Notes for a graduate-level course in

asymptotics for statisticians

David R. Hunter
Penn State University

June 2014

Contents
Preface

1 Mathematical and Statistical Preliminaries

1.1

Limits and Continuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.1.1

Limit Superior and Limit Inferior . . . . . . . . . . . . . . . . . . . .

1.1.2

Continuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.2

Differentiability and Taylors Theorem . . . . . . . . . . . . . . . . . . . . .

13

1.3

Order Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

18

1.4

Multivariate Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

26

1.5

Expectation and Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . .

33

2 Weak Convergence
2.1

2.2

41

Modes of Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

41

2.1.1

Convergence in Probability . . . . . . . . . . . . . . . . . . . . . . . .

41

2.1.2

Probabilistic Order Notation . . . . . . . . . . . . . . . . . . . . . . .

43

2.1.3

Convergence in Distribution . . . . . . . . . . . . . . . . . . . . . . .

45

2.1.4

Convergence in Mean . . . . . . . . . . . . . . . . . . . . . . . . . . .

48

Consistent Estimates of the Mean . . . . . . . . . . . . . . . . . . . . . . . .

51

2.2.1

52

The Weak Law of Large Numbers . . . . . . . . . . . . . . . . . . . .

2.3

2.2.2

Independent but not Identically Distributed Variables . . . . . . . . .

52

2.2.3

Identically Distributed but not Independent Variables . . . . . . . . .

54

Convergence of Transformed Sequences . . . . . . . . . . . . . . . . . . . . .

58

2.3.1

Continuous Transformations: The Univariate Case . . . . . . . . . . .

58

2.3.2

Multivariate Extensions . . . . . . . . . . . . . . . . . . . . . . . . .

59

2.3.3

Slutskys Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . .

62

3 Strong convergence
3.1

70

Strong Consistency Defined . . . . . . . . . . . . . . . . . . . . . . . . . . .

70

3.1.1

Strong Consistency versus Consistency . . . . . . . . . . . . . . . . .

71

3.1.2

Multivariate Extensions . . . . . . . . . . . . . . . . . . . . . . . . .

73

3.2

The Strong Law of Large Numbers . . . . . . . . . . . . . . . . . . . . . . .

74

3.3

The Dominated Convergence Theorem . . . . . . . . . . . . . . . . . . . . .

79

3.3.1

Moments Do Not Always Converge . . . . . . . . . . . . . . . . . . .

79

3.3.2

Quantile Functions and the Skorohod Representation Theorem . . . .

81

4 Central Limit Theorems

4.1

4.2

88

Characteristic Functions and Normal Distributions . . . . . . . . . . . . . .

88

4.1.1

The Continuity Theorem . . . . . . . . . . . . . . . . . . . . . . . . .

89

4.1.2

Moments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

90

4.1.3

The Multivariate Normal Distribution . . . . . . . . . . . . . . . . .

91

4.1.4

Asymptotic Normality . . . . . . . . . . . . . . . . . . . . . . . . . .

92

4.1.5

The Cramer-Wold Theorem . . . . . . . . . . . . . . . . . . . . . . .

94

The Lindeberg-Feller Central Limit Theorem . . . . . . . . . . . . . . . . . .

96

4.2.1

The Lindeberg and Lyapunov Conditions . . . . . . . . . . . . . . . .

97

4.2.2

Independent and Identically Distributed Variables . . . . . . . . . . .

98

ii

4.2.3

Triangular Arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

99

4.3

Stationary m-Dependent Sequences . . . . . . . . . . . . . . . . . . . . . . . 108

4.4

Univariate extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111

4.4.1

The Berry-Esseen theorem . . . . . . . . . . . . . . . . . . . . . . . . 112

4.4.2

Edgeworth expansions . . . . . . . . . . . . . . . . . . . . . . . . . . 113

5 The Delta Method and Applications

5.1

116

Local linear approximations . . . . . . . . . . . . . . . . . . . . . . . . . . . 116

5.1.1

Asymptotic distributions of transformed sequences

. . . . . . . . . . 116

5.1.2

Variance stabilizing transformations . . . . . . . . . . . . . . . . . . . 119

5.2

Sample Moments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121

5.3

Sample Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123

6 Order Statistics and Quantiles

127

6.1

Extreme Order Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

6.2

Sample Quantiles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134

6.2.1

Uniform Order Statistics . . . . . . . . . . . . . . . . . . . . . . . . . 134

6.2.2

Uniform Sample Quantiles . . . . . . . . . . . . . . . . . . . . . . . . 135

6.2.3

General sample quantiles . . . . . . . . . . . . . . . . . . . . . . . . . 137

7 Maximum Likelihood Estimation

140

7.1

Consistency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140

7.2

Asymptotic normality of the MLE . . . . . . . . . . . . . . . . . . . . . . . . 144

7.3

Asymptotic Efficiency and Superefficiency . . . . . . . . . . . . . . . . . . . 149

7.4

The multiparameter case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154

7.5

Nuisance parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159

iii

8 Hypothesis Testing

161

8.1

Wald, Rao, and Likelihood Ratio Tests . . . . . . . . . . . . . . . . . . . . . 161

8.2

Contiguity and Local Alternatives . . . . . . . . . . . . . . . . . . . . . . . . 165

8.3

The Wilcoxon Rank-Sum Test . . . . . . . . . . . . . . . . . . . . . . . . . . 175

9 Pearsons chi-square test

180

9.1

Null hypothesis asymptotics . . . . . . . . . . . . . . . . . . . . . . . . . . . 180

9.2

Power of Pearsons chi-square test . . . . . . . . . . . . . . . . . . . . . . . . 187

10 U-statistics

190

10.1 Statistical Functionals and V-Statistics . . . . . . . . . . . . . . . . . . . . . 190

10.2 Asymptotic Normality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
10.3 Multivariate and multi-sample U-statistics . . . . . . . . . . . . . . . . . . . 201
10.4 Introduction to the Bootstrap . . . . . . . . . . . . . . . . . . . . . . . . . . 205

iv

Preface
These notes are designed to accompany STAT 553, a graduate-level course in large-sample
theory at Penn State intended for students who may not have had any exposure to measuretheoretic probability. While many excellent large-sample theory textbooks already exist,
the majority (though not all) of them reflect a traditional view in graduate-level statistics
education that students should learn measure-theoretic probability before large-sample theory. The philosophy of these notes is that these priorities are backwards, and that in fact
statisticians have more to gain from an understanding of large-sample theory than of measure
theory. The intended audience will have had a year-long sequence in mathematical statistics,
along with the usual calculus and linear algebra prerequisites that usually accompany such
a course, but no measure theory.
Many exercises require students to do some computing, based on the notion that computing skills should be emphasized in all statistics courses whenever possible, provided that the
computing enhances the understanding of the subject matter. The study of large-sample theory lends itself very well to computing, since frequently the theoretical large-sample results
we prove do not give any indication of how well asymptotic approximations work for finite
samples. Thus, simulation for the purpose of checking the quality of asymptotic approximations for small samples is very important in understanding the limitations of the results
being learned. Of course, all computing activities will force students to choose a particular
computing environment. Occasionally, hints are offered in the notes using R (http://www.rproject.org), though these exercises can be completed using other packages or languages,
provided that they possess the necessary statistical and graphical capabilities.
Credit where credit is due: These notes originally evolved as an accompaniment to the book
Elements of Large-Sample Theory by the late Erich Lehmann; the strong influence of that
great book, which shares the philosophy of these notes regarding the mathematical level
at which an introductory large-sample theory course should be taught, is still very much
evident here. I am fortunate to have had the chance to correspond with Professor Lehmann
several times about his book, as my students and I provided lists of typographical errors
that we had spotted. He was extremely gracious and I treasure the letters that he sent me,
written out longhand and sent through the mail even though we were already well into the
1

era of electronic communication.

I have also drawn on many other sources for ideas or for exercises. Among these are the
fantastic and concise A Course in Large Sample Theory by Thomas Ferguson, the comprehensive and beautifully written Asymptotic Statistics by A. W. van der Vaart, and the classic
probability textbooks Probability and Measure by Patrick Billingsley and An Introduction to
Probability Theory and Its Applications, Volumes 1 and 2 by William Feller. Arkady Tempelman at Penn State helped with some of the Strong-Law material in Chapter 3, and it was
Tom Hettmansperger who originally convinced me to design this course at Penn State back
in 2000 when I was a new assistant professor. My goal in doing so was to teach a course
that I wished I had had as a graduate student, and I hope that these notes help to achieve
that goal.

Chapter 1
Mathematical and Statistical
Preliminaries
We assume that many readers are familiar with much of the material presented in this
chapter. However, we do not view this material as superfluous, and we feature it prominently
as the first chapter of these notes for several reasons. First, some of these topics may have
been learned long ago by readers, and a review of this chapter may remind them of knowledge
they have forgotten. Second, including these preliminary topics as a separate chapter makes
the notes more self-contained than if the topics were omitted: We do not have to refer
readers to a standard calculus textbook or a standard mathematical statistics textbook
whenever an advanced result relies on this preliminary material. Third, some of the topics
here are likely to be new to some readers, particularly readers who have not taken a course
in real analysis.
Fourth, and perhaps most importantly, we wish to set the stage in this chapter for a mathematically rigorous treatment of large-sample theory. By mathematically rigorous, we
do not mean difficult or advanced; rather, we mean logically sound, relying on arguments in which assumptions and definitions are unambiguously stated and assertions must
be provable from these assumptions and definitions. Thus, even well-prepared readers who
know the material in this chapter often benefit from reading it and attempting the exercises,
particularly if they are new to rigorous mathematics and proof-writing. We strongly caution
against the alluring idea of saving time by skipping this chapter when teaching a course,
telling students you can always refer to Chapter 1 when you need to; we have learned the
hard way that this is a dangerous approach that can waste more time in the long run than
it saves!

1.1

Limits and Continuity

Fundamental to the study of large-sample theory is the idea of the limit of a sequence. Much
of these notes will be devoted to sequences of random variables; however, we begin here by
focusing on sequences of real numbers. Technically, a sequence of real numbers is a function
from the natural numbers {1, 2, 3, . . .} into the real numbers R; yet we always write a1 , a2 , . . .
instead of the more traditional function notation a(1), a(2), . . ..
We begin by defining a limit of a sequence of real numbers. This is a concept that will be
intuitively clear to readers familiar with calculus. For example, the fact that the sequence
a1 = 1.3, a2 = 1.33, a3 = 1.333, . . . has a limit equal to 4/3 is unsurprising. Yet there are
some subtleties that arise with limits, and for this reason and also to set the stage for a
rigorous treatment of the topic, we provide two separate definitions. It is important to
remember that even these two definitions do not cover all possible sequences; that is, not
every sequence has a well-defined limit.
Definition 1.1 A sequence of real numbers a1 , a2 , . . . has limit equal to the real number a if for every  > 0, there exists N such that
|an a| <  for all n > N .
In this case, we write an a as n or limn an = a and we could say that
an converges to a.
Definition 1.2 A sequence of real numbers a1 , a2 , . . . has limit if for every real
number M , there exists N such that
an > M for all n > N .
In this case, we write an as n or limn an = and we could say
that an diverges to . Similarly, an as n if for all M , there exists
N such that an < M for all n > N .
Implicit in the language of Definition 1.1 is that N may depend on . Similarly, N may
depend on M (in fact, it must depend on M ) in Definition 1.2.
The symbols + and are not considered real numbers; otherwise, Definition 1.1 would
be invalid for a = and Definition 1.2 would never be valid since M could be taken to
be . Throughout these notes, we will assume that symbols such as an and a denote real
numbers unless stated otherwise; if situations such as a = are allowed, we will state this
fact explicitly.
A crucial fact regarding sequences and limits is that not every sequence has a limit, even
when has a limit includes the possibilities . (However, see Exercise 1.4, which asserts
4

that every nondecreasing sequence has a limit.) A simple example of a sequence without a
limit is given in Example 1.3. A common mistake made by students is to take the limit of
both sides of an equation an = bn or an inequality an bn . This is a meaningless operation
unless it has been established that such limits exist. On the other hand, an operation that is
valid is to take the limit superior or limit inferior of both sides, concepts that will be defined
in Section 1.1.1. One final word of warning, though: When taking the limit superior of a
strict inequality, < or > must be replaced by or ; see the discussion following Lemma
1.10.
Example 1.3 Define
an = log n;

bn = 1 + (1)n /n;

cn = 1 + (1)n /n2 ;

dn = (1)n .

Then an , bn 1, and cn 1; but the sequence d1 , d2 , . . . does not have a

limit. (We do not always write as n when this is clear from the context.)
Let us prove one of these limit statements, say, bn 1. By Definition 1.1, given
an arbitrary  > 0, we must prove that there exists some N such that |bn 1| < 
whenever n > N . Since |bn 1| = 1/n, we may simply take N = 1/: With this
choice, whenever n > N , we have |bn 1| = 1/n < 1/N = , which completes the
proof.
We always assume that log n denotes the natural logarithm, or logarithm base e,
of n. This is fairly standard in statistics, though in some other disciplines it is
more common to use log n to denote the logarithm base 10, writing ln n instead
of the natural logarithm. Since the natural logarithm and the logarithm base 10
differ only by a constant rationamely, loge n = 2.3026 log10 nthe difference is
often not particularly important. (However, see Exercise 1.27.)
Finally, note that although limn bn = limn cn in Example 1.3, there is evidently
something different about the manner in which these two sequences approach this
limit. This difference will prove important when we study rates of convergence
beginning in Section 1.3.
Example 1.4 A very important example of a limit of a sequence is

c n
= exp(c)
lim 1 +
n
n
for any real number c. This result is proved in Example 1.20 using lHopitals
rule (Theorem 1.19).
Two or more sequences may be added, multiplied, or divided, and the results follow intuitively pleasing rules: The sum (or product) of limits equals the limit of the sums (or
products); and as long as division by zero does not occur, the ratio of limits equals the limit
5

of the ratios. These rules are stated formally as Theorem 1.5, whose complete proof is the
subject of Exercise 1.1. To prove only the limit of sums equals sum of limits part of the
theorem, if we are given an a and bn b then we need to show that for a given  > 0,
there exists N such that for all n > N , |an + bn (a + b)| < . But the triangle inequality
gives
|an + bn (a + b)| |an a| + |bn b|,

(1.1)

and furthermore we know that there must be N1 and N2 such that |an a| < /2 for n > N1
and |bn b| < /2 for n > N2 (since /2 is, after all, a positive constant and we know an a
and bn b). Therefore, we may take N = max{N1 , N2 } and conclude by inequality (1.1)
that for all n > N ,
|an + bn (a + b)| <



+ ,
2 2

which proves that an + bn a + b.

Theorem 1.5 Suppose an a and bn b as n . Then an + bn a + b and
an bn ab; furthermore, if b 6= 0 then an /bn a/b.
A similar result states that continuous transformations preserve limits; see Theorem 1.16.
Theorem 1.5 may be extended by replacing a and/or b by , and the results remain true
as long as they do not involve the indeterminate forms , 0, or /.

1.1.1

Limit Superior and Limit Inferior

The limit superior and limit inferior of a sequence, unlike the limit itself, are defined for any
sequence of real numbers. Before considering these important quantities, we must first define
supremum and infimum, which are generalizations of the ideas of maximum and minumum.
That is, for a set of real numbers that has a minimum, or smallest element, the infimum
is equal to this minimum; and similarly for the maximum and supremum. For instance,
any finite set contains both a minimum and a maximum. (Finite is not the same as
bounded; the former means having finitely many elements and the latter means contained
in an interval neither of whose endpoints are .) However, not all sets of real numbers
contain a minimum (or maximum) value. As a simple example, take the open interval (0, 1).
Since neither 0 nor 1 is contained in this interval, there is no single element of this interval
that is smaller (or larger) than all other elements. Yet clearly 0 and 1 are in some sense
important in bounding this interval below and above. It turns out that 0 and 1 are the
infimum and supremum, respectively, of (0, 1).
An upper bound of a set S of real numbers is (as the name suggests) any value m such that
s m for all s S. A least upper bound is an upper bound with the property that no smaller
6

upper bound exists; that is, m is a least upper bound if m is an upper bound such that for
any  > 0, there exists s S such that s > m . A similar definition applies to greatest
lower bound. A useful fact about the real numbersa consequence of the completeness of
the real numbers which we do not prove hereis that every set that has an upper (or lower)
bound has a least upper (or greatest lower) bound.
Definition 1.6 For any set of real numbers, say S, the supremum sup S is defined to
be the least upper bound of S (or + if no upper bound exists). The infimum
inf S is defined to be the greatest lower bound of S (or if no lower bound
exists).
Example 1.7 Let S = {a1 , a2 , a3 , . . .}, where an = 1/n. Then inf S, which may also
be denoted inf n an , equals 0 even though 0 6 S. But supn an = 1, which is
contained in S. In this example, max S = 1 but min S is undefined.
If we denote by supkn ak the supremum of {an , an+1 , . . .}, then we see that this supremum is
taken over a smaller and smaller set as n increases. Therefore, supkn ak is a nonincreasing
sequence in n, which implies that it has a limit as n (see Exercise 1.4). Similarly,
inf kn ak is a nondecreasing sequence, which implies that it has a limit.
Definition 1.8 The limit superior of a sequence a1 , a2 , . . ., denoted lim supn an or
sometimes limn an , is the limit of the nonincreasing sequence
sup ak ,

sup ak ,

k1

k2

...

The limit inferior, denoted lim inf n an or sometimes limn an , is the limit of the
nondecreasing sequence
inf ak , inf ak , . . . .

k1

k2

Intuitively, the limit superior and limit inferior may be understood as follows: If we define
a limit point of a sequence to be any number which is the limit of some subsequence, then
lim inf and lim sup are the smallest and largest limit points, respectively (more precisely,
they are the infimum and supremum, respectively, of the set of limit points).
Example 1.9 In Example 1.3, the sequence dn = (1)n does not have a limit. However, since supkn dk = 1 and inf kn dk = 1 for all n, it follows that
lim inf dn = 1.

lim sup dn = 1 and

In this example, the set of limit points of the sequence d1 , d2 , . . . is simply {1, 1}.
Here are some useful facts regarding limits superior and inferior:
7

Lemma 1.10 Let a1 , a2 , . . . and b1 , b2 , . . . be arbitrary sequences of real numbers.

lim supn an and lim inf n an always exist, unlike limn an .
lim inf n an lim supn an
limn an exists if and only if lim inf n an = lim supn an , in which case
lim an = lim inf an = lim sup an .
n

Both lim sup and lim inf preserve nonstrict inequalities; that is, if an bn
for all n, then lim supn an lim supn bn and lim inf n an lim inf n bn .
lim supn (an ) = lim inf n an .
The next-to-last claim in Lemma 1.10 is no longer true if nonstrict inequalities is replaced
by strict inequalities. For instance, 1/(n + 1) < 1/n is true for all positive n, but the limit
superior of each side equals zero. Thus, it is not true that
lim sup
n

1
1
< lim sup .
n+1
n n

We must replace < by (or > by ) when taking the limit superior or limit inferior of both
sides of an inequality.

1.1.2

Continuity

Although Definitions 1.1 and 1.2 concern limits, they apply only to sequences of real numbers.
Recall that a sequence is a real-valued function of the natural numbers. We shall also require
the concept of a limit of a real-valued function of a real variable. To this end, we make the
following definition.
Definition 1.11 For a real-valued function f (x) defined for all points in a neighborhood of x0 except possibly x0 itself, we call the real number a the limit of f (x)
as x goes to x0 , written
lim f (x) = a,

xx0

if for each  > 0 there is a > 0 such that |f (x)a| <  whenever 0 < |xx0 | < .
First, note that Definition 1.11 is sensible only if both x0 and a are finite (but see Definition
1.13 for the case in which one or both of them is ). Furthermore, it is very important
to remember that 0 < |x x0 | < may not be replaced by |x x0 | < : The latter would
8

imply something specific about the value of f (x0 ) itself, whereas the correct definition does
not even require that this value be defined. In fact, by merely replacing 0 < |x x0 | <
by |x x0 | < (and insisting that f (x0 ) be defined), we could take Definition 1.11 to be
the definition of continuity of f (x) at the point x0 (see Definition 1.14 for an equivalent
formuation).
Implicit in Definition 1.11 is the fact that a is the limiting value of f (x) no matter whether
x approaches x0 from above or below; thus, f (x) has a two-sided limit at x0 . We may also
consider one-sided limits:
Definition 1.12 The value a is called the right-handed limit of f (x) as x goes to x0 ,
written
lim f (x) = a or f (x0 +) = a,

xx0 +

if for each  > 0 there is a > 0 such that |f (x) a| <  whenever 0 < x x0 < .
The left-handed limit, limxx0 f (x) or f (x0 ), is defined analagously: f (x0 ) =
a if for each  > 0 there is a > 0 such that |f (x)a| <  whenever < xx0 <
0.
The preceding definitions imply that
lim f (x) = a if and only if f (x0 +) = f (x0 ) = a;

xx0

(1.2)

in other words, the (two-sided) limit exists if and only if both one-sided limits exist and they
coincide. Before using the concept of a limit to define continuity, we conclude the discussion
of limits by addressing the possibilities that f (x) has a limit as x or that f (x) tends
to :
Definition 1.13 Definition 1.11 may be expanded to allow x0 or a to be infinite:
(a) We write limx f (x) = a if for every  > 0, there exists N such that
|f (x) a| <  for all x > N .
(b) We write limxx0 f (x) = if for every M , there exists > 0 such that
f (x) > M whenever 0 < |x x0 | < .
(c) We write limx f (x) = if for every M , there exists N such that f (x) > M
for all x > N .
Definitions involving are analogous, as are definitions of f (x0 +) = and
f (x0 ) = .

As mentioned above, the value of f (x0 ) in Definitions 1.11 and 1.12 is completely irrelevant;
in fact, f (x0 ) might not even be defined. In the special case that f (x0 ) is defined and equal
to a, then we say that f (x) is continuous (or right- or left-continuous) at x0 , as summarized
by Definition 1.14 below. Intuitively, f (x) is continuous at x0 if it is possible to draw the
graph of f (x) through the point [x0 , f (x0 )] without lifting the pencil from the page.
Definition 1.14 If f (x) is a real-valued function and x0 is a real number, then
we say f (x) is continuous at x0 if limxx0 f (x) = f (x0 );
we say f (x) is right-continuous at x0 if limxx0 + f (x) = f (x0 );
we say f (x) is left-continuous at x0 if limxx0 f (x) = f (x0 ).
Finally, even though continuity is inherently a local property of a function (since Definition 1.14 applies only to the particular point x0 ), we often speak globally of a continuous
function, by which we mean a function that is continuous at every point in its domain.

1.0

Statement (1.2) implies that every (globally) continuous function is right-continuous. However, the converse is not true, and in statistics the canonical example of a function that is
right-continuous but not continuous is the cumulative distribution function for a discrete
random variable.

F(t)

0.6

0.8

0.0

0.2

0.4

0.5

0.0

0.5

1.0

1.5

Figure 1.1: The cumulative distribution function for a Bernoulli (1/2) random variable is
discontinuous at the points t = 0 and t = 1, but it is everywhere right-continuous.

10

Example 1.15 Let X be a Bernoulli (1/2) random variable, so that the events X = 0
and X = 1 each occur with probability 1/2. Then the distribution function
F (t) = P (X t) is right-continuous but it is not continuous because it has
jumps at t = 0 and t = 1 (see Figure 1.1). Using one-sided limit notation of
Definition 1.12, we may write
0 = F (0) 6= F (0+) = 1/2 and 1/2 = F (1) 6= F (1+) = 1.
Although F (t) is not (globally) continuous, it is continuous at every point in the
set R \ {0, 1} that does not include the points 0 and 1.
We conclude with a simple yet important result relating continuity to the notion of the
limit of a sequence. Intuitively, this result states that continuous functions preserve limits
of sequences.
Theorem 1.16 If a is a real number such that an a as n and the real-valued
function f (x) is continuous at the point a, then f (an ) f (a).
Proof: We need to show that for any  > 0, there exists N such that |f (an ) f (a)| < 
for all n > N . To this end, let  > 0 be a fixed arbitrary constant. From the definition of
continuity, we know that there exists some > 0 such that |f (x) f (a)| <  for all x such
that |x a| < . Since we are told an a and since > 0, there must by definition be
some N such that |an a| < for all n > N . We conclude that for all n greater than this
particular N , |f (an ) f (a)| < . Since  was arbitrary, the proof is finished.

Exercises for Section 1.1

Exercise 1.1 Assume that an a and bn b, where a and b are real numbers.
(a) Prove that an bn ab
Hint: Show that |an bn ab| |(an a)(bn b)| + |a(bn b)| + |b(an a)| using
the triangle inequality.
(b) Prove that if b 6= 0, an /bn a/b.
Exercise 1.2 For a fixed real number c, define an (c) = (1 + c/n)n . Then Equation
(1.9) states that an (c) exp(c). A different sequence with the same limit is
obtained from the power series expansion of exp(c):
bn (c) =

n1 i
X
c
i=0

11

i!

For each of the values c {10, 1, 0.2, 1, 5}, find the smallest value of n such
that |an (c) exp(c)|/ exp(c) < .01. Now replace an (c) by bn (c) and repeat.
Comment on any general differences you observe between the two sequences.
Exercise 1.3 (a) Suppose that ak c as k for a sequence of real numbers
a1 , a2 , . . .. Prove that this implies convergence in the sense of Cesaro, which
means that
n
1X
ak c
as n .
(1.3)
n k=1
In this case, c may be real or it may be .
Hint: If c is real, consider the definition of ak c: There exists N such that
|ak c| <  for all k > N . Consider what happens when the sum in expression
(1.3) is broken into two sums, one for k N and one for k > N . The case
c = follows a similar line of reasoning.
(b) Is the converse true? In other words, does (1.3) imply ak c?
Exercise 1.4 Prove that if a1 , a2 , . . . is a nondecreasing (or nonincreasing) sequence,
then limn an exists and is equal to supn an (or inf n an ). We allow the possibility
supn an = (or inf n an = ) here.
Hint: For the case in which supn an is finite, use the fact that the least upper
bound M of a set S is defined by the fact that s M for all s S, but for any
 > 0 there exists s S such that s > M .
Exercise 1.5 Let an = sin n for n = 1, 2, . . ..
(a) What is supn an ? Does maxn an exist?
(b) What is the set of limit points of {a1 , a2 , . . .}? What are lim supn an and
lim inf n an ? (Recall that a limit point is any point that is the limit of a subsequence ak1 , ak2 , . . ., where k1 < k2 < .)
(c) As usual in mathematics, we assume above that angles are measured in
radians. How do the answers to (a) and (b) change if we use degrees instead (i.e.,
an = sin n )?
Exercise 1.6 Prove Lemma 1.10.
Exercise 1.7 For x 6 {0, 1, 2}, define
f (x) =

|x3 x|
.
x(x 1)(x 2)
12

(a) Graph f (x). Experiment with various ranges on the axes until you attain a
visually pleasing and informative plot that gives a sense of the overall behavior
of the function.
(b) For each of x0 {1, 0, 1, 2}, answer these questions: Is f (x) continuous at
x0 , and if not, could f (x0 ) be defined so as to make the answer yes? What are
the right- and left-hand limits of f (x) at x0 ? Does it have a limit at x0 ? Finally,
what are limx f (x) and limx f (x)?
Exercise 1.8 Define F (t) as in Example 1.15 (and as pictured in Figure 1.1). This
function is not continuous, so Theorem 1.16 does not apply. That is, an a does
not imply that F (an ) F (a).
(a) Give an example of a sequence {an } and a real number a such that an a
but lim supn F (an ) 6= F (a).
(b) Change your answer to part (a) so that an a and lim supn F (an ) = F (a),
but limn F (an ) does not exist.
(c) Explain why it is not possible to change your answer so that an a and
lim inf n F (an ) = F (a), but limn F (an ) does not exist.

1.2

Differentiability and Taylors Theorem

Differential calculus plays a fundamental role in much asymptotic theory. In this section
we review simple derivatives and one form of Taylors well-known theorem. Approximations
to functions based on Taylors Theorem, often called Taylor expansions, are ubiquitous in
large-sample theory.
We assume that readers are familiar with the definition of a derivative of a real-valued
function f (x):
Definition 1.17 If f (x) is continuous in a neighborhood of x0 and
lim

xx0

f (x) f (x0 )
x x0

(1.4)

exists, then f (x) is said to be differentiable at x0 and the limit (1.4) is called the
derivative of f (x) at x0 and is denoted by f 0 (x0 ) or f (1) (x0 ).
We use the standard notation for second- and higher-order derivatives. Thus, if f 0 (x) is
itself differentiable at x0 , we express its derivative as f 00 (x0 ) or f (2) (x0 ). In general, if the
kth derivative f (k) (x) is differentiable at x0 , then we denote this derivative by f (k+1) (x0 ). We
13

also write (dk /dxk )f (x) (omitting the k when k = 1) to denote the function f (k) (x), and to
denote the evaluation of this function at a specific point (say x0 ), we may use the following
notation, which is equivalent to f (k) (x0 ):


dk
f (x)
k
dx
x=x0
In large-sample theory, differential calculus is most commonly applied in the construction of
Taylor expansions. There are several different versions of Taylors Theorem, distinguished
from one another by the way in which the remainder term is expressed. The first form we
present here (Theorem 1.18), which is proved in Exercise 1.11, does not state an explicit form
for the remainder term. This gives it the advantage that it does not require that the function
have an extra derivative. For instance, a second-order Taylor expansion requires only two
derivatives using this version of Taylors Theorem (and the second derivative need only exist
at a single point), whereas other forms of Taylors Theorem require the existence of a third
derivative over an entire interval. The disadvantage of this form of Taylors Theorem is that
we do not get any sense of what the remainder term is, only that it goes to zero; however,
for many applications in these notes, this form of Taylors Theorem will suffice.
Theorem 1.18 If f (x) has d derivatives at a, then
f (x) = f (a) + (x a)f 0 (a) + +

(x a)d (d)
f (a) + rd (x, a),
d!

(1.5)

where rd (x, a)/(x a)d 0 as x a.

In some cases, we will find it helpful to have an explicit form of rd (x, a). This is possible
under stronger assumptions, namely, that f (x) has d + 1 derivatives on the closed interval
from x to a. In this case, we may write
Z x
(x t)d (d+1)
f
(t) dt
(1.6)
rd (x, a) =
d!
a
in equation (1.5). Equation (1.6) is often called the Lagrange form of the remainder. By the
Mean Value Theorem of calculus, there exists x somewhere in the closed interval from x to
a such that
rd (x, a) =

(x a)d+1 (d+1)
f
(x ).
(d + 1)!

(1.7)

Expression (1.7), since it follows immediately from Equation (1.6), is also referred to as the
Lagrange form of the remainder.
To conclude this section, we state the well-known calculus result known as lHopitals Rule.
This useful Theorem provides an elegant way to prove Theorem 1.18, among other things.
14

Theorem 1.19 lHopitals Rule: For a real number c, suppose that f (x) and g(x)
are differentiable for all points in a neighborhood containing c except possibly c
itself. If limxc f (x) = 0 and limxc g(x) = 0, then
f (x)
f 0 (x)
= lim 0
,
xc g(x)
xc g (x)
lim

(1.8)

provided the right-hand limit exists. Similarly, if limxc f (x) = and limxc g(x) =
, then Equation (1.8) also holds. Finally, the theorem also applies if c = , in
which case a neighborhood containing c refers to an interval (a, ) or (, a).
Example 1.20 Example 1.4 states that

c n
lim 1 +
= exp(c)
n
n

(1.9)

for any real number c. Let us prove this fact using lHopitals Rule. Care is
necessary in this proof, since lHopitals Rule applies to limits of differentiable
functions, whereas the left side of Equation (1.9) is a function of an integer-valued
n.
Taking logarithms in Equation (1.9), we shall first establish that n log(1+c/n)
c as n . Define f (x) = log(1 + cx) and g(x) = x. The strategy is to treat n
as 1/x, so we will see what happens to f (x)/g(x) as x 0. By lHopitals Rule,
we obtain
lim

x0

c/(1 + cx)
log(1 + cx)
= lim
= c.
x0
x
1

Since this limit must be valid no matter how x approaches 0, in particular we

may conclude that if we define xn = 1/n for n = 1, 2, . . ., then

log(1 + cxn )
c
lim
= lim n log 1 +
= c,
n
n
xn
n

(1.10)

which was to be proved. Now we use the fact that the exponential function
h(t) = exp t is a continuous function, so Equation (1.9) follows from Theorem
1.16 once we apply the exponential function to Equation (1.10).

Exercises for Section 1.2

Exercise 1.9 The well-known derivative of the polynomial function f (x) = xn for a
positive integer n is given by nxn1 . Prove this fact directly using Definition 1.17.
15

Exercise 1.10 For f (x) continuous in a neighborhood of x0 , consider

lim

xx0

f (x) f (2x0 x)
.
2(x x0 )

(1.11)

(a) Prove or give a counterexample: When f 0 (x0 ) exists, limit (1.11) also exists
and it is equal to f 0 (x0 ).
(b) Prove or give a counterexample: When limit (1.11) exists, it equals f 0 (x0 ),
which also exists.
Exercise 1.11 Prove Theorem 1.18.
Hint: Let Pd (x) denote the Taylor polynomial such that
rd (x, a) = f (x) Pd (x).
Then use lHopitals rule, Theorem 1.19, d 1 times. (You can do this because
the existence of f (d) (a) implies that all lower-order derivatives exist on an interval
containing a.) You cannot use lHopitals rule d times, but you wont need to if
you use Definition 1.17.
Exercise 1.12 Let f (t) = log t. Taking a = 1 and x = a + h, find the explicit
remainder term rd (x, a) in Equation (1.5) for all values of d {2, 3} and h
{0.1, 0.01, 0.001}. Give your results in a table. How does rd (x, a) appear to vary
with d? How does rd (a + h, a) appear to vary with h?
Exercise 1.13 The idea for Exercise 1.10 is based on a numerical trick for accurately
approximating the derivative of a function that can be evaluated directly but for
which no formula for the derivative is known.
(a) First, construct a first-order approximation to a derivative. Definition
1.17 with d = 1 suggests that we may choose a small h and obtain
f 0 (a)

f (a + h) f (a)
.
h

(1.12)

For f (x) = log x and a = 2, calculate the approximation to f 0 (a) in Equation

(1.12) using h {0.5, 0.05, 0.005}. How does the difference between the true
value (which you happen to know in this case) and the approximation appear to
vary as a function of h?
(b) Next, expand both f (a + h) and f (a h) using Taylors theorem with
d = 2. Subtract one expansion from the other and solve for f 0 (a). Ignore the
16

remainder terms and you have a second-order approximation. (Compare this

approximation with Exercise 1.10, substituting x0 and xx0 for a and h.) Repeat
the computations of part (a). Now how does the error appear to vary as a function
of h?
(c) Finally, construct a fourth-order approximation. Perform Taylor expansions of f (x + 2h), f (x + h), f (x h), and f (x 2h) with d = 4. Ignore the
remainder terms, then find constants C1 and C2 such that the second, third, and
fourth derivatives all disappear and you obtain
f 0 (a)

C1 [f (a + h) f (a h)] + C2 [f (a + 2h) f (a 2h)]

.
h

(1.13)

Repeat the computations of parts (a) and (b) using the approximation in Equation
(1.13).
Exercise 1.14 The gamma function (x) is defined for positive real x as
Z
(x) =
tx1 et dt

(1.14)

[in fact, equation (1.14) is also valid for complex x with positive real part]. The
gamma function may be viewed as a continuous version of the factorial function
in the sense that (n) = (n 1)! for all positive integers n. The gamma function
satisfies the identity
(x + 1) = x(x)

(1.15)

even for noninteger positive values of x. Since (x) grows very quickly as x increases, it is often convenient in numerical calculations to deal with the logarithm
of the gamma function, which we term the log-gamma function. The digamma
function (x) is defined to be the derivative of the log-gamma function; this function often arises in statistical calculations involving certain distributions that use
the gamma function.
(a) Apply the result of Exercise 1.13(b) using h = 1 to demonstrate how to
obtain the approximation
(x)

1
log [x(x 1)]
2

for x > 2.
Hint: Use Identity (1.15).

17

(1.16)

(b) Test Approximation (1.16) numerically for all x in the interval (2, 100) by
plotting the ratio of the approximation to the true (x). What do you notice about the quality of the approximation? If you are using R or Splus, then
digamma(x) gives the value of (x).
Exercise 1.15 The second derivative of the log-gamma function is called the trigamma
function:
0 (x) =

d2
log (x).
dx2

(1.17)

Like the digamma function, it often arises in statistical calculations; for example,
see Exercise 1.35.
(a) Using the method of Exercise 1.13(c) with h = 1 [that is, expanding f (x+2h),
f (x + h), f (x h), and f (x 2h) and then finding a linear combination that
makes all but the second derivative of the log-gamma function disappear], show
how to derive the following approximation to 0 (x) for x > 2:
"
15 
#
x
x

2
1
log
.
(1.18)
0 (x)
12
x1
x+1
(b) Test Approximation (1.18) numerically as in Exercise 1.14(b). In R or Splus,
trigamma(x) gives the value of 0 (x).

1.3

Order Notation

As we saw in Example 1.3, the limiting behavior of a sequence is not fully characterized
by the value of its limit alone, if the limit exists. In that example, both 1 + (1)n /n and
1 + (1)n /n2 converge to the same limit, but they approach this limit at different rates. In
this section we consider not only the value of the limit, but the rate at which that limit is
approached. In so doing, we present some convenient notation for comparing the limiting
behavior of different sequences.
Definition 1.21 We say that the sequence of real numbers a1 , a2 , . . . is asymptotically
equivalent to the sequence b1 , b2 , . . ., written an bn , if (an /bn ) 1 as n .
Equivalently, an bn if and only if


 an b n 


 an  0.
18

The expression |(an bn )/an | above is called the relative error in approximating an by bn .
The definition of asymptotic equivalence does not say that
lim an
= 1;
lim bn
the above fraction might equal 0/0 or /, or the limits might not even exist! (See Exercise
1.17.)
Example 1.22 A well-known asymptotic equivalence is Stirlings formula, which states

(1.19)
n! 2nn+.5 exp(n).
There are multiple ways to prove Stirlings formula. We outline one proof, based
on the Poisson distribution, in Exercise 4.5.
Example 1.23 For any k > 1,
n
X

ik

i=1

nk+1
.
k+1

(1.20)

This is proved in Exercise 1.19. But what about the case k = 1? Let us prove
that
n
X
1
i=1

log n.

(1.21)

Proof: Since 1/x is a strictly decreasing function of x, we conclude that

Z i+1
Z i
1
1
1
dx < <
dx
x
i
i
i1x
for i = 2, 3, 4, . . .. Summing on i (and using 1/i = 1 for i = 1) gives
Z
1+
2

n+1

X1
1
dx <
<1+
x
i
i=1

Z
1

1
dx.
x

Evaluating the integrals and dividing through by log n gives

Pn 1
1 + log(n + 1) log 2
1
< i=1 i <
+ 1.
log n
log n
log n
The left and right sides of this expression have limits, both equal to 1 (do you see
why?). A standard trick is therefore to take the limit inferior of the left inequality
19

and combine this with the limit superior of the right inequality (remember to
change < to when doing this; see the discussion following Lemma 1.10) to
obtain
Pn 1
Pn 1
i=1 i
1 lim inf
lim sup i=1 i 1.
n
log n
log n
n
This implies that the limit inferior and limit superior are in fact the same, so the
limit exists and is equal to 1. This is what we wished to show.
The next notation we introduce expresses the idea that one sequence is asymptotically negligible compared to another sequence.
Definition 1.24 We write an = o(bn ) (an is little-o of bn ) as n if an /bn 0
as n .
Among other advantages, the o-notation makes it possible to focus on the most important
terms of a sequence while ignoring the terms that are comparatively negligible.
Example 1.25 According to Definition 1.24, we may write
 
2
4
1
1
1
2 + 3 = +o
as n .
n n
n
n
n
This makes it clear at a glance how fast the sequence on the left tends to zero,
since all terms other than the dominant term are lumped together as o(1/n).
Some of the exercises in this section require proving that one sequence is little-o of another
sequence. Sometimes, lHopitals rule may be helpful; yet as in Example 1.20, care must be
exercised because lHopitals rule applies to functions of real numbers whereas a sequence is
a function of the positive integers.
Example 1.26 Let us prove that log log n = o(log n). The function (log log x)/ log x,
defined for x > 1, agrees with (log log n)/ log n on the positive integers; thus,
since lHopitals rule implies
log log x
= lim
lim
x
x log x

1
x log x
1
x

1
= 0,
x log x

= lim

we conclude that (log log n)/ log n must also tend to 0 as n tends to as an
integer.
Often, however, one may simply prove an = o(bn ) without resorting to lHopitals rule, as in
the next example.

20

Example 1.27 Prove that

!
n
X

n=o
i .

(1.22)

i=1

Proof: Letting bn/2c denote the largest integer less than or equal to n/2,
r
n
n
X
X

n jnk
i
i
.
2
2
i=1
i=bn/2c

Since n = o(n n), the desired result follows.

Equation (1.22) could have been proved using the result of Example 1.23, in which
Equation (1.20) with k = 1/2 implies that
n
X

2n3/2
.
i
3
i=1

(1.23)

However, we urge extreme caution when using asymptotic equivalences like Expression (1.23). It is tempting to believe that expressions that are asymptotically
equivalent may be substituted for one another under any circumstances, and this
is not true! In this particular example, we may write
!


n
3
2n3/2

Pn =
Pn ,
2
n
i
3
i
i=1
i=1
and because we know that the second fraction in parentheses tends to 1 by Expression (1.23) and the first fraction in parentheses tends to 0, we conclude that
the product of the two converges to 0 and Equation (1.22) is proved.
We define one additional order notation, the capital O.
Definition 1.28 We write an = O(bn ) (an is big-o of bn ) as n if there exist
M > 0 and N > 0 such that |an /bn | < M for all n > N .
In particular, an = o(bn ) implies an = O(bn ). In a vague sense, o and O relate to sequences
as < and relate to real numbers. However, this analogy is not perfect: For example, note
that it is not always true that either an = O(bn ) or bn = O(an ).
Although the notation above is very precisely defined, unfortunately this is not the case with
the language used to describe the notation. In particular, an is of order bn is ambiguous;
it may mean simply that an = O(bn ), or it may mean something more precise: Some authors
define an  bn or an = (bn ) to mean that |an | remains bounded between m|bn | and M |bn |
21

for large enough n for some constants 0 < m < M . Although the language can be imprecise,
it is usually clear from context what the speakers intent is.
This latter case, where an = O(bn ) but an 6= o(bn ), is one in which the ratio |an /bn | remains
bounded and also bounded away from zero: There exist positive constants m and M , and
an integer N , such that
 
 an 
(1.24)
m <   < M for all n > N .
bn
Some books introduce a special symbol for (1.24), such as an  bn or an = (bn ).
Do not forget that the use of o, O, or always implies that there is some sort of limit being
taken. Often, an expression involves n, in which case we usually assume n tends to even
if this is not stated; however, sometimes things are not so clear, so it helps to be explicit:
Example 1.29 According to Definition 1.24, a sequence that is o(1) tends to zero.
Therefore, Equation (1.5) of Taylors Theorem may be rewritten
f (x) = f (a) + (x a)f 0 (a) + +

(x a)d  (d)
f (a) + o(1)
d!

as x a.

It is important to write as x a in this case.

It is often tempting, when faced with an equation such as an = o(bn ), to attempt to apply a
function f (x) to each side and claim that f (an ) = o[f (bn )]. Unfortunately, however, this is
not true in general and it is not hard to find a counterexample [see Exercise 1.18(d)]. There
are certain circumstances in which it is possible to claim that f (an ) = o[f (bn )], and one such
circumstance is particularly helpful. It involves a convex function f (x), defined as follows:
Definition 1.30 We say that a function f (x) is convex if for all x, y and any [0, 1],
we have
f [x + (1 )y] f (x) + (1 )f (y).

(1.25)

If f (x) is everywhere differentiable and f 00 (x) > 0 for all x, then f (x) is convex (this is
proven in Exercise 1.24). For instance, the function f (x) = exp(x) is convex because its
second derivative is always positive.
We now see a general case in which it may be shown that f (an ) = o[f (bn )].
Theorem 1.31 Suppose that a1 , a2 , . . . and b1 , b2 , . . . are sequences of real numbers
such that an , bn , and an = o(bn ); and f (x) is a convex function such
that f (x) as x . Then f (an ) = o[f (bn )].
22

The proof of Theorem 1.31 is the subject of Exercise 1.25.

There are certain rates of growth toward that are so common that they have names, such
as logarithmic, polynomial, and exponential growth. If , , and are arbitrary positive
constants, then the sequences (log n) , n , and (1 + )n exhibit logarithmic, polynomial, and
exponential growth, respectively. Furthermore, we always have
(log n) = o(n ) and n = o([1 + ]n ).

(1.26)

Thus, in the sense of Definition 1.24, logarithmic growth is always slower than polynomial
growth and polynomial growth is always slower than exponential growth.
To prove Statement (1.26), first note that log log n = o(log n), as shown in Example 1.26.
Therefore, log log n = o( log n) for arbitrary positive constants and . Since exp(x) is
a convex function, Theorem 1.31 gives
(log n) = o(n ).

(1.27)

As a special case of Equation (1.27), we obtain log n = o(n), which immediately gives
log n = o[n log(1 + )] for arbitrary positive constants and . Exponentiating once again
and using Theorem 1.31 yields
n = o[(1 + )n ].

Exercises for Section 1.3

Exercise 1.16 Prove that an bn if and only if |(an bn )/an | 0.
Exercise 1.17 For each of the following statements, prove the statement or provide
a counterexample that disproves it.
(a) If an bn , then limn an / limn bn = 1.
(b) If limn an / limn bn is well-defined and equal to 1, then an bn .
(c) If neither limn an nor limn bn exists, then an bn is impossible.
Exercise 1.18 Suppose that an bn and cn dn .
(a) Prove that an cn bn dn .
(b) Show by counterexample that it is not generally true that an + cn bn + dn .
(c) Prove that |an | + |cn | |bn | + |dn |.
(d) Show by counterexample that it is not generally true that f (an ) f (bn ) for
a continuous function f (x).
23

Exercise 1.19 Prove the asymptotic relationship in Example 1.23.

Hint: One way to proceed is to prove that the sum lies between two simpleto-evaluate integrals that are themselves asymptotically equivalent. Consult the
proof of Expression (1.21) as a model.
Exercise 1.20 According to the result
Pn of Exercise 1.16, the limit (1.21) implies that
the relative difference between i=1 (1/i) and log n goes to zero. But this does
not imply that the difference itself goes to zero (in general, the difference may
not even have any limit at all). In this particular case, the difference converges to
a constant called Eulers constant that is sometimes used to define the complexvalued gamma function.
P
Evaluate ni=1 (1/i)log n for various large values of n (say, n {100, 1000, 10000})
to approximate the Euler constant.
Exercise 1.21 Let X1 , . . . , Xn be a simple random sample from an exponential distribution
with density f (x) = exp(x) and consider the estimator n (X) =
Pn
X
/(n+2)
of g() = 1/. Show that for some constants c1 and c2 depending
i
i=1
on ,
bias of n c1 (variance of n )

c2
n

as n . The bias of n equals its expectation minus (1/).

