( n(Y − µ0 )/σ ≥ d)
p
= Pµ (Y − µ0 ≥ σ d/n)
p
= Pµ (Y − µ ≥ σ d/n + µ0 − µ)
√
 
Y −µ µ0 − µ
= Pµ √ ≥ d+ √
σ/ n σ/ n
√ √
   
µ0 − µ µ0 − µ
= P Z ≥ d+ √ = 1−Φ d+ √
σ/ n σ/ n

where Z is a standard normal random variable with cumulative distribution function Φ. Now, as µ
µ0 −µ
√
increases to µ0 , σ/ √ decreases to 0, and Pµ ( Λ (Y1 , . . . , Yn ) ≥ d) increases to 1 − Φ ( d). In other
n
words, the maximum for calculating the p-value is achieved for µ = µ0 , giving p-value

√
 
X − µ0
1 − Φ ( d) = P Z ≥ √ .
σ/ n

Version: – November 19, 2024

380 hypothesis testing

As before, we can compute this p-value using R. To obtain a rejection cutoff, for example with
α = 0.05, we have

qnorm(0.95)

[1] 1.644854

We reject the null hypothesis if the p-value

 
X − µ0
P Z≥ √ < 0.05,
σ/ n

or equivalently,
√ X − µ0 σ
n > 1.645 ⇐⇒ X > µ0 + √ 1.645
σ n
Unlike the case where the null hypothesis allowed a single point value of µ, this is a one-sided test
in terms of X.

10.6.4 Normal Test for Mean When Variance is Unknown

As in the previous examples, suppose we have an i.i.d. sample X1 , X2 , . . . , Xn from the Normal(µ, σ 2 )
distribution, but now in addition to the mean µ, we assume more realistically that the variance
σ 2 is also unknown. Formally, the parameters of the problem are (µ, σ 2 ) ∈ P = R × (0, ∞).
As in Section 10.6.2, we are interested in testing for a specific value of µ. To simplify notation,
we will consider the null hypothesis µ = 0 instead of the more general µ = µ0 . As noted
earlier, this is not really a restriction; to test µ = µ0 , we can simply work with the transformed
data X1 − µ0 , X2 − µ0 , . . . , Xn − µ0 . Under the null hypothesis, the restricted parameter set is
P0 = {0} × (0, ∞).

Unlike the previous examples, we have two parameters in this problem, and we are interested
in testing a null hypothesis that only puts restrictions on one of them. It is easy to see that the
n
unrestricted MLEs of µ and σ 2 are given by µ̂ = X and σ̂ 2 = n1 (Xi − X )2 (see Exercise 9.2.7).
P
i=1

Version: – November 19, 2024

 10.6 specific examples 381

The restricted MLE of µ is of course µ̂0 = 0. It is easy to verify that the restricted MLE of σ 2 is
n n
σ̂02 = n1 (Xi − µ̂0 )2 = n1 Xi2 . It follows that
P P
i=1 i=1

L(µ̂, σ̂ 2 ; X1 , . . . , Xn )
Λ(X1 , . . . , Xn ) = 2 log
L(µ̂0 , σ̂02 ; X1 , . . . , Xn )
= 2 log L(µ̂, σ̂ 2 ; X1 , . . . , Xn ) − 2 log L(µ̂0 , σ̂02 ; X1 , . . . , Xn )
n n
1 X 1 X
= n log σ̂02 + 2 (Xi − µ̂0 )2 − n log σ̂ 2 − 2 (Xi − µ̂)2
σ̂0 i=1 σ̂
i=1
P n  P n 
X 2 ( X − X ) 2
1X 2
n
 i=1 i  n
 i=1 i
 − n log 1
X
= n log 2

Xi + n  n ( X i − X ) − n 
n

n  P 2
 n  P 2

i=1 Xi i=1 (Xi − X )
i=1 i=1
P 2
Xi
= n log P
(Xi − X )2

We are interested in the distribution of Λ(Y1 , . . . , Yn ) when Y1 , Y2 , . . . , Yn has the Normal 0, σ 2

distribution (i.e., µ = 0). In general, this distribution could depend on the other parameters,
namely, σ 2 in this example. As we saw in Section 10.4.3, this does not happen here; that is, the
distribution of Λ(Y1 , . . . , Yn ) does not depend on the value of σ 2 when µ = 0. To see this, notice
that
P 2
Yi
Λ(Y1 , . . . , Yn ) = n log P
(Yi − Y )2
(Yi /σ )2
P
= n log P
(Yi /σ − Y /σ )2
P 2
Zi
= n log P = Λ ( Z1 , . . . , Zn )
( Zi − Z ) 2

where Zi = Yi /σ ∼ Normal (0, 1) has distribution free of σ. In other words,

Pσ2 (Λ(Y1 , . . . , Yn ) ≥ d) = P1 (Λ(Y1 , . . . , Yn ) ≥ d)

for all σ 2 , where Pσ2 denotes probability calculations when Y1 , Y2 , . . . , Yn are i.i.d. from Normal(0, σ 2 ).
In particular, for an observed value d = Λ(X1 , . . . , Xn ), the p-value is given by
 P 2 
Zi
P ( Λ ( Z1 , . . . , Zn ) ≥ d ) = P n log P ≥ d
( Zi − Z ) 2
 P 2 
Zi d/n
= P P ≥e
( Zi − Z ) 2
( Zi − Z ) 2
P 
= P P 2 ≤ e−d/n
Zi

Version: – November 19, 2024

382 hypothesis testing

where Z1 , Z2 , . . . , Zn are i.i.d. from Normal(0, 1). Now, simple algebraic manipulation shows that

n n n
X X X 2
Zi2 = (Zi − Z + Z )2 = (Zi − Z )2 + nZ
i=1 i=1 i=1

2 √
Recall that Z ∼ Normal(0, n1 ) and so nZ = ( nZ )2 ∼ χ21 . An application of Theorem 8.1.10
2
further tells us that (i) (Zi − Z )2 ∼ χ2n−1 , and (ii) Z is independent of (Zi − Z )2 . Recall
P P

Example 8.1.5 to note that the χ2n distribution is the same as the Gamma( n2 , 12 ) distribution. In
other words,
n
n−1 1 1 1
   
X 2
(Zi − Z )2 ∼ Gamma , and nZ ∼ Gamma , ,
2 2 2 2
i=1

independently of each other. It then follows from Example 5.5.11 that

n n
( Zi − Z ) 2 (Zi − Z )2
P P
i=1 i=1
n = P
n 2
Zi2 (Zi − Z )2 + nZ
P
i=1 i=1

has the Beta 2 ,2

n−1 1
distribution. Thus, given an observed value for the likelihood ratio statistic

n
Xi2
P
i=1
d = Λ(X1 , . . . , Xn ) = n log n ,
(Xi − X )2
P
i=1

the corresponding p-value can be computed as

 
P U ≤ e−d/n

where U has the Beta 2 ,2

n−1 1
distribution.

For the data in Example 10.4.3, the p-value can be computed as follows.

x <- c(-5, -5, -4, -2, -2, 1, 2, 2, 3, 4, 4, 5, 7)

n <- length(x)
mu_x <- mean(x)
d <- n * log( sum(xˆ2) / sum((x - mu_x)ˆ2) )
d

[1] 0.5151229

pbeta(exp(-d/n), 0.5 * (n-1), 0.5)

Version: – November 19, 2024

 10.6 specific examples 383

[1] 0.4994148

To obtain the rejection region at level α, we need to start with the lower α quantile, say b, which
for α = 0.05 can be computed as follows.

qbeta(0.05, 0.5 * (n-1), 0.5)

[1] 0.7165366

We reject at level α if e−d/n ≤ b, that is, if

(Xi − X )2
P
2 ≤ b,
(Xi − X )2 + nX
P

or equivalently, if
2 2
(Xi − X )2 + nX 1
P
X
= 1 + nP ≥ ,
(Xi − X )2 2
P
(Xi − X ) b
2
n−1 1
 
X
⇐⇒ 1 ≥ −1
(Xi − X )2
P
n−1
n b
s
n−1 1
 
X
⇐⇒ ≥ −1
S n b
s
√ X 1
 
⇐⇒ n ≥ (n − 1) −1 ,
S b

where S 2 is the sample variance (see Definition 7.1.5 in Chapter 8). Note the similarity of the
rejection region with the one in Section 10.6.2 which addressed the analogous problem when σ 2 is
√
known; as in that case, this is also a two-sided test. It follows from Corollary 8.1.11 that n XS has
the tn−1 distribution when µ = 0, and so p-values of cutoffs can be computed using the t distribution
as well. It can be numerically verified that a two-sided t-test rejection cutoff is equivalent to the
Beta test cutoff as follows.

b <- qbeta(0.05, 0.5 * (n-1), 0.5)

sqrt((n-1) * (1/b - 1))

[1] 2.178813

qt(0.025, df = n-1, [Link] = FALSE)

[1] 2.178813

Version: – November 19, 2024

384 hypothesis testing

Before moving on, we make some observations and remarks about this important example. It
should be easy to see that for the general null hypothesis µ = µ0 , the Beta statistic derived above
generalizes to

(Xi − X )2 (Xi − X )2
P P
e−Λ(X1 ,...,Xn )/n = P 2
=
(Xi − µ0 ) (Xi − X )2 + n(X − µ0 )2
P

and the corresponding statistic with the tn−1 distribution is

√ X − µ0
n .
S
The latter form of the statistic is more conventional, not least due to its intuitively appealing
√
form and similarity with the Normal statistic n X−µ σ
0
in the case when σ 2 is known, as it can be
motivated as a modification where σ is replaced by an estimator when it is unknown. As with the
2

case where σ 2 is known, the test can be adjusted for a one-sided null hypothesis of the form µ ≤ µ0
or µ ≥ µ0 , resulting in a one-sided rejection region in terms of the t statistic. Even though this test
is derived under the assumption of Normality, it performs well for moderate departures from this
assumption, and is among the most common statistical tests used in practice.
Another equivalent form of the test is also important. In its squared form, the statistic

n ( X − µ0 ) 2
S2
has a F1,n−1 distribution when µ = µ0 (see Example 8.1.7). This form and its generalizations are
also commonly used. A typical example is the problem of testing for equality of means of two or
more populations, which we discuss next.

10.6.5 The Two-sample Test for Equality of Population Means

Connect this to a test of independence of treatment vs outcome, where outcome is now continuous
rather than discrete. Or maybe keep as-is and use another section or subsection to make the
connection and tie up other loose ends.
Hypothesis tests may be used to compare two samples to each other to see if the populations
they were derived from are similar. This is of particular use in many applications. For instance:
Are the political preferences of people of one region different from another? Are test scores at one
school better than those at another school? These questions could be approached by taking random
samples from each population and comparing them with each other.
Suppose X1 , X2 , . . . , Xm is an i.i.d. sample from a distribution X ∼ Normal(µ1 , σ12 ) and
Y1 , Y2 , . . . , Yn is an i.i.d. sample from a distribution Y ∼ Normal(µ2 , σ22 ) independent of the Xj
variables. Assume that µ1 and µ2 are unknown, as well as σ12 and σ22 . How might we test the null
hypothesis that µ1 = µ2 against the alternative hypothesis µ1 ̸= µ2 ? It turns out that this problem
is not easy to solve in its general form, and we will only consider the problem with the additional
assumption that σ12 = σ22 . We will use σ 2 to denote this common variance.

Version: – November 19, 2024

 10.6 specific examples 385

Based on the examples we have seen so far, we might guess that a good test would be based
on X − Y , as this should be close to 0 if the null hypothesis that µ1 = µ2 were true. It is simple
to check that X − Y has a Normal distribution with mean µ1 − µ2 and variance m 1
+ n1 σ 2 . If

we can obtain an estimator S 2 of σ 2 that is independent of X − Y and has a suitably scaled χ2

distribution, we could plausibly expect that

X −Y
q
1 1
S m + n

would have a t distribution. In fact, it is not difficult to find such an estimator S 2 . Consider the
standard unbiased estimators of σ 2 from the two independent populations:
m
1 X
S12 = (Xi − X )2
m−1
i=1
n
1 X
S22 = (Yj − Y )2
n−1
j =1

It follows from Theorem 8.1.10 that (m − 1)S12 /σ 2 ∼ χ2m−1 independently of X and (n − 1)S22 /σ 2 ∼
χ2n−1 independently of Y . This suggests the natural estimator
 
m n
1 X X
S2 =  (Xi − X )2 + (Yj − Y )2 
m+n−2
i=1 j =1

which is independent of both X and Y (and hence X − Y ). By Example 5.5.6, (m + n − 2)S 2 /σ 2 ∼

χ2m+n−2 , and so S 2 is an unbiased estimator of σ 2 . It follows from Corollary 8.1.11 that

X −Y
q
1 1
S m + n

has a tm+n−2 distribution. This is the basis for the standard two-sided test for this problem, and
this solution generalizes to the one-sided null hypotheses µ ≤ µ0 and µ ≥ µ0 in the usual way. As
in the test for the mean of a single Normal population, the squared version of the statistic
mn 2
m+n (X − Y )
S2
has the F1,m+n−2 distribution.

Version: – November 19, 2024

386 hypothesis testing

It turns out, not surprisingly, that this is equivalent to the test derived using the likeli-
hood ratio statistic. The likelihood function for this model is (suppressing the dependence on
X1 , . . . , Xm , Y1 , . . . , Yn for brevity)
m n
1 1 1 1
Y  Y  
L ( µ1 , µ2 , σ )
2
= √ exp − 2 (Xi − µ1 ) 2
√ exp − 2 (Yj − µ2 ) 2
nπσ 2σ nπσ 2σ
i=1 j =1
  
m + n m n
1 1
  X X 
= √ exp − 2  (Xi − µ1 )2 + (Yj − µ2 )2 
nπσ  2σ 
i=1 j =1

It is easy to verify, again following the approach of Exercise 9.2.7, that the unrestricted MLEs are

µ̂1 = X
µ̂2 = Y
 
m n
1 X X
σ̂ 2 = (Xi − µ̂1 )2 + (Yj − µ̂2 )2  ,
m+n
i=1 j =1

and therefore
  m n

2+ 2 
P P
m + n

 ( Xi − µ̂ 1 ) ( Yj − µ̂ )
2 
1
 
 m+n  

 i=1 j =1
L(µ̂1 , µ̂2 , σ̂ 2 ) = √ exp −

nπ σ̂ 2  m n
P 
(Xi − µ̂1 )2 + (Yj − µ̂2 )2 

 P 

 

i=1 j =1

m+n
2 −
= nπeσ̂ 2
.

Similarly, under the null hypothesis µ1 = µ2 , the restricted MLEs are

1
µ̂0 = µ̂10 = µ̂20 = (mX + nY )
m+n
 
m n
1  X X
σ̂02 = (Xi − µ̂0 )2 + (Yj − µ̂0 )2 
m+n
i=1 j =1

and therefore

− m + n
L(µ̂10 , µ̂20 , σ̂02 ) = nπeσ̂02 2
.

The likelihood ratio statistic is thus

L(µ̂1 , µ̂2 , σ̂ 2 ) σ̂02

 
Λ(X1 , . . . , Xm , Y1 , . . . , Yn ) = 2 log = (m + n) log ,
L(µ̂10 , µ̂20 , σ̂02 ) σ̂ 2

Version: – November 19, 2024

 10.6 specific examples 387

which is essentially equivalent to

m n
(Xi − µ̂0 )2 + (Yj − µ̂0 )2
P P
σ̂02 i=1 j =1
= Pm n .
σ̂ 2
− X )2 )2
P
(Xi + (Yj − Y
i=1 j =1

The test rejects the null hypothesis when this ratio is large, which is in line with our intuition
because if the null hypothesis is not true, then we would expect the deviations in the numerator
(from a common mean) to be larger than the deviations in the denominator (from group-specific
means). The denominator clearly has a χ2m+n−2 distribution (by Theorem 8.1.10 followed by
Example 5.5.6), but the distribution of the numerator is not immediately obvious. To simplify the
numerator, note that
m
X m
X m
X
(Xi − µ̂0 )2 = (Xi − X + X − µ̂0 )2 = (Xi − X )2 + m(X − µ̂0 )2 + 0.
i=1 i=1 i=1

Recall that µ̂0 = 1

m+n (mX + nY ), and so

(m + n)X − mX − nY 1 n
X − µ̂0 = = (mX + nX − mX − nY ) = (X − Y ),
m+n m+n m+n

and consequently,
m m
X X mn2
(Xi − µ̂0 )2 = (Xi − X )2 + (X − Y )2 .
(m + n)2
i=1 i=1

Analogously,
n n
X X m2 n
(Yj − µ̂0 )2 = (Yj − Y )2 + (X − Y )2
(m + n)2
j =1 j =1

and therefore the numerator term simplifies to

m n m n
X X X X mn
(Xi − µ̂0 )2 + (Yj − µ̂0 )2 = (Xi − X )2 + (Yj − Y )2 + (X − Y )2 .
m+n
i=1 j =1 i=1 j =1

As before, it follows from Theorem 8.1.10 and the discussion above that X − Y is independent of
the denominator
Xm Xn
(Xi − X )2 + (Yj − Y )2
i=1 j =1

and that mmn+n (X − Y ) has a χ1 distribution. Arguing similarly as we did in Section 10.6.4, it
2 2

follows that
m n
(Xi − X )2 + (Yj − Y )2
P P
σ̂ 2 σ̂02
i=1 j =1
= 1/ 2 = P m n
σ̂02 σ̂
(Xi − X )2 + (Yj − Y )2 + mmn 2
P
+n (X − Y )
i=1 j =1

Version: – November 19, 2024

388 hypothesis testing

has the Beta( m+2n−2 , 12 ) distribution (test rejects null when small), and

mn mn
m+n (X − Y )2 m+n (X − Y )
2

1
 =
S2
P
(Xi − X )2 + (Yj − Y )2
P
m+n−2

has the F1,m+n−2 distribution (test rejects null when large), and both are equivalent to the “intuitive”
two-sided t test derived above.

10.6.6 Equality of Population Means with Different Variances

The examples above may give the impression that the likelihood ratio approach always leads to a
useful test statistic. The next example, which is a simple and natural extension of the previous
problem, shows that this is not so.
Recall the setup of the previous problem, where we suppose that X1 , X2 , . . . , Xm is an i.i.d.
sample from X ∼ Normal(µ1 , σ12 ) and Y1 , Y2 , . . . , Yn is an independent i.i.d. sample from Y ∼
Normal(µ2 , σ22 ) independent of the Xj variables. We are still interested in testing the null hypothesis
that µ1 = µ2 against the alternative hypothesis µ1 ̸= µ2 , but this time we do not wish to assume
that σ12 = σ22 .
The unrestricted MLEs of the parameters are straightforward, as the X and Y observations
do not share any common parameters. The detailed calculations for the restricted case are more
involved. The details are left as an exercise, but it can be shown that the MLEs satisfy the following
equations:

2 X + nσ̂ 2 Y
mσ̂20 10
µ̂0 = µ̂10 = µ̂20 = 2 + nσ̂ 2
mσ̂20 10
m
2 1 X
σ̂10 = (Xi − µ̂0 )2
m
i=1
n
2 1X
σ̂20 = (Yj − µ̂0 )2
n
j =1

An exact solution can be obtained, although in practice an iterative approach works well. It can
be further shown that the likelihood ratio statistic can be expressed as follows in terms of these
estimates as follows.