Exercise 1.22 Let X1 , . . . , Xn be independent with identical density functions f (x) =
x1 I{0 < x < 1}.
(a) Let n be the posterior mean of , assuming a standard exponential prior for
(i.e., p() = e I{ > 0}). Compute n .
Hints: The posterior distribution of is gamma. If Y is a gamma random
variable, then f (y) y 1 ey and the mean of Y is /. To determine and
for the posterior distribution of , simply multiply the prior density times the
likelihood function to get an expression equal to the posterior density up to a
normalizing constant that is irrelevant in determining and .
(b) For each n {10, 50, 100, 500}, simulate 1000 different samples of size n from
the given distribution with = 2. Use these to calculate the value of n 1000
times for each n. Make a table in which you report, for each n, your estimate
of the bias (the sample mean of n 2) and the variance (the sample variance
of n ). Try to estimate the asymptotic order of the bias and the variance of
this estimator by finding nice positive exponents a and b such that na |biasn |
24

and nb variancen are roughly constant. (Nice here may be interpreted to mean
integers or half-integers.)
Hints: To generate a sample from the given distribution, use the fact that if
U1 , U2 , . . . is a sample from a uniform (0, 1) density and the continuous distribution function F (x) may be inverted explicitly, then letting Xi = F 1 (Ui ) results
in X1 , X2 , . . . being a simple random sample from F (x). When using Splus or R, a
sample from uniform (0, 1) of size, say, 50 may be obtained by typing runif(50).
Calculating n involves taking the sum of logarithms. Mathematically, this is
the same as the logarithm of the product. However, mathematically equivalent
expressions are not necessarily computationally equivalent! For a large sample,
multiplying all the values could result in overflow or underflow, so the logarithm
of the product wont always work. Adding the logarithms is safer even though
it requires more computation due to the fact that many logarithms are required
instead of just one.
Exercise 1.23 Let X1 , X2 , . . . be defined as in Exercise 1.22.
(a) Derive a formula for the maximum likelihood estimator of for a sample of
size n. Call it n .
(b) Follow the directions for Exercise 1.22(b) using n instead of n .
Exercise 1.24 Prove that if f (x) is everywhere twice differentiable and f 00 (x) 0 for
all x, then f (x) is convex.
Hint: Expand both f (x) and (1 )f (y) using Taylors theorem 1.18 with
d = 1, then add. Use the mean value theorem version of the Lagrange remainder
(1.7).
Exercise 1.25 Prove Theorem 1.31.
Hint: Let c be an arbitrary constant for which f (c) is defined. Then in inequality
(1.25), take x = bn , y = c, and = (an c)/(bn c). Be sure your proof uses all
of the hypotheses of the theorem; as Exercise 1.26 shows, all of the hypotheses
are necessary.
Exercise 1.26 Create counterexamples to the result in Theorem 1.31 if the hypotheses of the theorem are weakened as follows:
(a) Find an , bn , and convex f (x) with limx f (x) = such that an = o(bn )
but f (an ) =
6 o[f (bn )].
(b) Find an , bn , and convex f (x) such that an , bn , and an = o(bn )
25

but f (an ) 6= o[f (bn )].

(c) Find an , bn , and f (x) with limx f (x) = such that an , bn ,
and an = o(bn ) but f (an ) 6= o[f (bn )].
Exercise 1.27 Recall that log n always denotes the natural logarithm of n. Assuming
that log n means log10 n will change some of the answers in this exercise!
(a) The following 5 sequences have the property that each tends to as n ,
and for any pair of sequences, one is little-o of the other. List them in order of
rate of increase from slowest to fastest. In other words, give an ordering such that
first sequence = o(second sequence), second sequence = o(third sequence), etc.

Pn
3
n
log n!
i
2log n
(log n)log log n
i=1
Prove the 4 order relationships that result from your list.
Hint: Here and in part (b), using a computer to evaluate some of the sequences
for large values of n can be helpful in suggesting the correct ordering. However,
note that this procedure does not constitute a proof!
(b) Follow the directions of part (a) for the following 13 sequences.
n2

log(n!)
log(log n)

n
n!

nn
23 log n

log n
n

22

3n

nlog n

nn/2
(log n)n

Proving the 12 order relationships is challenging but not quite as tedious as it

sounds; some of the proofs will be very short.

1.4

Multivariate Extensions

We now consider vectors in Rk , k > 1. We denote vectors by bold face and their components
by regular type with subscripts; thus, a is equivalent to (a1 , . . . , ak ). For sequences of
vectors, we use bold face with subscripts, as in a1 , a2 , . . .. This notation has a drawback:
Since subscripts denote both component numbers and sequence numbers, it is awkward to
denote specific components of specific elements in the sequence. When necessary, we will
denote the jth component of the ith vector by aij . In other words, ai = (ai1 , . . . , aik )> for
i = 1, 2, . . .. We follow the convention that vectors are to be considered as columns instead
of rows unless stated otherwise, and the transpose of a matrix or vector is denoted by a
superscripted >.
26

The extension to the multivariate case from the univariate case is often so trivial that it is
reasonable to ask why we consider the cases separately at all. There are two main reasons.
The first is pedagogical: We feel that any disadvantage due to repeated or overlapping
material is outweighed by the fact that concepts are often intuitively easier to grasp in R
than in Rk . Furthermore, generalizing from R to Rk is often instructive in and of itself, as in
the case of the multivariate concept of differentiability. The second reason is mathematical:
Some one-dimensional results, like Taylors Theorem 1.18 for d > 2, need not (or cannot, in
some cases) be extended to multiple dimensions in these notes. In later chapters in these
notes, we will treat univariate and multivariate topics together sometimes and separately
sometimes, and we will maintain the bold-face notation for vectors throughout.
To define a limit of a sequence of vectors, we must first define a norm on Rk . We are
interested primarily in whether the norm of a vector goes to zero, a concept for which any
norm will suffice, so we may as well take the Euclidean norm:
v
u k
uX

def t
a2i = a> a.
kak =
i=1

We may now write down the analogue of Definition 1.1.

Definition 1.32 The sequence a1 , a2 , . . . is said to have limit c Rk , written an c
as n or limn an = c, if kan ck 0 as n . That is, an c means
that for any  > 0 there exists N such that kan ck <  for all n > N .
It is sometimes possible to define multivariate concepts by using the univariate definition on
each of the components of the vector. For instance, the following lemma gives an alternative
way to define an c:
Lemma 1.33 an c if and only if anj cj for all 1 j k.
Proof: Since
kan ck =

p
(an1 c1 )2 + + (ank ck )2 ,

the if part follows from repeated use of Theorem 1.5 (which says that the limit of a sum is
the sum of the limits and the limit of a product is the product of the limits) and Theorem 1.16
(which says that continuous functions preserve limits). The only if part follows because
|anj cj | kan ck for each j.
There is no multivariate analogue of Definition 1.2; it is nonsensical to write an .
However, since kan k is a real number, writing kan k is permissible. If we write
limkxk f (x) = c for a real-valued function f (x), then it must be true that f (x) tends
to the same limit c no matter what path x takes as kxk .
27

Suppose that the function f (x) maps vectors in some open subset U of Rk to vectors in R` ,
a property denoted by f : U R` . In order to define continuity, we first extend Definition
1.11 to the multivariate case:
Definition 1.34 For a function f : U R` , where U is open in Rk , we write
limxa f (x) = c for some a U and c R` if for every  > 0 there exists a
> 0 such that kf (x) ck <  whenever x U and 0 < kx ak < .
In Definition 1.34, kf (x) ck refers to the norm on R` , while kx ak refers to the norm on
Rk .
Definition 1.35 A function f : U R` is continuous at a U Rk if limxa f (x) =
f (a).
Since there is no harm in letting k = 1 or ` = 1 or both, Definitions 1.34 and 1.35 include
Definitions 1.11 and 1.14(a), respectively, as special cases.
The extension of differentiation from the univariate to the multivariate setting is not quite as
straightforward as the extension of continuity. Part of the difficulty lies merely in notation,
but we will also rely on a qualitatively different definition of the derivative in the multivariate
setting. Recall that in the univariate case, Taylors Theorem 1.18 implies that the derivative
f 0 (x) of a function f (x) satisfies
f (x + h) f (x) hf 0 (x)
0 as h 0.
h

(1.28)

It turns out that Equation (1.28) could have been taken as the definition of the derivative
f 0 (x). To do so would have required just a bit of extra work to prove that Equation (1.28)
uniquely defines f 0 (x), but this is precisely how we shall now extend differentiation to the
multivariate case:
Definition 1.36 Suppose that f : U R` , where U Rk is open. For a point a U ,
suppose there exists an ` k matrix Jf (a), depending on a but not on h, such
that
f (a + h) f (a) Jf (a)h
= 0.
h0
khk
lim

(1.29)

Then Jf (a) is unique and we call Jf (a) the Jacobian matrix of f (x) at a. We say
that f (x) is differentiable at the point a, and Jf (x) may be called the derivative
of f (x).
The assertion in Definition 1.36 that Jf (a) is unique may be proved as follows: Suppose that
(1)
(2)
Jf (a) and Jf (a) are two versions of the Jacobian matrix. Then Equation (1.29) implies

28

that

lim

h0


(1)
(2)
Jf (a) Jf (a) h
khk

= 0;

but h/khk is an arbitrary unit vector, which means that

zero matrix, proving the assertion.

(1)
Jf (a)

(2)
Jf (a)

must be the

Although Definition 1.36, sometimes called the Frechet derivative, is straightforward and
quite common throughout the calculus literature, there is unfortunately not a universally
accepted notation for multivariate derivatives. Various authors use notation such as f 0 (x),
f (x), Df (x), or f (x) to denote the Jacobian matrix or its transpose, depending on the
situation. In these notes, we adopt perhaps the most widespread of these notations, letting
f (x) denote the transpose of the Jacobian matrix Jf (x). We often refer to f as the
gradient of f .
When the Jacobian matrix exists, it is equal to the matrix of partial derivatives, which are
defined as follows:
Definition 1.37 Let g(x) be a real-valued function defined on a neighborhood of a
in Rk . For 1 i k, let ei denote the ith standard basis vector in Rk , consisting
of a one in the ith component and zeros elsewhere. We define the ith partial
derivative of g(x) at a to be

g(a + hei ) g(a)
g(x) 
def
,
= lim

h0
xi x=a
h
if this limit exists.
Now we are ready to state that the Jacobian matrix is the matrix of partial derivatives.
Theorem 1.38 Suppose f (x) is differentiable at a in the sense of Definition 1.36.
Define the gradient matrix f (a) to be the transpose of the Jacobian matrix
Jf (a). Then

f1 (x)

fx` (x)
x1

1
..


.
.
f (a) = .
(1.30)
. 
f1 (x)
f` (x)


xk

xk

x=a.

The converse of Theorem 1.38 is not true, in the sense that the existence of partial derivatives
of a function does not guarantee the differentiability of that function (see Exercise 1.31).
When f maps k-vectors to `-vectors, f (x) is a k ` matrix, a fact that is important to
memorize; it is often very helpful to remember the dimensions of the gradient matrix when
29

trying to recall the form of various multivariate results. To try to simplify the admittedly
confusing notational situation resulting from the introduction of both a Jacobian matrix
and a gradient, we will use only the gradient notation f (x), defined in Equation (1.30),
throughout these notes.
By Definition 1.36, the gradient matrix satisfies the first-order Taylor formula
f (x) = f (a) + f (a)> (x a) + r(x, a),

(1.31)

where r(x, a)/kx ak 0 as x a.

Now that we have generalized Taylors Theorem 1.18 for the linear case d = 1, it is worthwhile
to ask whether a similar generalization is necessary for larger d. The answer is no, except for
one particular case: We will require a second-order Taylor expansion (that is, d = 2) when
f (x) is real-valued but its argument x is a vector. To this end, suppose that U Rk is open
and that f (x) maps U into R. Then according to Equation (1.30), f (x) is a k 1 vector of
partial derivatives, which means that f (x) maps k-vectors to k-vectors. If we differentiate
once more and evaluate the result at a, denoting the result by 2 f (a), then Equation (1.30)
with /xi f (x) substituted for fi (x) gives

2 f (x)
2 f (x)



x1 xk
x21


.
.
2

.
..
(1.32)
f (a) =

2.
2

f (x)
f (x)

x
2
xk x1
k

x=a.

Definition 1.39 The k k matrix on the right hand side of Equation (1.32), when
it exists, is called the Hessian matrix of the function f (x) at a.
Twice differentiability guarantees the existence (by two applications of Theorem 1.38) and
symmetry (by Theorem 1.40 below) of the Hessian matrix. The Hessian may exist for a
function that is not twice differentiable, as seen in Exercise 1.33, but this mathematical
curiosity will not concern us elsewhere in these notes.
We state the final theorem of this section, which extends second-order Taylor expansions to
a particular multivariate case, without proof, but the interested reader may consult Magnus
and Neudecker (1999) for an encyclopedic treatment of this and many other topics involving
differentiation.
Theorem 1.40 Suppose that the real-valued function f (x) is twice differentiable at
some point a Rk . Then 2 f (a) is a symmetric matrix, and
1
f (x) = f (a) + f (a)> (x a) + (x a)> 2 f (a)(x a) + r2 (x, a),
2
where r2 (x, a)/kx ak2 0 as x a.
30

Exercises for Section 1.4

Exercise 1.28 (a) Suppose that f (x) is continuous at 0. Prove that f (tei ) is continuous as a function of t at t = 0 for each i, where ei is the ith standard basis
vector.
(b) Prove that the converse of (a) is not true by inventing a function f (x) that
is not continuous at 0 but such that f (tei ) is continuous as a function of t at
t = 0 for each i.
Exercise 1.29 Suppose that anj cj as n for j = 1, . . . , k. Prove that if
f : Rk R is continuous at the point c, then f (an ) f (c). This proves every
part of Exercise 1.1. (The hard work of an exercise like 1.1(b) is in showing that
multiplication is continuous).
Exercise 1.30 Prove Theorem 1.38.
Hint: Starting with Equation (1.29), take x = a + tei and let t 0, where ei
is defined in Definition 1.37.
Exercise 1.31 Prove that the converse of Theorem 1.38 is not true by finding a
function that is not differentiable at some point but whose partial derivatives at
that point all exist.
Exercise 1.32 Suppose that X1 , . . . , Xn comprises a sample of independent and identically distributed normal random variables with density
exp{ 21 2 (xi )2 }

.
f (xi ; , ) =
2 2
2

Let `(, 2 ) denote

the loglikelihood function; i.e., `(, 2 ) is the logarithm of the
Q
joint density i f (Xi ; , 2 ), viewed as a function of the parameters and 2 .
The score vector is defined to be the gradient of the loglikelihood. Find the score
vector for this example.
Hint: The score vector is a vector with two components and it is a function of
X1 , . . . , Xn , , and 2 . Setting the score vector equal to zero and solving for
and 2 gives
maximum likelihood estimators of and 2 , namely
P the well-known
1
2
X and n i (Xi X) .
Exercise 1.33 Define

f (x, y) =

x3 yxy 3
x2 +y 2

31

if x = y = 0;
otherwise.

Use Theorem 1.40 to demonstrate that f (x, y) is not twice differentiable at (0, 0)
by showing that 2 f (0, 0), which does exist, is not symmetric.
Exercise 1.34 (a) Find the Hessian matrix of the loglikelihood function defined in
Exercise 1.32.
(b) Suppose that n = 10 and that we observe this sample:
2.946 0.975
2.627 -0.628

1.333
2.476

4.484
2.599

1.711
2.143

Evaluate the Hessian matrix at the maximum likelihood estimator (

,
2 ). (A
formula for the MLE is given in the hint to Exercise 1.32).
(c) As we shall see in Chapter 7, the negative inverse of the Hessian matrix is a
reasonable large-sample estimator of the covariance matrix of the MLE (though
with only n = 10, it is not clear how good this estimator would be in this
example!). Invert your answer from part (b), then put a negative sign in front
and use the answer to give approximate standard errors (the square roots of the
diagonal entries) for
and
2.
Exercise 1.35 Suppose X1 , . . . , Xn is a sample of independent and identically distributed random variables from a Beta(, ) distribution, for which the density
function is
f (x; , ) =

( + ) 1
x (1 x)1
()()

for 0 < x < 1,

where and are assumed to be positive parameters.

(a) Calculate the score vector (the gradient of the loglikelihood) and the Hessian of the loglikelihood. Recall the definitions of the digamma and trigamma
functions in Exercises (1.14) and (1.15).
Exercise 1.36 The gamma distribution with shape parameter > 0 and rate parameter > 0 has density function
f (x; , ) =

1 x
x e
()

for x > 0.

(a) Calculate the score vector for an independent and identically distributed
gamma(, ) sample of size n.
(b) Using the approximation to the digamma function (x) given in Equation
(1.16), find a closed-form approximation to the maximum likelihood estimator
32

(obtained by setting the score vector equal to zero and solving for and ).
Simulate 1000 samples of size n = 50 from gamma(5, 1) and calculate this approximation for each. Give histograms of these estimators. Can you characterize
their performance?
The approximation of (x) in Equation (1.16) can be extremely poor for x < 2,
so the method above is not a reliable general-purpose estimation procedure.

1.5

Expectation and Inequalities

While random variables have made only occasional appearances in these notes before now,
they will be featured prominently from now on. We do not wish to make the definition of a
random variable rigorous hereto do so requires measure theorybut we assume that the
reader is familiar with the basic idea: A random variable is a function from a sample space
into R. (We often refer to random vectors rather than random variables if the range
space is Rk rather than R.)
For any random variable X, we denote the expected value of X, if this value exists, by
E X. We assume that the reader is already familiar with expected values for commonlyencountered random variables, so we do not attempt here to define the expectation operator
E rigorously. In particular, we avoid writing explicit formulas for E X (e.g., sums if X is
discrete or integrals if X is continuous) except when necessary. Much of the theory in these
notes may be developed using only the E X notation; exceptions include cases in which
we wish to evaluate particular expectations and cases in which we must deal with density
functions (such as the topic of maximum likelihood estimation). For students who have not
been exposed to any sort of a rigorous treatment of random variables and expectation, we
hope that the many applications of this theory presented here will pique your curiosity and
encourage you to delve further into the technical details of random variables, expectations,
and conditional expectations. Nearly any advanced probability textbook will develop these
details. For a quick, introductory-level exposure to these intricacies, we recommend the first
chapter of Lange (2003).
Not all random variables have expectations, even if we allow the possibilities E X = :
Let X + = max{X, 0} and X = max{X, 0} denote the positive and negative parts of X,
so that X = X + X . Now both E X + and E X are always well-defined if we allow as
a possibility, but if both X + and X have infinite expectation, then there is no sensible way
to define E X. It is easy to find examples of random variables X for which E X is undefined.
Perhaps the best-known example is a Cauchy random variable (whose density function is
given in Exercise 7.3), but we may construct other examples by taking any two independent
nonnegative random variables Y1 and Y2 with infinite expectatione.g., let Yi take the value
33

2n with probability 2n for all positive integers nand simply defining X = Y1 Y2 .

The expectation operator has several often-used properties, listed here as axioms because
we will not derive them from first principles. We assume below that X and Y are defined
on the same sample space and E X and E Y are well-defined.
1. Linearity: For any real numbers a and b, E (aX + bY ) = a E (X) + b E (Y ) (and if
a E (X) + b E (Y ) is undefined, then so is E (aX + bY )).
2. Monotonicity: If X() Y () for all , then E X E Y .
3. Conditioning: If E (X|Y ) denotes the conditional expectation of X given Y (which,
as a function of Y , is itself a random variable), then E X = E {E (X|Y )}.
As a special case of the conditioning property, note that if X and Y are independent, then
E (X|Y ) = E X, which gives the well-known identity
E XY = E {E (XY |Y )} = E {Y E (X|Y )} = E {Y E X} = E X E Y,
where we have used the fact that E (XY |Y ) = Y E (X|Y ), which is always true because
conditioning on Y is like holding it constant.
The variance and covariance operators are defined as usual, namely,
def

Cov (X, Y ) = E XY (E X)(E Y )

def

and Var (X) = Cov (X, X). The linearity property above extends to random vectors: For
scalars a and b we have E (aX + bY) = a E (X) + b E (Y), and for matrices P and Q with
dimensions such that P X + QY is well-defined, E (P X + QY) = P E (X) + Q E (Y). The
covariance between two random vectors is
def

Cov (X, Y) = E XY> (E X)(E Y)> ,

and the variance matrix of a random vector (sometimes referred to as the covariance matrix)
def
is Var (X) = Cov (X, X). Among other things, these properties imply that
Var (P X) = P Var (X)P >

(1.33)

for any constant matrix P with as many columns as X has rows.

Example 1.41 As a first application of the monotonicity of the expectation operator,
we derive a useful inequality called Chebyshevs inequality. For any positive
constants a and r and any random variable X, observe that
|X|r |X|r I{|X| a} ar I{|X| a},
34

where throughout these notes, I{} denotes the indicator function


1 if expression is true
def
I{expression} =
0 if expression is not true.

(1.34)

Since E I{|X| a} = P (|X| a), the monotonicity of the expectation operator

implies
P (|X| a)

E |X|r
.
ar

(1.35)

Inequality (1.35) is sometimes called Markovs inequality. In the special case that
X = Y E Y and r = 2, we obtain Chebyshevs inequality: For any a > 0 and
any random Y ,
P (|Y E Y | a)

Var Y
.
a2

(1.36)

Example 1.42 We now derive another inequality, Jensens, that takes advantage of
linearity as well as monotonicity. Jensens inequality states that
f (E X) E f (X)

(1.37)

for any convex function f (x) and any random variable X. Definition 1.30 tells
precisely what a convex function is, but the intuition is simple: Any line segment
connecting two points on the graph of a convex function must never go below the
graph (valley-shaped graphs are convex; hill-shaped graphs are not). To prove
inequality 1.37, we require another property of any convex function, called the
supporting hyperplane property. This property, whose proof is the subject of
Exercise 1.38, essentially guarantees that for any point on the graph of a convex
function, it is possible to construct a hyperplane through that point that puts
the entire graph on one side of that hyperplane.
In the context of inequality (1.37), the supporting hyperplane property guarantees
that there exists a line g(x) = ax + b through the point [E X, f (E X)] such
that g(x) f (x) for all x (see Figure 1.2). By monotonicity, we know that
E g(X) E f (X). We now invoke the linearity of the expectation operator to
conclude that
E g(X) = g(E X) = f (E X),
which proves inequality (1.37).

Exercises for Section 1.5

35

g(E X) = f(E X)

g(x)

f(x)

EX

Figure 1.2: The solid curve is a convex function f (x) and the dotted line is a supporting
hyperplane g(x), tangent at x = E X. This figure shows how to prove Jensens inequality.
Exercise 1.37 Show by example that equality can hold in inequality 1.36.
Exercise 1.38 Let f (x) be a convex function on some interval, and let x0 be any
point on the interior of that interval.
(a) Prove that
f (x) f (x0 )
xx0 +
x x0
exists and is finite; that is, a one-sided derivative exists at x0 .
lim

(1.38)

Hint: Using Definition 1.30, show that the fraction in expression (1.38) is nonincreasing and bounded below as x decreases to x0 .
(b) Prove that there exists a linear function g(x) = ax+b such that g(x0 ) = f (x0 )
and g(x) f (x) for all x in the interval. This fact is the supporting hyperplane
property in the case of a convex function taking a real argument.
Hint: Let f 0 (x0 +) denote the one-sided derivative of part (a). Consider the line
f (x0 ) + f 0 (x0 +)(x x0 ).
Exercise 1.39 Prove Holders inequality: For random variables X and Y and positive
p and q such that p + q = 1,

q

p
E |XY | E |X|1/p
E |Y |1/q .
(1.39)
36

(If p = q = 1/2, inequality (1.39) is also called the Cauchy-Schwartz inequality.)

Hint: Use the convexity of exp(x) to prove that |abXY | p|aX|1/p + q|bY |1/q
whenever aX 6= 0 and bY 6= 0 (the same inequality is also true if aX = 0 or
bY = 0). Take expectations, then find values for the scalars a and b that give the
desired result when the right side of inequality (1.39) is nonzero.
Exercise 1.40 Use Holders Inequality (1.39) to prove that if > 1, then
(E |X|) E |X| .
Hint: Take Y to be a constant in Inequality (1.39).
Exercise 1.41 Kolmogorovs inequality is a strengthening of Chebyshevs inequality
for a sum of independent random variables: If X1 , . . . , Xn are independent random
variables, define
Sk =

k
X

(Xi E Xi )

i=1

to be the centered kth partial sum for 1 k n. Then for a > 0, Kolmogorovs
inequality states that


Var Sn
.
(1.40)
P max |Sk | a
1kn
a2
(a) Let Ak denote the event that |Si | a for the first time when i = k; that is,
that |Sk | a and |Sj | < a for all j < k. Prove that

 X
n


2
a P max |Sk | a
E I{Ak }Sk2 .
1kn

i=1

Hint: Argue that

n
X


E I{Ai } = P

i=1


max |Sk | a

1kn

and E [I{Ak }Sk2 ] a2 E I{Ak }.

(b) Prove that
E Sn2

n
X



E I{Ak }{Sk2 + 2Sk (Sn Sk )} .

k=1

37

Hint: Use the fact that the Ak are nonoverlapping, which implies that 1
I(A1 ) + + I(An ). Also use Sn2 = Sk2 + 2Sk (Sn Sk ) + (Sn Sk )2 .
(c) Using parts (a) and (b), prove inequality (1.40).
Hint: By independence,
E [I{Ak }Sk (Sn Sk )] = E [I{Ak }Sk ] E (Sn Sk ).
What is E (Sn Sk )?
Exercise 1.42 Try a simple numerical example to check how much sharper Kolmogorovs inequality (1.40) is than Chebyshevs inequality (1.36).
(a) Take n = 8 and assume that X1 , . . . , Xn are independent normal random
variables with E Xi = 0 and Var Xi = 9 i. Take a = 12. Calculate the exact
values on both sides of Chebyshevs inequality (1.36).
(b) Simulate 104 realizations of the situation described in part (a). For each,
record the maximum value attained by |Sk | for k = 1, . . . , 8. Approximate the
probability on the left hand side of Kolmogorovs inequality (1.40). Describe
what you find when you compare parts (a) and (b). How does a histogram of the
maxima found in part (b) compare with the distribution of |Sn |?
Exercise 1.43 The complex
plane C consists of all points x + iy, where x and y are
real numbers and i = 1. The elegant result known as Eulers formula relates
the points on the unit circle to the complex exponential function:
exp{it} = cos t + i sin t

for all t R.

(1.41)

Because eit is on the unit circle for all real-valued t, the norm (also known as
the modulus) of eit , denoted |eit |, equals 1. This fact leads to the following
generalization of the triangle inequality: For any real-valued function g(x) and
any real number t,
Z t
 Z t
 Z


  
   t

 g(x)eix dx  g(x)eix  dx =  |g(x)| dx .
(1.42)

 
 

0

The inequalities below in parts (a) through (d) involving exp{it} will be used
in Chapter 4. Assume t is a real number, then use Equations (1.6) and (1.41),
together with Inequality (1.42), to prove them. [Since we only claim Equation
(1.6) to be valid for real-valued functions of real variables, it is necessary here
to use Eulers formula to separate eit into its real and imaginary parts, namely
cos t and sin t, then Taylor-expand them separately before reassembling the parts
using Eulers formula again.]
38

(a) In Equation (1.6), use a = 0 and d = 0 on both cos t and sin t to show that
for any t R,
| exp{it} 1| |t|.
(b) Proceed as above but with d = 1 to show that
| exp{it} 1 it| t2 /2.
(c) Proceed as above but with d = 2 to show that




1
2
exp{it} 1 it + t  |t|3 /6.

2 
(d) Proceed as above but using d = 1 for sin t, then d = 2 together with
integration by parts for cos t, to show that




exp{it} 1 it + 1 t2  t2 .

2 


Exercise 1.44 Refer to Exercise 1.43. Graph the functions exp{it} 1 it + 12 t2 ,
|t|3 /6, and t2 for t in the interval [10, 10]. Graph the three curves on the same
set of axes, using different plotting styles so they are distinguishable from one
another. As a check, verify that the inequalities in Exercises 1.43(c) and (d)
appear to be satisfied.
p
Hint: The modulus |z| of a complex number z = x + iy equals x2 + y 2 . Refer
to Equation (1.41) to deal with the expression exp{it}.
Exercise 1.45 For any nonnegative random variable Y with finite expectation, prove
that

P (Y i) E Y.

(1.43)

i=1

Hint: First, prove that equality holds if Y is supported on the nonnegative

integers. Then note for a general Y that E bY c E Y , where bxc denotes the
greatest integer less than or equal to x.
Though we will not do so here, it is possible to prove a statement stronger than
inequality (1.43) for nonnegative random variables, namely,
Z
P (Y t) dt = E Y.
0

39

(This equation
remains true Rif E Y = .) To sketch a proof, note that if we
R
can prove E f (Y, t) dt = E f (Y, t) dt, the result follows immediately by taking
f (Y, t) = I{Y t}.

40

Chapter 2
Weak Convergence
Chapter 1 discussed limits of sequences of constants, either scalar-valued or vector-valued.
Chapters 2 and 3 extend this notion by defining what it means for a sequence of random
variables to have a limit. As it turns out, there is more than one sensible way to do this.
Chapters 2 and 4 (and, to a lesser extent, Chapter 3) lay the theoretical groundwork for
nearly all of the statistical topics that will follow. While the material in Chapter 2 is
essential, readers may wish to skip Chapter 3 on a first reading. As is common throughout
the book, some of the proofs here have been relegated to the exercises.

2.1

Modes of Convergence

Whereas the limit of a sequence of real numbers is unequivocally expressed by Definition

1.32, in the case of random variables there are several ways to define the convergence of a
sequence. This section discusses three such definitions, or modes, of convergence; Section 3.1
presents a fourth. Because it is often easier to understand these concepts in the univariate
case than the multivariate case, we only consider univariate random vectors here, deferring
the analogous multivariate topics to Section 2.3.

2.1.1

Convergence in Probability

What does it mean for the sequence X1 , X2 , . . . of random variables to converge to, say,
the random variable X? Under what circumstances should one write Xn X? We begin
by considering a definition of convergence that requires that Xn and X be defined on the
same sample space. For this form of convergence, called convergence in probability, the
41

absolute difference |Xn X|, itself a random variable, should be arbitrarily close to zero
with probability arbitrarily close to one. More precisely, we make the following definition.
Definition 2.1 Let {Xn }n1 and X be defined on the same probability space. We
P
say that Xn converges in probability to X, written Xn X, if for any  > 0,
P (|Xn X| < ) 1 as n .

(2.1)

It is very common that the X in Definition 2.1 is a constant, say X c. In such cases, we
P
simply write Xn c. When we replace X by c in Definition 2.1, we do not need to concern
ourselves with the question of whether X is defined on the same sample space as Xn because
any constant may be defined as a random variable on any sample space. In the most common
statistical usage of convergence to a constant c, we take c to be some parameter and Xn
to be an estimator of :
P

Definition 2.2 If Xn , Xn is said to be consistent (or weakly consistent) for .

As the name suggests, weak consistency is weaker than (i.e., implied by) a condition called
strong consistency, which will be defined in Chapter 3. Consistency, used without the
word strong or weak, generally refers to weak consistency. Throughout this book, we
shall refer repeatedly to (weakly) consistent estimators, whereas strong consistency plays a
comparatively small role.
Example 2.3 Suppose that X1 , X2 , . . . are independent and identically distributed
(i.i.d.) uniform (0, ) random variables, where is an unknown positive constant.
For n 1, let X(n) be defined as the largest value among X1 through Xn : That
def

is, X(n) = max1in Xi . Then we may show that X(n) is a consistent estimator
of as follows:
By Definition 2.1, we wish to show that for an arbitrary  > 0, P (|X(n) | <
) 1 as n . In this particular case, we can evalulate P (|X(n) | < )
directly by noting that X(n) cannot possibly be larger than , so that
P (|X(n) | < ) = P (X(n) > ) = 1 P (X(n) ).
The maximum X(n) is less than some constant if and only if each of the random
variables X1 , . . . , Xn is less than that constant. Therefore, since the Xi are i.i.d.,

[1 (/)]n if 0 <  <
n
P (X(n) ) = [P (X1 )] =
0
if  .
Since 1 (/) is strictly less than 1, we conclude that no matter what positive
value  takes, P (Xn ) 0 as desired.
42

2.1.2

Probabilistic Order Notation

There are probabilistic analogues of the o and O notations of Section 1.3 that apply to
random variable sequences instead of real number sequences.
P

Definition 2.4 We write Xn = oP (Yn ) if Xn /Yn 0.

In particular, oP (1) is shorthand notation for a sequence of random variables that converges
to zero in probability, as illustrated in Equation (2.2) below.
Definition 2.5 We write Xn = OP (Yn ) if for every  > 0, there exist M and N such
that
 

 Xn 


P   < M > 1  for all n > N .
Yn
As a special case of Definition 2.5, we refer to any OP (1) sequence as a bounded in probability
sequence:
Definition 2.6 We say that X1 , X2 , . . . is bounded in probability if Xn = OP (1), i.e.,
if for every  > 0, there exist M and N such that P (|Xn | < M ) > 1 for n > N .
Definition 2.6 is primarily useful because of the properties of bounded in probability sequences established in Exercise 2.2.
Example 2.7 In Example 2.3, we showed that if X1 , X2 , . . . are independent and
identically distributed uniform (0, ) random variables, then
P

max Xi

as n .

1in

Equivalently, we may say that

max Xi = + oP (1)

1in

as n .

(2.2)

as n ,

(2.3)

It is also technically correct to write

max Xi = + OP (1)

1in

though Statement (2.3) is less informative than Statement (2.2). On the other
hand, we will see in Example 6.1 that Statement (2.3) may be sharpened considerably
and made more informative than Statement (2.2)by writing
 
1
as n .
max Xi = + OP
1in
n
43

Using the oP notation defined above, it is possible to rewrite Taylors theorem 1.18 in a
form involving random variables. This theorem will prove to be useful in later chapters; for
instance, it is used to prove the result known as the delta method in Section 5.1.1.
P

Theorem 2.8 Suppose that Xn 0 for a sequence of random variables X1 , X2 , . . .

and a constant 0 . Furthermore, suppose that f (x) has d derivatives at the point
0 . Then there is a random variable Yn such that
f (Xn ) = f (0 ) + (Xn 0 )f 0 (0 ) + +

(Xn 0 )d  (d)
f (0 ) + Yn (2.4)
d!

and Yn = oP (1) as n .
The proof of Theorem 2.8 is a useful example of an epsilon-delta proof (named for the 
and in Definition 1.11).
Proof: Let
(
Yn =

d!
(Xn 0 )d

h
f (Xn ) f (0 ) (Xn 0 )f 0 (0 )

(Xn 0 )d1
(d1)!

if Xn 6= 0
if Xn = 0 .
P

Then Equation (2.4) is trivially satisfied. We will show that Yn = oP (1), which means Yn 0,
by demonstrating that for an arbitrary  > 0, there exists N such that P (|Yn | < ) > 1 
for all n > N . By Taylors Theorem 1.18, there exists some > 0 such that |Xn 0 | <
implies |Yn | <  (that is, the event { : |Xn () 0 | < } is contained in the event { :
P
|Yn ()| < }). Furthermore, because Xn 0 , we know that there exists some N such that
P (|Xn 0 | < ) > 1  for all n > N . Putting these facts together, we conclude that for
all n > N ,
P (|Yn | < ) P (|Xn 0 | < ) > 1 ,
which proves the result.
In later chapters, we will generally write simply

(Xn 0 )d  (d)
f (Xn ) = f (0 ) + (Xn 0 )f (0 ) + +
f (0 ) + oP (1)
as n (2.5)
d!
when referring to the result of Theorem 2.8. A technical quibble with Expression (2.5) is
that it suggests that any random variable Yn satisfying (2.4) must also be oP (1). This is not
quite true: Since Yn may be defined arbitrarily in the event that Xn = 0 and still satisfy
(2.4), if
0

P (Xn = 0 ) > c

for all n

for some positive constant c, then Yn 6= oP (1) may still satisfy (2.4). However, as long as
one remembers what Theorem 2.8 says, there is little danger in using Expression (2.5).
44

2.1.3

Convergence in Distribution

As the name suggests, convergence in distribution (also known as convergence in law) has
to do with convergence of the distribution functions (or laws) of random variables. Given
a random variable X, the distribution function of X is the function
F (x) = P (X x).

(2.6)

Any distribution function F (x) is nondecreasing and right-continuous, and it has limits
limx F (x) = 0 and limx F (x) = 1. Conversely, any function F (x) with these properties is a distribution function for some random variable.
It is not enough to define convergence in distribution as simple pointwise convergence of a
sequence of distribution functions; there are technical reasons that such a simplistic definition
fails to capture any useful concept of convergence of random variables. These reasons are
illustrated by the following two examples.
Example 2.9 Let Xn be normally distributed with mean
0 and variance n. Then
the distribution function of Xn is Fn (x) = (x/ n), where (z) denotes the
standard normal distribution function. Because (0) = 1/2, we see that for any
fixed x, Fn (x) 1/2 as n . But the function that is constant at 1/2 is not
a distribution function. This example shows that not all convergent sequences of
distribution functions have limits that are distribution functions.
Example 2.10 By any sensible definition of convergence, 1/n should converge to 0
as n . But consider the distribution functions Fn (x) = I{x 1/n} and
F (x) = I{x 0} corresponding to the constant random variables 1/n and 0. We
do not have pointwise convergence of Fn (x) to F (x), since Fn (0) = 0 for all n but
F (0) = 1. However, Fn (x) F (x) is true for all x 6= 0. Not coincidentally, the
point x = 0 where convergence of Fn (x) to F (x) fails is the only point at which
the function F (x) is not continuous.
To write a sensible definition of convergence in distribution, Example 2.9 demonstrates that
we should require that the limit of distribution functions be a distribution function itself,
say F (x), while Example 2.10 suggests that we should exclude points where F (x) is not
continuous. We therefore arrive at the following definition:
Definition 2.11 Suppose that X has distribution function F (x) and that Xn has
distribution function Fn (x) for each n. Then we say Xn converges in distribution
d
to X, written Xn X, if Fn (x) F (x) as n for all x at which F (x) is
continuous. Convergence in distribution is sometimes called convergence in law
L
and written Xn X.
45

The notation of Definition 2.11 may be stretched a bit; sometimes the expressions on either
d
side of the symbol may be distribution functions or other notations indicating certain
distributions, rather than actual random variables as in the definition. The meaning is
always clear even if the notation is not consistent.
d

However, one common mistake should be avoided at all costs: If (or or any other limit
arrow) indicates that n , then n must never appear on the right side of the arrow. See
Expression (2.8) in Example 2.12 for an example of how this rule is sometimes violated.
Example 2.12 The Central Limit Theorem for i.i.d. sequences: Let X1 , . . . , Xn be
independent and identically distributed (i.i.d.) with mean and finite variance
2 . Then by a result that will be covered in Chapter 4 (but which is perhaps
already known to the reader),
!
n

1X
d
(2.7)
n
Xi N (0, 2 ),
n i=1
where N (0, 2 ) denotes a normal distribution with mean 0 and variance 2 .
[N (0, 2 ) is not actually a random variable; this is an example of stretching
d
the notation referred to above.]
Because Equation 2.7 may be interpreted as saying that the sample mean X n
has approximately a N (, 2 /n) distribution, it may seem tempting to rewrite
Equation (2.7) as
n

1X
d
Xi N
n i=1



2
,
.
n

(2.8)

Resist the temptation to do this! As pointed out above, n should never appear
on the right side of a limit arrow (as long as that limit arrow expresses the idea
that n is tending to ).
By the result of Exercise 2.2, the limit statement (2.7) implies that the left side
of
that statement is OP (1). We may therefore write (after dividing through by
n and adding )
n

1X
Xi = + OP
n i=1


as n .

(2.9)

Unlike Expression (2.8), Equation (2.9) is perfectly legal;and although it is less

specific than Expression (2.7), it expresses at a glance the n-rate of convergence
of the sample mean to .
46

Unlike Xn X, the expression Xn X does not require Xn X to be a random variable;

d
in fact, Xn X is possible even if Xn and X are not defined on the same sample space.
Even if Xn and X do have a joint distribution, it is easy to construct an example in which
d
Xn X but Xn does not converge to X in probability: Take Z1 and Z2 to be independent
and identically distributed standard normal random variables, then let Xn = Z1 for all n
d
and X = Z2 . Since Xn and X have exactly the same distribution by construction, Xn X
in this case. However, since Xn X is a N (0, 2) random variable for all n, we do not have
P
Xn X.
d

We conclude that Xn X cannot possibly imply Xn X (but see Theorem 2.14 for a special
case in which it does). However, the implication in the other direction is always true:
P

Theorem 2.13 If Xn X, then Xn X.

Proof: Let Fn (x) and F (x) denote the distribution functions of Xn and X, respectively.
P
Assume that Xn X. We need to show that Fn (t) F (t), where t is any point of continuity
of F (x).
Choose any  > 0. Whenever Xn t, it must be true that either X t +  or |Xn X| > .
This implies that
Fn (t) F (t + ) + P (|Xn X| > ).
Similarly, whenever X t , either Xn t or |Xn X| > , implying
F (t ) Fn (t) + P (|Xn X| > ).
We conclude that for arbitrary n and  > 0,
F (t ) P (|Xn X| > ) Fn (t) F (t + ) + P (|Xn X| > ).

(2.10)

Taking both the lim inf n and the lim supn of the above inequalities, we conclude [since
P
Xn X implies P (|Xn X| > ) 0] that
F (t ) lim inf Fn (t) lim sup Fn (t) F (t + )
n

for all . Since t is a continuity point of F (x), letting  0 implies

F (t) = lim inf Fn (t) = lim sup Fn (t),
n

so we conclude Fn (t) F (t) and the theorem is proved.

d

We remarked earlier that Xn X could not possibly imply Xn X because the latter
expression requires that Xn and X be defined on the same sample space for every n. However,
a constant c may be considered to be a random variable defined on any sample space; thus,
d
P
it is reasonable to ask whether Xn c implies Xn c. The answer is yes:
47

Theorem 2.14 Xn c if and only if Xn c.

d

Proof: We only need to prove that Xn c implies Xn c, since the other direction is a
special case of Theorem 2.13. If F (x) is the distribution function I{x c} of the constant
random variable c, then c +  and c  are points of continuity of F (x) for any  > 0.
d
Therefore, Xn c implies that Fn (c ) F (c ) = 0 and Fn (c + ) F (c + ) = 1 as
n . We conclude that
P ( < Xn c ) = Fn (c + ) Fn (c ) 1,
P

which means Xn c.
When we speak of convergence of random variables to a constant in this book, most commonly we refer to convergence in probability, which (according to Theorem 2.14) is equivalent
to convergence in distribution. On the other hand, when we speak of convergence to a random variable, we nearly always refer to convergence in distribution. Therefore, in a sense,
Theorem 2.14 makes convergence in distribution the most important form of convergence in
this book. This type of convergence is often called weak convergence.

2.1.4

Convergence in Mean

The third and final mode of convergence in this chapter is useful primarily because it is
sometimes easy to verify and thus gives a quick way to prove convergence in probability, as
Theorem 2.17 below implies.
Definition 2.15 Let a be a positive constant. We say that Xn converges in ath mean
a
to X, written Xn X, if
E |Xn X|a 0 as n .

(2.11)

Two specific cases of Definition 2.15 deserve special mention. When a = 1, we normally omit
mention of the a and simply refer to the condition E |Xn X| 0 as convergence in mean.
Convergence in mean is not equivalent to E Xn E X: For one thing, E Xn E X is
possible without any regard to the joint distribution of Xn and X, whereas E |Xn X| 0
clearly requires that Xn X be a well-defined random variable.
Even more important than a = 1 is the special case a = 2:
qm

Definition 2.16 We say that Xn converges in quadratic mean to X, written Xn X,

if
E |Xn X|2 0 as n .
48

Convergence in quadratic mean is important for two reasons. First, it is often quite easy
qm
to check; in Exercise 2.6, you are asked to prove that Xn c if and only if E Xn c
and Var Xn 0 for some constant c. Second, quadratic mean convergence (indeed, ath
mean convergence for any a > 0) is stronger than convergence in probability, which means
that weak consistency of an estimator may be established by checking that it converges in
quadratic mean. This latter property is a corollary of the following result:
qm

Theorem 2.17 (a) For a constant c, Xn c if and only if E Xn c and Var Xn 0.

a

(b) For fixed a > 0, Xn X implies Xn X.

Proof: Part (a) is the subject of Exercise 2.6. Part (b) relies on Markovs inequality (1.35),
which states that
1
(2.12)
P (|Xn X| ) a E |Xn X|a

a

for an arbitrary fixed  > 0. If Xn X, then by definition the right hand side of inequality
P
(2.12) goes to zero as n , so the left side also goes to zero and we conclude that Xn X
by definition.
Example 2.18 Any unbiased estimator is consistent if its variance goes to zero. This
fact follows directly from Theorem 2.17(a) and (b). As an example, consider a
sequence of independent and identically distributed random variables X1 , X2 , . . .
with mean and finite variance 2 . The sample mean
n

1X
Xn =
Xi
n i=1
has mean and variance 2 /n. Therefore, X n is unbiased and its variance goes
P
to zero, so we conclude that it is consistent; i.e., X n . This fact is the Weak
Law of Large Numbers (see Theorem 2.19) for the case of random variables with
finite variance.

Exercises for Section 2.1

P

Exercise 2.1 For each of the three cases below, prove that Xn 1:
(a) Xn = 1 + nYn , where Yn is a Bernoulli random variable with mean 1/n.
P
(b) Xn = Yn / log n, where Yn is a Poisson random variable with mean ni=1 (1/i).
P
(c) Xn = n1 ni=1 Yi2 , where the Yi are independent standard normal random
variables.
49

Exercise 2.2 This exercise deals with bounded in probability sequences; see Definition 2.6.
d

(a) Prove that if Xn X for some random variable X, then Xn is bounded in

probability.
Hint: You may use the fact that any interval of real numbers must contain a
point of continuity of F (x). Also, recall that F (x) 1 as x .
P

(b) Prove that if Xn is bounded in probability and Yn 0, then Xn Yn 0.

Hint: For fixed  > 0, argue that there must be M and N such that P (|Xn | <
M ) > 1 /2 and P (|Yn | < /M ) > 1 /2 for all n > N . What is then the
smallest possible value of P (|Xn | < M and |Yn | < /M )? Use this result to prove
P
Xn Yn 0.
Exercise 2.3 The Poisson approximation to the binomial:
(a) Suppose that Xn is a binomial random variable with n trials, where the
probability of success on each trial is /n. Let X be a Poisson random variable
d
with the same mean as Xn , namely . Prove that Xn X.
Hint: Argue that it suffices to show that P (Xn = k) P (X = k) for all
nonnegative integers k. Then use Stirlings formula (1.19).
(b) Part (a) can be useful in approximating binomial probabilities in cases where
the number of trials is large but the success probability is small: Simply consider
a Poisson random variable with the same mean as the binomial variable. Assume
that Xn is a binomial random variable with parameters n and 2/n. Create a plot
on which you plot P (X10 = k) for k = 0, . . . , 10. On the same set of axes, plot
the same probabilities for X20 , X50 , and the Poisson variable well denote by X .
Try looking at the same plot but with the probabilities transformed using the
logit (log-odds) transformation logit(t) = log(t) log(1 t). Which plot makes
it easier to characterize the trend you observe?
Exercise 2.4 Suppose that X1 , . . . , Xn are independent and identically distributed
Uniform(0, 1) random variables. For a real number t, let
Gn (t) =

n
X

I{Xi t}.

i=1

(a) What is the distribution of Gn (t) if 0 < t < 1?