L(µ̂1 , µ̂2 , σ̂12 , , σ̂22 )

 2   2 
σ̂10 σ̂20
Λ(X1 , . . . , Xm , Y1 , . . . , Yn ) = 2 log = m log + n log .
L(µ̂10 , µ̂20 , σ̂10 , σ̂20 )
2 2 2
σ̂1 σ̂22

This is where the usual procedure breaks down. Unlike in the previous examples, the distribution
of this quantity is not completely determined when the null hypothesis holds, because it depends
on the unknown ratio σ12 /σ22 . Thus, even using simulation to estimate the p-value is not feasible.
As alluded to earlier, one benefit of the likelihood ratio test is that even when the distribution
of the statistic is difficult to study, a powerful result gives its asymptotic distribution under fairly
general conditions. Applied to this problem, this result says that as m, n → ∞, the distribution of

Version: – November 19, 2024

 10.7 testing for goodness of fit 389

Λ(X1 , . . . , Xm , Y1 , . . . , Yn ) converges to a χ21 distribution under the null hypothesis. This is indeed
true; however, for small sample sizes m and n, calculating p-values using this null distribution leads
to substantially larger probability of Type I error than nominally specified. A modification works
quite well in practice, but the details are beyond our scope. For us, this example serves to illustrate
the limitations of the likelihood ratio test approach.

exercises

Ex. 10.6.1. Consider an i.i.d. sample X1 , X2 , . . . , Xn ∼ Normal µ, σ 2 . Show that the restricted

MLE µ̂0 for µ ∈ (−∞, µ0 ] is now given by


X if X ≤ µ0
µ̂0 = = min(X, µ0 )
µ 0 otherwise.

Hint: Notice that the likelihood function L(µ; X1 , . . . , Xn ), as a function of µ, is proportional to a

normal density with mean X, and is therefore maximum at X and symmetrically decreasing on
either side of it.

10.7 testing for goodness of fit

We now return to the problems discussed earlier in this chapter, in Sections 10.2 and 10.3, which we
left largely unresolved. We start with the “goodness of fit” problem, and specifically Example 10.2.1
which we can generalize as follows.

Consider a random experiment (such as rolling a die) which has k possible outcomes. As
in Example 3.2.12, let pj represent the probability that any individual trial results in the j-th
outcome, and let Xj represent the number of the n trials that result in the j-th outcome. The
joint distribution of the random variables (X1 , X2 , . . . , Xk ) is then a multinomial distribution with
parameters n and p = (p1 , p2 , . . . , pk ). Formally, the parameter space for the problem is
 
 k
X 
P= p = (p1 , p2 , . . . , pk ) : 0 ≤ pj ≤ 1 for all j, pj = 1 .
 
j =1

We wish to test the null hypothesis that p = p0 for some p0 = (p01 , p02 , . . . , p0k ). In Example 10.2.1,
where the experiment was rolling a die, we had k = 6 and p0 = (1/6, 1/6, . . . , 1/6).

As P0 = {p0 } is a singleton set, the MLE p̂0 under the null hypothesis is simply p0 . As we saw
in Exercise 9.2.11, the unrestricted MLE p̂ of p is given by the coordinatewise sample proportions,
that is, p̂j = Xj /n.

Version: – November 19, 2024

390 hypothesis testing

From the multinomial distribution obtained in Example 3.2.12, it follows easily that the
likelihood function, given observed data X1 , X2 , . . . , Xk , has the form

k
n! Y X
L(p; X1 , X2 , . . . , Xk ) = pj j , for p ∈ P
X1 !X2 ! . . . Xk !
j =1

and so, suppressing the dependence of L(p) on X1 , X2 , . . . , Xk , we have

k
X k
X
log L(p) = log n! − log Xj ! + Xj log pj , for p ∈ P.
j =1 j =1

Substituting p̂ and p̂0 , we have

k k
L(p̂) X p̂j X Xj
Λ(X1 , . . . , Xk ) = 2 log =2 Xj log =2 Xj log .
L(p̂0 ) p̂0j np0j
j =1 j =1

It is conventional in this context to define Ej = np0j , representing the expected value of Xj . This
leads to the test statistic
k
X Xj
Λ(X1 , . . . , Xk ) = 2 Xj log . (10.7.1)
Ej
j =1

Example 10.7.1. In a survey, a class of n = 71 students were asked what their birth month was,
and the following summary results were obtained:

January 7 February 4 March 5 April 8

May 2 June 6 July 6 August 7
September 3 October 8 November 9 December 6

Using this data, we wish to test the null hypothesis that all days in the year are equally likely as
birthdays. We will ignore the possibility of leap years for simplicity and assume that there are
365 days in a year. We also assume that the class represents an i.i.d. sample from some larger
population. We can set up the problem and calculate the test statistic in R as follows.

X <- c(7, 4, 5, 8, 2, 6, 6, 7, 3, 8, 9, 6)
n <- sum(X)
p0 <- c(31, 28, 31, 30, 31, 30, 31, 31, 30, 31, 30, 31) / 365
E <- n * p0
Lambda_x <- 2 * sum(X * log(X / E))
Lambda_x

[1] 9.008024

To compute the p-value for this test, we need the distribution of Λ(Y1 , . . . , Yk ) when (Y1 , Y2 , . . . , Yk )
follow the multinomial distribution with parameters n and p0 . Unfortunately, this distribution is
not easy to compute explicitly. However, as the distribution of (Y1 , Y2 , . . . , Yk ) is completely known

Version: – November 19, 2024

 10.7 testing for goodness of fit 391

(because P0 has only one element, p0 ), we can easily simulate values of Λ(Y1 , . . . , Yk ) and estimate
the p-value. We can do this, using 1000000 replications, as follows.

Lambda_sim <-
replicate(1000000,
{
Y <- rmultinom(1, size = n, prob = p0)
2 * sum(Y * log(Y / E), [Link] = TRUE) # Use [Link] to allow for Y = 0
})
uprob <- sum(Lambda_sim >= Lambda_x) / 1000000
uprob

[1] 0.648674

A remarkable result, a version of which is stated without proof below in Section 10.7.1, tells us
that even if we had been unable to estimate the exact p-value, the asymptotic distribution of
Λ(Y1 , . . . , Yk ) (as the sample size n → ∞) is known, and is in fact χ2k−1 . This asymptotic result
allows us to obtain an approximate p-value for this example as follows.

pchisq(Lambda_x, df = 11, [Link] = FALSE)

[1] 0.6211516

While not exactly the same, this is reasonably close to the exact p-value estimated using simulation.
■

10.7.1 Asymptotic Distribution of the Likelihood Ratio Test Statistic

Much of the appeal of the likelihood-based approach we have explored in this chapter comes from a
very powerful result that states that the null distribution of the likelihood ratio test statistic can
be obtained asymptotically, even if it cannot be computed exactly for any finite sample size. In
simplified form, the result may be stated as follows.

Theorem 10.7.2. (Wilks’ Theorem) Let X1 , X2 , . . . , Xn be an i.i.d. sample from a distribution

X with a probability mass function or probability density function f (x), where f (x) = f (x |
p1 , p2 , . . . , pd ) = f (x | p) depends on one or more unknown parameters p = (p1 , p2 , . . . , pd ) ∈ P ⊂
Rd for some d ≥ 1. We are interested in testing the null hypothesis that p ∈ P0 , where P0 is a
proper subset of P. Then, under certain conditions, the distribution of the likelihood ratio statistic
Λ(X1 , X2 , . . . , Xn ) defined in (10.5.2) converges in distribution to the χ2k distribution as n → ∞,
where k is the number of independent constraints that P0 puts on P, or equivalently, the difference
in the number of independent parameters in P and P0 .

The proof of this result is beyond the scope of this book. In fact, even a proper statement would be
unnecessarily complicated. For our purposes, it is sufficient to know that the result is applicable

Version: – November 19, 2024

392 hypothesis testing

in most situations, as long as two important conditions are met: The number of independent
parameters in P and P0 are fixed finite numbers, and no p ∈ P0 is on the “boundary” of P. We
will not get into the precise details of what it means for p to be in the boundary of P, but in the
goodness of fit example, an obvious way for this to happen would be if one of the components p0j
equals 0. If that were the case, the result would not be applicable.
To appreciate the power of this result, note that it applies even when the null parameter space
P0 is not a singleton set. Recall that to solve a testing problem we need two things: to find a
suitable test statistic, and then to find its null distribution. The likelihood ratio approach often gives
us a test statistic in situations where no natural candidate is available. In Example 10.7.1 above,
this approach gave a test statistic that we would probably not consider to be natural. However,
once we determine the test statistic, finding its null distribution is simple; even though this null
distribution is “unkown” in the sense that we cannot identify it as a standard distribution, it is
still possible to simulate from it because P0 is a singleton set, so any probabilities related to that
distribution can be computed as precisely as we wish.
The situation is fundamentally different when P0 contains multiple (usually infinitely many)
values. The likelihood ratio test statistic need not have a single “null distribution”, but rather a
different one for every p0 ∈ P0 , making the simulation approach impractical. Theorem 10.7.2 comes
to our rescue in such sitations, giving us a single null distribution that is at least asymptotically
valid. We will see an example of such a test in Section 10.8.