(b) Suppose c > 0. Find the distribution of a random variable X such that
50

Gn (c/n) X. Justify your answer.

(c) How does your answer to part (b) change if X1 , . . . , Xn are from a standard exponential distribution instead of a uniform distribution? The standard
exponential distribution function is F (t) = 1 et .
qm

Exercise 2.5 For each of the three examples in Exercise 2.1, does Xn 1? Justify
your answers.
Exercise 2.6 Prove Theorem 2.17(a).
Exercise 2.7 The converse of Theorem 2.17(b) is not true. Construct a counterexP
ample in which Xn 0 but E Xn = 1 for all n (by Theorem 2.17, if E Xn = 1,
then Xn cannot converge in quadratic mean to 0).
Hint: The mean of a random variable may be strongly influenced by a large
value that occurs with small probability (and if this probability goes to zero,
then the mean can be influenced in this way without destroying convergence in
probability).
Exercise 2.8 Prove or disprove this statement: If there exists M such that P (|Xn | <
qm
P
M ) = 1 for all n, then Xn c implies Xn c.
Exercise 2.9 (a) Prove that if 0 < a < b, then convergence in bth mean is stronger
a
b
than convergence in ath mean; i.e., Xn X implies Xn X.
Hint: Use Exercise 1.40 with = b/a.
(b) Prove by counterexample that the conclusion of part (a) is not true in general
if 0 < b < a.

2.2

Consistent Estimates of the Mean

For a sequence of random vectors X1 , X2 , . . ., we denote the nth sample mean by

n

Xn

1X
=
Xi .
n i=1

def

We begin with a formal statement of the weak law of large numbers for an independent
and identically distributed sequence. Later, we discuss some cases in which the sequence of
random vectors is not independent and identically distributed.
51

2.2.1

The Weak Law of Large Numbers

Theorem 2.19 Weak Law of Large Numbers (univariate version): Suppose that
X1 , X2 , . . . are independent and identically distributed and have finite mean .
P
Then X n .
The proof of Theorem 2.19 in its full generality is beyond the scope of this chapter, though it
may be proved using the tools in Section 4.1. However, by tightening the assumptions a bit,
a proof can be made simple. For example, if the Xi are assumed to have finite variance (not
a terribly restrictive assumption), the weak law may be proved in a single line: Chebyshevs
inequality (1.36) implies that

Var X n
Var X1
=
0,
P |X n | 
2

n2
P

so X n follows by definition. (This is an alternative proof of the same result in Example

2.18.)
P

Example 2.20 If X binomial(n, p), then X/n p. Although we could prove this
fact directly using the definition of convergence in probability, it follows immediately from the Weak Law of Large Numbers due to the fact that X/n is the
sample mean of n independent and identically distributed Bernoulli random variables, each with mean p.
Example 2.21 Suppose that X1 , X2 , . . . are independent and identically distributed
with mean and finite variance 2 . Then the estimator
n

n2

1X
(Xi )2
=
n i=1

is consistent for 2 because of the Weak Law of Large Numbers: The (Xi )2
are independent and identically distributed and they have mean 2 .
Ordinarily, of course, we do not know , so we replace by X n (and often replace
1
1
by n1
) to obtain the sample variance. We need a bit more theory to establish
n
the consistency of the sample variance, but we will revisit this topic later.

2.2.2

Independent but not Identically Distributed Variables

Let us now generalize the conditions of the previous section: Suppose that X1 , X2 , . . . are
independent but not necessarily identically distributed but have at least the same mean, so
52

that E Xi = and Var Xi = i2 . When is X n consistent for ? Since X n is unbiased, the

result of Example 2.18 tells us that we may conclude that X n is consistent as long as the
P
variance of X n tends to 0 as n . (However, X n does not imply Var X n 0; see
Exercise 2.10.) Since
Var X n =
P

we conclude that X n if

Pn

i=1

n
1 X 2
,
n2 i=1 i

(2.13)

i2 = o(n2 ).

What about alternatives to the sample mean? Suppose we restrict attention to weighted
mean estimators of the form
Pn
i=1 ci Xi

n = P
n
i=1 ci
for some sequence of positive constants c1 , c2 , . . .. The
n estimator above is unbiased, so we
consider whether its variance tends to zero. By independence, we may write
Pn 2 2
ci i
Var
n = Pi=1
2.
n
( i=1 ci )
How may
n ? To find the answer, we may set
Pwe obtain the smallest possible variance for
n with respect to 1 , . . . , n1 (after
i = ci / ni=1 ci and finding partial derivatives of Var
making the substitution n = 1 1 n1 ). Setting these partial derivatives equal to
zero gives the equations
i i2 = n n2

for 1 i n 1.

After checking to ensure that the solution is indeed a minimizer of the variance, we conclude that Var
n is minimized when each ci is proportional to 1/i2 . Thus, the variance is
minimized by
Pn
Xi /i2
n = Pi=1
,
(2.14)
n
2
j=1 1/j
which attains the variance
Var n = Pn

j=1

1/j2

(2.15)

An interesting fact about n Var X n and n Var n is that they are, respectively, the arithmetic
and harmonic means of 12 , . . . , n2 . In other words, our conclusion that Var n Var X n ,
53

with equality only when the i2 are all equal, is simply a restatement of a well-known mathematical inequality relating the harmonic and arithmetic means! See Exercise 2.11 for a
particularly compelling demonstration of the discrepancy between these two means.
Suppose for a sequence of independent random variables that instead of estimating their
mean , we wish to estimate their conditional mean, given a covariate. This is exactly the
case in regression:
Example 2.22 In the case of simple linear regression, let
Yi = 0 + 1 zi + i ,
where we assume the zi are known covariates and the i are independent and
identically distributed with mean 0 and finite variance 2 . (This implies that the
Yi are independent but not identically distributed.) If we define
(n)

wi

1
zi z n
(n)
(n)
=
and
v
z n wi ,
i
2
(z

z
)
n
n
j=1 j

= Pn

then the least squares estimators of 0 and 1 are

0n =

n
X

(n)
vi Yi and 1n =

i=1

n
X

(n)

w i Yi ,

(2.16)

i=1

respectively. One may prove, as in Exercise 2.14(a), that 0n and 1n are unbiased
estimators of 0 and 1 , so Example 2.18 tells us that they are consistent as long
as their variances tend to zero as n . It is therefore possible to show, as in
Exercise 2.14(b), that 0n is consistent if
z 2n
0
2
j=1 (zj z n )

(2.17)

1
0.
2
j=1 (zj z n )

(2.18)

Pn
and 1n is consistent if
Pn

2.2.3

Identically Distributed but not Independent Variables

Suppose that X1 , X2 , . . . have the same mean, say , but that they are not necessarily
independent. Since the unbiasedness of X n does not rely on independence, we still have
E X n = , so X n is consistent if its variance tends to zero. A direct calculation gives
n
n
1 XX
Cov (Xi , Xj ).
Var X n = 2
n i=1 j=1

54

(2.19)

Suppose we also assume that the Xi are identically distributed. In the presence of independence (the first i in i.i.d), the i.d. assumption is sufficient to specify the joint
distribution of the Xi . However, in the case of dependent random variables, assuming they
are identically distributed (with, say, variance 2 ) allows only a small simplification of Equation (2.19):
Var X n =

2 XX
2
+ 2
Cov (Xi , Xj ).
n
n
i<j

In order to deal with the n2 covariances above, all of which could in principle be distinct,
it would help to make some additional assumptions beyond identically distributed. One
possibility is to assume that the X1 , X2 , . . . are exchangeable:
Definition 2.23 Let denote an arbitrary permutation on n elements (that is, a
function that maps {1, . . . , n} onto itself). The finite sequence X1 , . . . , Xn is
said to be exchangeable if the joint distribution of the permuted random vector
(X(1) , . . . , X(n) ) is the same no matter which is chosen. The infinite sequence
X1 , X2 , . . . is said to be exchangeable if any finite subsequence is exchangeable.
Under exchangeability, the covariance between Xi and Xj is always the same, say Cov (X1 , X2 ),
when i 6= j. Therefore, Equation (2.19) reduces to
2 (n 1) Cov (X1 , X2 )
+
,
Var X n =
n
n
and we conclude that exchangeability implies Var X n Cov (X1 , X2 ) as n . Since this
is a nonzero limit unless the Xi are pairwise uncorrelated, exchangeability appears to be too
stringent a condition to place on the Xi in the context of searching for consistent estimators
of .
Thus, we turn to a weaker concept than exchangeability:
Definition 2.24 The sequence X1 , X2 , . . . is said to be stationary if, for a fixed k 0,
the joint distribution of (Xi , . . . , Xi+k ) is the same no matter what positive value
of i is chosen.
We see that i.i.d. implies exchangeability, which implies stationarity, which implies identically
distributed. To obtain an interesting simplification of Equation (2.19), it turns out that
stationarity is just about the right level in this hierarchy.
Under stationarity, Cov (Xi , Xj ) depends only on the gap j i. For example, stationarity
implies Cov (X1 , X4 ) = Cov (X2 , X5 ) = Cov (X5 , X8 ) = . Therefore, Equation (2.19)

55

becomes
Var X n

n1
2 X
2
+ 2
=
(n k) Cov (X1 , X1+k ).
n
n k=1

(2.20)

Lemma 2.25 Expression (2.20) tends to 0 as n if 2 < and Cov (X1 , X1+k )
0 as k .
Proof: It is immediate that 2 /n 0 if 2 < . Assuming that Cov (X1 , X1+k ) 0,
select  > 0 and note that if N is chosen so that | Cov (X1 , X1+k )| < /2 for all k > N , we
have


n1
N
n1
 2X
2 X
2 X


(n k) Cov (X1 , X1+k )
|Cov (X1 , X1+k )| +
|Cov (X1 , X1+k )| .
 2
 n
n
n k=N +1
k=1
k=1
The second term on the right is strictly less than /2, and the first term is a constant divided
by n, which may be made smaller than /2 by choosing n large enough. (This lemma is also
a corollary of the result stated in Exercise 1.3.)
Because of Lemma 2.25, it is sensible to impose conditions guaranteeing that Cov (X1 , X1+k )
tends to zero as k . For instance, we might consider sequences for which Cov (X1 , X1+k )
is exactly equal to zero for all k larger than some cutoff value, say m. This is the idea of
m-dependence:
Definition 2.26 For a fixed nonnegative integer m, the sequence X1 , X2 , . . . is called
m-dependent if the random vectors (X1 , . . . , Xi ) and (Xj , Xj+1 , . . .) are independent whenever j i > m.
Any stationary m-dependent sequence trivially satisfies Cov (X1 , X1+k ) 0 as k ,
so by Lemma 2.25, X n is consistent for any stationary m-dependent sequence with finite
variance. As a special case of m-dependence, any independent sequence is 0-dependent.

Exercises for Section 2.2

Exercise 2.10 The goal of this Exercise is to construct an example of an independent
P
sequence X1 , X2 , . . . with E Xi = such that X n but Var X n does not
converge to 0. There are numerous ways we could proceed, but let us suppose
that for some positive constants ci and pi , Xi = ci Yi (2Zi 1), where Yi and Zi
are independent Bernoulli random variables with E Yi = pi and E Zi = 1/2.
(a) Verify that E Xi = 0 and find Var X n .

56

(b) Show that X n 0 if

n

1X
ci pi 0.
n i=1

(2.21)

Hint: Use the triangle inequality to show that if Condition (2.21) is true, then
X n converges in mean to 0 (see Definition 2.15).
(c) Now specify ci and pi so that Var X n does not converge to 0 but Condition
(2.21) holds. Remember that pi must be less than or equal to 1 because it is the
mean of a Bernoulli random variable.
Exercise 2.11 Suppose that X1 , X2 , . . . are independent with mean zero and Var Xi =
(i + 1) log(i + 1). Let n be the minimum variance linear estimator defined in
Equation (2.14) and let X n denote the sample mean. Find the relative efficiency
of n with respect to X n (defined as Var X n / Var n ) for n = 10k , k = 1, . . . , 6.
What seems to be happening? Find, with proof, the limits of Var X n and Var n
as n to try to verify your conjecture.
Exercise 2.12 Suppose X1 , X2 , . . . are independentP
and identically distributed with
mean and finite variance 2 . Let Yi = X i = ( ij=1 Xj )/i.
P
(a) Prove that Y n = ( ni=1 Yi )/n is a consistent estimator of .
(b) Compute the relative efficiency eY n ,X n of Y n to X n , defined as Var (X n )/ Var (Y n ),
for n {5, 10, 20, 50, 100, } and report the results in a table. For n = , give
the limit (with proof) of the efficiency.
Exercise 2.13 Let Y1 , Y2 , . . . be independent and identically distributed with mean
and variance 2 < . Let
X 1 = Y1 ,

X2 =

Y2 + Y3
,
2

X3 =

Y4 + Y5 + Y6
,
3

etc.

Define n as in Equation (2.14).

(a) Show that n and X n are both consistent estimators of .
(b) Calculate the relative efficiency eX n ,n of X n to n , defined as Var (n )/ Var (X n ),
for n = 5, 10, 20, 50, 100, and and report the results in a table. For n = ,
give the limit (with proof) of the efficiency.
(c) Using Example 1.23, give a simple expression asymptotically equivalent
to eX n ,n . Report its values in your table for comparison. How good is the
approximation for small n?
57

Exercise 2.14 Consider the case of simple linear regression in Example 2.22.
(a) Prove that the least squares regression estimators defined in equation (2.16)
are unbiased. In other words, show that E 0n = 0 and E 1n = 1 .
P
P
(n)
(n)
Hint: Prove and use the facts that ni=1 wi = 0 and ni=1 wi zi = 1.
(b) Prove consistency of 0n and 1n under conditions (2.17) and (2.18), respectively.

2.3

Convergence of Transformed Sequences

Many statistical estimators of interest may be written as functions of simpler statistics whose
convergence properties are known. Therefore, results that describe the behavior of transformed sequences have central importance for the study of statistical large-sample theory.
We begin with some results about continuous transformations of univariate random variable
sequences. Yet the important result near the end of this section, called Slutskys theorem,
is intrinsically multivariate in nature. For this reason, after presenting a few results on continuous transformations, we will extend these and other univariate concepts from earlier in
this chapter to the k-dimensional setting for k > 1.

2.3.1

Continuous Transformations: The Univariate Case

Just as they do for sequences of real numbers, continuous functions preserve convergence
of sequences of random variables. We state this result formally for both convergence in
probability and convergence in distribution.
Theorem 2.27 Suppose that f (x) is a continuous function.
P

(a) If Xn X, then f (Xn ) f (X).

(b) If Xn X, then f (Xn ) f (X).
Theorem 2.27 is the random-variable analogue of Theorem 1.16, which is proved using a
straightforward - argument. It is therefore surprising that proving Theorem 2.27 is quite
difficult. Indeed, each of its two statements relies on an additional theorem for its proof.
For statement (a), this additional theorem (Theorem 3.10) involves almost sure convergence,
a mode of convergence not defined until Chapter 3. Statement (b) about convergence in
distribution, on the other hand, follows from a powerful characterization of convergence in
distribution (its proof is the subject of Exercises 2.15 and 2.16).
58

Theorem 2.28 Xn X if and only if E g(Xn ) E g(X) for all bounded and continuous real-valued functions g(x).
d

The forward half of Theorem 2.28, the fact that Xn X implies E g(Xn ) E g(X) if g(x)
is bounded and continuous, is called the Helly-Bray theorem. Taken as a whole, the theorem
establishes a condition equivalent to convergence in distribution, and in fact this condition is
d
sometimes used as the definition of Xn X. It is very important to remember that Theorem
d
2.28 does not say that Xn X implies E Xn E X (because the function g(t) = t, while
d
certainly continuous, is not bounded). In fact, the special conditions under which Xn X
implies E Xn E X are the subject of Section 3.3.
Using Theorem 2.28, Theorem 2.27(b) follows quickly.
Proof of Theorem 2.27(b) Let g(x) be any bounded and continuous function. By
Theorem 2.28, it suffices to show that E g[f (Xn )] E g[f (X)]. Since f (x) is continuous,
the composition x 7 g[f (x)] is bounded and continuous. Therefore, another use of Theorem
2.28 proves that E g[f (Xn )] E g[f (X)] as desired.

2.3.2

Multivariate Extensions

We now extend our notions of random-vector convergence to the multivariate case. Several
earlier results from this chapter, such as the weak law of large numbers and the results
on continuous functions, generalize immediately to this case. Note the use of bold type
to signify random vectors: Whereas Xn and X denote univariate random variables, the
possibly-multidimensional analogues are Xn and X.
A k-dimensional random vector is a function X(), usually abbreviated X, that maps a
probability space into k-dimensional Euclidean space Rk . As we remarked in Section
1.5, it is not possible to develop a coherent theory if we consider all possible functions
X() to be random vectors; therefore, strictly speaking we must restrict attention only to
measurable functions X(). Yet a reasonable treatment of measurability is beyond the scope
of this book, and instead of delving into technicalities we rest assured that basically every
interesting function X() is a legitimate random variable. (Indeed, it is a fairly challenging
mathematical exercise to construct a nonmeasurable function.)
The multivariate definitions of convergence in probability and convergence in ath mean
are both based on the sequence kXn Xk, which is a univariate sequence, so they are
straightforward and require no additional development:
P

Definition 2.29 Xn converges in probability to X (written Xn X) if for any  > 0,

P (kXn Xk < ) 1 as n .
59

Definition 2.30 For a > 0, Xn converges in ath mean to X (written Xn X) if

E kXn Xka 0 as n .
qm

Convergence in quadratic mean, written Xn X, is the special case of Definition 2.30 when
a = 2. Since kck2 = c> c, Xn converges in quadratic mean to X if


E (Xn X)> (Xn X) 0 as n .
(2.22)
Because Definitions 2.29 and 2.30 rely only on the univariate random variables kXn Xk,
qm
Theorem 2.17 immediately implies that (a) Xn c if and only if E kXn ck 0 and
a
P
Var kXn ck 0; and (b) if Xn X, then Xn X. However, fact (a) is less useful than
in the univariate setting; consider the comments following the proof of Theorem 2.31 below.
As an immediate application of the above results, we may extend the weak law of large
numbers to the multivariate case. For a sequence X1 , X2 , . . ., define the nth sample mean
to be
n

Xn

1X
=
Xi .
n i=1

def

Theorem 2.31 The Weak Law of Large Numbers: Suppose that X1 , X2 , . . . are inP
dependent and identically distributed and have finite mean . Then Xn .
Partial Proof: We do not prove this theorem in full generality until Section 4.1. However,
in the special (and very common) case in which the k k covariance matrix = Var X1
has only finite entries,
" n
#> " n
#
X
X
1
E (Xn )> (Xn ) =
E
(Xi )
(Xi )
n2
i=1
i=1
n
1
1 X
E (Xi )> (Xi ) = Tr() 0
=
2
n i=1
n
qm

as n . Therefore, Xn by definition, which implies that Xn .

It is instructive to compare the proof outlined in the univariate Example 2.18 to the multivariate proof above because the former method may not be adapted to the latter situation.
It is still true that any unbiased estimator whose covariance matrix converges to the zero
matrix is consistent [by Equation (2.22) with X replaced by E Xn ], but an argument for this
cannot easily be based on Theorem 2.17(a):
The fact
that an estimator like Xn is unbiased

for does not immediately imply that E Xn = 0.
60

To extend convergence in distribution to random vectors, we need the multivariate analogue

of Equation (2.6), the distribution function. To this end, let
def

F (x) = P (X x),
where X is a random vector in Rd and X x means that Xi xi for all 1 i d.
d

Definition 2.32 Xn converges in distribution to X (written Xn X) if for any point

c at which F (x) is continuous,
Fn (c) F (c) as n .
There is one subtle way in which the multivariate situation is not quite the same as the
univariate situation for distribution functions. In the univariate case, it is very easy to
characterize the points of continuity of F (x): The distribution function of the univariate
random variable X is continuous at x if and only if P (X = x) = 0. However, this simple
characterization no longer holds true for random vectors; a point x may be a point of
discontinuity yet still satisfy P (X = x) = 0. The task in Exercise 2.18 is to produce an
example of this phenomenon.
We may now extend Theorems 2.13, 2.14, and 2.27 to the multivariate case. The proofs
of these results do not differ substantially from their univariate counterparts; all necessary
modifications are straightforward. For instance, in the proofs of Theorems 2.13 and 2.14,
the scalar  should be replaced by the vector  = 1, each of whose entries equals ; with
this change, modified statements such as whenever Xn t, it must be true that either
X t +  or kXn Xk >  remain true.
P

Theorem 2.33 Xn X implies Xn X.

d
P
Xn c if and only if Xn c.

Furthermore, if c is a constant, then

Theorem 2.34 Suppose that f : S R` is a continuous function defined on some

subset S Rk , Xn is a k-component random vector, and P (X S) = 1.
P

(a) If Xn X, then f (Xn ) f (X).

(b) If Xn X, then f (Xn ) f (X).
The proof of Theorem 2.34 is basically the same as in the univariate case. For proving part
(b), we use the multivariate version of Theorem 2.28:
d

Theorem 2.35 Xn X if and only if E g(Xn ) E g(X) for all bounded and continuous real-valued functions g : Rk R.
61

Proving Theorem 2.35 involves a few complications related to the use of multivariate distribution functions, but the essential idea is the same as in the univariate case (the univariate
proof is the subject of Exercises 2.15 and 2.16).

2.3.3

Slutskys Theorem

As we have seen in the preceding few pages, many univariate definitions and results concerning convergence of sequences of random vectors are basically the same as in the univariate
case. Here, however, we consider a result that has no one-dimensional analogue.
At issue is the question of when we may stack random variables to make random vectors
while preserving convergence. It is here that we encounter perhaps the biggest surprise of
this section: Convergence in distribution is not preserved by stacking.
To understand what we mean by stacking preserving convergence, consider the case of
convergence in probability. By definition, it is straightforward to show that
   
Xn P X
P
P

.
(2.23)
Xn X and Yn Y together imply that
Y
Yn
Thus, two convergent-in-probability sequences Xn and Yn may be stacked, one on top of the
other, to make a vector, and this vector must still converge in probability to the vector of
stacked limits.
Example 2.36 Statement (2.23) gives a way to show that the multivariate Weak
Law of Large Numbers (Theorem 2.31) follows immediately from the univariate
version (Theorem 2.19).
Example 2.37 If X1 , X2 , . . . are independent and identically distributed with mean
and positive variance 2 , we often take as an estimator of 2 the so-called sample
variance
n

s2n

2
1 X
Xi X n .
=
n 1 i=1

We may use Statement (2.23) and Theorem 2.34(a), together with the univariate
Weak Law of Large Numbers (Theorem 2.19), to prove that s2n is a consistent
estimator of 2 .
To accomplish this, we first rewrite s2n as
" n
#
X
n
1
s2n =
(Xi )2 (X n )2 .
n 1 n i=1
62

The WLLN tells us that X n 0, from which we deduce that (X n )2 0

by Theorem 2.27(a). Since E (Xi )2 = 2 , the Weak Law also tells us that
P
P
(1/n) i (Xi )2 2 . Combining these two facts by stacking as in (2.23)
yields
 2
 1 Pn
2

P
i=1 (Xi )
n

.
2
0
(X n )
We now apply the continuous function f (a, b) = a b to both sides of the above
result, as allowed by Theorem 2.34(a), to conclude that
n

1X
P
(Xi )2 (X n )2 2 0.
n i=1

(2.24)

Finally, we may perform a similar stacking operation using Equation (2.24) toP
gether with the fact that n/(n 1) 1, whereupon multiplication yields the final
P
conclusion that s2n 2 .
If it seems as though we spent too much time in the above proof worrying about
obvious steps such as stacking followed by addition or multiplication, we did so
P
d
in order to make a point: When we replace by , the obvious is no longer
correct.
The converse of (2.23) is true by Theorem 2.27 because the function f (x, y) = x is a continuous function from R2 to R. By induction, we can therefore stack or unstack arbitrarily
many random variables or vectors without distrurbing convergence in probability. Combining this fact with Theorem 2.27 yields a useful result; see Exercise 2.22. By Definition 2.30,
P
a
Statement 2.23 remains true if we replace by throughout the statement for some a > 0.
P

However, Statement 2.23 is not true if is replaced by . Consider the following simple
counterexample.
Example 2.38 Take Xn and Yn to be independent standard normal random variables
for all n. These distributions do not depend on n at all, and it is correct to write
d
d
Xn Z and Yn Z, where Z N (0, 1). But it is certainly not true that
   
Xn d Z

,
(2.25)
Yn
Z
since the distribution on the left is bivariate normal with correlation 0, while the
distribution on the right is the (degenerate) bivariate normal with correlation 1.

63

Expression (2.25) is untrue precisely because the marginal distributions do not in general
uniquely determine the joint distribution. However, there are certain special cases in which
the marginals do determine the joint distribution. For instance, if random variables are
independent, then their marginal distributions uniquely determine their joint distribution.
Indeed, we can say that if Xn X and Yn Y , where Xn is independent of Yn and X is
P
d
independent of Y , then statement (2.23) remains true when is replaced by (see Exercise
2.23). As a special case, the constant c, when viewed as a random variable, is automatically
d
P
independent of any other random variable. Since Yn c is equivalent to Yn c by Theorem
2.14, it must be true that
   
Xn d X
d
P
Xn X and Yn c implies that

(2.26)
Yn
c
if Xn is independent of Yn for every n. The content of a powerful theorem called Slutskys
Theorem is that statement (2.26) remains true even if the Xn and Yn are not independent.
Although the preceding discussion involves stacking only random (univariate) variables, we
present Slutskys theorem in a more general version involving random vectors.
Theorem 2.39 Slutskys Theorem: For random vectors Xn , Yn , and X and a cond
P
stant c, if Xn X and Yn c as n , then
 
 
X
Xn
d

.
c
Yn
A proof of Theorem 2.39 is outlined in Exercise 2.24.
Putting several of the preceding results together yields the following corollary.
d

Corollary 2.40 If X is a k-vector such that Xn X, and Ynj cj for 1 j m,

then
 
 
Xn
X
d
f
f
Yn
c
for any continuous function f : S Rk+m R` .
It is very common practice in statistics to use Corollary 2.40 to obtain a result, then state
that the result follows by Slutskys Theorem. In fact, there is not a unanimously held
view in the statistical literature about what precisely Slutskys Theorem refers to; some
consider the Corollary itself, or particular cases of the Corollary, to be Slutskys Theorem.
These minor differences are unimportant; the common feature of all references to Slutskys
Theorem is some combination of one sequence that converges in distribution with one or
more sequences that converge in probability to constants.
64

Example 2.41 Asymptotic normality of the t-statistic: Let X1 , . . . , Xn be independent and identically distributed with mean and finite positive variance 2 . By

d
Example 2.12, n(X n ) N (0, 2 ). Suppose that
n2 is any consistent estiP
mator of 2 ; that is,
n2 2 . (For instance, we might take
n2 to be the usual
unbiased sample variance estimator s2n of Example 2.37, whose asymptotic properties will be studied later.) If Z denotes a standard normal random variable,
Theorem 2.39 implies

  
n(X n ) d Z

.
(2.27)

n2
2
Therefore, since f (a, b) = a/b is a continuous function for b > 0 (and 2 is
assumed positive),

n(X n ) d
p
Z.
(2.28)

n2
It is common practice to skip step (2.27), attributing equation (2.28) directly to
Slutskys Theorem.

Exercises for Section 2.3

d

Exercise 2.15 Here we prove half (the only if part) of Theorem 2.28: If Xn X
and g(x) is a bounded, continuous function on R, then E g(Xn ) E g(X). (This
half of Theorem 2.28 is sometimes called the univariate Helly-Bray Theorem.)
Let Fn (x) and F (x) denote the distribution functions of Xn and X, as usual. For
 > 0, take b < c to be constant real numbers such that F (b) <  and F (c) > 1.
First, we note that since g(x) is continuous, it must be uniformly continuous on
[b, c]: That is, for any  > 0 there exists > 0 such that |g(x)g(y)| <  whenever
|x y| < . This fact, along with the boundedness of g(x), ensures that there
exists a finite set of real numbers b = t0 < t1 < < tm = c such that:
Each ti is a continuity point of F (x).
F (t0 ) <  and F (tm ) > 1 .
For 1 i m, |g(x) g(ti )| <  for all x [ti1 , ti ].
(a) As in Figure 2.1, define
n
h(x) = g(ti ) if ti1 < x ti for some 1 i m.
0
otherwise.
65

M
t0 = b

t1 t2 t3

t4

t5

t6 = c

Figure 2.1: The solid curve is the function g(x), assumed to be bounded by M and M for
all x, and the horizontal segments are the function h(x). The points t0 , . . . , t6 are chosen so
that |g(x) h(x)| is always less than . The ti are continuity points of F (x), and both F (t0 )
and 1 F (t6 ) are less than .
Prove that there exists N such that | E h(Xn ) E h(X)| <  for all n > N .
Hint: Use the fact that for any random variable Y ,
E h(Y ) =

m
X

g(ti )P (ti1 < Y ti ).

i=1

Also, please note that we may not write E h(Xn ) E h(X) as E [h(Xn ) h(X)]
because it is not necessarily the case that Xn and X are defined on the same
sample space.
(b) Prove that E g(Xn ) E g(X).
Hint: Use the fact that
| E g(Xn ) E g(X)| | E g(Xn ) E h(Xn )| + | E h(Xn ) E h(X)|
+| E h(X) E g(X)|.
Exercise 2.16 Prove the other half (the if part) of Theorem 2.28, which states
that if E g(Xn ) E g(X) for all bounded, continuous functions g : R R, then
d
Xn X.
66

(a) Let t be any continuity point of F (x). Let  > 0 be arbitrary. Show that
there exists > 0 such that F (t ) > F (t)  and F (t + ) < F (t) + .
(b) Show how to define continuous functions g1 : R [0, 1] and g2 : R [0, 1]
such that for all x t, g1 (x) = g2 (x ) = 1 and for all x > t, g1 (x + ) =
g2 (x) = 0. Use these functions to bound the difference between Fn (t) and F (t)
in such a way that this difference must tend to 0.
Exercise 2.17 To illustrate a situation that can arise in the multivariate setting
that cannot arise in the univariate setting, construct an example of a sequence
(Xn , Yn ), a joint distribution (X, Y ), and a connected subset S R2 such that
d

(i) (Xn , Yn ) (X, Y );

(ii) every point of R2 is a continuity point of the distribution function of (X, Y );
(iii) P [(Xn , Yn ) S] does not converge to P [(X, Y ) S].
Hint: Condition (ii) may be satisfied even if the distribution of (X, Y ) is concentrated on a line.
Exercise 2.18 If X is a univariate random variable with distribution function F (x),
then F (x) is continuous at c if and only if P (X = c) = 0. Prove by counterexample that this is not true if variables X and c are replaced by vectors X and
c.
Exercise 2.19 Suppose that (X, Y ) is a bivariate normal vector such that both X and
Y are marginally standard normal and Corr (X, Y ) = . Construct a computer
program that simulates the distribution function F (x, y) of the joint distribution
of X and Y . For a given (x, y), the program should generate at least 50,000
random realizations from the distribution of (X, Y ), then report the proportion
for which (X, Y ) (x, y). (If you wish, you can also report a confidence interval
for the true value.) Use your function to approximate F.5 (1, 1), F.25 (1, 1), and
F.75 (0, 0). As a check of your program, you can try it on F0 (x, y), whose true
values are not hard to calculate directly for an arbitrary x and y assuming your
software has the ability to evaluate the standard normal distribution function.
Hint: To
1
matrix

X = U and

generate
a bivariate normal random vector (X, Y ) with covariance


, start with independent standard normal U and V , then take

1
p
Y = U + 1 2 V .

Exercise 2.20 Adapt the method of proof in Exercise 2.15 to the multivariate case,
d
proving half of Theorem 2.35: If Xn X, then E g(Xn ) E g(X) for any
67

bounded, continuous g : Rk R.
Hint: Instead of intervals (ai1 , ai ] as in Exercise 2.15, use small regions {x :
ai,j1 < xi ai,j for all i} of Rk . Make sure these regions are chosen so that their
boundaries contain only continuity points of F (x).
Exercise 2.21 Construct a counterexample to show that Slutskys Theorem 2.39 may
P
P
not be strengthened by changing Yn c to Yn Y .
P

Exercise 2.22 (a) Prove that if f : Rk R` is continuous and Xnj Xj for all
P
1 j k, then f (Xn ) f (X).
(b) Taking f (a, b) = a + b for simplicity, construct an example demonstrating
P
d
that part (a) is not true if is replaced by .
Exercise 2.23 Prove that if Xn is independent of Yn for all n and X is independent
of Y , then
   
Xn d X
d
d

.
Xn X and Yn Y implies that
Y
Yn
Hint: Be careful to deal with points of discontinuity: If Xn and Yn are independent, what characterizes a point of discontinuity of the joint distribution?
Exercise 2.24 Prove Slutskys Theorem, Theorem 2.39, using the following approach:
(a) Prove the following lemma:
Lemma 2.42 Let Vn and Wn be k-dimensional random vectors on the
same sample space.
d

If Vn V and Wn 0, then Vn + Wn V.
Hint: For  > 0, let  denote the k-vector all of whose entries are . Take
a Rk to be a continuity point of FV (v). Now argue that a, since it is a point
of continuity, must be contained in a neighborhood consisting only of points of
continuity; therefore,  may be taken small enough so that a  and a +  are
also points of continuity. Prove that
P (Vn a ) P (kWn k ) P (Vn + Wn a)
P (Vn a + ) + P (kWn k ).
Next, take lim supn and lim inf n . Finally, let  0.
68

(b) Show how to prove Theorem 2.39 using Lemma 2.42.

Hint: Consider the random vectors
 


Xn
0
Vn =
and Wn =
.
c
Yn c

69

Chapter 3
Strong convergence
There are multiple ways to define the convergence of a sequence of random variables. Chapter
2 introduced convergence in probability, convergence in distribution, and convergence in
quadratic mean. We now consider a fourth mode of convergence, almost sure convergence
or convergence with probability one. We will see that almost sure convergence implies both
convergence in probability and convergence in distribution, which is why we sometimes use
the term strong for almost sure convergence and weak for the other two.
The terms weak and strong do not indicate anything about their importance; indeed,
the weak modes of convergence are used much more frequently in asymptotic statistics
than the strong mode. Because weak convergence dominates the remainder of this book
beginning with Chapter 4, a reader may safely skip much of the material in the current
chapter if time is limited; however, the quantile function and the Dominated Convergence
Theorem of Section 3.3 are used elsewhere, and at least these topics should be reviewed
before moving on. Due to the technical nature of the material of this chapter, the exercises
are almost exclusively devoted to proofs.

3.1

Strong Consistency Defined

A random variable like Xn or X is a function on a sample space, say . Suppose that we

fix a particular element of that space, say 0 , so we obtain the real numbers Xn (0 ) and
X(0 ). If Xn (0 ) X(0 ) as n in the sense of Definition 1.1, then 0 is contained in
the event
S = { : Xn () X()}.

70

(3.1)

If the probability of Sthat is, E I{Xn X}equals 1, then we say that Xn converges
almost surely to X:
Definition 3.1 Suppose X and X1 , X2 , . . . are random variables defined on the same
sample space (and as usual P denotes the associated probability measure). If
P ({ : Xn () X()}) = 1,
then Xn is said to converge almost surely (or with probability one) to X, denoted
a.s.
Xn X or Xn X a.s. or Xn X w.p. 1.
In other words, convergence with probability one means exactly what it sounds like: The
probability that Xn converges to X equals one. Later, in Theorem 3.3, we will formulate
an equivalent definition of almost sure convergence that makes it much easier to see why
it is such a strong form of convergence of random variables. Yet the intuitive simplicity of
Definition 3.1 makes it the standard definition.
As in the case of convergence in probability, we may replace the limiting random variable
a.s.
X by any constant c, in which case we write Xn c. In the most common statistical usage
of convergence to a constant, the random variable Xn is some estimator of a particular
parameter, say :
a.s.

Definition 3.2 If Xn , Xn is said to be strongly consistent for .

As the names suggest, strong consistency implies consistency (also known as weak consistency), a fact we now explore in more depth.

3.1.1

Strong Consistency versus Consistency

As before, suppose that X and X1 , X2 , . . . are random variables defined on the same sample
space, . For given n and  > 0, define the events
An = { : |Xk () X()| <  for all k n}

(3.2)

Bn = { : |Xn () X()| < }.

(3.3)

and

First, note that An must be contained in Bn and that both An and Bn imply that Xn is close
to X as long as  is small. Therefore, both P (An ) 1 and P (Bn ) 1 seem like reasonable
ways to define the convergence of Xn to X. Indeed, as we have already seen in Definition
2.1, convergence in probability means precisely that P (Bn ) 1 for any  > 0.
Yet what about the sets An ? One fact is immediate: Since An Bn , we must have
P (An ) P (Bn ). Therefore, P (An ) 1 implies P (Bn ) 1. In other words, if we were to
71

take P (An ) 1 for all  > 0 to be the definition of a new form of convergence of random
sequences, then this form of convergence would be stronger than (i.e., it would imply) convergence in probability. By now, the reader may already have guessed that this new form of
convergence is actually equivalent to almost sure convergence:
Theorem 3.3 With An defined as in Equation (3.2), P (An ) 1 for any  > 0 if and
a.s.
only if Xn X.
Proving Theorem 3.3 is the subject of Exercise 3.1. The following corollary now follows from
the preceding discussion:
a.s.

Corollary 3.4 If Xn X, then Xn X.

The converse of Corollary 3.4 is not true, as the following example illustrates.

1
J1

J2

J5
J10

J3

J11

J6

J7

J12

J13

J4

J8

J14

J9

J15

J16

J17 J18 J19 J20 J21 J22 J23 J24 J25

...and so on

Figure 3.1: Example 3.5, in which P (Jn ) 0 as n , which means that I{Jn } 0.
However, the intervals Jn repeatedly cover the entire interval (0, 1], so the subset of (0, 1] on
which I{Jn } converges to 0 is empty!
Example 3.5 Take to be the half-open interval (0, 1], and for any interval J ,
say J = (a, b], take P (J) = b a to be the length of that interval. Define a
sequence of intervals J1 , J2 , . . . as follows (see Figure 3.1):
J1 = (0, 1]
J2 through J4 = (0, 13 ] , ( 13 , 23 ] , ( 32 ,1]
72

J5 through J9 = (0, 51 ] , ( 15 , 25 ] , ( 52 , 35 ] , ( 35 , 45 ] , ( 54 ,1]

..
. 



2m
1
,...,
,1
Jm2 +1 through J(m+1)2 =
0,
2m + 1
2m + 1
..
.

Note in particular that P

(Jn ) = 1/(2m + 1), where m = b n 1c is the largest
integer not greater than n 1. Now, define Xn = I{Jn } and take 0 <  < 1.
Then P (|Xn 0| < ) is the same as 1 P (Jn ). Since P (Jn ) 0, we conclude
P
Xn 0 by definition.
a.s.

However, it is not true that Xn 0. Since every is contained in infinitely

many Jn , the set An defined in Equation (3.2) is empty for all n. Alternatively,
consider the set S = { : Xn () 0}. For any , Xn () has no limit because
Xn () = 1 and Xn () = 0 both occur for infinitely many n. Thus S is empty.
This is not convergence with probability one; it is convergence with probability
zero!

3.1.2

Multivariate Extensions

We may extend Definition 3.1 to the multivariate case in a completely straightforward way:
Definition 3.6 Xn is said to converge almost surely (or with probability one) to X
a.s.
(Xn X) if
P (Xn X as n ) = 1.
Alternatively, since the proof of Theorem 3.3 applies to random vectors as well
a.s.
as random variables, we say Xn X if for any  > 0,
P (kXk Xk <  for all k n) 1 as n .

(3.4)

We saw in Theorems 2.27 and 2.34 that continuous functions preserve both convergence in
probability and convergence in distribution. Yet these facts were quite difficult to prove.
Fortunately, the analogous result for almost sure convergence follows immediately from the
results of Chapter 1. Similarly, unlike with convergence in distribution, there is no problem
stacking random sequences into vectors while preserving almost sure convergence. The
following theorem is really just a corollary of earlier results (specifically, Theorem 1.16 and
Lemma 1.33).

73

Theorem 3.7 (a) Suppose that f : S R` is a continuous function defined on some

subset S Rk , Xn is a k-component random vector, and the range of X and of
a.s.
a.s.
each Xn is contained in S with probability 1. If Xn X, then f (Xn ) f (X).
a.s.

a.s.

(b) Xn X if and only if Xnj Xj for all j.

We conclude this section with a simple diagram summarizing the implications among the
modes of convergence defined so far. In the diagram, a double arrow like means implies.
Note that the picture changes slightly when convergence is to a constant c rather than a
random vector X.
qm

qm

Xn X

a.s.
P
d
Xn X X n X Xn X

Xn c

a.s.
P
d
Xn c Xn c Xn c

Exercises for Section 3.1

a.s.

Exercise 3.1 Let S be the set defined in equation (3.1), so Xn X is equivalent to

P (S) = 1 by definition.
(a) Let An be defined as in Equation (3.2). Prove that
0
n=1 An for all  > 0
if and only if 0 S.
Hint: Use Definition 1.1.
(b) Prove Theorem 3.3.
Hint: Note that the sets An are increasing in n, so that by the lower continuity
of any probability measure (which you may assume without proof), limn P (An )
exists and is equal to P (
n=1 An ).
a.s.

Exercise 3.2 The diagram at the end of this section suggests that neither Xn X
qm
nor Xn X implies the other. Construct two counterexamples, one to show that
qm
qm
a.s.
Xn X does not imply Xn X and the other to show that Xn X does not
a.s.
imply Xn X.

3.2

The Strong Law of Large Numbers

Some of the results in this section are presented for univariate random variables and some are
presented for random vectors. Take note of the use of bold print to denote vectors. Nearly
74

all of the technical proofs are posed as exercises (with hints, of course).
Theorem 3.8 Strong Law of Large Numbers: Suppose that X1 , X2 , . . . are indepena.s.
dent and identically distributed and have finite mean . Then Xn .
It is possible to use fairly simple arguments to prove a version of the Strong Law under more
restrictive assumptions than those given above. See Exercise 3.4 for details of a proof of the
univariate Strong Law under the additional assumption that E Xn4 < . To aid the proof of
the Strong Law, with or without such an additional assumption, we first establish a useful
lemma.
P
a.s.
Lemma 3.9 If
k=1 P (kXk Xk > ) < for any  > 0, then Xn X.
Proof: The proof relies on the countable subadditivity of any probability measure, an axiom
stating that for any sequence C1 , C2 , . . . of events,
!

[
X
P
Ck
P (Ck ).
(3.5)
k=1

k=1

To prove the lemma using (3.4), we must demonstrate that P (kXk Xk  for all k n)
1 as n , which (taking complements) is equivalent to P (kXk Xk >  for some k n)
0. Letting Ck denote the event that kXk Xk > , countable subadditivity implies
!

X
[
P (Ck ) ,
P (Ck for some k n) = P
Ck
k=n

k=n

and the right hand side tends to 0 as n because it is the tail of a convergent series.
Lemma 3.9 is nearly the same as a famous result called the First Borel-Cantelli Lemma,
or sometimes simply the Borel-Cantelli Lemma; see Exercise 3.3. Lemma 3.9 is extremely
useful for establishing almost sure convergence of sequences. As an illustration of the type of
result this lemma helps to prove, consider the following theorem (see Exercise 3.8 for hints
on how to prove it).
P

Theorem 3.10 Xn X if and only if each subsequence Xn1 , Xn2 , . . . contains a further subsequence that converges almost surely to X.
Using Theorem 3.10, it is nowfinallypossible to prove that continuous transformations
preserve convergence in probability. This fact was stated in Theorem 2.27(a) (for the univariate case) and Theorem 2.34(a) (for the multivariate case). It suffices to complete the
proof for the multivariate case.
P

Proof of Theorem 2.34(a): If Xn X and f (x) is continuous, it suffices to prove

that each subsequence f (Xn1 ), f (Xn2 ), . . . contains a further subsequence that converges
75

almost surely to f (X). But we know that Xn1 , Xn2 , . . . must contain a further subsequence,
a.s.
a.s.
say Xni(1) , Xni(2) , . . ., such that Xni(j) X and therefore f (Xni(j) ) f (X) as j by
Theorem 3.7(a). This proves the theorem!
To conclude this section, we provide a proof of the Strong Law (Theorem 3.8). The approach
we give here is based on a powerful theorem of Kolmogorov:
Theorem 3.11 Kolmogorovs Strong Law of Large Numbers: Suppose that X1 , X2 , . . .
are independent with mean and

X
Var Xi
< .
i2
i=1
a.s.

Then X n .
Note that there is no reason the Xi in Theorem 3.11 P
must have the same means: If E Xi = i ,
a.s.
then the conclusion of the theorem becomes (1/n) i (Xi i ) 0. Theorem 3.11 may be
proved using Kolmogorovs inequality from Exercise 1.41; this proof is the focus of Exercise
3.7.
The key to completing the Strong Law of Large Numbers for an independent and identically
distributed sequence using Theorem 3.11 is to introduce truncated versions of X1 , X2 , . . . as
in the following lemmas, which are proved in Exercises 3.5 and 3.6.
Lemma 3.12 Suppose that X1 , X2 , . . . are independent and identically distributed
and have finite mean . Define Xi = Xi I{|Xi | i}. Then

X
Var Xi
< .
(3.6)
i2
i=1

Lemma 3.13 Under the assumptions of Lemma 3.12, let X n = (1/n)

a.s.
X n X n 0.

Pn

i=1

Xi . Then

Finally, it is possible to put the preceding results together to prove the Strong Law of Large
Numbers:
Proof of Theorem 3.8: Let X1 , X2 , . . . be independent and identically distributed with
finite mean , and let Xi = Xi I{|Xi | i}. Then Lemma 3.12 and Theorem 3.11 together
a.s.
a.s.
imply that X n . From Lemma 3.13, we obtain X n X n 0. Adding these two limit
statements (which is legal because of Theorem 3.7), we obtain

a.s.

X n = X n + (X n X n ) ,
which establishes the Strong Law for the univariate case. Since stacking sequences presents
no problems for almost sure convergence [Theorem 3.7(b)], the multivariate version follows
immediately.
76

Exercises for Section 3.2

Exercise 3.3 Let B1 , B2 , . . . denote a sequence of events. Let Bn i.o., which stands
for Bn infinitely often, denote the set
def

Bn i.o. = { : for every n, there exists k n such that Bk }.

P
Prove the First Borel-Cantelli Lemma, which states that if
n=1 P (Bn ) < ,
then P (Bn i.o.) = 0.
Hint: Argue that
Bn i.o. =

Bk ,

n=1 k=n

then adapt the proof of Lemma 3.9.

Exercise 3.4 Use the steps below to prove a version of the Strong Law of Large
Numbers for the special case in which the random variables X1 , X2 , . . . have a
finite fourth moment, E X14 < .
(a) Assume without loss of generality that E X1 = 0. Expand E (X1 + . . . + Xn )4
and then count the nonzero terms.
Hint: The only nonzero terms are of the form E Xi4 or (E Xi2 )2 .
(b) Use Markovs inequality (1.35) with r = 4 to put an upper bound on
 


P X n  > 
involving E (X1 + . . . + Xn )4 .
a.s.

(c) Combine parts (a) and (b) with Lemma 3.9 to show that X n 0.
P
2
Hint: Use the fact that
< .
n=1 n
Exercise 3.5 Lemmas 3.12 and 3.13 make two assertions about the random variables Xi = Xi I{|Xi | i}, where X1 , X2 , . . . are independent and identically
distributed with finite mean .
(a) Prove that for an arbitrary real number c,

X
1
c
I{|c| i} 1 + |c|.
i2
i=1
2

Hint: Bound the sum on the left hand side by an easy-to-evaluate integral.
77

(b) Prove Lemma 3.12, which states that

X
Var Xi
< .
i2
i=1

Hint: Use the fact that

Var Xi E (Xi )2 = E X12 I{|X1 | i}
together with part (a) and the fact that E |X1 | < .
Exercise 3.6 Assume the conditions of Exercise 3.5.
a.s.

(a) Prove that Xn Xn 0.

Hint: Note that Xn and Xn do not have bars here. Use Exercise 1.45 together
with Lemma 3.9.
(b) Prove Lemma 3.13, which states that
a.s.

X n X n 0.
Hint: Use Exercise 1.3.
Exercise 3.7 Prove Theorem 3.11. Use the following steps:
(a) For k = 1, 2, . . ., define
Yk =

max

2k1 n<2k

|X n |.

Use the Kolmogorov inequality from Exercise 1.41 to show that

P (Yk )

P2k

i=1 Var
4k 2

a.s.

Xi

.
a.s.

(b) Use Lemma 3.9 to show that Yk 0, then argue that this proves X n .
Hint: Letting dlog2 ie denote the smallest integer greater than or equal to log2 i
(the base-2 logarithm of i), verify and use the fact that

X
k=dlog2 ie

1
4
2.
k
4
3i
78

Exercise 3.8 Prove Theorem 3.10:

(a) To simplify notation, let Y1 = Xn1 , Y2 = Xn2 , . . . denote an arbitrary
subseqence of X1 , X2 , . . ..
P

Prove the only if part of Theorem 3.10, which states that if Xn X, then there
a.s.
exists a subsequence Ym1 , Ym2 , . . . such that Ymj X as j .
Hint: Show that there exist m1 , m2 , . . . such that
P (kYmj Xk > ) <

1
,
2j

then use Lemma 3.9.

(b) Now prove the if part of the theorem by arguing that if Xn does not
converge in probability to X, there exists a subsequence Y1 = Xn1 , Y2 = Xn2 , . . .
and  > 0 such that
P (kYk Xk > ) > 
for all k. Then use Corollary 3.4 to argue that Y1 , Y2 , . . . does not have a
subsequence that converges almost surely.

3.3

The Dominated Convergence Theorem

d

In this section, the key question is this: When does Xn X imply E Xn E X? The
answer to this question impacts numerous results in statistical large-sample theory. Yet
because the question involves only convergence in distribution, it may seem odd that it is
being asked here, in the chapter on almost sure convergence. We will see that one of the
d
most useful conditions under which E Xn E X follows from Xn X, the Dominated
Convergence Theorem, is proved using almost sure convergence.

3.3.1

Moments Do Not Always Converge

d

It is easy to construct cases in which Xn X does not imply E Xn E X, and perhaps

d
the easiest way to construct such examples is to recall that if X is a constant, then X and
P
P
X are equivalent: For Xn c means only that Xn is close to c with probability approaching
one. What happens to Xn when it is not close to c can have an arbitrarily extreme influence
P
on the mean of Xn . In other words, Xn c certainly does not imply that E Xn c, as the
next two examples show.
79

Example 3.14 Suppose that U is a standard uniform random variable. Define


n with probability 1/n
Xn = nI {U < 1/n} =
(3.7)
0 with probability 1 1/n.
P

We see immediately that Xn 0 as n , but the mean of Xn , which is 1 for

all n, is certainly not converging to zero!. Furthermore, we may generalize this
example by defining

cn with probability pn
Xn = cn I {U < pn } =
0 with probability 1 pn .
P

In this case, a sufficient condition for Xn 0 is pn 0. But the mean E Xn =

cn pn may be specified arbitrarily by an appropriate choice of cn , no matter what
nonzero value pn takes.
Example 3.15 Let Xn be a contaminated standard normal distribution with mixture
distribution function


1
1
(3.8)
(x) + Gn (x),
Fn (x) = 1
n
n
where (x) denotes the standard normal distribution function. No matter how
d
the distribution functions Gn are defined, Xn . However, letting n denote
the mean of Gn , E Xn = n /n may be set arbitrarily by an appropriate choice of
Gn .
Consider Example 3.14, specifically Equation (3.7), once again. Note that each Xn in that
example can take only two values, 0 or n. In particular, each Xn is bounded, as is any
random variable with finite support. Yet taken as a sequence, the Xn are not uniformly
boundedthat is, there is no single constant that bounds all of the Xn simultaneously.
On the other hand, suppose that a sequence X1 , X2 , . . . does have some uniform bound, say,
M such that |Xn | M for all n. Then define the following function:
(
M
if t > M
g(t) = t
if |t| M
M if t < M .
d

Note that g(t) is bounded and continuous. Therefore, Xn X implies that E g(Xn )
E g(X) by the Helly-Bray Theorem (see Theorem 2.28). Because of the uniform bound on
the Xn , we know that g(Xn ) is always equal to Xn . We conclude that E Xn E g(X), and
furthermore it is not difficult to show that |X| must be bounded by M with probability 1, so
E g(X) = E X. We may summarize this argument by the following Corollary of Theorem
2.28:
80

Corollary 3.16 If X1 , X2 , . . . are uniformly bounded (i.e., there exists M such that
d
|Xn | < M for all n) and Xn X, then E Xn E X.
Corollary 3.16 gives a vague sense of the type of conditiona uniform bound on the Xn
d
sufficient to ensure that Xn X implies E Xn E X. However, the corollary is not very
broadly applicable, since many common random variables are not bounded. In Example
3.15, for instance, the Xn could not possibly have a uniform bound (no matter how the Gn
are defined) because the standard normal component of each random vector is supported
on all of R and is therefore not bounded. It is thus desirable to generalize the idea of
a uniform bound of a sequence. There are multiple ways to do this, but probably the
best-known generalization is the Dominated Convergence Theorem introduced later in this
section. To prove this important theorem, we must first introduce quantile functions and
another theorem called the Skorohod Representation Theorem.

3.3.2

Quantile Functions and the Skorohod Representation Theorem

Roughly speaking, the q quantile of a variable X, for some q (0, 1), is a value q such
that P (X q ) = q. Therefore, if F (x) denotes the distribution function of X, we ought to
define q = F 1 (q). However, not all distribution functions F (x) have well-defined inverse
functions F 1 (q). To understand why not, consider Example 3.17.
Example 3.17 Suppose that U Uniform(0, 1) and V Binomial(2, 0.5) are independent random variables. Let X = V /4 + U V 2 /8. The properties of X are most
easily understood by noticing that X is either a constant 0, uniform on (1/4, 3/8),
or uniform on (1/2, 1), conditional on V = 0, V = 1, or V = 2, respectively. The
distribution function of X, F (x), is shown in Figure 3.2.
There are two problems that can arise when trying to define F 1 (q) for an arbitrary q (0, 1) and an arbitrary distribution function F (x), and the current
example suffers from both: First, in the range q (0, 1/4), there is no x for which
F (x) equals q because F (x) jumps from 0 to 1/4 at x = 0. Second, for q = 1/4
or q = 3/4, there is not a unique x for which F (x) equals q because F (x) is flat
(constant) at 1/4 and again at 3/4 for whole intervals of x values.
From Example 3.17, we see that a meaningful general inverse of a distribution function must
deal both with jumps and flat spots. The following definition does this.
Definition 3.18 If F (x) is a distribution function, then we define the quantile func-

81

1.0
0.5

0.0

0.0

0.5

1.0

Figure 3.2: The solid line is a distribution function F (x) that does not have a uniquely
defined inverse F 1 (q) for all q (0, 1). However, the dotted quantile function F (q) of
Definition 3.18 is well-defined. Note that F (q) is the reflection of F (x) over the line q = x,
so jumps in F (x) correspond to flat spots in F (q) and vice versa.
tion F : (0, 1) R by
def

F (q) = inf{x R : q F (x)}.

(3.9)

With the quantile function thus defined, we may prove a useful lemma:
Lemma 3.19 q F (x) if and only if F (q) x.
Proof: Using the facts that F () is nondecreasing and F [F (x)] x by definition,
q F (x) implies F (q) F [F (x)] x,
which proves the only if statement. Conversely, assume that F (q) x. Since F () is
nondecreasing, we may apply it to both sides of the inequality to obtain
F [F (q)] F (x).
Thus, q F (x) follows if we can prove that q F [F (q)]. To this end, consider the
set {x R : q F (x)} in Equation (3.9). Because any distribution function is rightcontinuous, this set always contains its infimum, which is F (q) by definition. This proves
that q F [F (q)].
82

Corollary 3.20 If X is a random variable with distribution function F (x) and U is

a uniform(0, 1) random variable, then X and F (U ) have the same distribution.
Proof: For any x R, Lemma 3.19 implies that
P [F (U ) x] = P [U F (x)] = F (x).

Assume now that X1 , X2 , . . . is a sequence of random variables converging in distribution to

X, and let Fn (x) and F (x) denote the distribution functions of Xn and X, respectively. We
a.s.
will show how to construct a new sequence Y1 , Y2 , . . . such that Yn Y and Yn Fn () and
Y F ().
Take = (0, 1) to be a sample space and adopt the probability measure that assigns to each
interval subset (a, b) its length (b a). [There is a unique probability measure on (0, 1)
with this property, a fact we do not prove here.] Then for every , define
def

def

Yn () = Fn () and Y () = F ().

(3.10)

Note that the random variable defined by U () = is a uniform(0, 1) random variable,

so Corollary 3.20 demonstrates that Yn Fn () and Y F (). It remains to prove that
a.s.
Yn Y , but once this is proven we will have established the following theorem:
Theorem 3.21 Skorohod Representation Theorem: Assume F, F1 , F2 , . . . are distrid
bution functions and Fn F . Then there exist random variables Y, Y1 , Y2 , . . .
such that
1. P (Yn t) = Fn (t) for all n and P (Y t) = F (t);
a.s.

2. Yn Y .
A completion of the proof of Theorem 3.21 is the subject of Exercise 3.10.
Having thus established the Skorohod Represenation Theorem, we now introduce the Dominated Convergence Theorem.
Theorem 3.22 Dominated Convergence Theorem: If for a nonnegative random varid
able Z, |Xn | Z for all n and E Z < , then Xn X implies that E Xn E X.
Proof: Use the Skorohod Representation Theorem to construct a sequence Yn converging to
d
d
Y almost surely such that Yn =Xn for all n and Y =X. Furthermore, construct a nonnegative
d
random variable Z on the same sample space satisfying Z =Z and |Yn | Z ; this is
possible by defining Z = supn |Yn | + W , where W is constructed to have the distribution
83

a.s.

of Z supn |Xn | using the idea of expression (3.10). In other words, we now have Yn Y ,
|Yn | Z , and E Z < . Since E Xn = E Yn for all n and E X = E Y , it suffices to prove
now that E Yn E Y .
Fatous Lemma (see Exercise 3.11) states that
E lim inf |Yn | lim inf E |Yn |.
n

(3.11)

A second application of Fatous Lemma to the nonnegative random variables Z |Yn | implies
E Z E lim sup |Yn | E Z lim sup E |Yn |.
n

Because E Z < , subtracting E Z preserves the inequality, so we obtain

lim sup E |Yn | E lim sup |Yn |.
n

(3.12)

Together, inequalities (3.11) and (3.12) imply

E lim inf |Yn | lim inf E |Yn | lim sup E |Yn | E lim sup |Yn |.
n

a.s.

Since Yn Y , both lim inf n |Yn | and lim supn |Yn | are equal to |Y | with probability one, so
we conclude that lim E |Yn | exists and is equal to E |Y |.
The Dominated Convergence Theorem essentially tells us when it is possible to interchange
the operations of limit and expectation, that is, when the limit of the expectations (of the
Xn ) equals the expectation of their limit.

Exercises for Section 3.3

Exercise 3.9 Prove that any nondecreasing function must have countably many points
of discontinuity. (This fact is used in proving the Skorohod Representation Theorem.)
Hint: Use the fact that the set of rational numbers is a countably infinite set
and that any real interval must contain a rational number.
Exercise 3.10 To complete the proof of Theorem 3.21, it only remains to show that
a.s.
Yn Y , where Yn and Y are defined as in Equation (3.10).
(a) Let > 0 and (0, 1) be arbitrary. Show that there exists N1 such that
Y () < Yn ()
84

(3.13)

for all n > N1 .

Hint: There exists a point of continuity of F (x), say x0 , such that
Y () < x0 < Y ().
Use the fact that Fn (x0 ) F (x0 ) together with Lemma 3.19 to show how this
fact leads to the desired conclusion.
(b) Take and as in part (a) and let  > 0 be arbitrary. Show that there
exists N2 such that
Yn () < Y ( + ) +

(3.14)

for all n > N2 .

Hint: There exists a point of continuity of F (x), say x1 , such that
Y ( + ) < x1 < Y ( + ) + .
Use the fact that Fn (x1 ) F (x1 ) together with Lemma 3.19 to show how this
fact leads to the desired conclusion.
(c) Suppose that (0, 1) is a continuity point of F (x). Prove that Yn ()
Y ().
Hint: Take lim inf n in Inequality (3.13) and lim supn in Inequality (3.14). Put
these inequalities together, then let 0. Finally, let  0.
a.s.

(d) Use Exercise 3.9 to prove that Yn Y .

Hint: Use countable subadditivity, Inequality (3.5), to show that the set of
discontinuity points of F (x) has probability zero.
Exercise 3.11 Prove Fatous lemma:
E lim inf |Xn | lim inf E |Xn |.
n

(3.15)

Hint: Argue that E |Xn | E inf kn |Xk |, then take the limit inferior of each
side. Use (without proof) the monotone convergence property of the expectation
operator: If
0 X1 () X2 ()

and Xn () X() for all ,

then E Xn E X.
85

Exercise 3.12 If Yn Y , a sufficient condition for E Yn E Y is the uniform

integrability of the Yn .
Definition 3.23 The sequence Y1 , Y2 , . . . of random variables is said to
be uniformly integrable if
sup E (|Yn | I{|Yn | }) 0 as .
n

Use the following steps to prove that if Yn Y and the Yn are uniformly integrable, then E Yn E Y .
(a) Prove that if A1 , A2 , . . . and B1 , B2 , . . . are both uniformly integrable sequences defined on the same probability space, then A1 + B1 , A2 + B2 , . . . is a
uniformly integrable sequence.
Hint: First prove that
|a + b|I{|a + b| } 2|a|I{|a| /2} + 2|b|I{|b| /2}.
(b) Define Zn and Z such that Zn has the same distribution as Yn , Z has the
a.s.
same distribution as Y , and Zn Z. (We know that such random variables exist
because of the Skorohod Representation Theorem.) Show that if Xn = |Zn Z|,
then X1 , X2 , . . . is a uniformly integrable sequence.
Hint: Use Fatous Lemma (Exercise 3.11) to show that E |Z| < , i.e., Z is
integrable. Then use part (a).
(c) By part (b), the desired result now follows from the following result, which
you are asked to prove: If X1 , X2 , . . . is a uniformly integrable sequence with
a.s.
Xn 0, then E Xn 0.
Hint: Use the Dominated Convergence Theorem 3.22 and the fact that
E Xn = E Xn I{|Xn | } + E Xn I{|Xn | < }.
Exercise 3.13 Prove that if there exists  > 0 such that supn E |Yn |1+ < , then
Y1 , Y2 , . . . is a uniformly integrable sequence.
Hint: First prove that
|Yn | I{|Yn | }

86

1
|Yn |1+ .


Exercise 3.14 Prove that if there exists a random variable Z such that E |Z| = <
and P (|Yn | t) P (|Z| t) for all n and for all t > 0, then Y1 , Y2 , . . . is a
uniformly integrable sequence. You may use the fact (without proof) that for a
nonnegative X,
Z
P (X t) dt.
E (X) =
0

Hints: Consider the random variables |Yn |I{|Yn | t} and |Z|I{|Z| t}. In
addition, use the fact that
E |Z| =

E (|Z|I{i 1 |Z| < i})

i=1

to argue that E (|Z|I{|Z| < }) E |Z| as .

87

Chapter 4
Central Limit Theorems
The main result of this chapter, in Section 4.2, is the Lindeberg-Feller Central Limit Theorem, from which we obtain the result most commonly known as The Central Limit Theorem
as a corollary. As in Chapter 3, we mix univariate and multivariate results here. As a general
summary, much of Section 4.1 is multivariate and most of the remainder of the chapter is
univariate. The interplay between univariate and multivariate results is exemplified by the
Central Limit Theorem itself, Theorem 4.9, which is stated for the multivariate case but
whose proof is a simple combination of the analagous univariate result with Theorem 4.12,
the Cramer-Wold theorem.
Before we discuss central limit theorems, we include one section of background material for
the sake of completeness. Section 4.1 introduces the powerful Continuity Theorem, Theorem
4.3, which is the basis for proofs of various important results including the Lindeberg-Feller
Theorem. This section also defines multivariate normal distributions.

4.1

Characteristic Functions and Normal Distributions

While it may seem odd to group two such different-sounding topics into the same section,
there are actually many points of overlap between characteristic function theory and the
multivariate normal distribution. Characteristic functions are essential for proving the Central Limit Theorems of this chapter, which are fundamentally statements about normal
distributions. Furthermore, the simplest way to define normal distributions is by using their
characteristic functions. The standard univariate method of defining a normal distribution
by writing its density does not work here (at least not in a simple way), since not all normal
distributions have densities in the usual sense. We even provide a proof of an important
resultthat characteristic functions determine their distributions uniquelythat uses nor88

mal distributions in an essential way. Thus, the study of characteristic functions and the
study of normal distributions are so closely related in statistical large-sample theory that it
is perfectly natural for us to introduce them together.

4.1.1

The Continuity Theorem

Definition 4.1 For a random vector X, we define the characteristic function X :

Rk C by
X (t) = E exp(it> X) = E cos(t> X) + i E sin(t> X),
where i2 = 1 and C denotes the complex numbers.
The characteristic function, which is defined on all of Rk for any X (unlike the moment
generating function, which requires finite moments), has some basic properties. For instance,
X (t) is always a continuous function with X (0) = 1 and |X (t)| 1. Also, inspection of
Definition 4.1 reveals that for any constant vector a and scalar b,
X+a (t) = exp(it> a)X (t) and bX (t) = X (bt).

(4.1)

Also, if X and Y are independent,

X+Y (t) = X (t)Y (t).

(4.2)

One of the main reasons that characteristic functions are so useful is the fact that they
uniquely determine the distributions from which they are derived. This fact is so important
that we state it as a theorem:
Theorem 4.2 The random vectors X1 and X2 have the same distribution if and only
if X1 (t) = X2 (t) for all t.
d

Now suppose that Xn X, which implies t> Xn t> X. Since both sin x and cos x are
bounded continuous functions, Theorem 2.28 implies that Xn (t) X (t). The converse,
which is much harder to prove, is also true:
d

Theorem 4.3 Continuity Theorem: Xn X if and only if Xn (t) X (t) for all t.
d

Here is a partial proof that Xn (t) X (t) implies Xn X. First, we note that the
distribution functions Fn must contain a convergent subsequence, say Fnk G as k ,
where G : R [0, 1] must be a nondecreasing function but G is not necessarily a true
distribution function (and, of course, convergence is guaranteed only at continuity points of
G). It is possible to define the characteristic function of Gthough we will not prove this
89

assertionand it must follow that Fnk (t) G (t). But this implies that G (t) = X (t)
because it was assumed that Xn (t) X (t). By Theorem 4.2, G must be the distribution
function of X. Therefore, every convergent subsequence of {Xn } converges to X, which gives
the result.
Theorem 4.3 is an extremely useful tool for proving facts about convergence in distribution.
Foremost among these will be the Lindeberg-Feller Theorem in Section 4.2, but other results
follow as well. For example, a quick proof of the Cramer-Wold Theorem, Theorem 4.12, is
possible (see Exercise 4.3).

4.1.2

Moments

One of the facts that allows us to prove results about distributions using results about
characteristic functions is the relationship between the moments of a distribution and the
derivatives of a characteristic function. We emphasize here that all random variables have
well-defined characteristic functions, even if they do not have any moments. What we will
see is that existence of moments is related to differentiability of the characteristic function.
We derive X (t)/tj directly by considering the limit, if it exists, of



exp{ihXj } 1
X (t + hej ) X (t)
>
= E exp{it X}
h
h
as h 0, where ej denotes the jth unit vector with 1 in the jth component and 0 elsewhere.
Note that



 Z Xj

 

exp{ihX
}

1
j
>
=
exp{it X}
exp{iht} dt |Xj |,



h
0
so if E |Xj | < then the dominated convergence theorem, Theorem 3.22, implies that






exp{ihXj } 1
>
X (t) = E lim exp{it X}
= i E Xj exp{it> X} .
h0
tj
h
We conclude that
Lemma 4.4 If E kXk < , then X (0) = i E X.
A similar argument gives
Lemma 4.5 If E X> X < , then 2 X (0) = E XX> .
It is possible to relate higher-order moments of X to higher-order derivatives of X (t) using
the same logic, but for our purposes, only Lemmas 4.4 and 4.5 are needed.
90

4.1.3

The Multivariate Normal Distribution

It is easy to define a univariate normal distribution. If and 2 are the mean and variance, respectively, then if 2 > 0 the corresponding normal distribution is by definition the
distribution whose density is the well-known function


1
1
2
f (x) =
exp 2 (x ) .
2
2 2
If 2 = 0, on the other hand, we simply take the corresponding normal distribution to be the
constant . However, it is not quite so easy to define a multivariate normal distribution. This
is due to the fact that not all nonconstant multivariate normal distributions have densities
on Rk in the usual sense. It turns out to be much simpler to define multivariate normal
distributions using their characteristic functions:
Definition 4.6 Let be any symmetric, nonnegative definite, k k matrix and let
be any vector in Rk . Then the normal distribution with mean and covariance
matrix is defined to be the distribution with characteristic function


t> t
>
.
(4.3)
X (t) = exp it
2
Definition 4.6 has a couple of small flaws. First, because it does not stipulate k 6= 1, it offers a
definition of univariate normality that might compete with the already-established definition.
However, Exercise 4.1(a) verifies that the two definitions coincide. Second, Definition 4.6
asserts without proof that equation (4.3) actually defines a legitimate characteristic function.
How do we know that a distribution with this characteristic function really exists for all
possible and ? There are at least two ways to mend this flaw. One way is to establish
sufficient conditions for a particular function to be a legitimate characteristic function, then
prove that the function in Equation (4.3) satisfies them. This is possible, but it would take
us too far from the aim of this section, which is to establish just enough background to
aid the study of statistical large-sample theory. Another method is to construct a random
variable whose characteristic function coincides with equation (4.3); yet to do this requires
that we delve into some linear algebra. Since this linear algebra will prove useful later, this
is the approach we now take.
Before constructing a multivariate normal random vector in full generality, we first consider
the case in which is diagonal, say = D = diag(d1 , . . . , dk ). The stipulation in Definition
4.6 that be nonnegative definite means in this special case that di 0 for all i. Now
take X1 , . . . , Xk to be independent, univariate normal random variables with zero means
and Var Xi = di . We assert without proofthe assertion will be proven laterthat X =
(X1 , . . . , Xk ) is then a multivariate normal random vector, according to Definition 4.6, with
mean 0 and covariance matrix D.
91

To define a multivariate normal random vector with a general (non-diagonal) covariance

matrix , we make use of the fact that any symmetric matrix may be diagonalized by an
orthogonal matrix. We first define orthogonal, then state the diagonalizability result as a
lemma that will not be proven here.
Definition 4.7 A square matrix Q is orthogonal if Q1 exists and is equal to Q> .
Lemma 4.8 If A is a symmetric k k matrix, then there exists an orthogonal matrix
Q such that QAQ> is diagonal.
Note that the diagonal elements of the matrix QAQ> in the matrix above must be the
eigenvalues of A. This follows since if is a diagonal element of QAQ> , then it is an
eigenvalue of QAQ> . Hence, there exists a vector x such that QAQ> x = x, which implies
that A(Q> x) = (Q> x) and so is an eigenvalue of A.
Taking and as in Definition 4.6, Lemma 4.8 implies that there exists an orgthogonal
matrix Q such that QQ> is diagonal. Since we know that every diagonal entry in QQ>
is nonnegative, we may define Y = (Y1 , . . . , Yk ), where Y1 , . . . , Yk are independent normal
random vectors with mean zero and Var Yi equal to the ith diagonal entry of QQ> . Then
the random vector
X = + Q> Y

(4.4)

has the characteristic function in equation (4.3), a fact whose proof is the subject of Exercise
4.1. Thus, Equation (4.3) of Definition 4.6 always gives the characteristic function of an
actual distribution. We denote this multivariate normal distribution by Nk (, ), or simply
N (, 2 ) if k = 1.
To conclude this section, we point out that in case is invertible, then Nk (, ) has a
density in the usual sense on Rk :


1
1
> 1
(4.5)
exp (x ) (x ) ,
f (x) = p
2
2k k ||
where || denotes the determinant of . However, this density will be of little value in the
large-sample topics to follow.

4.1.4

Asymptotic Normality

Now that Nk (, ) is defined, we may use it to state one of the most useful theorems in all of
statistical large-sample theory, the Central Limit Theorem for independent and identically
distributed (iid) sequences of random vectors. We defer the proof of this theorem to the next
section, where we establish a much more general result called the Lindeberg-Feller Central
Limit Theorem.
92

Theorem 4.9 Central Limit Theorem for independent and identically distributed multivariate sequences: If X1 , X2 , . . . are independent and identically distributed
with mean Rk and covariance , where has finite entries, then

d
n(Xn ) Nk (0, ).
Although we refer to several different theorems in this chapter as central limit theorems of
one sort or another, we also employ the standard statistical usage in which the phrase The
Central Limit Theorem, with no modifier, refers to Theorem 4.9 or its univariate analogue.
Before exhibiting some examples that apply Theorem 4.9, we discuss what is generally meant
by the phrase asymptotic distribution. Suppose we are given a sequence X1 , X2 , . . . of
random variables and asked to determine the asymptotic distribution of this sequence. This
d
might mean to find X such that Xn X. However, depending on the context, this might
d
not be the case; for example, if Xn c for a constant c, then we mean something else by
asymptotic distribution.
In general, the asymptotic distribution of Xn means a nonconstant random variable X,
d
along with real-number sequences {an } and {bn }, such that an (Xn bn ) X. In this case,
the distribution of X might be referred to as the asymptotic or limiting distribution of either
Xn or of an (Xn bn ), depending on the context.
Example 4.10 Suppose that Xn is the sum of n independent Bernoulli(p) random
P
variables, so that Xn binomial(n, p). Even though we know that Xn /n p
by the weak law of large numbers, this is not generally what we mean by the
asymptotic distribution of Xn /n. Instead, the asymptotic distribution of Xn /n
is expressed by



Xn
d
n
p N {0, p(1 p)},
n
which follows from the Central Limit Theorem because a Bernoulli(p) random
variable has mean p and variance p(1 p).
Example 4.11 Asymptotic distribution of sample variance: Suppose that X1 , X2 , . . .
are independent and identically distributed with E (Xi ) = , Var (Xi ) = 2 , and
Var {(Xi )2 } = 2 < . Define
n

Sn2

1X
=
(Xi X n )2 .
n i=1

We wish to determine the asymptotic distribution of Sn2 .

93

(4.6)

Since the distribution of Xi X n does not change if we replace each Xi by

Xi , we may assume without loss of generality that = 0. By the Central
Limit Theorem, we know that
!
n

1X 2
d
2
n
Xi N (0, 2 ).
n i=1

d
n(X n ) N (0, 2 )
Furthermore, the Central Limit Theorem and the Weak Lawimply


2
P
P
and X n 0, respectively, so Slutskys theorem implies n X n 0. Therefore,
since
!
n
X

 2

1
n(Sn2 2 ) = n
Xi2 2 + n X n ,
n i=1
Slutskys theorem implies that
sult.

n (Sn2 2 ) N (0, 2 ), which is the desired re-

Note that the definition of Sn2 in Equation 4.6 is not the usual unbiased sample
variance, which uses the denominator n 1 instead of n. However, since



n
n 2
2
2
2
2
n
Sn = n(Sn ) +
S
n1
n1 n

and n/(n 1) 0, we see that the simpler choice of n does not change the
asymptotic distribution at all.

4.1.5

The Cram
er-Wold Theorem

Suppose that X1 , X2 , . . . is a sequence of random k-vectors. By Theorem 2.34, we see

immediately that
d

Xn X implies a> Xn a> X for any a Rk .

(4.7)

This is because multiplication by a constant vector aT is a continuous transformation from

Rk to R. It is not clear, however, whether the converse of statement (4.7) is true. Such a
converse would be useful because it would give a means for proving multivariate convergence
in distribution using only univariate methods. As the counterexample in Example 2.38 shows,
multivariate convergence in distribution does not follow from the mere fact that each of the
components converges in distribution. Yet the converse of statement (4.7) is much stronger
than the statement that each component converges in distribution; could it be true that
requiring all linear combinations to converge in distribution is strong enough to guarantee
multivariate convergence? The answer is yes:
94

Theorem 4.12 Cramer-Wold Theorem: Xn X if and only if a> Xn a> X for all
a Rk .
Using the machinery of characteristic functions, to be presented in Section 4.1, the proof of
the Cramer-Wold Theorem is immediate; see Exercise 4.3. This theorem in turn provides
a straightforward method for proving cerain multivariate theorems using univariate results.
For instance, once we establish the univariate Central Limit Theorem (Theorem 4.19), we
will show how to use the Cramer-Wold Theorem to prove the multivariate CLT, Theorem 4.9.

Exercises for Section 4.1

Exercise 4.1 (a) Prove that if Y N (0, 2 ) with 2 > 0, then Y (t) = exp 21 t2 2 .
Argue that this demonstrates that Definition 4.6 is valid in the case k = 1.
Hint: Verify and solve the differential equation 0Y (t) = t 2 Y (t). Use integration by parts.
(b) Using part (a),
prove that if X is defined as in Equation (4.4), then X (t) =
exp it> 21 t> t .
Exercise 4.2 We will prove Theorem 4.2, which states that chacteristic functions
uniquely determine their distributions.
(a) First, prove the Parseval relation for random X and Y:


E exp(ia> Y)X (Y) = E Y (X a).
Hint: Use conditioning to evaluate E exp{i(X a)> Y}.
(b) Suppose that Y = (Y1 , . . . , Yk ), where Y1 , . . . , Yk are independent and identically distributed normal random variables with mean 0 and variance 2 . That
is, Y has density

fY (y) = ( 2 2 )k exp(y> y/2 2 ).

Show that X + Y has density
fX+Y (s) = E fY (s X).
(c) Use the result of Exercise 4.1 along with part (b) to show that



X
s
k
fX+Y (s) = ( 2 2 ) E Y

.
2 2
95

Argue that this fact proves X (t) uniquely determines the distribution of X.
Hint: Use parts (a) and (b) to show that the distribution of X + Y depends on
d
X only through X . Then note that X + Y X as 2 0.
Exercise 4.3 Use the Continuity Theorem to prove the Cramer-Wold Theorem, Theorem 4.12.
d

Hint: a> Xn a> X implies that a> Xn (1) a> X (1).

Exercise 4.4 Suppose X Nk (, ), where is invertible. Prove that
(X )> 1 (X ) 2k .
Hint: If Q diagonalizes , say QQ> = , let 1/2 be the diagonal, nonnegative
matrix satisfying 1/2 1/2 = and consider Y> Y, where Y = (1/2 )1 Q(X).
Exercise 4.5 Let X1 , X
2 , . . . be independent Poisson random variables with mean
= 1. Define Yn = n(X n 1).
(a) Find E (Yn+ ), where Yn+ = Yn I{Yn > 0}.
(b) Find, with proof, the limit of E (Yn+ ) and prove Stirlings formula

n!
2 nn+1/2 en .
Hint: Use the result of Exericse 3.12.
Exercise 4.6 Use the Continuity Theorem to prove Theorem 2.19, the univariate
Weak Law of Large Numbers.
Hint: Use a Taylor expansion (1.5) with d = 2 for both the real and imaginary
parts of the characteristic function of X n .
Exercise 4.7 Use the Cramer-Wold Theorem along with the univariate Central Limit
Theorem (from Example 2.12) to prove Theorem 4.9.

4.2

The Lindeberg-Feller Central Limit Theorem

The Lindeberg-Feller Central Limit Theorem states in part that sums of independent random
variables, properly standardized, converge in distribution to standard normal as long as a
certain condition, called the Lindeberg Condition, is satisfied. Since these random variables
do not have to be identically distributed, this result generalizes the Central Limit Theorem
for independent and identically distributed sequences.
96

4.2.1

The Lindeberg and Lyapunov Conditions

Suppose that X1 , X2 , . . . are independent random variables such that E Xn = n and

Var Xn = n2 < . Define
Yn = Xn n ,
Pn
Tn =
i=1 Yi ,
P
2
sn = Var Tn = ni=1 i2 .
Instead of defining Yn to be the centered version of Xn , we could have simply taken n to be
zero without loss of generality. However, when these results are used in practice, it is easy
to forget the centering step, so we prefer to make it explicit here.
Note that Tn /sn has mean zero and variance 1. We wish to give sufficient conditions that
d
ensure Tn /sn N (0, 1). We give here two separate conditions, one called the Lindeberg condition and the other called the Lyapunov condition. The Lindeberg Condition for sequences
states that
n


1 X
for every  > 0, 2
E Yi2 I {|Yi | sn } 0 as n ;
sn i=1

(4.8)

the Lyapunov Condition for sequences states that

there exists > 0 such that

n
1 X

s2+
n

E |Yi |2+ 0 as n .

(4.9)

i=1

We shall see later (in Theorem 4.16, the Lindeberg-Feller Theorem) that Condition (4.8)
implies Tn /sn N (0, 1). For now, we show only that Condition (4.9)the Lyapunov
Conditionis stronger than Condition (4.8). Thus, the Lyapunov Condition also implies
Tn /sn N (0, 1):
Theorem 4.13 The Lyapunov Condition (4.9) implies the Lindeberg Condition (4.8).
Proof: Assume that the Lyapunov Condition is satisfied and fix  > 0. Since |Yi | sn
implies |Yi /sn | 1, we obtain
n
n




1 X
1 X
2
2+
E
Y
I
{|Y
|

s
}

E
|Y
|
I
{|Y
|

s
}
i
n
i
i
n
i
s2n i=1
 s2+
n
i=1
n


1 X
2+
E |Yi |2+ .
 sn i=1

Since the right hand side tends to 0, the Lindeberg Condition is satisfied.
97

Example 4.14 Suppose that we perform a series of independent Bernoulli trials with
possibly different success probabilities. Under what conditions will the proportion
of successes, properly standardized, tend to a normal distribution?
Let Xn Bernoulli(pn ), so that Yn = Xn pn and n2 = pn (1 pn ). As we shall
see later (Theorem 4.16), either the Lindeberg Condition (4.8) or the Lyapunov
P
d
Condition (4.9) will imply that ni=1 Yi /sn N (0, 1).
Let us check the Lyapunov Condition for, say, = 1. First, verify that
E |Yn |3 = pn (1 pn )3 + (1 pn )p3n = n2 [(1 pn )2 p2n ] n2 .
P
Using this upper bound on E |Yn |3 , we obtain ni=1 E |Yi |3 s2n . Therefore, the
Lyapunov condition is satisfied whenever s2n /s3n 0, which implies sn . We
conclude that the proportion of successes tends to a normal distribution whenever
s2n =

n
X

pn (1 pn ) ,

i=1

which will be true as long as pn (1 pn ) does not tend to 0 too fast.

4.2.2

Independent and Identically Distributed Variables

We now set the stage for proving a central limit theorem for independent and identically
distributed random variables by showing that the Lindeberg Condition is satisfied by such
a sequence as long as the common variance is finite.
Example 4.15 Suppose that X1 , X2 , . . . are independent and identically distributed
with E (Xi ) = and Var (Xi ) = 2 < . The case 2 = 0 is uninteresting, so we
assume 2 > 0.
P
Let Yi = Xi and s2n = Var ni=1 Yi = n 2 . Fix  > 0. The Lindeberg
Condition states that
n


1 X
2
E
Y
I{|Y
|


n} 0 as n .
i
i
n 2 i=1

(4.10)

Since the Yi are identically distributed, the left hand side of expression (4.10)
simplifies to


1
2
E
Y
I{|Y
|


n} .
1
1
2
98

(4.11)


To simplify notation, let Zn denote the random variable Y12 I{|Y1 |  n}.
Thus, we
wish to prove that E Zn 0. Note that Zn is nonzero if and only if
|Y1 |  n. Since this event has probability tending to zero as n , we
P
conclude that Zn 0 by the definition of convergence in probability. We can also
see that |Zn | Y12 , and we know that E Y12 < . Therefore, we may apply the
Dominated Convergence Theorem, Theorem 3.22, to conclude that E Zn 0.
This demonstrates that the Lindeberg Condition is satisfied.
The preceding argument, involving the Dominated Convergence Theorem, is quite common
in proofs that the Lindeberg Condition is satisfied. Any beginning student is well-advised
to study this argument carefully.
Note that the assumptions of Example 4.15 are not strong enough to ensure that the Lyapunov Condition (4.9) is satisfied. This is because there are some random variables that
have finite variances but no finite 2 + moment for any > 0. Construction of such an
example is the subject of Exercise 4.10. However, such examples are admittedly somewhat
pathological, and if one is willing to assume that X1 , X2 , . . . are independent and identically
distributed with E |X1 |2+ = < for some > 0, then the Lyapunov
Condition is much
easier to check than the Lindeberg Condition. Indeed, because sn = n, the Lyapunov
Condition reduces to

n
= /2 2+ 0,
2
1+/2
(n )
n
which follows immediately.

4.2.3

Triangular Arrays

It is sometimes the case that X1 , . . . , Xn are independent random variablespossibly even

identically distributedbut their distributions depend on n. Take the simple case of the
binomial(n, pn ) distribution as an example, where the probability pn of success on any trial
changes as n increases. What can we say about the asymptotic distribution in such a
case? It seems that what we need is some way of dealing with a sequence of sequences, say,
Xn1 , . . . , Xnn for n 1. This is exactly the idea of a triangular array of random variables.
Generalizing the concept of sequence of independent random variables, a triangular array
or random variables may be visualized as follows:
X11
independent
X21 X22
independent
X31 X32 X33 independent
..
.
99

Thus, we assume that for each n, Xn1 , . . . , Xnn are independent. Carrying over the notation
2
< . Let
from before, we assume E Xni = ni and Var Xni = ni
Yni = Xni ni ,
Pn
Tn =
i=1 Yni ,
P
2
2
.
sn = Var Tn = ni=1 ni
As before, Tn /sn has mean 0 and variance 1; our goal is to give conditions under which
Tn d
N (0, 1).
sn

(4.12)

Such conditions are given in the Lindeberg-Feller Central Limit Theorem. The key to this
theorem is the Lindeberg condition for triangular arrays:
For every  > 0,

n


1 X
2
E
Y
I
{|Y
|

s
}
0 as n .
ni
n
ni
s2n i=1

(4.13)

Before stating the Lindeberg-Feller theorem, we need a technical condition that says essentially that the contribution of each Xni to s2n should be negligible:
1
max 2 0 as n .
s2n in ni

(4.14)

Now that Conditions (4.12), (4.13), and (4.14) have been written, the main result may be
stated in a single line:
Theorem 4.16 Lindeberg-Feller Central Limit Theorem: Condition (4.13) holds if
and only if Conditions (4.12) and (4.14) hold.
A proof of the Lindeberg-Feller Theorem is the subject of Exercises 4.8 and 4.9. In most
practical applications of this theorem, the Lindeberg Condition (4.13) is used to establish
asymptotic normality (4.12); the remainder of the theorems content is less useful.
Example 4.17 As an extension of Example 4.14, suppose Xn binomial(n, pn ). The
calculations here are not substantially different from those in Example 4.14, so
we use the Lindeberg Condition here for the purpose of illustration. We claim
that
X npn
d
p n
N (0, 1)
(4.15)
npn (1 pn )
whenever npn (1 pn ) as n . In order to use Theorem 4.16 to prove
this result, let Yn1 , . . . , Ynn be independent and identically distributed with
P (Yni = 1 pn ) = 1 P (Yni = pn ) = pn .
100

P
Then with Xn = npn + ni=1 Yni , we obtain Xn binomial(n, pn ) as specified.
Furthermore, E Yni = 0 and Var Yni = pn (1 pn ), so the Lindeberg condition
says that for any  > 0,
n
o
n

X
p
1
E Yni2 I |Yni |  npn (1 pn )
0.
npn (1 pn ) i=1

(4.16)

Since |Yni | p
1, the left hand side of expression (4.16) will be identically zero
whenever  npn (1 pn ) > 1. Thus, a sufficient condition for (4.15) to hold is
that npn (1 pn ) . One may show that this is also a necessary condition
(this is Exercise 4.11).
Note that any independent sequence X1 , X2 , . . . may be considered a triangular array by
simply taking Xn1 = X1 for all n 1, Xn2 = X2 for all n 2, and so on. Therefore,
Theorem 4.16 applies equally to the Lindeberg Condition (4.8) for sequences. Furthermore,
the proof of Theorem 4.13 is unchanged if the sequence Yi is replaced by the array Yni .
Therefore, we obtain an alternative means for checking asymptotic normality:
Corollary 4.18 Asymptotic normality (4.12) follows if the triangular array above
satisfies the Lyapunov Condition for triangular arrays:
there exists > 0 such that

n
1 X

s2+
n

E |Yni |2+ 0 as n .

(4.17)

i=1

Combining Theorem 4.16 with Example 4.15, in which the Lindeberg condition is verified for
a sequence of independent and identically distributed variables with finite positive variance,
gives the result commonly referred to simply as The Central Limit Theorem:
Theorem 4.19 Univariate Central Limit Theorem for iid sequences: Suppose that
X1 , X2 , . . . are independent and identically distributed with E (Xi ) = and
Var (Xi ) = 2 < . Then

d
n X n N (0, 2 ).

(4.18)

The case 2 = 0 is not covered by Example 4.15, but in this case limit (4.18) holds automatically.
We conclude this section by generalizing Theorem 4.19 to the multivariate case, Theorem
4.9. The proof is straightforward using theorem 4.19 along with the Cramer-Wold theorem,
theorem 4.12. Recall that the Cramer-Wold theorem allows us to establish multivariate
convergence in distribution by proving univariate convergence in distribution for arbitrary
linear combinations of the vector components.
101

Proof of Theorem 4.9: Let X Nk (0, ) and take any vector a Rk . We wish to show
that


 d
a> n Xn a> X.
But this follows immediately from the univariate Central Limit Theorem, since a> (X1
), a> (X2 ), . . . are independent and identically distributed with mean 0 and variance
a> a.
We will see many, many applications of the univariate and multivariate Central Limit Theorems in the chapters that follow.

Exercises for Section 4.2

Exercise 4.8 Prove that (4.13) implies both (4.12) and (4.14) (the forward half of
the Lindeberg-Feller Theorem). Use the following steps:
(a) Prove that for any complex numbers a1 , . . . , an and b1 , . . . , bn with |ai | 1
and |bi | 1,
|a1 an b1 bn |

n
X

|ai bi | .

(4.19)

i=1

Hint: First prove the identity when n = 2, which is the key step. Then use
mathematical induction.
(b) Prove that

  

2
2 2

 |t|3 ni


t
t2
t

ni
2
Y

+
E
Y
I{|Y
|

s
}
.

ni
n
ni
ni

sn
2s2n 
s2n
s2n

(4.20)

Hint: Use the results of Exercise 1.43, parts (c) and (d), to argue that for any
Y,


 
  3  
  2
2 2 

 tY 
Y 
itY
itY
t
Y
exp
   I   <  + tY
1+

I{|Y | sn }.

 sn 
2
sn
sn
2sn   sn 
sn
(c) Prove that (4.13) implies (4.14).
Hint: For any i, show that
2
ni
E (Yni2 I{|Yni | sn })
2
<

+
.
s2n
s2n

102

2
(d) Use parts (a) and (b) to prove that, for n large enough so that t2 maxi ni
/s2n
1,

  Y

n
n
n 
Y
2


t
t2 ni
t2 X


3
2
E
Y
I
{|Y
|

s
}
.

|t|
+
 Yni

ni
n
ni

sn
2s2n 
s2n i=1
i=1
i=1

(e) Use part (a) to prove that

 
 Y
 2 2 
n
n
n
Y
2 2
t4 X 4
t4
t ni
t ni 

2

.
1
exp 2  4
ni 2 max ni

2
1in

2sn
2sn  4sn i=1
4sn
i=1
i=1
Hint: Prove that for x 0, |1 + x exp(x)| x2 .
(f ) Now put it all together. Show that

  Y
 2 2 
n
n
Y
t
t



exp 2ni  0,
 Yni


s
2s
n
n
i=1
i=1
proving (4.12).
Exercise 4.9 In this problem, we prove the converse of Exercise 4.8, which is the
part of the Lindeberg-Feller Theorem due to Feller: Under the assumptions of
the Exercise 4.8, the variance condition (4.14) and the asymptotic normality
(4.12) together imply the Lindeberg condition (4.13).
(a) Define
ni = Yni (t/sn ) 1.
Prove that
max |ni | 2 max P (|Yni | sn ) + 2|t|

1in

1in

and thus
max |ni | 0 as n .

1in

Hint: Use the result of Exercise 1.43(a) to show that | exp{it}1| 2 min{1, |t|}
for t R. Then use Chebyshevs inequality along with condition (4.14).

103

(b) Use part (a) to prove that

n
X

|ni |2 0

i=1

as n .
2
Hint: Use the result of Exercise 1.43(b) to show that |ni | t2 ni
/s2n . Then
write |ni |2 |ni | maxi |ni |.

(c) Prove that for n large enough so that maxi |ni | 1/2,


n
n
n
 X
Y
Y


|ni |2 .
 exp(ni ) (1 + ni )


i=1

i=1

i=1

Hint: Use the fact that |exp(z 1)| = exp(Re z 1) 1 for |z| 1 to argue
that Inequality (4.19) applies. Also use the fact that | exp(z) 1 z| |z|2 for
|z| 1/2.
(d) From part (c) and condition (4.12), conclude that
n
X
i=1

1
Re (ni ) t2 .
2

(e) Show that

0

n
X


E

i=1

tYni
t2 Yni2
cos
1+
sn
2s2n


0.