10.7.2 The Standard χ2 Test for Goodness of Fit

We conclude this section with a discussion of a much more well known test for the goodness of fit
problem, given by the following test statistic.

k
X (Xj − Ej )2
T (X1 , . . . , Xk ) =
Ej
j =1

This is an intutively appealing test statistic. As with the likelihood ratio test described above,
the exact distribution of T (X1 , . . . , Xk ) cannot be computed explicitly, but can be studied using
simulation for any specific problem. This distribution also converges to the same χ2k−1 distribution
as n → ∞, which as we will see soon, is not a coincidence. This asymptotic result can be proved
without appealing to the more general Theorem 10.7.2, although the proof is still beyond the scope
of this book. For Example 10.7.1, the resulting test statistic and the corresponding p-value can be
obtained as follows.

T_x <- sum((X-E)ˆ2 / E)

T_x

[1] 8.110577

Version: – November 19, 2024

 10.7 testing for goodness of fit 393

pchisq(T_x, df = 11, [Link] = FALSE)

[1] 0.7033655

One may wonder, given the same asymptotic null distribution of Λ(X1 , . . . , Xk ) and T (X1 , . . . , Xk ),
whether they are related to each other. The answer is that they are indeed related. To see this,
define εj = Xj − Ej , so that we can write (10.7.1) as

k
X Xj
Λ(X1 , . . . , Xk ) = 2 Xj log
Ej
j =1
k  
X εj
= 2 (Ej + εj ) log 1 +
Ej
j =1

εj
Now, for x close to 0, we can write log(1 + x) ≈ x − 12 x2 , so we can write, provided Ej ≈ 0,

k
" !#
2
X εj 1 εj
Λ(X1 , . . . , Xk ) ≈ 2 (Ej + εj ) − (10.7.2)
Ej 2 Ej2
j =1

It is easy to see that

!
2
1 εj 1 (Xj − Ej )2
 
εj Xj − Ej
(Ej + εj ) − = (Xj − Ej ) + 1−
Ej 2 Ej2 2 Ej Ej

Substituting in (10.7.2) and noting that Ej = n and hence (Xj − Ej ) = 0, we have

P P P
Xj =

k 
1 X (Xj − Ej )2
 
Xj − Ej
Λ(X1 , . . . , Xk ) ≈ 0 + 2 1−
2 Ej Ej
j =1

εj Xj −Ej
This approximation is of course only valid when Ej = Ej ≈ 0. In that case, we can further
Xj −Ej
approximate Λ(X1 , . . . , Xk ) by assuming that 1 − Ej ≈ 1, to get

k
X (Xj − Ej )2
Λ(X1 , . . . , Xk ) ≈ = T (X1 , . . . , Xk ).
Ej
j =1

ε p
It is not particularly difficult to show that Ejj −→ 0 as n → 0, and thus establish using Slutsky’s
theorem that Λ(X1 , . . . , Xk ) and T (X1 , . . . , Xk ) have the same asymptotic distribution.

A natural next question is to ask which of these tests is better in terms of their power to identify
situations where the null hypothesis does not hold. The results we have cited so far do not give a
clear answer, but simulation studies can be used to get an indication.

Version: – November 19, 2024

394 hypothesis testing

10.8 testing for independence of categorical attributes

Next we consider the problem of testing whether two categorical attributes are independent. There
are several formulations of this problem that give potentially different answers. We start with the
multinomial formulation discussed earlier in Section 10.3.

10.8.1 The Multinomial Model

Recall that in Section 10.3, we described a multinomial model for this problem parameterized by
the parameter vector p = (p11 , p21 , p12 , p22 ). This model is natural when the units studied in the
problem can be viewed as an i.i.d. sample from some population, as would be appropriate in an
observational study. An alternative formulation that is more appropriate for randomized controlled
trials is parameterized by p = (π1 , π2 , q11 , q12 , q21 , q22 ), where π1 and π2 are the probabilities of
a specific unit being assigned to treatment 1 and treatment 2 respectively, and q1ℓ and q2ℓ are
corresponding conditional probabilities of outcome ℓ. As the two formulations are equivalent (as
long as units are allocated treatment independently), we will only consider the first formulation.

We have already obtained (see Section 10.3.2) maximum likelihood estimates of p under the
unconstrained model, as well as under the null hypothesis of independence which restricts the
parameter values to
P0 = {p : pkℓ = pk◦ p◦ℓ for k, ℓ = 1, 2} .

To obtain the likelihood ratio statistic, we can follow the same calculations as in the goodness of fit
problem, to write the test statistic as

2 2
L(p̂) XX p̂
Λ(X11 , X12 , X21 , X22 ) = 2 log =2 Xkℓ log kℓ
L(p̂0 ) p̂0,kℓ
k =1 ℓ=1

Substituting the maximum likelihood estimates of p̂ and p̂0 , we have

2 X
2
X Xkℓ /n
Λ(X11 , X12 , X21 , X22 ) = 2 Xkℓ log .
(Xk1 + Xk2 )(X1ℓ + X2ℓ )/n2
k =1 ℓ=1

As in the case of the goodness of fit test, it is conventional to write this statistic as

2 X
2
X Xkℓ
Λ(X11 , X12 , X21 , X22 ) = 2 Xkℓ log , (10.8.1)
Ekℓ
k =1 ℓ=1

where Ekℓ = n1 (Xk1 + Xk2 )(X1ℓ + X2ℓ ) = np̂k◦ p̂◦ℓ is interpreted as the “expected” value of Xkℓ if
the null hypothesis is true. The more popular Pearson’s χ2 test of independence, with test statistic

2 X
2
X (Xkℓ − Ekℓ )2
T (X11 , X12 , X21 , X22 ) = 2 ,
Ekℓ
k =1 ℓ=1

Version: – November 19, 2024

 10.8 testing for independence of categorical attributes 395

can be similarly viewed as an approximation of Λ(X11 , X12 , X21 , X22 ). Both these tests can be
easily generalized to situations where there more than two treatments or more than two outcomes.
Unlike the goodness of fit problem, the null parameter space P0 is no longer a singleton set.
Also, unlike in the examples involving the Normal distribution, the distribution of the test statistic
does not become independent of the choice of p0 ∈ P0 . This effectively rules out the simulation
approach to obtain the null distribution, as there are infinitely many choices of p0 which need to be
considered if we wish to compute the p-value using (10.5.3).
Fortunately, Theorem 10.7.2 is still applicable in this case. As the number of independent
parameters is 3 in P and 2 in P0 , both the likelihood ratio statistic and Pearson’s test statistic
have the χ21 distribution asymptotically. In general, the degrees of freedom of the asymptotic null
distribution will depend on the number of treatments and outcomes.

10.8.2 Binomial Model with Fixed Row Margins

A common strategy when designing clinical trials such as the one in Example 10.3.1 is to fix the
number of individuals in each treatment group in advance. In other words, the numbers n1 and n2
to be given treatment 1 and treatment 2 respectively are fixed in advance.
This still leaves open the question of how to choose the n1 individuals
  to be given treatment
n
1. A natural choice is to choose them randomly, uniformly from the possibilities. Such an
n1
allocation scheme forms the basis of the approach described in Section 10.8.3.
Here, we consider the model where the treatment attribute is not random at all, but rather
fixed in advance. One may think of this as comparing two different populations, based on samples
of size n1 and n2 . In this setup, the null hypothesis of independence of treatment and outcome can
be reinterpreted to mean that the distribution of the outcome attribute does not depend on the
population from which the individual comes. It is similar in that sense to the two-sample test for
equality of population means discussed in Section 10.6.5. Alternatively, this model can be thought
to have been derived from the multinomial model by conditioning on the treatment attribute of all
individuals. In terms of Lemma 10.3.2, we have

Yi = 1 | Ti = 1 ∼ Bernoulli(q11 ),
Yi = 1 | Ti = 2 ∼ Bernoulli(q21 ).

By exchangeability of individuals within each treatment group, this is equivalent to conditioning on

the totals n1 and n2 in each treatment group. This interpretation allows us to use this (conditional)
model wherever the multinomial model is appropriate. In this conditional model, the parameter π1
is not required, so the parameter space reduces to P = {p = (q11 , q21 ) : 0 ≤ q11 , q21 ≤ 1}, and and
the null hypothesis is represented by the subset P0 = {p ∈ P : q11 = q21 }.
The likelihood function for this model is simply the product of two Binomial likelihoods, and the
maximum likelihood estimators of p are suitable sample proportions. Specifically, the unconstrained
estimators are given by q̂k1 = Xk1 /nk for k = 1, 2, and the estimators under the null hypothesis are
q̂0,11 = q̂0,21 = (X11 + X21 )/n. It is left as an exercise to verify that the likelihood ratio statistic

Version: – November 19, 2024

396 hypothesis testing

is exactly the same as (10.8.1), the statistic obtained in the multinomial model. Although both P
and P0 have one fewer independent parameter in this case, P0 is still not a singleton set, so the
null distribution will depend on the unknown common value of q11 = q21 . Again, the asymptotic
distribution of the test statistic is χ21 .