(f ) For arbitrary  > 0, choose t large enough so that t2 /2 > 2/2 . Show that


t2
2
2
2




n
n

X
tYni
1 X
t2 Yni2
2
E Yni I{|Yni | sn }
E cos
1+
,
2
s2n i=1
s
2s
n
n
i=1

which completes the proof.

Hint: Bound the expression in part (e) below by using the fact that 1 is a
lower bound for cos x. Also note that |Yni | sn implies 2 2Yni2 /(2 s2n ).
104

Exercise 4.10 Give an example of an independent and identically distributed sequence to which the Central Limit Theorem 4.19 applies but for which the Lyapunov condition is not satisfied.
Exercise 4.11 In Example 4.17, we show that npn (1 pn ) is a sufficient condition for (4.15) to hold. Prove that it is also a necessary condition. You may
assume that pn (1 pn ) is always nonzero.
Hint: Use the Lindeberg-Feller Theorem.
Exercise 4.12 (a) Suppose that X1 , X2 , . . . are independent and identically distributed with E Xi = and 0 < Var Xi = 2 < . Let an1 , . . . , ann be constants
satisfying
maxin a2ni
Pn
0 as n .
2
j=1 anj
Let Tn =

Pn

i=1

d
ani Xi , and prove that (Tn E Tn )/ Var Tn N (0, 1).

(b) Reconsider Example 2.22, the simple linear regression case in which
0n =

n
X

(n)
vi Yi

and 1n =

i=1

n
X

(n)

w i Yi ,

i=1

where
(n)

wi

zi z n
1
(n)
(n)
and vi = z n wi
2
n
j=1 (zj z n )

= Pn

for constants z1 , z2 , . . .. Using part (a), state and prove sufficient

conditions
on the constants zi that ensure the asymptotic normality of n(0n 0 ) and

n(1n 1 ). You may assume the results of Example 2.22, where it was shown
that E 0n = 0 and E 1n = 1 .
Exercise 4.13 Give an example (with proof) of a sequence of independent
random

variables Z1 , Z2 , . . . with E (Zi ) = 0, Var (Zi ) = 1 such that n(Z n ) does not
converge in distribution to N (0, 1).
Exercise 4.14 Let (a1 , . . . , an ) be a random permutation of the integers 1, . . . , n. If
aj < ai for some i < j, then the pair (i, j) is said to form an inversion. Let Xn
be the total number of inversions:
Xn =

j1
n X
X
j=2 i=1

105

I{aj < ai }.

For example, if n = 3 and we consider the permutation (3, 1, 2), there are 2
inversions since 1 = a2 < a1 = 3 and 2 = a3 < a1 = 3. This problem asks you to
find the asymptotic distribution of Xn .
(a) Define Y1 = 0 and for j > 1, let
Yj =

j1
X

I{aj < ai }

i=1

be the number of ai greater than aj to the left of aj . Then the Yj are independent
(you dont have to show this; you may wish to think about why, though). Find
E (Yj ) and Var Yj .
(b) Use Xn = Y1 + Y2 + + Yn to prove that


3
4Xn
d
n
1 N (0, 1).
2
2
n
(c) For n = 10, evaluate the distribution of inversions as follows. First, simulate
1000 permutations on {1, 2, . . . , 10} and for each permutation, count the number
of inversions. Plot a histogram of these 1000 numbers. Use the results of the
simulation to estimate P (X10 24). Second, estimate P (X10 24) using a
normal approximation. Can you find the exact integer c such that 10!P (X10
24) = c?
Exercise 4.15 Suppose that X1 , X2 , X3 is a sample of size 3 from a beta (2, 1) distribution.
(a) Find P (X1 + X2 + X3 1) exactly.
(b) Find P (X1 + X2 + X3 1) using a normal approximation derived from the
central limit theorem.
(c) Let Z
I{X1 + X2 + X3 1}. Approximate E Z = P (X1 + X2 + X3 1)
P=
1000
by Z =
i=1 Zi /1000, where Zi = I{Xi1 + Xi2 + Xi3 1} and the Xij are
independent beta (2, 1) random variables. In addition to Z, report Var Z for
your sample. (To think about: What is the theoretical value of Var Z?)
(d) Approximate P (X1 + X2 + X3 32 ) using the normal approximation and
the simulation approach. (Dont compute the exact value, which is more difficult
to than in part (a); do you see why?)
Exercise 4.16 Lindeberg and Lyapunov impose sufficient conditions on moments so
that asymptotic normality occurs. However, these conditions are not necessary;
106

it is possible to have asymptotic normality even if there are no moments at all.

Let Xn assume the values +1 and 1 with probability (1 2n )/2 each and the
value 2k with probability 2k for k > n.
(a) Show that E (Xnj ) = for all positive integers j and n.

d

(b) Show that n X n N (0, 1).

Exercise 4.17 Assume that elements (coupons) are drawn from a population of
size n, randomly and with replacement, until the number of distinct elements
that have been sampled is rn , where 1 rn n. Let Sn be the drawing on which
this first happens. Suppose that rn /n , where 0 < < 1.
(a) Suppose k 1 distinct coupons have thus far entered the sample. Let Xnk
be the waiting time until the next distinct one appears, so that
Sn =

rn
X

Xnk .

k=1

Find the expectation and variance of Xnk .

(b) Let mn = E (Sn ) and n2 = Var (Sn ). Show that
Sn mn d
N (0, 1).
n
Tip: One approach is to apply Lyapunovs condition with = 2. This involves
demonstrating an asymptotic expression for n2 and a bound on E [Xnk E(Xnk )]4 .
There are several ways to go about this.
Exercise 4.18 Suppose that X1 , X2 , . . . are independent binomial(2, p) random variables. Define Yi = I{Xi = 0}.
(a) Find a such that the joint asymptotic distribution of
 


Xn
n
a
Yn
is nontrivial, and find this joint asymptotic distribution.
(b) Using the
Cramer-Wold Theorem, Theorem 4.12, find the asymptotic distribution of n(X n + Y n 1 p2 ).

107

4.3

Stationary m-Dependent Sequences

Here we consider sequences that are identically distributed but not independent. In fact, we
make a stronger assumption than identically distributed; namely, we assume that X1 , X2 , . . .
is a stationary sequence. (Stationary is defined in Definition 2.24.) Denote E Xi by and
let 2 = Var Xi .

We seek sufficient conditions for the asymptotic normality of n(X n ). The variance of
X n for a stationary sequence is given by Equation (2.20). Letting k = Cov (X1 , X1+k ), we
conclude that
Var

n1

2X
2
n(X n ) = +
(n k)k .
n k=1

(4.21)

Suppose that
n1

2X
(n k)k
n k=1

(4.22)

as n . Then based on Equation (4.21), it seems reasonable to ask whether

n(X n ) N (0, 2 + ).

The answer, in many cases, is yes. This section explores one such case.
Recall from Definition 2.26 that X1 , X2 , . . . is m-dependent for some m 0 if the vector
(X1 , . . . , Xi ) is independent of (Xi+j , Xi+j+1 , . . .) whenever j > m. Therefore, for an mdependent sequence we have k = 0 for all k > m, so limit (4.22) becomes
n1
m
X
2X
(n k)k 2
k .
n k=1
k=1

For a stationary m-dependent sequence, the following theorem asserts the asymptotic normality of X n as long as the Xi are bounded:
Theorem 4.20 If for some m 0, X1 , X2 , . . . is a stationary m-dependent sequence
of bounded random variables with E Xi = and Var Xi = 2 , then
!
m
X

d
n(X n ) N 0, 2 + 2
Cov [X1 , X1+k ] .
k=1

108

The assumption in Theorem 4.20 that the Xi are bounded is not necessary, as long as 2 < .
However, the proof of the theorem is quite tricky without the boundedness assumption, and
the theorem is strong enough for our purposes as it stands. See, for instance, Ferguson (1996)
for a complete proof. The theorem may be proved using the following strategy: For some
integer kn , define random variables V1 , V2 , . . . and W1 , W2 , . . . as follows:
V1 = X1 + + Xkn ,
V2 = Xkn +m+1 + + X2kn +m ,

..
.

W1 = Xkn +1 + + Xkn +m ,
W2 = X2kn +m+1 + + X2kn +2m ,

(4.23)

In other words, each Vi is the sum of kn of the Xi and each Wi is the sum of m of the Xi .
Because the sequence of Xi is m-dependent, we conclude that the Vi are independent. For
this reason, we may apply the Lindeberg-Feller theorem to the Vi . If kn is defined carefully,
then the contribution of the Wi may be shown to be negligible. This strategy is implemented
in Exercise 4.19, where a proof of Theorem 4.20 is outlined.
Example 4.21 Runs of successes: Suppose X1 , X2 , . . . are independent Bernoulli(p)
variables. Let Tn denote the number of runs of successes in X1 , . . . , Xn , where
a run of successes is defined as a sequence of consecutive Xi , all of which equal
1, that is both preceded and followed by zeros (unless the run begins with X1 or
ends with Xn ). What is the asymptotic distribution of Tn ?
We note that
Tn =

n
X

I{run starts at ith position}

i=1

= X1 +

n
X

Xi (1 Xi1 ),

i=2

since a run starts at the ith position for i > 1 if and only if Xi = 1 and Xi1 = 0.
Letting Yi = Xi+1 (1 Xi ), we see immediately that Y1 , Y2 , . . . is a stationary

1-dependent sequence with E Yi = p(1 p), so that by Theorem 4.20, n{Y n

d
p(1 p)} N (0, 2 ), where
2 = Var Y1 + 2 Cov (Y1 , Y2 )
= E Y12 (E Y1 )2 + 2 E Y1 Y2 2(E Y1 )2
= E Y1 3(E Y1 )2 = p(1 p) 3p2 (1 p)2 .
Since
Tn np(1 p)
X1 Yn

= n{Y n p(1 p)} +

,
n
n
109

we conclude that
Tn np(1 p) d

N (0, 2 ).
n

Exercises for Section 4.3

Exercise 4.19 We wish to prove theorem 4.20. Suppose X1 , X2 , . . . is a stationary m-dependent sequence of bounded random variables such that Var Xi =
2 . Without loss of generality, assume E Xi = 0. We wish to prove that

d
n(X n ) N (0, 2 ), where
2

= +2

m
X

Cov (X1 , X1+k ).

k=1

For all n, define kn = bn1/4 c and `n = bn/(kn + m)c and tn = `n (kn + m). Define
V1 , . . . , V`n and W1 , . . . , W`n as in Equation (4.23). Then

`n
`n
n
1 X
1 X
1 X

n(X n ) =
Vi +
Wi +
Xi .
n i=1
n i=1
n i=t +1
n

(a) Prove that

n
1 X
P

Xi 0.
n i=t +1

(4.24)

Hint: Bound the left hand side of expression (4.24) using Markovs inequality
(1.35) with r = 1. What is the greatest possible number of summands?
(b) Prove that
`

n
1 X
P

Wi 0.
n i=1

Hint: For kn > m, the Wi are independent and identically distributedwithP

disn
tributions that do not depend on n. Use the central limit theorem on (1/ `n ) `i=1
Wi .
110

(c) Prove that

`

n
1 X
d

Vi N (0, 2 ),
n i=1

then use Slutskys theorem to prove theorem 4.20.

Hint: Use the Lindeberg-Feller theorem.
Exercise 4.20 Suppose X0 , X1 , . . . is an independent sequence of Bernoulli trials with
success probability p. Suppose Xi is the indicator of your teams success on rally
i in a volleyball game. Your teamP
scores a point each time it has a success that
follows another success. Let Sn = ni=1 Xi1 Xi denote the number of points your
team scores by time n.
(a) Find the asymptotic distribution of Sn .
(b) Simulate a sequence X0 , X1 , . . . , X1000 as above and calculate S1000 for p = .4.
Repeat this process 100 times, then graph the empirical distribution of S1000 obtained from simulation on the same axes as the theoretical asymptotic distribution
from (a). Comment on your results.
Exercise 4.21 Let X0 , X1 , . . . be independent and identically distributed random
variables from a continuous distribution F (x). Define Yi = I{Xi < X
i1 and Xi < Xi+1 }.
P
Thus, Yi is the indicator that Xi is a relative minimum. Let Sn = ni=1 Yi .
(a) Find the asymptotic distribution of Sn .
(b) Let n = 5000. For a sample X0 , . . . , X5001 of size 5002 from the uniform (0, 1)
random number generator in R, compute an approximate two-sided p-value based
on the observed value of Sn and the answer to part (a). The null hypothesis is
that the sequence of random numbers generated is independent and identically
distributed. (Naturally, the random numbers are not random at all, but are
generated by a deterministic formula that is supposed to mimic randomness.)

4.4

Univariate extensions

This section discusses two different extensions of Theorem 4.19, the univariate Central Limit
Theorem. As in the statement of that theorem, we assume here that X1 , X2 , . . . are independent and identically distributed with E (Xi ) = and Var (Xi ) = 2 .

111

4.4.1

The Berry-Esseen theorem

Let us define Yi = (Xi )/ and Sn = nY n . Furthermore, let Gn (s) denote the cumulative
distribution function of Sn , i.e., Gn (s) = P (Sn s). Then the Central Limit Theorem tells
us that for any real number s, Gn (s) (s) as n , where as usual we let denote the
cumulative distribution function of the standard normal distribution. Since (s) is bounded
and continuous, we know that this convergence is uniform, which is to say that
sup |Gn (s) (s)| 0 as n .
sR

Of course, this limit result says nothing about how close the left and right sides must be
for any fixed (finite) n. However, theorems discovered independently by Andrew Berry and
Carl-Gustav Esseen in the early 1940s do just this (each of these mathematicians actually
proved a result that is slightly more general than the one that typically bears their names).
The so-called Berry-Esseen Theorem is as follows:
Theorem 4.22 There exists a constant c such that if Y1 , Y2 , . . . are independent and
identically distributed random variables with mean 0 and variance 1, then
sup |Gn (s) (s)|
sR

c E |Y13 |

for all n, where Gn (s) is the cumulative distribution function of

nY n .

Notice that the inequality is vacuously true whenever E |Y13 | is infinite. In terms of the
original sequence X1 , X2 , . . . , the theorem is therefore sometimes stated by saying that when
= E |X13 | < ,
c
sup |Gn (s) (s)| 3 .
n
sR
We will not give a proof of Theorem 4.22 here, though the interested reader might wish to
consult papers by Ho and Chen (1978, Annals of Probability, pp. 231249) and Stein (1972,
Proceedings of the Sixth Berkeley Symposium on Mathematical Statistics and Probability,
Volume 2, pp. 583602). The former authors give a proof of Theorem 4.22 based on the
Stein paper that gives the value c = 6.5. However, they do not prove that 6.5 is the
smallest possible value of c, and in fact an interesting aspect of the Berry-Esseen Theorem
is that the smallest possible value is not known. Currently, one author (Irina Shevtsova,
arXiv:1111.6554v1) has shown that the inequality is valid for c = 0.4748. Furthermore,
Esseen himself proved that c cannot be less than 0.4097. For the sake of simplicity, we may
exploit the known results by taking c = 1/2 to state with certainty that
sup |Gn (s) (s)|
sR

112

E |Y13 |
.
2 n

4.4.2

Edgeworth expansions

As in the previous section, Let us define Yi = (Xi )/ and Sn =

= E Yi3

nY n . Furthermore, let

and = E Yi4

and suppose that < . The Central Limit Theorem says that for every real y,
P (Sn y) = (y) + o(1) as n .
But we would like a better approximation to P (Sn y) than (y), and we begin by constructing the characteristic function of Sn :
(
)
n

X
n
Sn (t) = E exp (it/ n)
(4.25)
Yi = Y (t/ n) ,
i=1

where Y (t) = E exp{itY } is the characteristic function of Yi .

Before proceeding with an examination of Equation (4.25), we first establish four preliminary
facts:
1. Sharpening a well-known limit: We already know that (1 + a/n)n ea . But how
good is this approximation? The binomial theorem shows (after quite a bit of algebra)
that for a fixed nonnegative integer k,

 


1
a nk
a(a + 2k)
a
+o
(4.26)
1+
=e 1
n
2n
n
as n .
2. Hermite polynomials: If (x) denotes the standard normal density function, then
we define the Hermite polynomials Hk (x) by the equation
(1)k

dk
(x) = Hk (x)(x).
dxk

(4.27)

Thus, by simply differentiating (x) repeatedly, we may verify that H1 (x) = x, H2 (x) =
x2 1, H3 (x) = x3 3x, and so on. By differentiating Equation (4.27) itself, we obtain
the recursive formula
d
[Hk (x)(x)] = Hk+1 (x)(x).
dx

113

(4.28)

3. An inversion formula for characteristic

Z (t)
R functions: Suppose Z G(z) and
0
denotes the characteristic function of Z. If |Z (t)| dt < , then g(z) = G (z) exists
and
Z
1
g(z) =
eitz Z (t) dt.
(4.29)
2
We wont prove Equation (4.29) here, but a proof can be found in most books on
theoretical probability.
4. An identity involving (x): For any positive integer k,
Z
Z
1
(1)k dk
itx t2 /2
k
itx t2 /2
e e
(it) dt =
e
e
dt
2
2 dxk
dk
= (1)k k (x)
dx
= Hk (x)(x),

(4.30)
(4.31)

where (4.30) follows from (4.29) since et /2 is the characteristic function for a standard
normal distribution, and (4.31) follows from (4.27).

Returning to Equation (4.25), we next use a Taylor expansion of exp{itY / n}: As n ,


 



(it)2 Y 2 (it)3 Y 3 (it)4 Y 4
1
t
itY
+
+
Y
+o
= E 1+ +
2
2n
24n
n2
n
n
6n n




1
t2
(it4 )
(it)3
=
1
+o
+ +
.
2
2n
24n
n2
6n n
If we raise this tetranomial to the nth power, most terms are o(1/n):



Y

n


n 
n1  3

t2
t2
(it) (it)4
+
=
1
+ 1
2n
2n
24n
6 n





n2
t2
(n 1)(it)6 2
1
+ 1
+o
.
2
2n
72n
n

By Equations (4.26) and (4.32), we conclude that


 

(it)3 (it)4
(it)6 2
1
t4
t2 /2
+o
Sn (t) = e
1
+ +
+
8n
24n
72n
n
6 n


 
3
4
6 2
(it) (it) ( 3) (it)
1
2
= et /2 1 + +
+
+o
.
24n
72n
n
6 n
114

(4.32)

(4.33)

If we apply these three approximations to equation (4.32), we obtain


 
n

 
t4
t
(it)3 (it)4
(it)6 2
1
t2 /2
1
X
= e
+ +
+
+o
8n
24n
72n
n
n
6 n


 
3
6 2
4
(it) (it) ( 3) (it)
1
2
+
+o
.
= et /2 1 + +
24n
72n
n
6 n
Putting (4.33) together with (4.29), we obtain the following density function as an approximation to the distribution of Sn :
Z
Z
1

2
ity t2 /2
g(y) =
e e
dt +
eity et /2 (it)3 dt
2
6 n

Z
Z
3
2
itx t2 /2
4
ity t2 /2
6
+
e e
(it) dt +
e e
(it) dt .
(4.34)
24n
72n
Next, combine (4.34) with (4.31) to yield


H3 (y) ( 3)H4 (y) 2 H6 (y)
+
.
g(y) = (y) 1 + +
24n
72n
6 n

(4.35)

By (4.28), the antiderivative of g(y) equals



H2 (y) ( 3)H3 (y) 2 H5 (y)
+
G(y) = (y) (y)
+
24n
72n
6 n


2
3
(y 1) ( 3)(y 3y) 2 (y 5 10y 3 + 15y)

+
+
= (y) (y)
.
24n
72n
6 n
The expression above is called the second-order Edgeworth expansion. By carrying out the
expansion in (4.33) to more terms, we may obtain higher-order Edgeworth expansions. On
the other hand, the first-order Edgeworth expansion is


(y 2 1)

.
(4.36)
G(y) = (y) (y)
6 n
(see Exercise 4.23). Thus, if the distribution of Y is symmetric, we obtain = 0 and
therefore in this case, the usual (zero-order) central limit theorem approximation given by
(y) is already first-order accurate.
Incidentally, the second-order Edgeworth expansion explains why the standard definition of
kurtosis of a distribution with mean 0 and variance 1 is the unusual-looking 3.

Exercises for Section 4.4

Exercise 4.22 Verify Equation (4.26).
Exercise 4.23 Verify that Equation (4.36) is the first-order Edgeworth approximation to the distribution function of Sn .
115

Chapter 5
The Delta Method and Applications
5.1

Local linear approximations

Suppose that a particular random sequence converges in distribution to a particular constant. The idea of using a first-order (linear) Taylor expansion of a known function, in the
neighborhood of that constant limit, is a very useful technique known as the delta method.
This chapter introduces the method, named for the in g(x + x) g(x) + xg 0 (x), and
discusses some of its applications.

5.1.1

Asymptotic distributions of transformed sequences

In the simplest form of the Central Limit Theorem, Theorem 4.19, we consider a sequence
X1 , X2 , . . . of independent and identically distributed (univariate) random variables with
finite variance 2 . In this case, the Central Limit Theorem states that

d
n(X n ) Z,
(5.1)
where = E X1 and Z is a standard normal random variable.
In this chapter, we wish to consider the asymptotic distribution of some function of X n . In
the simplest case, the answer depends on results already known: Consider a linear function
g(t) = at+b for some known constants a and b. Since E X n = , clearly E g(X n ) = a+b =
g()
by the linearity of the expectation operator. Therefore, it is reasonable to ask whether
n[g(X n ) g()] tends to some distribution as n . But the linearity of g(t) allows one
to write





n g(X n ) g() = a n X n .
116

We conclude by Theorem 2.27 that

 d

n g(X n ) g() aZ.
Of course, the distribution on the right hand side above is N (0, a2 2 ).
None of the preceding development is especially deep; one might even say that it is obvious
that a linear transformation of the random variable X n alters its asymptotic distribution
by a constant multiple. Yet what if the function g(t) is nonlinear? It is in this nonlinear
case that a strong understanding of the argument above, as simple as it may be, pays real
dividends. For if X n is consistent for (say), then we know that, roughly speaking, X n
will be very close to for large n. Therefore, the only meaningful aspect of the behavior of
g(t) is its behavior in a small neighborhood of . And in a small neighborhood of , g()
may be considered to be roughly a linear function if we use a first-order Taylor expansion.
In particular, we may approximate
g(t) g() + g 0 ()(t )
for t in a small neighborhood of . We see that g 0 () is the multiple of t, and so the logic of
the linear case above suggests
d

(5.2)
n g(X n ) g() g 0 ()Z.
Indeed, expression (5.2) is a special case of the powerful theorem known as the delta method,
which we now state and prove:
d

Theorem 5.1 Delta method: If g 0 (a) exists and nb (Xn a) X for b > 0, then
d

nb {g(Xn ) g(a)} g 0 (a)X.

P

Proof: By Slutskys Theorem, Xn a 0 because Xn a = nb nb (Xn a) and nb (X) 0(X) =

0. Therefore, we may apply Theorem 2.8, which is Taylors theorem as it applies to random
variables. Taking d = 1 in Equation (2.5) gives
nb {g(Xn ) g(a)} = nb (Xn a) {g 0 (a) + oP (1)}
d

as n . Therefore, Slutskys theorem together with the fact that nb (Xn a) X proves
Theorem 5.1.
Expression (5.2) may be reexpressed as a corollary of Theorem 5.1:
Corollary 5.2 The often-used special case of Theorem 5.1 in which X is normally

d
distributed states that if g 0 () exists and n(X n ) N (0, 2 ), then
d



n g(X n ) g() N 0, 2 g 0 ()2 .
117

Ultimately, we will extend Theorem 5.1 in two directions: Theorem 5.5 deals with the special
case in which g 0 (a) = 0, and Theorem 5.6 is the multivariate version of the delta method. But
we first apply the delta method to a couple of simple examples that illustrate a principle that
we discussed in Section 4.1.4: When we speak of the asymptotic distribution of a sequence
of random variables, we generally refer to a nontrivial (i.e., nonconstant) distribution. For
example, in the case of an independent and identically distributed sequence X1 , X2 , . . . of
random variables with finite variance, the phrase asymptotic distribution of X n generally
refers to the fact that

d

n X n E X1 N (0, Var X1 ),
P

not the fact that X n E X1 .

2

Example 5.3 Asymptotic distribution of X n Suppose X1 , X2 , . . . are independent

and identically distributed with mean and finite variance 2 . Then by the
central limit theorem,

n(X n ) N (0, 2 ).

Therefore, the delta method gives

n(X n 2 ) N (0, 42 2 ).

(5.3)

However, this is not necessarily the end of the story. If = 0, then the normal

2
limit in (5.3) is degeneratethat is, expression (5.3) merely states that n(X n )
converges in probability to the constant 0. This is not what we mean by the
asymptotic distribution! Thus, we must treat the case = 0 separately, noting

d
in that case that nX n N (0, 2 ) by the central limit theorem, which implies
that
2 d

nX n 2 21 .
Example 5.4 Estimating binomial variance: Suppose Xn binomial(n, p). Because
Xn /n is the maximum likelihood estimator for p, the maximum likelihood estimator for p(1 p) is n = Xn (n Xn )/n2 . The central limit theorem tells us that

d
n(Xn /n p) N {0, p(1 p)}, so the delta method gives



d
n {n p(1 p)} N 0, p(1 p)(1 2p)2 .

Note that in the case p = 1/2, this does not give the asymptotic distribution of
n . Exercise 5.1 gives a hint about how to find the asymptotic distribution of n
in this case.
118

We have seen in the preceding examples that if g 0 (a) = 0, then the delta method gives
something other than the asymptotic distribution we seek. However, by using more terms
in the Taylor expansion, we obtain the following generalization of Theorem 5.1:
Theorem 5.5 If g(t) has r derivatives at the point a and g 0 (a) = g 00 (a) = =
d
g (r1) (a) = 0, then nb (Xn a) X for b > 0 implies that
d

nrb {g(Xn ) g(a)}

1 (r)
g (a)X r .
r!

It is straightforward using the multivariate notion of differentiability discussed in Definition

1.36 to prove the following theorem:
Theorem 5.6 Multivariate delta method: If g : Rk R` has a derivative g(a) at
a Rk and
d

nb (Xn a) Y
for some k-vector Y and some sequence X1 , X2 , . . . of k-vectors, where b > 0,
then
d

nb {g (Xn ) g (a)} [g(a)]> Y.

The proof of Theorem 5.6 involves a simple application of the multivariate Taylor expansion
of Equation (1.31).

5.1.2

Variance stabilizing transformations

Often, if E (Xi ) = is the parameter of interest, the central limit theorem gives

n(X n ) N {0, 2 ()}.

In other words, the variance of the limiting distribution is a function of . This is a problem
if we wish to do inference for , because ideally the limiting distribution should not depend
on the unknown . The delta method gives a possible solution: Since
d



n g(X n ) g() N 0, 2 ()g 0 ()2 ,
we may search for a transformation g(x) such that g 0 ()() is a constant. Such a transformation is called a variance stabilizing transformation.

119

Example 5.7 Suppose that X1 , X2 , . . . are independent normal random variables

with mean 0 and variance 2 . Let us define 2 = Var Xi2 , which for the normal
distribution may be seen to be 2 4 . (To verify this, try showing that E Xi4 = 3 4
by differentiating the normal characteristic function four times and evaluating at
zero.) Thus, Example 4.11 shows that
!
n
X

1
d
n
Xi2 2 N (0, 2 4 ).
n i=1
To do inference for 2 when we believe that our data are truly independent and
identically normally distributed, it would be helpful if the limiting distribution
did not depend on the unknown 2 . Therefore, it is sensible in light of Corollary
2 2 4
5.2 to search for a function g(t) such that 2[g 0 (
)] is not a function of 2 . In
other words, we want g 0 (t) to be proportional to t2 = |t|1 . Clearly g(t) = log t
is such a function. Therefore, we call the logarithm function a variance-stabilizing
function in this example, and Corollary 5.2 shows that
(
!
)
n

1X 2
d
n log
Xi log 2 N (0, 2).
n i=1

Exercises for Section 5.1

Exercise 5.1 Let n be defined as in Example 5.4. Find the asymptotic distribution
of n in the case p = 1/2. That is, find real-valued sequences an and bn and a
d
nontrivial random variable X such that an (n bn ) X.
Hint: Let Yn = Xn (n/2). Apply the central limit theorem to Yn , then
transform both sides of the resulting limit statement so that a statement involving
n results.
Exercise 5.2 Prove Theorem 5.5.
Exercise 5.3 Suppose Xn binomial(n, p), where 0 < p < 1.
(a) Find the asymptotic distribution of g(Xn /n)g(p), where g(x) = min{x, 1
x}.

(b) Show that h(x) = sin1 ( x) is a variance-stabilizing transformation for

Xn /n. This is called the arcsine transformation of a sample proportion.

Hint: (d/du) sin1 (u) = 1/ 1 u2 .

120

Exercise 5.4 Let X1 , X2 , . . . be independent from N (, 2 ) where 6= 0. Let

n

Sn2

1X
=
(Xi X n )2 .
n i=1

Find the asymptotic distribution of the coefficient of variation Sn /X n .

Exercise 5.5 Let Xn binomial(n, p), where p (0, 1) is unknown. Obtain confidence intervals for p in two different ways:

d
(a) Since n(Xn /np) N [0, p(1p)], the variance of the limiting distribution
P
depends only on p. Use the fact that Xn /n p to find a consistent estimator of
the variance and use it to derive a 95% confidence interval for p.
(b) Use the result of problem 5.3(b) to derive a 95% confidence interval for p.
(c) Evaluate the two confidence intervals in parts (a) and (b) numerically for
all combinations of n {10, 100, 1000} and p {.1, .3, .5} as follows: For 1000
realizations of X bin(n, p), construct both 95% confidence intervals and keep
track of how many times (out of 1000) that the confidence intervals contain p.
Report the observed proportion of successes for each (n, p) combination. Does
your study reveal any differences in the performance of these two competing
methods?

5.2

Sample Moments

The weak law of large numbers tells us that If X1 , X2 , . . . are independent and identically
distributed with E |X1 |k < , then
n

1X k P
X E X1k .
n i=1 i
That is, sample moments are (weakly) consistent. For example, the sample variance, which
we define as
n

s2n =

1X
1X 2
(Xi X n )2 =
X (X n )2 ,
n i=1
n i=1 i

is consistent for Var Xi = E Xi2 (E Xi )2 .

121

(5.4)

However, consistency is not the end of the story. The central limit theorem and the delta
method will prove very useful in deriving asymptotic distribution results about sample moments. We consider two very important examples involving the sample variance of Equation
(5.4).
Example 5.8 Distribution of sample T statistic: Suppose X1 , X2 , . . . are independent and identically distributed with E (Xi ) = and Var (Xi ) = 2 < . Define
s2n as in Equation (5.4), and let

n(X n )
Tn =
.
sn
Letting

n(X n )
An =

and Bn = /sn , we obtain Tn = An Bn . Therefore, since An N (0, 1) by the

P
central limit theorem and Bn 1 by the weak law of large numbers, Slutskys
d
theorem implies that Tn N (0, 1). In other words, T statistics are asymptotically
normal under the null hypothesis.
Example 5.9 Let X1 , X2 , . . . be independent and identically distributed with mean
, variance 2 , third central moment E (Xi )3 = , and Var (Xi
)2 =
2
2
< . Define Sn as in Equation (4.6). We have shown earlier that n(Sn2
d
2 ) N (0, 2 ). The same fact may be proven using Theorem 4.9 as follows.
First, let Yi = Xi and Zi = Yi2 . We may use the multivariate central limit
theorem to find the joint asymptotic distribution of Y n and Z n , namely
   
  2


Yn
0

d
n
N2 0,

.
2
Zn
2
Note that the above result uses the fact that Cov (Y1 , Z1 ) = .
We may write Sn2 = Z n (Y n )2 . Therefore, define the function g(a, b) = b a2
and observe that this gives g(a, b) = (2a, 1)> . To use the delta method, we
should evaluate
 2

 2
 
0

2 >
2
g(0, )
g(0, ) = ( 0 1 )
= 2
2
2


1
We conclude that
  
 

Yn
0
d
n g
g 2
= n(Sn2 2 ) N (0, 2 )
Zn

as we found earlier (using a different argument) in Example 4.11.

122

Exercises for Section 5.2

Exercise 5.6 Suppose that X1 , X2 , . . . are independent and identically distributed
Normal (0, 2 ) random variables.
(a) Based on the result of Example 5.7, Give an approximate test at = .05 for
H0 : 2 = 02 vs. Ha : 2 6= 02 .
(b) For n = 25, estimate the true level of the test in part (a) for 02 = 1 by
simulating 5000 samples of size n = 25 from the null distribution. Report the
proportion of cases in which you reject the null hypothesis according to your test
(ideally, this proportion will be about .05).

5.3

Sample Correlation

Suppose that (X1 , Y1 ), (X2 , Y2 ), . . . are independent and identically distributed vectors with
E Xi4 < and E Yi4 < . For the sake of simplicity, we will assume without loss of
generality that E Xi = E Yi = 0 (alternatively, we could base all of the following derivations
on the centered versions of the random variables).
We wish to find the asymptotic distribution of the sample correlation coefficient, r. If we let

Pn

X
mx
i
Pi=1
n

my

Pni=1 Yi2
1
mxx =

Xi
(5.5)

Pi=1

n
n
2
myy

Pni=1 Yi
mxy
i=1 Xi Yi
and
s2x = mxx m2x , s2y = myy m2y , and sxy = mxy mx my ,

(5.6)

then r = sxy /(sx sy ). According to the central limit theorem,

mx
0
0

Cov
(X
,
X
)

Cov
(X
,
X
Y
)
1
1
1
1 1

my 0

2 d
Cov (Y1 , X1 )

Cov (Y1 , X1 Y1 )

(. 5.7)
n
..
..
..
mxx x2 N5 0 ,

.
.
.

myy
y
0

Cov
(X
Y
,
X
)

Cov
(X
Y
,
X
Y
)
1
1
1
1
1
1
1
0
mxy
xy
Let denote the covariance matrix in expression (5.7). Define a function g : R5 R3 such
that g applied to the vector of moments in Equation (5.5) yields the vector (s2x , s2y , sxy ) as
123

defined in expression (5.6). Then

a
2a
0
b
b
0
2b a

g c = 1
0
0
.
d
0
1
0
e
0
0
1
Therefore, if we let
>

0
0

x g x2
=
g

y
xy
xy

2
2
Cov (X12 , X1 Y1 )
Cov (X12 , Y12 )
Cov (X1 , X1 )
Cov (Y12 , X1 Y1 ) ,
Cov (Y12 , Y12 )
= Cov (Y12 , X12 )
2
2
Cov (X1 Y1 , X1 ) Cov (X1 Y1 , Y1 ) Cov (X1 Y1 , X1 Y1 )

then by the delta method,

2 2
x

sx2
d
2

y
sy
N3 (0, ).

xy
sxy

(5.8)

As
note that expression (5.8) gives the same marginal asymptotic distribution for
an2 aside,
2
n(sx x ) as was derived using a different approach in Example 4.11, since Cov (X12 , X12 )
is the same as 2 in that example.

Next, define the function h(a, b, c) = c/ ab, so that we have h(s2x , s2y , sxy ) = r. Then


1
c
c
2
>
, ,
[h(a, b, c)] =
,
2
ab
a3 b ab3
so that
[h(x2 , y2 , xy )]>


=

1
xy xy
,
,
2x3 y 2x y3 x y


=

1
,
,
2x2 2y2 x y


.

(5.9)

Therefore, if A denotes the 1 3 matrix in Equation (5.9), using the delta method once
again yields

d
n(r ) N (0, A A> ).
To recap, we have used the basic tools of the multivariate central limit theorem and the
multivariate delta method to obtain a univariate result. This derivation of univariate facts
via multivariate techniques is common practice in statistical large-sample theory.
124

Example 5.10 Consider the special case of bivariate normal (Xi , Yi ). In this case,
we may derive

2x3 y
22 x2 y2
2x4
.
2y4
2x y3
(5.10)
= 22 x2 y2
3
3
2
2 2
2x y 2x y (1 + )x y
In this case, A A> = (1 2 )2 , which implies that

d
n(r ) N {0, (1 2 )2 }.

(5.11)

In the normal case, we may derive a variance-stabilizing transformation. According to Equation (5.11), we should find a function f (x) satisfying f 0 (x) = (1x2 )1 .
Since
1
1
1
=
+
,
2
1x
2(1 x) 2(1 + x)
we integrate to obtain
f (x) =

1
1+x
log
.
2
1x

This is called Fishers transformation; we conclude that




1
1+r 1
1+ d
n
log
log
N (0, 1).
2
1r 2
1

Exercises for Section 5.3

Exercise 5.7 Verify expressions (5.10) and (5.11).
Exercise 5.8 Assume (X1 , Y1 ), . . . , (Xn , Yn ) are independent and identically distributed
from some bivariate normal distribution. Let denote the population correlation
coefficient and r the sample correlation coefficient.
(a) Describe a test of H0 : = 0 against H1 : 6= 0 based on the fact that

d
n[f (r) f ()] N (0, 1),
where f (x) is Fishers transformation f (x) = (1/2) log[(1 + x)/(1 x)]. Use
= .05.
(b) Based on 5000 repetitions each, estimate the actual level for this test in the
case when E (Xi ) = E (Yi ) = 0, Var (Xi ) = Var (Yi ) = 1, and n {3, 5, 10, 20}.
125

Exercise 5.9 Suppose that X and Y are jointly distributed such that X and Y
are Bernoulli (1/2) random variables with P (XY = 1) = for (0, 1/2).
Let (X1 , Y1 ), (X2 , Y2 ), . . . be independent and identically distributed with (Xi , Yi )
distributed as (X, Y ).


(a) Find the asymptotic distribution of n (X n , Y n ) (1/2, 1/2) .
(b) If rn is the sample correlation
coefficient for a sample of size n, find the

asymptotic distribution of n(rn ).

(c) Find a variance stabilizing transformation for rn .
(d) Based on your answer to part (c), construct a 95% confidence interval for .
(e) For each combination of n {5, 20} and {.05, .25, .45}, estimate the true
coverage probability of the confidence interval in part (d) by simulating 5000
samples and the corresponding confidence intervals. One problem you will face
is that in some samples, the sample correlation coefficient is undefined because
with positive probability each of the Xi or Yi will be the same. In such cases,
consider the confidence interval to be undefined and the true parameter therefore
not contained therein.
Hint: To generate a sample of (X, Y ), first simulate the Xs from their marginal
distribution, then simulate the Y s according to the conditional distribution of
Y given X. To obtain this conditional distribution, find P (Y = 1 | X = 1) and
P (Y = 1 | X = 0).

126

Chapter 6
Order Statistics and Quantiles
Consider an ordering function on n real numbers, a vector-valued function f n that maps
Rn Rn so that if we let y = f n (x), then the values y1 , . . . , yn are simply a permutation
of the values x1 , . . . , xn such that y1 yn . In this chapter, we consider the order
statistics, which are the result of applying this ordering function to a simple random sample
X1 , . . . , X n .
We introduce a specialized notation for these random variables, which we call the order
statistics. Given a finite sample X1 , . . . , Xn , define the values X(1) , . . . , X(n) to be a permutation of X1 , . . . , Xn such that X(1) X(2) X(n) . We call X(i) the ith order statistic
of the sample.
Even though the notation X(i) does not explicitly use the sample size n, the distribution of
X(i) depends essentially on n. For this reason, some textbooks use slightly more complicated
notation such as
X(1:n) , X(2;n) , . . . , X(n;n)
for the order statistics of a sample. We choose to use the simpler notation here, though it is
important to remember that we will always understand the sample size to be n.

6.1

Extreme Order Statistics

The asymptotic distributions of order statistics at the extremes of a sample may be derived
without any specialized knowledge other than the limit formula

c n
ec as n
(6.1)
1+
n
127

and its generalization


b
cn n
1+
ec if cn c and bn
bn

(6.2)

(see Example 1.20). Recall that by asymptotic distribution of X(1) , we mean sequences an
d

and bn , along with a nondegenerate random variable X, such that an (X(1) bn ) X. This
section consists mostly of a series of illustrative examples.
Example 6.1 Suppose X1 , . . . , Xn are independent and identically distributed uniform(0,1) random variables. What is the asymptotic distribution of X(n) ?
Since X(n) t if and only if X1 t, X2 t, . . . , and Xn t, by independence
we have
(
0 if t 0
P (X(n) t) = tn if 0 < t < 1
(6.3)
1 if t 1.
P

From Equation (6.3), it is apparent that X(n) 1, though this limit statement
does not fully reveal the asymptotic distribution of X(n) . We desire sequences
an and bn such that an (X(n) bn ) has a nondegenerate limiting distribution.
Evidently, we should expect bn = 1, a fact we shall rederive below.
Computing the distribution function of an (X(n) bn ) directly, we find


u
+ bn
F (u) = P {an (X(n) bn ) u} = P X(n)
an
as long as an > 0. Therefore, we see that

n
u
u
F (u) =
+ bn
+ bn < 1.
for 0 <
an
an

(6.4)

We would like this expression to tend to a limit involving only u as n .

Keeping expression 6.2 in mind, we take bn = 1 and an = n so that F (u) =
(1 + u/n)n , which tends to eu .
However, we are not quite finished, since we have not determined which values
of u make the above limit valid. Equation 6.4 required that 0 < bn + (u/an ) < 1,
which in this case becomes 1 < u/n < 0. This means u may be any negative
real number, since for any u < 0, 1 < u/n < 0 for all n > |u|. We conclude
that if the random variable U has distribution function

exp(u) if u 0
F (u) =
1
if u > 0,
128

then n(X(n) 1) U . Since U is simply a standard exponential random variable,

we may also write
d

n(1 X(n) ) Exponential(1).

Example 6.2 Suppose X1 , X2 , . . . are independent and identically distributed exponential random variables with mean 1. What is the asymptotic distribution of
X(n) ?
As in Equation 6.3, if u/an + bn > 0 then
 

n



u
u
+ bn = 1 exp bn
P an (X(n) bn ) u = P X(n)
.
an
an
Taking bn = log n and an = 1, the rightmost expression above simplifies to
n

eu
,
1
n
which has limit exp(eu ). The condition u/an + bn > 0 becomes u + log n > 0,
which is true for all u R as long as n > exp(u). Therefore, we conclude that
d
X(n) log n U , where
def

P (U u) = exp{ exp(u)} for all u.

(6.5)

The distribution of U in Equation 6.5 is known as the extreme value distribution

or the Gumbel distribution.
In Examples 6.1 and 6.2, we derived the asymptotic distribution of a maximum from a simple
random sample. We did this using only the definition of convergence in distribution without
relying on any results other than expression 6.2. In a similar way, we may derive the joint
asymptotic distribution of multiple order statistics, as in the following example.
Example 6.3 Range of uniform sample: Let X1 , . . . , Xn be a simple random sample
from Uniform(0, 1). Let Rn = X(n) X(1) denote the range of the sample. What
is the asymptotic distribution of Rn ?
To answer this question, we begin by finding the joint asymptotic distribution of
(X(n) , X(1) ), as follows. For sequences an and bn , as yet unspecified, consider
P (an X(1) > x and bn (1 X(n) ) > y) = P (X(1) > x/an and X(n) < 1 y/bn )
= P (x/an < X(1) < < X(n) < 1 y/bn ),

129

where we have assumed that an and bn are positive. Since the probability above
is simply the probability that the entire sample is to be found in the interval
(x/an , 1 y/bn ), we conclude that as long as
0<

x
y
<1
< 1,
an
bn

(6.6)

we have

P (an X(1) > x and bn (1 X(n) ) > y) =

y
x
1

b n an

n
.

Expression (6.1) suggests that we set an = bn = n, resulting in


y
x n
P (nX(1) > x and n(1 X(n) ) > y) = 1
.
n n
Expression 6.6 becomes
0<

y
x
< 1 < 1,
n
n

which is satisfied for large enough n if and only if x and y are both positive. We
conclude that for x > 0, y > 0,
P (nX(1) > x and n(1 X(n) ) > y) ex ey .
Since this is the joint distribution of independent standard exponential random
variables, say, Y1 and Y2 , we conclude that

  
Y1
nX(1)
d

.
Y2
n(1 X(n) )
Therefore, applying the continuous function f (a, b) = a + b to both sides gives
d

n(1 X(n) + X(1) ) = n(1 Rn ) Y1 + Y2 Gamma(2, 1).

Let us consider a different example in which the asymptotic joint distribution does not
involve independent random variables.
Example 6.4 As in Example 6.3, let X1 , . . . , Xn be independent and identically distributed from uniform(0, 1). What is the joint asymptotic distribution of X(n1)
and X(n) ?
Proceeding as in Example 6.3, we obtain

  

n(1 X(n1) )
x
x
y
. (6.7)
P
>
= P X(n1) < 1 and X(n) < 1
n(1 X(n) )
y
n
n
130

We consider two separate cases: If 0 < x < y, then the right hand side of (6.7)
is simply P (X(n) < 1 y/n), which converges to ey as in Example 6.1. On the
other hand, if 0 < y < x, then


x
y
x
P X(n1) < 1 and X(n) < 1
= P X(n) < 1
n
n
n

x
y
+P X(n1) < 1 < X(n) < 1
n
n


x n
x n1  x y 

=
1
+n 1
n
n
n n
ex (1 + x y).
The second equality above arises because X(n1) < a < X(n) < b if and only if
exactly n1 of the Xi are less than a and exactly one is between a and b. We now
know the joint asymptotic distribution of n(1 X(n1) ) and n(1 X(n1) ); but
can we describe this joint distribution in a simple way? Suppose that Y1 and Y2
are independent standard exponential variables. Consider the joint distribution
of Y1 and Y1 + Y2 : If 0 < x < y, then
P (Y1 + Y2 > x and Y1 > y) = P (Y1 > y) = ey .
On the other hand, if 0 < y < x, then
P (Y1 + Y2 > x and Y1 > y) = P (Y1 > max{y, x Y2 }) = E e max{y,xY2 }
Z xy
y
= e P (y > x Y2 ) +
etx et dt
0

= ex (1 + x y).
Therefore, we conclude that


 
n(1 X(n1) ) d Y1 + Y2

.
Y1
n(1 X(n) )
d

Notice that marginally, we have shown that n(1 X(n1) ) Gamma(2, 1).
Recall that if F is a continuous, invertible distribution function and U is a standard uniform
random variable, then F 1 (U ) F . The proof is immediate, since P {F 1 (U ) t} =
P {U F (t)} = F (t). We may use this fact in conjunction with the result of Example 6.4
as in the following example.
Example 6.5 Suppose X1 , . . . , Xn are independent standard exponential random variables. What is the joint asymptotic distribution of (X(n1) , X(n) )?

131

The distribution function of a standard exponential distribution is F (t) = 1et ,

whose inverse is F 1 (u) = log(1 u). Therefore,

 

log(1 U(n1) ) d X(n1)
=
,
log(1 U(n) )
X(n)
d

where = means has the same distribution. Thus,






log n(1 U(n1) )
X(n1) log n
d


=
.
X(n) log n
log n(1 U(n) )
We conclude by the result of Example 6.4 that

 

X(n1) log n d log(Y1 + Y2 )

,
X(n) log n
log Y1
where Y1 and Y2 are independent standard exponential variables.

Exercises for Section 6.1

Exercise 6.1 For a given n, let X1 , . . . , Xn be independent and identically distributed
with distribution function
P (Xi t) =

t3 + 3
23

for t [, ].