10.8.3 Fisher’s Exact Test of Independence

The tests of independence derived above are approximate tests, valid asymptotically. An interesting
and somewhat natural extension of this setup does allow us to obtain an exact test, although its
formulation is such that it does not fit nicely into our usual parametric setup. Nonetheless, we end
this section with a discussion of this formulation because it provides a useful perspective on the
testing problem in general. The resulting test is known as Fisher’s exact test of independence.
Technically, the setup of Fisher’s test can be obtained from the multinomial model by condi-
tioning on a certain event. In this sense, it is an extension of the previous two models. Specifically,
the multinomial model in Section 10.3.1 defines a probability distribution on the set of all 2 × 2
tables of the form
Outcome 1 Outcome 2 Total
Treatment 1 X11 X12 N1◦
Treatment 2 X21 X22 N2◦
Total N◦1 N◦2 n
where Xkℓ , k, ℓ = 1, 2 are non-negative integer counts, Nk◦ and N◦ℓ are row and column totals, and
N◦ℓ = n, the total number of participants. The row and column totals Nk◦
P P P
Xkℓ = Nk◦ =
k,ℓ k ℓ
and N◦ℓ are random, although the total sum n is fixed as the total number of units (the sample
size) is fixed in advance. The model considered in Section 10.8.2, as we have noted above, can be
derived from the multinomial model by conditioning on the treatment attribute of each participant,
or equivalently by conditioning on N1◦ and N2◦ . The interpretation of this conditioning from the
perspective of the random experiment is that the number of individuals in each treatment group
(for instance, the number of individuals given the placebo and the vaccine in Example 10.3.1) is
fixed in advance. To derive Fisher’s exact test, we further condition on the column totals N◦1 and
N◦2 .
If both row and column totals are fixed in advance, then it is immediate that any one element
of the table determines the others. This solves one of our problems, namely, finding a test statistic:
without loss of generality, we can take X11 to be our test statistic, as it completely defines the entire
table. In the multinomial setup, the conditional distribution of X11 given the row and column
totals turns out to be the familiar hypergeometric distribution (Example 2.3.1).
It is not, however, immediately obvious how the conditioning on column totals can be interpreted
from the perspective of the underlying random experiment. Fixing the total number of individuals
to be assigned treatment 1 and treatment 2 in advance is reasonable. However, it is completely
unreasonable to expect that the number of individuals who have outcome 1 and outcome 2 would
also be fixed in advance. To link the conditional model above to a reasonable experimental setup,
we have to view the experiment from a different, nonparametric perspective.

Version: – November 19, 2024

 10.8 testing for independence of categorical attributes 397

Recall that there are four possible treatment-outcome combinations for each individual or unit,
namely (1, 1), (1, 2), (2, 1), and (2, 2), identified respectively with the four matrices
" # " # " # " #
1 0 0 0 0 1 0 0
, , , and .
0 0 1 0 0 0 0 1

As in Lemma 10.3.2, let (Ti , Yi ) denote the treatment and outcome pair for the i-th individual.
The treatment attribute Ti for each unit is usually random, but the nature of the randomness
depends on the type of experiment. For observational studies, individuals are sampled from a
population and both “treatment” and “outcome” attributes are observed; neither are under the
control of the experimenter. However, for controlled trials such as the vaccine trial described in
Example 10.3.1, the treatment is assigned randomly as part of the experiment, typically by ensuring
that all participants are equally likely to get a particular treatment. The outcome Yi is also random,
presumably in a way that depends on the individual involved. In the vaccine example, the outcome
may depend on age, gender, and other attributes of a participant that are not available to us. The
outcome is also possibly affected by the treatment assigned; in fact, in the vaccine trial, we hope
that being vaccinated makes an individual less likely to become affected with the disease.
However, we are interested in testing the hypothesis of independence of the two attributes,
which posits that the outcome is not affected by the treatment, even if it does depend on the
individual. How can we formulate this idea as a probability model? The multinomial model with
fixed row totals described in Section 10.8.2 assumes a parametric model where the distribution
of Yi | Ti = t depends only on the parameter q11 or q21 , depending on whether t = 1 or t = 2.
Fisher’s exact test uses a different probability formulation which in principle allows the distribution
of Yi | Ti = t to vary from individual to individual, without explicitly providing a parametric model.
We describe this formulation next.
Normally, probability statements are interpreted in terms of the outcome in repeated perfor-
mances of an experiment. Here, repeating the experiment may mean selecting a new sample of
units (participants) via some random sampling mechanism, randomly assigning them treatments,
and observing the outcomes. However, let us suppose that we simplify the process of repeating the
experiment by skipping the first step: instead of selecting a new set of units on which to perform
the experiment, suppose we use the same set of n participants. However, we do still randomize
the
 allocation
 of treatments, by randomly selecting a new subset of size n1 uniformly from the
n
possibilities, to receive treatment 1. Here n1 , the number of participants who get treatment
n1
1, is assumed to be fixed in advance, and thus n1 can be viewed as the first row total N1◦ in
the multinomial formulation. Of course, we cannot actually perform this experiment and observe
the outcomes on the same units again (after all, someone who has been vaccinated cannot be
un-vaccinated), but we can conjecture about what could have happened if the experiment had been
performed with this new treatment assignment.
In general, we cannot say what the outcome would have been, as they could have changed if
different treatments had been received. However, suppose we restrict ourselves to the case when
the outcome is independent of the treatment, which is what the position of the skeptic would be.
Interpreting the notion of independence literally rather than probabilistically, we can then say

Version: – November 19, 2024

398 hypothesis testing

that for each unit, the outcome depends only on the individual and not the treatment, so would
have remained the same as the outcome observed with the original assignment. In other words,
for the i-th unit if we had Yi = 1 in the original experiment, we would again have Yi = 1 in the
new hypothetical experiment, regardless of whether the value of Ti had changed. For each such
hypothetical experiment then, we can recreate the summary 2 × 2 table without requiring any new
information other than the new treatment assignments. An important point, which is easy to see, is
that the row and column totals of these tables remain unchanged from the original by construction.
The row totals are the number of units assigned treatments 1 and 2, which are always n1 and
n − n1 . The column totals are the total number of units with outcome 1 and 2, which also remain
unchanged, provided that outcomes are not affected by treatment assignment.
The argument above gives us a very concrete (if still somewhat abstract) procedure to test the
null hypothesis (the skeptic’s conjecture) that treatment does not affect outcome: If this conjecture
were true, then we can randomly choose a hypothetical treatment assignment to get a random
summary table, forming a probability distribution whose sample space consists of all 2 × 2 summary
tables with the given marginal row and column totals, or equivalently, just its first entry X11 . Even
if we could not say anything more about this distribution, we could always simulate as many such
tables as we wanted to get an empirical distribution of X11 . As it happens, we can actually say
more, because the distribution of X11 again turns out to be the same hypergeometric distribution.
Finally, to decide whether or not to reject the null hypothesis, we use the same approach as
earlier. Depending on the problem at hand, departure from the null may be indicated either by
high values of X11 , or low values of X11 , or both. Accordingly, the test is one-sided or two-sided.
For one-sided tests, the p-value is given by the corresponding tail probability of the null distribution
starting from the observed value of X11 . For two-sided tests, the computation is less obvious as
the null distribution is not symmetric. Following the principle described in Example 10.6.1, the
p-value in this case is obtained by adding up the individual probabilities of all outcomes in the null
distribution that are at most as likely as the observed value of X11 .

Version: – November 19, 2024

S O M E M AT H E M AT I C A L D E TA I L S
A
a.1 transformation of continuous random variables- jacobian

Suppose the random variable X : S → R is a continuous random variable, with probability density
function fX : R → R. Let g : R → R and Y = g (X ). In general it may be hard to find the
distribution of Y . For some specific class of g the random variable Y will also be a continous random
variable and we can calculate its probability density function. We recall the method discussed in
Section 5.3. One immediately observes that the distribution function of Y is given by

P (Y ≤ y ) = P (g (X ) ∈ (−∞, y ]) = P (X ∈ g −1 (−∞, y ]).

Thus the above formula provides a theoretical expression for the distribution function of Y provided
for all y the function g is such that g −1 (−∞, y ] is an event. Now, let us assume that g is strictly
increasing and differentiable function with g ′ being continuous and g ′ (x) > 0 for all x ∈ R. This
implies that g −1 : R → R exists and is differentiable. The distribution function of Y is given by
Z g −1 (y )
P (Y ≤ y ) = P (g (X ) ≤ y ) = P (X ≤ g −1 (y )) = fX (x)dx.
−∞

From the above, using the fundamental theorem of calculus, we see that Y has a probability density
function fY : R → given by
dg −1
fY (y ) = (y )fX (g −1 (y )), (A.1.1)
dy
for all y ∈ R. Now, let us assume that g is strictly decreasing and differentiable function with
g ′ being continuous and g ′ (x) < 0 for all x ∈ R. This implies that g −1 : R → R exists and is
differentiable. The distribution function of Y is given by
Z ∞
P (Y ≤ y ) = P (g (X ) ≤ y ) = P (X ≥ g −1 (y )) = fX (x)dx.
g −1 (y )

From the above, using the fundamental theorem of calculus, we see that Y has a probability density
function fY : R → given by
dg −1
fY ( y ) = − (y )fX (g −1 (y )), (A.1.2)
dy
for all y ∈ R. We now present the above deductions as a theorem below.

399

Version: – November 19, 2024

400 some mathematical details

Theorem A.1.1. Let X : S → R be a continuous random variable with probability density

function fX (·) with I = {x ∈ R : fX (x) > 0}. Let g : I → R be a differentiable function
on I such that g ′ : I → R is continuous with g ′ (·) ̸= 0 on I. Then the random variable
Y = g (X )a has a density fY : R → R given by

dg −1

|

 dy (y ) | fX (g −1 (y )), y ∈ Range(g )
fY (y ) = (A.1.3)

0

otherwise


or equivalently 
 f (y ) = 1
f (x), with y = g (x), x ∈ I
 Y |g ′ (x)| X


(A.1.4)


0

otherwise.
a
We can assume without loss of generality that X (s) ∈ I for all s ∈ S so that Y (s) = g (X (s)) is
well defined for all s ∈ S.

Example A.1.2. Let X ∼ Uniform (−1, 1). Let g : (−1, 1) → R given by g (x) = x2 and Y = g (X ).
Observe that g is differentiable on (−1, 1) with g ′ : (−1, 1) → R given g ′ (x) = 2x. Again g ′ (0) = 0
so Theorem A.1.3 is not applicable. As we have seen before we can calculate the probability density
function of Y . We first calculate the distribution function of Y .

0 y ≤ 0
P (Y ≤ y ) =
1 y ≥ 1

For 0 < y < 1, using the probability density function of X we have

√
√ √ y
1
Z
1
2
P (Y ≤ y ) = P (X ≤ y ) = P (− y ≤ X ≤ y) = dz = y 2 .
√
− y 2

We note that the distribution function of Y is piecewise differentiable and hence Y has a probability
density function given by 
 1 y − 21 y ∈ (0, 1),
fY ( y ) = 2
0 otherwise

In the above example the transformation g was not one-one and hence was not invertible. We
note that the function g has a well defined inverse in the interval (−1, 0) and (0, 1). Intuitively one
should be able to apply Theorexm A.1.3 on each of these intervals. The next theorem formalises
this.

Version: – November 19, 2024

 A.1 transformation of continuous random variables- jacobian 401

Theorem A.1.3. Let X : S → R be a continuous random variable with probability density

function fX (·) with I = {x ∈ R : fX (x) > 0}. Let g : I → R be a differentiable function on
I such that

(a) Z = {x ∈ I : g ′ (x) = 0} is finite.

(b) g ′ : I \ Z → R is continuous with g ′ (·) ̸= 0 on I \ Z.

Let B = {g (x) : x ∈ I \ Z}. Then Zy = {x ∈ I \ Z : g (x) = y} is necessarily a finite

non-empty set for y ∈ B and the random variable Y = g (X ) has a density fY : R → R
given by 
f (x )
f (y ) = xi ∈Zy |gX′ (x i)| , with y ∈ B
 P
 Y

i

(A.1.5)


0

otherwise.

a.1.1 Multiple Continuous Random Variables and the Jacobian

An alternate way of viewing the change of distribution formula in the previous subsection is as the
familiar u-substitution from calculus. If g (x) is a differentiable function with differentiable inverse
and Y = g (X ) then using the subsitution y = g (x) (so x = g −1 (y )) we have
Z
fY (y ) dy = P (Y ∈ A)
A
= P (g (X ) ∈ A)
= P (X ∈ g −1 (A))
Z
= fX (x) dx
g −1 (A)
dg −1
Z
= fX (g −1 (y )) (y ) dy
A dy

dg −1
But if −1 dy for every event A, then the integrands must be
R R
A fY (y ) dy = A fX (g (y )) dy
−1
the same and fY (y ) = fX (g −1 (y )) dgdy (y ).

When multiple random variables are involved a similar formula may be derived using the
multivariate change of variables formula invovling the Jacobian. Recall the following result from
multivariate calculus:

Version: – November 19, 2024

402 some mathematical details

Theorem A.1.4. Let S, T ⊂ Rn and let h : S → T be an invertible function

h ( y1 , y2 , . . . , yn ) = ( x1 , x2 , . . . , xn )

∂ (x1 ,x2 ,...,xn )

Suppose the Jacobian J = ∂ (y1 ,y2 ,...,yn )
exists and is never zero. Then
Z Z Z
··· f (x1 , x2 , . . . , xn ) dx1 dx2 . . . dxn
h(A)
Z Z Z
= ··· f (x1 , x2 , . . . , xn ) | J (y1 , y2 , . . . , yn ) | dy1 dy2 . . . dyn
A

where in the final line it is understood that the xj variables have been written in terms of
y1 , y2 , . . . , yn via the h−1 function.

Now let n ≥ 1 and suppose X1 , X2 , . . . , Xn are random variables with a joint density fX (x1 , x2 , . . . , xn ).
Let h : S → T be as in the theorem above and define an Rn -valued random vector

(Y1 , Y2 , . . . , Yn ) = h(X1 , X2 , . . . , Xn ).

Let A be an event in Rn . Then,

P ((Y1 , Y2 , . . . , Yn ) ∈ A) = P (h(X1 , X2 , . . . , Xn ) ∈ A)
= P ((X1 , X2 , . . . , Xn ) ∈ h−1 (A))
Z Z Z
= ··· fX (x1 , x2 , . . . , xn ) dx1 dx2 , . . . dxn
h−1 (A)
Z Z Z
= ··· fX (x1 , x2 , . . . , xn ) | J (y1 , y2 , . . . , yn ) dy1 dy2 , . . . dyn
A

where, as in the prior theorem, the xj variables are understood to have been written in terms of
y1 , y2 , . . . , yn .
Let fY (y1 , y2 , . . . , yn ) represent the joint density for the (Y1 , Y2 , . . . , Yn ). Since that density is
defined by the equation
Z Z Z
P ((Y1 , Y2 , . . . , Yn ) ∈ A) = ··· fY (y1 , y2 , . . . , yn ) dy1 dy2 , . . . dyn
A

it must be that fY (y1 , y2 , . . . , yn ) = fX (x1 , x2 , . . . , xn ) | J (y1 , y2 , . . . , yn ) which provides an

equation relating the joint density of the (X1 , X2 , . . . , Xn ) with the joint density of (Y1 , Y2 , . . . , Yn ).

Example A.1.5. Let X1 , X2 ∼ Uniform((0, 1)) be independent random variables. Let Y1 = X1 + X2

and let Y2 = X1 − X2 . This is a globally invertable linear transformation for which X1 = 21 Y1 + 12 Y2
and X2 = 21 Y1 − 12 Y2 . The Jacobian is the determinant

Version: – November 19, 2024

 A.1 transformation of continuous random variables- jacobian 403

" #
∂ ( x1 , x2 ) 1/2 1/2 1
J= = det =− .
∂ (y1 , y2 ) 1/2 −1/2 2

Since the joint density for the X variables is


1 if 0 < x1 < 1 and 0 < x2 < 1
fX ( x1 , x2 ) =
0 otherwise

then the joint density for the Y variables is


1/2 if 0 < y1 + y2 < 2 and 0 < y1 − y2 < 2
fY (y1 , y2 ) = fX (x1 , x2 )|J (y1 , y2 )| =
0 otherwise

where the region in which the density is non-zero is the square with corners (0, 0), (1, 1), (2, 0),
and (1, −1) in the (y1 , y2 )-plane. In particular we could use this joint density to provide another
derivation of the density for the sum of two independent Uniform(0, 1) random variables. That is
simply the marginal distribution of Y1 alone which we can now calculatate.
When 0 < y1 < 1 we have

Z ∞
fY1 (y1 ) = fY (y1 , y2 ) dy2
−∞
1
Z y1
= dy2
−y1 2
= y1

while when 1 < y1 < 2 we have

Z ∞
fY1 (y1 ) = fY (y1 , y2 ) dy2
−∞
1
Z 2−y1
= dy2
y1 −2 2
= 2 − y1

In other words, we have the familiar result


y1

 if 0 < y1 < 1
fY1 (y1 ) = 2 − y1 if 1 < y1 < 2

0 otherwise



which was proven in Example 5.5.2 using a different technique. ■

The change of variables formula requries that the number of X-variables be equal to the
number of Y -variables in order for the Jacobian determinant to exist. For some problems this may

Version: – November 19, 2024

404 some mathematical details

not originally be the case, but this problem can often be alleviated by inserting extra variables.
For instance, consider Exercise 5.5.4 from Chapter 5. In that problem (X1 , X2 , X3 ) are given as
independent and uniformly distributed on (0, 1). The problem asks for the value of P (X1 X3 < X22 ).
While that problem can be solved using the techniques from that chapter, it can also be solved
using the Jacobian method.

Example A.1.6. Let Y1 = X1 X3 and let Y2 = X22 . Note that these are the quantities of interest
in the probability we are asked to compute. To use the Jacobian technique we will also introduce
Y3 = X1 simply to maintain an equal number of variables. On the region X1 , X2 , X3 ∈ (0, 1) where
the density is non-zero, this transformation is invertible. Solving for the X-variables gives: X1 = Y3 ,
√
X2 = Y2 , and X3 = YY31 . Therefore,
 
0 0 1
∂ ( x1 , x2 , x3 ) 0  = − √1 .
= det  0 √1
 
J=  2 y2
∂ (y1 , y2 , y3 )  2 y2 y3
1
y3 0 − yy12
3

Since the joint density for the X-variables is fX (x1 , x2 , x3 ) = 1 whenever 0 < x1 , x2 , x3 < 1
(and 0 otherwise) we have

√

y1
 √1
2 y2 y3 if 0 < y1 < 1 and 0 < y2 < 1 and 0 < y3 <1
fY (y1 , y2 , y3 ) = fX (x1 , x2 , x3 )|J (y1 , y2 , y3 )| =
0 otherwise

which may be written more compactly as


 √1
2 y2 y3 if 0 < y1 < y3 < 1 and 0 < y2 < 1
fY (y1 , y2 , y3 ) =
0 otherwise

At that point we may calculate the desired probability using the new joint density.

P (X1 X3 < X22 ) = P (Y1 < Y2 )

Z Z Z
= fY (y1 , y2 , y3 ) dy1 dy2 dy3
{y1 <y2 }
1Z y3 Z 1
1
Z
= √ dy2 dy1 dy3
0 0 y1 2 y2 y3
√
1
Z 1 Z y3
y1
= − dy1 dy3
0 0 y 3 y3
2√
Z 1
= 1− y3 dy3
0 3
5
=
9
■

Version: – November 19, 2024

 A.2 strong law of large numbers 405

a.2 strong law of large numbers

In this section we shall state and prove the strong law of large numbers.