Let X(1) denote the first order statistic from the sample of size n; that is, X(1) is
the smallest of the Xi .
(a) Prove that X(1) is consistent for .
(b) Prove that
d

n( + X(1) ) Y,
where Y is a random variable with an exponential distribution. Find E (Y ) in
terms of .
(c) For a fixed , define


,n = 1 +
X(1) .
n
Find, with proof, such that
d

n ( ,n ) Y E (Y ),
132

where Y is the same random variable as in part (b).

(d) Compare the two consistent -estimators ,n and X(1) empirically as
follows. For n {102 , 103 , 104 }, take = 1 and simulate 1000 samples of size
n from the distribution of Xi . From these 1000 samples, estimate the bias and
mean squared error of each estimator. Which of the two appears better? Do your
empirical results agree with the theoretical results in parts (c) and (d)?
Exercise 6.2 Let X1 , X2 , . . . be independent uniform (0, ) random variables. Let
X(n) = max{X1 , . . . , Xn } and consider the three estimators

2
n
n
2
0
1
X(n) n =
n = X(n) n =
X(n) .
n1
n1
(a) Prove that each estimator is consistent for .
(b) Perform an empirical comparison of these three estimators for n = 102 , 103 , 104 .
Use = 1 and simulate 1000 samples of size n from uniform (0, 1). From these
1000 samples, estimate the bias and mean squared error of each estimator. Which
one of the three appears to be best?
(c) Find the asymptotic distribution of n( ni ) for i = 0, 1, 2. Based on your
results, which of the three appears to be the best estimator and why? (For the
latter question, dont attempt to make a rigorous mathematical argument; simply
give an educated guess.)
Exercise 6.3 Find, with proof, the asymptotic distribution of X(n) if X1 , . . . , Xn are
independent and identically distributed with each of the following distributions.
(That is, find an , bn , and a nondegenerate random variable X such that an (X(n)
d

bn ) X.)
(a) Beta(3, 1) with distribution function F (x) = x2 for x (0, 1).
(b) Standard logistic with distribution function F (x) = ex /(1 + ex ).
Exercise 6.4 Let X1 , . . . , Xn be independent
uniform(0, 1) random
variables. Find


the joint asymptotic distribution of nX(2) , n(1 X(n1) ) .
Hint: To find a probability such as P (a < X(2) < X(n) < b), consider the
trinomial distribution with parameters [n; (a, b a, 1 b)] and note that the
probability in question is the same as the probability that the numbers in the
first and third categories are each 1.
Exercise 6.5 Let X1 , . . . , Xn be a simple random sample from the distribution function F (x) = [1 (1/x)]I{x > 1}.
133

(a) Find the joint asymptotic distribution of (X(n1) /n, X(n) /n).
Hint: Proceed as in Example 6.5.
(b) Find the asymptotic distribution of X(n1) /X(n) .
Exercise 6.6 If X1 , . . . , Xn are independent and identically distributed uniform(0, 1)
d
variables, prove that X(1) /X(2) uniform(0, 1).
Exercise 6.7 Let X1 , . . . , Xn be a simple random sample from a logistic distribution
with distribution function F (t) = et/ /(1 + et/ ) for all t.
(a) Find the asymptotic distribution of X(n) X(n1) .
Hint: Use the fact that log U(n) and log U(n1) both converge in probability to
zero.
(b) Based on part (a), construct an approximate 95% confidence interval for
. Use the fact that the .025 and .975 quantiles of the standard exponential
distribution are 0.0253 and 3.6889, respectively.
(c) Simulate 1000 samples of size n = 40 with = 2. How many confidence
intervals contain ?

6.2

Sample Quantiles

To derive the distribution of sample quantiles, we begin by obtaining the exact distribution
of the order statistics of a random sample from a uniform distribution. To facilitate this
derivation, we begin with a quick review of changing variables. Suppose X has density fX (x)
and Y = g(X), where g : Rk Rk is differentiable and has a well-defined inverse, which we
denote by h : Rk Rk . (In particular, we have X = h[Y].) The density for Y is
fY (y) = |Det[h(y)]| fX [h(y)],

(6.8)

where |Det[h(y)]| is the absolute value of the determinant of the k k matrix h(y).

6.2.1

Uniform Order Statistics

We now show that the order statistics of a uniform distribution may be obtained using ratios
of gamma random variables. Suppose X1 , . . . , Xn+1 are independent standard exponential,

134

or Gamma(1, 1), random variables. For j = 1, . . . , n, define

Pj
i=1 Xi
Yj = Pn+1
.
i=1 Xi

(6.9)

We will show that the joint distribution of (Y1 , . . . , Yn ) is the same as the joint distribution of
the order statistics (U(1) , . . . , U(n) ) of a simple random sample from uniform(0, 1) by demonstrating that their joint density function is the same as that of the uniform order statistics,
namely n!I{0 < u(1) < . . . < u(n) < 1}.
We
Pj derive the joint density of (Y1 , . . . , Yn ) as follows. As an intermediate step, define Zj =
i=1 Xi for j = 1, . . . , n + 1. Then

Zi
if i = 1
Xi =
Zi Zi1 if i > 1,
which means that the gradient of the transformation from Z to X is upper triangular with
ones on the diagonal, a matrix whose determinant is one. This implies that the density for
Z is
fZ (z) = exp{zn+1 }I{0 < z1 < z2 < < zn+1 }.
Next, if we define Yn+1 = Zn+1 , then we may express Z in terms of Y as

Yn+1 Yi if i < n + 1
Zi =
Yn+1
if i = n + 1.

(6.10)

The gradient of the transformation in Equation (6.10) is lower triangular, with yn+1 along
n
,
the diagonal except for a 1 in the lower right corner. The determinant of this matrix is yn+1
so the density of Y is
n
fY (y) = yn+1
exp{yn+1 }I{yn+1 > 0}I{0 < y1 < < yn < 1}.

(6.11)

Equation (6.11) reveals several things: First, (Y1 , . . . , Yn ) is independent of Yn+1 and the
marginal distribution of Yn+1 is Gamma(n + 1, 1). More important for our purposes, the
marginal joint density of (Y1 , . . . , Yn ) is proportional to I{0 < y1 < < yn < 1}, which is
exactly what we needed to prove. We conclude that the vector Y defined in Equation (6.9)
has the same distribution as the vector of order statistics of a simple random sample from
uniform(0, 1).

6.2.2

Uniform Sample Quantiles

Using the result of Section 6.2.1, we may derive the joint asymptotic distribution of a set of
sample quantiles for a uniform simple random sample.
135

Suppose we are interested in the p1 and p2 quantiles, where 0 < p1 < p2 < 1. The following
argument may be generalized to obtain the joint asymptotic distribution of any finite number
of quantiles. If U1 , . . . , Un are independent uniform(0, 1) random variables, then the p1 and
p2 sample quantiles may be taken to be the an th and bn th order statistics, respectively, where
def
def
an = b.5 + np1 c and bn = b.5 + np2 c (b.5 + xc is simply x rounded to the nearest integer).
Next, let
a

def

An =

n
1X
Xi ,
n i=1

def

Bn =

bn
1 X
Xi ,
n i=a +1

def

and Cn =

n+1
1 X
Xi .
n i=b +1
n

We proved in Section 6.2.1 that (U(an ) , U(bn ) ) has the same distribution as


An + Bn
An
def
.
,
g(An , Bn , Cn ) =
An + Bn + Cn An + Bn + Cn

(6.12)

The asymptotic distribution of g(An , Bn , Cn ) may be determined using the delta method if
we can determine the joint asymptotic distribution of (An , Bn , Cn ).
A bit of algebra reveals that
r
r





an
an
an np1
nAn np1
nAn
.
n (An p1 ) =
an

=
an
1 +
n
an
an
n
an
n

d
By the central limit theorem, an (nAn /an 1) N (0, 1) because the Xi have mean 1 and
variance 1. Furthermore, an /n p1 and the rightmost term above goes to 0, so Slutskys
theorem gives

n (An p1 ) N (0, p1 ).

Similar arguments apply to Bn and to Cn . Because An and Bn and Cn are independent of

one another, we may stack them as in Exercise 2.23 to obtain

p1
p1
0
0

An
d

Bn
p2 p1
0 p2 p1
0
n

N3
0 ,

Cn
1 p2
0
0
1 p2
0
by Slutskys theorem. For g : R3 R2 defined in Equation (6.12), we obtain

b+c
c
1
a
c .
g(a, b, c) =
(a + b + c)2
a a b

136

Therefore,


>

[g(p1 , p2 p1 , 1 p2 )] =

1 p1
1 p2

p1
1 p2

p1
p2


,

so the delta method gives


  
  


U(an )
p1
p1 (1 p1 ) p1 (1 p2 )
0
d
n

N2
,
.
p1 (1 p2 ) p2 (1 p2 )
0
U(bn )
p2

(6.13)

The method used above to derive the joint distribution (6.13) of two sample quantiles may
be extended to any number of quantiles; doing so yields the following theorem:
Theorem 6.6 Suppose that for given constants p1 , . . . , pk with 0 < p1 < < pk < 1,
we define sequences {a1n }, . . . , {akn } such that for all 1 i k,

 ain
n
pi 0.
n
Then if U1 , . . . , Un is a sample from Uniform(0,1),

U(a1n )
p1 (1 p1 ) p1 (1 pk )
p1
0

.
d
.
..
..
.
n ... .. Nk .. ,
.
.

0
pk
p1 (1 pk ) pk (1 pk )
U(akn )
Note that for i < j, both the (i, j) and (j, i) entries in the covariance matrix above equal
pi (1 pj ) and pj (1 pi ) never occurs in the matrix.

6.2.3

General sample quantiles

Let F (x) be the distribution function for a random variable X. The quantile function F (u)
of Definition 3.18 is nondecreasing on (0, 1) and it has the property that F (U ) has the same
distribution as X for U uniform(0, 1). This property follows from Lemma 3.19, since
P [F (U ) x] = P [U F (x)] = F (x)
for all x. Since F (u) is nondecreasing, it preserves ordering; thus, if X1 , . . . , Xn is a random
sample from F (x), then

d 
(X(1) , . . . , X(n) ) = F (U(1) ), . . . , F (U(n) ) .
d

(The symbol = means has the same distribution as.)

137

Now, suppose that at some point , the derivative F 0 () exists and is positive. Then F (x)
must be continuous and strictly increasing in a neighborhood of . This implies that in this
neighborhood, F (x) has a well-defined inverse, which must be differentiable at the point
def
p = F (). If F 1 (u) denotes the inverse that exists in a neighborhood of p, then
dF 1 (p)
1
= 0 .
dp
F ()

(6.14)

Equation (6.14) may be derived by differentiating the equation F [F 1 (p)] = p. Note also
that whenever the inverse F 1 (u) exists, it must coincide with the quantile function F (u).
Thus, the condition that F 0 () exists and is positive is a sufficient condition to imply that
F (u) is differentiable at p. This differentiability is important: If we wish to transform the
uniform order statistics U(ain ) of Theorem 6.6 into order statistics X(ain ) using the quantile
function F (u), the delta method requires the differentiability of F (u) at each of the points
p1 , . . . , p k .
The delta method, along with Equation (6.14), yields the following corollary of Theorem 6.6:
Theorem 6.7 Let X1 , . . . , Xn be a simple random sample from a distribution function
F (x) such that F (x) is differentiable at each of the points 1 < < k and
F 0 (i ) > 0 for all i. Denote F (i ) by pi . Then under the assumptions of Theorem
6.6,
p1 (1p1 )

k)
F 0p(1 (1p

X(a1n )
0
1
0
F 0 (1 )2

1 )F (k )

.
d

.
.
.
.
.

.
.
..
..
n

.
. Nk . ,
.

p
(1p
)
p
(1p
)
1
k
k
k
0
k
X(akn )

F 0 (1 )F 0 (k )
F 0 (k )2

Exercises for Section 6.2

Exercise 6.8 Let X1 , . . . , Xn be independent uniform(0, 2) random variables.
(a) Let M = (X(1) + X(n) )/2. Find the asymptotic distribution of n(M ).
(b) Compare the asymptotic performance of the three estimators M , X n , and
n by considering their relative efficiencies.
the sample median X
(c) For n {101, 1001, 10001}, generate 500 samples of size n, taking = 1.
n for each sample. Construct a 3 3 table in which
Keep track of M , X n , and X
you report the sample variance of each estimator for each value of n. Do your
simulation results agree with your theoretical results in part (b)?

138

Exercise 6.9 Let X1 be Uniform(0, 2) and let X2 be standard

exponential,

independent of X1 . Find the joint distribution of (Y1 , Y2 ) = ( 2X2 cos X1 , 2X2 sin X1 ).
Note: Since log U has a standard exponential distribution if U uniform(0, 1),
this problem may be used to simulate normal random variables using simulated
uniform random variables.
Exercise 6.10 Suppose X1 , . . . , Xn is a simple random sample from a distribution
that is symmetric about , which is to say that P (Xi x) = F (x ), where
F (x) is the distribution function for a distribution that is symmetric about zero.
We wish to estimate by (Qp + Q1p )/2, where Qp and Q1p are the p and 1 p
sample quantiles, respectively. Find the smallest possible asymptotic variance for
the estimator and the p for which it is achieved for each of the following forms of
F (x):
(a) Standard Cauchy
(b) Standard normal
(c) Standard double exponential
Hint: For at least one of the three parts of this question, you will have to solve
for a minimizer numerically.
Exercise 6.11 When we use a boxplot to assess the symmetry of a distribution, one
of the main things we do is visually compare the lengths of Q3 Q2 and Q2 Q1 ,
where Qi denotes the ith sample quartile.
(a) Given a random sample of size n from N (0, 1), find the asymptotic distribution of (Q3 Q2 ) (Q2 Q1 ).
(b) Repeat part (a) if the sample comes from a standard logistic distribution.
(c) Using 1000 simulations from each distribution, use graphs to assess the
accuracy of each of the asymptotic approximations above for n = 5 and n = 13.
(For a sample of size 4k + 1, define Qi to be the (ik + 1)th order statistic.) For
each value of n and each distribution, plot the empirical distribution function
against the theoretical limiting distribution function.
Exercise 6.12 Let X1 , . . . , Xn be a random sample from Uniform(0, 2). Find the
asymptotic distributions of the median, the midquartile range, and 32 Q3 , where
Q3 denotes the third quartile and the midquartile range is the mean of the 1st
and 3rd quartiles. Compare these three estimates of based on their asymptotic
variances.

139

Chapter 7
Maximum Likelihood Estimation
7.1

Consistency

If X is a random variable (or vector) with density or mass function f (x) that depends on
a parameter , then the function f (X) viewed as a function of is called the likelihood
function of . We often denote this function by L(). Note that L() = f (X) is implicitly
a function of X, but we suppress this fact in the notation. Since repeated references to
the density or mass function would be awkward, we will use the term density to refer
to f (x) throughout this chapter, even if the distribution function of X is not continuous.
(Allowing noncontinuous distributions to have density functions may be made technically
rigorous; however, this is a measure theoretic topic beyond the scope of this book.)
then
Let the set of possible values of be the set . If L() has a maximizer in , say ,
is called a maximum likelihood estimator or MLE. Since the logarithm function is a strictly
def
increasing function, any maximizer of L() also maximizes `() = log L(). It is often much
easier to maximize `(), called the loglikelihood function, than L().
Example 7.1 Suppose = (0, ) and X binomial(n, e ). Then
 
n
`() = log
X + (n X) log(1 e ),
X
so
`0 () = X +

X n
.
1 e

Thus, setting `0 () = 0 yields = log(X/n). It isnt hard to verify that

`00 () < 0, so that log(X/n) is in fact a maximizer of `().
140

As the preceding example demonstrates, it is not always the case that a MLE existsfor if
X = 0 or X = n, then log(X/n) is not contained in . This is just one of the technical
details that we will consider. Ultimately, we will show that the maximum likelihood estimator
is, in many cases, asymptotically normal. However, this is not always the case; in fact, it is
not even necessarily true that the MLE is consistent, as shown in Problem 7.1.
We begin the discussion of the consistency of the MLE by defining the so-called KullbackLeibler divergence.
Definition 7.2 If f0 (x) and f1 (x) are two densities, the Kullback-Leibler divergence
from f0 to f1 equals
K(f0 , f1 ) = E 0 log

f0 (X)
.
f1 (X)

If P0 (f0 (X) > 0 and f1 (X) = 0) > 0, then K(f0 , f1 ) is defined to be .

The Kullback-Leibler divergence is sometimes called the Kullback-Leibler information number or the relative entropy of f1 with respect to f0 . Although it is nonnegative, and takes
the value zero if and only if f1 (x) = f0 (x) except possibly for a set of x values having measure zero, The K-L divergence is not a true distance because K(f0 , f1 ) is not necessarily
the same as K(f1 , f0 ).
We may show that the Kullback-Leibler information must be nonnegative by noting that
E 0

f1 (X)
= E 1 I{f0 (X) > 0} 1.
f0 (X)

Therefore, by Jensens inequality (1.37) and the strict convexity of the function log x,
K(f0 , f1 ) = E 0 log

f (X)
f1 (X)
log E 0 1
0,
f0 (X)
f0 (X)

(7.1)

with equality if and only if P0 {f0 (X) = f1 (X)} = 1. Inequality (7.1) is sometimes called
the Shannon-Kolmogorov information inequality.
In the (admittedly somewhat bizarre) case in which the parameter space contains only
finitely many points, the Shannon-Kolmogorov information inequality may be used to prove
the consistency of the maximum likelihood estimator. For the proof of the following theorem,
note that if X1 , . . . , Xn areP
independent and identically distributed with density f0 (x), then
the loglikelihood is `() = ni=1 log f0 (xi ).
Theorem 7.3 Suppose contains finitely many elements and that X1 , . . . , Xn are
independent and identically distributed with density f0 (x). Furthermore, suppose that the model parameter is identifiable, which is to say that different values
of lead to different distributions. Then if n denotes the maximum likelihood
P
estimator, n 0 .
141

Proof: The Weak Law of Large Numbers (Theorem 2.19) implies that
n

1X
f (Xi ) P
f (Xi )
log
= K(f0 , f )
E 0 log
n i=1
f0 (Xi )
f0 (Xi )

(7.2)

for all . The value of K(f0 , f ) is strictly negative for 6= 0 by the identifiability of
. Therefore, since = n is the maximizer of the left hand side of Equation (7.2),
!
!
n
n
X
X
X
f
(X
)
f
(X
)
1
1

i
log
>0
P
log
> 0 0.
P (n 6= 0 ) = P max
6=0 n
f
(X
)
n
f
(X
)

i
0
0
i=1
i=1
6=
0

P
This implies that n 0 .

The result of Theorem 7.3 may be extended in several ways; however, it is unfortunately not
true in general that a maximum likelihood estimator is consistent, as demonstrated by the
example of Problem 7.1.
If we return to the simple Example 7.1, we found that the MLE was found by solving the
equation
`0 () = 0.

(7.3)

Equation (7.3) is called the likelihood equation, and naturally a root of the likelihood equation is a good candidate for a maximum likelihood estimator. However, there may be no
root and there may be more than one. It turns out the probability that at least one root
exists goes to 1 as n . Consider Example 7.1, in which no MLE exists whenever X = 0
or X = n. In that case, both P (X = 0) = (1 e )n and P (X = n) = en go to zero as
n . In the case of multiple roots, one of these roots is typically consistent for 0 , as
stated in the following theorem.
Theorem 7.4 Suppose that X1 , . . . , Xn are independent and identically distributed
with density f0 (x) for 0 in an open interval R, where the parameter is identifiable (i.e., different values of give different distributions). Furthermore,
suppose that the loglikelihood function `() is differentiable and that the support
{x : f (x) > 0} does not depend on . Then with probability approaching 1 as
P
n , there exists n = n (X1 , . . . , Xn ) such that `0 (n ) = 0 and n 0 .
Stated succinctly, Theorem 7.4 says that under certain regularity conditions, there is a
consistent root of the likelihood equation. It is important to note that there is no guarantee
that this consistent root is the MLE. However, if the likelihood equation only has a single
root, we can be more precise:

142

Corollary 7.5 Under the conditions of Theorem 7.4, if for every n there is a unique
root of the likelihood equation, and this root is a local maximum, then this root
is the MLE and the MLE is consistent.
Proof: The only thing that needs to be proved is the assertion that the unique root is
the MLE. Denote the unique root by n and suppose there is some other point such that
`() `(n ). Then there must be a local minimum between n and , which contradicts the
assertion that n is the unique root of the likelihood equation.

Exercises for Section 7.1

Exercise 7.1 In this problem, we explore an example in which the MLE is not consistent. Suppose that for (0, 1), X is a continuous random variable with
density


x
1
h
,
(7.4)
f (x) = g(x) +
()
()
where () > 0 for all , g(x) = I{1 < x < 1}/2, and
3(1 x2 )
h(x) =
I{1 < x < 1}.
4
(a) What condition on () ensures that {x : f (x) > 0} does not depend on ?
(b) With () = exp{(1 )4 }, let = .2. Take samples of sizes n
{50, 250, 500} from f (x). In each case, graph the loglikelihood function and find
the MLE. Also, try to identify the consistent root of the likelihood equation in
each case.
Hints: To generate a sample from f (x), note that f (x) is a mixture density,
which means you can start by generating a standard uniform random variable.
If its less than , generate a uniform variable on (1, 1). Otherwise, generate a
variable with density 3( 2 x2 )/4 3 on (, ) and then add . You should be able
to do this by inverting the distribution function or by using appropriately scaled
and translated beta(2, 2) variables. If you use the inverse distribution function
method, verify that


4 1
1
1
+ cos (1 2u) .
H (u) = 2 cos
3
3
Be very careful when graphing the loglikelihood and finding the MLE. In particular, make sure you evaluate the loglikelihood analytically at each of the sample
143

points in (0, 1) and incorporate these analytical calculations into your code; if
you fail to do this, youll miss the point of the problem and youll get the MLE
incorrect. This is because the correct loglikelhood graph will have tall, extremely
narrow spikes.
Exercise 7.2 In the situation of Exercise 7.1, prove that the MLE is inconsistent.
Exercise 7.3 Suppose that X1 , . . . , Xn are independent and identically distributed
with density f (x), where (0, ). For each of the following forms of f (x),
prove that the likelihood equation has a unique solution and that this solution
maximizes the likelihood function.
(a) Weibull: For some constant a > 0,
f (x) = aa xa1 exp{(x)a }I{x > 0}
(b) Cauchy:
f (x) =

2
x + 2

(c)

32 3
f (x) =
I{x > 0}
2(x3 + 3 )
Exercise 7.4 Find the MLE and its asymptotic distribution given a random sample
of size n from f (x) = (1 )x , x = 0, 1, 2, . . ., (0, 1).
Hint: For the asymptotic distribution, use the central limit theorem.

7.2

Asymptotic normality of the MLE

As seen in the preceding section, the MLE is not necessarily even consistent, let alone
asymptotically normal, so the title of this section is slightly misleadinghowever, Asymptotic normality of the consistent root of the likelihood equation is a bit too long! It will be
necessary to review a few facts regarding Fisher information before we proceed.
Definition 7.6 Fisher information: For a density (or mass) function f (x), the
Fisher information function is given by
2

d
log f (X) .
(7.5)
I() = E
d
144

If = g() for some invertible and differentiable function g(), then since
d
d d
1 d
=
= 0
d
d d
g () d
by the chain rule, we conclude that
I() =

I()
.
{g 0 ()}2

(7.6)

Loosely speaking, I() is the amount of information about contained in a single observation
from the density f (x). However, this interpretation doesnt always make sensefor example,
it is possible to have I() = 0 for a very informative observation. See Exercise 7.6.
Although we do not dwell on this fact in this course because it has measure-theoretic underpinnings, expectation may be viewed as integration (even when, say, the distribution is
discrete and the density is actually a mass function). Suppose that f (x) is twice differentiable with respect to and that the operations of differentiation and integration may be
interchanged in the following sense:
Z
Z
d
d
f (x) dx
(7.7)
f (x) dx =
d
d
and
d2
d2

Z
f (x) dx =

d2
f (x) dx.
d2

(7.8)

R
(It is awkward to express the above ideas using our usual E operator!) Since f (x) dx = 1,
the left-hand sides of Equations (7.7) and (7.8) are both zero and this fact leads to two
additional expressions for I(). From Equation (7.7) follows


d
I() = Var
log f (X) ,
(7.9)
d
and Equation (7.8) implies

I() = E


d2
log f (X) ;
d2

(7.10)

see Exercise 7.5. In many cases, Equation (7.10) is the easiest form of the information to
work with.
Equations (7.9) and (7.10) make clear a helpful property of the information, namely that
for independent random variables, the information about contained in the joint sample is
145

simply the sum of the individual information components. In particular, if we have a simple
random sample of size n from f (x), then the information about equals nI().
The reason that we need the Fisher information is that we will show that under certain
assumptions (often called regularity conditions),



1
d
n(n 0 ) N 0,
,
(7.11)
I(0 )
where n is the consistent root of the likelihood equation guaranteed to exist by Theorem
7.4.
Example 7.7 Suppose that X1 , . . . , Xn are independent Poisson(0 ) random variables. Then the likelihood equation has a unique root, namely n = X n , and we

d
know that by the central limit theorem n(n 0 ) N (0, 0 ). Furthermore, the
Fisher information for a single observation in this case is

 2
1
d
X
E
log
f
(X)
=
E
=
.

d2
2

Thus, in this example, Equation (7.11) holds.

Rather than stating all of the regularity conditions necessary to prove Equation (7.11), we
work backwards, figuring out the conditions as we go through the steps of the proof. The
first step is to expand `0 (n ) in a Taylor series around 0 . Let us introduce the notation `i ()
to denote the contribution to the loglikelihood from the ith observation; that is, `i () =
log f (Xi ). Thus, we obtain
`() =

n
X

`i ().

i=1

For the Taylor expansion of `0 (n ), let ei (n , 0 ) denote the remainder for a first-order expansion of `0i (n ). That is, we define ei (n , 0 ) so that
`0i (n ) = `0i (0 ) + (n 0 )`00i (0 ) + ei (n , 0 ).
Summing over i, we obtain
`0 (n ) = `0 (0 ) + (n 0 ) [`00 (0 ) + En ] ,

(7.12)

P
where En = ni=1 ei (n , 0 )/(n 0 ), or, in the event n = 0 , En = 0. (Remember, n and
En are random variables.)

146

Rewriting Equation (7.12) gives

n(n 0 ) =

n )}
1 {`0 (0 ) `0 (
n{`0 (n ) `0 (0 )}
n
=
.
`00 (0 ) + En
n1 `00 (0 ) n1 En

(7.13)

Lets consider the pieces of Equation (7.13) individually. If Equation (7.7) holds and I(0 ) <
, then
!
n

1 0
1X d
d
` (0 ) = n
log f0 (Xi ) N {0, I(0 )}
n i=1 d
n
by the central limit theorem and Equation (7.9). If Equation (7.8) holds, then
n

1
1 X d2
P
`00 (0 ) =
log f0 (Xi ) I(0 )
n
n i=1 d2
by the weak law of large numbers and Equation (7.10). And we relied on the conditions of
Theorem 7.4 to guarantee the existence of n such that `0 (n ) = 0 with probability approaching one and n 0 (do you see where we used the latter fact?).
P

Finally, we need a condition that ensures that n1 En 0. One way this is often done is as
follows: If we assume that the third derivative `000
i () exists and is uniformly bounded in a
neighborhood of 0 , say by the constant K0 , we may write the Taylor theorem remainder
ei (n , 0 ) in the form of equation (1.7) to obtain
n
n
1 X ei (n , 0 )
n 0 X 000
1
En =
=
` ( ),
n
n i=1 n 0
2n i=1 i in
P

where each in
is between n and 0 . Therefore, since n 0 , we know that with probability
approaching 1 as n ,


n
 1  n 0 X
K0
 En 
K0 =
(n 0 ),
n 
2n i=1
2
P

which means that n1 En 0.

In conclusion, if all of our assumptions hold, then the numerator of (7.13) converges in
distribution to N {0, I(0 )} by Slutskys theorem. Furthermore, the denominator in (7.13)
converges to I(0 ), so a second use of Slutskys theorem gives the following theorem.
Theorem 7.8 Suppose that the conditions of Theorem 7.4 are satisfied, and let n
denote a consistent root of the likelihood equation. Assume also that Equations
147

(7.7) and (7.8) hold and that 0 < I(0 ) < . Finally, assume that `i () has three
derivatives in some neighborhood of 0 and that `000
i () is uniformly bounded in
this neighborhood. Then



1
d

.
n(n 0 ) N 0,
I(0 )
Sometimes, it is not possible to find an exact zero of `0 (). One way to get a numerical
approximation to a zero of `0 () is to use Newtons method, in which we start at a point 0
and then set
1 = 0

`0 (0 )
.
`00 (0 )

(7.14)

Ordinarily, after finding 1 we would set 0 equal to 1 and apply Equation (7.14) iteratively.

However, we may show that by using a single step of Newtons method, starting from a nconsistent estimator
of 0 , we may obtain an estimator with the same asymptotic distribution

as n . A n-consistent estimator is an estimator of 0 , say n , with the property that

n(n 0 ) is bounded in probability. For the full

definition of bounded in probability, refer
to Exercise 2.2, but a sufficient condition is that n(n 0 ) converges in distribution to
any random variable.
The proof of the following theorem is left as an exercise:

Theorem 7.9 Suppose that n is any n-consistent estimator of 0 . Then under the
conditions of Theorem 7.8, if we set
`0 (n )
n = n
,
`00 (n )

(7.15)

then


n(n 0 ) N 0,
d

1
I(0 )


.

Exercises for Section 7.2

Exercise 7.5 Show how assumptions (7.7) and (7.8) establish Equations (7.9) and (7.10),
respectively.
Exercise 7.6 Suppose that X is a normal random variable with mean 3 and known
variance 2 . Calculate I(), then argue that the Fisher information can be zero
in a case in which there is information about in the observed value of X.
148

Exercise 7.7 (a) Show that under the conditions of Theorem 7.8, then if n is a
consistent root of the likelihood equation, P0 (n is a local maximum) 1.
(b) Using the result of part (a), show that for any two sequences 1n and 2n of
consistent roots of the likelihood equation, P0 (1n = 2n ) 1.
Exercise 7.8 Prove Theorem 7.9.

Hint: Start with n(n 0 ) = n(n n ) + n(n 0 ), then expand `0 (n )

in a Taylor series about 0 and substitute the result into Equation (7.15). After
simplifying, use the result of Exercise 2.2 along with arguments similar to those
leading up to Theorem 7.8.
Exercise 7.9 Suppose that the following is a random sample from a logistic density
with distribution function F (x) = (1 + exp{ x})1 (Ill cheat and tell you that
I used = 2.)
1.0944
1.2853
1.0422

6.4723
1.0439
0.1690

3.1180
1.7472
3.6111

3.8318
4.9483
0.9970

4.1262
1.7001
2.9438

(a) Evaluate the unique root of thelikelihood equation numerically. Then, taking
the sample median as our known n-consistent estimator n of , evaluate the
estimator n in Equation (7.15) numerically.

(b) Find the asymptotic distributions of n(n 2) and n(n 2). Then,
simulate 200 samples of size n = 15 from the logistic distribution with = 2.
Find the sample variances of the resulting sample medians and n -estimators.
How well does the asymptotic theory match reality?

7.3

Asymptotic Efficiency and Superefficiency

In Theorem 7.8, we showed that a consistent root n of the likelihood equation satisfies



1
d

n(n 0 ) N 0,
.
I(0 )

In Theorem 7.9, we stated that if n is a n-consistent estimator of 0 and n = n

`0 (n )/`00 (n ), then


n(n 0 ) N 0,
d

149

1
I(0 )


.

(7.16)

Note the similarity in the last two asymptotic limit statements. There seems to be something
special about the limiting variance 1/I(0 ), and in fact this is true.
Much like the Cramer-Rao lower bound states that (under some regularity conditions) an
unbiased estimator of 0 cannot have a variance smaller than 1/I(0 ), the following result is
true:
Theorem 7.10 Suppose that the conditions of theorem 7.8 are satisfied and that n
is an estimator satisfying

d
n(n 0 ) N {0, v(0 )}
for all 0 , where v() is continuous. Then v() 1/I() for all .
In other words, 1/I() is, in the sense of Theorem 7.10, the smallest possible asymptotic
variance for an estimator. For this reason, we refer to any estimator n satisfying (7.16) for
all 0 an efficient estimator.
One condition in Theorem 7.10 that may be a bit puzzling is the condition that v() be
continuous. If this condition is dropped, then a well-known counterexample, due to Hodges,
exists:
Example 7.11 Suppose that n is an efficient estimator of 0 . Then if we define

0 if n(n )4 < 1;

n =
n otherwise,
it is possible to show (see Exercise 7.10) that n is superefficient in the sense that



1
d
n(n 0 ) N 0,
I(0 )

d
for all 0 6= 0 but n(n 0 ) 0 if 0 = 0. In other words, when the true value
of the parameter 0 is 0, then n does much better than an efficient estimator;
yet when 0 6= 0, n does just as well.
Just as the invariance property of maximum likelihood estimation states that the MLE of
a function of equals the same function applied to the MLE of , a function of an efficient
estimator is itself efficient:
Theorem 7.12 If n is efficient for 0 , and if g() is a differentiable and invertible
function with g 0 (0 ) 6= 0, g(n ) is efficient for g(0 ).
The proof of the above theorem follows immediately from the delta method, since the Fisher
information for g() is I()/{g 0 ()}2 by Equation (7.6). In fact, if one simply remembers
150

the content of Theorem 7.12, then it is not necessary to memorize Equation (7.6), for it is
always possible to rederive this equation quickly using the delta method!

We have already noted that (under suitable regularity conditions) if n is a n-consistent

estimator of 0 and
`0 (n )
n = n
,
`00 (n )

(7.17)

then n is an efficient estimator of 0 . Alternatively, we may set

`0 (n )
n = n +
nI(n )

(7.18)

and n is also an efficient estimator of 0 . Problem 7.8 asked you to prove the former
fact regarding n ; the latter fact regarding n is proved in nearly the same way because
P
n1 `00 (n ) I(0 ). In Equation (7.17), as already remarked earlier, n results from a single step of Newtons method; in Equation (7.18), n results from a similar method called
Fisher scoring. As is clear from comparing Equations (7.17) and (7.18), scoring differs from
Newtons method in that the Fisher information is used in place of the negative second
derivative of the loglikelihood function. In some examples, scoring and Newtons method
are equivalent.
A note on terminology: The derivative of `() is sometimes called the score function.
Furthermore, nI() and `00 () are sometimes referred to as the expected information
and the observed information, respectively.
Example 7.13 Suppose X1 , . . . , Xn are independent from a Cauchy location family
with density
f (x) =

1
.
{1 + (x )2 }

Then
0

` () =

n
X
i=1

2(xi )
,
1 + (xi )2

so the likelihood equation is very difficult to solve.

However, an efficient estimator

may still be created by starting with some n-consistent estimator n , say the
sample median, and using either Equation (7.17) or Equation (7.18). In the latter
case, we obtain the simple estimator
2 0
` (n ),
n
verification of which is the subject of Problem 7.11.
n = n +

151

(7.19)

For the remainder of this section, we turn our attention to Bayes estimators, which give
yet another source of efficient estimators. A Bayes estimator is the expected value of the
posterior distribution, which of course depends on the prior chosen. Although we do not
prove this fact here (see Ferguson 21 for details), any Bayes estimator is efficient under some
very general conditions. The conditions are essentially those of Theorem 7.8 along with the
stipulation that the prior density is positive everywhere on . (Note that if the prior density
is not positive on , the Bayes estimator may not even be consistent.)
Example 7.14 Consider the binomial distribution with beta prior, say X binomial(n, p)
and p beta(a, b). Then the posterior density of p is proportional to the product of the likelihood and the prior, which (ignoring multiplicative constants not
involving p) equals
pa1 (1 p)b1 pX (1 p)nX .
Therefore, the posterior distribution of p is beta(a + X, b + n X). The Bayes
estimator is the expectation of this distribution, which equals (a + X)/(a + b + n).
If we let n denote the Bayes estimator here, then





X
X
n(n p) = n
p + n n
.
n
n
We may see that the rightmost term converges to zero in probability by writing





X
n
X
n n
=
a (a + b)
,
n
a+b+n
n
P

since a (a + b)X/n a (a + b)p by the weak law of large numbers. In other

words, the Bayes estimator in this example has the same limiting distribution as
the MLE, X/n. It is possible to verify that the MLE is efficient in this case.
The central question when constructing a Bayes estimator is how to choose the prior distribution. We consider one class of prior distributions, called Jeffreys priors. (A common
grammatical mistake is to write Jeffreys priors, but this is incorrect because they are
named for Harold Jeffreys and the letter s is not possessive. Analogously, Bayes estimators
are named for Thomas Bayes.)
For a Bayes estimator n of 0 , we have


n(n 0 ) N 0,
d

1
I(0 )


.

Since the limiting variance of n depends on I(0 ), if I() is not a constant, then some values
of may be estimated more precisely than others. In analogy with the idea of variancestabilizing transformations seen in Section 5.1.2, we might consider a reparameterization
152

 = g() such that {g 0 ()}2 /I() is a constant. More precisely, if g 0 () = c

o
n
d
n g(n ) 0 N (0, c2 ).

I(), then

So as not to influence the estimation of , we choose as the Jeffreys prior a uniform prior on
. p
Therefore, the Jeffreys prior density on is proportional to g 0 (), which is proportional
to I(). Note that this may lead to an improper prior.
Example 7.15 In the case of Example 7.14, we may verify that for a Bernoulli(p)
observation,
d2
I(p) = E 2 {X log p + (1 X) log(1 p)}
dp


1X
1
X
+
.
=
= E
2
2
p
(1 p)
p(1 p)
Thus, the Jeffreys prior on p in this case has a density proportional to p1/2 (1
p)1/2 . In other words, the prior is beta( 21 , 21 ). Therefore, the Bayes estimator
corresponding to the Jeffreys prior is
X + 12
n =
.
n+1

Exercises for Section 7.3

Exercise 7.10 Verify the claim made in Example 7.11.
Exercise 7.11 If f (x) forms a location family, so that f (x) = f (x ) for some
density f (x), then the Fisher information I() is a constant (you may assume
this fact without proof).
(a) Verify that for the Cauchy location family,
f (x) =

1
,
{1 + (x )2 }

we have I() = 21 .
(b) For 500 samples of size n = 51 from a standard Cauchy distribution, calculate
the sample median n and the efficient estimator n of Equation (7.19). Compare
the variances of n and n with their theoretical asymptotic limits.
153

Exercise 7.12 (a) Derive the Jeffreys prior on for a random sample from Poisson().
Is this prior proper or improper?
(b) What is the Bayes estimator of for the Jeffreys prior? Verify directly that
this estimator is efficient.
Exercise 7.13 (a) Derive the Jeffreys prior on 2 for a random sample from N (0, 2 ).
Is this prior proper or improper?
(b) What is the Bayes estimator of 2 for the Jeffreys prior? Verify directly that
this estimator is efficient.

7.4

The multiparameter case

Suppose now that the parameter is the vector = (1 , . . . , k ). If X f (x), then I(), the
information matrix, is the k k matrix


I() = E h (X)h>
(X) ,
where h (x) = [log f (x)]. This is a rare instance in which its probably clearer to use
component-wise notation than vector notation:
Definition 7.16 Given a k-dimensional parameter vector and a density function
f (x), the Fisher information matrix I() is the k k matrix whose (i, j) element
equals



log f (X)
log f (X) ,
Iij () = E
i
j
as long as the above quantity is defined for all (i, j).
Note that the one-dimensional Definition 7.6 is a special case of Definition 7.16.
Example 7.17 Let = (, 2 ) and suppose X N (, 2 ). Then

1
(x )2
2,
log f (x) = log 2

log
2
2 2
so

x
log f (x) =
and

1
log f (x) = 2
2

154


(x )2
1 .
2

Thus, the entries in the information matrix are as follows:



(X )2
1
I11 () = E
=
,
4
2


(X )3 X
= 0,
I21 () = I12 () = E

2 6
2 4


1
1
(X )2 (X )4
2
3 4
1
I22 () = E
=

+
= 4.
2
6
8
4
6
8
4
2
4
4
2
4
2
As in the one-dimensional case, Definition 7.16 is often not the easiest form of I() to work
with. This fact is illustrated by Example 7.17, in which the calculation of the information
matrix requires the evaluation of a fourth moment as well as a lot of algebra. However, in
analogy with Equations (7.7) and (7.8), if

f (X)
f (X)
2
f (X)

i
E
= E
and 0 =
E
= E
0=
i
f (X)
f (X)
i j
f (X)

for all i and j, then the following alternative forms of I() are valid:



Iij () = Cov
log f (X),
log f (X)
i
j


2
= E
log f (X) .
i j

2
f (X)
i j

f (X)

(7.20)

(7.21)
(7.22)

Example 7.18 In the normal case of Example 7.17, the information matrix is perhaps
a bit easier to compute using Equation (7.22), since we obtain
1
2
x
2
log
f
(x)
=

,
log f (x) = 4 ,

2
2
2

and
2
1
(x )2
log
f
(x)
=

( 2 )2
2 4
6
Taking expectations gives

I() =

1
2

1
2 4

as before but without requiring any fourth moments.

155

By Equation (7.21), the Fisher information matrix is nonnegative definite, just as the Fisher
information is nonnegative in the one-parameter case. A further fact the generalizes into
the multiparameter case is the additivity of information: If X and Y are independent,
then IX () + IY () = I(X,Y ) (). Finally, suppose that = g() is a reparameterization,
where g() is invertible and differentiable. Then if J is the Jacobian matrix of the inverse
transformation (i.e., Jij = i /j ), then the information about is
I() = J > I()J.
As well see later, almost all of the same efficiency results that applied to the one-parameter
case apply to the multiparameter case as well. In particular, we will see that an efficient
is one that satisfies
estimator

d
0 )
Nk {0, I( 0 )1 }.
n(
is an
Note that the formula for the information under a reparameterization implies that if
0
1
efficient estimator for and the matrix J is invertible, then g (
) is an efficient estimator
for 0 = g 1 ( 0 ), since

d
n{g 1 (
) g 1 ( 0 )} Nk {0, J(J > I( 0 )J)1 J > },
and the covariance matrix above equals I( 0 )1 .
As in the one-parameter case, the likelihood equation is obtained by setting the derivative of
`() equal to zero. In the multiparameter case, though, the gradient `() is a 1 k vector,
so the likelihood equation becomes `() = 0. Since there are really k univariate equations
implied by this likelihood equation, it is common to refer to the likelihood equations (plural),
which are

`() = 0 for i = 1, . . . , k.
i
In the multiparameter case, we have essentially the same theorems as in the one-parameter
case.
Theorem 7.19 Suppose that X1 , . . . , Xn are independent and identically distributed
random variables (or vectors) with density f0 (x) for 0 in an open subset of Rk ,
where distinct values of 0 yield distinct distributions for X1 (i.e., the parameter
is identifiable). Furthermore, suppose that the support set {x : f (x) > 0} does

not depend on . Then with probability approaching 1 as n , there exists

P
= 0 and
0 .
such that `()
guaranteed by Theorem 7.19 as a
As in the one-parameter case, we shall refer to the
consistent root of the likelihood equations. Unlike Theorem 7.4, however, Corollary 7.5 does
156

 is the unique solution

not generalize to the multiparameter case because it is possible that
of the likelihood equations and a local maximum but not the MLE. The best we can say is
the following:
Corollary 7.20 Under the conditions of Theorem 7.19, if there is a unique root of
the likelihood equations, then this root is consistent for 0 .
The asymptotic normality of a consistent root of the likelihood equation holds in the multiparameter case just as in the single-parameter case:

Theorem 7.21 Suppose that the conditions of Theorem 7.19 are satisfied and that
denotes a consistent root of the likelihood equations. Assume also that Equation
(7.20) is satisfied for all i and j and that I( 0 ) is positive definite with finite
entries. Finally, assume that 3 log f (x)/i j k exists and is bounded in a
neighborhood of 0 for all i, j, k. Then

d
0 )
n(
Nk {0, I 1 ( 0 )}.
As in the one-parameter case, there is some terminology associated with the derivatives of
the loglikelihood function. The gradient vector `() is called the score vector. The negative
second derivative 2 `() is called the observed information, and nI() is sometimes called
the expected information. And the second derivative of a real-valued function of a k-vector,
such as the loglikelihood function `(), is called the Hessian matrix.
Newtons method (often called the Newton-Raphson method in the multivariate case) and
n , one step
scoring work just as they do in the one-parameter case. Starting from the point
of Newton-Raphson gives
n
o1
n 2 `(
n)
n)
n =
`(
(7.23)
and one step of scoring gives
n )`(
n ).
n + 1 I 1 (
n =
n

(7.24)

n is a n-consistent estiTheorem 7.22 Under the assumptions of Theorem 7.21, if

mator of 0 , then the one-step Newton-Raphson estimator n defined in Equation
(7.23) satisfies

n( n 0 ) Nk {0, I 1 ( 0 )}

and the one-step scoring estimator n defined in Equation (7.24) satisfies

n( n 0 ) Nk {0, I 1 ( 0 )}.
157

 n as one that satisfies

As in the one-parameter case, we define an efficient estimator

d
n 0 )
n(
Nk {0, I 1 ( 0 )}.
This definition is justified by the fact that if n is any estimator satisfying

d
n( n ) Nk {0, ()},
where I 1 () and () are continuous, then () I 1 () is nonnegative definite for all .
This is analagous to saying that 2 () I 1 () is nonnegative in the univariate case, which
is to say that I 1 () is the smallest possible variance.
Finally, we note that Bayes estimators are efficient, just as in the one-parameter case. This
means that the same three types of estimators that are efficient in the one-parameter case
the consistent root of the likelihood equation, the one-step scoring and Newton-Raphson
estimators, and Bayes estimatorsare also efficient in the multiparameter case.

Exercises for Section 7.4

Exercise 7.14 Let X multinomial(1, p), where p is a k-vector for k > 2. Let
p = (p1 , . . . , pk1 ). Find I(p ).
Exercise 7.15 Suppose that R R+ (that is, 1 R and 2 (0, )) and


x 1
1
f (x) = f
2
2
for some continuous, differentiable density f (x) that is symmetric about the origin. Find I().
Exercise 7.16 Prove Theorem 7.21.
Hint: Use Theorem 1.40.
Exercise 7.17 Prove Theorem 7.22.
Hint: Use Theorem 1.40.
Exercise 7.18 The multivariate generalization of a beta distribution is a Dirichlet
distribution, which is the natural prior distribution for a multinomial
Plikelihood.
If p is a random (k+1)-vector constrained so that pi > 0 for all i and k+1
i=1 pi = 1,
then (p1 , . . . , pk ) has a Dirichlet distribution with parameters a1 > 0, . . . , ak+1 > 0
if its density is proportional to
( k
)
n
o
X
p1a1 1 pakk 1 (1 p1 pk )ak+1 1 I min pi > 0 I
pi < 1 .
i

158

i=1

Prove that if G1 , . . . , Gk+1 are independent random variables with Gi gamma(ai , 1),
then
1
(G1 , . . . , Gk )
G1 + + Gk+1
has a Dirichlet distribution with parameters a1 , . . . , ak+1 .