Theorem A.2.1. (Strong Law of Large Numbers) Let X1 , X2 , . . . be a sequence of

i.i.d. random variables. Assume that X1 has finite mean µ and finite variance σ 2 . Let
A = {limn→∞ X n = µ}. Then
P (A) = 1. (A.2.1)

As remarked in Chapter 8, the above results states that the convergence of sample mean to µ
actually happens with Probability one. This mode of convergence of the sample mean to the true
mean is called “convergence with probability 1.” We define it precisely below.

Definition A.2.2. A sequence X1 , X2 , . . . is said to converge with probability one to a

random variable X if A = {limn→∞ X n = X}.

P (A) = 1. (A.2.2)

The following notation

w.p.1
Xn −→ X

is typically used to convey that the sequence X1 , X2 , . . . converges with probability one to X.

As alluded earlier that this is a stronger mode of convergence. We prove it in the next
proposition.

Proposition A.2.3. Let X1 , X2 , . . . be a sequence of random variables on a sample space S.

Suppose Xn converges to a random variable X with probability 1 then Xn converges to a random
variable X in probability.

Proof- Let ϵ > 0 and δ > 0 be given. We need to show ∃N such that

P (|Xm − X| > ϵ) < δ, ∀m ≥ N . (A.2.3)

Let A = {ω ∈ S : limn→∞ Xn (ω ) = X}. We are given that

P (A) = 1. (A.2.4)

Suppose we denote, for η > 0 and n ≥ 1,

Aηn = {ω ∈ S : |Xn (ω ) − X (ω )| ≤ ϵ }.

then
A = ∩η>0 ∪∞ ∞
k =1 ∩n=k An .
η

Version: – November 19, 2024

406 some mathematical details

This can be verified using the fact that ω ∈ A if and only if for all η > 0, there is a k ≡ k (ω ) such
that
|Xn (ω ) − X (ω )| ≤ ϵ, ∀ n ≥ k.

For m ≥ 1, define Bm
ϵ = ∩∞ Aϵ . Note
n=m n

+1 , (A.2.5)
ϵ ϵ
Bm ⊂ Bm

for all m ≥ 1. So by Exercise 1.1.13, we have

lim P (Bm
ϵ
) ↑ P (∪∞
m = 1 Bm ) .
ϵ
(A.2.6)
m→∞

As A ⊂ ∪∞
m=1 Bm , using (A.2.4) we have 1 = P (A) ≤ P (∪m=1 Bm ) ≤ 1. So
ϵ ∞ ϵ

P (∪∞
m=1 Bm ) = 1.
ϵ
(A.2.7)

By (A.2.6) and (A.2.7) ∃N such that

ϵ0
P ( Bm ) > 1 − δ, ∀m ≥ N .

ϵ ⊂ Aϵ ,
As Bm m
P (Aϵm ) > 1 − δ, ∀m ≥ N .

Therefore by considering the complement of Aϵm we obtain (A.2.3). ■

We will need a technical Lemma regarding convergence in probabilty which we state and prove
below.
Lemma A.2.4. Suppose a sequence of random variables Xn is such that

p p
Xn −→ X and Xn −→ Y

for some random variables X, Y then P (X = Y ) = 1.

Proof- Let k ≥ 1. Let Ak = {| X − Y |≥ k1 }. Notice that Ak ⊂ Ak+1 and ∪∞k =1 Ak = {X ̸= Y }.
p p
Let k ≥ 1, δ > 0 be given. As Xn −→ X and Xn −→ Y , (applying Definition 8.2.6 with ϵ = 2k 1
),
there exists N such that for all n ≥ N

1 1
   
δ δ
0 ≤ P | Xn − X |> < and 0 ≤ P | Xn − Y |> < . (A.2.8)
2k 2 2k 2

Using the triangle inequality we observe that | X − Y |≤| X − Xn | + | Xn − Y | for all n ≥ 1. So,

1 1
Ak ⊂ {| Xn − X |> } ∪ {| Xn − X |> } (A.2.9)
2k 2k
for all n ≥ 1. Combining (A.2.8) and (A.2.9) we have (using any n ≥ N )

1 1
   
δ δ
0 ≤ P ( Ak ) ≤ P | Xn − X |> +P | Xn − Y |> ≤ + = δ.
2k 2k 2 2

Version: – November 19, 2024

 A.2 strong law of large numbers 407

As δ > 0 was arbitrary we have P (Ak ) = 0. Further by Exercise 1.1.13,

P (X ̸= Y ) = lim P (Ak ) = 0.
k→∞

Hence P (X = Y ) = 1. ■

Proof of Theorem A.2.1(Special Case)- We provide a complete proof of Theorem A.2.1 in the
special case when the random variables are i.i.d Bernoulli (p) random variables. We will proceed in
two steps.

Step 1: X n converges with probability one to a random variable X.

Let S = lim supn→∞ X n and S = lim inf n→∞ X n . Clearly,

0 ≤ S ≤ S ≤ 1.

Fix ϵ > 0, then for every k define

Xk + Xk+1 . . . + Xk+n−1
Nk = inf{n ∈ N : ≥ S − ϵ}.
n

The random variable Nk , in some sense, measures how close we are to S and our main effort will
be to control the size Nk . It is easy to see that Nk is finite a.e. and are all identically distributed
(because of independence of Xi ). Hence we can choose an m such that P (Nk > m) < ϵ for all k.
Define random variables Yk and NkY by the following mechanism:
(
Xk if Nk ≤ m
Yk = (A.2.10)
1 if Nk > m
Yk + Yk+1 . . . + Yk+n−1
NkY = inf{n ∈ N : ≥ S − ϵ}. (A.2.11)
n

Clearly NkY ≤ Nk and if k is such that Nk ≥ m then NkY = 1 (since setting Yk = 1 ensures
that we are above S − ϵ immediately). So we have

NkY ≤ m. a.e.

So for large enough n ∈ N we can break up nk=1 Yk into pieces of lengths atmost M such that
P

the average over each piece is atleast S − ϵ. Then finally stop at the n-th term. Then it is clear
that,
Xn
Yk ≥ (n − m)(S − ϵ). (A.2.12)
k =1

By our choice of m

E (Yk ) = E (Xk 1(Nk ≤ m)) + P (Nk > m) < E (Xk ) + ϵ = E (X ) + ϵ,

Version: – November 19, 2024

408 some mathematical details

for any k. Take expectations in (A.2.12) and use the above inequality to obtain

n(E (X ) + ϵ) ≥ (n − m)(E (S ) − ϵ).

Divide by n and first let n → ∞ followed by ϵ → 0, to get

E (S ) ≤ E (X ). (A.2.13)

Let X
fk = 1 − Xk . Applying the above argument to X
e (verify this) we have

e ).
E (Se) ≤ E (X

Since S = −Se this implies

E (S ) ≥ E (X ). (A.2.14)

Now, S ≤ S a.e. So only way (A.2.14) and (A.2.13) can hold only if S = Sa.e. Therefore limn→∞ X n
exists almost everywhere and let us call it X. This completes step 1.
Step 2: We shall now use the Weak Law of Large numbers (Theorem 8.2.1), along with
Proposition A.2.3, and Lemma A.2.4 to complete the proof. The weak law implies that

p
X n −→ µ as n → ∞.

From Step 1, we know that

w.p.1
X n −→ X as n → ∞.

Proposition A.2.3 then implies that

p
X n −→ X as n → ∞.

Finally Lemma A.2.4 implies P (X = µ) = 1. Therefore

w.p.1
X n −→ µ as n → ∞.

■
Proof of Theorem A.2.1(General Case) The essence of the proof is contained in the special
case proven above. We provide a sketch of the proof.
Case 1:(0 ≤ X ≤ 1) An imitation of Step 1 of the proof for Bernoulli p random variables will
show that there is a limit. Step 2 of the above proof follows readily.
Case 2: Bounded Case When the random variable X is bounded, i.e. | X |≤ M for some
Xi −M
M > 0. One can consider Y = X−M 2M and Yi = 2M . As 0 ≤ Y ≤ 1 then one can use Case 1 for
Yi to establish that there is a limit. Step 2 of the above proof follows readily.
Case 3: (General Case by Truncation) One fixes α, β > 0 and defines

(β )
S (α) = min{S, α}, X (β ) = max{X, −β} and Xk = max{Xk , −β} ∀k ∈ N.

Version: – November 19, 2024

 A.2 strong law of large numbers 409

The above quantities are all bounded. One imitates Step 1 of the above proof and this will result in
w.p.1
inequalities depending on α, β. One then allows α, β approach infinity to establish that X n −→ X
for a random variable X. Step 2 of the above proof follows readily. We refer the reader to [AS09]
for the complete proof. ■