7.5

Nuisance parameters

This section is the one intrinsically multivariate section in this chapter; it does not have an
analogue in the one-parameter setting. Here we consider how efficiency of an estimator is
affected by the presence of nuisance parameters.
Suppose is the parameter vector but 1 is the only parameter of interest, so that 2 , . . . , k
are nuisance parameters. We are interested in how the asymptotic precision with which we
may estimate 1 is influenced by the presence of the nuisance parameters. In other words, if
is efficient for , then how does 1 as an estimator of 1 compare to an efficient estimator

of 1 , say , under the assumption that all of the nuisance parameters are known?
Assume I() is positive definite. Let ij denote the (i, j) entry of I() and let ij denote
the (i, j) entry of I 1 (). If all of the nuisance
are known, then I(1 ) = 11 ,
parameters

which means that the asymptotic variance of n( 1 ) is 1/11 . On

the other hand, if the
) is I 1 (),
nuisance parameters are not known then the asymptotic
variance
of
n(

which means that the marginal asymptotic variance of n(1 1 ) is 11 . Of interest here
is the comparison between 11 and 1/11 .
The following theorem may be interpreted to mean that the presence of nuisance parameters
always increases the variance of an efficient estimator.
Theorem 7.23 11 1/11 , with equality if and only if 12 = = 1k = 0.
Proof: Partition I() as follows:

I() =

11 >


,

where and are (k 1) 1 and (k 1) (k 1), respectively. Let = 11 > 1 .

We may verify that if > 0, then


1
1
> 1
1
I () =
.
1 1 > 1 + 1

159

This proves the result, because the positive definiteness of I() implies that 1 is positive
definite, which means that
11 =

1
1

,
>
1
11
11

with equality if and only if = 0. Thus, it remains only to show that > 0. But this is
immediate from the positive definiteness of I(), since if we set


1
v=
,
> 1
then = v> I()v.
The above result shows that it is important to take nuisance parameters into account in
estimation. However, it is not necessary to estimate the entire parameter vector all at once,
since (1 , . . . , k ) is efficient for if and only if each of the i is efficient for i in the presence
of the other nuisance parameters (see problem 7.19).

Exercises for Section 7.5

Exercise 7.19 Letting ij denote the (i, j) entry of I 1 (), we say that i is efficient
for i in the presence of the nuisance parameters 1 , . . . , i1 , i+1 , . . . , k if the
asymptotic variance of n(i i ) is ii .
Prove that (1 , . . . , k ) is efficient for if and only if for all i, the estimator i is
efficient for i in the presence of nuisance parameters 1 , . . . , i1 , i+1 , . . . , k .

160

Chapter 8
Hypothesis Testing
8.1

Wald, Rao, and Likelihood Ratio Tests

Suppose we wish to test H0 : = 0 against H1 : 6= 0 . The likelihood-based results of

Chapter 7 give rise to several possible tests.
To this end, let `() denote the loglikelihood and n the consistent root of the likelihood
equation. Intuitively, the farther 0 is from n , the stronger the evidence against the null
hypothesis. But how far is far enough? Note that if 0 is close to n , then `(0 ) should
also be close to `(n ) and `0 (0 ) should be close to `0 (n ) = 0.
If we base a test upon the value of (n 0 ), we obtain a Wald test.
If we base a test upon the value of `(n ) `(0 ), we obtain a likelihood ratio test.
If we base a test upon the value of `0 (0 ), we obtain a (Rao) score test.
Recall that in order to prove Theorem 7.8, we argued that under certain regularity conditions,
the following facts are true under H0 :



1
d

n(n 0 ) N 0,
;
(8.1)
I(0 )
1
d
`0 (0 ) N {0, I(0 )};
(8.2)
n
1
P
`00 (0 ) I(0 ).
(8.3)
n
161

Equation (8.2) is proved using the central limit theorem; Equation (8.3) is proved using
the weak law of large numbers; and Equation (8.1) is the result of a Taylor expansion
together with Equations (8.2) and (8.3). The three equations above will help us to justify
the definitions of Wald, score, and likelihood ratio tests to follow.
Equation (8.1) suggests that if we define
p
Wn = nI(0 )(n 0 ),

(8.4)

then Wn , called a Wald statistic, should converge in distribution to standard normal under
H0 . Note that this fact remains true if we define
p
n 0 ),
Wn = nI(
(8.5)
where I is consistent for 0 ; for example, I could be I(n ).
Definition 8.1 A Wald test is any test that rejects H0 : = 0 in favor of H1 : 6= 0
when |Wn | u/2 for Wn defined as in Equation (8.4) or Equation (8.5). As usual,
u/2 denotes the 1 2 quantile of the standard normal distribution.
Equation (8.2) suggests that if we define
Rn = p

1
nI(0 )

`0 (0 ),

(8.6)

then Rn , called a Rao score statistic or simply a score statistic, converges in distribution to
standard normal under H0 . We could also replace I(0 ) by a consistent estimator I as in
Equation (8.5), but usually this is not done: One of the main benefits of the score statistic
is that it is not necessary to compute n , and using I(n ) instead of I(0 ) would defeat this
purpose.
Definition 8.2 A score test, sometimes called a Rao score test, is any test that
rejects H0 : = 0 in favor of H1 : 6= 0 when |Rn | u/2 for Rn defined as in
Equation (8.6).
The third type of test, the likelihood ratio test, requires a bit of development. It is a test
based on the statistic
n = `(n ) `(0 ).
If we Taylor expand `0 (n ) = 0 around the point 0 , we obtain
(
)
n 0

`0 (0 ) = (n 0 ) `00 (0 ) +
`000 ( ) .
2
162

(8.7)

(8.8)

We now use a second Taylor expansion, this time of `(n ), to obtain

"
#
n 0

1
n = (n 0 )`0 (0 ) + (n 0 )2 `00 (0 ) +
`000 ( ) .
2
3

(8.9)

If we now substitute Equation (8.8) into Equation (8.9), we obtain



1 00
1 00
2

n = n(n 0 ) ` (0 ) + ` (0 ) + (n 0 )OP (1) ,

n
2n

(8.10)

where the OP (1) term consists of the sum of `000 ( )/(2n) and `000 ( )/(6n), which is
bounded in probability under the third-derivative assumption of Theorem 7.8.
d

By Equations (8.1) and (8.3) and Slutskys theorem, this implies that 2n 21 under the
null hypothesis. Noting that the 1 quantile of a 21 distribution is u2/2 , we make the
following definition.
Definition 8.3 A likelihood ratio test is any test that rejects
H0 : = 0 in favor

2
of H1 : 6= 0 when 2n u/2 or, equivalently, when 2n u/2 .

P
P
Since it may be shown that 2n |Wn | 0 and Wn Rn 0, the three tests defined above
Wald tests, score tests, and likelihood ratio tests are asymptotically equivalent in the
sense that under H0 , they reach the same decision with probability approaching 1 as n .
However, they may be quite different for a fixed sample size n, and furthermore they have
some relative advantages and disadvantages with respect to one another. For example,
It is straightforward to create one-sided Wald and score tests (i.e., tests of H0 : = 0
against H1 : > 0 or H1 : < 0 ), but this is more difficult with a likelihood ratio
test.
The score test does not require n whereas the other two tests do.
The Wald test is most easily interpretable and yields immediate confidence intervals.
The score test and likelihood ratio test are invariant under reparameterization, whereas
the Wald test is not.
Example 8.4 Suppose that X1 , . . . , Xn are independent with density f (x) = ex I{x >
0}. Then `() = n(log X n ), which yields


1
n
0
` () = n
Xn
and `00 () = 2 .

163

From the form of `00 (), we see that I() = 2 , and setting `0 () = 0 yields
n = 1/X n . From these facts, we obtain as the Wald, score, and likelihood ratio
statistics


n
1
0 ,
Wn =
0 X n



1
Wn
Rn = 0 n
Xn =
, and
0
0 X n


n = n X n 0 1 log(0 X n ) .
Thus, we rejectH0 : = 0 in favor of H1 : 6= 0 whenever |Wn | u/2 ,
|Rn | u/2 , or 2n u/2 , depending on which test were using.
Generalizing the three tests to the multiparameter setting is straightforward. Suppose we
wish to test H0 : = 0 against H1 : 6= 0 , where Rk . Then
d
0 )> I( 0 )(
0 )
n(
2k ,
(8.11)
1
d
def
`( 0 )> I 1 ( 0 )`( 0 ) 2k , and
(8.12)
Rn =
n
d 1 2
def
`( 0 )
n = `()
.
(8.13)
2 k
Therefore, if ck denotes the 1 quantile of the 2k distribution, then the multivariate Wald
test, score test, and likelihood ratio test reject H0 when Wn ck , Rn ck , and 2n ck ,
respectively. As in the one-parameter case, the Wald test may also be defined with I( 0 )

replaced by a consistent estimator I.

Wn

def

Exercises for Section 8.1

Exercise 8.1 Let X1 , . . . , Xn be a simple random sample from a Pareto distribution
with density
f (x) = c x(+1) I{x > c}
for a known constant c > 0 and parameter > 0. Derive the Wald, Rao, and
likelihood ratio tests of = 0 against a two-sided alternative.
Exercise 8.2 Suppose that X is multinomial(n, p), where p Rk . In order to satisfy
the regularity condition that the parameter space be an open set, define =
(p1 , . . . , pk1 ). Suppose that we wish to test H0 : = 0 against H1 : 6= 0 .
(a) Prove that the Wald and score tests are the same as the usual Pearson chisquare test.
(b) Derive the likelihood ratio statistic 2n .
164

8.2

Contiguity and Local Alternatives

Suppose that we wish to test the hypotheses

H0 : = 0 and H1 : > 0 .

(8.14)

The test is to be based on a statistic Tn , where as always n denotes the sample size, and we
shall decide to
reject H0 in (8.14) if Tn Cn

(8.15)

for some constant Cn . The test in question may be one of the three types of tests introduced
in Section 8.1, or it may be an entirely different test. We may define some basic asymptotic
concepts regarding tests of this type.
Definition 8.5 If P0 (Tn Cn ) for test (8.15), then test (8.15) is said to have
asymptotic level .
Definition 8.6 If two different tests of the same hypotheses reach the same conclusion
with probability approaching 1 under the null hypothesis as n , the tests
are said to be asymptotically equivalent.
The power of test (8.15) under the alternative is defined to be
n () = P (Tn Cn ).
We expect that the power should approach 1.
Definition 8.7 A test (or, more precisely, a sequence of tests) is said to be consistent
against the alternative if n () 1.
Note that in some contexts, is used to denote the type II error probability, which is
actually one minus the power. We admit that the inconsistent usage of in the literature is
confusing, but we hope that bringing attention to this inconsistency will help to allay much
of this confusion. Here, will always refer to power.
Unfortunately, the concepts we have defined so far are of limited usefulness. If we wish to
compare two different tests of the same hypotheses, then if the tests are both sensible they
should be asymptotically equivalent and consistent. Thus, consistency is nice but it doesnt
tell us much; asymptotic equivalence is nice but it doesnt allow us to compare tests.
We make things more interesting by considering, instead of a fixed alternative , a sequence
of alternatives 1 , 2 , . . .. Let us make some assumptions about the asymptotic distribution

165

of the test statistic Tn in (8.15). First, define () and 2 ()/n to be the mean and variance,
respectively, of Tn when is the true value of the parameter. In other words, let
() = E (Tn ) and 2 () = n Var (Tn ).
We assume that if the null hypothesis is true, which means is fixed and equal to 0 , then

n{Tn (0 )} d
p
N (0, 1)
(8.16)
2 (0 )
as n . Furthermore, we assume that if the alternatives are true, which means that the
distribution of Tn is determined by = n for each n, then

n{Tn (n )} d
p
N (0, 1)
(8.17)
2 (n )
as n . The limit in (8.17) is trickier than the one in (8.16) because it assumes that the
underlying parameter is changing along with n. Consider Example 8.8.
Example 8.8 Suppose we have a sequence of independently and identically distributed
random variables X1 , X2 , . . . with common distribution F . In a one-sample ttest, the test statistic Tn may be taken to be the sample mean X n . In this case,
the limit in (8.16) follows immediately from the central limit theorem. Yet to
verify (8.17), it is necessary to consider a triangular array of random variables:
X11 F1
X21 , X22 F2
X31 , X32 , X33 F3
...
We may often check that the Lyapunov or Lindeberg condition is satisfied, then
use the results of Section 4.2 to establish (8.17). In fact, the existence of, say,
a finite third absolute central moment, () = E |X1 E X1 |3 , is generally
sufficient because
n
X
(n )
1
p
E |Xni E Xni |3 = 3
2
3
n (n )
( n (n )) i=1

and the Lyapunov condition holds as long as (n )/ 3 (n ) tends to some finite

limit. We generally assume that n 0 , so as long as () and () are continuous, (n )/ 3 (n ) tends to (0 )/ 3 (0 ).

166

If (8.16) and (8.17) are true, we may calculate the power of the test against the sequence
of alternatives 1 , 2 , . . . in a straightforward way using normal distribution calculations. As
mentioned in Example 8.8, we typically assume that this sequence converges to 0 . Such
a sequence is often referred to as a contiguous sequence of alternatives, since contiguous
means next to; the idea of contiguity is that we choose not a single alternative hypothesis
but a sequence of alternatives next to the null hypothesis.
First, we should determine a value for Cn so that test (8.15) has asymptotic level . Define
u to be the 1 quantile of the standard normal distribution. By limit (8.16),


(0 )u
;
P0 Tn (0 )
n
therefore, we define a new test, namely
reject H0 in (8.14) if Tn (0 ) +

(0 )u

(8.18)

and conclude that test (8.18) has asymptotic level as desired.

We now calculate the power of test (8.18) against the alternative n :


(0 )u
n (n ) = Pn Tn (0 ) +
n



n{Tn (n )} (n )
n(n 0 ) (n ) (0 )

u .(8.19)
= P n
(n )
(0 )
(0 )
n 0
Thus,
n (n ) tends to an interesting limit (i.e., a limit between and 1) if (n ) (0 );

n(n 0 ) tends to a nonzero, finite limit; and () is differentiable at 0 . This fact is

summarized in the following theorem:
Theorem 8.9 Let n > 0 for all n. Suppose that limits
(8.17) and (8.16) hold, ()
is continuous at 0 , () is differentiable at 0 , and n(n 0 ) for some
finite > 0. If 0 (0 ) or (0 ) depends on n, then suppose that 0 (0 )/ (0 )
tends to a nonzero, finite limit. Then if n (n ) denotes the power of test (8.18)
against the alternative n ,


0 (0 )
n (n ) lim
u .
n
(0 )
The proof of Theorem 8.9 merely uses Equation (8.19) and Slutskys theorem, since the
hypotheses of the theorem imply that (n )/ (0 ) 1 and {(n )(0 )}/(n 0 ) 0 (0 ).

167


Example 8.10 Let X Binomial(n, pn ), where pn = p0 + / n and Tn = X/n. To
test H0 : p = p0 against H1 : p > p0 , note that

n(T p0 ) d
p n
N (0, 1)
p0 (1 p0 )
p
p0 (1 p0 )/n.
under H0 . Thus, test (8.18) says to reject H0 whenever
T

p
+u
n
0

p
This test has asymptotic level . Since (p) = p(1 p) is continuous and
(p) = p is differentiable, Theorem 8.9 applies in this case as long as we can
verify the limit (8.17).
Let Xn1 , . . . , Xnn be independent Bernoulli(pn ) random variables. Then if Xn1
pn , . . . , Xnn pn can be shown to satisfy the Lyapunov condition, we have

n(Tn pn ) d
N (0, 1)
(pn )
and so Theorem 8.9 applies. The Lyapunov condition follows since |Xni pn | 1
implies
n
X
n
1
E |Xni pn |4
0.
2
{Var (nTn )} i=1
{npn (1 pn )}2

Thus, we conclude by Theorem 8.9 that

n (pn )

p
u
p0 (1 p0 )

!
.

To apply this result, suppose that we wish to test whether a coin is fair by flipping
it 100 times. We reject H0 : p = 1/2 in favor of H1 : p > 1/2 if the number of
successes divided by 100 is at least as large as 1/2 + u.05 /20, or 0.582. The power
of this test against the alternative p = 0.6 is approximately
!

100(0.6 0.5)

1.645 = (2 1.645) = 0.639.

0.52
Compare this asymptotic approximation with the exact power: The probability
of at least 59 successes out of 100 for a binomial(100, 0.6) random variable is
0.623.
Starting from Equation (8.19) and using the fact that

n{Tn (n )} (n )

(n )
(0 )
168

is approximately distributed as standard normal, we may obtain the following approximation

to the power for a fixed sample size n and a fixed alternative as follows:


n{() (0 )}
n ()
u .
(8.20)
(0 )
Note that in the case of Example 8.10, the approximation of Equation
(8.20) is the same as

the approximation obtained from Theorem 8.9 by setting = n(p p0 ).

There is an alternative formulation that yields a slightly different approximation. Starting
from



(0 )
n{Tn ()}
n{() (0 )}
u

,
n () = P
()
()
()
we obtain



n{() (0 )}
(0 )
n ()
u
.
()
()

(8.21)

Applying approximation (8.21) to the binomial case of Example 8.10, we obtain 0.641 instead
of 0.639 for the approximate power.
We may invert approximations (8.20) and (8.21) to obtain approximate sample sizes required
to achieve desired power against alternative . From (8.20) we obtain

(u u ) (0 )
() (0 )

(8.22)

and from (8.21) we obtain

u (0 ) u ()
.
n
() (0 )

(8.23)

We may compare tests by considering the relative sample sizes necessary to achieve the same
power at the same level against the same alternative.
Definition 8.11 Given tests 1 and 2 of the same hypotheses with asymptotic level
and a sequence of alternatives {k }, suppose that
(1)
m
(k ) and n(2)
(k )
k
k

as k for some sequences {mk } and {nk } of sample sizes. Then the asymptotic relative efficiency (ARE) of test 1 with respect to test 2 is
nk
e1,2 = lim
,
k mk
assuming this limit exists.
169

In Examples 8.12 and 8.13, we consider two different tests for the same hypotheses. Then,
in Example 8.15, we compute their asymptotic relative efficiency.
Example 8.12 Suppose we have paired data (X1 , Y1 ), . . . , (Xn , Yn ). Let Zi = Yi Xi
for all i. Assume that the Zi are independent and identically distributed with
distribution function P (Zi z) = F (z ) for some , where f (z) = F 0 (z) exists
and is symmetric about 0. Let W1 , . . . , Wn be a permutation of Z1 , . . . , Zn such
that |W1 | |W2 | |Wn |.
We wish to test H0 : = 0 against H1 : > 0. First, consider the Wilcoxon
signed rank test. Define
Rn =

n
X

iI{Wi > 0}.

i=1

Then under H0 , the I{Wi > 0} variables are independent Bernoulli(1/2) random
variables. Thus,
n
n
X
X
n(n + 1)
i2
n(n + 1)(2n + 1)
i
=
and Var Rn =
=
.
E Rn =
2
4
4
24
i=1
i=1

Furthermore, one may prove that

Rn E Rn d

N (0, 1)
Var Rn
under H0 by verifying, say, the Lindeberg condition. Thus, a test with asymptotic
level rejects H0 when
Rn

n(n + 1) u (0)
+ ,
4
n

p
where (0) = n (n + 1)(2n + 1)/24. Furthermore, it is possible to prove that

n[Rn (n )] d
p
N (0, 1),
2 (n )
where (n ) and 2 (n) are the mean and n times the variance of Rn under the
alternatives n = / n, though we will not prove this fact here (it is not as
simple as checking a Lindeberg condition because under n > 0, the I{Wi > 0}
variables are not quite independent). This means that it is possible to apply
Theorem 8.9.

170


Now we must find E Rn under the alternative n = / n. First, we note that
since |Wi | |Wj | for i j, Wi + Wj > 0 if and only if Wj > 0. Therefore,
P
j
i=1 I{Wi + Wj > 0} = jI{Wj > 0} and so we may rewrite Rn in the form
Rn =

j
n X
X

I{Wi + Wj > 0} =

j=1 i=1

j
n X
X

I{Zi + Zj > 0}.

j=1 i=1

Therefore, we obtain
(n ) =

j
n X
X
j=1 i=1

 
n
Pn (Zi + Zj > 0) = nPn (Z1 > 0) +
Pn (Z1 + Z2 > 0).
2

Since Pn (Z1 > 0) = Pn (Z1 n > n ) = 1 F (n ) and

Pn (Z1 + Z2 > 0) = Pn {(Z1 n ) + (Z2 n ) > 2n }
= E Pn {Z1 n > 2n (Z2 n ) | Z2 }
Z
{1 F (2n z)}f (z) dz,
=

we conclude that
 Z
n
2f (2 z)f (z) dz.
() = nf () +
2

Thus, because f (z) = f (z) by assumption,

Z

(0) = nf (0) + n(n 1)

f 2 (z) dz.

Letting
Z
K=

f 2 (z) dz,

we obtain
0 (0)
lim
= lim
n (0)
n

24{f (0) + (n 1)K}

p
= K 12.
(n + 1)(2n + 1)

Therefore, Theorem 8.9 shows that

n (n ) (K 12 u ).

171

(8.24)

Example 8.13 As in Example 8.12, suppose we have paired data (X1 , Y1 ), . . . , (Xn , Yn )
and Zi = Yi Xi for all i. The Zi are independent and identically distributed
with distribution function P (Zi z) = F (z ) for some , where f (z) = F 0 (z)
exists and is symmetric about 0. Suppose that the variance of Zi is 2 .
Since the t-test (unknown variance) and z-test (known variance) have the same
asymptotic properties, lets consider the z-test for simplicity. Then () = for
all . The relevant statistic is merely Z n , and the central limit theorem implies

d
nZ n / N (0, 1) under the null hypothesis. Therefore, thez-test in this case
rejects H0 : = 0 in favor of H1 : > 0 whenever Z n > u / n. A check of the
Lindeberg condition on the triangular array given by Z1 under 1 ; Z1 , Z2 under
2 ; Z1 , Z2 , Z3 under 3 ; and so on shows that

n(Z n n ) d
N (0, 1)

under the alternatives n = / n. Therefore, by Theorem 8.9, we obtain

n (n ) (/ u )

(8.25)

since 0 () = 1 for all .

Before finding the asymptotic relative efficiency (ARE) of the Wilcoxon signed rank test and
the t-test, we prove a lemma that enables this calculation.
Suppose that for two tests, called test 1 and test 2, we use sample sizes m and n, respectively.
We want m and n to tend to infinity together, an idea we make explicit by setting m = mk
and n = nk for k = 1, 2, . . .. Suppose that we wish to apply both tests to the same sequence
of alternative hypotheses 1 , 2 , . . .. As usual, we make the assumption that (k 0 ) times
the square root of the sample size tends to a finite, nonzero limit as k . Thus, we
assume

mk (k 0 ) 1 and nk (k 0 ) 2 .
Then if Theorem 8.9 may be applied to both tests, define c1 = lim 01 (0 )/1 (0 ) and c2 =
lim 02 (0 )/2 (0 ). The theorem says that
mk (k ) {1 c1 u } and nk (k ) {2 c2 u } .

(8.26)

To find the ARE, then, Definition 8.11 specifies that we assume that the two limits in (8.26)
are the same, which implies 1 c1 = 2 c2 , or
nk
c2
12 .
mk
c2
Thus, the ARE of test 1 with respect to test 2 equals (c1 /c2 )2 . This result is summed up in
the following lemma, which defines a new term, efficacy.
172

Lemma 8.14 For a test to which Theorem 8.9 applies, define the efficacy of the test
to be
c = lim 0 (0 )/ (0 ).
n

(8.27)

Suppose that Theorem 8.9 applies to each of two tests, called test 1 and test 2.
Then the ARE of test 1 with respect to test 2 equals (c1 /c2 )2 .
Example 8.15 Using the results of Examples 8.12 and 8.13, we conclude that the
efficacies of the Wilcoxon signed rank test and the t-test are
Z 2
1
12
f (z) dz and ,

respectively. Thus, Lemma 8.14 implies that the ARE of the signed rank test to
the t-test equals
2
Z
2
2
f (z) dz .
12

In the case of normally distributeddata, we may verify without too much difficulty
that the integral above equals (2 )1 , so the ARE is 3/ 0.9549. Notice how
close this is to one, suggesting that for normal data, we lose very little efficiency
by using a signed rank test instead of a t-test. In fact, it may be shown that this
asymptotic relative efficiency has a lower bound of 0.864. However, there is no
upper bound on the ARE in this case, which means that examples exist for which
the t-test is arbitrarily inefficient compared to the signed rank test.

Exercises for Section 8.2

Exercise 8.3 For the hypotheses considered in Examples 8.12 and 8.13, the sign test

d
is based on the statistic N+ = #{i : Zi > 0}. Since 2 n(N+ /n 12 ) N (0, 1)
under the null hypothesis, the sign test (with continuity correction) rejects H0
when

1
u n n
N+
+ .
2
2
2
(a) Find the efficacy of the sign test. Make sure to indicate how you go about
verifying Equation (8.17).
(b) Find the ARE of the sign test with respect to the signed rank test and the
t-test. Evaluate each of these for the case of normal data.
173

Exercise 8.4 Suppose X1 , . . . , Xn are a simple random sample from a uniform (0, 2)
distribution. We wish to test H0 : = 0 against H1 : > 0 at = .05.
Define Q1 and Q3 to be the first and third quartiles of the sample. Consider test
A, which rejects when
Q3 Q1 0 An ,
and test B, which rejects when
X 0 Bn .
Based on the asymptotic distribution of X and the joint asymptotic distribution
of (Q1 , Q3 ), find the values of An and Bn that correspond with the test in (8.18).
Then find the asymptotic relative efficiency of test A relative to test B.
Exercise 8.5 Let P be a family of probability distributions indexed by a real parameter . If X P , define () = E (X) and 2 () = Var (X). Now let 1 , 2 , . . .
be a sequence of parameter values such that n 0 as n . Suppose that
E n |X|2+ < M for all n for some positive and M . Also suppose that
Pfor each n,
Xn1 , . . . , Xnn are independent with distribution Pn and define X n = ni=1 Xni /n.
Prove that if 2 (0 ) < and 2 () is continuous at the point 0 , then

n[X n (n )] N (0, 2 (0 ))

as n .
Hint:


|a + b|2+ 22+ |a|2+ + |b|2+ .
Exercise 8.6 Suppose X1 , X2 , . . . are independent exponential random variables with
mean . Consider the test of H0 : = 1 vs H1 : > 1 in which we reject H0 when
u
Xn 1 + ,
n
where = .05.
(a) Derive an asymptotic approximation to the power of the test for a fixed
sample size n and alternative . Tell where you use the result of Problem 8.5.
(b) Because the sum of independent exponential random variables is a gamma
random variable, it is possible to compute the power exactly in this case. Create
a table in which you compare the exact power of the test against the alternative
= 1.2 to the asymptotic approximation in part (a) for n {5, 10, 15, 20}.
174

Exercise 8.7 Let X1 , . . . , Xn be independent from Poisson (). Create a table in

which you list the exact power along with approximations (8.20) and (8.21) for
the test that rejects H0 : = 1 in favor of H1 : > 1 when

n(X n 0 )

u ,
0
where n = 20 and = .05, against each of the alternatives 1.1, 1.5, and 2.
Exercise 8.8 Let X1 , ..., Xn be an independent sample from an exponential distribution with mean , and Y1 , ..., Yn be an independent sample from an exponential
distribution with mean . Assume that Xi and Yi are independent. We are interested in testing the hypothesis H0 : = versus H1 : > . Consider the
statistic
n
X

Tn = 2
(Ii 1/2)/ n,
i=1

where Ii is the indicator variable Ii = I(Xi > Yi ).

(a) Derive the asymptotic distribution of Tn under the null hypothesis.
(b) Use the Lindeberg Theorem to show that, under the local alternative hypothesis (n , n ) = ( + n1/2 , ), where > 0,
Pn
(Ii n ) L
n
+ n1/2
p i=1
N (0, 1), where n =
.
=
n + n
2 + n1/2
nn (1 n )
(c) Using the conclusion of part (b), derive the asymptotic distribution of Tn
under the local alternative specified in (b).
Exercise 8.9 Suppose X1 , . . . , Xm is a simple random sample and Y1 , . . . , Yn is another simple random sample independent of the Xi , with P (Xi t) = t2 for
t [0, 1] and P (Yi t) = (t )2 for t [, + 1]. Assume m/(m + n) as
m, n and 0 < < 1.

Find the asymptotic distribution of m + n[g(Y X) g()].

8.3

The Wilcoxon Rank-Sum Test

Suppose that X1 , . . . , Xm and Y1 , . . . , Yn are two independent simple random samples, with
P (Xi t) = P (Yj t + ) = F (t)
175

(8.28)

for some continuous distribution function F (t) with f (t) = F 0 (t). Thus, the distribution of
the Yj is shifted by from the distribution of the Xi . We wish to test H0 : = 0 against
H1 : > 0.
To do the asymptotics here, we will assume that n and m are actually both elements of
separate sequences of sample sizes, indexed by a third variable, say k. Thus, m = mk and
n = nk both go to as k , and we suppress the subscript k on m and n for convenience
of notation. Suppose that we combine the Xi and Yj into a single sample of size m + n.
Define the Wilcoxon rank-sum statistic to be
Sk =

n
X

Rank of Yj among combined sample.

j=1

Letting Y(1) , . . . , Y(n) denote the order statistics for the sample of Yj as usual, we may rewrite
Sk in the following way:
Sk =
=

n
X
j=1
n
X

Rank of Y(j) among combined sample

j + #{i : Xi < Y(j) }

j=1
n

n(n + 1) X X
+
I{Xi < Y(j) }
=
2
j=1 i=1
n(n + 1) X X
=
+
I{Xi < Yj }.
2
j=1 i=1

(8.29)

Let N = m + n, and suppose that m/N as k for some constant (0, 1). First,
we will establish the asymptotic behavior of Sk under the null hypothesis. Define
p
() = E Sk and () = N Var Sk .
P
To evaluate (0 ) and (0 ), where 0 = 0, let Zi = nj=1 I{Xi < Yj }. Then the Zi are
identically distributed but not independent, and we have E 0 Zi = n/2 and
n
+ n(n 1) Cov 0 (I{Xi < Y1 }, I{Xi < Y2 })
4
n n(n 1) n(n 1)
=
+

4
3
4
n(n + 2)
=
.
12

Var 0 Zi =

176

Furthermore,
E 0 Zi Zj =

n X
n
X

P0 (Xi < Yr and Xj < Ys ) =

r=1 s=1

n(n 1)
+ nP0 (Xi < Y1 and Xj < Y1 ),
4

which implies
Cov 0 (Zi , Zj ) =

n(n 1) n n2
n
+
= .
4
3
4
12

Therefore,
(0 ) =

n(n + 1) mn
n(N + 1)
+
=
2
2
2

and
2 (0 ) = N m Var Z1 + N m(m 1) Cov (Z1 , Z2 )
= [N mn(n + 2) + N m(m 1)n] /12
= [N mn(N + 1)] /12.

Let 1 , 2 , . . . be a sequence of alternatives such that N (k 0 ) for a positive, finite

constant . It is possible to show that

N {Sk (0 )} d
N (0, 1)
(8.30)
(0 )
under H0 (see Exercise 8.10) and

N {Sk (k )} d
N (0, 1)
(k )

(8.31)

under the alternatives {k }. As in the case of the signed-rank test of the previous section, we
will not prove the asymptotic normality under the sequence of alternatives here because it
is not a direct consequence of any of the results we have seen thus far. Yet it may be proven
using the Hoeffding projection idea described in Chapter 10, by which Sk may be expressed
as a sum of independent random variables plus an asymptotically negligible term.
By expression (8.30), the test based on Sk withasymptotic level rejects H0 : = 0 in favor
of Ha : > 0 whenever Sk (0 ) + u (0 )/ N , or
r
n(N + 1)
mn(N + 1)
Sk
+ u
.
2
12
177

The Wilcoxon rank-sum test is sometimes referred to as the Mann-Whitney test. This
alternative name helps to distinguish this test from the similarly named Wilcoxon signed
rank test.
To find the limiting power of the rank-sum test, we may use Theorem 8.9 to conclude that


0 (0 )
k (k ) lim
u .
(8.32)
k
(0 )
According to expression (8.32), we must evaluate 0 (0 ). To this end, note that
P (X1 < Y1 ) = E {P (X1 < Y1 | Y1 )}
= E F (Y1 )
Z
F (y)f (y ) dy
=

Z
=
F (y + )f (y) dy.

Therefore,
d
P (X1 < Y1 ) =
d

f (y + )f (y) dy.

This gives
0

(0) = mn

f 2 (y) dy.

Thus, the efficacy of the Wilcoxon rank-sum test is

R
Z
mn 12 f 2 (y) dy p
0 (0 )
lim
= lim p
= 12(1 )
f 2 (y) dy.
k (0 )
k
mnN (N + 1)

The asymptotic power of the test follows immediately from (8.32).

Exercises for Section 8.3

Exercise 8.10 Prove expression (8.30) under the null hypothesis H0 : = 0.
Hint: Verify either the Lindeberg condition or the Lyapunov condition.

178

Exercise 8.11 Suppose Var Xi = Var Yi = 2 < and we wish to test the hypotheses H0 : = 0 vs. H1 : > 0 using the two-sample Z-statistic
Y X
q
.
1
1
m+n
Note that this Z-statistic is s/ times the usual T-statistic, where s is the pooled
sample standard deviation, so the asymptotic properties of the T-statistic are the
same as those of the Z-statistic.
(a) Find the efficacy of the Z test. Justify your use of Theorem 8.9.
(b) Find the ARE of the Z test with respect to the rank-sum test for normally
distributed data.
(c) Find the ARE of the Z test with respect to the rank-sum test if the data
1 |t/|
e
.
come from a double exponential distribution with f (t) = 2
(d) Prove by example that the ARE of the Z-test with respect to the rank-sum
test can be arbitrarily close to zero.

179

Chapter 9
Pearsons chi-square test
9.1

Null hypothesis asymptotics

Let X1 , X2 , be independent from a multinomial(1, p) distribution, where p is a k-vector

with nonnegative entries that sum to one. That is,
P (Xij = 1) = 1 P (Xij = 0) = pj for all 1 j k

(9.1)

and each Xi consists of exactly k1 zeros and a single one, where the one is in the component
of the success category at trial i. Note that the multinomial distribution is a generalization
of the binomial distribution to the case in which there are k categories of outcome instead
of only 2.
The purpose of this section is to derive the asymptotic distribution of the Pearson chi-square
statistic
2 =

k
X
(nj npj )2
,
np
j
j=1

(9.2)

where nj is the random variable nX j , the number of successes in the jth category for trials
1, . . . , n. In a real application, the true value of p is not known, but instead we assume
that p = p0 for some null value p0 . We will show that 2 converges in distribution to the
chi-square distribution on k 1 degrees of freedom, which yields to the familiar chi-square
test of goodness of fit for a multinomial distribution.
Equation (9.1) implies that Var Xij = pj (1 pj ). Furthermore, Cov (Xij , Xi` ) = E Xij Xi`

180

pj p` = pj p` for j 6= `. Therefore, the random vector Xi has covariance matrix

p1 (1 p1 )
p1 p2

p1 pk
p1 p2
p2 (1 p2 )
p2 pk

=
.
..
..
.
.

.
.
.
p1 pk
p2 pk
pk (1 pk )

(9.3)

Since E Xi = p, the central limit theorem implies

d

n Xn p Nk (0, ).

(9.4)

Note that the sum of the jth column of is pj pj (p1 + + pk ) = 0, which is to say that
the sum of the rows of is the zero vector, so is not invertible.
We now present two distinct derivations of this asymptotic distribution of the 2 statistic
in equation (9.2), because each derivation is instructive. One derivation avoids dealing with
the singular matrix , whereas the other does not.
In the first approach, define for each i Yi = (Xi1 , . . . , Xi,k1 ). That is, let Yi be the k 1vector consisting of the first k 1 components of Xi . Then the covariance matrix of Yi is
the upper-left (k 1) (k 1) submatrix of , which we denote by . Similarly, let p
denote the vector (p1 , . . . , pk1 ).
One may verify that is invertible and that
1
1
+ p1k
p1
pk
1
1

+
pk
p2

( )1 =
..

1
pk

1
pk

1
pk

...

1
pk
1
pk

..
.
1
pk1
+

1
pk

(9.5)

Furthermore, the 2 statistic of equation (9.2) by be rewritten as

2 = n(Y p )> ( )1 (Y p ).

(9.6)

The facts in Equations (9.5) and (9.6) are checked in Problem 9.2. If we now define

Zn = n( )1/2 (Y p ),
d

then the central limit theorem implies Zn Nk1 (0, I). By definition, the 2k1 distribution
is the distribution of the sum of the squares of k 1 independent standard normal random
variables. Therefore,
d

2 = (Zn )> Zn 2k1 ,

181

(9.7)

which is the result that leads to the familiar chi-square test.

In a second approach to deriving the limiting distribution (9.7), we use some properties of
projection matrices.
Definition 9.1 A symmetric matrix P is called a projection matrix if it is idempotent;
that is, if P 2 = P .
The following lemmas, to be proven in Problem 9.3, give some basic facts about projection
matrices.
Lemma 9.2 Suppose P is a projection matrix. Then every eigenvalue of P equals 0
or 1. Suppose that r denotes the number of eigenvalues of P equal to 1. Then if
Z Nk (0, P ), Z> Z 2r .
Lemma 9.3 The trace of a square matrix M , Tr (M ), is equal to the sum of its
diagonal entries. For matrices A and B whose sizes allow them to be multiplied
in either order, Tr (AB) = Tr (BA).
Recall (Lemma 4.8) that if a square matrix M is symmetric, then there exists an orthogonal
matrix Q such that QM Q> is a diagonal matrix whose entries consist of the eigenvalues of
M . By Lemma 9.3, Tr (QM Q> ) = Tr (Q> QM ) = Tr (M ), which proves yet another lemma:
Lemma 9.4 If M is symmetric, then Tr (M ) equals the sum of the eigenvalues of M .
Define = diag (p), and let be defined as in Equation (9.3). Equation (9.4) implies

n1/2 (X p) Nk (0, 1/2 1/2 ).

Since may be written in the form pp> ,

1/2 1/2 = I 1/2 pp> 1/2 = I

>
p p

(9.8)

has trace k 1; furthermore,

(I p p> )(I p p> ) = I 2 p p> + p p> p p> = I p p>

because p> p = 1, so the covariance matrix (9.8) is a projection matrix.

Define An = n1/2 (X p). Then we may check (in problem 9.3) that
2 = (An )> An .

(9.9)

Therefore, since the covariance matrix (9.8) is a projection with trace k 1, Lemma 9.4 and
d
Lemma 9.2 prove that 2 2k1 as desired.
182

Exercises for Section 9.1

Exercise 9.1 Hotellings T 2 . Suppose X(1) , X(2) , . . . are independent and identically
distributed from some k-dimensional distribution with mean and finite nonsingular covariance matrix . Let Sn denote the sample covariance matrix
n

Sn

1 X (j)
=
(X X)(X(j) X)> .
n 1 j=1

To test H0 : = 0 against H1 : 6= 0 , define the statistic

T 2 = (V(n) )> Sn1 (V(n) ),
where V(n) =

n(X 0 ). This is called Hotellings T 2 statistic.

[Notes: This is a generalization of the square of a unidimensional t-statistic. If the

sample is multivariate normal, then [(n k)/(nk k)]T 2 is distributed as Fk,nk .
A Pearson chi square statistic may be shown to be a special case of Hotellings
T 2. ]
P

(a) You may assume that Sn1 1 , which follows from the Weak Law of Large
Numbers since P (Sn is nonsingular) 1. Prove that under the null hypothesis,
d
T 2 2k .
(b) An approximate 1 confidence set for based on the result in part (a)
may be formed by plotting the elliptical set
{ : n(X )> Sn1 (X ) = c },
where c is defined by the equation tP (2k > c ) = . For a random sample of
size 100 from N2 (0, ), where


1 3/5
=
,
3/5 1
produce a scatterplot of the sample and plot 90% and 99% confidence sets on this
scatterplot.
Hints: In part (b), to produce a random vector with the N2 (0, ) distribution,
take a N2 (0, I) random vector and left-multiply by a matrix A such that AA> =
. It is not hard to find such an A (it may be taken to be lower triangular). One
way to graph the ellipse is to find a matrix B such that B > Sn1 B = I. Then note
that
{ : n(X )> Sn1 (X ) = c } = {X B : > = c /n},
183

so it remains only to find points , closely spaced, such that > equals a constant. To find a matrix B such as the one specified, note that the matrix of
eigenvectors of Sn , properly normalized, gives an orthogonal matrix that diagonalizes.
Exercise 9.2 Verify Equations (9.5) and (9.6).
Exercise 9.3 Prove Lemma 9.2 and Lemma 9.3, then verify Equation (9.9).
Exercise 9.4 Pearsons chi-square for a 2-way table: Product multinomial model.
If A and B are categorical variables with 2 and k levels, respectively, and we
collect random samples of size m and n from levels 1 and 2 of A, then classify
each individual according to its level of the variable B, the results of this study
may be summarized in a 2 k table. The standard test of the independence of
variables A and B is the Pearson chi-square test, which may be written as
X
all cells in

(Oj Ej )2
,
Ej
table

where Oj is the observed count in cell j and Ej is the estimate of the expected
count under the null hypothesis. Equivalently, we may set up the problem as
follows: If X and Y are independent Multinomial(m, p) and Multinomial (n, p)
random vectors, respectively, then the Pearson chi-square statistic is
W


k 
X
(Xj mZj /N )2 (Yj nZj /N )2
+
,
=
mZ
/N
nZ
/N
j
j
j=1

where Z = X + Y and N = n + m. (Note: I used W2 to denote the chi-square

statistic to avoid using yet another variable that looks like an X.)
Prove that if N in such a way that n/N (0, 1), then
d

W2 2k1 .
Exercise 9.5 Pearsons chi-square for a 2-way table: Multinomial model. Now consider the case in which (X, Y) is a single multinomial (N, q) random 2k-vector.
Xi will still denote the (1, i) entry in a 2 k table, and Yi will still denote the
(2, i) entry.
(a) In this case, q is a 2k-vector. Let = q1 /(q1 + qk+1 ) and define p to be the
k-vector such that (q1 , . . . , qk ) = p. Prove that under the usual null hypothesis
that variable A is independent of variable B (i.e., the row variable and the column
variable are independent), q = (p, (1 )p) and p1 + + pk = 1.
184

(b) As in Problem 9.4, let Z = X + Y. Assume the null hypothesis is true and
suppose that for some reason is known. The Pearson chi-square statistic may
be written as

k 
X
(Xj Zj )2 (Yj (1 )Zj )2
2
W =
+
.
(9.10)
Z
(1

)Z
j
j
j=1
Find the joint asymptotic distribution of


p
X1
Y1
Xk
Yk
N (1 )

,...,

N N (1 )
N n(1 )
d

and use this result to prove that W2 2k .

Exercise 9.6 In Problem 9.5(b), it was assumed that was known. However, in most
problems this assumption is unrealistic.
Therefore, we replace all occurrences of
Pk
in Equation (9.10) by
= i=1 Xi /N . This results in a different asymptotic
2
distribution for the W statistic. Suppose we are given the following multinomial
probabilities for a 2 2 table with independent row and column variables:
P (X1 = 1) = .1 P (X2 = 1) = .15 .25
P (Y1 = 1) = .3 P (Y2 = 1) = .45 .75
.4
.6
1
Note that = .25 in the above table. Let N = 50 and simulate 1000 multinomial
random vectors with the above probabilities. For each, calculate the value of
W2 using both the known value = .25 and the value
estimated from the
data. Plot the empirical distribution function of each of these two sets of 1000
values. Compare with the theoretical distribution functions for the 21 and 22
distributions.
Hint: To generate a multinomial random variable with expectation vector
matching the table above, because of the independence inherent in the table you
can generate two independent Bernoulli random variables with respective success
probabilities equal to the margins: That is, let P (A = 2) = 1 P (A = 1) = .6
and P (B = 2) = 1 P (B = 1) = .75, then classify the multinomial observation
into the correct cell based on the random values of A and B.
Exercise 9.7 The following example comes from genetics. There is a particular characteristic of human blood (the so-called MN blood group) that has three types:
M, MN, and N. Under idealized circumstances known as Hardy-Weinberg equi2
librium, these three types occur in the population with probabilities p1 = M
,
185

2
p2 = 2M N , and p3 = N
, respectively, where M is the frequency of the M
allele in the population and N = 1 M is the frequency of the N allele.

We observe data X1 , . . . , Xn , where Xi has one of three possible values: (1, 0, 0)T ,
(0, 1, 0)T , or (0, 0, 1)T , depending on whether the ith individual has the M, MN,
or N blood type. Denote the total number of individuals of each of the three
types by n1 , n2 , and n3 ; in other words, nj = nX j for each j.
If the value of M were known, then the results of this section would show that
the Pearson 2 statistic converges in distribution to a chi-square distribution on
2 degrees of freedom. However, of course we usually dont know M . Instead,
we estimate it using the maximum likelihood estimator
M = (2n1 + n2 )/2n.
=
By the invariance principle of maximum likelihood estimation, this gives p
2
2 T
(
M
, 2
M
N ,
N
) as the maximum likelihood estimator of p.

). Use the delta method to derive the asymptotic

(a) Define Bn = n(X p
1/2
distribution of
Bn , where = diag(p1 , p2 , p3 ).
to be the diagonal matrix with entries p1 , p2 , p3 along its diagonal.
(b) Define
1/2 Bn .
Derive the asymptotic distribution of
(c) Derive the asymptotic distribution of the Pearson chi-square statistic

3
X
(nj n
pj )2
=
.
n
pj
j=1

(9.11)

Exercise 9.8 Take M = .75 and n = 100 in the situation described in Problem 9.7.
Simulate 500 realizations of the data.
(a) Compute
3
X
(nj npj )2
npj
j=1

for each of your 500 datasets. Compare the empirical distribution function of
these statistics with both the 21 and 22 distribution functions. Comment on
what you observe.
(b) Compute the 2 statistic of Equation (9.11) for each of your 500 datasets.
Compare the empirical distribution function of these statistics with both the 21
and 22 distribution functions. Comment on what you observe.

186

9.2

Power of Pearsons chi-square test

Suppose the k-vector X is distributed as multinomial (n, p), and we wish to test the null
hypothesis H0 : p = p0 against the alternative H1 : p 6= p0 using the
chi-square
Pearson
(n)
(n)
test. We are given a sequence of specific alternatives
p
satisfying
n(p

p)
for
Pk
some constant matrix . Note that this means i=1 i = 0, a fact that will be used later.
Our task is to derive the limit of the power of the sequence of tests under the sequence of
alternatives p(n) .
The notion of a noncentral chi-square distribution will be important in this development, so
we first give a definition.
Definition 9.5 If A1 , . . . , An are independent random variables with Ai N (i , 1),
then the distribution of A21 + A22 + + A2n is noncentral chi-square with n degrees
of freedom and noncentrality parameter = 21 + + 2n . (In particular, the
distribution depends on the i only through .) We denote this distribution
2n (). Equivalently, we can say that if A Nn (, I), then A> A 2n () where
= > .
In some references, the noncentrality parameter is defined to be equal to > /2. The form
of the actual parameter is not important, though it is of course necessary to know in a
particular context which parameterization is used.
Actually, Definition (9.5) is not a valid definition unless we may prove that the distribution
of A> A depends on only through = > . We prove this as follows.
First, note that

if = 0 then there is nothing to prove. Otherwise, define = / . Next, find an

orthogonal matrix Q whose first row is ( )> . (It is always possible to do this, though we do
not explain the details here. One method is the process of Gram-Schmidt
orthogonalization).