Version: – November 19, 2024

410 some mathematical details

Version: – November 19, 2024

TA B L E S
B

411

Version: – November 19, 2024

412 tables

0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09
0.0 0.5000 0.5040 0.5080 0.5120 0.5160 0.5199 0.5239 0.5279 0.5319 0.5359
0.1 0.5398 0.5438 0.5478 0.5517 0.5557 0.5596 0.5636 0.5675 0.5714 0.5753
0.2 0.5793 0.5832 0.5871 0.5910 0.5948 0.5987 0.6026 0.6064 0.6103 0.6141
0.3 0.6179 0.6217 0.6255 0.6293 0.6331 0.6368 0.6406 0.6443 0.6480 0.6517
0.4 0.6554 0.6591 0.6628 0.6664 0.6700 0.6736 0.6772 0.6808 0.6844 0.6879
0.5 0.6915 0.6950 0.6985 0.7019 0.7054 0.7088 0.7123 0.7157 0.7190 0.7224
0.6 0.7257 0.7291 0.7324 0.7357 0.7389 0.7422 0.7454 0.7486 0.7517 0.7549
0.7 0.7580 0.7611 0.7642 0.7673 0.7704 0.7734 0.7764 0.7794 0.7823 0.7852
0.8 0.7881 0.7910 0.7939 0.7967 0.7995 0.8023 0.8051 0.8078 0.8106 0.8133
0.9 0.8159 0.8186 0.8212 0.8238 0.8264 0.8289 0.8315 0.8340 0.8365 0.8389
1.0 0.8413 0.8438 0.8461 0.8485 0.8508 0.8531 0.8554 0.8577 0.8599 0.8621
1.1 0.8643 0.8665 0.8686 0.8708 0.8729 0.8749 0.8770 0.8790 0.8810 0.8830
1.2 0.8849 0.8869 0.8888 0.8907 0.8925 0.8944 0.8962 0.8980 0.8997 0.9015
1.3 0.9032 0.9049 0.9066 0.9082 0.9099 0.9115 0.9131 0.9147 0.9162 0.9177
1.4 0.9192 0.9207 0.9222 0.9236 0.9251 0.9265 0.9279 0.9292 0.9306 0.9319
1.5 0.9332 0.9345 0.9357 0.9370 0.9382 0.9394 0.9406 0.9418 0.9429 0.9441
1.6 0.9452 0.9463 0.9474 0.9484 0.9495 0.9505 0.9515 0.9525 0.9535 0.9545
1.7 0.9554 0.9564 0.9573 0.9582 0.9591 0.9599 0.9608 0.9616 0.9625 0.9633
1.8 0.9641 0.9649 0.9656 0.9664 0.9671 0.9678 0.9686 0.9693 0.9699 0.9706
1.9 0.9713 0.9719 0.9726 0.9732 0.9738 0.9744 0.9750 0.9756 0.9761 0.9767
2.0 0.9772 0.9778 0.9783 0.9788 0.9793 0.9798 0.9803 0.9808 0.9812 0.9817
2.1 0.9821 0.9826 0.9830 0.9834 0.9838 0.9842 0.9846 0.9850 0.9854 0.9857
2.2 0.9861 0.9864 0.9868 0.9871 0.9875 0.9878 0.9881 0.9884 0.9887 0.9890
2.3 0.9893 0.9896 0.9898 0.9901 0.9904 0.9906 0.9909 0.9911 0.9913 0.9916
2.4 0.9918 0.9920 0.9922 0.9925 0.9927 0.9929 0.9931 0.9932 0.9934 0.9936
2.5 0.9938 0.9940 0.9941 0.9943 0.9945 0.9946 0.9948 0.9949 0.9951 0.9952
2.6 0.9953 0.9955 0.9956 0.9957 0.9959 0.9960 0.9961 0.9962 0.9963 0.9964
2.7 0.9965 0.9966 0.9967 0.9968 0.9969 0.9970 0.9971 0.9972 0.9973 0.9974
2.8 0.9974 0.9975 0.9976 0.9977 0.9977 0.9978 0.9979 0.9979 0.9980 0.9981
2.9 0.9981 0.9982 0.9982 0.9983 0.9984 0.9984 0.9985 0.9985 0.9986 0.9986
3.0 0.9987 0.9987 0.9987 0.9988 0.9988 0.9989 0.9989 0.9989 0.9990 0.9990
3.1 0.9990 0.9991 0.9991 0.9991 0.9992 0.9992 0.9992 0.9992 0.9993 0.9993
3.2 0.9993 0.9993 0.9994 0.9994 0.9994 0.9994 0.9994 0.9995 0.9995 0.9995
3.3 0.9995 0.9995 0.9995 0.9996 0.9996 0.9996 0.9996 0.9996 0.9996 0.9997
3.4 0.9997 0.9997 0.9997 0.9997 0.9997 0.9997 0.9997 0.9997 0.9997 0.9998

Rz x2
−
Table B.1: Normal tables evaluating : 2π
1
−∞ e
2 dx

Version: – November 19, 2024

BIBLIOGRAPHY

[AS09] Siva, Athreya, Sunder, V.S. Measure and Probability CRC Press (Outside India) 2009 ISBN
14 3980 126 6.

[CasBer90] Casella, George; Berger, Roger [Link] inference. The Wadsworth & Brooks/Cole
Statistics/Probability Series. Wadsworth & Brooks/Cole Advanced Books & Software, Pacific
Grove, CA, 1990. xviii+650 pp. ISBN: 0-534-11958-1

[Fel68] W. Feller, An introduction to probability theory and its applications Vol. I. Third edition
John Wiley & Sons, Inc., New York-London-Sydney 1968.

[Fel71] W. Feller, An introduction to probability theory and its applications Vol. II. Second edition
John Wiley & Sons, Inc., New York-London-Sydney 1971.

[FPP98] [Link], [Link], [Link], Statistics Third edition [Link] and Company,
Inc., New York-London, 1998.

[FG97] B. Fristedt and L. Gray, A Modern Approach to Probability Thoery Birkhauser, Boston 1997.

[Ghah00] S. Ghahramani Fundamentals of probability. Second edition Prentice Hall, New Jersey,
2000.

[HMC05] [Link], J. McKean, A. Craig, Introduction to Mathematical Statistics, Sixth Edition,

Pearson Education, Inc., New Jersey, 2005.

[HPS72] P.G. Hoel, S.C. Port, and C.J. Stone, Introduction to Probability Theory, Houghton
Mifflin Company, 1972.

[Keane] M. Keane, The Essence of the Law of Large Numbers, , Pages 125–129, Algorithms, Fractals,
and Dynamics, Springer USA, 1995.

[Pit89] Pitman, Probability, Springer-verlag, New York, 1989.

[Ram97] Ramasubramanian, The Normal Distribution 1. From Binomial to Normal Resonance,

Vol 2, No.6, (June 1997), pp 15-14.

413

Version: – November 19, 2024

414 BIBLIOGRAPHY

[Rao73] [Link], Linear Statistical Inference and its Application , Second Edition, John Wiley,
New York, 1973.

[Ross84] S. Ross, A first course in probability, Second edition Macmillan Publishing Company,
New York, 1984.

[Ser09] Serfling RJ., Approximation theorems of mathematical statistics, John Wiley & Sons; 2009
Sep 25.

[Stig84] Stephen M. Stigler, Kruskal’s Proof of the Joint Distribution of X n and s2n , The American
Statistician, Vol 38, No. 2, (May 1984), pp 134-135.

[Wil27] Wilson E.B., Probable inference, the law of succession, and statistical inference, Journal of
the American Statistical Association. 1927 Jun 1;22(158):209-12.

Version: – November 19, 2024

INDEX

σ-field, 148 semicircular distribution, 167

arcsine distribution, 169
average, 93 Cauchy Distribution, 176, 205
distribution function, 156
Bayes’ theorem, 21, 22
exponential distribution, 159
Bernoulli distribution, 67, 100, 114
F Distribution, 279
Bernoulli trials, 37
Gamma distribution, 199
binomial distribution, 41, 68, 100, 114
Gaussian distribution, 161
binomial expansion, 38
normal distribution, 161
birthday problem, 14
Pareto Distribution, 179
bivariate normal, 239
Semi-circular law, 187
Borel sets, 148
uniform distribution, 158
Uniform distribution in the plane, 182
Central Limit Theorem
Weibull distribution, 167
De Moivre-Laplace, 162
distribution (of a random variable), 64
central limit theorem, 299
distribution function, 156
Chebychev’s inequality, 125, 218
Chi-square distribution, 279 empirical cummulative distribution, 248
conditional density, 190 empirical distribution, 248
conditional distribution, 72 equaltiy of distribution, 67
conditional expected value, 128 event (temporary definition), 2
conditional probability, 17 expected value, 211, 214
conditional variance, 128 discrete, 93
convergence in distribution, 293
convergence in probability, 290 fair game, 98
convolution, 84
correlation geometric distribution, 42, 68, 101, 115
continuous, 222
hypergeometric distribution, 56
discrete, 140
Expectation, 103
covariance
hypergoemetric distribution, 69
discrete, 136
cumulative distribution function, 156 i.i.d. random variables, 71
independence
density, 181 discrete random variables, 71
density function, 151, 181 independence (of events), 26, 28
distribution, 10
Beta Distribution, 206 joint density, 180, 181

415

Version: – November 19, 2024

416 INDEX

multivariate continuous, 272 positive correlation, 136

joint distribution probability, 2
bivariate-discrete, 74 Probability Generating Function, 238
multivariate-discrete, 74 probability mass function, 66
multivariate continuous, 272 probability space, 149
joint distribution function
bivariate, 181 quantile, 268
multivariate, 272
random variable, 63, 156
likelihood absolutely continuous, 157
profiled, 329 continuous, 156
discrete, 66
m.g.f. convergence theorem, 296
m.g.f. uniqueness theorem, 236 sample mean, 249
marginal density, 185 sample space, 1
marginal distribution sample variance, 250
Continuous, 184 sampling with replacement, 54
Discrete, 75 sampling without replacement, 55
Markov’s inequality, 124, 218 [Link] paradox, 109
Maximum Likelihood Estimator, 326 standard deviation, 112
median, 170 continuous, 215
memoryless property discrete, 110
Geometric, 76 standard units, 122
mode, 41 standardized random variable, 118
moment, 231 strong law of large numbers, 405
Moment Generating Function, 232
moment generating function t Distribution, 281
joint, 235
multinomial distribution, 77 unbiased estimator, 250
mutual independence uncorrelated, 136
discrete random variables), 71 uniform distribution, 10, 67

negative binomial distribution, 47 variance, 112

negative binomial distribution, 68 continuous, 215
negative correlation, 136 discrete, 110
variance stabilizing transformation, 313
Poisson distribution, 50, 68, 102
Polya urn, 20, 146 weak law of large numbers, 286

Version: – November 19, 2024