Then QA Nk (Q, I). Since Q is a vector with first element and remaining elements
0, QA has a distribution that depends on only through . But A> A = (QA)> (QA),
proving that the distribution of A> A depends on the i only through .
We will derive the power of the chi-square test by adapting the projection matrix technique
of Section 9.1. First, we prove a lemma that generalizes Lemma 9.2.
Lemma 9.6 Suppose Z Nk (, P ), where P is a projection matrix of rank r k
and P = . Then Z> Z 2r (> ).
Proof: Since P is a covariance matrix, it is symmetric, which means that there exists an
orthogonal matrix Q with QP Q1 = diag (), where is the vector of eigenvalues of P .
Since P is a projection matrix, all of its eigenvalues are 0 or 1. Since P has rank r, exactly
r of the eigenvalues are 1. Without loss of generality, assume that the first r entries of are
1 and the last k r are 0. The random vector QZ is Nn (Q, diag ()), which implies that
187

Z> Z = (QZ)> (QZ) is by definition distributed as 2r () + , where

=

r
X

(Q)2i

and =

i=1

k
X

(Q)2i .

i=r+1

Note, however, that

Q = QP = QP Q> Q = diag ()Q.

(9.12)

Since entries r + 1 through k of are zero, the corresponding entries of Q must be zero
because of Equation (9.12). This implies two things: First, = 0; and second,
=

r
X
i=1

(Q)2i

k
X
=
(Q)2i = (Q)> (Q) = > .
i=1

Thus, Z> Z 2r (> ), which proves the result.

Define = diag (p0 ). Let = p0 (p0 )> be the usual multinomial covariance matrix

d
under the null hypothesis; i.e., n(X(n) /n p0 ) Nk (0, ) if X(n) multinomial(n, p0 ).
Consider X(n) to have
a multinomial (n, p(n) ) distribution. Under the assumption
instead
made earlier that n(p(n) p0 ) , it may be shown that

d
n(X(n) /n p(n) ) Nk (0, ).

(9.13)

We claim that the limit (9.13) implies that the chi square statistic n(X(n) /np0 )> 1 (X(n) /n
p0 ) converges in distribution to 2k1 ( > 1 ), a fact that we now prove.
First, recall that we
have already shown that 1/2 1/2 is a projection matrix of rank
k 1. Define V(n) = n(X(n) /n p0 ). Then

V(n) = n(X(n) /n p(n) ) + n(p(n) p0 ).

The first term on the right hand side converges in distribution to Nk (0, ) and the second
d
term converges to . Therefore, Slutskys theorem implies that V(n) Nk (, ), which gives
d

1/2 V(n) Nk (1/2 , 1/2 1/2 ).

Thus, if we can show that (1/2 1/2 )(1/2 ) = (1/2 ), then the result we wish to
prove follows from Lemma 9.6. But
1/2 1 = 1/2 [ p0 (p0 )> ]1 = 1/2 [ p0 (1)> ] = 1/2
Pk
since 1> =
i=1 i = 0. Thus, we conclude that the chi-square statistic converges in
distribution to 2k1 ( > 1 ) under the sequence of alternatives p(1) , p(2) , . . ..
188

Example 9.7 For a particular trinomial experiment with n = 200, suppose the null
hypothesis is H0 : p = p0 = ( 41 , 12 , 14 ). (This hypothesis might arise in the
context of a genetics experiment.) We may calculate the approximate power of
the Pearson chi-square test at level = 0.01 against the alternative p = ( 31 , 13 , 31 ).

1
1
First, set = n(p p0 ) = 200( 12
, 61 , 12
). Under the alternative p, the chi
square statistic is approximately noncentral 22 with noncentrality parameter


4
2
4
200
>
0 1
diag (p ) = 200
+
+
=
.
144 36 144
9
Since the test rejects H0 whenever the statistic is larger than the .99 quantile of
) > 9.210} = 0.965.
22 , namely 9.210, the power is approximated by P {22 ( 200
9
These values were found using R as follows:
> qchisq(.99,2)
[1] 9.21034
> 1-pchisq(.Last.value, 2, ncp=200/9)
[1] 0.965006

Exercises for Section 9.2

Exercise 9.9 Suppose we have a tetranomial experiment and wish to test the hypothesis H0 : p = (1/4, 1/4, 1/4, 1/4) against the alternative H1 : p 6= (1/4, 1/4, 1/4, 1/4)
at the .05 level.
(a) Approximate the power of the test against the specific alternative (1/10, 2/10, 3/10, 4/10)
for a sample of size n = 200.
(b) Give the approximate sample size necessary to give power of 80% against
the alternative in part (a).

Exercise 9.10 In Exercise 9.1, let {(n) } be alternatives such that n((n) 0 )
. You may assume that under {(n) },

n(X (n) ) Nk (0, ).

Find (with proof) the limit of the power against the alternatives {(n) } of the
test that rejects H0 : = 0 when T 2 c , where P (2k > c ) = .

189

Chapter 10
U-statistics
When one is willing to assume the existence of a simple random sample X1 , . . . , Xn , Ustatistics generalize common notions of unbiased estimation such as the sample mean and
the unbiased sample variance (in fact, the U in U-statistics stands for unbiased). Even
though U-statistics may be considered a bit of a special topic, their study in a large-sample
theory course has side benefits that make them valuable pedagogically. The theory of Ustatistics nicely demonstrates the application of some of the large-sample topics presented
thus far. Furthermore, the study of U-statistics enables a theoretical discussion of statistical
functionals, which gives insight into the common modern practice of bootstrapping.

10.1

Statistical Functionals and V-Statistics

Let S be a set of cumulative distribution functions and let T denote a mapping from S into
the real numbers R. Then T is called a statistical functional. If, say, we are given a simple
random sample from a distribution with unknown distribution function F , we may want
to learn the value of = T (F ) for a (known) functional T . In this way, we may think of
the value of a statistical functional as a parameter we wish to estimate. Some particular
instances of statistical functionals are as follows:
If T (F ) = F (c) for some constant c, then T is a statistical functional mapping each F
to PF (Y c).
If T (F ) = F 1 (p) for some constant p, where F 1 (p) is defined in Equation (3.18),
then T maps F to its pth quantile.
If T (F ) = E F (Y ) or T (F ) = Var F (Y ), then T maps F to its mean or its variance 2 ,
190

respectively.
Suppose X1 , . . . , Xn is an independent and identically distributed sequence in other words,
a simple random sample with distribution function F (x). We define the empirical distribution function Fn to be the distribution function for a discrete uniform distribution on
{X1 , . . . , Xn }. In other words,
n

1X
1
I{Xi x}.
Fn (x) = #{i : Xi x} =
n
n i=1
Since Fn (x) is a legitimate distribution function, a reasonable estimator of T (F ) is the socalled plug-in estimator T (Fn ). For example, if T (F ) = E F (Y ), then the plug-in estimator
given a simple random sample X1 , X2 , . . . from F is
n

1X
T (Fn ) = E Fn (Y ) =
Xi = X n .
n i=1

(10.1)

In Equation (10.1), Y is a random variable whose distribution is the same as the empirical
distribution of the data, which means that the true population mean of Y equals the sample
mean of X1 , . . . , Xn . This equation illustrates how we distinguish between the notational
use of X and Y in this chapter: We use X and Xi whenever we must refer specifically to
the data X1 , . . . , Xn ; but we use Y and Yi whenever we refer generally to a functional and
no specific reference to the data is made.
As we will see later, a plug-in estimator, such as X n above, is also known as a V-statistic or a
V-estimator when the functional T (F ) is of a particular type called an expectation functional.
Suppose that for some real-valued function (y), we define T (F ) = E F (Y ). In this case,
we find
T {F1 + (1 )F2 } = E F1 (Y ) + (1 ) E F2 (Y ) = T (F1 ) + (1 )T (F2 ).
For this reason, such a functional is sometimes called a linear functional; see Definition 10.1.
To generalize this idea, we consider a real-valued function taking more than one real argument, say (y1 , . . . , ya ) for some a > 1, and define
T (F ) = E F (Y1 , . . . , Ya ),

(10.2)

which we take to mean the expectation of (Y1 , . . . , Ya ) where Y1 , . . . , Ya is a simple random sample from the distribution function F . Letting denote some permutation mapping {1, . . . , a} onto itself, the fact that the Yi are independent and identically distributed
191

means that the joint distribution of (Y1 , . . . , Ya ) is the same as the joint distribution of
(Y(1) , . . . , Y(a) ). Therefore,
E F (Y1 , . . . , Ya ) = E F (Y(1) , . . . , Y(a) ).
Since there are a! such permutations, consider the function
X
def 1
(y1 , . . . , ya ) =
(y(1) , . . . , y(a) ).
a! all
Since E F (Y1 , . . . , Ya ) = E F (Y1 , . . . , Ya ) and is symmetric in its arguments, we see
that in Equation (10.2) we may assume without loss of generality that is symmetric in
its arguments. In other words, (y1 , . . . , ya ) = (y(1) , . . . , y(a) ) for any permutation of
the integers 1 through a. A function defined as in Equation (10.2) is called an expectation
functional, as summarized in the following definition:
Definition 10.1 For some integer a 1, let : Ra R be a function symmetric in its a arguments. The expectation of (Y1 , . . . , Ya ) under the assumption
that Y1 , . . . , Ya are independent and identically distributed from some distribution F will be denoted by E F (Y1 , . . . , Ya ). Then the functional T (F ) =
E F (Y1 , . . . , Ya ) is called an expectation functional. If a = 1, then T is also
called a linear functional.
Expectation functionals are important in this chapter because they are precisely the functionals that give rise to V-statistics and U-statistics. The function (y1 , . . . , ya ) in Definition
10.1 is used so frequently that we give it a special name:
Definition 10.2 Let T (F ) = E F (Y1 , . . . , Ya ) be an expectation functional, where
: Ra R is a function that is symmetric in its arguments. Then is called the
kernel function associated with T (F ).
Suppose T (F ) is an expectation functional defined according to Equation (10.2). If we have
a simple random sample of size n from F , then as noted earlier, a natural way to estimate
T (F ) is by the use of the plug-in estimator T (Fn ). This estimator is called a V-estimator or a
V-statistic. It is possible to write down a V-statistic explicitly: Since Fn assigns probability
1
to each Xi , we have
n
Vn

n
n
X
1 X

(Xi1 , . . . , Xia ).
= T (Fn ) = E Fn (Y1 , . . . , Ya ) = a
n i =1 i =1
1

(10.3)

In the case a = 1, Equation (10.3) becomes

n

Vn =

1X
(Xi ).
n i=1
192

(10.4)

It is clear in Equation (10.4) that E Vn = T (F ), which we denote by . Furthermore, if

2 = Var F (Y ) < , then the central limit theorem implies that

d
n(Vn ) N (0, 2 ).
For a > 1, however, the sum in Equation (10.3) contains some terms in which i1 , . . . , ia
are not all distinct. The expectation of such terms is not necessarily equal to = T (F )
because in Definition 10.1, requires a independent random variables from F . Thus, Vn is
not necessarily unbiased for a > 1.
Example 10.3 Let a = 2 and (y1 , y2 ) = |y1 y2 |. It may be shown (Problem 10.2)
that the functional T (F ) = E F |Y1 Y2 | is not linear in F . Furthermore, since
|Yi1 Yi2 | is identically zero whenever i1 = i2 , it may also be shown that the
V-estimator of T (F ) is biased:
n1
1 XX
EF |Yi Yj | =
T (F )
E Vn = 2
n
n
i6=j
because there are exactly n(n 1) pairs (i, j) for which i 6= j.
Since the bias in Vn is due to the duplication among the subscripts i1 , . . . , ia , one way to
correct this bias is to restrict the summation in Equation (10.3) to sets of subscripts i1 , . . . , ia
that contain no duplication. For example, we might sum instead over all possible subscripts
satisfying i1 < < ia . The result is the U-statistic, which is the topic of Section 10.2.

Exercises for Section 10.1

Exercise 10.1 Let X1 , . . . , Xn be a simple random sample from F . For a fixed t for
which 0 < F (t) < 1, find the asymptotic distribution of Fn (t).
Exercise 10.2 Let T (F ) = E F |Y1 Y2 |. Show that T (F ) is not a linear functional
by exhibiting distributions F1 and F2 and a constant (0, 1) such that
T {F1 + (1 )F2 } =
6 T (F1 ) + (1 )T (F2 ).
Exercise 10.3 Let X1 , . . . , Xn be a random sample from a distribution F with finite
third absolute moment.
(a) For a = 2, find (y1 , y2 ) such that E F (Y1 , Y2 ) = Var F Y . Your function
should be symmetric in its arguments.
Hint: The fact that = E Y12 E Y1 Y2 leads immediately to a non-symmetric
function. Symmetrize it.
193

(b) For a = 3, find (y1 , y2 , y3 ) such that E F (Y1 , Y2 , Y3 ) = E F (Y E F Y )3 .

As in part (a), should be symmetric in its arguments.

10.2

Asymptotic Normality

Recall that X1 , . . . , Xn are independent and identically distributed random variables. Because the V-statistic
Vn =

n
n
X
1 X

(Xi1 , . . . , Xia )
na i =1 i =1
a

is in general a biased estimator of the expectation functional T (F ) = E F (Y1 , . . . , Ya ) due

to the presence of summands in which there are duplicated indices on the Xik , one way to
produce an unbiased estimator is to sum only over those (i1 , . . . , ia ) in which no duplicates
occur. Because is assumed to be symmetric in its arguments, we may without loss of
generality restrict attention to the cases in which 1 i1 < < ia n. Doing this, we
obtain the U-statistic Un :
Definition 10.4 Let a be a positive integer and let (y1 , . . . , ya ) be the kernel function
associated with an expectation functional T (F ) (see Definitions 10.1 and 10.2).
Then the U-statistic corresponding to this functional equals
Un =

X
1 X

(Xi1 , . . . , Xia ),
n
a

(10.5)

1i1 <<ia n

where X1 , . . . , Xn is a simple random sample of size n a.

The U in U-statistic stands for unbiased (the V in V-statistic stands for von Mises,
who was one of the originators of
this theory in the late 1940s). The unbiasedness of Un
follows since it is the average of na terms, each with expectation T (F ) = E F (Y1 , . . . , Ya ).
Example 10.5 Consider a random sample X1 , . . . , Xn from F , and let
Rn =

n
X

jI{Wj > 0}

j=1

be the Wilcoxon signed rank statistic, where W1 , . . . , Wn are simply X1 , . . . , Xn

reordered in increasing absolute value. We showed in Example 8.12 that
Rn =

n X
i
X

I{Xi + Xj > 0}.

i=1 j=1

194

Letting (a, b) = I{a + b > 0}, we see that is symmetric in its arguments and
thus it is a legitimate kernel function for an expectation functional. We find that
n

1
1 X




I{Xi > 0} = Un + OP ( n1 ) ,
n Rn = Un + n
2

i=1

where Un is the U-statistic corresponding to the expectation functional T (F ) =

PF (Y1 + Y2 > 0). Therefore, some asymptotic properties of the signed rank test
that we have already derived elsewhere can also be obtained using the theory of
U-statistics.
In the special case a = 1, the V-statistic and the U-statistic coincide. In this case, we have
already seen that both Un and Vn are asymptotically normal by the central limit theorem.
However, for a > 1, the two statistics do not coincide in general. Furthermore, we may no
longer use the central limit theorem to obtain asymptotic normality because the summands
are not independent (each Xi appears in more than one summand).
To prove the asymptotic normality of U-statistics, we shall use a method sometimes known as
the H-projection method after its inventor, Wassily Hoeffding. If (y1 , . . . , ya ) is the kernel
function of an expectation functional T (F ) = E F (Y1 , . . . , Ya ), suppose X1 , . . . , Xn is a
simple random sample from the distribution F . Let = T (F ) and let Un be the U-statistic
defined in Equation (10.5). For 1 k a, suppose that the values of Y1 , . . . , Yk are held
constant, say, Y1 = y1 , . . . , Yk = yk . This may be viewed as projecting the random vector
(Y1 , . . . , Ya ) onto the (a k)-dimensional subspace in Ra given by {(y1 , . . . , yk , ck+1 , . . . , ca ) :
(ck+1 , . . . , ca ) Rak }. If we take the conditional expectation, the result will be a function
of y1 , . . . , yk , which we will denote by k . To summarize, for k = 1, . . . , a we shall define
k (y1 , . . . , yk ) = E F (y1 , . . . , yk , Yk+1 , . . . , Ya ).

(10.6)

Equivalently, we may use conditional expectation notation to write

k (Y1 , . . . , Yk ) = E F {(Y1 , . . . , Ya ) | Y1 , . . . , Yk } .

(10.7)

From Equation (10.7), we see that E F k (Y1 , . . . , Yk ) = E F (Y1 , . . . , Ya ) = for all k.

The variances of the k functions will be useful in what follows. Therefore, we introduce
new notation, letting
k2 = Var F k (Y1 , . . . , Yk ).

(10.8)

The importance of the k2 values, particularly 12 , is seen in the following theorem, which
gives a closed-form expression for the variance of a U-statistic:

195

Theorem 10.6 The variance of a U-statistic is


a  
1 X a na 2
Var F Un = n

.
k
ak k
a k=1
If 12 , . . . , a2 are all finite, then
a2 12
Var F Un =
+O
n

1
n2


.

Theorem 10.6 is proved in Exercise 10.4. This theorem shows that

the variance of nUn tends

to a2 12 , and indeed we may well wonder whether it is true that n(Un ) is asymptotically
normal with this limiting variance. It is the goal of Hoeffdings H-projection method to prove
exactly that fact.
We shall derive the asymptotic normality of Un in a sequence of steps. The basic idea will
be to show that Un has the same limiting distribution as the sum
Un =

n
X

E F (Un | Xj )

(10.9)

j=1

of projections. The asymptotic distribution of Un follows from the central limit theorem
because Un is the sum of independent and identically distributed random variables.
Lemma 10.7 For all 1 j n,
E F (Un | Xj ) =

a
{1 (Xj ) } .
n

Proof: Expanding Un using the definition (10.5) gives

E F (Un | Xj ) =

X
1 X

E F {(Xi1 , . . . , Xia ) | Xj } ,
n
a

1i1 <<ia n

where from equation (10.7) we see that

if j {i1 , . . . , ia }
otherwise.


The number of ways to choose {i1 , . . . , ia } so that j is among them is n1
, so we obtain
a1


n1
a
a1

{1 (Xj ) } =
E F (Un | Xj ) =
{1 (Xj ) } .
n
n
a
E F {(Xi1 , . . . , Xia ) | Xj } =

1 (Xj )
0

196

Lemma 10.8 If 12 < and Un is defined as in Equation (10.9), then

d
nUn N (0, a2 12 ).
Proof: Lemma 10.8 follows immediately from Lemma 10.7 and the central limit theorem
since a1 (Xj ) has mean a and variance a2 12 .
Now that we know the asymptotic distribution of Un , it remains to show that Un and
Un have the same asymptotic behavior.
Lemma 10.9
n
o
E F Un (Un ) = E F Un2 .
Proof: By Equation (10.9) and Lemma 10.7, E F Un2 = a2 12 /n. Furthermore,
n
n
o
aX

E F Un (Un ) =
E F {(1 (Xj ) )(Un )}
n j=1
n

aX
=
E F E F {(1 (Xj ) )(Un ) | Xj }
n j=1
n
a2 X
E F {1 (Xj ) }2
=
2
n j=1

a2 12
,
n
where the third equality above follows from Lemma 10.7.
=

Lemma 10.10 If k2 < for k = 1, . . . , a, then



P
n Un Un 0.
Proof: Since convergence in quadratic mean implies convergence in probability (Theorem
2.17), it suffices to show that
n
o2
EF
n(Un Un ) 0.


2

By Lemma 10.9, n E F Un Un = n Var F Un E F Un2 . But n Var F Un = a2 12 +
O(1/n) by Theorem 10.6, and n E F Un2 = a2 12 , proving the result.

Finally, since n(Un ) = nUn + n(Un Un ), Lemmas 10.8 and 10.10 along with
Slutskys theorem result in the theorem we originally set out to prove:
197

Theorem 10.11 If k2 < for k = 1, . . . , a, then

n(Un ) N (0, a2 12 ).

(10.10)

Example 10.12 Consider the expectation functional defined by the kernel function
(y1 , y2 ) = (y1 y2 )2 /2. We obtain
T (F ) = E F (Y1 , Y2 ) = E F (Y12 + Y22 2Y1 Y2 )/2 = EF Y 2 (E F Y )2 = Var F Y.
Given a simple random sample X1 , . . . , Xn from F , let us derive the asymptotic
distribution of the associated U-statistic. First, we obtain
1 XX
Un = n

(Xi , Xj ) =
2

1i<jn

XX
1
(Xi , Xj ),
n(n 1) i=1 j=1

where we have used the fact that (Xi , Xj ) = 0 in this example whenever i = j
in order to allow both i and j to range from 1 to n. Continuing, we obtain
n

Un

XX
1
=
(X 2 + Xj2 2Xi Xj )2 /2
n(n 1) i=1 j=1 i
n

X
X
XX
1
1
1
=
Xi2 +
Xj2
Xi Xj .
2(n 1) i=1
2(n 1) j=1
n(n 1) i=1 j=1
P P
2
By observing that X n = i j Xi Xj /n2 , we may now conclude that
" n
#
n
X
1
1 X
2
2
Un =
Xi nX n =
(Xi X n )2 ,
n 1 i=1
n 1 i=1
which is the usual unbiased sample variance.
We have already derived the asymptotic distribution of the sample variance
twice! in Examples 4.11 and 5.9, though we used a biased version of the sample
variance in each of those examples. Now, we may obtain the same result a third
time using the theory of U-statistics we just developed. Since a = 2 here, we
know that

d
n(Un 2 ) N (0, 412 ).
It remains to find 12 . To this end, we must define 1 (y). Letting = E F Y and
2 = Var F Y , we obtain
1 (y) = E F (y, Y2 ) = E F (y 2 2yY2 + Y22 )/2 = (y 2 2y + 2 + 2 )/2.
198

Therefore (since adding the constant 2 does not change the variance),
1
1
12 = Var F 1 (Y ) = Var F (Y 2 2Y + 2 ) = Var F (Y )2 .
4
4
We conclude that

d
n(Un 2 ) N [0, Var F (Y )2 ].
This confirms the results obtained in Examples 4.11 and 5.9.

Exercises for Section 10.2

Exercise 10.4 Prove Theorem 10.6, as follows:
(a) Prove that for 1 k a,
E F (Y1 , . . . , Ya )(Y1 , . . . , Yk , Ya+1 , . . . , Ya+(ak) ) = k2 + 2
and thus Cov F {(Y1 , . . . , Ya ), (Y1 , . . . , Yk , Ya+1 , . . . , Ya+(ak) )} = k2 .
Hint: Use conditioning! In this case, it makes sense to condition on Y1 , . . . , Yk
because conditional on those random variables, the expression above is the product of independent realizations of k .
(b) Show that
 
n
Var F
Un =
a
 X

a  
n
a na
a

k=1

ak

Cov F {(X1 , . . . , Xa ), (X1 , . . . , Xk , Xa+1 , . . . , Xa+(ak) )}

and then use part (a) to prove the first equation of theorem 10.6.
(c) Verify the second equation of theorem 10.6.
Exercise 10.5 Suppose a kernel function (y1 , . . . , ya ) satifies E F |(Yi1 , . . . , Yia )| <
for any (not necessarily distinct) i1 , . . . , ia . Prove that if Un and Vn are the
corresponding U- and V-statistics for a simple random sample X1 , . . . , Xn , then

P
n(Vn Un ) 0 so that Vn has the same asymptotic distribution as Un .
Hint: Verify and use the equation

X
X
1
Vn Un = Vn a

(Xi1 , . . . , Xia )
n all i distinct
j
"
#
X
X
1
1


+ a

(Xi1 , . . . , Xia ).
n
a! na all i distinct
j

199

Exercise 10.6 For the kernel function of Example 10.3, (a, b) = |a b|, the corresponding U-statistic is called Ginis mean difference and it is denoted Gn . For a
random sample from uniform(0, ), find the asymptotic distribution of Gn .
Exercise 10.7 Let (y1 , y2 , y3 ) have the property
(a + by1 , a + by2 , a + by3 ) = (y1 , y2 , y3 )sgn(b) for all a, b. (10.11)
Let = E (Y1 , Y2 , Y3 ). The function sgn(b) is defined as the sign of b, which
may be expressed as I{b > 0} I{b < 0}.
(a) We define the distribution F to be symmetric if for Y F , there exists some
(the center of symmetry) such that Y and Y have the same distribution.
Prove that if F is symmetric then = 0.
(b) Let y and y denote the mean and median of y1 , y2 , y3 . Let (y1 , y2 , y3 ) =
sgn(y y). Show that this function satisfies criterion (10.11), then find the
asymptotic distribution for the corresponding U-statistic if F is the standard
uniform distribution.
Exercise 10.8 If the arguments of the kernel function (y1 , . . . , ya ) of a U-statistic
are vectors instead of scalars, note that Theorem 10.11 still applies with no
modification. With this in mind, consider for y, z R2 the kernel (y, z) =
I{(y1 z1 )(y2 z2 ) > 0}.
(a) Given a simple random sample X1 , . . . , Xn , if Un denotes the U-statistic
corresponding to the kernel above, the statistic 2Un 1 is called Kendalls tau
statistic. Suppose the marginal distributions of Xi1 and Xi2 are both
continuous,
with Xi1 and Xi2 independent. Find the asymptotic distribution of n(Un )
for an appropriate value of .
(b) To test the null hypothesis that a sample W1 , . . . , Wn is independent and
identically distributed against the alternative hypothesis that the Wi are stochastically increasing in i, suppose we reject the null hypothesis if the number of pairs
(Wi , Wj ) with Wi < Wj and i < j is greater than cn . This test is called Manns
test against trend. Based on your answer to part (a), find cn so that the test has
asymptotic level .05.
(c) Estimate the true level of the test in part (b) for a simple random sample
of size n from a standard normal distribution for each n {5, 15, 75}. Use 5000
samples in each case.
Exercise 10.9 Suppose that X1 , . . . , Xn is a simple random sample from a uniform(0, )
distribution. For some fixed a, let Un be the U-statistic associated with the kernel
200

function
(y1 , . . . , ya ) = max{y1 , . . . , ya }.
Find the asymptotic distribution of Un .

10.3

Multivariate and multi-sample U-statistics

In this section, we generalize the idea of U-statistics in two different directions. First, we
consider single U-statistics for situations in which there is more than one sample. Next, we
consider the joint asymptotic distribution of two (single-sample) U-statistics.
We begin by generalizing the idea of U-statistics to the case in which we have more than
one random sample. Suppose that Xi1 , . . . , Xini is a simple random sample from Fi for all
1 i s. In other words, we have s random samples, each potentially from a different
distribution, and ni is the size of the ith sample. We may define a statistical functional
= E (Y11 , . . . , Y1a1 ; Y21 , . . . , Y2a2 ; ; Ys1 , . . . , Ysas ) .

(10.12)

Notice that the kernel in Equation (10.12) has a1 + a2 + + as arguments; furthermore,

we assume that the first a1 of them may be permuted without changing the value of , the
next a2 of them may be permuted without changing the value of , etc. In other words, there
are s distinct blocks of arguments of , and is symmetric in its arguments within each
of these blocks. Finally, notice that in Equation (10.12), we have dropped the subscripted
F on the expectation operator used in the previous section, when we wrote E F this is
because there are now s different distributions, F1 through Fs , and writing E F1 ,...,Fs would
make a bad notational situation even worse!
Letting N = n1 + + ns denote the total sample size, the U-statistic corresponding to the
expectation functional (10.12) is
X
X


1
1

X1i1 , . . . , X1ia1 ; ; Xsr1 , . . . , Xsras .

(10.13)
UN = n1
ns

a1

as

1i1 <<ia1 n1

1r1 <<ras ns

As we did in the case of single-sample U-statistics, define for 0 k1 ai , . . . , 0 ks as

k1 ks (Y11 , . . . , Y1k1 ; ; Ys1 , . . . , Ysks ) =
E {(Y11 , . . . , Y1a1 ; ; Ys1 , . . . , Ysas ) | Y11 , . . . , Y1k1 , , Ys1 , . . . , Ysks }

(10.14)

and
k21 ks = Var k1 ks (Y11 , . . . , Y1k1 ; ; Ys1 , . . . , Ysks ).
201

(10.15)

By an argument similar to the one used in the proof of Theorem 10.6, but much more tedious
notationally, we can show that
k21 ks = Cov {(Y11 , . . . , Y1a1 ; ; Ys1 , . . . , Ysas ),
(Y11 , . . . , Y1k1 , Y1,a1 +1 , . . . ; ; Ys1 , . . . , Xsks , Ys,as +1 , . . .)} .

(10.16)

Notice that some of the ki may equal 0. This was not true in the single-sample case, since
0 would have merely been the constant , so 02 would have been 0.
In the special case when s = 2, Equations (10.14), (10.15) and (10.16) become
jk (Y1 , . . . , Yj ; Z1 , . . . , Zk ) = E {(Y1 , . . . , Ya1 ; Z1 , . . . , Za2 ) | Y1 , . . . , Yj , Z1 , . . . , Zk } ,
2
jk
= Var jk (Y1 , . . . , Yj ; Z1 . . . , Zk ),
and
2
jk
= Cov {(Y1 , . . . , Ya1 ; Z1 , . . . , Za2 ),

(Y1 , . . . , Yj , Ya1 +1 , . . . , Ya1 +(a1 j) ; Z1 , . . . , Zk , Za2 +1 , . . . , Za2 +(a2 k) ) ,
respectively, for 0 j a1 and 0 k a2 .
Although we will not derive it here as we did for the single-sample case, there is an analagous
asymptotic normality result for multisample U-statistics, as follows.
Theorem 10.13 Suppose that for i = 1, . . . , s, Xi1 , . . . , Xini is a random sample
from the distribution Fi and that these s samples are independent of each other.
Suppose further that there exist constants 1 , . . . , s in the interval (0, 1) such
that ni /N i for all i and that a21 as < . Then

N (UN ) N (0, 2 ),

where
2 =

a21 2
a2 2
1000 + + s 0001
.
1
s

Although the notation required for the multisample U-statistic theory is nightmarish, life
becomes considerably simpler in the case s = 2 and a1 = a2 = 1, in which case we obtain
n1 X
n2
1 X
UN =
(X1i ; X2j ).
n1 n2 i=1 j=1

202

Equivalently, we may assume that X1 , . . . , Xm are a simple random sample from F and
Y1 , . . . , Yn are a simple random sample from G, which gives
m

UN =

1 XX
(Xi ; Yj ).
mn i=1 j=1

(10.17)

In the case of the U-statistic of Equation (10.17), Theorem 10.13 states that


2
2

01
10
d
+
,
N (UN ) N 0,

1
2
2
= Cov {(X1 ; Y1 ), (X1 ; Y2 )}, and 01
= Cov {(X1 ; Y1 ), (X2 ; Y1 )}.
where = lim m/N , 10

Example 10.14 For independent random samples X1 , . . . Xm from F and Y1 , . . . , Yn

from G, consider the Wilcoxon rank-sum statistic W , defined to be the sum of
the ranks of the Yi among the combined sample. We may show that
m

XX
1
W = n(n + 1) +
I{Xi < Yj }.
2
i=1 j=1
Therefore, if we let (a; b) = I{a < b}, then the corresponding two-sample Ustatistic UN is related to W by W = 12 n(n + 1) + mnUN . Therefore, we may use
Theorem 10.13 to obtain the asymptotic normality of UN , and therefore of W .
However, we make no assumption here that F and G are merely shifted versions
of one another. Thus, we may now obtain in principle the asymptotic distribution
of the rank-sum statistic for any two distributions F and G that we wish, so long
as they have finite second moments.
The other direction in which we will generalize the development of U-statistics is consideration of the joint distribution of two single-sample U-statistics. Suppose that there are
two kernel functions, (y1 , . . . , ya ) and (y1 , . . . , yb ), and we define the two corresponding
U-statistics
X
1 X

(Xi1 , . . . , Xia )
Un(1) = n

a

1i1 <<ia n

and
Un(2) =

X
1 X

(Xj1 , . . . , Xjb )
n
b

1j1 <<jb n
(1)

(2)

for a single random sample X1 , . . . , Xn from F . Define 1 = E Un and 2 = E Un .

Furthermore, define jk to be the covariance between j (Y1 , . . . , Yj ) and k (Y1 , . . . , Yk ), where
j and k are defined as in Equation (10.7). Letting ` = min{j, k}, it may be proved that


jk = Cov (Y1 , . . . , Ya ), (Y1 , . . . , Y` , Ya+1 , . . . , Ya+(b`) ) .
(10.18)
203

Note in particular that jk depends only on the value of min{j, k}.

(1)

The following theorem, stated without proof, gives the joint asymptotic distribution of Un
(2)
and Un .
Theorem 10.15 Suppose X1 , . . . , Xn is a random sample from F and that : Ra
R and : Rb R are two kernel functions satisfying Var (Y1 , . . . , Ya ) < and
Var (Y1 , . . . , Yb ) < . Define 12 = Var 1 (Y1 ) and 22 = Var 1 (Y1 ), and let jk
be defined as in Equation (10.18). Then
(
 )
   2 2

(1) 

Un
1
0
a 1 ab11
d
n
N
,
.
(2)
ab11 b2 22
2
0
Un

Exercises for Section 10.3

Exercise 10.10 Suppose X1 , . . . , Xm and Y1 , . . . , Yn are independent random samples
from distributions Unif(0, ) and Unif(, + ), respectively. Assume m/N
as m, n and 0 < < .
(a) Find the asymptotic distribution of the U-statistic of Equation (10.17), where
(x; y) = I{x < y}. In so doing, find a function g(x) such that E (UN ) = g().
(b) Find the asymptotic distribution of g(Y X).
(c) Find the range of values of for which the Wilcoxon estimate of g() is
asymptotically more efficient than g(Y X). (The asymptotic relative efficiency
in this case is the ratio of asymptotic variances.)
Exercise 10.11 Solve each part of Problem 10.10, but this time under the assumptions that the independent random samples X1 , . . . , Xm and Y1 , . . . , Yn satisfy
P (X1 t) = P (Y1 t) = t2 for t [0, 1] and 0 < < 1. As in Problem
10.10, assume m/N (0, 1).
Exercise 10.12 Suppose X1 , . . . , Xm and Y1 , . . . , Yn are independent random samples
from distributions N (0, 1) and N (, 1), respectively. Assume m/(m + n) 1/2
as m, n . Let UN be the U-statistic of Equation (10.17), where (x; y) =
I{x < y}. Suppose that () and 2 () are such that

N [UN ()] N [0, 2 ()].

Calculate () and 2 () for {.2, .5, 1, 1.5, 2}.

204

Hint: This problem requires a bit of numerical integration. There are a couple
of ways you might do this. A symbolic mathematics program like Mathematica
or Maple will do it. There is a function called integrate in R and Splus and
one called quad in MATLAB for integrating a function. If you cannot get any of
these to work for you, let me know.
Exercise 10.13 Suppose X1 , X2 , . . . are independent and identically distributed with
finite variance. Define
n

Sn2

1 X
(xi x)2
=
n 1 i=1

and let Gn be Ginis mean difference, the U-statistic defined in Problem 10.6.
Note that Sn2 is also a U-statistic, corresponding to the kernel function (x1 , x2 ) =
(x1 x2 )2 /2.
(a) If Xi are distributed as Unif(0, ), give the joint asymptotic distribution of
Gn and Sn by first finding the joint asymptotic distribution of the U-statistics
Gn and Sn2 . Note that the covariance matrix need not be positive definite; in this
problem, the covariance matrix is singular.
(b) The singular asymptotic covariance matrix in this problem implies that as
n , the joint distribution of Gn and Sn becomes concentrated on a line. Does
this appear to be the case? For 1000 samples of size n from Uniform(0, 1), plot
scatterplots of Gn against Sn . Take n {5, 25, 100}.

10.4

Introduction to the Bootstrap

This section does not use very much large-sample theory aside from the weak law of large
numbers, and it is not directly related to the study of U-statistics. However, we include
it here because of its natural relationship with the concepts of statistical functionals and
plug-in estimators seen in Section 10.1, and also because it is an increasingly popular and
often misunderstood method in statistical estimation.
Consider a statistical functional Tn (F ) that depends on n. For instance, Tn (F ) may be some
property, such as bias or variance, of an estimator n of = (F ) based on a random sample
of size n from some distribution F .


As an example, let (F ) = F 1 21 be the median of F . Take n to be the mth order statistic
from a random sample of size n = 2m 1 from F .
Consider the bias TnB (F ) = E F n (F ) and the variance TnV (F ) = E F n2 (E F n )2 .
205

Theoretical properties of TnB and TnV are very difficult to obtain. Even asymptotics arent

d
very helpful, since n(n ) N {0, 1/(4f 2 ())} tells us only that the bias goes to zero
and the limiting variance may be very hard to estimate because it involves the unknown
quantity f (), which is hard to estimate.
Consider the plug-in estimators TnB (Fn ) and TnV (Fn ). (Recall that Fn denotes the empirical
distribution function, which puts a mass of n1 on each of the n sample points.) In our median
example,
TnB (Fn ) = E Fn n n
and
TnV (Fn ) = E Fn (n )2 (E Fn n )2 ,
where n is the sample median from a random sample X1 , . . . , Xn from Fn .
To see how difficult it is to calculate TnB (Fn ) and TnV (Fn ), consider the simplest nontrivial
case, n = 3: Conditional on the order statistics (X(1) , X(2) , X(3) ), there are 27 equally likely
possibilities for the value of (X1 , X2 , X3 ), the sample of size 3 from Fn , namely
(X(1) , X(1) , X(1) ), (X(1) , X(1) , X(2) ), . . . , (X(3) , X(3) , X(3) ).
Of these 27 possibilities, exactly 1 + 6 = 7 have the value X(1) occurring 2 or 3 times.
Therefore, we obtain
7
13
7
P (n = X(1) ) = , P (n = X(2) ) = , and P (n = X(3) ) = .
27
27
27
This implies that
1
1
2
2
2
E Fn n =
(7X(1) + 13X(2) + 7X(3) ) and E Fn (n )2 =
(7X(1)
+ 13X(2)
+ 7X(3)
).
27
27
Therefore, since n = X(2) , we obtain
1
TnB (Fn ) = (7X(1) 14X(2) + 7X(3) )
27
and
TnV (Fn ) =

14
2
2
2
(10X(1)
+ 13X(2)
+ 10X(3)
13X(1) X(2) 13X(2) X(3) 7X(1) X(3) ).
729

To obtain the sampling distribution of these estimators, of course, we would have to consider
the joint distribution of (X(1) , X(2) , X(3) ). Naturally, the calculations become even more
difficult as n increases.
206

Alternatively, we could use resampling in order to approximate TnB (Fn ) and TnV (Fn ). This is
the bootstrapping idea, and it works like this: For some large number B, simulate B random
samples from Fn , namely

X11
, . . . , X1n
,
..
.

XB1 , . . . , XBn
,

and approximate a quantity like E Fn n by the sample mean

B
1 X
,
B i=1 in

. Notice that the

, . . . , Xin
is the sample median of the ith bootstrap sample Xi1
where in
weak law of large numbers asserts that
B
1 X P
E Fn n .
B i=1 in

To recap, then, we wish to estimate some parameter Tn (F ) for an unknown distribution F

based on a random sample from F . We estimate Tn (F ) by Tn (Fn ), but it is not easy to
evaluate Tn (Fn ) so we approximate Tn (Fn ) by resampling B times from Fn and obtain a

bootstrap estimator TB,n

. Thus, there are two relevant issues:

1. How good is the approximation of Tn (Fn ) by TB,n

? (Note that Tn (Fn ) is NOT an
unknown parameter; it is known but hard to evaluate.)

2. How precise is the estimation of Tn (F ) by Tn (Fn )?

Question 1 is usually addressed using an asymptotic argument using the weak law or the
central limit theorem and letting B . For example, if we have an expectation functional
Tn (F ) = E F h(X1 , . . . , Xn ), then

TB,n
=

B
1 X
P

h(Xi1
, . . . , Xin
) Tn (Fn )
B i=1

as B .
Question 2, on the other hand, is often tricky; asymptotic results involve letting n and
are handled case-by-case. We will not discuss these asymptotics here. On a related note,
207

however, there is an argument in Lehmanns book (on pages 432433) about why a plug-in
estimator may be better than an asymptotic estimator. That is, if it is possible to show
Tn (F ) T as n , then as an estimator of Tn (F ), Tn (Fn ) may be preferable to T .
We conclude this section by considering the so-called parametric bootstrap. If we assume that
the unknown distribution function F comes from a family of distribution functions indexed
by a parameter , then Tn (F ) is really Tn (F ). Then, instead of the plug-in estimator Tn (Fn ),
we might consider the estimator Tn (F ), where
is an estimator of .
Everything proceeds as in the nonparametric version of bootstrapping. Since it may not be
easy to evaulate Tn (F ) explicitly, we first find
and then take B random samples of size

n, X11
, . . . , X1n
through XB1
, . . . , XBn
, from F . These samples are used to approximate
Tn (F ).
Example 10.16 Suppose X1 , . . . , Xn is a random sample from Poisson(). Take
=
X. Suppose Tn (F ) = Var F
. In this case, we happen to know that Tn (F ) =
/n, but lets ignore this knowledge and apply a parametric bootstrap. For some
large B, say 500, generate B samples from Poisson(
) and use the sample variance

of
as an approximation to Tn (F ). In R, with = 1 and n = 20 we obtain
x <- rpois(20,1) # Generate the sample from F
muhat <- mean(x)
muhat
[1] 0.85
muhatstar <- rep(0,500) # Allocate the vector for muhatstar
for(i in 1:500) muhatstar[i] <- mean(rpois(20,muhat))
var(muhatstar)
[1] 0.04139177
Note that the estimate 0.041 is close to the known true value 0.05. This example
is simplistic because we already know that Tn (F ) = /n, which makes
/n a more
natural estimator. However, it is not always so simple to obtain a closed-form
expression for Tn (F ).
Incidentally, we could also use a nonparametric bootstrap approach in this example:
for (i in 1:500) muhatstar2[i] <- mean(sample(x,replace=T))
var(muhatstar2)
[1] 0.0418454
Of course, 0.042 is an approximation to Tn (Fn ) rather than Tn (F ). Furthermore,
we can obtain a result arbitrarily close to Tn (Fn ) by increasing the value of B:

208

muhatstar2_rep(0,100000)
for (i in 1:100000) muhatstar2[i] <- mean(sample(x,replace=T))
var(muhatstar2)
[1] 0.04136046
In fact, it is in principle possible to obtain an approximate variance for our
estimates of Tn (Fn ) and Tn (F ), and, using the central limit theorem, construct
approximate confidence intervals for these quantities. This would allow us to
specify the quantities to any desired level of accuracy.

Exercises for Section 10.4

Exercise 10.14 (a) Devise a nonparametric bootstrap scheme for setting confidence
intervals for in the linear regression model Yi = + xi + i . There is more
than one possible answer.
(b) Using B = 1000, implement your scheme on the following dataset to obtain a
95% confidence interval. Compare your answer with the standard 95% confidence
interval.
Y
x

21
460

16
498

20
512

34
559

33
614

43
675

47
719

(In the dataset, Y is the number of manatee deaths due to collisions with powerboats in Florida and x is the number of powerboat registrations in thousands for
even years from 1978-1990.)
Exercise 10.15 Consider the following dataset that lists the latitude and mean August temperature in degrees Fahrenheit for 7 US cities. The residuals are listed
for use in part (b).
City
Miami
Phoenix
Memphis
Baltimore
Pittsburgh
Boston
Portland, OR

Latitude
26
33
35
39
40
42
46

Temperature
83
92
81
76
71
72
69

Residual
-5.696
10.116
1.062
-0.046
-4.073
-1.127
-0.235

Minitab gives the following output for a simple linear regression:

209

Predictor
Constant
latitude

Coef
113.99
-0.9730

S = 5.546

SE Coef
13.01
0.3443

R-Sq = 61.5%

T
8.76
-2.83

P
0.000
0.037

R-Sq(adj) = 53.8%

Note that this gives an asymptotic estimate of the variance of the slope parameter
as .34432 = .1185.
In (a) through (c) below, use the described method to simulate B = 500 boot

) for 1 b B. For each b, refit the model

), . . . , (xb7 , yb7
strap samples (xb1 , yb1

to obtain b . Report the sample variance of 1 , . . . , B and compare with the

asymptotic estimate of .1185.

(a) Parametric bootstrap. Take xbi = xi for all b and i. Let ybi
= 0 +1 xi +i ,
where i N (0,
2 ). Obtain 0 , 1 , and
2 from the above output.

(b) Nonparametric bootstrap I. Take xbi = xi for all b and i. Let ybi
=

0 + 1 xi + rbi , where rb1 , . . . , rb7 is an iid sample from the empirical distribution
of the residuals from the original model (you may want to refit the original model
to find these residuals).

(c) Nonparametric bootstrap II. Let (xb1 , yb1

), . . . , (xb7 , yb7
) be an iid sample
from the empirical distribution of (x1 , y1 ), . . . , (x7 , y7 ).

Note: In R or Splus, you can obtain the slope coefficient of the linear regression
of the vector y on the vector x using lm(y~x)$coef[2].
Exercise 10.16 The same resampling idea that is exploited in the bootstrap can be
used to approximate the value of difficult integrals by a technique sometimes
called Monte Carlo integration. Suppose we wish to compute
Z 1
2
=2
ex cos3 (x) dx.
0

(a) Use numerical integration (e.g., the integrate function in R and Splus) to
verify that = 1.070516.
2

(b) Define g(t) = 2et cos3 (t). Let U1 , . . . , Un be an iid uniform(0,1) sample.
Let
n
1X

1 =
g(Ui ).
n i=1
P
Prove that 1 .

210

(c) Define h(t) = 2 2t. Prove that if we take Vi = 1

is a random variable with density h(t). Prove that with

Ui for each i, then Vi

1 X g(Vi )
2 =
,
n i=1 h(Vi )
P
we have 2 .

(d) For n = 1000, simulate 1 and 2 . Give estimates of the variance for each
estimator by reporting
2 /n for each, where
2 is the sample variance of the g(Ui )
or the g(Vi )/h(Vi ) as the case may be.
(e) Plot, on the same set of axes, g(t), h(t), and the standard uniform density
for t [0, 1]. From this plot, explain why the variance of 2 is smaller than the
variance of 1 . [Incidentally, the technique of drawing random variables from a
density h whose shape is close to the function g of interest is a variance-reduction
technique known as importance sampling.]
Note: This was sort of a silly example, since numerical methods yield an exact value for . However, with certain high-dimensional integrals, the curse of
dimensionality makes exact numerical methods extremely time-consuming computationally; thus, Monte Carlo integration does have a practical use in such
cases.

211